text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
arXiv:1404.4304v1 [cs.CE] 16 Apr 2014
Automated Classification of Airborne Laser
Scanning Point Clouds
Christoph Waldhauser
Ronald Hochreiter
Johannes Otepka
Norbert Pfeifer
Sajid Ghuffar
Karolina Korzeniowska
Gerald Wagner
April 2014
Abstract
Making sense of the physical world has always been at the core of mapping. Up until recently, this has always dependent on using the human
eye. Using airborne lasers, it has become possible to quickly “see” more of
the world in many more dimensions. The resulting enormous point clouds
serve as data sources for applications far beyond the original mapping
purposes ranging from flooding protection and forestry to threat mitigation. In order to process these large quantities of data, novel methods
are required. In this contribution, we develop models to automatically
classify ground cover and soil types. Using the logic of machine learning, we critically review the advantages of supervised and unsupervised
methods. Focusing on decision trees, we improve accuracy by including
beam vector components and using a genetic algorithm. We find that
our approach delivers consistently high quality classifications, surpassing
classical methods.
1
Introduction
Surveying the very planet we live on has been an ongoing effort since the dawn of
mankind. From the early maps of Anatolia to modern geospatial intelligence, the
mission of any map was always to make sense of the world around us. Boosting
map drawing with the latest advances of machine learning has the potential
to largely facilitate the generation of maps and extend their usefullness into
application domains beyond path finding.
In this chapter, we present the combined efforts of academia and industry
to create a framework for the automated generation of maps. The basis for this
project is airborne laser scanning: the systematic recording and digitizing of
ground by means of laser emitted from aircraft. The resulting point clouds of
the environment are then automatically classified into ground cover types, using
supervised learning and evolutionary computation approaches.
1
This chapter is organized as follows. In a first section we describe the technical background of airborne laser scanning. Section 3 details the work related
to develop automated classification models. There we will compare the practical aspects of supervised and unsupervised approaches as well as detail the
supervised classification approach we implemented and evolutionary computation extensions to it. That section also features a description of the data set we
used to empirically test our approaches. The aspects of implementing our model
in industry applications is discussed in the subsequent section 4. We close with
some concluding remarks pointing to future research.
2
2.1
Airborne Laser Scanning Point Clouds
Measurement principle
Airborne Laser Scanning (ALS) is a remote sensing method for obtaining geometrical and additional information about objects not in contact with the sensor,
i.e. the laser scanner. A laser scanner emits a short pulse of infrared light which
travels through the atmosphere and is scattered and partially absorbed by any
objects in the instantaneous field of view of the laser beam. If diffuse reflection occurs, which is the standard case for many object surfaces, including, e.g.,
vegetation, bare ground, and building surfaces, a portion of the incident light
is scattered back to the sensor. There, the backscattered signal is detected and
recorded. The time lag between emission of the pulse and detection of its echo
is the two way travel time from the sensor to the object. With the known speed
of light this time lag is turned into the distance from sensor to object. This is
also called laser range finding (LRF). In laser scanning, the beam is scanned
across the entire field of view, thus covering a larger extent . Rotating mirrors
and comparable devices are used to deflect the laser beam and cover large areas.
With the known orientation of the mirror and the known position of the laser
scanner in a global Earth fixed coordinate system (e.g. WGS84, in UTM projection), the location of the objects at which the laser pulse was scattered can
be computed. This provides a so-called 3D point cloud: a set of points, each
with 3 co-ordinates x, y, z. These points are obtained in the sensor co-ordinate
system.
In airborne laser scanning the scanner is mounted on a flying platform (fixed
wing or helicopter). Its position is measured with Global Navigation Satellite
Systems (GNSS, e.g., GPS). The angular attitude of the sensor platform inside
the aircraft is observed with Inertial Measurement Unit (IMU, comprised of accelerometers and gyros). The laser scanner is mounted to look downwards and
the beam is scanned at right angles to the flight direction (see Figure 11 .). Together with the forward motion of the aircraft, larger areas can be scanned. Even
larger areas are measured by flying strip-wise above the terrain. This six degree
of freedom trajectory defines a moving co-ordinate system for the observation
1 Full
color,
high resolution
http://www2.wu.ac.at/alsopt
versions
2
of
each
figure
can
be
found
at
Figure 1: Diagram explaining the principles of ALS [3]
of range and angle from the laser scanner. With an Euclidean transformation
the points can be transformed from the sensor co-ordinate system to the global
co-ordinate system. Typical results for the accuracy of such points is in the
order of 10cm (single standard deviation in each coordinate).
Besides the observation of distance between the sensor and an object point
by the time lag of emission and detection of its echo, also other observations can
be retrieved from the received echo. Firstly, it is not always the case that the
laser beam hits exactly one object. Due to the diameter of the beam, e.g., 50cm,
multiple objects may be within the beam, but at different heights. Examples
include vegetation canopy and ground below. While a part of the signal is
reflected at the leaves of the canopy of a tree, other parts of the signal continue
traveling downwards until they hit lower vegetation or the ground, from which
they are reflected. Thus, each emitted pulse may give rise to several echoes.
Other examples, next to vegetation, are power lines and house edges, where a
part of the signal is reflected on the roof, while the other part is reflected from
the ground.
Furthermore, the backscattered echo can be sampled as a function of time,
so-called waveform digitizing. The recorded amplitude depends on the range, on
laser scanner device parameters, e.g. the receiving aperture diameter, but also
on object properties, i.e., how much of the incident signal is absorbed, scattered
diffusely, etc. By means of calibration [26] the parameters of the object like
the backscatter cross-section, can be determined [27, 20]. The received echo
may also be deformed relative to the emitted pulse. An increased echo width
3
[11] is a hint for either hitting a slanted surface or, more often the case, hitting
vegetation. Within the footprint a number of leaves may be found which have
similar by not identical height. Thus the echoes from all the leaves overlap and
form a single widened echo.
2.2
Additional point descriptors
A point cloud P is a collection of points pi = (x, y, z) ∈ P in a three dimensional space. The laser scanning point cloud can be analysed locally to enhance
the description of each point further. For instance, given a point density of
4 points/m2 , a local surface model [13] can be computed using e.g. the ten nearest points. This model may be an inclined plane, with it’s normal vector being
an additional description for the point. The equation of a plane through the
point (x0 , y0 , z0 ) having a normal vector n = (a, b, c) is given as
a (x − x0 ) + b (y − y0 ) + c (z − z0 ) = 0
(1)
To compute the three components (a, b, c) of the normal vector, three equations, i.e. three (non collinear) points are required. To add robustness, generally
more than three points are used. A subset of points in the neighbourhood are
typically selected based on k-nearest neighbours or points with in the sphere of
a pre-defined radius. If k nearest neighbours are selected than there are k + 1
points and subsequently k + 1 equations (1). A least squares solution of this
overdetermined system of equations estimates an optimal plane by minimizing
squared sum of distances between the points and the estimated plane. In the
matrix form this equation system is written as
A · β = 0,
(2)
where each row of matrix A contains the coordinates of a point relative to
T
the center [xn − x0 , yn − y0 , zn − z0 ], here n = 1..k + 1 and β = [a, b, c] is the
unknown normal vector. The least squares solution for a system of equations
of this (2) form is equivalent to solving the eigenvalue problem of the matrix
AT A. The unknown normal vector β of the estimated plane is the eigenvector
corresponding to the smallest eigenvalue of AT A. The matrix AT A is often
called structure tensor [7]. The mathematical form of the structure tensor is:
AT A = T =
k+1
1X
T
(pi − p̄) (pi − p̄)
k i=1
(3)
here p̄ = (x0 , y0 , z0 ) is the center of the points in the neighborhood.
For house roofs or street surfaces the normal vectors have been shown to
reach an accuracy of a few degrees. The normal vector further allows to convert
the backscatter cross section into the so-called diffuse reflectance. This value
assumes a certain (Lambertian) scattering behavior of the object. This scattering mechanism is described by the reflectance (a unit-less value) and the normal
4
vector of the surface. A surface reflecting all incoming light perfectly diffuse has
a reflectance of 1.
The quality of the plane fitting, e.g., the root mean square distances between
the optimal plane and the given points indicates the roughness of the surface
[10]. The smallest eigenvalue of T gives the variance of the distances between
the points and the estimated plane.
The structure tensor T holds plenty more useful information about the distribution of points in the neighborhood. The geometric information encoded in
T is essential in the characterization and classification of natural and artificial
objects. Three widely used features derived from T are linearity, planarity and
omnivariance. The linearity feature reflects how well the distribution of points
can be modeled by a 3D line. Points over power lines exhibit such a characteristic, therefore, the linearity feature is essential in classifying power lines and
similar structures. The planarity describes the smoothness of the surface which
is directly related to the roughness measure and the quality of plane fitting
for normal vector estimation. In contrast to power lines and smooth surfaces,
laser echoes from trees often spread inhomogeneously across a larger 3D volume.
This volumetric point distribution is described by the concept of omnivariance.
These features are computed using the three eigen values λ1 ≥ λ2 ≥ λ3 ≥ 0 of
the matrix T :
LT =
λ1 − λ2
λ1
λ2 − λ3
λ1
p
OT = 3 λ1 λ2 λ3
PT =
(4)
(5)
(6)
In addition to LT , PT and OT , features like anisotropy, eigenentropy and
curvature are also derived using the eigenvalues of the structure tensor T [14,
8, 21, 28].
More information about the characteristics of the surfaces can be derived
using features like echo ratio, ZRange, ZRank, NormalizedZ and PointDistance.
Echo ratio represents the vertical penetration of the surface [9]. ZRange represents the maximum height difference between the points in the neighborhood,
while ZRank is the rank of the point corresponding to its height in the neighboorhood. NormalizedZ is the rank of the point (between 0 and 1) multiplied
by the height range in the neighboorhood. PointDistance is the average of all
shortest distances between the points in the neigboorhood. A more detailed
description of these features can be found in [16, 19].
Thus, the point cloud can be augmented by additional parameters besides
the coordinates x, y, z: the echo ID (first, second, . . . last echo of a sequence
of echoes) and overall length of the echo sequence, echo amplitude, echo width,
backscatter cross section, diffuse reflectance, roughness (NormalSigma), normal
vector (NormalX, NormalY, NormalZ ) echo ratio (ER), ZRange, ZRank, NormalizedZ, PointDensity, PointDistance, linearity, planarity, and omnivariance.
5
3
Classification
A major application for the automated processing of point clouds is the classification of points. In this application, every point is assigned a class due to its
inherent laser return characteristics and its derived features. If successful, any
such endeavor promises massive savings in terms of human resources and time,
and thus ultimately in cost.
In the past, the remote sensing community focused on classifying data obtained from satellite measurements [15]. They report that results in general
have been only somewhat satisfactory with large portions being continuously
misclassified. In contrast we work with air- and not satellite-borne data. This
allows for a much higher resolution and considerable less atmospheric interference when measuring. Further, the used full waveform data contains much
more information than traditional approaches using laser solely for range measurements. Finally, the method of actively illuminating the ground with a laser
beam is superior to passively recording reflections of sun light.
Classification tasks can be grouped into human and machine based classifications. Machine based classification itself can be split into knowledge- and
learning based systems. The former is today’s industry standard in ALS point
cloud processing, the latter the eventual developmental goal. The main disadvantage of knowledge-based systems over machine learning classifiers is their
requirement of explicit definitions of ontologies and classification rules. Machine learning classifiers, on the other hand base their classifications on rules
automatically deduced from the available data with minimal (or no) human intervention. A machine learning classifier with human intervention uses initial
human input to deduce automatically classification rules from it, that then can
be used to autonomously classify points of previously unseen point clouds.
When charging humans with point classification, a number of factors come
into play. Foremost, there is the need for additional data. Usually, this data is
provided by means of orthophotos that are (ideally) taken in parallel to the laser
scanning. Secondly, the qualification, endurance and accuracy of the employed
human has to be taken into consideration as well. That person needs to be an
expert user of geographical information systems and trained to recognize the
subtleties of orthophotos.
This confluence of laser scanning data, external data via orthophotos and
human experience allows for rather precise classifications of points. So far,
human performance has not been surpassed by machines in terms of accuracy.
Naturally, human classification is a very time consuming process. And equally
naturally, machines outperform humans in the time domain by many orders
of magnitude. Therefore, investigating algorithms for automated point cloud
classification is an active area of research.
When turning to learning based classification, two approaches following the
classical machine-learning dichotomy of supervised vs. unsupervised learning
come to mind. The former requires initial human classifier input to derive a
classification of unseen points, while the latter does not. The advantage of
supervised classification is, that the resulting classes correspond with target
6
classes provided through human input. Since unsupervised classification lacks
any human interaction, the classes found may or may not be interpretable or
relateable to classes that humans would come up with. In the remainder of
this section, we will focus on supervised classification based on initial human
interaction and the difficulties that arise from it.
As detailed above and elsewhere [16], point characteristics can be grouped
according to the way they were obtained: by direct measurement, by calibrated
or spatial improved measurements, by deriving them computationally, by linking
with meta data. For the former three groups, problems can arise. Directly
measured point features are subject to the specifics of the laser scanner used.
The predominant method employed for airborne laser scanning enterprises is a
laser that is being deflected off the vertical by a rotating prism or a swinging
mirror. This allows to scan a range perpendicular to the flight path and is
essential for obtaining complete laser scans. However, this method changes
the characteristics of the laser return signal, as the angle of the return signal
not only depends on the characteristics of the surface, but also the angle of
the inbound signal. For instance, when scanning directly below the aircraft
only little occlusion will occur, while at extreme scan angles, the laser beam
will be obscured by any objects between the aircraft and the ground. This
distortion needs to be taken into consideration when working with point cloud
data. Subsection 3.4 below discusses the detection of and compensation for
these effects in greater detail.
A further question that needs addressing is rooted in the way derived attributes are being computed. Many such attributes are computed taking in
account a neighborhood of points. Here, neighborhood size becomes a defining
factor. Choosing an appropriate neighborhood size is far from trivial. However,
neighborhood size theoretically affects the classification quality that can be derived. Further complications arise from different neighborhood sizes that can
be chosen for each attribute. In subsection 3.5 we present a genetic algorithm
for finding optimal neighborhood sizes for all neighborhood dependent features
involved in the classification.
Before turning to the problems described above, we will briefly introduce
the data set we worked with and describe how supervised classification works
from human and machine perspectives, respectively.
3.1
Data set and example
From the industrial side of view, the motivation for this project was to find a
new, fast and reliable algorithm for the classification of point clouds, which can
minimize the manual checking and correction, because every manual manipulation is a very time consuming task. The scenario described in subsection 4.4
below was the basis for the development of the models used to automatically
classify point clouds.
The data set used was taken from the project DGM-W Niederrhein with kind
permission of the Bundesanstalt für Gewässerkunde, Germany. Four predefined
areas have been selected, each not bigger than 60 hectares, with different content
7
like bridge, power lines, houses, coniferous and deciduous trees, concrete, gravel,
bare earth, groynes and water.
The flight was done by airplane with the use of a Riegl LMS-Q560 200 KHz
Laser Scanner. Flight speed was 100 knots at an altitude above ground of 600
meters. The distance between the flight lines was 300 meters. The effective
scanning rate was set to 150 KHz with 80 lines per second. The resulting mean
point density was about 6 points/m2 over the whole area (except water areas).
A radiometric calibration was computed using asphalt streets in each flight
session as calibration reference. The calibration parameters were then applied
to compute the reflectance, a normalized intensity value. Roughness shapes
(derived from digital orthophotos), which define different ground classes of all
areas were known and used as support.
For the classification a list of classes was discussed and defined. First a highorder list with standard classes (level 1), which are most common in the majority
of laser scanning projects was generated; then each standard class was refined
into subclasses (level 2) to better represent the different kinds of environment.
The classes used for this project can be seen in Table 1.
Following the refined class list and taking the roughness shapes into consideration, a three dimensional classification was done manually using the TerraScan
software package. This was done to provide reference data to generate training
and testing data sets for the supervised classification method. Later the manual classification was also used as gold standard for assessing the results of the
classification done with both the supervised and unsupervised methods.
3.2
Reference data generation through manual classification
The term reference data refers to data that is manually classified by humans
using external data sources like orthophotos. It serves two purposes: providing
training data for supervised classification and representing a gold standard that
can be used to test the automatic classification’s accuracy.
One method to generate reference data is a manual classification of the data
set [24, 12]. This process requires a thorough visual analysis of the data and a
labeling of each point. As this process is time consuming, typically only small
parts of the data set are manually labeled. Thus, this part should represent the
diversity of the terrain surface (flat, hilly, etc.) as well as a large amount of the
different target classes with a variety of geometric appearance and distribution of
other measured or derived objects. These target classes often comprise natural
objects (bare-earth, water, vegetation, etc.) and man-made objects (buildings,
roads, bridges, ramps, power and other transmission lines, fences, cars and
other moving objects, etc.). Vegetation, as one targeted class for example, can
be tall and low and have different density. The variety of vegetation has to be
included in the manually labeled part for both, accuracy assessment and machine
learning. The diversity of classes depends on the purpose of the classification.
Reference data generation can be performed in a number of different ways.
Firstly, an automatic classification based on a selection from available algorithms
8
Table 1: Defined classes in two levels of granularity.
First level
Second level
Class
Code
Class
Code
unclassified
0
unclassified
0
undefined
1
undefined
1
ground
2
ground
sand
gravel
stone, rock
asphalt
cement
river dam, groyne
2
18
3
4
22
21
28
vegetation
5
deciduous forest
coniferous forest
mixed forest
5
6
7
building
8
building roof
wall, building wall
8
24
water
9
water
9
10
artificial objects
10
car, other moving object
temporary object (under construction)
bridge
power line
tower, power pole
bridge cable
road protection fence
bridge construction
technical
23
technical, e.g. concrete part
of a bridge
23
ground, vegetation
20
ground, vegetation
20
error
99
error
99
9
11
12
13
14
15
16
17
can be performed, followed by a manual improvement of the results. Secondly,
only manual classification of an unclassified data set can be performed. In the
case of a large number of classes the second way is recommended. A third
option in the generation of reference data by using existing data sets. As those
data sets were often acquired at a different time, with a different measurement
technology, and often with other applications in mind, the transferability of such
a classification is limited. Therefore, the next paragraphs will concentrate on
the methods for manual classification.
The most common methods for visualization and reference data generation
are described below. The basic and most common method uses a 2D profile
(Figure 2). Profiles are sets of points cut out from the entire point cloud with
a vertical rectangular prism, not bound in height. The width of the prism is
typically small, e.g. 2 m, whereas its length is larger, e.g. 50 m. These values
are sensible when working with point densities ranging from 1 to 20 points/m2 .
Profiles allow the user to see a part of the terrain from a side view which enables
her to distinguish the points within different classes but also to identify the
border between different objects, e.g. building and ground, or vegetation and
ground. These borders are harder to identify in a top view. In order to classify
larger areas in an organized manner, transects are used. This means that a set
of parallel profiles is generated which cover a rectangular area. Advancing in the
manual classification from one profile to the next accelerates the entire process.
The second method uses a shaded relief map (hillshade) of the surface generated
by the points of one class. A hillshade requires an artificial illumination source,
which is set in a standard manner to an azimuth of 315 degrees, lighting the area
from the northwest. Lighting from different directions can substantially help to
notice the terrain slope as well as objects located on the ground, especially
in the case of mountainous regions. Hillshades can be generated for the bareearth class, in which the surface represents the digital terrain model (DTM).
An example for transects and hillshades can be found in the top panel of Figure
2. That figure’s bottom panel exhibits a DTM. Also combinations of classes,
e.g. bare-earth and buildings, can be used. This method can be applied for
refining a manual classification, i.e. reclassifying points. This is especially suited
to remove small small artifacts which occur when close spatial proximity between
two classes led to a misclassification in an earlier step.
3.3
Supervised classification
The idea behind supervised classification is to automatically derive from a small
training set enough classification rules, so that a larger, unseen data set can be
classified automatically using the model derived from the former. For that
purpose, the training data needs to be classified already. Usually, this initial
classification is achieved by manually classifying the points. This training data
is then used to build a model or equally train the classifying algorithm. In
supervised classification, the interpretation of the model comes second, therefore
more complex models are favored over simplistic ones that would ease human
interpretation; in fact, the boundary to model complexity is dictated only by
10
Figure 2: Methods for visualizing the data during manual classification; a –
shaded relief for DTM; b – shaded relief for DTM, buildings and vegetation; P1,
P2 – 2D profiles; T(P1), T(P2) – transects for 2D profiles; SA – sun azimuth
for shaded relief
11
overfitting avoidance. This model is then used to classify unseen data. To
evaluate model performance, true classification information for the unseen data
is required as well. However, in production environments, model evaluation for
the entire data set is usually not performed. Therefore, supervised classification
promises to save a considerable amount of costs.
The method of choice for supervised classification here is classification trees.
The tree is a predictive model that links up point features with that point’s class.
Structurally, the tree consists of leaves and branches. The leaves represent the
final class labels and the branches the conjunctions of features that lead up
to these class labels. Literature suggests a number of different algorithms for
growing a tree [23, 15]. For the purpose of classifying point clouds, we have found
Breiman et al’s Classification and Regression Trees (CART) [1] to strike a good
balance between computational complexity and reliability. The implementation
we used was that of rpart [25]. In terms of Friedl et al. [5] these trees are
univariate classification trees.
Conceptually, a classification tree seeks to partition the entire feature space
of a data set, one variable at a time. It does that by selecting a variable and an
appropriate splitting value that will contribute maximally to node purity. Node
purity is computed using the Gini impurity coefficient:
IG (f ) = 1 −
m
X
fi2
(7)
i=1
with fi being the fraction of items labeled to be of class i for a set of m class
labels.
This splitting and branch growing continues, until no variable can be found
that further increases node purity. The resulting trees can become quite large
which hinders interpretation (not a problem for point cloud classification) and
are prone to overfitting. This latter limitation can become troublesome when
trying to classify point clouds, as the learned model does not generalize well anymore for unseen data. However, using cross-validation and pruning off branches
that are not occurring in a significant number of replications, proves to be an
effective tool against overfitting.
As stated above, the performance of a classification tree can be gauged if
not only training but also test data contain true class labels. A measurement
statistic of classification performance is the misclassification rate. Let M be
a cross-classification matrix between true and predicted class labels and its
elements being the counts of the predicted elements and J the number of all
points in the point cloud, then
M CR = 1 − tr(M )/J
(8)
is the misclassification rate.
When selecting training data, two factors need consideration: the randomness of the selection process and its stratification. The former factor becomes
important once large sets of random numbers need to be created. While computers can always only generate pseudo random numbers, most of them are
12
sufficiently strong for point cloud processing.2 However, strong random number generation with guaranteed randomness does not suffice to select a suitable
training data set, if the classes are not evenly distributed. In that common
case, single classes—say temporary construction structures—have only very few
points associated with them. When choosing points at random, it is extremely
unlikely that many of the rare class points will end up in the training data set.
And if a class does not show up in the training data set, the supervised classification algorithm cannot learn the rules required to classify it. Therefore simple
random sampling schemes do not work in the presence of rare classes.
To enable the supervised classification of rare classes, stratified sampling
needs to be applied. In its simplest form, stratified sampling guarantees that
numerous points from each class are selected for the training data set. This,
at the expense of having the entire training data set being representative for
the point cloud it has been sampled from. The heuristic used for our stratified
sampling approach sets the size of the sample for stratum c (sc ) to be either
half of the points of that class (Sc ) or the overall sample size (k) divided by the
number of classes in the point cloud (|A|):
sc = min(
Sc k
,
)
2 |A|
(9)
As noted above, the resulting stratified sample is not representative for the
entire point cloud anymore: rare classes occur much more often in the training
data set than they do in the point cloud. It is therefore necessary to inform the
supervised classification algorithm of that misrepresentation.
Perhaps obviously, the performance of a tree depends on the number of data
points it is allowed to learn from: the larger the training data set, the better
(usually) the classification of test data will be. However, manually classifying
points is expensive. Therefore, it is crucial to find a training data set size that
is just large enough to produce reliable predictions. Figure 3 depicts this relationship. As can be seen, there is a sharp drop between 10,000 and 20,000
points as training data set size with respect to mean misclassification rate and
its dispersion. After about 50,000 points, the improvement gained by adding additional points subsides. We therefore settled for 50,000 points as training data
set size. The resulting mean misclassification rate of 0.065 is a usable starting
point. In the following, we will discuss aspects of improving this achievement
even further.
When classifying point cloud data into predetermined classes, not all classes
that appear to be epistemologically justified to humans can be sufficiently identified using laser return signals. For the problem at hand, the points were to be
partitioned into 26 classes. Logically, these classes could be broken down into
coarsely and finely grained classes. While the coarse classes were successfully
classified (MCR: 0.02, σ = 0.002), the finer classification exhibited the 6.5%
MCR as described above. Table 2 lists the finely grained classes that were no2 We
used the R [18] implementation of the Mersenne twister, which has a period of 219937 −
1.
13
Figure 3: Misclassification rate as a function of training data size; classification
of a 3 million strong point cloud, results bootstrapped with 50 replications
14
Table 2: Classes that were hard to predict. Percentage of points that ended
up in that class. Remainder to 100 percent is scatter in all classes.
True class
Predicted classes
Building, wall
Deciduous forest (67%)
Building roof (17%)
Building, wall (17%)
Temporary object
Temporary object (78%)
Road protection fence (14%)
Power pole
Power pole (75%)
Road protection fence (16%)
Error class points
Scattered in all classes
Figure 4: Predicted (left) and true (right) classification of a sample area
toriously troublesome. Figure 4 shows the differences between automatic and
true human classification results.
When casting a more detailed look at these misclassifications, it becomes
evident that many of them are conceivably caused by imprecise classifications
of humans in the first place. Consider for example a road in winter: the asphalt
tarmac is at places covered with grit sand to prevent the icing of the road.
Grit and asphalt tarmac will differ in texture and material. Therefore, the
laser return signal for patches of road that contain more grit sand than others
will exhibit different characteristics. In manual classification based on aerial
photography, these patches of grit sand are unlikely to be identified and marked
as such by the human classifier. To a certain extent, the misclassification rate
achieved by supervised classification of finely grained classes can be explained
by the algorithm outperforming human classification. This is obviously very
dependent on the quality of human classification.
A similar argument holds for the error class. Here, points were classified as
errors if some of their measurements exceeded a valid measurement range. The
15
Figure 5: Misclassification at a power line where pole and vegetation cannot
be separated reliably; automatically classified point cloud on top, bottom panel
shows the manually classified one
algorithm was informed about these missing values. On the other hand, classification trees are able to cope with missing information by substituting it with
the second best split. Therefore, points that a human would not classify because
it contained obviously faulty measurements, were classified by the algorithm.
Another problem that is rooted in the difficulty of epistemological concepts is
the misclassification of many temporary object points as road protection fences.
It is difficult for any automated classifier to learn the concept of an object
being temporary in nature. While the algorithm successfully classifies almost
all temporary objects as some kind of artificial objects, it cannot differentiate
between these objects being permanent or temporary (road protection fences).
However, the largest problem in misclassification cannot possibly be rooted
in epistemological complexities: Buildings and walls are being classified predominantly as trees. From a geometrical point of view, trees and buildings do indeed
share some properties related to their height and volume. On the other hand,
distinctive characteristics like texture and material should have been picked up
by the algorithm. This type of mistake is also represented in Figure 5. To some
extent, the misclassifications can be explained by snow or leaf covered roofs
on top of buildings. Still, this unsatisfactory performance can most likely only
be overcome by implementing geometrical shape detection in a post-processing
step. This is the focus of ongoing research.
16
Table 3: Model quality in mean MCR for models with different kinds of border
effect components. Results bootstrapped with 50 replications.
Model type
µM CR σM CR
no border effects
beam vector components
scan angle
beam vector components
and scan angle
3.4
0.081
0.065
0.074
0.004
0.005
0.005
0.063
0.004
Border effects
Airborne laser scanning is limited by the principles of optics: dependent on the
incident angle, the characteristics of a laser return signal varies. For example,
hitting vegetation from the side will produce many more laser echoes than hitting it straight from above. Also, the shape of the beam’s cross section depends
on that angle. Additional distortion in the characteristics of points may arise
from the method of aerial laser scanning. Due to the limited field of view of
airborne laser scanners wider areas are scanned by multiple overlapping strips.
Typically, these strips overlap to achieve full coverage even in case of wind
sheer or minor navigation errors. In these overlapping areas, the properties of
the measurement process change (as there are multiple overpasses); a change
that needs to be accounted for.
One method to compensate for the different return signal quality/properties
is to take the deflection of the laser into account. There are two approaches
available. One uses the raw beam vector components (vx , vy , vz ) that indicate
the deflection of the laser beam for a given point. The other method combines
these components to derive the scan angle φ:
q
vx2 + vy2
φ = arctan(
)
|vz |
The following Table 3 shows the effect beam vector components and scan angle have on the misclassification rate. Starting with the simplest model without
any compensation for border effects, the mean classification rate lies at 8.1%.
Adding the scan angle to the model improves its quality by one, beam vector
components by two percentage points. Adding both compensation terms to the
model barely improves classification quality with respect to a pure beam vector
components model.
3.5
Scale space selection
A number of point cloud features are not directly measured but computed with
respect to any points immediate neighborhood. In general, the local neighborhood of a point can be defined in 2D or 3D. Furthermore, a certain number
17
Table 4: Parameters of the genetic algorithm.
Parameter
Value
Population size
Tournament size
Mutation probability
Elite proportion
Reseed proportion
100
5
0.05
0.1
0.1
of closest neighbors, a fixed distance or a combination of both can be used as
neighborhood definition. For the following analysis a cylinder (i.e. 2D fixed
distance neighborhood) for each point is formed. Obviously, larger radii lead
to a stronger averaging effect while smaller ones are prone to overfitting. It,
therefore, is important to find the optimal radius for each feature in order to
minimize misclassification rate.
To discover the optimal radii for neighborhood-dependent features, a genetic
search algorithm [6] was used. In the following we will describe the genetic algorithm used for this optimization and its parameters. We then turn our attention
towards evaluating the algorithm’s performance in terms of convergence and solution stability. The former examines the relation of improvement achieved due
to and time spent on optimization. The latter analyzes the stability of recommended radii across a number of optimizations.
The 13 neighborhood-dependent features were computed each with radii
ranging from 1 to 6 m in 0.5 m increments resulting in 11 versions of each feature. The algorithm’s genomes were then modeled to be integer vectors of length
13 with each gene being an integer from 1 to 11, encoding the chosen neighborhood size for each feature. The algorithm was initialized with 100 random
genomes as starting solutions. The standard genetic operators of single-point
cross-over breeding and mutation were employed for evolutionary optimization.
Further, pairing genomes for mating was done using tournament selection and
a proportion of the top performing solutions was cloned directly into each new
generation. To ensure that the gene pool remained fresh and to safeguard
against local optima traps, some random genomes were introduced with each
generation. Table 4 gives the parameters of the genetic algorithm, which were
established by experiment.
The fitness function to be optimized was the misclassification rate as described above. In order to ensure comparability, MCR was computed using
the same training–test data split each time. The initial split was generated using a stratified sampling scheme and included 5137 points in the training data
set. Using a random sample of 100,000 points, the algorithm was allowed 500
generations to find the optimum combination of radii for the 13 neighborhooddependent features. In order to ensure computability within reasonable time,
not the entire point cloud could be processed. Therefore, a very large simple random sample of 100,000 points was drawn from the point cloud, and all
18
operations were performed on that sample.
As genetic optimization is essentially stochastic in nature, the optimization
was repeated 34 times. Of these 34 replications, 30 reached the same optimum
while 4 stayed behind (by a very small margin). Almost all replications had converged to the optimum after 50 generations. By generation 75 all 30 successful
replications had converged. The optimum discovered implied a misclassification
rate of 0.022. When compared to the best misclassification achieved using a
constant radius of 6 meters (0.065) this is a notable improvement by more than
60 percent.
Turning to solution stability, it is of interest whether each replication’s terminal solution leads to the same combination of radii or not. Figure 6 displays
a heat map of cylinder radii per feature chosen in each (optimal) replication.
Features that exhibit the same color shades for the entire column can be considered stable. These are the variables NormalizedZ, NormalZ and PointDensity.
For each of these features, the optimal cylinder radius is at 1 meter. At the
other end of the spectrum, very colorful columns, Linearity, Planarity and ZRange, are indicative of features whose neighborhood size has no impact on
misclassification rate.
The genetic algorithm delivers a definite improvement of the misclassification
rate. The remaining two percent are most likely due to measurement and human
classification error. With respect to solution stability, it became obvious that
while some features are computationally dependent on neighborhood size, the
outcome is not affected by them. On the other hand, there are features that
clearly exhibit a strong dependence on neighborhood size. Conceptually, the
genetic algorithm can be improved by implementing consensus voting when
delivering radii recommendations. This too, is an ongoing research effort.
We conclude that supervised classification of point clouds is definitely an idea
worthwhile pursuing. The data quality obtained from airborne laser scanning
allows for a very precise analysis of the ground. In combination with the sophisticated computation of derived point cloud features, advanced classification
algorithms sampling schemes as well as evolutionary optimization strategies, we
are able to produce classification accuracies that surpass classical satellite based
classification. While the classical approaches rarely ever reach above 90 percent
accuracy, our approach delivers consistently accuracies close to 100 percent.
While there are challenges that remain to be overcome, the achieved accuracy
is already good enough for many applications. In the following we will discuss
these applications further.
4
Industrial applications
Airborne Laser Scanning is in use for industrial purposes since the mid 1990s
and has dramatically improved since then. For expample: in the beginning
there have been laser scanners with a fixed array of fibre optical conductors,
which brought a good point density in the direction of flight, but very poor
density in the transverse direction. So a detection of embankments along the
19
Figure 6: Solution stability of 30 genetic optimization replications
20
flight direction was very hard. Technological advances like the steadily increased
measurement rates, improved apertures and new detection algorithms prepared
the way for a wide field of applications.
There are different technologies at work in today’s laser scanners: they provide sampling rates of up to 600,000 laser pulses per second. Also modern
apertures are able to detect more than just one single return per pulse and
provide reflectance, echo ID and echo width for each return; some can even penetrate water surfaces and give information on submarine ground and submerged
objects.
Higher point densities result in better environment depicting. With today’s
high point densities, embankments can be well detected by extracting breaklines
within the point clouds. Normally 4 points/m2 will be ordered, but customers
more often want 8 or more points/m2 . This gives the opportunity to model the
ground more precisely. But customers are not only interested in the presence of
ground, they also want to know what kind of ground they are looking at.
Classification is mostly a semiautomatic process, consisting of an automatic
step and a manual checking and correction step. One of the aims is to minimize
the need of manual correction, due to its cost. Another aim is to improve the
automatic detection of more than a standard set of classes to cater to future
customer’s requirements.
In the following we will present some examples of airborne laser scanning
applications.
4.1
Digital Terrain Model
Often a plain model of the ground is needed for planning or research purposes.
These models are of great importance for e.g. road- or railway planning offices,
to know how much material has to be removed or added for street or railway
planning. Therefore the point cloud has to be classified with special emphasis on
detecting erroneous echoes. The DTM classes mostly consist of ground, water
and unclassified points, which have no influence on the model.
4.2
Digital Surface Model
The DSM features ground, vegetation, buildings, bridges and sometimes power
lines and describes the earth’s surface including natural and artificial objects. By
subtracting the DTM from the DSM the result will be a normalized DSM. This
can then be used for e.g. easy measurement of building or vegetation heights.
4.3
Avalanche prediction
In mountainous areas avalanches (snow or boulders) are a common threat, so
prediction and subsequently protection is an important task. For aviation purposes it is also necessary to know the position of power lines or cable-cars.
Therefore each point needs to be classified along the lines of ground, various
vegetation, water, building, power lines, . . .
21
To compute the pathways and probabilties of avalanches in certain areas, one
not only needs to know point classes, but also inclination, roughness (in this
case roughness refers to a parameter, which will tell how fluids will be slowed
on a surface), azimuth, . . .
All these features can be derived out of the point cloud by classifying using
the above algorithm.
4.4
Flooding prediction
To protect people and environment in areas that are in danger of flooding around
rivers, it is vital to know, how water is flowing over different types of ground.
Therefore ground has to be classified in different roughness classes, that have
known properties for flowing or seeping. The classification of roughness areas is
normally done by digitizing digital orthophotos [17, 4]. In respect to the classification methods described in section 3, roughness can be set in direct relation
with different ground classes. Taking into account the derived DTM together
with the digitized breaklines [2], a triangulated surface can be computed.
By combining the DTM surface with information of the different point classes
from ground detection, there can be defined areas with varying roughness. This
classification is normally done by using digital orthophotos as reference. By
classifying the roughness purely from the data contained within a laser point
cloud, the high cost of extra orthophotos can be skipped.
4.5
Forestry and Vegetation
The detection of forested areas is an important part of environmental applications. Especially time series analyses, e.g. to estimate deforestation, was often
carried out using analog or digital orthophotos so far. However, Airborne Laser
Scanning gets more popular for such applications, because it is not restricted
to the canopy. The laser beam can often penetrate the vegetation returning
multiple echoes. This provides information about the vertical structure of the
forest including good knowledge of the ground, which is needed to compute
high quality DTMs, tree heights, stem volumes, etc. In urban areas the knowledge of classified vegetation is used in applications for 3D visualizations, urban
planning, noise emission charts, etc. [22].
5
Conclusion
In this chapter we presented an overview of advances in processing and automatically classifying point clouds from airborne laser scanning. Particularly,
the accuracy of the classification of point clouds can be improved greatly using
machine learning based methods like decision trees. There, manually classified
training data—a small subset of the entire point cloud—is used to build a classification model. This then in turn can be used to classify the remainder of the
point cloud or a fresh one.
22
These advances in classification accuracy are chiefly due to our making use
of the entire full wave form of the laser echoes. Using advanced radiometric
and computational methods, for every echo additional properties or features are
computed from that echo’s wave form, external data and the echo’s immediate
neighborhood. Using an evolutionary algorithm we were able to identify features where the size of that neighborhood influenced classification accuracy and
establish optimal neighborhood size values for these features.
The model presented in this chapter has applications ranging from forestry
to avalanche and flooding protection. A more immediate application is the automatic generation of maps. However, this is but the beginning of our journey. We
already pointed to the inclusion of shape detection for improving classification
accuracy and consensus voting the genetic algorithm to optimize neighborhood
size recommendations as current research goals. Further extensions focus on
better understanding how the scan angle affects echo properties when analyzing
the flights strip-wise. A major issue is the possibility to learn from multiple
but possibly unreliably sources. Often, orthophotos related to a point cloud are
out-of-date or older maps are used to provide external reference data. Ideally,
if we were able to use these data sources to speed up training data and model
generation, the entire remote sensing work flow could be revolutionized.
References
[1] Breiman, L., Friedman, J., Olshen, R., Stone, C.: Classification and Regression Trees. Monterey, CA (1984)
[2] Briese, C., Mandlburger, G., Ressl, C., Brockman, H.: Automatic break
line determination for the generation of a DTM along the river Main. In:
ISPRS Workshop Laserscanning 2009, pp. 236–241 (2009)
[3] Doneus, M., Briese, C., Fera, M., Janner, M.: Archaeological prospection
of forested areas using full-waveform airborne laser scanning. Journal of
Archaeological Science 35(4), 882 – 893 (2008)
[4] Dorninger, P.: Eine praktikable und genaue Methode zur Bestimmung von
Wasser-Land-Grenzen aus Laser-Scanner-Daten. BfG-Veranstaltungen 3,
93–100 (2011)
[5] Friedl, M., Brodley, C.: Decision tree classification of land cover from remotely sensed data. Remote Sensing of Environment 61(3), 399–409 (1997)
[6] Goldberg, D., Holland, J.: Genetic algorithms and machine learning. Machine Learning 3(2), 95–99 (1988)
[7] Gressin, A., Mallet, C., David, N.: Improving 3D LiDAR point cloud registration using optimal neighborhood knowledge. ISPRS Annals of Photogrammetry, Remote Sensing and Spatial Information Sciences I-3 pp.
111–116 (2012)
23
[8] Gross, H., Thoennessen, U.: Extraction of lines from laser point clouds. In:
Symposium of ISPRS Commission III: Photogrammetric Computer Vision
PCV06. International Archives of Photogrammetry, Remote Sensing and
Spatial Information Sciences, vol. 36, pp. 86–91 (2006)
[9] Höfle, B., Hollaus, M., Hagenauer, J.: Urban vegetation detection using
radiometrically calibrated small-footprint full-waveform airborne LiDAR
data. ISPRS Journal of Photogrammetry and Remote Sensing 67(1), 134–
147 (2012)
[10] Hollaus, M., Aubrecht, C., Höfle, B., Steinnocher, K., Wagner, W.: Roughness mapping on various vertical scales based on full-waveform airborne
laser scanning data. Remote Sensing 3(3), 503–523 (2011)
[11] Höfle, B., Hollaus, M., Lehner, H., Pfeifer, N., Wagner, W.: Area-based
parameterization of forest structure using full-waveform airborne laser scanning data. In: Silvilaser 2008, p. 9. Edinburgh, Scotland (2008)
[12] Kobler, A., Pfeifer, N., Orginc, P., Todorovski, L., Oštir, K., Džeroski,
S.: Repetitive interpolation: A robust algorithm for DTM generation from
aerial laser scanner data in forested terrain. Remote Sensing of Environment 108(1), 9–23 (2007)
[13] Kraus, K., Pfeifer, N.: Advanced DTM generation from LiDAR data. In:
International Archives of the Photogrammetry, Remote Sensing and Spatial
Information Sciences, pp. 23–30. Annapolis, MD, USA (2001)
[14] Mallet, C., Bretar, F., Soergel, U.: Analysis of full-waveform LiDAR data
for classification of urban areas. Photogrammetrie Fernerkundung Geoinformation 5, 337–349 (2008)
[15] Mather, P., Tso, B.: Classification Methods for Remotely Sensed Data.
CRC press (2010)
[16] Otepka, J., Ghuffar, S., Waldhauser, C., Hochreiter, R., Pfeifer, N.: Georeferenced point clouds: Data model, features and management. ISPRS
International Journal of Geo-Information 2(4), 1038–1065 (2013)
[17] Prinz, R.: DGM-W-Modellierung unter Einbeziehung erfasster Buhnen und
Buhnenfelder. BfG-Veranstaltungen 3, 48–57 (2011)
[18] R Core Team: R: A Language and Environment for Statistical Computing.
R Foundation for Statistical Computing, Vienna, Austria (2013)
[19] Research Groups Photogrammetry and Remote Sensing: OPALS: Orientation and Processing of Airborne Laser Scanning data. Department of
Geodesy and Geoinformation, TU Vienna, Vienna, Austria (2013)
24
[20] Roncat, A., Bergauer, G., Pfeifer, N.: B-spline deconvolution for differential target cross-section determination in full-waveform laser scanning data.
ISPRS Journal of Photogrammetry and Remote Sensing 66(4), 418–428
(2011)
[21] Rusu, R.: Semantic 3D object maps for everyday manipulation in human
living environments. Ph.D. thesis, München, Techn. Univ., Diss., 2009
(2009)
[22] Rutzinger, M., Höfle, B., Pfeifer, N.: Detection of high urban vegetation
with airborne laser scanning data (2007). Proceedings of ForestSat 2007
[23] Safavian, S., Landgrebe, D.: A survey of decision tree classifier methodology. IEEE Transactions on Systems, Man and Cybernetics 21(3), 660–674
(1991)
[24] Sithole, G., Vosselman, G.: Experimental comparison of filter algorithms
for bare-earth extraction from airborne laser scanning point clouds. ISPRS
Journal of Photogrammetry and Remote Sensing 59(1–2), 85–101 (2004)
[25] Therneau, T., Atkinson, B., Ripley, B.: rpart: Recursive Partitioning
(2013). R package version 4.1-3
[26] Wagner, W.: Radiometric calibration of small-footprint full-waveform airborne laser scanner measurements: Basic physical concepts. ISPRS Journal
of Photogrammetry and Remote Sensing 65(6), 505–513 (2010). ISPRS
Centenary Celebration Issue
[27] Wagner, W., Ullrich, A., Ducic, V., Melzer, T., Studnicka, N.: Gaussian
decomposition and calibration of a novel small-footprint full-waveform digitising airborne laser scanner. ISPRS Journal of Photogrammetry and Remote Sensing 60(2), 100–112 (2006)
[28] Weinmann, M., Jutzi, B., Mallet, C.: Feature relevance assessment for the
semantic interpretation of 3D point cloud data. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences II-5/W2
pp. 313–318 (2013)
25
| 5cs.CE
|
A probabilistic framework for the control of systems
with discrete states and stochastic excitation
Gianluca Meneghello∗ and Thomas Bewley
Flow Control Lab, UC San Diego, La Jolla, CA 92093-0411, USA.
Paolo Luchini
arXiv:1701.01777v1 [cs.SY] 6 Jan 2017
DIMEC, Università di Salerno, Via Ponte don Melillo, 84084 Fisciano (SA), Italy
(Dated: January 10, 2017)
A probabilistic framework is proposed for the optimization of efficient switched control strategies
for physical systems dominated by stochastic excitation. In this framework, the equation for the
state trajectory is replaced with an equivalent equation for its probability distribution function in
the constrained optimization setting. This allows for a large class of control rules to be considered,
including hysteresis and a mix of continuous and discrete random variables. The problem of steering
atmospheric balloons within a stratified flowfield is a motivating application; the same approach can
be extended to a variety of mixed-variable stochastic systems and to new classes of control rules.
INTRODUCTION
Optimal control theory is concerned with minimizing the energy required to maintain a feasible phase-space trajectory within a fixed time-average measure from a target trajectory [2]. This may be achieved by solving the constrained
optimization problem [3]
min |u|Q
u
(
|x − x̄|R = constant
with
ẋ = f (x, u) + ξ,
(1a)
(1b)
where u(x) is a given feedback control rule, x = x(t) and x̄ = x̄(t) are the actual and target trajectories in phase
space, and the additive noise term ξ models the unknown or uncertain components of the dynamical system. The
norms | · |Q and | · |R must be chosen to reflect the actual control energy and the specific measure of interest of the
system state, but are often limited to L2 or L∞ norms to make the optimization problem tractable. The present work
is motivated by the general inability of the formulation (1) to treat problems with mixed continuous and discrete
random variables, hysteretic behavior, and/or norms others than L2 or L∞ .
As a motivating application, consider a balloon in a stably stratified turbulent flowfield whose time-averaged velocity
is a function of height only, as depicted by thin arrows in Figure 1a. This is a good approximation for the radial flow
within a hurricane, as depicted in Figure 1b. The balloon’s density can be changed to control its vertical velocity
(and, hence, its altitude), and the balloon’s motion can be well approximated as the motion of a massless particle
carried by the flowfield
Ẋ = αZ + ξ,
(2a)
Ż = u(X, Z),
(2b)
where X and Z are random variables denoting the horizontal and vertical positions, α is the vertical gradient of the
time-averaged horizontal velocity (i.e., αz is the time-averaged horizontal velocity at height z), and the turbulent
fluctuations of the horizontal velocities are characterized by a white Gaussian noise ξ with zero mean and spectral
density c2 . Neglecting the vertical velocity fluctuations, the balloon moves in the horizontal direction according to a
Brownian motion with a probability distribution function (PDF) pX,Z (x, z) with horizontal mean µX (t) = αzt and
2
variance σX
(t) = c2 t. In the uncontrolled case, the variance of the balloon’s horizontal position grows linearly with
time. The vertical velocity u can then be used, leveraging the background flow stratification α, to return the balloon
to its original position.
We are thus interested in designing a control strategy u(x, z) to limit the variance of the horizontal position of the
2
balloon to a target value σ̄X
, while minimizing the control cost |u(x, z)|Q . More specifically, we consider a three-level
control (TLC) feedback rule, depicted by thick lines in Figure 1a, consisting of step-changes of altitude ±h in the
2
(b)
(a)
z
12
8
6
10 −3
4
2
−2
α=
−h
α
−d
x
z
d
(km)
10
h
1
2
3
4
5
0
−20 −15 −10
0
α=1
−5
0
5
10
Radial velocity (m/s)
FIG. 1: Left: flowfield model (thin arrows) and the three-level control (TLC) rule (thick lines). Right: radial velocity profiles,
composite from dropsonde measurements between 1996 and 2012 within 200 km of the hurricane center [5], binned into 50 m
altitude intervals and sorted according to hurricane category (1 to 5). The jaggedness of the profiles above 2 km altitude is due
to the reduced number of available measurements with respect to the lower region. Dashed lines estimate the mean velocity
gradient α.
vertical position, which are applied when the balloon reaches a distance of ∓d from the target trajectory x = 0. In such
a setting, the vertical coordinate Z̄ ∈ {−h, 0, h} is discrete, and the control rule exhibits hysteresis in the horizontal
coordinate; we additionally note that the L1 norm, measuring the step size h, is a more appropriate measure of the
energy |u|Q required by the balloon to change altitude than the classical L2 norm. Despite the apparent simplicity
of this control rule, it cannot be optimized as formulated in (1).
Rather than constraining the problem by the state-space representation of the system (2), as is done in (1), we thus
instead use an equivalent condition on the PDF pX,Z (x, z), and restate the optimization problem (1) as
min E [ |u| ]
u
h
2 i
E X − X̄
= constant
with
2
∂ p + f (X, u) · ∇p + c ∇2 p = 0,
t X
X
X
2
(3a)
(3b)
where we have replaced the state equation in (1b) with an equivalent Fokker-Plank equation for the PDF pX (x) [4],
and the norms are interpreted as expected values. The solution of the optimization problem as stated in (3) is the
principal contribution of this work.
The remainder of this paper is concerned with the solution of the optimization problem (3) for the TLC rule of
Figure 1a, and with comparison to the classical linear control rule u = k1 x + k2 z, whose optimal solution is given
by the Linear Quadratic Regulator (LQR) [2]. To facilitate comparison, we first derive the functional form of the
solution by dimensional analysis.
DIMENSIONAL ANALYSIS
The control problem is governed by three parameters: the velocity gradient
density c2 , and the √
target
pα, the spectral
2
−1
2
horizontal variance σ̄X . Take the length, time, and velocity scales as L = c /α, T = α , and U = L/T = c2 α.
A single dimensionless parameter can be defined as
2
R = σ̄X
α/c2 ,
(4)
and the dimensionless control cost can be written as w/U = E [ |u| ] /U = F(R) where F(R) is an unknown
dimensionless function. Similar expressions can be written for d/L and h/L. The system (2) is additionally invariant
with respect to a rescaling of the vertical coordinate by the time scale α−1 , thus reducing the parameters governing
3
the problem to the standard deviation σ̄X and the spectral density c2 . Dimensional analysis [1] can then be used to
write the control cost as
1 c4
w
− 32
.
(5)
= γw
3 = γw R
U
U ασX
Expressions for the control parameters can also be obtained as
1
1
d
Z
σX
1 c2
(6)
= γd
= γd R 2 ,
= γh
= γh R − 2 .
L
L
L
L ασX
The solution is then summarized by the optimal value of γ(·) for each control parameter. Similarly, for the linear
feedback control rule u = k1 x + k2 z, we can write T k1 = γk1 R−2 and T k2 = γk2 R−1 . Note that the solution (5) is
independent of the specific choice of the control rule u(x, z); a comparison between different rules can be obtained by
comparing the respective values of γw .
THREE-LEVEL CONTROL (TLC) RULE
x≤0
z = −h
x ≤ −d
z=0
z=h
x≥0
x≥d
FIG. 2: Implementation of the three-level control rule.
We now proceed in seeking the optimal values for the parameters d and h [equivalently, γd and γh in (6)] in the
TLC rule indicated by thick lines in Figure 1a, corresponding to step changes in altitude h at x = 0, ±d. In this limit,
the governing equations (2) can be restated as
Ẋ = −αZ̄ + ξ,
(7)
where X ∈ R is the same as in the original problem, but Z̄ ∈ {−h, 0, h} is now a discrete rather than a continuous
random variable, and the corresponding equation is replaced by the automata
√ in Figure 2. This is a valid approximation
when the control velocity u is larger than the horizontal velocity scale c2 α. Note that the dynamics of (7) is quite
simple, and is dominated by the effect of the stochastic excitation by ξ.
P
Let pX,Z̄ (x, z̄) be the PDF of the balloon position, and pz̄ (x) = pX|Z̄ (x|z̄) pZ̄ (z̄), so that pX (x) = z̄ pz̄ (x) is the
marginal probability. The governing equations for the PDFs can be written as
∂t pz̄ (x) + ∂x αz̄ pz̄ (x) − ∂xx
∂x pX (x)|x− = ∂x pX (x)|x+
c2
pz̄ (x) = 0
2
for z̄ ∈ {−h, 0, h},
for x ∈ {−d, 0, d},
(8a)
(8b)
where (8a) are three Fokker-Plank equations for each discrete altitude z̄ obtained by considering the transition
probabilities implied by (7), and (8b) represent the transition probabilities in the z direction, and imposes the
conservation of the probability fluxes represented by arrows in Figure 2. Equations (8) now take the role of the
optimization problem constraint (3b).
The statistically steady-state solution of (8a) can actually be computed analytically, as shown in Figure 3, and is
given by
p0 (x) =
ph (x) =
c1 x + c2
0 < x < d,
0
d < x < ∞,
2x
q1 λ2 e− λ + q2
2x
r1 λ2 e− λ + r2
(9a)
0 < x < d,
(9b)
d < x < ∞,
4
z
h
x
−h
−d
d
FIG. 3: Steady state probability distribution functions for c2 = h = α = 1. Filled: p0 (x) and p±h (x). Solid: marginal
probability pX (x) = p0 (x) + p+h (x) + p−h (x). Dashed: normal distribution with the same variance.
2
c
where λ = gh
is twice the e-folding scale of the PDF (the symmetry of the problem at x = 0 can be used to compute
the solution for x < 0, Z̄ = h). The integration constants c1 , c2 , q1 , q2 , r1 , r2 can be obtained imposing (8b) together
the boundary conditions p0 (x = d) = 0, ph (x = 0) = 0, limx→∞ pX (x) = 0, and the normalization condition
Rwith
∞
p
(x)dx = 1, resulting in
−∞ X
2d
1
1
eλ −1
, q1 = −
, r1 =
,
d(d + λ)
d(d + λ)
d(d + λ)
1
λ
c2 =
,
q2 =
, r2 = 0.
d+λ
2d(d + λ)
c1 = −
(10)
Upon substitution of the coefficients (10) into the expressions for p0 and ph in (9), the variance can be written
Z ∞
d3 + 2λd2 + 3λ2 d + 3λ3
2
σX
=
x2 pX (x)dx =
.
(11)
6(d + λ)
−∞
The control cost can be computed as the total transition probability between the states Z̄ = 0 and Z̄ = h, multiplied
by the cost of the single control activation h:
w = 2αhc2 ∂x p0 |x=d =
2ghc2
.
d(d + λ)
(12)
The control parameters h and d, and the corresponding control cost w, can then be computed by minimization of the
objective functional (3a). The corresponding dimensionless constants in (5) and (6) are
γw = 0.5432,
γd = 1.6288,
f = 0.4864 c
2
/σx2
γh = 1.1166,
= 0.4864 T
−1
R
−1
,
(13a)
(13b)
where f is the frequency at which the control has to be activated, and can be computed by considering the times
tout = d2 /c2 and tin = d/(gZ) to reach the location x = d from x = 0 and back, respectively.
It is also of interest to compute the minimum variance attainable for given values of d and h, and from there obtain
the limiting values of d and h for a specified σ̄X :
h→∞
2
lim σX
=
d→0
√
d2
6
⇒
d<
c4
2α2 h2
⇒
1 c2
h> √
.
2 σ̄X α
2
lim σX
=
6 σ̄X ,
(14a)
(14b)
In both limits the control cost tends to infinity, in the first case because h → ∞, and in the second because the
frequency of the steps increases without bound. Reasonable (finite) values of h and d, away from these limiting
values, are thus important in application. Results are summarized in Figure 4.
5
(b)
(c)
Linear rule
101
101
∝
3/
−
R
2
T
10−2
re
r
10−1
ep
ea
lin
st
e-
w
U
L
Tk
2
h/
L
10−1
100
R=
z(km)
10−2
10−1
100
R=
101
2
σ̄X
g/c2
100
k
1
10−1
101
100
2
σ̄X
g/c2
R=
(d)
10−1
101
d/
100
th
100
Three-step rule
R = 6
R = 6
102
R = 6
(a)
101
2
σ̄X
g/c2
(e)
2
2
0
−2
0
−2
−8 −6 −4 −2 0 2 4 6 8
−4.89
x(km)
0
4.89
x(km)
FIG. 4: Comparison of control rules: (a) control cost w = E [ |u| ] as a function of the dimensionless parameter R for the two
different control rules; the two control costs are almost indistinguishable, differing by less than 5%. (b,c) Control parameters
and (d,e) corresponding probability density for the hurricane case (R = 6) for the linear and TLC control rules; gray patterns
in (c) define the limit of each curve for which a given value of R is attainable, given by (14).
APPLICATION TO THE CONTROL OF BALLOONS WITHIN A HURRICANE
Finally, we compute optimal values of the control parameters for atmospheric balloons within an idealized hurricane
flowfield, and compare them with results using the linear control rule u = k1 x+k2 z. A velocity gradient of α = 10−3 s−1
(see Figure 1b) and a spectral density of c2 = 1500 m2 s−1 [6] are assumed.
We additionally impose a target standard deviation of σ̄X = 3 km, resulting in R = 6 [see (4)]. Control parameters
and control cost can be obtained using (6) together with the values in (13). For the TLC rule,
d = 4886.4 m,
w = 4.54 × 10
−2
h = 558.3 m,
(15a)
−5 −1
f = 8.11 × 10
m/s,
s
,
(15b)
where f corresponds to a period of about 3.5 hours.
The optimal solution for the linear control rule can be readily obtained by solving (1), resulting in
k1 = 3.125 × 10−5 s−1 ,
w = 4.32 × 10
−2
k2 = 2.5 × 10−4 s−1 ,
m/s.
(16a)
(16b)
Simulations for both control rules are shown in Figure 4d and 4e.
CONCLUSIONS
This paper introduces a probabilistic framework for the optimization of physical systems dominated by stochastic
excitation in the presence of mixed continuous and discrete random variables, non-linearities, and hysteresis. Its
application has been demonstrated by addressing the problem of controlling atmospheric balloons within a stratified
flowfield; the same framework can be extended to a class of problems that, to the authors’ knowledge, were previously
intractable from an optimization point of view. From an application perspective, the three-level control (TLC) rule
of Figure 2 has many advantages: holding rather than continuously adjusting a position is often an easier solution to
implement, possibly requiring less energy in real life applications. For an observational platform like a sensor balloon,
it has the additional advantage of performing measurements which are not disturbed by continuous control actions.
6
∗
[1]
[2]
[3]
[4]
[5]
[6]
Corresponding author: [email protected]; current address: Department of Earth, Atmospheric and Planetary
Sciences, Massachusetts Institute of Technology, Cambridge, Massachusetts, 02139, USA.
Barenblatt, G. I. (1996). Scaling, self-similarity, and intermediate asymptotics: dimensional analysis and intermediate
asymptotics, volume 14. Cambridge University Press.
Lewis, F. L. and Syrmos, V. L. (1995). Optimal control. John Wiley & Sons.
Nocedal, J. and Wright, S. (2006). Numerical optimization. Springer Science & Business Media.
Risken, H. (1984). Fokker-planck equation. In The Fokker-Planck Equation, pages 63–95. Springer.
Wang, J., Young, K., Hock, T., Lauritsen, D., Behringer, D., Black, M., Black, P. G., Franklin, J., Halverson, J., Molinari,
J., et al. (2015). A long-term, high-quality, high-vertical-resolution GPS dropsonde dataset for hurricane and other studies.
Bulletin of the American Meteorological Society, 96(6):961–973.
Zhang, J. A. and Montgomery, M. T. (2012). Observational estimates of the horizontal eddy diffusivity and mixing length
in the low-level region of intense hurricanes. Journal of the Atmospheric Sciences, 69(4):1306–1316.
| 3cs.SY
|
UCL DEPARTMENT OF
COMPUTER SCIENCE
Research Note
RN/12/03
Evolving Human Competitive
Spectra-Based
Fault Localisation
Research
Note
Techniques
arXiv:1306.5667v1 [cs.NE] 24 Jun 2013
RN/13/12
08/05/2012 to Model Software
Using Genetic Programming
24 June 2013
Shin Yoo
W. B. Langdon and M. Harman
Abstract
Abstract
Spectra-Based Fault Localisation (SBFL) aims to assist de- bugging by applying risk
evaluation formulæ (sometimes called suspiciousness metrics) to program spectra and
study a generic
programto
to investigate
the scope
automatically
it for aformula
vital
ranking We
statements
according
the predicted
risk.forDesigning
a customising
risk evaluation
is
task,
which was
not by
considered
it wasengineer.
first written.
In detail,
show genetic
often ancurrent
intuitive
process
done
human when
software
This
paperwe
presents
a Genetic
programming
(GP) can
models
of aspects
of BLAST’s
output The
when empirical
it is used toevaluation
map
Programming
approach
for evolve
evolving
risk
assessment
formulæ.
1
Solexa
Next-Gen
DNA
sequences
to
the
human
genome.
using 92 faults from four Unix utilities produces promising results . GP-evolved equations
can consistently
thegenetic
human-designed
formulæ,
as Tarantula,
Keywords: outperform
1000 genomesmany
project of
1KG,
programming, GIP,
GISMOE,such
automatic
softOchiai, Jaccard,
Ample, and Wong1/2, up to 5.9 times. More importantly, they can perform
ware re-engineering, SBSE, software modelling, collaborative coevolution, teams of agents,
equally soft
as memory,
well as bioinformatics,
Op2, which approximate
was recently
proved to be optimal against If-Then-Else-2
string lookup, sequence alignment, SNP
(ITE2) structure, or even outperform it against other program structures.
1
The program spectra data used in the paper, as well as the complete empirical results, are available from:
http://www.cs.ucl.ac.uk/staff/s.yoo/evolving-sbfl.html.
Using Genetic Programming to Model Software
1
W. B. Langdon and M. Harman
Introduction
Hand made software is expensive. Therefore to make the widest possible use of it we usually try to make
programs as generic as possible. Even though a task specific program might be better at that task, it is often
too expensive to write bespoke software and the user must make do with the generic program. However if
automated software production were to be successful, the balance could turn from single generic programs
to multiple bespoke programs, each tailored to a particular task or even tailored to each user.
We study a popular generic program BLAST [Altschul et al., 1997] to investigate the scope for automatically customising it for a popular current demand, mapping next generation DNA sequences. BLAST
is too generic, too slow to be used with Next-Gen sequences. Indeed they were unknown when it was
first written. Yet matching them against the human genome is a vital component of an increasingly
important part of modern biology and doubtless this will also shortly play a vital role in medicine.
If it was practical, BLAST would be the tool of choice for Next-Gen sequence alignments. However it is far too slow to cope with the hundreds of millions of DNA sequences generated by current,
let alone future, sequencing machines. Instead new hand made programs are used. The new software re-uses ideas from BLAST (e.g. string hashing and partial matches) but sometimes also includes
newer ideas (e.g. data compression algorithms [Langmead and Salzberg, 2012; Langdon and Harman, ;
Langdon, 2013]). Also often the new performance comes at the price of losing some functionality [Langmead et al., 2009]. In the future both combining existing code in new ways and performance tradeoffs may
be within the scope of software optimisation techniques. In detail, we show genetic programming (GP) can
model aspects of BLAST’s output which are specific to Solexa Next-Gen DNA sequences and the human
genome.
BLAST is perhaps the most important Bioinformatics tool [Karplus et al., 1998, p848]1 . Originally designed to find optimal matches between protein and amino acid sequences, it and its variants are now
widely used for many kinds of sequence comparison, in particular for DNA sequences. With the advent of
the sequencing of the human genome and next generation sequencing the demands of sequence matching
have changed enormously since BLAST was written. With this in mind we have been investigating to what
extent it might be possible to automatically re-engineer a bespoke version for BLAST for Next-Gen DNA
sequence look up. We are somewhat short of our goal of having evolutionary computing rebuild a version BLAST which is more efficient at aligning next generation DNA sequence with the human genome.
However there are certain aspects of the input data which are critical to BLAST’s performance which we
have been able to model with genetic programming. These include: the quality and length of the match
found and indeed the number of matches found or even if the DNA sequence does not occur in the human
genome.
In the next section we briefly summarise BLAST and related programs, including important properties that
influence run time. Then in Section 3 we will describe the source of the training and verification data.
(All of these data are available online, e.g. via ftp.1000genomes.ebi.ac.uk.) Section 5 shows it is possible
(using genetic programming, Section 4) to estimate in advance important properties of the BLAST’s output.
Although GP has been used many times before this is the first time it has been used to model a substantial
man made software tool like BLAST2 as part of plan to reverse engineer it.
2
Background
In June 2013 Wikipedia listed more than 140 Bioinformatics tools which perform some aspect of string
analysis either on protein databases or DNA sequences. This represents a huge investment in manual
labour. Many of these tools can be regarded as finding a specific trade-off between performance (typically
in terms of accuracy and number of matches found) and non-functional requirements (typically speed, or
1
Google Scholar says BLAST had been cited more than 47 000 times and it is top of the Nucleic Acids Research journal’s
(1 June 2013) list of most cited articles.
2
BLAST contains about 20 000 lines of C++.
RN/13/12
Page 1
Using Genetic Programming to Model Software
W. B. Langdon and M. Harman
speed and memory requirements). As part of the GISMOE program [Langdon, 2012; Harman et al., 2012]
[Harman et al., ; Petke et al., 2013; Langdon and Harman, 2010; Langdon, 2013; Langdon and Harman, ]
to mechanise aspects of software production we are investigating to what extent these trade-offs can be
automated.
The performance of many sequencing tools is dominated by the length and number of exact matches. Run
time can grow rapidly, e.g. O(n3 ), with number of matches (n) [Langdon and Harman, ]. Hash techniques
allow exact matches to be found very quickly but approximate matches are time consuming. Therefore
we investigated if GP could say in advance if a look up is likely to match exactly and with what length
of match and if multiple matches would be encountered. In sequence alignment multiple matches are less
useful but are very time consuming, thus (since they have redundant data) biologists often ignore such
sequences [Cheung et al., 2011, page 1].
Where redundant data are available, if we could identify high quality matches (cf. E value, Section 4.1) in
advance it would make sense to use them first. Also if we suspect a sequence will not match it can be given
a lower priority. Similarly if it occurs more than once in the human genome the problem of deciding how to
resolve this ambiguity can be delayed until after more promising sequences have been tried. Finally perfect
match “seeds” are critical components of sequence look up. Knowing in advance how long an exactly
matching region is likely to be would make hash based search more efficient.
Although we model BLAST, since it incorporates many Biological sequencing heuristics and is widely
used, it has become the de facto standard for sequence matching which other tools seek to emulate. Thus
a model of BLAST is also a model of the desired output of other (more modern) sequence look up tools.
However although GP models are fast enough to make prioritisation of DNA sequence matching feasible,
BLAST itself is still not competitive with more modern tools specifically designed for Next-Gen sequences.
3
Training Data – NCBI Human Genome, BLAST and the 1000 Genomes Project
We used the official USA’s National Center for Biotechnology Information (NCBI) release of the reference
human genome. (Release 37 patch 5, GRCh37.p5, was down loaded via update blastdb.pl from http://www.
ncbi.nlm.nih.gov/blast/docs.) Similarly we used the official NCBI 64 bit Linux version of BLAST (version
2.2.25+ was downloaded via ftp://ftp.ncbi.nlm.nih.gov/blast/executables/blast+/)
The 1000 genomes project [Durbin, et al., 2010] is a consortium in which DNA from more than one
thousand individuals has been sequenced by several institutions using a variety of scanners. To avoid unknown data inconsistencies during training we decided to concentrate upon a single scanner used by one
laboratory. Similarly we wanted to minimise true biological variation and concentrated upon the software
technology. Therefore we chose one well studied family and the Solexa data provided by the Broad Institute, Cambridge, MA. This gives a large pool of homogeneous data. The Solexa data includes sequences of
a number of lengths (from 25 to 489 DNA bases) and so we further limited ourselves to all the Solexa data
with exactly 36 DNA bases per sequence. These sequences are “single ended” and give “high coverage”.
We initially trained our GP using data from 11 different Solexa-3623 scans of the same female (NA12878)
from Utah (a total of 89 087 344 DNA sequences).
The Solexa data quality is highly variable. In addition to its quality indicators, the Solexa data use “N” to
indicate a DNA base which it cannot decide which of the four bases (A, C, G, T) it really is. In one dataset
less than 1 in a thousand sequences has an N. In the worst training dataset (SRR001752), every record had
at least one (typically two or three). We selected data from the best, the worst and an intermediate dataset
for training.
RN/13/12
Page 2
Using Genetic Programming to Model Software
4
W. B. Langdon and M. Harman
Genetic Programming
This section describes in sufficient detail to allow reproduction of our results the genetic programming
system we used. Section 5 (page 8) describes the models it evolved, their performance and also explores
how they work.
4.1
Preparing the Training Data
Solexa scanners optically read DNA sequences from 60 or more square tiles. It has been suggested that
data quality may be worse near the borders of the tiles. However we do not see this effect. Nevertheless for
each of the three training datasets, we selected sequences in narrow (11 pixels wide) vertical and horizontal
(10 pixels) strips crossing the whole of a tile (at pixel 511,779) (Tile 119 was randomly selected for two
scanner runs but proved to give a small data sample in the third case, so tile 78 was randomly chosen for
the last set of NA12878 training data.) This gave 606, 565 and 1186 (total 2357) DNA sequences each with
36 bases for training.
In addition to BLAST’s default parameters, we used -task blastn-short and -num threads 6,
as it looked up each DNA sequence in the NCBI reference human genome. (Using 6 threads BLAST
took approximately 7 hours on a 32GB 8 CPU 3GHz Intel server to process the 2357 sequences. The GP
runs used a single CPU each.) In total for the 2357 queries BLAST generated 14.5 gigabytes of output
containing 60 151 902 partial matches. For each partial match BLAST supplies a number of statistics.
These include its expectation (or E-value), which is an estimate of how likely the match would be at
random. (The E-values lie in the range 10−10 to 6.6.) BLAST also reports the number of bases which
match exactly between the query sequence and the human reference genome (0-36).
The NCBI reference database for the human genome contains DNA sequences from a number of sources
[Aldecoa-Otalora et al., 2009; Langdon and Arno, 2012]. In many cases these sources overlap. Therefore
to count the number of matches found by BLAST, we include only high quality matches (i.e. E< 10−5 )
and restrict BLAST to the reference sequence for human chromosomes 1 to 22 and the human X and Y
chromosomes.
4.2
Randomised Test Suite Sub-sampling
When a large volume of training data is available we had previously used a random sub-sample of the
test data at each generation [Langdon, 2010] to reduce the volume of testing but also found it helped with
generalisation. Here we also use it to spread the training data more uniformly.
We divided the training data into non-overlapping bins using the value to be predicted. (In the case of the
two binary classification problems, Sections 4.5 and 4.6, there are just two bins.) Each generation equal
numbers of training examples are randomly chosen from each bin. Where a bin contains more examples
than are needed the examples are kept in the same random order and the next group are taken. Except where
noted, this ensures the examples used in the next generation are not the same ones as used in the previous
generation. If there are insufficient examples left, the bin’s examples are first put in a new random order.
(This is somewhat reminiscent of Gathercole’s DSS [Gathercole and Ross, 1994], as used commercially
[Foster, 2001].)
4.3
Predicting the Quality (E value) of DNA Matches
Typically BLAST reports many matches. However it sorts them so that the best match comes first. Therefore we simply took the E value of the first match to be GP’s target value. Where BLAST was unable to
find any match this was presented to GP as an E value of 100. Due to the wide range of E values, GP works
with log10 (E) rather than the E value directly. Using the integer part of log10 (E) we divided the training
data into 13 groups (separated by powers of ten). Each generation 35 examples are randomly chosen from
each group. This means each generation 12 × 35 + 11 = 431 (of 2357) examples are used. (The twelfth bin
RN/13/12
Page 3
Using Genetic Programming to Model Software
W. B. Langdon and M. Harman
Table 1: GP Parameters for Predicting BLAST results with 1000 Genomes Project Solexa short DNA
sequences.
Pass0 Terminals: 1037 random constants pos len=36 A C G T Self Complement Samesize Opposite N
Qual S Run CountN X Y Aux1 Aux2
Pass1 Terminals: 1037 random constants pos len=36 A C G T Self Complement Samesize Opposite N
Qual S Run CountN X Y Aux1 Aux2 M Sum0
Result Terminals: 1037 random constants len=36 S Run CountN X Y Aux1 Aux2 Sum0 Sum1
Functions:
IFLTE ORN ADD SUB MUL DIV LOOK01 set Aux1 set Aux2 sum Aux1 sum Aux2
Fitness:
If predicting E or length, fitness is correlation. If classifying high quality match or
repeated high quality match then fitness is the number of correct answers
Population:
Panmictic, generational. 10 000 members. 4 members per selection tournament. New
training sample each generation.
Parameters:
Initial pop ramped half-and-half 6:2. 50% subtree crossover, 22.5% point mutation,
22.5% mutation which swaps constants, 2.5% shrink mutation, 2.5% subtree mutation.
No depth limit, max size 1022. Stop after 100 generations.
contains only 11 examples.) Note, the bins are only used for training, data used to test the evolved predictors are drawn uniformly. The GP fitness function is the Pearson correlation between the value calculated
by GP and log10 (E).
4.4
Predicting the size of the Best BLAST Match
Again we took the best match to be GP’s target and this time asked GP to predict its length (i.e. the length
of the first match reported by BLAST). There are 137 Solexa training sequences where BLAST was unable
to find any match. These are treated as having a length of zero and placed in a group by themselves. All
reported matches have a length of at least 18 bases. They are placed in ten groups: 18-19, 20-21, 22-23, ...,
34-35 and length 36.) So 11 × 35 = 385 of 2357 examples are randomly chosen for use in each generation.
Fitness is the correlation between GP’s value and the actual length.
4.5
Predicting High Quality BLAST Matches
We evolve two classifiers for the number of BLAST matches. The first (this section) returns a positive
value if GP thinks at least one high quality match will be found (i.e. E < 10−5 ). Conversely, there are 997
(42%) training DNA sequences which either BLAST could not find any matches or where the matches had
poor E values (E≥ 10−5 ).
With both classifiers (i.e. this section and the next), each generation 300 positive and 300 negative randomly
chosen examples are used to assess every individual’s fitness. Fitness is the number of training cases
correctly classified.
4.6
Predicting Repeats in the Human Genome
When predicting repeats, a positive value is used to indicate GP expects there to be multiple instances of the
DNA sequences in the Human genome. (Repeated sequences are quite common [IHGSC, 2001].) However
there are only 324 training sequences with multiple high quality matches. Selecting 300 positive examples
means almost all of them are used every generation. In contrast each of the 2033 negative examples will
have to wait at least six generations (b2033/300c = 6) before it is even eligible to be reused by the fitness
function. (On average each negative example is used in 16 generations.) As with all our other predictors,
the separate data used to validate the evolved predictor are drawn uniformly.
RN/13/12
Page 4
Using Genetic Programming to Model Software
5’
W. B. Langdon and M. Harman
3’
m4
GP prediction
36
Sum0
prepas0
Sum1
prepas1
Expect
Figure 1: Schematic of three tree architecture used to evolve prediction of BLAST results. Tree prepas0
is called once for each base in the DNA sequence in 5’–3’ order. Various memory cells (not all shown)
allow it to communicate partial results both with itself and with the other trees. Similarly the evolved tree
prepas1 is called once per base and can access information about the DNA sequence and results passed to
it by prepas0. Finally the Expect tree is called once. It can use data provided by both prepas0 and prepas1
to return the overall GP prediction for the whole DNA sequence. Figures 7, 8, 9 and 10 contain examples
of actual evolved code.
4.7
Genetic Programming Architecture and Primitives
The rest of Section 4 describes the GP representation, data flows within its three components, terminal and
function sets and the speed of the GP system. We used a multiple tree GP (based on [Langdon, 1998])
with functions and terminals inspired by Koza’s transmembrane prediction work [Koza, 1994]. (Details
are given in Table 1.)
4.7.1
Co-Evolving Three Trees
It seems clear that the evolved classifiers will need to scan the DNA sequence. It is unclear how many
times they will want to scan it. To support serial processing of the sequence and still allow evolution
some freedom of choice, we adopted a novel three tree architecture (Figure 1). The first two trees scan the
sequence and pass data via memory to the final one. GP is free to decide how to use this architecture. It
can readily use it to process the complete DNA sequence 0, 1 or 2 times. The first tree is called once for
each member of the sequence from the start to the end. (In Bioinformatics this is known as 5’–3’ order).
Thus the first GP tree is used 36 times. Then the second tree is called for each DNA base in the sequence
in the same order. (Again making 36 times per fitness evaluation.) Finally the last tree is called once and
its result is used to determine the fitness of the whole team.
The trees can also be thought of as team members or agents. They are independently evolved in that
crossover and mutation act on a single agent but whole teams are forced to coevolve collaboratively. Each
member is locked into its team and cannot act outside it. In earlier experiments with evolving memory we
showed GP can coevolve such teams of agents. [Langdon, 1998] contains examples of teams of five and 15
agents. Recently GP has since been used to evolve even bigger teams [Moreno-Torres et al., 2009]. GP can
evolve individuals which pass results returned by earlier tree calls and previously calculated intermediate
values amongst team members by using a variety of special purpose and general purpose memory cells. (As
is common in machine learning, GP individuals cannot affect the fitness of other members of the population
or influence their own behaviour on other test cases.)
4.7.2
Passing Data Between Trees via Memory
It is expected that passing along the DNA sequence will enable GP to calculate complex statistics somewhat
similar to running averages. To assist this the sum of the 36 results produced by each of the prepass trees
is stored and made available to the following parts of the GP individual via read-only terminals Sum0 and
Sum1.
RN/13/12
Page 5
Using Genetic Programming to Model Software
W. B. Langdon and M. Harman
The GP leaf M refers to 36 read-only memory cells, which allows the second prepass to use the result
returned by the first pass. It can be thought of a similar to an ADF [Koza, 1994] with a cache [Langdon,
1998].
We also have more conventional indexed memory [Teller, 1994]. There are two floating point memory
cells, called Aux1 and Aux2, and routines to write to them, set Aux1 and set Aux2. As with [Langdon,
1998], functions set Aux1 and set Aux2 take one argument, which is assigned to the memory cell and
return the memory’s new value.
Inspired by the work on “memory with memory” [Poli et al., 2009] we also introduce two functions:
sum Aux1 and sum Aux2. Instead of overwriting their memory cell’s contents they add their input’s value
to it. (Again returning the memory’s new value.) Alternatively these can be thought of as parametrised
memory increment instructions [Langdon, 1998].
4.7.3
DNA bases
Inspired by Koza’s work on protein sequences [Koza, 1994] we use five terminals (A, C, G, T and N) to
indicate (in both prepass trees) if the current bases is an A, C, G or T. N indicates the scanner did not know
which of the four bases it is. The leaf has the value 1 if is true and -1 otherwise. (Thus leaf N is always -1
unless the DNA scanner has indicated it cannot tell which of the four bases it is.)
4.7.4
Look Ahead
It may be convenient for GP to be able to compare nearby parts of the sequence. The function LOOK
moves the active position (denoted by terminal pos) ahead one. Within the overall limit on program size,
LOOK can be as deeply nested as evolution likes and so it allows arbitrary look ahead. I.e. special inputs,
like A, when invoked by LOOK’s single argument, respond as they would do when the prepass tree is
called later along the sequence. LOOK applies to all position dependent primitives (including Qual, Run,
M, S and pos itself). Protection against looking pass the end of the DNA sequence is given.
4.7.5
DNA Specific Terminals
The complementary binding of A–T and C–G is well known. We sought to give ready access to such
information. The four terminals: Self, Complement, Samesize and Opposite each compare the current base
(as updated by LOOK) with the base at the standard position (i.e. without LOOK). Like A, C, G, T and N,
they signal true and false with 1 and -1.
Self returns 1 if the current DNA base and the “LOOKed” at DNA base are the same. Complement returns 1
if they are members of a complementary pair. Samesize returns 1 if they contain the same number of rings.
(C and T are “small” and have one ring. A and G are large and have 2 rings. Arbitrarily N is given 0 rings.)
Opposite is true only if Self, Complement and Samesize are all false.
4.7.6
Solexa Quality Terminals
As well as N, the Solexa scanner includes a quality value for each base. These are presented to the GP via
the Qual terminal as values in the range 0, 0.1, 0.2, ..., 3.9, 4.0. (4.0 indicates the DNA scanner has the
highest confidence in its output. 0 means it has no confidence in it.)
4.7.7
Runs of the Same DNA Base
It is known that sequences of the same base can affect Bioinformatics equipment [Upton et al., 2008], so
we provide GP with a primitive which counts the length of identical bases up to and including the current
point. (Run’s value is therefore in the range 1, 2, 3, ... theoretically up to 36.) No special treatment for N
is provided. I.e. an N value terminates a run of other letters but we can also have runs of Ns.
RN/13/12
Page 6
Using Genetic Programming to Model Software
4.7.8
W. B. Langdon and M. Harman
CountN
CountN is simply the number of “N” (i.e. unknown) bases in the current DNA sequence. Typically CountN
takes the value 0, 1, 2, or 3. Since CG ratio is known to be important to the Solexa scanner [Cheung et al.,
2011], we might also have provided similar CountA, CountC, CountG and CountT terminals.
4.7.9
Entropy S
The terminal S holds the entropy or information content of the string in bits from its start up to the current
P
position. I.e. S= A,C,G,T −pi log2 (pi ). Where, for example, pA is the number of As in the current DNA
sequence from its start up to the current position divided by the number of all four bases up to the current
position. Since N indicates an unknown base, it is accounted for by adding 1/4 to each of the four bases’
totals. S in the third tree yields the entropy of the whole sequence.
4.7.10
X and Y
The leafs X and Y tell the GP where the DNA sequence was located on the Solexa tile. (See Section 4.1).
They are both normalised to the range 0..1 by dividing by the width of the tile in pixels.
4.7.11
Ephemeral Random Constants
1000 values were randomly selected from a tangent distribution [Langdon, 1998]. (I.e. a value is randomly
chosen from the range 0..π and its tangent is taken.) This gives a few very large numbers (the largest was
632.124324 and the smallest -425.715953) but about half the values lie in the range −1.. + 1. This was
supplemented by the 37 integer values: 0, 1, ... 36.
4.7.12
Functions
In addition to LOOK and the four usual arithmetic operations (+, -, × and protected division, which returns
1 on divide by zero) we include if-less-than-or-equal [Langdon, 1998] and Koza’s ORN [Koza, 1994].
ORN takes two arguments. If the first is true (i.e. > 0) the second is skipped and ORN returns 1. Otherwise
it evaluates its second argument and returns 1 if it is true. If the second argument is false (i.e. ≤ 0), ORN
returns -1.
Mostly default behaviour is applied for special values like NaN or infinities. However care is needed with
calculating correlation based fitness and protecting against rounding errors etc. causing numerical instabilities (e.g. leading to negative variance). Members of the GP population causing unresolved numerical
problems are given low fitness values.
4.7.13
GP Speed
The special GPquick interpreter [Singleton, 1994], [Langdon, 1998] we used processes on average 213
million GP primitives per second. In [Langdon, 2010] we reported a GPU based GP interpreter running
more than a thousand times faster, however this was build to exploit both the graphics hardware accelerator
and the bit level parallelism inherent in Boolean problems [Poli and Langdon, 1999], rather than floating
point numbers as used here. Our GPquick compares well with tinyGP [Poli et al., 2008], which processes
in the region of 20–80 MGPops−1 (C version). [Langdon, 2011] discusses the speed of recent compiled and
interpreted GP approaches, particularly those using graphics hardware accelerators. (See also [Langdon, ,
Table 15.3].)
RN/13/12
Page 7
Using Genetic Programming to Model Software
W. B. Langdon and M. Harman
0.8
Best in population Correlation
0.75
0.7
0.65
0.6
0.55
All 2357 training cases
per generation 431 training cases
0.5
0
10
20
30
40
50
Generation
60
70
80
90
100
Figure 2: Evolution of performance on training data of best of generation BLAST E-value predictor.
Overall training performance (lines) follows performance on 431 tests being used in the current generation
(lines with +).
5
Evolved Prediction of BLAST Matches
In each of the four problems, we extracted the best of generation individuals at generation ten and generation one hundred. (Details are given in Table 1.) Figure 2 shows the evolution of the best of generations
fitness, both on its per generation training data (lines with +) and on the whole of the training data (solid
lines).
In each of the four prediction tasks and at generation ten and generation one hundred, we tested the smallest
of the best of population individuals on a validation set of 31 000 Solexa DNA sequences. The validation
data came from completely new sets of sequences firstly chosen uniformly at random from the eleven
NA12878 data sets (a 1000 from each). There were only two other individuals in the 1000 genomes project
with more than ten data scans from similar Illumina2 Solexa scanners. These are NA12891 and NA12892,
who are her father and mother respectively. Ten data scan were chosen at random for each and then 1000
DNA sequences were chosen from each scan. (This entailed running BLAST in total another 31 000 times,
taking about 108 hours.)
In all four cases GP was to some degree successful at predicting BLAST’s output. Although there are
statistically significant differences between the thirty one validation DNA scans (e.g. SRR001770), performance of the evolved predictors held up well. See Figures 3, 4, 5, 6 and 11. In contrast to the other scans
of NA12878’s DNA, scan SRR001770 has about half as many high quality matches. (I.e. BLAST matches
with E = 10−10 .) The reduction in high quality matches, being approximately the same as the increased
fraction of intermediate quality matches, i.e. E ≈ 10−7 . Even on the most difficult problem, predicting the
existence of repeated short sequences in the Human genome, across all 31 000 verification sequences, GP’s
predictions are much better than random guessing, χ2 = 108 (1 dof).
RN/13/12
Page 8
Using Genetic Programming to Model Software
W. B. Langdon and M. Harman
Validation datasets
0.85
0.8
E generation 100
Correlation
0.75
E generation 10
0.7
0.65
0.6
Expect Generation 10
Expect Generation 100
0.55
NA12878
1770
NA12891
NA12892
Figure 3: Validation of evolved predictors of BLAST E-values evolved at generations 10 and 100 on 31
DNA scans from three individuals (NA12878, NA12891, NA12892). Apart from NA12878’s DNA scan
SRR001770, performance on unseen data is close to that on 2357 training data (horizontal lines). Error bars
indicate standard error [Hotelling, 1953, p212]. For comparison the horizontal lines indicate performance
on the 2357 training data.
Validation datasets
0.7
0.65
Length generation 100
Correlation
0.6
0.55
Length generation 10
0.5
0.45
Length Generation 10
Length Generation 100
0.4
NA12878
1770
NA12891
NA12892
Figure 4: Validation of evolved predictors of length of BLAST matches from generations 10 and 100
on 31 DNA scans from three individuals (NA12878, NA12891, NA12892). As with predicting E values,
Figure 3, apart from NA12878’s DNA scan SRR001770, performance on unseen data is close to that on
2357 training data (horizontal lines). Error bars are standard error.
RN/13/12
Page 9
Using Genetic Programming to Model Software
W. B. Langdon and M. Harman
Validation dataset
0.9
0.85
Fraction Correct
0.8
Match generation 100
0.75
Match generation 10
0.7
0.65
0.6
0.55
Match Generation 10
Match Generation 100
0.5
NA12878
1770
NA12891
NA12892
Figure 5: Performance when predicting whether BLAST will find matches or not on unseen data. Apart
from NA12878’s DNA scan SRR001770, performance is close to that on the 2357 training data (horizontal
lines). As with Figures 3 and 4, we plot validation scores for predictors evolved at generation 10 and at
generation 100. Error bars are standard error.
Validation dataset
0.7
0.65
0.6
Fraction Correct
Multiple generation 10
0.55
Multiple generation 100
0.5
0.45
0.4
0.35
Multiple Generation 10
Multiple Generation 100
0.3
NA12878
1770
NA12891
NA12892
Figure 6: Predicting repeated matches in the validation data. As with Figures 3, 4 and 5, we plot validation of GP results, as evolved at generations 10 and 100. Unfortunately, particularly for generation 100,
performance is worse on the validation data. However Figure 11 shows this is compensated for a little by
GP’s being able to predict most of the minority class. Again horizontal lines show performance on the
2357 training data. Also NA12878’s DNA scan SRR001770 is again not typical. (Actually SRR001770 is
better than average in this case.) Error bars show standard error.
RN/13/12
Page 10
Using Genetic Programming to Model Software
W. B. Langdon and M. Harman
prepas0 = (sum_Aux2 (DIV (set_Aux2
(LOOK (ADD (LOOK
Samesize)
(ADD Qual ((ADD
(LOOK (ADD (ADD ((
(LOOK (ADD
(LOOK (ADD Qual (ORN (MUL Samesize T)
(LOOK C)))
)
((ADD Qual (ORN (MUL Samesize T) (ADD T (SUB (ORN Opposite Self) T))))))
)))
(SUB (ORN Opposite Self) T))
(ADD Qual ( (ADD Qual (ORN (MUL Samesize T) (LOOK C))))))
)
(MUL Samesize T))))))
) Qual))
prepas1 = (0)
Expect = (DIV -0.551039 (ADD 11 (Sum0)))
Figure 7: Simplified best of generation 100 BLAST E-value predictor. GP gives its prediction via the
value returned by the Expect tree. Sum0 contains the sum of the 36 values returned by tree prepas0
when it is called in turn for each of the 36 bases in the DNA sequence generated by the Solexa scanner. The
average correlation with BLAST’s E-value is 0.76. Much of its predictive power comes from the Solexa
quality indicators (via the Qual terminals). See Section 5.1.
5.1
Predicting BLAST “E” Values
After 100 generations GP evolved a program whose correlation with BLAST’s E-value is r=0.76 on the
training data and whose median correlation is r=0.77 on all 31 000 validation sequences, see × in Figure 3.
Generation 100’s 77 primitive program can be simplified to an almost equivalent program (see Figure 7)
whose final output depends upon the sum over all 36 position in the Solexa DNA sequence of the value
calculated by prepas0. The use of sum Aux2 at the root of prepas0 frequently occurs and gives more weight
to the 3’ end of the sequence. GP uses nested LOOKs to compare nearby DNA bases and their reported
quality. In fact the best of generation 10 individual from the same run uses only Solexa’s quality indicator to
predict BLAST’s E-value. Although performance is not as good, it still obtains r=0.68 on the training data
P 0
P 0
and is almost equivalent to 350 iQual × 350 i/Qual. (Remember division is protected, Section 4.7.12.)
The formula suggests lower quality Solexa data contains more mistakes which in turn leads to BLAST
matches where more DNA bases do not align exactly and hence BLAST assigns them poorer E-values.
Notice again the 3’ end is weighted 36 times more important than the 5’ start of the Solexa DNA sequence.
This suggests that GP has learnt that the probability of errors accumulate as the scanner moves along the
DNA sequence from the 5’ start. Thus GP assigns more importance to data from the 3’ end of the sequence
than from the 5’ start of the sequence.
The evolved model (Figure 7) gives a quick way of predicting how good the match between the DNA
sequence and the human genome is. Thus redundant sequences which are predicted to give poor quality
matches could be given reduced priority or even ignored in favour of better DNA sequences generated by
the scanner.
5.2
Predicting Match Length
Predicting the length of the match found by BLAST proved more difficult with the smallest best of generation 100 predictor having a correlation on all the training data of r=0.61. However again this holds up
well on the validation data, with a median correlation r=0.60, see × in Figure 4. The error bars confirm the
results are highly statistically significantly different from the null hypothesis (no correlation).
RN/13/12
Page 11
Using Genetic Programming to Model Software
W. B. Langdon and M. Harman
prepas0 = (1.090291)
prepas1 = (IFLTE (IFLTE (MUL N A) Qual (sum_Aux2 Sum0)
(sum_Aux2 (set_Aux2 (sum_Aux1 1.784881))))
(IFLTE (IFLTE (MUL N A) Qual (sum_Aux2 Sum0)
(sum_Aux2 (set_Aux2 (sum_Aux1 2.105103))))
Qual
(set_Aux2 (sum_Aux1 2.235319))
(sum_Aux2 (set_Aux2 (MUL T Aux1))))
(sum_Aux2 Sum0)
(IFLTE (MUL N (SUB Opposite X)) Qual (sum_Aux2 Sum0)
(sum_Aux2 N)))
Expect = (ADD (ADD (sum_Aux1 (DIV (ADD (IFLTE CountN Y len Aux2) Sum1)
(IFLTE CountN Y len Aux2)))
Sum1) Sum1)
Figure 8: Smallest best of generation 100 predictor of the length of the best match found by BLAST.
GP’s prediction is given by the Expect tree. In this program, the Expect tree’s return value is largely
determined by the second prepass along the DNA sequence (via variables aux1 and aux2 and sum1). Sum1
P
= 36
i=1 prepas1. Its predictive power comes to a large extent from the quality (via Qual and CountN
leafs) of the Solexa data. See Section 5.2. Although position data X and Y are included they play a minor
role.
The best of generation 100 program is shown in Figure 8. The best program at generation ten has a similar
performance (r=0.60) and being smaller it is easier to simplify. It is similar to a predictor which returns
P30
50 if(Quali >1) i else -i. Notice that again the 3’ end is weighted 36 times more important than the
5’ start. This formula means high quality levels (i.e. bigger than 1.0) especially near the end of the DNA
sequence, suggest longer BLAST matches (r=0.54). The evolved formula includes ignoring low quality
values on A bases. If we add just this, we can get back almost all the performance (r=0.59). Whereas only
using Quali >1 by itself, excluding the location weighting i, still gives a correlation of r=0.48.
Again we see the quality of the sequence generated by the scanner playing a dominant role. Poorer quality
sequences tend to have more errors leading to BLAST finding sequences in the human genome which do
not match exactly. A single mismatch at the end of the query sequence reduces the length of the reported
match marginally, whereas if its in the middle it can half the length of the exactly matching region.
Typically sequencing tools like BLAST assume a minimum length of exactly matching sequences. (If
the query sequence is disrupted so the number of consecutive matching DNA bases is less than this, then
the sequence will not be found.) Sub-sequences of the minimum length are used as hash keys into the
reference genome. Longer keys tend to mean the hash tables are smaller and there are fewer hits which
speeds the search. Also both fewer hits and longer hash keys tend to mean the subsequent fuzzy matching
has less work to do and so is faster. BLAST relies on the user to set the correct hash key size. However
users typically use the supplied default, which is fixed (for a given class of queries). If GP could tell us in
advance how long the exact matching region was going to be, the optimum hash key size could be used.
5.3
Predicting High Quality BLAST Matches
The smallest of the best of generation 100 programs evolved to predict if BLAST will find a high quality
match (Figure 9), gets 80% of the training data correct. The × plot in Figure 5 shows this holds up well on
the verification sequences (median 79%). Again this is highly statistically significant. (Across all 31 000
verification sequences χ2 = 8400, 1 dof).
RN/13/12
Page 12
Using Genetic Programming to Model Software
W. B. Langdon and M. Harman
prepas0 = (SUB 1.744647 pos)
prepas1 =
(IFLTE (LOOK (IFLTE 1.370267 N 0.496664 Qual))
(ORN (sum_Aux1 (DIV (LOOK (LOOK Aux2))
(sum_Aux2 (DIV (MUL Aux2 G) (sum_Aux2 N)))))
(ORN Aux1 Samesize))
(DIV (IFLTE (sum_Aux1 (DIV (ORN (set_Aux1 T) (ORN Complement G))
(sum_Aux2 (LOOK A))))
(ORN (set_Aux1 T) (ORN Complement G))
(DIV (MUL Aux2 G) (sum_Aux2 N))
(set_Aux1 (DIV (set_Aux1 X) (ORN (ORN (set_Aux1 T)
(ORN Complement G)) (ORN Complement G)))))
(sum_Aux2 N))
(set_Aux1 (LOOK (LOOK Aux2))))
Expect = (sum_Aux2 len)
Figure 9: Smallest best of generation 100 predictor of whether BLAST will find a good match. Here the
evolved classifier has not been simplified. Figure 10 shows an almost equivalent program.
prepas1 = (DIV (LOOK (SUB Qual G)) Qual)
Expect = (SUB 72 (ADD Sum1 (MUL S 1.177709)))
Figure 10: Simplified best of generation 10 predictor of whether BLAST will find a good match (cf. Figure 9). Section 5.3 describes how it works.
Since the generation 100 code is large (Figure 9), we again start from the smallest best of generation 10
program which is slightly less accurate (74%) but smaller. It can be simplified to the code shown in Figure 10. The simplified code also has a training accuracy of 74%. Remember the prediction is given by the
sign of the “Expect” tree and so essentially the predictor works by comparing the sum calculated by the secP 0
ond scan of the sequence with 72.0. The second prepass calculates Sum1= 350 (Quali+1 − Gi+1 )/Quali .
Sum1 takes values spread widely around a median value of 71.1. The average value of S is 1.9 so on
average the last term (MUL S 1.177709) comes to about 2.2. This enables entropy (Section 4.7.9) to
P 0
play a marginal role where 350 (Quali+1 − Gi+1 )/Quali is near its average value but the evolved predictor’s performance comes mostly from the prepass through the whole DNA sequence. The prepass uses the
Solexa quality values and whether or not the next DNA base is a G. However the presence or absence of G
bases is relatively unimportant.
Again we see the Solexa quality data playing a dominant role in the evolved predictor. Since, in this case,
we are asking GP to predict the quality of the match found by BLAST this is perhaps less surprising.
However evolution has managed to find other information and weave it into the classifier to improve it
above simply using the quality indicators provided by Solexa.
5.4
Predicting Repeats in the Human Genome
After one hundred generations GP evolved a program which scores 54% on all the training data. Its median
score is 46% on the validation data (see × in Figure 6). Although obviously reduced from its performance
on its training data, again this is highly significant. Across all 31 000 verification sequences χ2 = 910
(1 dof).
Only 14% of the training data give rise to multiple high quality BLAST matches. So although each generation the fitness function tests the evolving population on a 50:50 balanced dataset (see Section 4.6)
good multiple matches are very much the minority class. Despite this class imbalance the median fraction
of multiple matches correctly predicted is 83% on training data and 83% across the verification datasets
(Figure 11).
RN/13/12
Page 13
Using Genetic Programming to Model Software
W. B. Langdon and M. Harman
Validation dataset
1
0.9
Multiple ok
Fraction Correct
median 83%
0.7
0.6
Multiple matches found Generation 100
0.5
0.4
NA12878
1770
NA12891
NA12892
Figure 11: Plot on 31 verification datasets showing GP on average correctly predicts 83% of repeated
DNA sequences (the minority class). Error bars indicate standard error [Hotelling, 1953, p212].
The three trees of the smallest best of generation 100 program have a combined size of 254. Hence we try to
explain each part of it individually. GP’s overall prediction is given by the sign of Sum1−2M36 . However
M36 tends to be small compared to Sum1. Thus the sign of Sum1 plays the major role in predicting repeated
Human DNA sequences.
Even the simplified prepas1 tree used to calculate Sum1 has 96 elements. Each DNA base it forms a nonlinear composition of data calculated by prepas0, data from the last time it was called and G (i.e. if the
current base is a G or not, Section 4.7.3). Thus the prediction depends only on G and, via prepas0, Qual
and Run. In fact G only plays a minor role, suggesting GP has found a way to predict if BLAST will find
multiple high quality matches which mostly uses the quality of the Solexa scan and BLAST’s dislike of
simple patterns (which tend to give Run larger values).
Obviously predicting repeated patterns in the three billion Human DNA bases from a sample of just 36 is
very hard. So it is surprising that the evolved predictor can find 83% of those reported by BLAST. However
the above analysis hints that GP is not only using the 36 bases but also inferring something about the way
BLAST deals with and reports repeats.
RN/13/12
Page 14
Using Genetic Programming to Model Software
6
W. B. Langdon and M. Harman
Conclusions
Considerable manual effort is needed to create programs. In many cases problems do not have clear cut
solutions but instead there may be a range of solutions which make different compromises between various
benefits and differing types of cost. Where the solutions involve software, it may be impossible to know
how good a trade-off a solution offers until the software has been coded. Today even identify new operating
points is often highly labour intensive and few projects can afford to even explore more than one possibility by hand. Automated software production offers the prospect of exploring complete Pareto trade-off
surfaces, for example, between functionality and speed.
As a prelude to this, we have explored modelling critical aspects of an important non-trivial program
(BLAST). Since in future we hope to consider automatically generating bespoke software, we deliberately
restricted our analysis to one of the many applications where the program might be applied. We chose
mapping human genetic variation using Next-Gen DNA sequences from the 1000 genomes project. This
is an important task for which the data are both plentiful and freely available online and it is well beyond
our target program’s current abilities. Our goal is not to make better generic code but to improve existing
code by specialising it to each task. The importance of the 1000 genomes project and the shear volumes of
data make this an ideal candidate for dedicated software. If software was cheap enough other applications
would each have their own bespoke code.
Although our models were created by considering BLAST’s outputs, BLAST is the de facto standard for
Biological string matching which other programs seek to emulate. Thus although our models deliberately
apply only to short (36 bases) human DNA sequences generated by the Solexa scanners used by the 1000
genomes project, they can be thought of as embodying to a greater or less extent aspects of an idealised
mapping between this Solexa data and the reference human genome. Thus they might also apply to other
tools (e.g. BWA and Bowtie2) which emulate BLAST but are much faster. Our models cannot replace such
programs but where data are plentiful they might be used to prioritise them. E.g. so that DNA sequences
which are expected to yield unique high quality matches are processed first.
One of the critical components of string matching is the use of prefect match “seed” regions. These are
short fixed length sequences which exactly match the database and so correspond to our length of matching
region. As explained in Section 5.2, correctly choosing the seed length has a considerable impact on
performance. Thus, although our models are not perfect, being able to predict the length of the region
where the string match exactly would be of great benefit.
This modelling work shows GP can predict the quality of the match (E) (Section 5.1), length of matching
sequences (Section 5.2), if a sequence will match (Section 5.3), and can even find 83% of multiple matches
(Section 5.4). Analysis of the evolved predictors of E value and length of match shows a strong dependence
on Solexa’s two quality indicators. Solexa’s Qual also figures in the evolved prediction of whether there is
a match or not. It also depends upon the sequence’s entropy. Whilst other factors seem to relate to quality
of the input data, the use of entropy hints at a degree of reverse engineering of BLAST, which does not
return low entropy sequences.
There are many repeated sequences in the human genome. They make it difficult to map genetic differences,
such as SNPs, (one of the main goals of the 1000 genomes project) and therefore “reads that map to multiple
sites in the genome are usually discarded” [Cheung et al., 2011, page 1]. Not only are such matches useless
they also slow down search and yet GP can predict 83% of them in advance.
Although it took BLAST 108 hours, excluding I/O, to process 31 000 DNA sequences (Section 5), the GP
predictors take on average less than 0.3 seconds.
Acknowledgements
I would like to thank Caroline Johnston, Jeremy Leipzig, Keith James, Pablo Pareja, Pierre Lindenbaum
and Larry Parnell. Funded by EPSRC grant EP/I033688/1.
RN/13/12
Page 15
Using Genetic Programming to Model Software
W. B. Langdon and M. Harman
References
[Aldecoa-Otalora et al., 2009] Estibaliz Aldecoa-Otalora, William B. Langdon, Phil Cunningham, and
Matthew J. Arno. Unexpected presence of mycoplasma probes on human microarrays. BioTechniques,
47(6):1013–1016, December 2009.
[Altschul et al., 1997] Stephen F. Altschul, Thomas L. Madden, Alejandro A. Schaffer, Jinghui Zhang,
Zheng Zhang, Webb Miller, and David J. Lipman. Gapped BLAST and PSI-BLAST a new generation
of protein database search programs. Nucleic Acids Research, 25(17):3389–3402, 1997.
[Cheung et al., 2011] Ming-Sin Cheung, Thomas A. Down, Isabel Latorre, and Julie Ahringer. Systematic bias in high-throughput sequencing data and its correction by BEADS. Nucleic Acids Research,
39(15):e103, August 2011.
[Durbin, et al., 2010] Richard M. Durbin, et al. A map of human genome variation from population-scale
sequencing. Nature, 467(7319):1061–1073, 28 Oct 2010.
[Foster, 2001] James A. Foster. Review: Discipulus: A commercial genetic programming system. Genetic
Programming and Evolvable Machines, 2(2):201–203, June 2001.
[Gathercole and Ross, 1994] Chris Gathercole and Peter Ross. Some training subset selection methods for
supervised learning in genetic programming. Presented at ECAI’94 Workshop on Applied Genetic and
other Evolutionary Algorithms, 1994.
[Harman et al., ] Mark Harman, Yue Jia, William B. Langdon, Tim Menzies, and Shin Yoo. ALERTI:
adaptive learning to evolve radical tuning iteratively. International Journal on Software Tools for Technology Transfer. Invited opinion corner.
[Harman et al., 2012] Mark Harman, William B. Langdon, Yue Jia, David R. White, Andrea Arcuri, and
John A. Clark. The GISMOE challenge: Constructing the Pareto program surface using genetic programming to find better programs. In The 27th IEEE/ACM International Conference on Automated
Software Engineering (ASE 12), pages 1–14, Essen, Germany, September 3-7 2012. ACM.
[Hotelling, 1953] Harold Hotelling. New light on the correlation coefficient and its transforms. Journal of
the Royal Statistical Society. Series B (Methodological), 15(2):193–232, 1953.
[IHGSC, 2001] International Human Genome Sequencing Consortium. Initial sequencing and analysis of
the human genome. Nature, 409(6822):860–921, 15 Feb 2001.
[Karplus et al., 1998] K Karplus, C Barrett, and R Hughey. Hidden Markov models for detecting remote
protein homologies. Bioinformatics, 14(10):846–856, 1998.
[Koza, 1994] John R. Koza. Genetic Programming II Automatic Discovery of Reusable Programs. MIT
Press, 1994.
[Langdon, ] William B. Langdon. Large scale bioinformatics data mining with parallel genetic programming on graphics processing units. In Shigeyoshi Tsutsui and Pierre Collet, editors, Massively Parallel
Evolutionary Computation on GPGPUs, Natural Computing Series, chapter 15. Springer. Forthcoming.
[Langdon and Arno, 2012] W. B. Langdon and M.J. Arno. In Silico infection of the human genome. In
Mario Giacobini, Leonardo Vanneschi, and William S. Bush, editors, 10th European Conference on
Evolutionary Computation, Machine Learning and Data Mining in Bioinformatics, EvoBIO 2012, volume 7246 of LNCS, pages 245–249, Malaga, Spain, 11-13 April 2012. Springer Verlag.
[Langdon and Harman, ] William B. Langdon and Mark Harman. Optimising existing software with genetic programming. IEEE Transactions on Evolutionary Computation. Accepted.
RN/13/12
Page 16
Using Genetic Programming to Model Software
W. B. Langdon and M. Harman
[Langdon and Harman, 2010] W. B. Langdon and M. Harman. Evolving a CUDA kernel from an nVidia
template. In Pilar Sobrevilla, editor, 2010 IEEE World Congress on Computational Intelligence, pages
2376–2383, Barcelona, 18-23 July 2010. IEEE.
[Langdon, 1998] William B. Langdon. Genetic Programming and Data Structures. Kluwer, Boston, 1998.
[Langdon, 2010] W. B. Langdon. A many threaded CUDA interpreter for genetic programming. In
Anna Isabel Esparcia-Alcazar, Aniko Ekart, Sara Silva, Stephen Dignum, and A. Sima Uyar, editors,
Proceedings of the 13th European Conference on Genetic Programming, EuroGP 2010, volume 6021
of LNCS, pages 146–158, Istanbul, 7-9 April 2010. Springer.
[Langdon, 2011] W. B. Langdon. Graphics processing units and genetic programming: An overview. Soft
Computing, 15:1657–1669, August 2011.
[Langdon, 2012] W. B. Langdon. Genetic improvement of programs. In Radomil Matousek, editor, 18th
International Conference on Soft Computing, MENDEL 2012, Brno, Czech Republic, 27-29 June 2012.
Brno University of Technology. Invited keynote.
[Langdon, 2013] W. B. Langdon. Which is faster: Bowtie2GP > Bowtie > Bowtie2 > BWA. In Francisco
Luna, editor, GECCO 2013 Late breaking abstracts workshop, Amsterdam, 6-10 July 2013. ACM.
[Langmead and Salzberg, 2012] Ben Langmead and Steven L Salzberg. Fast gapped-read alignment with
Bowtie 2. Nature Methods, 9(4):357–359, 4 March 2012.
[Langmead et al., 2009] Ben Langmead, Cole Trapnell, Mihai Pop, and Steven Salzberg. Ultrafast
and memory-efficient alignment of short DNA sequences to the human genome. Genome Biology,
10(3):R25, 2009.
[Moreno-Torres et al., 2009] Jose Garcia Moreno-Torres, Xavier Llora, David E. Goldberg, and Rohit
Bhargava. On the homogenization of data from two laboratories using genetic programming. In
Jaume Bacardit, Will N. Browne, Jan Drugowitsch, Ester Bernadó-Mansilla, and Martin V. Butz, editors, Learning Classifier Systems, volume 6471 of Lecture Notes in Computer Science, pages 185–197.
Springer, 2009.
[Petke et al., 2013] Justyna Petke, William B. Langdon, and Mark Harman. Applying genetic improvement to MiniSAT. In Gordon Fraser, editor, Symposium on Search-Based Software Engineering,
Leningrad, August 24-26 2013. Short Papers.
[Poli and Langdon, 1999] Riccardo Poli and William B. Langdon. Sub-machine-code genetic programming. In Lee Spector, William B. Langdon, Una-May O’Reilly, and Peter J. Angeline, editors, Advances
in Genetic Programming 3, chapter 13, pages 301–323. MIT Press, 1999.
[Poli et al., 2008] Riccardo Poli, William B. Langdon, and Nicholas Freitag McPhee.
A field
guide to genetic programming.
Published via http://lulu.com and freely available at
http://www.gp-field-guide.org.uk, 2008. (With contributions by J. R. Koza).
[Poli et al., 2009] Riccardo Poli, Nicholas Freitag McPhee, Luca Citi, and Ellery Crane. Memory with
memory in genetic programming. Journal of Artificial Evolution and Applications, 2009. Article ID
570606.
[Singleton, 1994] Andy Singleton. Genetic programming with C++. BYTE, pages 171–176, February
1994.
[Teller, 1994] Astro Teller. The evolution of mental models. In Kenneth E. Kinnear, Jr., editor, Advances
in Genetic Programming, chapter 9, pages 199–219. MIT Press, 1994.
[Upton et al., 2008] Graham JG Upton, William B Langdon, and Andrew P Harrison. G-spots cause incorrect expression measurement in Affymetrix microarrays. BMC Genomics, 9:613, 2008.
RN/13/12
Page 17
| 9cs.NE
|
1
Random Caching in Backhaul-Limited
Multi-Antenna Networks: Analysis and Area
Spectrum Efficiency Optimization
arXiv:1709.06278v1 [cs.IT] 19 Sep 2017
Sufeng Kuang, Student Member, IEEE, Nan Liu, Member, IEEE
Abstract
Caching at base stations is a promising technology to satisfy the increasing capacity requirements
and reduce the backhaul loads in future wireless networks. Careful design of random caching can fully
exploit the file popularity and achieve good performance. However, previous works on random caching
scheme usually assumed single antenna at BSs and users, which is not the case in practical multi-antenna
networks. In this paper, we consider the analysis and optimization in the cache-enabled multi-antenna
networks with limited backhaul. We first derive a closed-form expression and a simple tight upper
bound of the successful transmission probability, using tools from stochastic geometry and a gamma
approximation. Based on the analytic results, we then consider the area spectrum efficiency maximization
by optimizing design parameters, which is a complicated mixed-integer optimization problem. After
analyzing the optimal properties, we obtain a local optimal solution with lower complexity. To further
simplify the optimization, we then solve an asymptotic optimization problem in the high user density
region, using the upper bound as the objective function. Numerical simulations show that the asymptotic
optimal caching scheme achieves better performance over existing caching schemes. The analysis and
optimization results provide insightful design guidelines for random caching in practical networks.
Index Terms
Cache, limited backhaul, multi-antenna, Poisson point process, stochastic geometry
I. I NTRODUCTION
The deployment of small base stations (SBSs) or network densification, is proposed as a key
method for meeting the tremendous capacity increase in 5G networks [1]. Dense deployment of
S.Kuang and N.Liu are with the National Mobile Communications Research Laboratory, Southeast University, Nanjing 210096,
China.(Email: [email protected] and [email protected]).
2
small cells can significantly improve the performance of the network by bringing the BSs closer
to the users. On the other hand, MIMO technology, especially the deployment of massive number
of antennas at BSs, is playing an essential role in 5G networks to satisfy the increasing data
requirements of the users. Therefore, the combination of small cells and novel MIMO techniques
is inevitable in 5G networks [2]. However, this approach aggravates the transmission loads of
the backhaul links connecting the SBSs and the core networks.
Caching at BSs is a promising method to alleviate the heavy backhaul loads in small cell
networks [3]. Therefore, jointly deploying cache and multi antennas at BSs is proposed to
achieve 1000x capacity increase for 5G networks [4]–[6]. In [4] and [5], the authors considered
the optimization of the cooperative MIMO for video streaming in cache-enabled networks. In
[6], the authors considered the optimal multicasting beamforming design for cache-enabled cloud
radio access network (C-RAN). Note that the focuses of the above works were on the parameter
optimization for cache-enabled networks under the traditional grid model.
Recently, Poisson point process (PPP) is proposed to model the BS locations to capture the
irregularity and randomness of the small cell networks [7]. Based on the random spatial model,
the authors in [8] and [9] considered storing the most popular files in the cache for small
cell networks and heterogeneous cellular networks (HetNets). In [10], the authors considered
uniformly caching all the files at BSs, assuming that the file request popularity follows a uniform
distribution. In [11] , the authors considered storing different files at the cache in an i.i.d. manner,
in which each file is selected with its request probability. Note that in [8]–[11], the authors did
not consider the optimal cache placement, rather, they analyzed the performance under a given
cache placement, and thus, the results of the papers might not yield the best performance of the
system.
In view of the above problem, random caching strategy is proposed to achieve the optimal
performance [12]–[15]. In [12] and [13], the authors considered the caching and multicasting in
the small cell networks or the HetNets, assuming random caching at SBSs. In [14], the authors
considered the analysis and optimization of the cache-enabled multi-tier HetNets, assuming
random caching for all tiers of the BSs. Note that in [12]–[14], the authors obtained a waterfilling-type optimal solutions in some special cases due to the Rayleigh distribution of the fading.
In [15] and [16], the authors considered the helper cooperation for the small cell networks or
the HetNets, in which the locations of BSs are modeled as Poisson cluster Process.
However, the works mentioned above considered the random caching in networks equipped
3
with a single antenna at BSs and users, which is rarely the case in practical networks. In [17],
the authors considered the cache distribution optimization in HetNets, where multi-antennas are
deployed at the MBSs. However, the authors considered a special case of zero forcing precoding,
i.e., the number of the users is equal to the number of BS antennas, in which the equivalent
channel gains from the BSs to its served users follow the Rayleigh distribution. This property
does not hold for the general MIMO scenario, where the equivalent channel gains from the BSs
to its served users follow the Gamma distribution [18], [19].
The main difficulty of the analysis of the multi-antenna networks stems from the complexity
of the random matrix channel [20]–[22]. In [20], the authors utilized the stochastic ordering to
compare MIMO techniques in the HetNets, but the authors did not comprehensively analyze
the SINR distribution. In [21], the authors proposed to utilize a Toeplitz matrix representation
to obtain a tractable expression of the successful transmission probability in the multi-antenna
small cell networks. In [22]–[24], the authors extended the approach of the Toeplitz matrix
representation to analyze the MIMO mutli-user HetNets, MIMO networks with interference
nulling and millimeter wave networks with directional antenna arrays. However, this expression
involves the matrix inverse, which is difficult for analysis and optimization. In [25], a gamma
approximation [26] was utilized to facilitate the analysis in the millimeter wave networks.
In this paper, we consider the analysis and optimization of random caching in backhaullimited multi-antenna networks. Unlike the previous works [12]–[14] focusing on the successful
transmission probability of the typical user, we analyze and optimize the area spectrum efficiency,
which is a widely-used metric to describe the network capacity. The optimization is over file
allocation strategy and cache placement strategy, where a file allocation strategy dictates which
file should be stored at the cache of the BSs and which file should be transmitted via the
backhaul, and a cache placement strategy is to design the probability vector according to which
the files are randomly stored in the cache of the BSs.
First, we derive an exact expression of the successful transmission probability in cacheenabled multi-antenna networks with limited backhaul, using tools from stochastic geometry
and a Toeplitz matrix representation. We then utilize a gamma approximation to derive a tight
upper bound on the performance metrics to facilitate the parameter design. The exact expression
involves the inverse of a lower triangular Toeplitz matrix and the upper bound is a sum of a
series of fractional functions of the caching probability. These expressions reveal the impacts of
the parameters on the performance metrics, i.e., the successful transmission probability and the
4
area spectrum efficiency. Furthermore, the simple analytical form of the upper bound facilitates
the parameter design.
Next, we consider the area spectrum efficiency maximization by jointly optimizing the file
allocation and cache placement, which is a very challenging mixed-integer optimization problem.
We first prove that the area spectrum efficiency is an increasing function of the cache placement.
Based on this characteristic, we then exploit the properties of the file allocation and obtain a
local optimal solution in the general region, in which the user density is moderate. To further
reduce the complexity, we then solve an asymptotic optimization problem in the high user density
region, using the upper bound as objective function. Interestingly, we find that the optimal file
allocation for the asymptotic optimization is to deliver the most B popular files via the backhaul
and store the rest of the files at the cache, where B is the largest number of files the backhaul
can deliver at same time. In this way, the users requesting the most B popular files are associated
with the nearest BSs, who obtain the most B popular files via the backhaul, and therefore achieve
the optimal area spectrum efficiency.
Finally, by numerical simulations, we show that the asymptotic optimal solution with low
complexity achieves a significant gain in terms of area spectrum efficiency over previous caching
schemes.
II. S YSTEM M ODEL
A. Network Model
We consider a downlink cache-enabled multi-antenna network with limited backhaul, as shown
in Fig. 1, where BSs, equipped with N antennas, are distributed according to a homogeneous
Poisson point process (PPP) Φb with density λb . The locations of the single-antenna users are
distributed as an independent homogeneous PPP Φu with density λu . According to Slivnyak’s
theorem [27], we analyze the performance of the typical user who is located at the origin without
loss of generality. All BSs are operating on the same frequency band and the users suffer intercell
interference from other BSs. We assume that all the BSs are active due to high user density.
We assume that each user requests a certain file from a finite content library which contains
F files. Let F = {1, 2, 3, · · · , F } denotes the set of the files in the network. The popularity of
the requested files is known a priori and is modeled as a Zipf distribution [28]
f −γ
qf = PF −γ ,
i=1 i
(1)
5
Backhaul File
Cache File
BS
User
Cache
Backhhaul
Distance-based Association.
Content-centric Association.
Fig. 1. Illustration of the system model. For the backhaul file, the distance-based association scheme is adopted. For the cached
file, the content-centric association scheme is adopted.
where qf is the probability that a user requests file f and γ is the shape parameter of the Zipf
distribution. We assume that all the files have same size and the size of a file is normalized to
1 for simplicity.
Each BS is equipped with a cache with C segments and the cache can store at most C different
files out of the content library. For the files which are not stored in the cache, the BSs can fetch
them from the core network via backhaul links which can transmit at most B files at same time.
We refer to C as cache size and B as backhaul capability. We assume B + C ≤ F to illustrate
the resource limitation.
B. Caching and Backhaul Delivery
The set of all the files is partitioned into two disjoint sets, and we define the set of files stored
in the cache of all the BSs and the set of files not stored in any of the BSs, i.e., the files in it
must be transmitted via the backhaul, as cached file set and backhaul file set, which are denoted
as Fc and Fb , respectively. The number of files in Fx is Fx , x = c, b. Since Fc and Fb form a
partition of F, we have
Fc
[
Fb = F, Fc
\
Fb = ∅.
(2)
We define the process of designing Fc and Fb as file allocation.
Each BS can cache C different files from a total Fc files via the random caching scheme, in
which the BS stores a certain file i ∈ Fc randomly with probability ti . Let t = (ti )i∈Fc denotes
6
the caching distribution of all the files. Then we have the following constraints
ti ∈ [0, 1], ∀i ∈ Fc ,
X
ti ≤ C.
(3)
(4)
i∈Fc
We refer to the specification of t as cache placement.
For further analysis, we define the BSs that have the cached file f ∈ Fc in the cache as the
f-cached BSs. According to the thinning theory of the PPP, the density of the f -cached BSs is
λfb = tf λb . We denote the set of the f -cached BSs as Φfb and the set of the remaining BSs that
do not have file f in their cache as Φ−f
b . When a user requests the file f out of the cached
file set, the user is associated with the nearest f -cached BSs which will provide the required
file f from its cache. We refer to the BS associated with the typical user as the tagged BS and
this association scheme is called content-centric association scheme. Content-centric association
scheme is different from the traditional distance-based association scheme, where the user is
associated with the nearest BS. We denote the tagged BS as BS 1.
When a user requests a file out of the backhaul file set, the distance-based association is
adopted and the user is associated with the nearest BS. We define the set of backhaul requested
r
files as the set of the backhaul files requested by the users of BS i, which is denoted as Fb,i
⊆ Fb .
r
The number of the backhaul request file is denoted as Fb,i
. If BS i needs to transmit less than
r
r
B backhaul files via backhaul, i.e., Fb,i
≤ B, then BS i gets all Fb,i
files from the backhaul and
transmits all of them to the designated users; otherwise, BS i will randomly select B different
r
files from Fb,i
according to the uniform distribution, where these files will be transmitted from
the backhaul links and then passed on to the designated users.
III. P ERFORMANCE M ETRIC AND P ROBLEM F ORMULATION
A. Performance Metric
In this part, we define the successful transmission probability (STP) and the area spectrum
efficiency (ASE) of the typical user when the single-user maximal ratio combination (MRT)
beamforming is adopted. We consider the single-user MRT beamforming due to the low complexity of the beamforming design, which is suitable for the scenario when a large number
of antennas are deployed at the BSs. For other MIMO scenarios, the equivalent channel gain
follows the gamma distribution with different parameters [18], [19]. Therefore, the method of the
7
analysis here can be extended to the cache-enabled MIMO networks with different precoding and
combining strategies. We assume that the transmitter can get the perfect channel state infomation
(CSI) through the feedback from the users. The BSs do not have the CSI of the other cells due
to the high BS density.
We refer to the typical user as user 0 and it is served by the tagged BS located at x1 . Due to
the assumption of single-user MRT, each BS serves one user per resource block (RB). Hence,
the received signal of the typical user on its resource block is
β
y0f = kx1 k− 2 h∗0,1 w1 s1 +
X
β
kxi k− 2 h∗0,i wi si + n0 ,
(5)
i∈{Φb \1}
where wi is the beamforming vector of BS i to its served user, f is the file requested by the
typical user, xi is the location of BS i, h0,i ∈ CN ×1 is the channel coefficient vector from BS i
to the typical user, si ∈ C1×1 is the transmitting message of BS i and n0 is the additive white
Gaussian noise (AWGN) at the receiver. We assume that E[si ∗ si ] = P for any i and P is the
transmit power of the BS. The elements of the channel coefficient vector h0,i are independent
and identical complex Gaussian random variables, i.e., CN (0, 1). β > 2 is the pathloss exponent.
For single-user MRT, to maximize the channel gain from the BS to its served user, the
beamforming at BS i is wi =
hi
khi k
[29], where hi ∈ CN ×1 is the channel coefficient from
BS i to its served user. Thus the SIR of the typical user requiring file f , whether f is in the
cached file set or the backhaul file set, is given by
SIRf = P
where gi =
kh∗0,i hi k2
khi k2
P kx1 k−β g1
,
−β g
i
i∈{Φb \1} P kxi k
(6)
is the equivalent channel gain (including channel coefficient and the
beamforming) from BS i to the typical user. It is shown that the equivalent channel gain from the
tagged BS to its served user, i.e., g1 ∼ Gamma(N, 1) and gi ∼ Exp(1), ∀i > 1 [29]. In this paper,
we consider SIR for performance analysis rather than SINR due to the dense deployments of the
SBSs. In simulations, we include the noise to illustrate that in dense networks the consideration
of the SIR achieves nearly the same performance as that of the SINR.
For the single-user MRT, the successful transmission probability (STP) of the typical user is
defined as the probability that the SIR is larger than a threshold, i.e.,
Ps (Fc , t) =
X
f ∈Fc
qf P (SIRf > τ ) +
X
f ∈Fb
qf P (SIRf > τ, f transmitted through backhaul) , (7)
8
where τ is the SIR threshold. As mentioned before, f is transmitted by the backhaul if the
r
number of the backhaul requested files of the tagged BS Fb,1
is no more than the backhaul
capability B, or if it is been chosen to be transmitted according to the uniform distribution in
r
r
the event where Fb,1
> B. Note that the STP is related to the random variables Fb,1
and SIRf .
We use the area spectrum efficiency (ASE) as the metric to describe the average spectrum
efficiency per area. The ASE of the single-user MRT is defined as [30], [31]
R(Fc , t) = λb Ps (Fc , t) log2 (1 + τ ).
(8)
where the unit is bit/s/Hz/km2 . Note that the ASE reveals the relationship between the BS density
and the network capacity.
B. Problem Formulation
Under given backhaul capability B and cache size C, the caching strategy, i.e., the file
allocation strategy and the cache placement strategy, fundamentally affects the ASE. We study the
problem of maximizing the ASE via a careful design of file allocation Fc and cache placement
t as follows
max R(Fc , t)
Fc ,t
s.t
(9)
(2), (3), (4).
We will derive the expressions of the STP and the ASE in Section IV and solve the ASE
maximization problem, i.e., the problem in (9), in Section V.
IV. P ERFORMANCE A NALYSIS
In this section, we first derive an exact expression of the STP and ASE under given file
allocation and cache placement strategy, i.e., under given Fc and t. Then we utilize a gamma
distribution approximation to obtain a simpler upper bound of the STP and ASE.
A. Exact Expression
In this part, we derive an exact expression of the STP and the ASE using tools from stochastic
geometry. In general, the STP is related to the number of the backhaul request file of the tagged
9
r
BS, i.e., Fb,1
. Therefore, to obtain the STP, we first calculate the probability mass function (PMF)
r
of Fb,1
.
r
r
Lemma 1. (pmf of Fb,1
) When f ∈ Fb is requested by the tagged BS, the pmf of Fb,1
is given by
r
b
= k = g ({Fb \ f }, k − 1) ,
PF
Fb,1
f
ä
Ä
k ∈ {1, 2, · · · , Fb },
(10)
where g(B, k) is given by
4
X
g (B, k) =
Ç
q i λu
1− 1+
3.5λb
Y
Y∈{X ⊆B:|X |=k} i∈Y
å−4.5 !
Y
Ç
qi λu
1+
3.5λb
i∈B\Y
å−4.5
.
(11)
Proof: See Appendix A.
Based on Lemma 1, we have the following corollary.
r
r
Corollary 1. (The pmf of Fb,1
when λu → ∞). When λu → ∞, the pmf of Fb,1
is
0,
k = 1, 2, · · · , Fb − 1
1,
k = Fb
r
b
lim PF
Fb,1
=k =
f
λu →∞
Ä
ä
.
(12)
r
converges to constant Fb in distribution as λu → ∞. The
Corollary 1 interprets that Fb,1
asymptotic result is consistent with the fact that when the user density is high, each BS will
have many users connected to it, and thus, each BS will require all the backhaul files.
We then calculate the STP, which is defined in (7). Base on Lemma 1, we can rewrite the
r
. Therefore, we can obtain the
STP as a combination of the STP conditioned on the given Fb,1
STP in Theorem 1.
Theorem 1. (STP) The STP is given by
Ps (Fc , t) =
X
qf Psf,c (tf ) +
f ∈Fc
X
f ∈F \Fc
qf
Fb
Ä
X
PF b F r
b,1
f
=k
k=1
ä
B
P b,
max (k, B) s
(13)
r
b
where PF
Fb,1
= k is given in (10), qf is given in (1), Psf,c (tf ) and Psb are the STPs of the
f
Ä
ä
cached file f ∈ Fc and the backhaul file f ∈ F \ Fc , which are given by
Psf,c (tf )
tf
=
tf + l0c,f
1
Psb =
1 + l0b,f
τ 2\β
I−
tf + l0c,f
"
τ 2\β
I−
1 + l0b,f
"
#−1
!
D
c,f
,
(14)
1
#−1
!
Db,f
,
1
(15)
10
where k · k1 is the l1 induced matrix norm (i.e, ||B||1 = max1≤j≤n
Pm
i=1
|bij |, B ∈ Rm×n ), I is
an N × N identity matrix, Dc,f and Db,f are N × N Toeplitz matrices of the cached file f ∈ Fc
and the backhaul file f ∈ F \ Fc , which are given by
n,f
l
1
n,f
l2
.
..
,
0
Dn,f =
0
l1n,f
0
..
.
..
.
n ∈ {c, b},
(16)
n,f
n,f
n,f
0
lN
lN
−1, · · · l1
where l0c,f , l0b,f , lic,f and lib,f are given by
l0c,f
l0b,f
lic,f
lib,f
ñ
ô
Ç
2
2
2π
2τ
; 2 − ; −τ + (1 − tf ) csc
= tf
2 F1 1, 1 −
β−2
β
β
β
ñ
ô
2τ
2
2
=
; 2 − ; −τ ,
2 F1 1, 1 −
β−2
β
β
ñ
2τ i−2\β
2 2
2
= (1 − tf ) B( + 1, i − ) + tf
2 F1 i + 1, i −
β β
β
iβ − 2
ñ
ô
2τ i−2\β
2
2
=
; i + 1 − ; −τ , ∀i ≥ 1.
2 F1 i + 1, i −
iβ − 2
β
β
å
2π
τ 2\β ,
β
(17)
(18)
2
2
; i + 1 − ; −τ
β
β
ô
, ∀i ≥ 1, (19)
(20)
Here, 2 F1 (·) is the Gauss hypergeometric function and B(·) is the Beta function.
Proof. Considering the equivalent channel gain g1 ∼ Gamma(N, 1), the STP is a complex n-th
derivative of the interference Laplace transform LI (s) [20]. Utilizing the approach in [21], we
obtain the expressions of the STPs in lower triangular Toeplitz matrix representation as (14) and
(15). For details, please see Appendix B.
According to (8), we then obtain the ASE under given Fc and t
R(Fc , t) = λb Ps (Fc , t) log2 (1 + τ ).
(21)
For backhaul-limited multi-antenna networks, the change of the BS density λb and user density
Ä
ä
r
λu influence Ps (Fc , t) via the pmf of the backhaul request file, i.e., P Fb,1
= k . However, when
λu approaches infinity and λb remains finite, since all the backhaul files are requested by the
r
b
tagged BS (shown in Corollary 1), PF
Fb,1
= k is no longer related to λb . Therefore, when
f
Ä
ä
the user density is high and the design parameter Fc and t are given, deploying more BSs will
11
always increase the ASE.
From Theorem 1 and the definition of the ASE, we can derive a tractable expression of the
ASE for N = 1, i.e., the backhaul-limited single-antenna networks. The ASE of the backhaullimited single-antenna networks is given in the following corollary.
Corollary 2. (ASE of Single-Antenna Networks) The ASE of the cache-enabled single-antenna
networks with limited backhaul is given by
Ñ
RSA (Fc , t) = λb log2 (1 + τ )
Fb
r
b
X X
= k qf B
PF
Fb,1
qf tf
f
+
ζ1 (τ )tf + ζ2 f ∈F \Fc k=1 max (k, B) (ζ1 (τ ) + ζ2 )
ä
Ä
X
f ∈Fc
é
, (22)
where ζ1 (τ ) and ζ2 (τ ) satisfy
ñ
2τ
2
2
ζ1 (τ ) = 1 +
; 2 − ; −τ
2 F1 1, 1 −
β−2
β
β
Ç
å
2π
2π
ζ2 (τ ) =
csc
τ 2\β .
β
β
ô
Ç
å
2π
2π
−
csc
τ 2\β ,
β
β
(23)
(24)
The proof of Corollary 2 is similar to the proof of Theorem 1 except that the equivalent
channel gain g1 ∼ Exp(1).
B. Upper Bound and Asymptotic Analysis
In this part, we first derive an upper bound of the STP under given Fc and t. We then give
the asymptotic analytic results in high user density region. First, we introduce a useful lemma
to present a lower bound of the gamma distribution.
Lemma 2. [26]: Let g be a gamma random variable follows Gamma(M, 1). The probability
P(g < τ ) can be lower bounded by
P(g < τ ) > 1 − e−aτ
î
óM
,
(25)
1
where α = (M !)− M .
Utilizing the above lemma, we then obtain the upper bound of the STP as follows.
Theorem 2. (Upper Bound of STP) The upper bound of the STP is given by
Psu (Fc , t) =
X
f ∈Fc
qf Psu,f,c (tf ) +
X
f ∈F \Fc
qf
Fb
Ä
X
PF b F r
f
k=1
b,1
=k
ä
B
P u,b ,
max (k, B) s
(26)
12
where Psu,f,c (tf ) and Psu,b are the upper bounds of the STPs of the cached file f ∈ Fc and the
backhaul file f ∈ F \ Fc , which are given by
Psu,f,c (tf )
Psu,b
(−1)i+1 Ni tf
=
,
i=1 (θA (i) tf + θC (i))
Ä ä
N
X
(27)
(−1)i+1 Ni
=
,
i=1 (θA (i) + θC (i))
Ä ä
N
X
(28)
where
ñ
ô
2τ
2
2
θA (i) = 1 +
; 2 − ; −iατ
2 F1 1, 1 −
β−2
β
β
Ç
å
2π
2π
θC (i) =
csc
(iατ )2\β .
β
β
Ç
å
2π
2π
−
csc
(iατ )2\β
β
β
(29)
(30)
1
Here, α = (N !)− N is a constant related to the number of BS antennas N .
Proof. The STP of the cached file f ∈ Fc is
Psf,c (tf ) = P g1 > τ Ikx1 kβ
Ä
Ä
(a)
≤ 1 − EIkx1 kβ
=
N
X
=
(−1)
=
1 − exp −ατ Ikx1 k
β
ääN ò
Ä
ä
N
EIkx1 kβ exp −iατ Ikx1 kβ
i
!
i+1
(−1)
i=1
(b)
Ä
!
i+1
i=1
N
X
ïÄ
ää
Ä
ä
N
Ekx1 kβ LI iατ kx1 kβ
i
(−1)i+1
N
X
i=1 tf
2τ
F
β−2 2 1
h
1, 1 − β2 ; 2 − β2 ; −iατ
i
ÄN ä
i
tf
+ 1 + (1 − tf ) 2π
csc
β
2π
β
(iατ )2\β
.
(31)
where (a) follows from g1 ∼ Gamma(N, 1) and Lemma 2, (b) follows from the PDF of kx1 k
fkx1 k (r) = 2πtf λb r exp {−πtf λb r2 } for f ∈ Fc and LI iτ kx1 kβ is given in Appendix B as
Ä
LI iτ kx1 kβ = exp −πλb
Ä
ä
2τ tf
F
β−2 2 1
h
1, 1 − β2 ; 2 − β2 ; −iατ
ä
i
+ (1 − tf ) 2π
csc
β
2π
β
(iατ )2\β kx1 k2 .
1
Here, α = (N !)− N .
Therefore, we obtain an upper bound of Psu,c,f (tf ) , ∀f ∈ Fc . We can obtain the expression
of Psu,b similarly and then finish the proof of Theorem 2.
The upper bound of the STP is a series of fractional functions of the cache placement t.
Comparing with the exact expression of the STP in Theorem 1, the upper bound approximates
the STP in a simpler manner, and therefore facilitates the analysis and further optimization.
13
According to Theorem 2, the upper bound of the ASE is given by
Ru (Fc , t) = λb Psu (Fc , t) log2 (1 + τ ).
(32)
To obtain design insights, we then analyze the upper bound of the ASE in the asymptotic
r
region, i.e., the high user density region. When λu → ∞, the discrete random variable Fb,1
→ Fb
in distribution as shown in Corollary 1. Therefore, we have the following corollary according to
Theorem 2.
Corollary 3. (Asymptotic Upper Bound of ASE) In high user density region, i.e., λu → ∞, the
asymptotic upper bound of the ASE is given by
Ö
Ru,∞ (Fc , t) = λb log2 (1 + τ )
è
X
qf Psu,f,c (tf ) +
f ∈Fc
X
f ∈F \Fc
qf B
P u,b
max (Fb , B) s
,
(33)
where Psu,f,c (tf ) and Psu,b are given in (27) and (28).
We plot Fig. 2 and Fig. 3 to validate the correctness of the analytic results. Fig. 2 plots the
successful transmission probability vs. the number of BS antennas and target SIR. Fig. 2 verifies
Theorem 1 and Theorem 2, and demonstrates the tightness of the upper bound. Fig. 2 also
shows that the successful transmission probability increases with the number of BS antennas
and decreases with the SIR threshold. Moreover, Fig. 2 (a) indicates that when the user density
is large, the increase of the BS density increases the successful transmission probability. Fig. 3
plots the area spectrum efficiency vs. user density by showing that when the user density is larger
than a certain threshold, i.e., 6 × 10−3 m−2 for the single-antenna networks and 4 × 10−3 m−2 for
the multi-antenna networks, the asymptotic upper bound of the ASE is nearly same as the ASE.
V. ASE O PTIMIZATION
A. General ASE Optimization
In this part, we solve the ASE maximization problem, i.e., maximize R(Fc , t) via optimizing
the file allocation Fc and the cache placement t. Based on the relationship between R(Fc , t)
and Ps (Fc , t), the ASE optimization problem is formulated as follows.
14
0.7
λ b =5 × 10 -4 m-2
0.8
0.7
0.6
λ b =10 -4 m-2
0.5
Upper Bound
Theo.
Monte Carlo
0.4
2
3
4
5
6
7
8
9
Theoritical
Monte Carlo
0.5
N=1
0.4
N=2
0.3
0.2
N=3
0.1
0.3
1
Upper Bound
0.6
Successful Transmission Probability
Successful Transmission Probability
0.9
0
-10
10
-5
0
5
10
15
20
SIR Threshold (dB)
The number of BS antennas N
(b) STP vs. SIR threshold τ at λb = 10−4 m−2 .
(a) STP vs. the number of BS antennas N at τ = 0dB.
Fig. 2. STP vs. the number of BS antennas N and SIR threshold τ . λu = 10−3 m−2 , β = 4, F = 8, B = C = 2,
Fb = {1, 2, 3, 4}, Fc = {5, 6, 7, 8}, t = (0.8, 0.6, 0.4, 0.2), γ = 1. In this paper, the transmit power is 6.3W, the
noise power in the Monte Carlo simulation is σn = −97.5dBm [32], the theoretical results are obtained without
consideration of noise. The Monte Carlo results are obtained by averaging over 106 random realizations.
55
Area Spectrum Efficiency(bits/Hz/km 2 )
N=3
Upper Bound
Theo.
Monte Carlo
Asym.
50
45
40
N=2
35
N=1
30
25
1
2
3
4
5
6
User Density λu (m -2 )
7
8
9
10
× 10-3
Fig. 3. The ASE vs. user density λu . λb = 10−4 m−2 , β = 4, F = 8, B = C = 2, Fb = {1, 2, 3, 4}, Fc = {5, 6, 7, 8},
t = (0.8, 0.6, 0.4, 0.2), γ = 1, τ = 0dB.
Problem 1. (ASE Optimization)
R∗ , max λb Ps (Fc , t) log2 (1 + τ )
Fc ,t
s.t.
(34)
(2), (3), (4).
where Ps (Fc , t) is given in (13).
The above problem is a mixed-integer problem to optimize the discrete parameter Fc and the
continuous parameter t. To solve the complex problem, we first explore optimal properties of
the discrete variable Fc and then optimize the continuous variable t. To optimally design the
15
file allocation Fc , we first study the properties of the STPs of the cached file and backhaul file,
i.e., Psf,c (tf ) and Psb . Based on the properties of the lower triangular Toeplitz matrix form in
(14) and (15), we then obtain the following lemma.
Lemma 3. Psf,c (tf ) and Psb have following properties:
1) Psf,c (tf ) is bounded by
tf
tf
≤ Psf,c (tf ) ≤
,
µA tf + νA
µB tf + νB
(35)
where µA , νA , µB and νB are given by
Ç
å
ñ
2π
2τ
2
2
2π
csc
τ 2\β +
; 2 − ; −τ
µA = 1 −
2 F1 1, 1 −
β
β
β−2
β
β
+
N
−1
X
i=1
N −i
2τ i
2
2
; i + 1 − ; −τ
2 F1 i + 1, i −
N
iβ − 2
β
β
ñ
ñ
ô
ô
ô
2 2
2
− B( + 1, i − )τ 2\β , (36)
β β
β
N
−1
X
2π
2π
N −i2 2
2
νA =
csc
τ 2\β −
B( + 1, i − )τ 2\β ,
β
β
N β β
β
i=1
Ç
å
Ç
å
ñ
2π
2π
2τ
2
2
µB = 1 −
csc
τ 2\β +
; 2 − ; −τ
2 F1 1, 1 −
β
β
β−2
β
β
+
N
−1 ñ
X
i=1
2
2
2τ i
; i + 1 − ; −τ
2 F1 i + 1, i −
iβ − 2
β
β
ñ
ô
(37)
ô
ô
2 2
2
− B( + 1, i − )τ 2\β ,
β β
β
N
−1
X
2π
2π
2
2 2
2\β
νB =
csc
τ
−
B( + 1, i − )τ 2\β ).
β
β
β
β
i=1 β
Ç
(38)
å
(39)
Furthermore, νA , νB are positive and µA , µB are no larger than 1.
2) Psf,c (tf ) is an increasing function of tf .
3) Psf,c (tf ) is no larger than Psb and no smaller than 0, i.e., 0 ≤ Psf,c (tf ) ≤ Psb .
Proof: See Appendix C.
Property 1 in Lemma 3 shows that the bounds of Psf,c (tf ) are fractional functions of tf . The
expressions of the bounds are similar to the expression of the STP of the cached file in singleantenna networks given in Theorem 2. Based on the properties of Psf,c (tf ) and Psb , we have the
following theorem to reveal the properties of the optimal file allocation Fc∗ .
Theorem 3. (Property of the Optimal File Allocation Fc∗ ) To optimize cached file set Fc for
Problem 1, we should allocate at least C files and at most F − B files to store in the cache,
that is to say, Fc∗ = {C, C + 2, · · · , F − B}.
16
Proof. We utilize the contradiction to prove Theorem 3. More specifically, we consider the cases
in which we cache less than C files or more than F − B files, and we then prove that the cases
are not optimal in terms of the ASE. For details, please see Appendix D.
The above theorem interprets that for optimal file allocation, the number of cached files
should be larger than the cache size and the number of the backhaul files shoud be larger than
the backhaul capability, indicating that we should make full use of the resources.
Based on Theorem 3, we can reduce the complexity of search for Fc∗ . Otherwise, we have
to check the cases such that Fc < C or Fc > F − B. When B or C is large, Theorem 3 will
largely reduce the complexity. Under given Fc , we optimize the cache placement t. The ASE
optimization over t under given Fc is formulated as follows
Problem 2. (Cache Placement Optimization under Given Fc )
R∗ (Fc ) , max λb
t
s.t.
X
qf Psf,c (tf ) log2 (1 + τ )
(40)
f ∈Fc
(3), (4).
where Psf,c (tf ) is given in (14).
The optimal solution for Problem 2 is denoted as t∗ (Fc ). The above problem is a continuous
optimization problem of a differentiable function over a convex set and we can use the gradient
projection method to obtain a local optimal solution. Under given Fc , we can obtain optimal
cache placement t∗ (Fc ) using Algorithm 1.
Algorithm 1 Optimal Solution to Problem 2
1.
2.
3.
4.
5.
6.
7.
8.
Initialization: n = 1, nmax = 104 and ti (1) = F1c for all i ∈ Fc . Constant lower triangular
Toeplitz matrix ∂B
is given in (55).
∂ti
repeat
CalculateÄDc,i with täi = ti (n) according to (16) for all i ∈ Fc .
Bi (n) = ti (n) + l0c,i I − τ 2\β Dc,i for all i ∈ Fc .
0
−1
∂B −1
ti (n+1) = ti (n)+s(n) B−1
λb log2 (1+τ )qi for all i ∈ Fc .
i (n) − ti (n)Bi (n) ∂ti Bi (n)
1
î
0
ó1
ti (n + 1)
=
ti (n + 1) − u∗ 0 for all i
∈
Fc , where
î
ó
P
0
∗ 1
1
i∈Fc ti (n + 1) − u 0 = C and [x]0 denotes max{min{1, x}, 0}.
n = n + 1.
until Convergence or n is larger than nmax .
u∗
satisfying
17
From Corollary 2, we can easily observe that when N = 1, i.e., in the single-antenna networks,
the objective function of Problem 2 is concave and the Slaters condition is satisfied, implying
that strong duality holds. In this case, we can obtain a closed-form optimal solution to the convex
optimization problem using KKT conditions. After some manipulations, we obtain the optimal
cache placement t∗ under given Fc for single-antenna networks.
Corollary 4. (Optimal Cache Placement of Single-Antenna Networks) When the cached file set
Fc is fixed, the optimal cache placement t∗ of the single-antenna networks is given by
1
t∗f =
ζ1 (τ )
Ñs
é1
λb log2 (1 + τ )qf ζ2 (τ )
− ζ2 (τ ) , f ∈ Fc ,
u∗
(41)
0
where [x]10 denotes max{min{1, x}, 0} and u∗ is the optimal dual variable for the constraint
∗ ti
i∈FC
P
≤ C, which satisfies
1
f ∈FC ζ1 (τ )
X
Ñs
é1
λb log2 (1 + τ )qf ζ2 (τ )
− ζ2 (τ ) = C.
u∗
(42)
0
Here ζ1 (τ ) and ζ2 (τ ) are given in (23) and (24), respectively.
Finally, combining the analysis of the properties of Fc∗ in Theorem 3 and the optimization of
t under given Fc , we can obtain (Fc∗ , t∗ ) to Problem 1. The process of solving Problem 1 is
summarized in Algorithm 2.
Algorithm 2 Optimal Solution to Problem 1
1. Initialization: R∗ = ∞.
2. for Fc = C : F − B do
3.
Choose Fc ∈ {X ⊆ F : |X | = Fc }.
4.
Obtain the optimal solution t∗ (Fc ) to Problem 2 using Algorithm 1 (when N > 1) or
Corollary 4 (when N = 1).
5.
if R∗ < R (Fc , t∗ (Fc )) then
6.
R∗ = R (Fc , t∗ (Fc )), (Fc∗ , t∗ ) = (Fc , t∗ (Fc )).
7.
end if
8. end for
Algorithm 2 includes two layers. In the outer layer, we search Fc∗ by checking all the possible
PF −B ÄF ä
i=C
i
choices, and in the inner layer, we utilize Algorithm 1 or Corollary 4 to obtain t∗ (Fc )
under given Fc . We refer to the optimal solution based on Algorithm 2 as Exact Opt..
18
B. Asymptotic ASE Optimization based on the Upper Bound in (33)
In Algorithm 2, we need to consider
PF −B ÄF ä
i=C
i
choices and the complexity is Θ(F F ). When F
is very large, Algorithm 2 is not acceptable due to high complexity. Furthermore, the expression
of Psf,c (tf ) is complex and we need to calculate the matrix inverse to obtain the derivative of
tf . Thus Algorithm 2 requires high complexity if the number of antennas is large. Note that
the upper bound of the STP closely approximates the STP, as illustrated in Fig. 2, therefore we
utilize the upper bound of the STP, i.e., Psu,f,c (tf ) to approximate Psf,c (tf ). To facilitate the
optimization, we formulate the problem of optimizing the asymptotic upper bound of the ASE
to provide insightful guidelines for the parameter design in the high user density region. Based
on the Corollary 3, the asymptotic optimization problem is formulated as follows.
Problem 3. (Asymptotic ASE Optimization)
∗
Ru,∞
, max Ru,∞ (Fc , t)
Fc ,t
s.t
(2), (3), (4).
(43)
The above problem is a mixed-integer problem. By carefully investigating the characteristic
of Ru,∞ (Fc , t), which is a series of fractional functions of t in (33), we obtain the following
lemma to reveal the properties of Ru,∞ (Fc , t).
Lemma 4. Under a given Fc , Ru,∞ (Fc , t) is an increasing function of tf for any f ∈ Fc .
Proof: See Appendix E.
Based on the properties of the asymptotic upper bound Ru,∞ (Fc , t), we then analyze the
properties of the optimal file allocation Fc∗ and obtain Fc∗ as a unique solution.
Theorem 4. (Asymptotic Optimal File Allocaiton) In high user density region, i.e., λu → ∞,
the optimal cached file FC∗ is given by FC∗ = {B + 1, B + 2, · · · , F }.
Proof. We first prove that the number of optimal cached files is F − B and then prove that the
optimal F − B cached files are the least F − B popular files. For details, please see Appendix
F.
Theorem 4 interprets that we should transmit B most popular files via the backhaul and cache
the remaining files when the user density is very large. Compared to the process of checking
Area Spectrum Efficiency(bps/Hz/km 2 )
19
80
70
60
Exact.Opt.
Asym.Opt.
50
40
4
3
2
Number of Antennas N
1
4000
5000
6000
7000
8000
9000
10000
User Density λu (km -2 )
Fig. 4. Comparison between Exact Opt. and Asym. Opt.. λb = 10−4 m−2 , β = 4, F = 6, B = C = 2, γ = 0.6, τ = 0dB.
PF −B ÄF ä
i=C
choices in Algorithm 2, we get a unique optimal solution of file allocation and thus
i
largely reduce the complexity. When Fc∗ is given, we only need to optimize the continuous
variable t. We then use the gradient projection method get the local optimal solution. The
algorithm is summarized in Algorithm 3.
Algorithm 3 Optimal Solution to Problem 3
1.
Initialization: n = 1, nmax = h 104 , Fc∗ = {B + 1, Bi + 2, · · · , F }, ti (1) = F1c for all
2τ
2
2
i ∈ Fc∗ , θA (j) = 1 + β−2
− 2π
csc 2π
(jατ )2\β , θC (j) =
2 F1 1, 1 − β ; 2 − β ; −jατ
β
β
2π
β
2.
3.
4.
5.
6.
csc
repeat
2π
β
(jατ )2\β for all j ∈ {1, 2, · · · , N }.
Ç
0
ti (n + 1) = ti (n) + s(n)
î
0
PN
(−1)j+1 (N
θ (j)
j) C
j=1 (θA (j)ti (n)+θC (j))2
ó
∗ 1
å
λb log2 (1 + τ )qi for all i ∈ Fc .
∈
Fc , where
ti (n + 1)
=
ti (n + 1) − u 0 for all i
î 0
ó
P
∗ 1
1
i∈Fc ti (n + 1) − u 0 = C and [x]0 denotes max{min{1, x}, 0}.
n = n + 1.
until Convergence or n is larger than nmax .
u∗
satisfying
Algorithm 3 is guaranteed to converge because the gradient projection method converges to
a local optimal point for solving a problem whose feasible set is convex. In Algorithm 3, the
calculation of the matrix inverse (Step 5 in Algorithm 1) and the search for Fc∗ (Step 3 in
Algorithm 2 ) are avoided. Therefore, Algorithm 3 achieves lower complexity comparing to
Algorithm 2. We refer to the caching scheme based on Algorithm 3 as Asym. Opt..
Now we utilize simulations to compare the proposed Exact Opt. (the optimal solution obtained
by Algorithm 2) and Asym. Opt. (the asymptotic optimal solution obtained by Algorithm 3). From
Fig. 4, we can see that the performance of Asym. Opt. is very close to that of Exact Opt., even
20
when the user density is low. Therefore, Algorithm 3 with low complexity is applicable and
effective for parameter design in general region.
VI. N UMERICAL R ESULTS
In this section, we compare the proposed asymptotic optimal caching scheme given by Algorithm 3 with three caching schemes, i.e., the MPC (most popular caching) scheme [8], the UC
(uniform caching) scheme [10] and the IID (identical independent distributed caching) scheme
[11]. In the MPC scheme, the BSs cache or use backhaul to deliver the most popular B +C files.
In the IID scheme, the BSs select B + C files to cache or transmit via the backhaul in an i.i.d
manner with probability qi for file i. In the UC scheme, the BSs select B+C files according to the
uniform distribution to cache or deliver via the backhaul. Note that in simulations, we consider the
noise. Unless otherwise stated, our simulation environment parameters are as follows: P = 6.3W,
σn = −97.5dBm, λb = 10−4 m−2 , λu = 5 × 10−3 m−2 , β = 4, F = 500, τ = 0dB.
Fig. 5 and Fig. 6 illustrate the area spectrum efficiency vs. different parameters. We observe
that the proposed asymptotic optimal scheme outperforms all previous caching schemes. In
addition, out of the previous caching schemes, the MPC scheme achieves the best performance
and the UC scheme achieves the worst performance.
Fig. 5 (a) plots the ASE vs. the number of BS antennas. We can see that the ASE of all the
schemes increases with the number of BS antennas. This is because the increase of the number
of BS antennas leads to larger spatial diversity and thus achieves better performance. It is shown
that the increase of the number of BS antennas leads to an increasing gap between the proposed
asymptotic optimal caching scheme and previous caching schemes. This is because the better
performance of a larger number of BS antennas leads to a larger gain when we exploit the file
diversity. Furthermore, for asymptotic optimal caching scheme, the less popular files are more
likely to be stored when the number of BS antennas is large. Fig. 5 (b) plots the ASE vs. the
Zipf parameter γ. We can see that the ASE of the proposed asymptotic optimal caching scheme,
the MPC scheme and the IID scheme increases with the increase of the Zipf parameter γ. This
is because when γ increases, the probability that the popular files are requested increases, and
hence, the users are more likely to require the popular files from the nearby BSs who cache the
files or obtain the files via backhaul. The change of the Zipf parameter γ have no influence to
the ASE of the UC scheme. This is because in the UC scheme, all the files are stored/fetched
40
80
35
70
Area Spectrum Efficiency(bps/Hz/km 2 )
Area Spectrum Efficiency(bps/Hz/km 2 )
21
30
25
20
15
Asym.Opt.
MPC
IID
UC
10
5
1
2
3
4
5
6
7
8
9
60
50
40
Asym.Opt.
MPC
IID
UC
30
20
10
0.3
10
0.4
0.5
0.6
The number of BS antennas N
0.7
0.8
0.9
1
1.1
1.2
Zipf Parameter γ
(a) Number of BS antennas at C = 30, B = 20, γ = 0.6.
(b) Zipf parameter at N = 8, C = 30, B = 20.
Fig. 5. ASE vs. the number of BS antennas N and Zipf parameter γ.
50
60
Area Spectrum Efficiency(bps/Hz/km 2 )
Area Spectrum Efficiency(bps/Hz/km 2 )
55
50
45
40
35
30
25
Asym.Opt.
MPC
IID
UC
20
15
10
10
45
40
35
30
25
20
Asym.Opt.
MPC
IID
UC
15
10
20
30
40
50
60
70
80
90
100
Cache Size C
(a) Cache size at N = 8, B = 20, γ = 0.6.
110
0
5
10
15
20
25
30
35
40
45
50
Backhaul Capacity B
(b) Backhaul capability at N = 8, C = 30, γ = 0.6.
Fig. 6. ASE vs. cache size C and backhaul capability B.
with the same probability, and the change of the file popularity by altering γ has no influence
to the ASE.
Fig. 6 plots the ASE vs. the cache size C or the backhaul capability B. From Fig. 6, we can
see that the ASE of all the schemes increases with the cache size and the backhaul capability
because the probability that a randomly requested file is cached at or delivered via the backhaul
increases. The increase of the cache size leads to the increase of the gap between the proposed
asymptotic optimal caching scheme and the MPC scheme. This is because when the cache size
is small, the ASEs of the proposed caching scheme and the MPC scheme mainly come from the
spectrum efficiency of the backhaul files, which is same for the proposed caching scheme and the
MPC scheme. The increase of the cache size can bring larger gains of caching diversity. However,
the increase of the backhaul capability has no influence to the gap between the proposed caching
22
scheme and the previous caching schemes.
VII. C ONCLUSION
In this paper, we consider the analysis and optimization of random caching in backhaul-limited
multi-antenna networks. We propose a file allocation and cache placement design to effectively
improve the network performance. We first derive an exact expression and an upper bound of the
successful transmission probability, using tools from stochastic geometry. Then, we consider the
area spectrum efficiency maximization problem with continuous variables and integer variables.
We obtain a local optimal solution with reduced complexity by exploring optimal properties,
and we also solve an asymptotic optimization problem in high user density region, utilizing the
upper bound as the objective function. Finally, we show that the proposed asymptotic optimal
caching scheme achieves better performance compared with the existing caching schemes, and
the gains are larger when the number of the antennas is larger and/or the Zipf parameter is
smaller.
A PPENDIX
A. Proof of Lemma 1
According to the thinning theory of the PPP, the density of the users that require file f
is qf λu . We define Pi as the probability that file i is requiring by the BS. Note that Pi is
equivalent to the probability that the number of the users requiring file i is not zero, which is
Pi = 1 − 1 +
qi λu −4.5
3.5λb
according to [33]. When file f is requested by a certain BS, to calculate
r
, we need to consider the rest Fb − 1 files out of {Fb \ f } because file f is always
the pmf of Fb,1
requested by the BS. We define the file from the set {Fb \ f } as rest backhaul file and the file
r
from the set {Fb,1
\ f } as rest backhaul request file. To calculate the probability that the number
r
b
of the rest backhaul request files is k, i.e., PF
Fb,1
− 1 = k , we combine all the probabilities
f
Ä
ä
of the cases when any given k − 1 rest backhaul files are required and the Fb − k rest backhaul
r
b
Fb,1
files are not required. Therefore, PF
= k is given by
f
Ä
r
b
PF
Fb,1
−1=k =
f
Ä
ä
X
ä
Y
Y∈{X ⊆{Fb \f }:|X |=k} i∈X
Therefore we finish the proof of Lemma 1.
Pi
Y
i∈B\X
(1 − Pi ) , k = {0, 1, · · · , Fb − 1}.
(44)
23
B. Proof of Theorem 1
The STP of f ∈ Fc is given by
Ä
Psf,c (tf ) = P SIRf > τ
ä
= P g1 > τ Ikx1 kβ
Ä
(a)
= Ex1
ä
ä
(−1)m τ m kx1 kmβ (m) Ä
LI
τ kx1 kβ ,
m!
m=0
"N −1
X
#
where (a) follows from g1 ∼ Gamma(N, 1), I ,
P
i∈{Φb \1}
(45)
kxi k−β gi , LI (s) , EI [exp(−Is)] is
(m)
the Laplace transform of I, and LI (s) is the mth derivative of LI (s).
Denote s = τ kx1 kβ and ym =
(−1)m sm (m)
LI (s),
m!
guish the the interference, we define I f =
we have Psf,c (tf ) = Ex1
P
i∈Φfb \Bf,0
kx1 k−β gi and I −f =
N −1
m=0
îP
ó
ym . To distin-
P
i∈Φ−f
\Bf,0
b
kx1 k−β gi .
We then have LI (τ kx1 kβ )) = LI f (τ kx1 kβ ))LI −f (τ kx1 kβ )).
Therefore, we can derive the expression of LI f τ kx1 kβ ) as follows
Ä
X
i∈{Φfb \1}
β
LI f τ kx1 kβ ) = E
exp −τ kx1 k
Ä
ä
"
ä
kxi k−β gi
1
=
E
1 + τ kx1 kβ kxi k−β
i∈{Φfb \1}
(a)
#
Y
!
Z ∞
!
1
1−
= exp −2πtf λb
rdr
kx1 k
1 + τ kx1 kβ r−β
Ç
ô
ñ
å
2τ
2
2
2
= exp −πtf λb
; 2 − ; −τ kx1 k ,
2 F1 1, 1 −
β−2
β
β
(b)
(46)
where (a) follows from gi ∼ exp(1) due to the random beamforming effect, (b) follows from
the probability generating functional (PGFL) of the PPP [27]. Similarly, we have
LI −f τ kx1 kβ =
Ä
ä
Y
E exp −τ kx1 kβ kxi k−β gi
î
¶
©ó
i∈Φ−f
b
Z ∞
1
= exp −2π(1 − tf )λb
1−
0
1 + τ kx1 kβ r−β
å
Ç
Ç
å
2π
2π
2
2\β
= exp −π(1 − tf )λb csc
τ kx1 k .
β
β
!
!
rdr
(47)
24
Therefore, the expression of LI (τ kx1 kβ )) is given by
Ç
β
ñ
Ç
ô
2
2
2τ
; 2 − ; −τ
tf
2 F1 1, 1 −
β−2
β
β
LI (τ kx1 k )) = exp −πλb
Ç
å
å
To further calculate the STP, we need to calculate the nth derivative of Laplace transform
(n)
LI (s). After some calculations, we can derive the following recursive relationship
(n)
LI (s) =
n−1
X
i=0
n−1
X
+
i=0
!
n−1
(−1)n−i (n − i)!
πtf λb
i
Z ∞
kx1 k2
v
!
n−1
(−1)n−i (n − i)!
π(1 − tf )λb
i
n−i
− β2
1 + sv
Z ∞
0
− β2
dv (i)
n−i+1 LI (s)
β
v− 2
n−i
1 + sv
− β2
dv
(i)
n−i+1 LI (s).
(48)
According to the definition of yn and s, we have
yn = a
n−1
X
i=0
Ö
where li =
(1 − tf )
R∞
0
β
w− 2
i
β
1+w− 2
n−i
l y,
n n−i i
dw
+ tf
i+1
R∞
τ −2\β
β
w− 2
i
β
1+w− 2
(49)
è
dw
, i = {1, 2, · · · , N − 1} and a =
i+1
πλb kx1 k2 τ 2\β . Note that li can be expressed as the combination of the Gauss hypergeometric
function and the Beta function, which is presented in Theorem 1. Let l0 =
2
; −τ
β
ò
+ (1 − tf ) 2π
csc
β
Å
2π
β
ã
ã
Å
2τ
tf β−2
2 F1
ï
1, 1− β2 ; 2−
τ 2\β and we then have y0 = LI (τ kx1 kβ )) = exp −πλb l0 kx1 k2 .
Ä
ä
To get the expression of yn , we need to solve a series of linear equality. We then construct a
Toeplitz matrix as [21] and after some manipulations, we obtain Psf,c (tf ) as follows
Psf,c (tf )
= Ex1 y0
N
−1
X
i=0
1 i i
aD
,
i!
(50)
1
where kk1 is the l1 induced matrix norm and the expression of D is
l
1
l2
.
..
.
0
D=
0
l1
0
..
.
..
lN lM −2 · · ·
l1
.
å
2π
2π
+ (1 − tf ) csc
τ 2\β kx1 k2 .
β
β
(51)
0
According to the thinning of the PPP, the PDF of the distance of the closest f -cached BS
25
2
to the typical user is 2πtf λb kx1 ke−πtf λb kx1 k . After taking expectation over x1 and utilizing the
Taylor expansion, the STP is given by
Psf,c
"
tf
(tf ) =
tf + l0
τ 2\β
I−
tf + l0
#−1
!
D
.
(52)
1
We can obtain Psb , f ∈ Fb similarly and we omit the details due to space limitation.
C. Proof of Lemma 3
ï
Firstly, we prove the property 1. Let A , I −
tf
tf +l0c,f
Å
τ 2\β
tf +l0c,f
ã
Dc,f
ò
and we have Psf,c (tf ) =
kA−1 k1 . We first derive the lower bound of kA−1 k1 . For any x and y satisfying y = A−1 x,
we have kA−1 k1 ≥
kyk1
kxk1
due to the inequality kyk1 ≤ kA−1 k1 kxk1 . Let y = [1, 1, · · · , 1]T and
then we have
A−1
1
≥
kyk1
=
kxk1
N−
N
τ 2\β
tf +l0c,f
PN −1
i=1
(N − i)lic,f
.
(53)
We then derive the upper bound of kA−1 k1 . Noticing that A−1 = (I − A) A−1 + I and using
the triangle inequality, we have kA−1 k1 ≤ k(I − A)k1 kA−1 k1 + kIk1 . Therefore, we obtain the
upper bound of kA−1 k1 as follows
A−1
1
≤
kIk1
=
1 − k(I − A)k1
1−
1
τ 2\β
tf +l0c,f
PN −1 c,f .
li
i=1
tf
kA−1 k1 , we get the bounds of Psf,c (tf ) as
tf +l0c,f
c,f
lic,f and l0c,f in Theorem 1. It can be shown that li+1,
(54)
Note that Psf,c (tf ) =
(35) after substituting
the expression of
≤ lic,f for i ∈ N and
l0c,f =
c,f
i=1 li , ∀tf
P∞
∈ [0, 1]. Moreover, lic,f decreases with tf for i ∈ N. Therefore, after carefully
checking the properties of νA , µA , νB and µB , we obtain property 1.
Secondly, we prove property 2. We define B , tf + l0c,f A and then we have Psf,c (tf ) =
Ä
ä
ktf B−1 k1 . Furthermore, the derivative of B w.r.t. tf is a lower triangular Toeplitz matrix and
,
1 − k0
∂B
=
∂tf
k1
1 − k0
k2
k1
..
.
..
.
kN −1
kN −2
1 − k0
...
···
k1 1 − k0
(55)
26
where k0 and ki , i ∈ {1, 2, · · · , N − 1} are given by
Ç
å
ñ
ô
2π
2τ
2
2
2π
csc
τ 2\β −
; 2 − ; −τ ,
(56)
k0 =
2 F1 1, 1 −
β
β
β−2
β
β
ñ
ô
2
2τ 2\β
2
2
2τ i
2
ki =
B( + 1, i − ) −
; i + 1 − ; −τ , 1 ≤ i ≤ N − 1. (57)
2 F1 i + 1, i −
β
β
β
iβ − 2
β
β
Then we derive the derivative of Psf,c (tf ) w.r.t. tf as follows
∂Psf,c (tf )
∂(tf B−1 )
=
∂tf
∂tf
1
∂B −1
B
∂tf
= B−1 − tf B−1
1
≥
tf
≥
1
1
tf B−1
tf
1
tf B−1
=
tf
Note that 1 − k0 +
PN −1
i=1
ki ≤ 1 − k0 +
1
∂Psf,c (tf )
≥
tf B−1
∂tf
tf
−1 ∂B
−1
tf B
1
1
1
1
− tf B
1−
∂tf
!
−1
B tf
1
!
∂B
∂tf
tf B−1
1
1
1 − 1 − k0 +
N
−1
X
!
!
−1
ki
tf B
P∞
i=1
.
1
i=1
(58)
ki = 1, therefore we have
1 − tf B−1
1
=
Ä
ä
1 f,c
Ps (tf ) 1 − Psf,c (tf ) .
tf
∂Psf,c (tf )
∂tf
Note that Psf,c (tf ) ∈ [0, 1], we then finish the proof of property 2 because
(59)
≥0
Finally, we prove property 3. Note that Psf,c (tf ) = Psb if and only if tf = 1. According to
property 2, we have Psf,c (tf ) ≤ Psb . According to property 1, the upper bound and the lower
bound of Psf,c (tf ) are both 0 if tf = 0, therefore we have Psf,c (tf ) = 0 if tf = 0.
D. Proof of Theorem 3
To prove property 1 is equivalent to prove that B ≤ |Fb∗ | ≤ F − C for optimizing Ps (Fc , t).
We first prove that |Fb∗ | ≤ F − C. Suppose that there exists optimal (Fb∗ , t∗ ) to problem 1
0
Ä
0
satisfying |Fb | > F − C. Now we construct a feasible solution Fb , t
ä
to problem 1, where
0
0
Fb is the set of the most popular F − C files of Fb∗ , the elements of t are same as t∗ if
0
0
Ä
0
f ∈ {F \ Fb∗ } and are one if f ∈ {Fb∗ \ Fb }. Due to the fact that Fc ≤ C, Fb , t
0
ä
is a
feasible solution satisfying the constraints. Note that Psf,c (tf ) = Psb if tf = 1 , we then have
0
0
Ps (t , F \ Fb ) − Ps (t∗ , F \ Fb∗ ) =
P
0
f ∈Fb∗ \Fb
qf Psf,c (1) −
PFb
k=1
Ä
r
P Fb,1
=k
ä
B
Pb
max(k,B) s
which contradicts with the optimality of (Fb∗ , t∗ ). Therefore, we prove |Fb∗ | ≤ F − C.
> 0,
27
We then prove that |Fb∗ | ≥ B. Suppose that there exist optimal (Fb∗ , t∗ ) to problem 1 satisfying
0
Ä
0
0
ä
|Fb | < B. Now we construct a feasible solution Fb , t to problem 1, where Fb is the combining
0
0
of Fb∗ and any B − |Fb∗ | files in {F \ Fb∗ }, the elements of t are same as t∗ if f ∈ {F \ Fb }.
When |Fb | ≤ B, Ps (Fc , t) =
0
P
f ∈Fc
qf Psf,c (tf ) +
0
then have Ps (t , F \ Fb ) − ps (t∗ , F \ Fb∗ ) =
P
f ∈F \Fc
P
f ∈Fb, \Fb∗
qf Psb . Note that Psb ≥ Psf,c (tf ) , we
Ä
ä
qf Psb − Psf,c (tf ) ≥ 0, which contradicts
with the optimality of (Fb∗ , t∗ ). Therefore, we prove |Fb∗ | ≥ B.
E. Proof of Lemma 4
To prove Lemma 4 is equivalent to prove Psu,f,c (tf ) is increasing w.r.t. tf for f ∈ Fc .
qf tf
ζ1 (ατ )tf +ζ2 (ατ )
When N = 1, we have Psu,f,c (tf ) =
and it is increasing w.r.t. tf . We then
consider the scenario when N ≥ 2. According to the proof of the upper bound, Psu,f,c (tf ) =
1 − EIkx1 kβ
ïÄ
Ä
1 − exp −ατ Ikx1 k
Psu,f,c (tf ) = 1 − Ekx1 kβ
β
ääN ò
. After taking expectation over I, we have
ïÄ
1 − exp πλb ((1 − ζ1 (ατ )) tf − ζ2 (ατ )) kx1 k2
Ä
ääN ò
,
(60)
where ζ1 (ατ ) and ζ2 (ατ ) are given in (23) and (24).
Let U (tf |x1 ) , 1 − exp πλb ((1 − ζ1 (ατ )) tf − ζ2 (ατ )) kx1 k2
Ä
Ä
ää
and we have
Ä
ä
Ä
ä
∂U N (tf |x1 )
= N πλb kx1 k2 (ζ1 (ατ ) − 1) U N −1 (tf |x1 ) exp πλb ((1 − ζ1 (ατ )) tf − ζ2 (ατ )) kx1 k2 .
∂tf
∂U N (tf |x1 )
∂tf
Note that ζ1 (ατ ) < 1 and ζ2 (ατ ) > 0, we have
< 0 and U N (tf |x1 ) is a decreasing
î
ó
function of tf for any x1 . Therefore, Psu,f,c (tf ) = 1 − Ekx1 kβ U N (tf |x1 ) is an increasing
î
ó
function of tf . This is because Ekx1 kβ U N (tf |x1 ) can be interpreted as a combination of a
series of U N (tf |x1 ) with different x1 and the monotonicity holds.
F. Proof of Theorem 4
To prove Fc∗ = {B + 1, B + 2, · · · , F } is equivalent to prove Fb∗ = {1, 2, · · · , B}.
Firstly we prove that |Fb∗ | ≤ B. Suppose that there exist optimal (Fb∗ , t∗ ) to problem 1
satisfying |Fb∗ | > B and then we have max (Fb∗ , B) = Fb . Now we construct a feasible solution
Ä
0
0
Fb , t
ä
0
to problem 3, where Fb is the set of the most popular B files of Fb∗ , the elements of
0
0
t are same as t∗ if f ∈ {F \ Fb∗ } and are zero if f ∈ {Fb∗ \ Fb }. Note that Psu,f,c (tf ) = 0
0
0
u
u
if and only if tf = 0 , we then have Rs,∞
(t , F \ Fb ) − Rs,∞
(t∗ , F \ Fb∗ ) = λb log2 (1 +
τ)
Å
P
0
f ∈Fb
qf Psu,b
B
−
ã
u,b
P
f ∈Fb∗
qf Ps
Fb
B > 0 because Psu,b > 0, which contradicts with the optimality
of (Fb∗ , t∗ ). Therefore, we can prove |Fb∗ | ≤ B.
28
Secondly we prove that |Fb∗ | ≥ B. Note that Psu,f,c (tf ) increases w.r.t. tf and Psu,f,c (tf ) ≤
Psu,b , therefore we can easily prove |Fb∗ | ≥ B similarly as Appendix D.
Combining the results |Fb∗ | ≤ B and |Fb∗ | ≥ B, we have |Fb∗ | = B. Finally we prove
Fb∗ = {1, 2, · · · , B}. Suppose that there exists optimal (Fb∗ , t∗ ) to problem 1 satisfying |Fb∗ | = B
0
Ä
0
and Fb 6= {1, 2, · · · B}. We construct a feasible solution Fb , t
0
Ä
{1, 2, · · · B}, the elements of t are same as t∗ if f ∈ {F \ Fb∗
0
¶
in order for the rest files, i.e., tfn = t∗fm , fn ∈ Fb∗
¶
n and m denote the order of the file in Fb∗
0
u
Fb ) − Rs,∞
(t∗ , F \ Fb∗ ) = λb log2 (1 + τ )
τ)
P
fn ∈{Fb∗
T
Fc0 } qfn
we then obtain
q fn
Psu,b − Psc,b t∗fm
Ä
Ä
0
0
0
©
P
©
0
fm ∈{Fb
¶
T
to problem 3, where Fb =
S
0
¶
Fc , fm ∈ Fb
0
Fc and Fb
0
ä
Fb } and are same as t∗
T
©
Fc∗ , n = m. Note that
0
©
u
Fc∗ . We then have Rs,∞
(t , F \
T
Fc∗ } qfm
Ä
Ä
Psu,b − Psc,b t∗fm
Ä
ää
ää
Ä
ä
− λb log2 (1 +
0
ä
Psu,b − Psc,b tfn . Note that Psu,b −Psc,b t∗fm = Psu,b −Psc,b tfn if m = n,
0
0
u
u
(t∗ , F \Fb∗ )
(t , F \Fb )−Rs,∞
Rs,∞
Å
ãÅ
Ä
T
T
0
ä
ãã
= λb log2 (1+τ )
Å
P
0 T
m∈{1,2,··· ,|{Fb
Fc∗ }|}
q fm −
> 0. The last inequality is because qfm > qfn if n = m and Psu,b >
ä
Psc,b t∗fm .
R EFERENCES
[1] N. Bhushan, J. Li, D. Malladi, R. Gilmore, D. Brenner, A. Damnjanovic, R. T. Sukhavasi, C. Patel, and S. Geirhofer,
“Network densification: the dominant theme for wireless evolution into 5G,” IEEE Communications Magazine, vol. 52,
no. 2, pp. 82–89, February 2014.
[2] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K. Soong, and J. C. Zhang, “What will 5G be?” IEEE
Journal on Selected Areas in Communications, vol. 32, no. 6, pp. 1065–1082, June 2014.
[3] X. Wang, M. Chen, T. Taleb, A. Ksentini, and V. C. M. Leung, “Cache in the air: exploiting content caching and delivery
techniques for 5G systems,” IEEE Communications Magazine, vol. 52, no. 2, pp. 131–139, February 2014.
[4] A. Liu and V. K. N. Lau, “Cache-enabled opportunistic cooperative MIMO for video streaming in wireless systems,” IEEE
Transactions on Signal Processing, vol. 62, no. 2, pp. 390–402, Jan 2014.
[5] ——, “Exploiting base station caching in MIMO cellular networks: Opportunistic cooperation for video streaming,” IEEE
Transactions on Signal Processing, vol. 63, no. 1, pp. 57–69, Jan 2015.
[6] M. Tao, E. Chen, H. Zhou, and W. Yu, “Content-centric sparse multicast beamforming for cache-enabled cloud RAN,”
IEEE Transactions on Wireless Communications, vol. 15, no. 9, pp. 6118–6131, Sept 2016.
[7] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to coverage and rate in cellular networks,” IEEE
Transactions on Communications, vol. 59, no. 11, pp. 3122–3134, November 2011.
[8] E. Baştuğ, M. Bennis, M. Kountouris, and M. Debbah, “Cache-enabled small cell networks: Modeling and tradeoffs,”
EURASIP Journal on Wireless Communications and Networking, vol. 2015, no. 1, pp. 1–11, 2015.
[9] C. Yang, Y. Yao, Z. Chen, and B. Xia, “Analysis on cache-enabled wireless heterogeneous networks,” IEEE Transactions
on Wireless Communications, vol. 15, no. 1, pp. 131–145, Jan 2016.
[10] S. T. ul Hassan, M. Bennis, P. H. J. Nardelli, and M. Latva-Aho, “Modeling and analysis of content caching in wireless
small cell networks,” in 2015 International Symposium on Wireless Communication Systems (ISWCS), Aug 2015, pp.
765–769.
29
[11] B. B. Nagaraja and K. G. Nagananda, “Caching with unknown popularity profiles in small cell networks,” in 2015 IEEE
Global Communications Conference (GLOBECOM), Dec 2015, pp. 1–6.
[12] Y. Cui, D. Jiang, and Y. Wu, “Analysis and optimization of caching and multicasting in large-scale cache-enabled wireless
networks,” IEEE Transactions on Wireless Communications, vol. 15, no. 7, pp. 5101–5112, July 2016.
[13] Y. Cui and D. Jiang, “Analysis and optimization of caching and multicasting in large-scale cache-enabled heterogeneous
wireless networks,” IEEE Transactions on Wireless Communications, vol. 16, no. 1, pp. 250–264, Jan 2017.
[14] J. Wen, K. Huang, S. Yang, and V. O. K. Li, “Cache-enabled heterogeneous cellular networks: Optimal tier-level content
placement,” IEEE Transactions on Wireless Communications, vol. 16, no. 9, pp. 5939–5952, Sept 2017.
[15] Z. Chen, J. Lee, T. Q. S. Quek, and M. Kountouris, “Cooperative caching and transmission design in cluster-centric small
cell networks,” IEEE Transactions on Wireless Communications, vol. PP, no. 99, pp. 1–1, 2017.
[16] S. Kuang and N. Liu, “Cache-enabled base station cooperation for heterogeneous cellular network with dependence,” in
2017 IEEE Wireless Communications and Networking Conference (WCNC), March 2017, pp. 1–6.
[17] D. Liu and C. Yang, “Caching policy toward maximal success probability and area spectral efficiency of cache-enabled
HetNets,” IEEE Transactions on Communications, vol. PP, no. 99, pp. 1–1, 2017.
[18] X. Yu, C. Li, J. Zhang, and K. B. Letaief, “A tractable framework for performance analysis of dense multi-antenna
networks,” arXiv preprint arXiv:1702.04573, 2017.
[19] L. H. Afify, H. ElSawy, T. Y. Al-Naffouri, and M. S. Alouini, “A unified stochastic geometry model for MIMO cellular
networks with retransmissions,” IEEE Transactions on Wireless Communications, vol. 15, no. 12, pp. 8595–8609, Dec
2016.
[20] H. S. Dhillon, M. Kountouris, and J. G. Andrews, “Downlink MIMO HetNets: Modeling, ordering results and performance
analysis,” IEEE Transactions on Wireless Communications, vol. 12, no. 10, pp. 5208–5222, October 2013.
[21] C. Li, J. Zhang, and K. B. Letaief, “Throughput and energy efficiency analysis of small cell networks with multi-antenna
base stations,” IEEE Transactions on Wireless Communications, vol. 13, no. 5, pp. 2505–2517, May 2014.
[22] C. Li, J. Zhang, J. G. Andrews, and K. B. Letaief, “Success probability and area spectral efficiency in multiuser MIMO
HetNets,” IEEE Transactions on Communications, vol. 64, no. 4, pp. 1544–1556, April 2016.
[23] C. Li, J. Zhang, M. Haenggi, and K. B. Letaief, “User-centric intercell interference nulling for downlink small cell
networks,” IEEE Transactions on Communications, vol. 63, no. 4, pp. 1419–1431, April 2015.
[24] X. Yu, J. Zhang, M. Haenggi, and K. B. Letaief, “Coverage analysis for millimeter wave networks: The impact of directional
antenna arrays,” arXiv preprint arXiv:1702.04493, 2017.
[25] T. Bai and R. W. Heath, “Coverage and rate analysis for millimeter-wave cellular networks,” IEEE Transactions on Wireless
Communications, vol. 14, no. 2, pp. 1100–1114, Feb 2015.
[26] H. Alzer, “On some inequalities for the incomplete gamma function,” Mathematics of Computation of the American
Mathematical Society, vol. 66, no. 218, pp. 771–778, 1997.
[27] M. Haenggi, Stochastic geometry for wireless networks.
Cambridge University Press, 2012.
[28] L. Breslau, P. Cao, L. Fan, G. Phillips, and S. Shenker, “Web caching and Zipf-like distributions: evidence and implications,”
in INFOCOM ’99. Eighteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings.
IEEE, vol. 1, Mar 1999, pp. 126–134 vol.1.
[29] A. M. Hunter, J. G. Andrews, and S. Weber, “Transmission capacity of ad hoc networks with spatial diversity,” IEEE
Transactions on Wireless Communications, vol. 7, no. 12, pp. 5058–5071, December 2008.
[30] F. Baccelli, B. Blaszczyszyn, and P. Muhlethaler, “An aloha protocol for multihop mobile wireless networks,” IEEE
Transactions on Information Theory, vol. 52, no. 2, pp. 421–436, Feb 2006.
30
[31] W. C. Cheung, T. Q. S. Quek, and M. Kountouris, “Throughput optimization, spectrum allocation, and access control in
two-tier femtocell networks,” IEEE Journal on Selected Areas in Communications, vol. 30, no. 3, pp. 561–574, April 2012.
[32] H. Holma and A. Toskala, LTE for UMTS-OFDMA and SC-FDMA based radio access.
John Wiley & Sons, 2009.
[33] S. Singh, H. S. Dhillon, and J. G. Andrews, “Offloading in heterogeneous networks: Modeling, analysis, and design
insights,” IEEE Transactions on Wireless Communications, vol. 12, no. 5, pp. 2484–2497, May 2013.
| 7cs.IT
|
Nonsparse learning with latent variables †
Zemin Zheng,
University of Science and Technology of China, China
Jinchi Lv
arXiv:1710.02704v1 [stat.ME] 7 Oct 2017
University of Southern California, Los Angeles, USA
and Wei Lin
Peking University, China
Summary. As a popular tool for producing meaningful and interpretable models, large-scale sparse learning
works efficiently when the underlying structures are indeed or close to sparse. However, naively applying the
existing regularization methods can result in misleading outcomes due to model misspecification. In particular, the
direct sparsity assumption on coefficient vectors has been questioned in real applications. Therefore, we consider
nonsparse learning with the conditional sparsity structure that the coefficient vector becomes sparse after taking
out the impacts of certain unobservable latent variables. A new methodology of nonsparse learning with latent
variables (NSL) is proposed to simultaneously recover the significant observable predictors and latent factors as
well as their effects. We explore a common latent family incorporating population principal components and derive
the convergence rates of both sample principal components and their score vectors that hold for a wide class of
distributions. With the properly estimated latent variables, properties including model selection consistency and
oracle inequalities under various prediction and estimation losses are established for the proposed methodology.
Our new methodology and results are evidenced by simulation and real data examples.
Keywords: High dimensionality; Nonsparse coefficient vectors; Latent variables; Conditional sparsity; Principal
component analysis; Spiked covariance; Model selection
1. Introduction
Advances of information technologies have made high-dimensional data increasingly frequent particularly
in the domains of genomic and metagenomic data in biology, imaging data in machine learning, and high
frequency transaction data in economics. The key assumption that enables high-dimensional statistical
inference is that the regression function lies in a low-dimensional manifold [23, 19, 8], meaning that the model
parameter vector is sparse with many zero components. Based on this sparsity assumption, a long list of
regularization methods have been developed to generate meaningful and interpretable models, including [44,
17, 53, 10, 42, 13], among many others. Theoretical guarantees such as oracle properties, oracle inequalities,
model selection consistency, asymptotic distributions, and false discovery rate control were also established
for various regularization methods. See, for example, [51, 6, 43, 15, 16, 25, 46, 50, 5, 9, 31].
Although large-scale sparse learning works efficiently when the underlying structures are indeed or close
to sparse, naively applying the existing regularization methods can result in misleading outcomes due to
model misspecification [48, 36]. In particular, it was imposed in most high-dimensional inference methods
that the coefficient vectors are sparse, which has been questioned in real applications. For instance, [7]
suggested the omnigenic model that the genes associated with complex traits tend to be spread across
most of the genome. Similarly, it was conjectured earlier in [40] that instead of being sparse, the causal
variants responsible for a trait can be distributed. Under such cases, making correct statistical inference is
an important yet challenging task. Though it is generally impossible to accurately estimate large numbers
of nonzero parameters with relatively low sample size, nonsparse learning may be achieved by considering a
natural extension of the sparse scenario, that is, the conditional sparsity structure. Specifically, we assume
the coefficient vector to be sparse after taking out the impacts of certain unobservable latent variables. A
†This work was supported by a grant from the Simons Foundation, National Natural Science Foundation of
China Grants 11601501, 11671018, 71532001, and 71731010, and Anhui Provincial Natural Science Foundation Grant
1708085QA02.
2
Z. Zheng, J. Lv and W. Lin
similar idea was exploited in [18] by the low rank plus sparse representation for large covariance estimation,
where a sparse error covariance structure is imposed after extracting common but unobservable factors.
To characterize the impacts of latent variables, various methods have been proposed under different
model settings. For instance, the latent and observed variables were assumed to be jointly Gaussian in [11]
for graphical model selection. To control for confounding in genetical genomics studies, [33] used genetic
variants as instrumental variables. [38] characterized latent variables by confirmatory factor analysis (CFA)
in survival analysis and estimated them using the EM algorithm. Despite the growing literature, relatively
few studies deal with latent variables in high dimensions. In this paper, we focus on high-dimensional linear
regression and allow for both the numbers of observable predictors and potential latent variables to be
large, where the latent variables are nonsparse linear combinations of a group of observable covariates (not
necessarily the predictors). To the best of our knowledge, this is a new contribution to the case of highdimensional latent variables. Under this setup, our goals are to (i) effectively estimate latent factors given
the observable covariates, (ii) identify the significant features and estimate their effects from a large pool of
predictors in the presence of the underlying factors, (iii) recover the whole structure by further finding out
how many important confounding factors exist as well as their impacts.
We would like to provide some partial answers to the aforementioned questions by considering an important class of potential latent variables incorporating the population principal components of the observable
covariates. The main reasons are as follows. Theoretically, when latent factors are uncorrelated with each
other, factor loadings can generally be obtained through the singular value decomposition (SVD) in factor
analysis. Then population principal components would give the unobservable factors. Even if the observed
covariates are subject to measurement errors, principal components yield the maximum likelihood estimates
of unobservable factors when the measurement errors follow Gaussian distribution with the same variance
[37]. Practically, principal components evaluate orthogonal directions that reflect maximal variations in
the data. In genome-wide association studies, the first few principal components are typically extracted to
adjust for population substructures. Similarly, [30] employed principal components as surrogate variables
to estimate the unobservable factors in genome-wide expression studies. Practical examples of such latent
variables include the underlying cell types of patients, which can be reflected by the expressions of genes in
pathways [4].
The major contributions of this paper are threefold. First, we propose nonsparse learning with latent
variables based on the aforementioned conditional sparsity structure to simultaneously recover the significant
observable predictors and latent factors as well as their effects. By exploring population principal components
as common latent variables, it will be helpful in attenuating collinearity and facilitating dimension reduction.
Second, to estimate population principal components, we use the sample counterparts and provide the
convergence rates of both sample principal components and their score vectors that hold for a wide class
of distributions. The convergence property of sample score vectors is critical to the estimation accuracy of
latent variables. This is, however, much less studied in the literature compared with the principal components
and our work is among the first attempts in the high-dimensional case. Third, we characterize the model
identifiability condition and show that the proposed methodology is applicable to general families with
properly estimated latent variables. In particular, under some regularity conditions, NSL via the thresholded
regression is proved to enjoy model selection consistency and oracle inequalities under various prediction and
estimation losses.
The rest of this paper is organized as follows. Section 2 presents the new methodology of nonsparse
learning with latent variables. We establish asymptotic properties of sample principal components and their
score vectors in high dimensions, as well as theoretical properties of the proposed methodology via the
thresholded regression in Section 3. Simulated and real data examples are provided in Section 4. Section
5 discusses extensions and possible future work. All the proofs of the main results and additional technical
details are included in the Supplementary Material.
2. Nonsparse learning with latent variables
2.1. Model setting
Denote by y = (y1 , . . . , yn )T the n-dimensional response vector, X = (x1 , . . . , xp ) the n × p random design
matrix with p predictors, and F = (f1 , . . . , fK ) the n × K random matrix consisting of K potential latent
variables. The latent variables are unobservable but depend on q observable covariates, given by the n × q
NSL
3
random matrix W = (w1 , . . . , wq ). Assume that the rows of X and W have mean zero and covariance
matrices ΣX and ΣW , respectively. We consider the following high-dimensional linear regression model with
latent variables,
y = Xβ + Fγ + ε,
(1)
where β = (β1 , . . . , βp )T and γ = (γ1 , . . . , γK )T are respectively the regression coefficient vectors of observable predictors and latent variables, and ε ∼ N (0, σ 2 In ) is an n-dimensional error vector independent
of X and F. The conditional sparsity structure is imposed in model (1) such that after taking out the
impacts of unobservable latent factors Fγ, the true regression coefficient vector of observable predictors
β 0 = (β0,1 , . . . , β0,p )T is sparse. Both the dimensionality p and number of covariates q are allowed to grow
nonpolynomially fast with the sample size n.
The main difference between model (1) and linear regression model in high dimensions lies in the extra
latent part Fγ, which accounts for some nonsparse effects of covariates W. Moreover, when confounding
factors potentially exist beyond the original model Xβ, omitting them will cause inconsistency in both
variable selection and parameter estimation. See [2, Chapter 2] for detailed discussions of the impacts of
confounding variables. Although analyzing the effects of latent variables is challenging as they cannot be
measured directly, we will make use of the observable covariates W to adjust for confounding as well as
the nonsparse effects. Generally speaking, the predictors X in model (1) stand for features with individual
effects while W are covariates reflecting the confounding substructures. In practice, which variables should be
chosen as X and which should be chosen as W depend on research interests and the underlying mechanism.
Our analysis also allows for a special case that the features X and W are identical, meaning that the
latent factors are nonsparse combinations of the original predictors. The identifiability of this model will be
discussed after Condition 4 in Section 3.3.
Now we illustrate model (1) with a practical application. In the analysis of body mass index (BMI),
both nutrient intake and gut microbiome composition are believed to be important and they share strong
associations [12]. These two groups of features cannot be packed together as predictors in a single model due
to their fairly different structures and strong correlations, but we may use nutrient intake as predictors while
adjusting for latent factors involving gut microbiome composition. The results of this real data analysis will
be presented in Section 4.2.
As discussed in Section 1, we focus on one potential family of latent variables incorporating population
principal components of the observable covariates W. Specifically, denote by {ui }K
i=1 the top-K principal
components (eigenvectors) of the covariance matrix ΣW , where K is the number of spiked eigenvalues (to be
discussed in Section 3.1) of ΣW , and is allowed to diverge with the sample size. Then each potential latent
variable will be the population principal component score vector, that is, fi = Wui for 1 ≤ i ≤ K.
2.2. Estimation procedure by NSL
With unobservable latent factors F, it is challenging to consistently estimate and recover the support of the
true regression coefficient vector β0 for observable predictors and the true effects of confounding variables
γ 0 = (γ0,1 , . . . , γ0,K )T . We partially overcome this difficulty by assuming that the confounding factors
appear in an unknown linear form of the observable covariates. Then F can be estimated by the sample
principal component scores of covariate matrix W. As W has mean zero, the sample covariance matrix
S = n−1 WT W is an unbiased estimate of ΣW with top-K principal components {b
ui } K
i=1 . So the estimated
b = (bf1 , . . . , bfK ) with bfi = Wb
latent variables are F
ui for 1 ≤ i ≤ K. To ensure model identifiability, both
fi and bfi are rescaled to have a common L2 -norm n1/2 , matching that of the constant predictor 1 for the
intercept. For future prediction, we can transform the coefficient vector γ back by multiplying the scalars
n1/2 kWb
ui k−1
2 . The notation k · kq denotes the Lq -norm of a given vector for q ∈ [0, ∞].
To produce a joint estimate for the true coefficient vectors β 0 and γ 0 , we suggest nonsparse learning with
latent variables which minimizes
n
o
n
o
2
b
Q (β T , γ T )T = (2n)−1 y − Xβ − Fγ
+ pλ (βT∗ , γ T )T
,
(2)
2
1
the penalized residual sum of squares with penalty function pλ (·). Here β∗ = (β∗,1 , . . . , β∗,p )T is the
Hadamard (componentwise) product of two p-dimensional vectors β and (n−1/2 kxk k2 )1≤k≤p . It corresponds
to the design matrix with each column rescaled to have a common L2 -norm n1/2 . The penalty function pλ (t)
4
Z. Zheng, J. Lv and W. Lin
is defined on t ∈ [0, ∞), indexed by λ ≥ 0, and assumed to be increasing in both λ and t with pλ (0) = 0.
We use a compact notation for
o n
oT
n
pλ (β T∗ , γ T )T = pλ (|β∗,1 |), . . . , pλ (|β∗,p |), pλ (|γ1 |), . . . , pλ (|γK |) .
The proposed methodology in (2) enables the possibility to simultaneously estimate β and γ, identifying
the significant observable predictors and latent factors altogether. However, it is still difficult to obtain
b and the correlations
accurate estimates since the confounding factors F are replaced by the estimate F,
between the observable predictors and latent variables can aggravate the difficulty. To prevent the estimation
errors being further magnified in prediction, we consider γ in an L∞ ball BT = {γ ∈ RK : kγk∞ ≤ T },
where any component of γ is assumed to be no larger than T in magnitude. We allow T to diverge slowly
such that it will not deteriorate the overall prediction accuracy.
2.3. Comparisons with existing methods
The proposed methodology can be regarded as a realization of the aforementioned low rank plus sparse
representation [18] in the linear regression setting, but there are significant differences lying behind them.
First, the latent variables in our setup are not necessarily a part of the original predictors, but can stem
from any sources related to the underlying features. Second, unlike the typical assumption in factor analysis
that the factors and the remaining part are uncorrelated, we allow latent variables to share correlations with
the observable predictors. In the extreme case, the latent variables can be nonsparse linear combinations of
the predictors. Third, latent variables are employed to recover the information beyond the sparse effects of
predictors, and thus we do not modify or assume simplified correlations between the original predictors even
after accounting for the latent substructures.
Another method proposed in [29] also incorporated principal components as extra predictors in penalized
regression, but differs from ours in the following aspects. First of all, based on the framework of factor
analysis, the observed predictors in [29] were mixtures of individual features and common factors, both of
which were unobservable. In view of this, we aim at different scopes of applications. Moreover, [29] suggested
sparse regression on the projected model, where individual features were recovered as residuals of projecting
the observed predictors on the factors. In contrast, we keep the original predictors such that they will not be
contaminated when the estimated latent variables are irrelevant. Last but not least, benefitting from factor
analysis, the individual features in [29] were uncorrelated with each other and also shared no correlation
with the factors. But we do not impose such assumptions as explained before.
The proposed methodology is also closely related to principal component regression (PCR). PCR suggests
regressing the response vector on a subset of principal components instead of all explanatory variables, and
comprehensive properties have been established in the literature for its importance in reducing collinearity
and enabling prediction in high dimensions. For instance, [14] explored the situations where the response can
be regressed on the leading principal components of predictors with little loss of information. Probabilistic
explanation was provided in [3] to support the phenomenon that the response is often highly correlated
with the leading principal components. Our new methodology takes advantage of the strengths of principal
components to extract the most relevant information from additional sources and adjust for confounding
and nonsparse effects, while the model interpretability is also retained by exploring the individual effects of
observable predictors.
3. Theoretical properties
We will first establish the convergence properties of sample principal components and their score vectors
for a wide class of distributions under the spiked covariance structure. With the aid of them, properties
including model selection consistency and oracle inequalities will be proved for the proposed methodology
via the thresholded regression using hard-thresholding.
3.1. Spiked covariance model
High-dimensional principal component analysis (PCA) particularly in the context of spiked covariance model,
introduced by [26], has been studied in [39, 28, 41, 47], among many others. This model assumes that
NSL
5
the first few eigenvalues of the population covariance matrix deviate from one while the rest are equal to
one. Although sample principal components are generally inconsistent without strong conditions when the
number of covariates is comparable to or larger than the sample size [27], with the aid of spiked covariance
structure, consistency of sample principal components was established in the literature under different highdimensional settings. For instance, in the high dimension, low sample size context, [28] proved the consistency
of sample principal components for spiked eigenvalues. When both the dimensionality and sample size are
diverging, phase transition of sample principal components was studied in [39, 41] for multivariate Gaussian
observations. The asymptotic distributions of spiked principal components were established in [47] for subGaussian distributions with a finite number of distinguishable spiked eigenvalues.
In this section, we adopt the generalized version of spiked covariance model studied in [28] for the
covariance structure of covariate matrix W, where the population covariance matrix ΣW is assumed to
contain K spiked eigenvalues that can be divided into m groups. The eigenvalues grow at the same rate
within each group while the orders of magnitude of the m groups are different from each other. To be specific,
there are positive constants α1 > α2 > · · · > αm > 1 such that the eigenvalues in the lth group grow at the
rate of q αl , 1 ≤ l ≤ m, where q is the dimensionality or number of covariates in W. The constants αl are
larger than 1 since otherwise the sample eigenvectors
can be strongly inconsistent [28]. Denote the group
Pm
sizes by positive integers k1 , . . . , km satisfying l=1 kl = K < n. Set km+1 = q − K, which is the number of
non-spiked eigenvalues. Then the set of indices for the lth group of eigenvalues is
l−1
l−1
o
n
X
X
kj , l = 1, . . . , m + 1.
kj , . . . , kl +
Jl = 1 +
j=1
(3)
j=1
Although the above eigen-structure looks almost the same as that in [28], the key difference lies in the
magnitudes of the sample size n and the number of spiked eigenvalues K, both of which are allowed to
diverge in our setup instead of being fixed. It makes the original convergence analysis of sample eigenvalues
and eigenvectors invalid since the number of entries in the dual matrix SD = n−1 WWT is no longer finite.
We will overcome this difficulty by conducting a delicate analysis on the deviation bound of the entries such
that the corresponding matrices converge in Frobenius norm. Our theoretical results are applicable to a
wide class of distributions including sub-Gaussian distributions. For multivariate Gaussian or sub-Gaussian
observations with a finite number of spiked eigenvalues, the phase transition of PCA consistency was studied
in, for instance, [41, 47]. Nevertheless, the convergence property of sample principal component score vectors
was not provided in the aforementioned references and needs further investigation.
Assume that the eigen-decomposition of the population covariance matrix ΣW is given by ΣW = UΛUT ,
where Λ is a diagonal matrix of eigenvalues λ1 ≥ λ2 ≥ · · · ≥ λq ≥ 0 and U = (u1 , . . . , uq ) is an orthogonal matrix consisting of the population principal components. Analogously, the eigen-decomposition of
b1 ≥ λ
b2 ≥ · · · ≥ λ
bq ≥ 0 and the orbΛ
bU
b T provides the diagonal matrix Λ
b of sample eigenvalues λ
S = U
b = (b
b q ) consisting of sample principal components. We always assume that the
thogonal matrix U
u1 , . . . , u
sample principal components take the correct directions such that the angles between sample and population
principal components are no more than a right angle. Our main focus is the high-dimensional setting where
the number of covariates q is no less than the sample size n. Denote by Z = Λ−1/2 UT WT the sphered data
matrix. It is clear that the columns of Z are independent and identically distributed (i.i.d.) with mean zero
and covariance matrix Iq . To build our theory, we will impose a tail probability bound on the entry of Z
and make use of the n-dimensional dual matrix SD = n−1 ZT ΛZ, which shares the same nonzero eigenvalues
with S.
3.2. Thresholded regression using hard-thresholding
As discussed in Section 1, there are a large spectrum of regularization methods for sparse learning in high dimensions. It has been demonstrated in [21] that the popular L1 -regularization of Lasso and concave methods
can be asymptotically equivalent in thresholded parameter space for polynomially growing dimensionality,
meaning that they share the same convergence rates in the oracle inequalities. For exponentially growing
dimensionality, concave methods can also be asymptotically equivalent and have faster convergence rates
than the Lasso. Therefore, we will show theoretical properties of the proposed methodology via a specific
concave regularization method, the thresholded regression using hard-thresholding [52]. It utilizes either
6
Z. Zheng, J. Lv and W. Lin
the hard-thresholding penalty pH,λ (t) = 12 λ2 − (λ − t)2+ or the L0 -penalty pH0 ,λ (t) = 2−1 λ2 1{t6=0} in the
penalized least squares (2), both of which enjoy the hard-thresholding property [52, Lemma 1] that facilitates
sparse modeling and consistent estimation.
A key concept for characterizing model identifiability in [52] is the robust spark rsparkc (X) of a given
n × p design matrix X with bound c, defined as the smallest possible number τ such that there exists a
submatrix consisting of τ columns from n−1/2 X̃ with a singular value less than the given positive constant
c, where X̃ is obtained by rescaling the columns of X to have a common L2 -norm n1/2 . The bound on the
magnitude of rsparkc (X) was established in [21] for Gaussian design matrices and further studied by [35] for
more general random design matrices. Under mild conditions, M = c̃n/(log p) with some positive constant
c̃ will provide a lower bound on rsparkc (X, F) for the augmented design matrix (see Condition 4 in Section
3.3 for details). Following [21] and [52], we consider the regularized estimator on the union of coordinate
subspaces SM/2 = {(β T , γ T )T ∈ Rp+K : k(βT , γ T )T k0 < M/2} to ensure model identifiability and reduce
bT , γ
b T )T is defined as the global minimizer of the penalized
estimation instability. So the joint estimator (β
least squares (2) constrained on space SM/2 .
3.3. Technical conditions
Here we list a few technical conditions and discuss their relevance. Denote ∆ = min1≤l≤m−1 (αl − αl+1 ).
Then q ∆ reflects the minimum gap between the magnitudes of spiked eigenvalues in two successive groups.
The first two conditions are imposed for Theorem 1, while the rest are needed in Theorem 2 to be presented
in Section 3.4.
Condition 1. There exist positive constants ci and C such that uniformly over i ∈ Jl , 1 ≤ l ≤ m,
λi /q αl = ci + O(q −∆ ) with ci ≤ C,
and λj ≤ C for any j ∈ Jm+1 .
Condition 2. (a) There exists some positive α < min{∆, αm − 1} such that uniformly over 1 ≤ i ≤ n
and 1 ≤ j ≤ q, the (j, i)th entry of Z, denoted by zji , satisfies
2
> K −1 q α ) = o(q −1 n−1 ).
P (zji
(b) For any 1 ≤ l ≤ m, kn−1 Zl ZTl − Ikl k∞ = o(kl−1 ), where Zl is a submatrix of Z consisting of the rows
with indices in Jl .
Condition 3. Uniformly over j, 1 ≤ j ≤ K, the angle ωjj between the jth estimated latent vector bfj
c2 log n
2
and its population counterpart fj satisfies cos(ωjj ) ≥ 1 − 8K
2 T 2 n with probability 1 − θ1 that converges to one
as n → ∞.
Condition 4. The inequality kn−1/2 (X, F)δk2 ≥ ckδk2 holds for any δ satisfying kδk0 < M with probability 1 − θ2 approaching one as n → ∞.
Condition 5. There exists some positive constant L such that
kxj k2
P ∩pj=1 L−1 ≤ √
≤ L = 1 − θ3 ,
n
where θ3 converges to zero as n → ∞.
Condition 6. Denote by s = kβ0 k0 + kγ 0 k0 and b0 = minj∈supp(β ) (|β0,j |) ∧ minj∈supp(γ 0 ) (|γ0,j |) the
0
number of overall significant predictors and overall minimum signal strength, respectively. It holds that
s < M/2 and
p
√
−1
b0 > [( 2c−1
1 ) ∨ 1]c1 c2 L (2s + 1)(log p)/n
√
for some positive constants c1 defined in Proposition 1 in Section 3.4 and c2 > 2 2σ.
NSL
7
Condition 1 requires that the orders of magnitude of spiked eigenvalues in each group be the same while
their limits can be different, depending on the constants ci . It is weaker than those usually imposed in the
literature such as [41], where the spiked eigenvalues in each group share exactly the same limit. Nevertheless,
we will prove the consistency of spiked sample eigenvalues under very mild conditions. To distinguish the
eigenvalues in different groups, convergence to the corresponding limit is assumed to be at a rate of O(q −∆ ).
As the number of spiked eigenvalues diverges with q, we impose a constant upper bound C on ci for simplicity,
and our technical argument still applies when C diverges slowly with q. Without loss of generality, the upper
bound C also controls the non-spiked eigenvalues.
As pointed out earlier, the columns of the sphered data matrix Z are i.i.d. with mean zero and covariance
matrix Ip . Then part (a) of Condition 2 holds as long as the entries in any column of Z satisfy the tail
probability bound. Moreover, it is clear that this tail bound decays polynomially, so that it holds for a
wide class of distributions including sub-Gaussian distributions. With this tail bound, the larger sample
eigenvalues would dominate the sum of all eigenvalues in the smaller groups regardless of the randomness.
Furthermore, by definition we know that the columns of Zl are i.i.d. with mean zero and covariance matrix
Ikl such that n−1 Zl ZTl → Ikl entrywise as n → ∞. Hence, part (b) of Condition 2 is a very mild assumption
to deal with the possibly diverging group sizes kl .
Condition 3 imposes a convergence rate of log n/(K 2 T 2 n) for the estimation accuracy of confounding
b will not deteriorate the overall estimation and prediction powers.
factors, so that the estimation errors in F
This rate is easy to satisfy in view of the results in Theorem 1 in Section 3.4 since the sample principal
component score vectors are shown to converge to the population counterparts in polynomial orders of q,
which is typically larger than n in high-dimensional settings.
Condition 4 assumes the robust spark of matrix (X, F) with bound c to be at least M = c̃n/(log p)
with significant probability. It is the key for characterizing the model identifiability in our conditional
sparsity structure and also controls the correlations between the observable predictors X and latent factors
F. Consider a special case where F consists of nonsparse linear combinations of the original predictors
X. Then model (1) cannot be identified if we allow for nonsparse regression coefficients. However, if we
constrain the model size by certain sparsity level, such as rsparkc (X, F), the model will become identifiable
since F cannot be represented by sparse linear combinations of X. Utilizing the same idea, if we impose
conditions such as the minimum eigenvalue for the covariance matrix of any M1 features in (X, F) being
bounded from below, where M1 = c̃1 n/(log p) with c̃1 > c̃ denotes the sparsity level, then [35, Theorem 2]
ensures that the robust spark of any submatrix consisting of less than M1 columns of (X, F) will be no less
than M = c̃n/(log p). It holds for general distributions with tail probability decaying exponentially fast with
the sample size n, and the constant c̃ depending only on c. This justifies the inequality in Condition 4.
While no distributional assumptions are imposed on the random design matrix X, Condition 5 puts a
mild constraint that the L2 -norm of any column vector of X divided by its common scale n1/2 is bounded
with significant probability. It can be satisfied by many distributions and is needed due to the rescaling of β ∗
in (2). Condition 6 is similar to that of [52] for deriving the global properties via the thresholded regression.
The first part puts a sparsity constraint on the true model size s for model identifiability as discussed after
Condition 4, while the second part gives a lower bound O{[s(log p)/n]1/2 } on the minimum signal strength
to distinguish the significant predictors from the others.
3.4. Main results
We provide two main theorems in this section. The first one is concerned with the asymptotic properties
of sample principal components and their score vectors, which serves as a bridge for establishing the global
properties in the second theorem.
A sample principal component is said to be consistent with its population counterpart if the angle
between them converges to zero asymptotically. However, when several population eigenvalues belong to
the same group, the corresponding principal components may not be distinguishable. In that case, subspace
ui , span{uj : j ∈
consistency is essential to characterizing the asymptotic properties [28]. Denote θil = Angle(b
Jl }) for i ∈ Jl , 1 ≤ l ≤ m, which is the angle between the ith sample principal component and the subspace
spanned by population principal components in the corresponding spiked group. The following theorem
presents the convergence rates of sample principal components in terms of angles under the aforementioned
generalized spiked covariance model. Moreover, for the identifiability of latent factors, we assume each group
8
Z. Zheng, J. Lv and W. Lin
size to be one for the spiked eigenvalues when studying the principal component score vectors. That is, kl = 1
for 1 ≤ l ≤ m, implying K = m.
Theorem 1 (Convergence rates). Under Conditions 1 and 2, with probability approaching one, the
following statements hold.
(a) Uniformly over i ∈ Jl , 1 ≤ l ≤ m, θil = Angle(b
ui , span{uj : j ∈ Jl }) is no more than
arccos([1 −
l−1
l−1
X
Y
t=1
i=t+1
(1 + ki ) O{kt A(t)} − O{A(l)}]1/2 ),
(4)
Pj
Qj
Pm
αl
where A(t) =
+ km+1 K −1 q α−αt and we define t=i st = 0 and t=i st = 1 if j < i for any
l=t+1 kl q
sequence {st }.
(b) If each group of spiked eigenvalues has size one, then uniformly over 1 ≤ i ≤ K, ωii = Angle(Wb
ui , Wui )
is no more than
arccos([1 −
i−1
X
t=1
2i−t−1 O{A(t)} − O{A(i)}]1/2 ).
Part (a) of Theorem 1 provides the uniform convergence rates of sample principal components to the
corresponding subspaces for general spiked covariance structure with possibly tiered eigenvalues under mild
2
conditions. Since the convergence rates of θil
to zero and cos2 (θil ) to one are the same by L’Hospital’s rule,
Pl−1 Ql−1
both of them are t=1
i=t+1 (1 + ki ) O{kt A(t)} + O{A(l)} in view of (4). Thus, when the group sizes
kl are relatively small, the convergence rates are determined by A(t), which decays polynomially with q and
converges to zero fairly fast. It shows the “blessing of dimensionality” under the spiked covariance structure
since the larger q gives faster convergence rates. Furthermore, it is clear that when the gaps between the
magnitudes of different spiked groups are large, A(t) decays quickly with q to accelerate the convergence of
sample principal components.
The uniform convergence rates of sample principal component score vectors are given in part (b) of
Theorem 1 when each group contains only one spiked eigenvalue such that the latent factors are separable.
In fact, the proof of Theorem 1 shows that the sample score vectors converge at least as fast as the sample
principal components. Then the results in part (b) are essentially the convergence rates in part (a) with
kl = 1. Since the number of spiked eigenvalues K is much smaller than q, the sample principal component
score vectors will converge to the population counterparts polynomially with q. The convergence property
of sample score vectors is critical to our purpose of nonsparse learning since it offers the estimation accuracy
of latent variables, which is much less well studied in the literature. To the best of our knowledge, our work
is a first attempt in high dimensions.
The established asymptotic property of sample principal component score vectors justifies the estimation
accuracy assumption in Condition 3. Together with Condition 4, it leads to the following proposition.
Proposition 1. Under Conditions 3 and 4, the inequality
b
kn−1/2 (X, F)δk
2 ≥ c1 kδk2
holds for some positive constant c1 and any δ satisfying kδk0 < M with probability at least 1 − θ1 − θ2 .
From the proof of Proposition 1, we see that the constant c1 is smaller than but can be very close to c
when n is relatively large. Therefore, Proposition 1 shows that the robust spark of the augmented design
b will be close to that of (X, F) when F is accurately estimated by F.
b We are now ready to
matrix (X, F)
present theoretical properties for the proposed methodology.
Theorem 2 (Global properties). Assume that Conditions 3–6 hold and
p
√
c1−1 c2 (2s + 1)(log p)/n < λ < L−1 b0 [1 ∧ (c1 / 2)].
Then for both the hard-thresholding penalty pH,λ (t) and L0 -penalty pH0 ,λ (t), with probability at least 1 −
c2
2
−1/2 1− 8σ2
−1/2
4σ(2/π)1/2 c−1
p
− 2σ(2/π)1/2 c−1
2 (log p)
2 s(log n)
c2
2
T
b ,γ
b T )T satisfies that:
· n− 8σ2 − θ1 − θ2 − θ3 , the regularized estimator (β
NSL
T T T
b ,γ
b )
(a) (Model selection consistency) supp (β
9
= supp (β T0 , γ T0 )T ;
√ p
bT , γ
b β
b T )T − (X, F)(β T0 , γ T0 )T k2 ≤ (c2 /2 + 2c2 c−1
(b) (Prediction loss) n−1/2 k(X, F)(
s) (log n)/n;
1
p
p
1/q
b − β kq ≤ 2c−2 c2 Ls1/q (log n)/n, kb
(log n)/n for
(c) (Estimation losses) kβ
γ − γ 0 kq ≤ 2c−2
0
1
1 c2 s
b
q ∈ [1, 2]. The upper bounds with q = 2 also hold for kβ − β0 k∞ and kb
γ − γ 0 k∞ .
The model selection consistency in Theorem 2 shows that we can recover both the significant observable
predictors and the latent variables, so that the whole model would be identified by combining these two parts
even if it contains nonsparse coefficients. The prediction loss of the joint estimator is shown to be within a
logarithmic factor (log n)1/2 of that of the oracle estimator when the regularization parameter λ is properly
chosen, which is similar to the result in [52]. It means that the prediction accuracy is maintainedpregardless
of the hidden effects as long as the latent factors are properly estimated. The extra term (c2 /2) (log n)/n
in the prediction bound reflects the price we pay in estimating the confounding factors. Furthermore, the
b and γ
b under Lq -estimation losses with q ∈ [1, 2] ∪ {∞} are also established in
oracle inequalities for both β
Theorem 2. Although the estimation accuracy for the nonsparse coefficients Uγ of W are obtainable, we
omit the results here since their roles in inferring the individual effects and prediction are equivalent to those
of the latent variables.
The proposed methodology of nonsparse learning with latent variables under the conditional sparsity
structure is not restrictive to the potential family of population principal components. It is more broadly
applicable to any latent family provided that the estimation accuracy of latent factors in Condition 3 and
the correlations between the observable predictors and latent factors characterized by the robust spark in
Condition 4 hold similarly. The population principal component provides a common and concrete example
to extract the latent variables from additional covariates. A significant advantage of this methodology is that
even if the estimated latent factors are irrelevant, they rarely affect the variable selection and effect estimation
of the original predictors since the number of potential latent variables is generally a small proportion of
that of the predictors. This is a key difference between our methodology and those based on factor analysis,
which renders it useful for combining additional sources.
4. Numerical studies
In this section, we discuss the implementation and investigate the finite sample performance of NSL via three
regularization methods of the Lasso [44], SCAD [17], and the thresholded regression using hard-thresholding
(Hard) [52]. The oracle procedure (Oracle) which knew the true model in advance is also conducted as a
benchmark. We will explore two different models, where model M1 involves only observable predictors and
model M2 incorporates estimated latent variables as extra predictors. The case of linear regression model
(1) with the confounding factor as nonsparse combination of the existing predictors is considered in the first
example, while in the second example multiple latent factors stem from additional observable covariates and
the error vector is relatively heavy-tailed with t-distribution.
4.1. Simulation examples
4.1.1. Simulation example 1
In the first simulation example, we consider a special case of linear regression model (1) with potential
latent factors F coming from the existing observable predictors, that is, W = X. Then Fγ represents the
nonsparse effects of the predictors X, and it will be interesting to check the impacts of latent variables
when they are dense linear combinations of the existing predictors. The sample size n was chosen to be
100 with true regression coefficient vectors β 0 = (vT , . . . , vT , 0)T , γ 0 = (0.5, 0)T , and Gaussian error vector
ε ∼ N (0, σ 2 In ), where v = (0.6, 0, 0, −0.6, 0, 0)T is repeated k times and γ 0 is a K-dimensional vector with
one nonzero component 0.5, denoting the effect of the significant confounding factor. We generated 200 data
sets and adopted the setting of (p, k, K, σ) = (1000, 3, 10, 0.4) such that there are six nonzero components
with magnitude 0.6 in the true coefficient vector β 0 and ten potential latent variables.
The key point in the design of this simulation study is to construct a population covariance matrix Σ with
spiked structure. Therefore, for each data set, the rows of the n × p design matrix X were sampled as i.i.d.
10
Z. Zheng, J. Lv and W. Lin
Table 1. Means and standard errors (in parentheses) of different performance
measures by all methods over 200 simulations in Section 4.1.1; M1 : model with
only observable predictors, M2 : model includes estimated latent variables
Model Measure
Lasso
SCAD
Hard
Oracle
M1
PE
65.27 (1.35) 65.29 (1.40) 68.80 (6.45)
—
L2 -loss
1.61 (0.24)
1.61 (0.25)
2.25 (1.07)
—
L1 -loss
4.69 (1.84)
4.70 (1.88)
5.03 (2.31)
—
L∞ -loss
0.65 (0.13)
0.65 (0.15)
1.48 (1.10)
—
FP
4.45 (7.15)
4.45 (7.16)
0.51 (0.90)
—
FN
5.93 (0.26)
5.93 (0.26)
5.98 (0.16)
—
σ
b
7.88 (0.57)
7.88 (0.57)
7.78 (0.60)
—
M2
PE
0.39 (0.16)
0.19 (0.01)
0.19 (0.01)
0.17 (0.01)
L2 -loss
0.43 (0.13)
0.13 (0.03)
0.10 (0.03)
0.10 (0.03)
L1 -loss
1.52 (0.40)
0.44 (0.07)
0.21 (0.13)
0.21 (0.06)
L∞ -loss
0.23 (0.07)
0.07 (0.02)
0.07 (0.02)
0.07 (0.02)
FP
28.79(6.52)
15.99 (5.63)
0.02 (0.28)
0 (0)
FN
0.02 (0.16)
0 (0)
0 (0)
0 (0)
σ
b
0.47 (0.06)
0.38 (0.03)
0.41 (0.03)
0.40 (0.03)
copies from a multivariate normal distribution N (0, Σ) with Σ = 21 (Σ1 + Σ2 ), where Σ1 = (0.5|i−j| )1≤i,j≤p
and Σ2 = 0.5Ip + 0.511T . The choice of Σ1 allows for correlation between the predictors at the population
level and Σ2 has an eigen-structure such that the spiked eigenvalue is comparable with p. Based on the
construction of Σ1 and Σ2 , it is easy to check that Σ has the largest eigenvalue 251.75 and the others are
all below 1.75. For regularization methods, model M2 involved the top-K sample principal components
as estimated latent variables while the oracle procedure used the true confounding factor instead of the
estimated one. We applied the Lasso, SCAD, and Hard for both M1 and M2 to produce a sequence of sparse
models and selected the regularization parameter λ by minimizing the prediction error calculated based on
an independent validation set for fair comparison of all methods.
To compare the performance of the aforementioned methods under two different models, we consider
b 2 in
several performance measures. The first measure is the prediction error (PE) defined as E(Y − xT β)
T
T
T
b − bf γ
b or (β
b ,γ
b )2 in model M2 , where β
b )T are the estimated coefficients in
model M1 and as E(Y − xT β
T
the corresponding models, (x , Y ) is an independent test sample of size 10, 000, and bf is the sample principal
component score vector. For the oracle procedure, bf is replaced by the true confounding factor f. The second
b − β kq with q = 2, 1, and ∞, respectively.
to fourth measures are the Lq -estimation losses of β0 , that is, kβ
0
The fifth and sixth measures are the false positives (FP), falsely selected noise predictors, and false negatives
(FN), missed true predictors with respect to β 0 . We also calculated the estimated error standard deviation
σ
b by all methods in both models. The results are summarized in Table 1. For the selection and effect
estimation of latent variables in model M2 , we display in Table 2 the measures similar to those defined in
Table 1 but with respect to γ 0 . They are Lq -estimation losses kb
γ − γ 0 kq with q = 2, 1, and ∞, FPγ , and
FNγ .
In view of Table 1, it is clear that compared with model M2 , the performance measures in variable selection, estimation, and prediction all deteriorated seriously in model M1 , where most of important predictors
were missed and both the estimation and prediction errors were quite large. We want to emphasize that in
this first example, the latent variables are linear combinations of the observable predictors initially included
in the model, which means that the nonsparse effects would not be captured without the help of estimated
confounding factors. On the other hand, the prediction and estimation errors of all regularization methods
were reasonably small in the latent variable augmented model M2 . And the performance of Hard was comparable to that of the oracle procedure regardless of the estimation errors of latent features, which is in line
with the theoretical results in Theorem 2. Furthermore, we can see from Table 2 that all methods with the
estimated latent variables correctly identified the true confounding factor and accurately recovered its effect.
NSL
11
Table 2. Means and standard errors (in parentheses) of different performance measures for regression coefficients of confounding factors
by all methods over 200 simulations in Section 4.1.1 (The notation
0.00 denotes a number less than 0.005.)
Measure
Lasso
SCAD
Hard
Oracle
L2 -loss
0.02 (0.00) 0.01 (0.00) 0.01 (0.00) 0.00 (0.00)
L1 -loss
0.02 (0.01) 0.01 (0.00) 0.01 (0.00) 0.00 (0.00)
L∞ -loss 0.02 (0.00) 0.01 (0.00) 0.01 (0.00) 0.00 (0.00)
FPγ
0.29 (0.55) 0.21 (0.43)
0 (0)
0 (0)
FNγ
0 (0)
0 (0)
0 (0)
0 (0)
Table 3. Means and standard errors (in parentheses) of different performance measures by all methods over 200 simulations in Section 4.1.2; M1 : model with only
observable predictors, M
p2 : model includes estimated latent variables, population error standard deviation σ df /(df − 2) equals to 0.45
Model Measure
Lasso
SCAD
Hard
Oracle
M1
PE
72.33 (1.53)
72.33 (1.53)
76.04 (6.62)
—
L2 -loss
1.58 (0.24)
1.58 (0.24)
2.25 (1.09)
—
L1 -loss
4.49 (1.86)
4.49 (1.86)
5.00 (2.10)
—
L∞ -loss
0.64 (0.13)
0.64 (0.13)
1.50 (1.15)
—
FP
3.59 (6.52)
3.59 (6.52)
0.49 (0.72)
—
FN
5.95 (0.23)
5.95 (0.23)
6.00 (0.07)
—
Error SD
8.33 (0.59)
8.33 (0.59)
8.20 (0.63)
—
M2
PE
1.74 (1.08)
1.10 (1.05)
1.04 (0.99)
0.22 (0.01)
L2 -loss
0.70 (0.22)
0.25 (0.22)
0.16 (0.18)
0.11 (0.03)
L1 -loss
2.18 (0.50)
0.87 (0.50)
0.39 (0.53)
0.23 (0.07)
L∞ -loss
0.37 (0.12)
0.13 (0.10)
0.10 (0.09)
0.08 (0.03)
FP
20.63 (13.60) 23.29 (12.52)
0.70 (3.34)
0 (0)
FN
0.09 (0.38)
0.15 (0.94)
0.09 (0.63)
0 (0)
Error SD
0.74 (0.20)
0.48 (0.21)
0.50 (0.12)
0.45 (0.04)
4.1.2. Simulation example 2
Now we consider a more general case where the latent variables stem from a group of observable covariates
instead of the original predictors. Moreover, we also want to see whether similar results hold when more
significant confounding factors are involved and the errors become relatively heavy-tailed. Thus, there are
three main changes in the setting of this second example. First, the predictors X and observable covariates
W are different, as well as their covariance structures which will be specified later. Second, there are two
significant latent variables and the K-dimensional true coefficient vector γ 0 = (0.5, −0.5, 0)T . Third, the
error vector ε = ση, where the components of the n-dimensional random vector η are independent and
follow the t-distribution with df = 10 degrees of freedom. The settings of β0 and (n, p, K, σ) are the same
as in the first simulation example in Section 4.1.1, while the dimensionality q of covariates W equals 1000,
which is also large.
For the covariance structure of X, we set ΣX = (0.5|i−j| )1≤i,j≤p to allow for correlation at the population
level. On the other hand, in order to estimate the principal components in high dimensions, the population
covariance matrix of W should have multiple spiked eigenvalues. Thus, we constructed it using the block
diagonal structure such that
Σ11
0
,
ΣW =
0
Σ22
where Σ11 = 34 (Σ1 + Σ2 )1≤i,j≤200 and Σ22 = 12 (Σ1 + Σ2 )1≤i,j≤800 with the defintions of Σ1 and Σ2 similar
to those in Section 4.1.1 except for different dimensions. Under such construction, the two largest eigenvalues
of ΣW are 201.75 and 77.61, respectively, while the others are less than 2.63. Based on the aforementioned
covariance structures, for each data set, the rows of X and W were sampled as i.i.d. copies from the
corresponding multivariate normal distribution.
We included the top-K sample principal components in model M2 as potential latent factors and compared
the performance of the Lasso, SCAD, Hard, and Oracle by the same performance measures as defined in
12
Z. Zheng, J. Lv and W. Lin
Table 4. Means and standard errors (in parentheses) of different performance measures for regression coefficients of confounding factors
by all methods over 200 simulations in Section 4.1.2 (The notation
0.00 denotes a number less than 0.005.)
Measure
Lasso
SCAD
Hard
Oracle
L2 -loss
0.08 (0.03) 0.07 (0.04) 0.07 (0.04) 0.01 (0.00)
L1 -loss
0.10 (0.04) 0.09 (0.05) 0.08 (0.05) 0.01 (0.00)
L∞ -loss 0.08 (0.03) 0.06 (0.03) 0.06 (0.03) 0.01 (0.00)
FPγ
0.21 (0.45) 0.34 (0.60) 0.01 (0.10)
0 (0)
FNγ
0 (0)
0 (0)
0 (0)
0 (0)
Section 4.1.1. The results are summarized in Tables 3 and 4. From Table 3, it is clear that the methods
which replied only on the observable predictors still suffered a lot under this more difficult setting, where all
true predictors were missed, prediction errors were large, and the error standard deviation (SD) was poorly
estimated. In contrast, the new NSL methodology via the Lasso, SCAD, and Hard was able to tackle the
issues associated with variable selection, coefficient estimation, prediction, and error SD estimation. With
the latent variable augmented model M2 , Hard almost recovered the exact underlying model. Similar to
the first example, in view of Table 4, all methods correctly identified the significant confounding factors
and estimated their effects accurately. However, compared with Tables 1 and 2, most of the performance
measures deteriorated in this second example. This is mainly due to the relatively heavy-tailed random
errors, as well as the difficulty in estimating multiple high-dimensional principal components.
4.2. Application to nutrient intake with gut microbiome data
Nutrient intake strongly affects human health or diseases such as obesity, while gut microbiome composition
is an important factor in energy extraction from the diet. We illustrate the usefulness of our proposed
methodology by applying it to the data set reported in [49] and previously studied by [12] and [34], where
a cross-sectional study of 98 healthy volunteers was carried out to investigate the habitual diet effect on the
human gut microbiome. The nutrient intake consisted of 214 micronutrients collected from the volunteers
by a food frequency questionnaire. The values were normalized by the residual method to adjust for caloric
intake and then standardized to have mean zero and standard deviation one. Similar to [12], we used one
representative for a set of highly correlated micronutrients whose correlation coefficients are larger than 0.9,
resulting in 119 representative micronutrients in total. Furthermore, stool samples were collected and DNA
samples were analyzed by 454/Roche pyrosequencing of 16S rDNA gene segments from the V1–V2 region.
After taxonomic assignment of the denoised pyrosequences, the operational taxonomic units were combined
into 87 genera which appeared in at least one sample. We are interested in identifying the important
micronutrients and potential latent factors from the gut microbiome genera that are associated with the
body mass index (BMI).
Due to the high correlations between the micronutrients, we applied NSL via the elastic net [53] to this
data set by treating BMI, nutrient intake, and gut microbiome composition (after the centered log-ratio
transformation [1]) as the response, predictors, and covariates of confounding factors, respectively. The data
set was split 100 times into a training set of 60 samples and a validation set of the remaining samples. For
each splitting of the data set, we explored two different models M1 and M2 as defined in Section 4.1 with the
top-20 sample principal components (PCs) of gut microbiome composition included in model M2 to estimate
the potential latent factors. All predictors were rescaled to have a common L2 -norm of n1/2 and the tuning
parameter was chosen by minimizing the prediction error calculated on the validation set. We summarize
in Table 5 the selection probabilities and coefficients of the significant micronutrients and latent variables
whose selection probabilities were above 0.9 in M1 or above 0.85 in M2 . The means (with standard errors in
parentheses) of the prediction errors averaged over 100 random splittings were 167.9 (7.2) in model M1 and
110.3 (4.0) in model M2 , while the median model size also reduced from 93 to 69 after applying the NSL
methodology. It shows that the prediction performance was improved after utilizing the information of gut
microbiome genera.
In view of the model selection results in Table 5, many significant micronutrients in model M1 became
insignificant after adjusting for the latent substructures, which implies that either they affect BMI through
NSL
13
Table 5. Selection probabilities and rescaled coefficients (in parentheses) of the most
frequently selected predictors by each model across 100 random splittings in Section
4.2; M1 : model with only micronutrients as predictors, M2 : model includes latent variables from gut microbiome composition
Predictor
Sodium
Eicosenoic acid
Vitamin B12
Gallocatechin
Riboflavin pills
Acrylamide
Naringenin
Pelargonidin
Lauric acid
Model M1
0.98 (1.35)
0.98 (-2.47)
0.96 (0.43)
0.96 (-4.81)
0.94 (1.71)
0.94 (-0.34)
0.94 (1.11)
0.94 (-1.15)
0.93 (1.88)
Model M2
0.67 (0.55)
0.80 (-1.24)
0.62 (0.30)
0.84 (-1.70)
0.55 (0.61)
0.62 (0.32)
0.58 (0.32)
0.75 (-1.03)
0.71 (0.50)
Predictor
PC(7th)
PC(6th)
Apigenin
PC(9th)
PC(10th)
Iron
Aspartame
Vitamin C
Vitamin E
Model M1
———
———
0.95 (-1.67)
———
———
0.93 (1.22)
0.93 (-0.46)
0.93 (-0.71)
0.92 (0.45)
Model M2
0.99 (1.76)
0.96 (-1.21)
0.93 (-1.88)
0.88 (-0.87)
0.86 (0.78)
0.86 (0.75)
0.79 (0.59)
0.76 (-0.39)
0.65 (-0.29)
Table 6. Major gut microbiome genera in the compositions of the two
significant latent variables identified by the model-free knockoffs in
Section 4.2
Latent variable Phylum
Genus
Weight
PC(7th)
Firmicutes
Dialister
-0.40
Firmicutes
Eubacterium
0.39
Bacteroidetes
Barnesiella
-0.28
PC(9th)
Firmicutes
Acidaminococcus -0.51
Firmicutes
Megasphaera
-0.36
Firmicutes
Ruminococcus
-0.30
the gut microbiome genera or their combinative effects are captured by the latent variables. This was also
evidenced by the reduction in the model size mentioned before. Moreover, the effects of some micronutrients
changed signs in model M2 and the subsequent associations with BMI are consistent with scientific discoveries
[22]. For instance, aspartame is a sugar substitute widely used in beverages such as the diet coke, and it was
negatively associated with BMI in model M1 but tended to share a positive association after accounting for
the gut microbiome genera. A potential reason is that the people who drink diet coke can have a relatively
healthy habitual diet and gut microbiome composition which in turn lower the BMI, but the diet coke itself
does not reduce fats. Similar phenomena happened to acrylamide and vitamin E as well.
We also applied the model-free knockoffs [9] with the target FDR level 0.2 on model M2 , and the most
significant predictors identified were the latent variables of 7th and 9th PCs. The major gut microbiome
genera in the compositions of these two latent variables are displayed in Table 6. At the phylum level, the
latent factors mainly consist of bacteroidetes and firmicutes, whose relative proportion has been shown to
affect human obesity [32]. In view of the associations with BMI, both the 7th and 9th PCs confirm the claim
that firmicutes-enriched microbiome holds a greater metabolic potential for energy gain from the diet which
results in the gain of weight [45]. Furthermore, one of the major microbiome genera in the latent factor of
9th PC, Acidaminococcus, was also found to be positively associated with the BMI in [34], which shows that
human obesity can be affected at the genus level.
5. Discussions
In this paper, we have introduced a new methodology NSL for prediction and variable selection in the
presence of nonsparse coefficient vectors through the conditional sparsity structure, where latent variables
are exploited to capture the nonsparse combinations of either the original predictors or additional covariates.
The suggested methodology is ideal for the applications including two sets of predictors that cannot be
packed directly for analysis, as in our BMI study. Both theoretical guarantees and empirical performance
of the potential latent family incorporating population principal components have been demonstrated. And
our methodology is also applicable to more general families with properly estimated latent variables and
identifiable models.
It would be interesting to further investigate several problems such as hypothesis testing and false discovery rate control in nonsparse learning by the idea of NSL. Based on the established model identifiability
14
Z. Zheng, J. Lv and W. Lin
condition which characterizes the correlations between observable and latent predictors, hypothesis testing
can be proceeded using the de-biasing idea in [25, 46, 50], and false discovery rate could be controlled by
applying the knockoffs inference procedures [5, 9, 20] on the latent variable augmented model. The main
difficulty lies in analyzing how the estimation errors of unobservable factors affect the corresponding procedures. Another possible direction is to explore more general ways of modeling the latent variables to deal
with the nonsparse coefficient vectors. These problems are beyond the scope of the current paper and will
be interesting topics for future research.
References
[1] Aitchison, J. (1983). Principal component analysis of compositional data. Biometrika, 70, 57–65.
[2] Anderson, S., Auquier, A., Hauck, W. W., Oakes, D., Vandaele, W., I., W. H., Bryk, A. S., and Kleinman,
J. (1980). Statistical Methods for Comparative Studies. New York: Wiley.
[3] Artemiou, A. and Li, B. (2009). On principal components and regression: a statistical explanation of a
natural phenomenon. Statist. Sinica, 19, 1557–1565.
[4] Bair, E., Hastie, T., Paul, D., and Tibshirani, R. (2006). Prediction by supervised principal components.
J. Amer. Statist. Assoc., 101, 119–137.
[5] Barber, R. and Candès, E. J. (2015). Controlling the false discovery rate via knockoffs. Ann. Statist.,
43, 2055–2085.
[6] Bickel, P. J., Ritov, Y., and Tsybakov, A. B. (2009). Simultaneous analysis of Lasso and Dantzig selector.
Ann. Statist., 37, 1705–1732.
[7] Boyle, E., Li, Y., and Pritchard, J. (2017). An expanded view of complex traits: from polygenic to
omnigenic. Cell, 169, 1177–1186.
[8] Bühlmann, P. and van de Geer, S. (2011). Statistics for High-Dimensional Data: Methods, Theory and
Applications. Springer.
[9] Candès, E. J., Fan, Y., Janson, L., and Lv, J. (2016). Panning for gold: Model-free knockoffs for
high-dimensional controlled variable selection. Manuscript, arXiv:1610.02351.
[10] Candès, E. J. and Tao, T. (2007). The Dantzig selector: Statistical estimation when p is much larger
than n (with discussion). Ann. Statist., 35, 2313–2404.
[11] Chandrasekaran, V., Parrilo, P. A., and Willsky, A. S. (2012). Latent variable graphical model selection
via convex optimization (with discussion). Ann. Statist., 40, 1935–1967.
[12] Chen, J. and Li, H. (2013). Variable selection for sparse Dirichlet-multinomial regression with an
application to microbiome data analysis. Annals of Applied Statistics, 7, 418–442.
[13] Chen, M., Ren, Z., Zhao, H., and Zhou, H. H. (2016). Asymptotically normal and efficient estimation
of covariate-adjusted Gaussian graphical model. J. Amer. Statist. Assoc., 111, 394–406.
[14] Cook, R. D. (2007). Fisher lecture: Dimension reduction in regression. Statist. Sci., 22, 1–40.
[15] Fan, J., Guo, S., and Hao, N. (2012). Variance estimation using refitted cross-validation in ultrahigh
dimensional regression. J. Roy. Statist. Soc. Ser. B, 74, 37–65.
[16] Fan, J., Han, X., and Gu, W. (2012). Control of the false discovery rate under arbitrary covariance
dependence (with discussion). J. Amer. Statist. Assoc., 107, 1019–1045.
[17] Fan, J. and Li, R. (2001). Variable selection via nonconcave penalized likelihood and its oracle properties.
J. Amer. Statist. Assoc., 96, 1348–1360.
NSL
15
[18] Fan, J., Liao, Y., and Micheva, M. (2013). Large covariance estimation by thresholding principal
orthogonal complements (with discussion). J. Roy. Statist. Soc. Ser. B, 75, 603–680.
[19] Fan, J. and Lv, J. (2010). A selective overview of variable selection in high dimensional feature space.
Statistica Sinica, 20, 101–148.
[20] Fan, Y., Demirkaya, E., Li, G., and Lv, J. (2017). RANK: large-scale inference with graphical nonlinear
knockoffs. Manuscript, arXiv:1709.00092.
[21] Fan, Y. and Lv, J. (2013). Asymptotic equivalence of regularization methods in thresholded parameter
space. J. Amer. Statist. Assoc., 108, 247–264.
[22] Gul, S., Hamilton, A., Munoz, A., Phupitakphol, T., Liu, W., Hyoju, S., Economopoulos, K., Morrison,
S., Hu, D., Zhang, W., Gharedaghi, M., Huo, H., Hamarneh, S., and Hodin, R. (2017). Inhibition of the
gut enzyme intestinal alkaline phosphatase may explain how aspartame promotes glucose intolerance and
obesity in mice. Applied Physiology, Nutrition, and Metabolism, 42, 77–83.
[23] Hastie, T., Tibshirani, R., and Friedman, J. (2009). The Elements of Statistical Learning: Data Mining,
Inference, and Prediction. 2nd ed. Springer.
[24] Horn, R. A. and Johnson, C. R. (1990). Matrix Analysis. 2nd ed. Cambridge University Press.
[25] Javanmard, A. and Montanari, A. (2014). Confidence intervals and hypothesis testing for highdimensional regression. Journal of Machine Learning Research, 15, 2869–2909.
[26] Johnstone, I. (2001). On the distribution of the largest eigenvalue in principal components analysis.
Ann. Statist., 29, 295–327.
[27] Johnstone, I. and Lu, A. (2009). On consistency and sparsity for principal components analysis in high
dimensions. J. Amer. Statist. Assoc., 104, 682–693.
[28] Jung, S. and Marron, J. (2009). PCA consistency in high dimension, low sample size context. Ann.
Statist., 37, 4104–4130.
[29] Kneip, A. and Sarda, P. (2011). Factor models and variable selection in high-dimensional regression
analysis. Ann. Statist., 39, 2410–2447.
[30] Leek, J. and Storey, J. (2007). Capturing heterogeneity in gene expression studies by surrogate variable
analysis. PLoS Genetics, 3, 1724–1735.
[31] Lee, J., Sun, D., Sun, Y., and Taylor, J. (2016). Exact post-selection inference with the lasso. Ann.
Statist., 44, 907–927.
[32] Ley, R. E., Turnbaugh, P. J., Klein, S., and Gordon, J. I. (2006). Human gut microbes associated with
obesity. Nature, 444, 1022–1023.
[33] Lin, W., Feng, R., and Li, H. (2015). Regularization methods for high-dimensional instrumental variables
regression with an application to genetical genomics. J. Amer. Statist. Assoc., 110, 270–288.
[34] Lin, W., Shi, P., Feng, R., and Li, H. (2014). Variable selection in regression with compositional
covariates. Biometrika, 101, 785–797.
[35] Lv, J. (2013). Impacts of high dimensionality in finite samples. Ann. Statist., 41, 2236–2262.
[36] Lv, J. and Liu, J. S. (2014). Model selection principles in misspecified models. J. Roy. Statist. Soc. Ser.
B, 76, 141–167.
[37] Mardia, K., Kent, J., and Bibby, J. (1979). Multivariate Analysis. New York: Academic Press.
[38] Pan, D., He, H., Song, X., and Sun, L. (2015). Regression analysis of additive hazards model with latent
variables. J. Amer. Statist. Assoc., 110, 1148–1159.
16
Z. Zheng, J. Lv and W. Lin
[39] Paul, D. (2007). Asymptotics of sample eigenstructure for a large dimensional spiked covariance model.
Statist. Sinica, 17, 1617–1642.
[40] Pritchard, J. (2001). Are rare variants responsible for susceptibility to complex diseases? Am. J. Hum.
Genet., 69, 124–137.
[41] Shen, D., Shen, H., and Marron, J. (2016). A general framework for consistency of principal component
analysis. Journal of Machine Learning Research, 17, 5218–5251.
[42] Sun, T. and Zhang, C.-H. (2012). Scaled sparse linear regression. Biometrika, 99, 879–898.
[43] Tang, C. Y. and Leng, C. (2010). Penalized high-dimensional empirical likelihood. Biometrika, 97,
905–920.
[44] Tibshirani, R. (1996). Regression shrinkage and selection via the Lasso. J. Roy. Statist. Soc. Ser. B,
58, 267–288.
[45] Turnbaugh, P. J., Ley, R. E., Mahowald, M. A., Magrini, V., Mardis, E. R., and Gordon, J. I. (2006).
An obesity-associated gut microbiome with increased capacity for energy harvest. Nature, 444, 1027–1031.
[46] van de Geer, S., Bühlmann, P., Ritov, Y., and Dezeure, R. (2014). On asymptotically optimal confidence
regions and tests for high-dimensional models. Ann. Statist., 42, 1166–1202.
[47] Wang, W. and Fan, J. (2017). Asymptotics of empirical eigen-structure for high dimensional spiked
covariance. Ann. Statist., 45, 1342–1374.
[48] White, H. (1982). Maximum likelihood estimation of misspecified models. Econometrica, 50, 1–25.
[49] Wu, G. D., Chen, J., Hoffmann, C., Bittinger, K., Chen, Y. Y., Keilbaugh, S. A., Bewtra, M., Knights,
D., Walters, W. A., Knight, R., Sinha, R., Gilroy, E., Gupta, K., Baldassano, R., Nessel, L., Li, H., Bushman, F. D., and Lewis, J. D. (2011). Linking long-term dietary patterns with gut microbial enterotypes.
Science, 334, 105–108.
[50] Zhang, S. and Zhang, C.-H. (2014). Confidence intervals for low dimensional parameters in high dimensional linear models. J. Roy. Statist. Soc. Ser. B, 76, 217–242.
[51] Zhao, P. and Yu, B. (2006). On model selection consistency of Lasso. Journal of Machine Learning
Research, 7, 2541–2563.
[52] Zheng, Z., Fan, Y., and Lv, J. (2014). High-dimensional thresholded regression and shrinkage effect. J.
Roy. Statist. Soc. Ser. B, 76, 627–649.
[53] Zou, H. and Hastie, T. (2005). Regularization and variable selection via the elastic net. J. Roy. Statist.
Soc. Ser. B, 67, 301–320.
NSL
1
Supplementary Material to “Nonsparse learning with latent variables”
Zemin Zheng, Jinchi Lv and Wei Lin
This Supplementary Material consists of two parts. Section A lists the key lemmas and presents the proofs
for main results. Additional technical proofs for the lemmas are provided in Section B.
A. Proofs of main results
A.1. Lemmas
The following lemmas are used in the proofs of main results.
Lemma 1 (Consistency of spiked sample eigenvalues). Under Conditions 1 and 2, with asymptotic probability one, the eigenvalues of the sample covariance matrix S satisfy that for any l, 1 ≤ l ≤ m,
uniformly over i ∈ Jl ,
bi → ci as q → ∞.
q −αl λ
b 0 the submatrices of X and F
b consisting of columns in supp(β ) and
Lemma 2. Denote by X0 and F
0
b
supp(γ 0 ), respectively, and e
ε = (F − F)γ + ε. For the following two events
o
n
p
b Te
and
Ee = kn−1 (X, F)
εk∞ ≤ c2 (log p)/n
o
n
p
b 0 )T e
Ee0 = kn−1 (X0 , F
εk∞ ≤ c2 (log n)/n
√
b in Condition 3 holds and the columns of X
with constant c2 > 2 2σ, when the estimation error bound of F
1/2
adopt a common scale of L2 -norm n , we have
√
√
c2
c2
2
2 2σs
4 2σ
1− 8σ22
e
e
p
n− 8σ2 ,
− √
P (E ∩ E0 ) ≥ 1 − √
c2 π log p
c2 π log n
which converges to one as n → ∞.
A.2. Proof of Theorem 1
Proof of part (a). In this part, we will focus on the convergence rates of the sample eigenvectors. The
key ingredient of this proof is to link the angle between the sample eigenvector and the space spanned by
population eigenvectors with the sum of inner products between the sample and population eigenvectors
by the cos(·) function. In this way, it suffices to show that the sum of inner products converges to one for
subspace consistency, and at the same time, deriving the convergence rates by induction. To ease readability,
we will finish the proof in four steps.
Step 1: Analysis of the subspace consistency. We first show that for any i ∈ Jl , 1 ≤ l ≤ m, the
b i is equivalent to
subspace consistency of the sample eigenvector u
X
p2ji → 1,
(A.1)
j∈Jl
b i is the inner product between the population eigenvector uj (the jth column of U) and u
bi
where pji = uTj u
b
(the ith column of U).
b are obtained through eigen-decomposition, we know that kuj k2 = 1 and kb
Since U and U
ui k2 = 1 for
P
b i )uj is the projection of u
b i on the space span{uj : j ∈ Jl }.
any i and j, 1 ≤ i, j ≤ q. Note that j∈Jl (uTj u
It gives
P
o
n
b i )uj ]
b Ti [ j∈Jl (uTj u
u
P
=
Angle(b
ui , span{uj : j ∈ Jl }) = arccos
T
b i )uj k2
kb
ui k2 · k j∈Jl (uj u
s
Tb 2
o
n P
X T
X 2 1/2
i)
j∈Jl (uj u
2 = arccos
P
b
=
arccos
arccos
.
u
)
(u
pji
i
j
T
b i )2 ]1/2
[ j∈Jl (uj u
j∈J
j∈J
l
l
2
Z. Zheng, J. Lv and W. Lin
P
Thus, Angle(b
ui , span{uj : j ∈ Jl }) → 0 is equivalent to j∈Jl p2ji → 1 as q → ∞ for any i ∈ Jl , 1 ≤ l ≤ m.
P
Moreover, the convergence rate of j∈Jl p2ji indeed provides the convergence rate of the sample eigenvector
b i to the corresponding space of population eigenvectors.
u
We will then prove the convergence rates by induction. Hereafter our analysis will be conditional on the
event E, which is defined in the proof of Lemma 1 for the consistency of the spiked sample eigenvalues and
enjoys asymptotic probability one.
Step 2: Convergence rates of sample eigenvectors
with indices in J1 . This step aims at proving
P
that uniformly over i ∈ J1 , the convergence rate of j∈J1 p2ji is given by
X
j∈J1
m
X
p2ji ≥ 1 − O{
l=2
kl q αl + km+1 K −1 q α−α1 } = 1 − O{A(1)},
(A.2)
−1 α−α
Pm
αl
t
is defined in Theorem 1. It is also the first part of induction.
q
where A(t) =
l=t+1 kl q + km+1 K
Pq
T b
2
Let P = U U = {pij }1≤i,j≤q . We have j=1 pji = 1 for any i since P is a unitary matrix. To prove (A.2),
it suffices to show
X
p2ji ≤ O{A(1)}.
j∈J2 ∪···∪Jm+1
bΛ
bU
b T . Therefore, we get a connection between Z
Recall that Z = Λ−1/2 UT WT , S = n−1 WT W = U
and P that
b T Λ−1/2 .
n−1 ZZT = n−1 Λ−1/2 UT WT WUΛ−1/2 = Λ−1/2 PΛP
For any j, 1 ≤ j ≤ q, in view of the (j, j)th entry, the above equality gives
λ−1
j
q
X
i=1
bi p2 = n−1 zT zj ,
λ
ji
j
(A.3)
−1 T
b 2
where zj is the jth column vector of ZT . It implies for any i, 1 ≤ i ≤ q, λ−1
zj zj . Based on this
j λi pji ≤ n
fact, we have
X
j∈J2 ∪···∪Jm+1
p2ji ≤
X
j∈J2 ∪···∪Jm+1
bi =
n−1 zTj zj λj /λ
n
X
X
t=1 j∈J2 ∪···∪Jm+1
2
bi ),
zjt
λj /(nλ
(A.4)
where zjt is the (j, t)th entry of Z. Conditional on the event E, by Lemma 1, Conditions 1 and 2, we have
n
X
X
t=1 j∈J2 ∪···∪Jm+1
= O{K −1 q α C
2
bi ) ≤
zjt
λj /(nλ
m
X
l=2
X
j∈J2 ∪···∪Jm+1
bi
K −1 q α λj /λ
kl q αl + km+1 /q α1 } = O{A(1)}.
(A.5)
bi are uniform over i ∈ J1 by Lemma 1, the above inequality holds uniformly
Since the convergences of λ
P
over i ∈ J1 . Inequalities (A.4) and (A.5) together entail j∈J2 ∪···∪Jm+1 p2ji ≤ O{A(1)} uniformly over i ∈ J1 ,
which implies the convergence rate in (A.2) for the sample eigenvectors with indices in J1 . It shows that
when s = 1, the convergence rate coincides with our claim that uniformly over i ∈ Jl , 1 ≤ l ≤ s,
X
j∈Jl
Note that we define
Pb
t=a st
p2ji ≥ 1 −
= 0 and
l−1
l−1
X
Y
t=1
Qb
t=a st
(1 + ki ) O{kt A(t)} − O{A(l)}.
(A.6)
i=t+1
= 1 if b < a for any positive sequence {st }.
Step 3: Convergence rates of sample eigenvectors with indices P
in J2 . Before formally completing
the proof by induction, we would like to derive the convergence rates of j∈J2 p2ji directly for i ∈ J2 to get
the basic idea of induction.
NSL
3
Since we already proved the convergence rate in (A.2) uniformly over i ∈ J1 in Step 2, summing over
i ∈ J1 gives
X X
p2ji ≥ k1 (1 − O{A(1)}) = k1 − O{k1 A(1)}.
(A.7)
i∈J1 j∈J1
Along with the fact that
X
Pq
i=1
p2ji = 1, we get
X
p2ji =
i∈J2 ∪···∪Jm+1 j∈J1
= k1 −
X X
i∈J1 j∈J1
q X
X
p2ji −
i=1 j∈J1
X X
p2ji =
i∈J1 j∈J1
q
XX
j∈J1 i=1
p2ji −
X X
p2ji
i∈J1 j∈J1
p2ji ≤ k1 − (k1 − O{k1 A(1)}) = O{k1 A(1)}.
The above result is important as it also implies that uniformly over i ∈ J2 ,
X
p2ji ≤ O{k1 A(1)}.
(A.8)
(A.9)
j∈J1
For the sample eigenvector u
bi with index i ∈ J2 , in order to find a lower bound for
it as
X
X
X
p2ji = 1 −
p2ji −
p2ji .
j∈J2
j∈J1
P
j∈J2
p2ji , we write
(A.10)
j∈J3 ∪···∪Jm+1
P
P
The upper bound of j∈J1 p2ji was provided in (A.9). For the second term j∈J3 ∪···∪Jm+1 p2ji , similar to
(A.4) and (A.5) in Step 2, by Lemma 1, Conditions 1 and 2, we have uniformly over i ∈ J2 ,
X
j∈J3 ∪···∪Jm+1
m
X
p2ji ≤ O{K −1 q α C
l=3
kl q αl + km+1 /q α2 } = O{A(2)}.
Plugging the above two bounds into (A.10) gives
X
p2ji ≥ 1 − O{k1 A(1)} − O{A(2)},
j∈J2
which shows that the uniform convergence rate of the sample eigenvectors u
bi over i ∈ J2 . Together with the
uniform convergence rate over i ∈ J1 established in Step 2, our claim in (A.6) gives the uniform convergence
rates of the sample eigenvectors u
bi over i ∈ J1 ∪ J2 .
Step 4: Convergence rates of sample eigenvectors with indices in J3 to Jm . In this step, we
will complete the proof by induction. Specifically, we show that the claim in (A.6) holds for any fixed s,
3 ≤ s ≤ m, based on the induction assumption that the claim holds for s − 1.
By the induction assumption, we have uniformly over i ∈ Jl , 1 ≤ l ≤ s − 1,
X
j∈Jl
p2ji ≥ 1 −
l−1
l−1
X
Y
t=1
(1 + ki ) O{kt A(t)} − O{A(l)}.
i=t+1
By a similar argument as in (A.7) and (A.8), it follows that
X
X
i∈Jl+1 ∪···∪Jm+1 j∈Jl
p2ji ≤ kl
l−1
l−1
X
Y
t=1
i=t+1
(1 + ki ) O{kt A(t)} + O{A(l)} .
Similarly as in Step 3, for any i ∈ Js , to get the convergence rate of
X
j∈Js
p2ji = 1 −
X
j∈J1 ∪···∪Js−1
p2ji −
P
X
j∈Js
p2ji , we write it as
j∈Js+1 ∪···∪Jm+1
p2ji .
(A.11)
4
Z. Zheng, J. Lv and W. Lin
P
We will first derive the convergence rate of j∈J1 ∪···∪Js−1 p2ji . When 1 ≤ l ≤ s − 1, we have i ∈ Js ⊂
Jl+1 ∪ · · · ∪ Jm+1 . In view of (A.11), it gives that uniformly over i ∈ Js ,
X
j∈Jl
p2ji ≤ kl
l−1
l−1
X
Y
t=1
i=t+1
(1 + ki )]O{kt A(t)} + O{A(l)} .
Summing over l = 1, . . . , s − 1, we get
X
p2ji
j∈J1 ∪···∪Js−1
≤
s−1
X
l−1
Y
X
l−1
kl
t=1
l=1
(1 + ki ) O{kt A(t)} + O{A(l)} .
i=t+1
To simplify the above expression, exchanging the summation order with respect to l and t gives
X
j∈J1 ∪···∪Js−1
p2ji ≤
=
l−1
s−1 X
X
kl
l=1 t=1
s−2 X
s−1
X
i=t+1
kl
t=1 l=t+1
l−1
Y
l−1
Y
s−1
X
O{kl A(l)}
(1 + ki ) O{kt A(t)} +
l=1
(1 + ki ) O{kt A(t)} +
i=t+1
s−1
X
O{kt A(t)}.
t=1
Then we combine the coefficients of kt A(t) to get
X
j∈J1 ∪···∪Js−1
p2ji ≤
s−2
X
O{kt A(t)} 1 +
t=1
s−1
X
l=t+1
kl
l−1
Y
i=t+1
X
(1 + ki ) +
O{kt A(t)}.
t=s−1
Since it is immediate to conclude by induction that
1+
s−1
X
kl
l=t+1
l−1
Y
i=t+1
(1 + ki ) = 1 + kt+1 + kt+2 (1 + kt+1 ) + · · ·
+ ks−1 (1 + ks−2 )(1 + ks−3 ) · · · (1 + kt+2 )(1 + kt+1 ) =
s−1
Y
(1 + ki ),
i=t+1
we then have
X
j∈J1 ∪···∪Js−1
p2ji ≤
=
s−2
Y
X
s−1
t=1
i=t+1
s−1
X
s−1
Y
t=1
i=t+1
X
(1 + ki ) O{kt A(t)} +
O{kt A(t)}
t=s−1
(1 + ki ) O{kt A(t)}.
On the other hand, similar to (A.4) and (A.5) in Step 2, we have uniformly over i ∈ Js ,
X
p2ji ≤ O{A(s)}.
j∈Js+1 ∪···∪Jm+1
Combining the above two bounds gives the convergence rate of
X
j∈Js
p2ji
≥1−
s−1
Y
X
s−1
t=1
i=t+1
P
j∈Js
p2ji uniformly over i ∈ Js as
(1 + ki ) O{kt A(t)} − O{A(s)}.
Together with the induction assumption that our claim in (A.6) holds uniformly over i ∈ Jl , 1 ≤ l ≤ s − 1,
we know that the claim also holds uniformly over i ∈ Jl , 1 ≤ l ≤ s. Therefore, by induction, the results in
part (a) of Theorem 1 hold uniformly over i ∈ Jl , 1 ≤ l ≤ m.
NSL
5
Proof of part (b). In this part, we will show that when each group of spiked eigenvalues has size one
(that is, kl = 1 for any l, 1 ≤ l ≤ m), the convergence rates of the angles between the sample score vectors
Wb
ui and the population score vectors Wui , 1 ≤ i ≤ K, are at least as fast as those of the angles between
the corresponding sample and population eigenvectors established in part (a) of Theorem 1. The key idea
is to conduct delicate analysis on the cos(·) function of the angles between the sample score vectors and
population score vectors, where some results about the sample eigenvalues derived in the proof of Lemma 1
will be used.
b i (i ∈ Jl ) to the space
When each group has size one, we have K = m and the convergence rates of u
b i to ui , 1 ≤ i ≤ K. Denote by θii = Angle(b
span{uj : j ∈ Jl } become the convergence rates of u
ui , ui ) and
ωii = Angle(Wb
ui , Wui ). Then the results in part (a) give that uniformly over 1 ≤ i ≤ K,
cos2 (θii ) = p2ii ≥ 1 −
i−1
X
t=1
2i−t−1 O{A(t)} − O{A(i)}.
(A.12)
T
bi
bΛ
bU
b ,u
b i would be the eigenvector of WT W corresponding to the eigenvalue nλ
Since S = n−1 WT W = U
with L2 -norm 1. It follows that
q
bi cos(θii )
Tb
T
b
nλ
nλi ui ui
(Wui ) Wb
ui
=
=q
.
cos(ωii ) =
kWui k2 kWb
ui k 2
kWui k2
bi kWui k2
nλ
Squaring both sides above gives
cos2 (ωii ) =
bi .
Therefore, it suffices to show kWui k22 ≤ nλ
bi cos2 (θii )
nλ
.
kWui k22
(A.13)
T
bΛ
bU
b that
For the term kWui k22 , it follows from WT W = nU
T
bΛ
bU
b ui = n
kWui k22 = uTi WT Wui = nuTi U
q
X
j=1
By further making use of equality (A.3), we have
kWui k22 = n
q
X
j=1
2
bj (uT u
λ
i bj ) = n
q
X
j=1
bj p2 .
λ
ij
bj p2 = λi zT zi ,
λ
ij
i
where zi is the ith column vector of ZT . On the other hand, inequality (A.33) in the proof of Lemma 1 gives
bi , 1 ≤ i ≤ K. Under the current setting that each group has size
a lower bound for the sample eigenvalues λ
one, it gives
bi ≥ ϕ1 (n−1 λi zi zT ) = ϕ1 (n−1 λi zT zi ) = n−1 λi zT zi ,
λ
i
i
i
where ϕ1 (·) denotes the largest eigenvalue of a given matrix. It follows that
Therefore, in view of (A.13), we get
bi ≥ λi zT zi = kWui k2 .
nλ
i
2
cos2 (ωii ) ≥ cos2 (θii ),
which means that the convergence rate of the sample score vector is at least as good as that of the corresponding sample eigenvector. Then it follows from (A.12) that uniformly over 1 ≤ i ≤ K,
cos2 (ωii ) ≥ 1 −
i−1
X
t=1
2i−t−1 O{A(t)} − O{A(i)},
which completes the proof of part (b) of Theorem 1.
6
Z. Zheng, J. Lv and W. Lin
A.3. Proof of Proposition 1
By Condition 4, the inequality kn−1/2 (X, F)δk2 ≥ ckδk2 holds for any δ satisfying kδk0 < M with signifb by analyzing the estimation errors of
icant probability 1 − θn,p . We now derive a similar result for (X, F)
confounding factors F.
By the estimation error bound in Condition 3, we have for any 1 ≤ j ≤ K,
kfj − bfj k22 ≤ kfj k22 + kbfj k22 − 2 f′jbfj = n + n − 2kfj k2 kbfj k2 cos(ωjj )
= 2n − 2n cos(ωjj ) = 2n{1 − cos(ωjj )} ≤
c22 log n
.
4K 2T 2
Since the above bound does not vary with the index j, it gives the uniform confounding factor estimation
error bound
c2 p
log n.
(A.14)
max kfj − bfj k2 ≤
1≤j≤K
2KT
b First of all, it follows from Condition 4 and the
Now we proceed to prove the inequality for (X, F).
triangular inequality that
−1/2
b
b
kn−1/2 (X, F)δk
(X, F)δk2 − kn−1/2 (X, F)δ − n−1/2 (X, F)δk
2 ≥ kn
2
−1/2
−1/2
b
b
≥ ckδk2 − n
k(F − F)δ 1 k2 ≥ ckδk2 − n
max k(fj − fj )k2 kδ 1 k1 ,
1≤j≤K
√
√
where δ 1 is a subvector of δ consisting of the last K components. Note that kδ 1 k1 ≤ Kkδ 1 k2 ≤ Kkδk2 .
Further applying inequality (A.14) yields
√
c2 p
−1/2
b
log n · Kkδk2 ≥ c1 kδk2 ,
kn−1/2 (X, F)δk
·
2 ≥ ckδk2 − n
2KT
q
log n
c2
where c1 is some positive constant no larger than c − 2T
nK . It is clear that c1 is smaller than but close
to c when n is relatively large. In view of the tail probabilities in Conditions 3 and 4, the above inequality
holds with probability at least 1 − θ1 − θ2 . Thus, we finish the proof of Proposition 1.
A.4. Proof of Theorem 2
With Proposition 1, we will apply a similar idea as in [52] to prove the global properties. The proof consists
of two parts. The first part shows the model selection consistency property with the range of λ given in
Theorem 2. Based on the first part, several oracle inequalities will then be induced. We will first prove the
properties when the columns of design matrix X have a common scale of L2 -norm n1/2 as a benchmark,
meaning that β ∗ = β and L = 1, and then illustrate the results in general cases.
Part 1: Model selection consistency. This part contains two steps. In the first step, it will be shown
p
b T b T )T is no larger
that when c−1
1 c2 (2s + 1)(log p)/n < λ < b0 , the number of nonzero elements in (β , γ
than s conditioning on the event Ee defined in Lemma 2. We prove this by using the global optimality of
bT , γ
b T )T .
(β
By the hard-thresholding property [52, Lemma 1] and λ < b0 , any nonzero component of the true
bT , γ
b T )T is greater than λ, which ensures
regression coefficient vector (β T , γ T )T or of the global minimizer (β
0
0
bT , γ
bT , γ
b T )T }k1 = λ2 k(β
b T )T k0 /2 and kpλ {(βT0 , γ T0 )T }k1 = sλ2 /2. Thus,
that kpλ {(β
o
n
n
o
o
n T
bT , γ
b ,γ
b T )T k0 − s λ2 /2.
b T )T
= k(β
− pλ (β T0 , γ T0 )T
pλ (β
1
1
T
b ,γ
b T )T − (β T0 , γ T0 )T . Direct calculation yields
Denote by δ = (β
o
o
n
n T
1
2
b ,γ
b
b
b T )T − Q (βT0 , γ T0 )T = 2−1 n− 2 (X, F)δ
Q (β
− n−1 e
εT (X, F)δ
2
n
o
bT , γ
b T )T k0 − s λ2 /2,
+ k(β
(A.15)
NSL
b + ε, the sum of the random error vector ε and estimation errors (F − F)γ.
b
where e
ε = (F − F)γ
e we have
On the other hand, conditional on event E,
b
b ∞ kδk1
|n−1 e
εT (X, F)δ|
≤ kn−1e
εT (X, F)k
p
p
1
≤ c2 (log p)/nkδk1 ≤ c2 (log p)/nkδk02 kδk2 .
7
(A.16)
bT , γ
b T )T k0 < M ,
In addition, by Condition 6 and the definition of SM/2 , we obtain kδk0 ≤ k(βT0 , γ T0 )T k0 +k(β
b with bound c1 by Proposition 1. Thus, we have
where M is the robust spark of (X, F)
1
b
kn− 2 (X, F)δk
2 ≥ c1 kδk2 .
(A.17)
Plugging inequalities (A.16) and (A.17) into (A.15) gives that
o
o
n
n T
p
1
b ,γ
b T )T − Q (βT0 , γ T0 )T ≥ 2−1 c21 kδk22 − c2 (log p)/nkδk02 kδk2
Q (β
n
o
bT , γ
b T )T k0 − s λ2 /2.
+ k(β
(A.18)
bT , γ
b T )T ensures that
Thus, the global optimality of (β
r
n
o
1
log p
2
−1 2
bT , γ
b T )T k0 − s λ2 /2 ≤ 0.
kδk02 kδk2 + k(β
2 c1 kδk2 − c2
n
After completing the squares in the above inequality, we get
r
2
n
o
h
i
1 2
log p
c2 log p
c2
bT , γ
b T )T k0 − s λ2 ≤ 0.
kδk02 −
kδk0 + k(β
c1 kδk2 −
c1
n
c1
n
q
h
i
1 2
Since c1 kδk2 − cc21 logn p kδk02 ≥ 0, it gives
2
n
o
T
c2
log p
T T
2
b
b ) k0 − s λ ≤
k(β , γ
kδk0 .
c1
n
(A.19)
bT , γ
bT , γ
b T )T k0 by the above inequality. Let k = k(β
b T )T k0 . Then
We continue to bound the value of k(β
T
b ,γ
b T )T − (β T0 , γ T0 )T k0 ≤ k + s. Thus, it follows from (A.19) that
kδk0 = k(β
(k − s)λ2 ≤
c2
c1
2
log p
(k + s).
n
Organizing it in terms of k and s, we get
2
k λ −
Since λ > c1−1 c2
Thus we have
c2
c1
2
log p
n
!
2
≤s λ +
c2
c1
2
log p
n
!
.
(A.20)
p
log p
2 2
2
2
2
(2s + 1) log p/n, we have λ2 − (c−1
1 c2 ) (2s+ 1) n > 0 and λ c1 n− c2 log p > 2c2 s log p.
2c22 log p
λ2 c21 n−c22 log p
< 1/s. Then it follows from inequality (A.20) that
k≤s
(λ2 + ( cc12 )2 logn p )
(λ2 − ( cc12 )2 logn p )
=s 1+
2c22 log p
2
λ2 c1 n − c22 log p
T
b ,γ
b T )T satisfies
Therefore, the number of nonzero elements in (β
T
b ,γ
b T )T
(β
0
≤ s.
< s + 1.
8
Z. Zheng, J. Lv and W. Lin
The second step is based on the first step, where we will use proof by contradiction to show that
√
bT , γ
b T )T ) with the additional assumption λ < b0 c1 / 2 in the theorem. Supsupp((β T0 , γ T0 )T ) ⊂ supp((β
bT , γ
b T )T , and we denote the number of missed true coefficients as
pose that supp((β T , γ T )T ) 6⊂ supp(β
0
0
T T T
b ,γ
b )
k = supp (β T0 , γ T0 )T \supp (β
≥ 1.
bT , γ
bT , γ
b T )T k0 ≥ s − k and kδk0 ≤ k(β
b T )T k0 + k(βT0 , γ T0 )T k0 ≤ 2s by the first step.
Then we have k(β
Combining these two results with inequality (A.18) yields
!
r
o
o
n
n T
2s
log
p
T
b ,γ
b )T − Q (β T0 , γ T0 )T ≥ 2−1 c21 kδk2 − c2
kδk2 − kλ2 /2.
(A.21)
Q (β
n
bT , γ
b T )T ), we have |δj | ≥ b0 with b0 the lowest signal
Note that for each j ∈ supp((β T0 , γ T0 )T ) \ supp((
β
√
strength defined in Condition 6. Thus, kδk2 ≥ kb0 , which together with Condition 6 entails
√
p
4−1 c21 kδk2 ≥ 4−1 c21 kb0 ≥ 4−1 c21 b0 > c2 (2s log p)/n.
Thus, it follows from (A.21) that
o
o
n
n T
b ,γ
b T )T − Q (β T0 , γ T0 )T ≥ 4−1 c21 kδk22 − kλ2 /2 ≥ 4−1 c21 kb20 − kλ2 /2 > 0,
Q (β
√
where the last step is because of the additional assumption λ < b0 c1 / 2. The above inequality contradicts
bT , γ
bT , γ
b T )T . Thus, we have supp((β T0 , γ T0 )T ) ⊂ supp((β
b T )T ). Combining
with the global optimality of (β
T
T
b ,γ
b ,γ
b T )T k0 ≤ s from the first step, we know that supp{(β
b T )T } = supp{(β T , γ T )T }.
this with k(β
0
0
Part 2: Prediction and estimation losses. In this part, we will bound the prediction and estimation
bT , γ
b T )T , conditional
losses. The idea is to get the L2 -estimation loss bound by the global optimality of (β
on the event Ee ∩ Ee0 defined in Lemma 2. Then by similar techniques as in the first part, we would derive
bounds for the prediction and estimation losses.
b 0 are the submatrices of X and F
b consisting of columns in supp(β0 ) and supp(γ 0 ),
Recall that X0 , F
respectively. Conditioning on Ee∩ Ee0 , we have kδk0 ≤ s by the model selection consistency established before.
Thus, applying the Cauchy-Schwarz inequality and definition of Ee0 gives
b 0 )δ| ≤ kn−1 e
b 0 )k∞ kδk1
|n−1 e
εT (X0 , F
εT (X0 , F
r
r
log n
s log n
kδk1 ≤ c2
kδk2 .
≤ c2
n
n
(A.22)
In views of (A.15) and (A.17), it follows from inequality (A.22) and the model selection consistency property
bT , γ
b T )T k0 = s that
k(β
o
o
n
n T
b ,γ
b T )T − Q (β T0 , γ T0 )T
Q (β
2
−1 T
bT , γ
b
b + k(β
e
b T )T k0 − s λ2 /2
= 2−1 kn−1 (X, F)δk
ε (X, F)δ
2−n
r
−1 2
2
−1 T
−1
2
b 0 )δ ≥ 2 c kδk2 − c2 s log n kδk2 .
≥ 2 c1 kδk2 − n e
ε (X0 , F
1
n
bT , γ
b T )T is the global optimizer of Q, we have
Since (β
2−1 c21 kδk2
− c2
r
s log n
≤ 0,
n
NSL
9
which gives the L2 and L∞ estimation loss bounds as
T
b ,γ
b T )T − (β T0 , γ T0 )T k2 = kδk2 ≤ 2c−2
k(β
1 c2
p
(s log n)/n,
bT , γ
bT , γ
b T )T − (β T0 , γ T0 )T k∞ ≤ k(β
b T )T − (β T0 , γ T0 )T k2 ≤ 2c−2
k(β
1 c2
For Lq -estimation losses with 1 ≤ q < 2, applying Hölder’s inequality gives
p
(s log n)/n.
X
X
X 2 1 1
1
bT , γ
b T )T −(β T0 , γ T0 )T kq = (
|δj |2 ) 2 (
|δj |q )1/q ≤ (
k(β
1 2−q ) q − 2
j
j
1
1
q−2
= kδk2 kδk0
δj 6=0
p
(log n)/n.
≤ 2c−2
1 c2 s
1
q
T
b ,γ
b T )T is the global minimizer, it follows
Next we prove the bound for oracle prediction loss. Since (β
from (A.15) and the model selection consistency property that
bT , γ
b β
b T )T − (β T0 , γ T0 )T }k2
n−1/2 k(X, F){(
o1/2
n
o1/2 n
p
b
b 0 )T ε̃k∞ kδk1
≤ 2n−1e
εT (X, F)δ
≤ 2c2 c−1
≤ 2kn−1 (X0 , F
s(log n)/n,
1
where the last step is because of the L1 estimation loss bound proved before. Then for the oracle prediction
loss, together with (A.34) in the proof of Lemma 2, it follows that
b β
bT , γ
b T )T − (X, F)(β T0 , γ T0 )T k2
n−1/2 k(X, F)(
p
p
−1 √
b
≤ 2c2 c−1
s(log n)/n + n−1/2 k(F − F)γ
s + c2 /2) (log n)/n.
0 k2 ≤ (2c2 c1
1
Last we will derive our results for general cases when the L2 -norms of columns of X are not of the
common scale n1/2 . Note that the penalized least squares in (2) can be rewritten as
e ∗ − Fγk
b 22 + kpλ (β T∗ , γ T )T k1 ,
= (2n)−1 ky − Xβ
Q (β T , γ T )T
e is the matrix with the L2 -norm of each column rescaled to n1/2 and
where X
β∗ = n−1/2 (β1 kx1 k2 , . . . , βp kxp k2 )T
is the corresponding coefficient vector defined in (2). By Conditions 5 and 6, the same argument applies to
derive the model selection consistency property and the bounds on oracle prediction and estimation losses
T T
bT , γ
for (β
∗ b ) since the relationship between λ and signal strength keeps the same even if L 6= 1. Based
b implies that of β.
b And the bound on
on Condition 5, it is clear that the model selection consistency of β
∗
b
b
b they can be
e
prediction loss does not change since Xβ∗ = Xβ. As for the bounds of estimation losses on β,
deduced as
p
p
b − β k2 ≤ 2c−2 c2 L (s log n)/n, kβ
b − β kq ≤ 2c−2 c2 Ls q1 (log n)/n,
kβ
0
0
1
1
p
b − β k∞ ≤ kβ
b − β k2 ≤ 2L−1 c−2 c2 (s log n)/n.
kβ
0
0
1
The tail probability for these results to hold is at most the sum of the tail probabilities in Conditions 3-5
and Lemma 2. Thus, we know that these properties hold simultaneously with probability at least
1−
√
√
c2
c2
2
2
4 2σ
2 2σs
√
p1− 8σ2 + √
n− 8σ2 − θ1 − θ2 − θ3 .
c2 π log p
c2 π log n
It concludes the proof of Theorem 2.
10
Z. Zheng, J. Lv and W. Lin
B. Additional technical details
The following lemma is needed in proving Lemma 1.
Lemma 3 (Weyl’s inequality [24]). If A and B are m × m real symmetric matrices, then for all
k = 1, . . . , m,
ϕk (A) + ϕm (B)
ϕk (A) + ϕ1 (B)
ϕk−1 (A) + ϕ2 (B)
ϕk+1 (A) + ϕm−1 (B)
,
≤ ϕk (A + B) ≤
..
..
.
.
ϕm (A) + ϕk (B)
ϕ1 (A) + ϕk (B)
where ϕi (·) is the function that takes the ith largest eigenvalue of a given matrix.
B.1. Proof of Lemma 1
The main idea of proving Lemma 1 is to use induction to show that the sample eigenvalues divided by their
corresponding orders of q will be convergent in an event with asymptotic probability one. To ease readability,
the proof is divided into three steps.
Step 1: Large probability event E. In this step, we will define an event E and show that its probability
approaches one when q increases to infinity. Our later discussion will be conditional on this event. Denote
a series of events by Ejt , 1 ≤ j ≤ q, 1 ≤ t ≤ n, such that
2
Ejt = {zjt
≤ K −1 q α },
where zjt is the (j, t)th entry of Z. By Condition 2, the events Ejt satisfy a uniform tail probability bound
c
P (Ejt
) = o(q −1 n−1 ). Let E = ∩nt=1 ∩qj=1 Ejt be the intersection of all events in the series. Then the probability
of event E converges to one since
c
P (E c ) = P (∪nt=1 ∪qj=1 Ejt
)≤
q
n X
X
t=1 j=1
c
P (Ejt
) = nq · o(q −1 n−1 ) → 0, as q → ∞.
Step 2: Convergence of eigenvalues with indices in J1 . This is the first part of induction. We will
bi → ci , as q → ∞.
show that conditional on event E, uniformly over i ∈ J1 , q −α1 λ
Denote by C the q × q diagonal matrix with the first K diagonal components equaling to cj , 1 ≤ j ≤ K,
and the rest diagonal components 1. We decompose Z, C and Λ into block matrices according to the index
sets J1 , J2 , . . . , Jm+1 such that
Z1
C1 O · · ·
O
Λ1 O · · ·
O
Z2
O C2 · · ·
O Λ2 · · ·
O
O
Z = . ,C = .
,Λ = .
(A.23)
.
.
.
.. .
.
.
..
..
..
..
..
..
..
..
.
Zm+1
O
O
···
Cm+1
O
O
···
Λm+1
Then for the dual matrix SD , we have
SD = n
−1
T
Z ΛZ = n
−1
m+1
X
ZTl Λl Zl .
(A.24)
l=1
Divided by q α1 on both sides of (A.24) gives
q −α1 SD = n−1 q −α1 ZT1 Λ1 Z1 + n−1 q −α1
m
X
ZTl Λl Zl + n−1 q −α1 ZTm+1 Λm+1 Zm+1 .
(A.25)
l=2
We will show the sum of the last two terms above converges to the zero matrix in Frobenius norm, where
the Frobenius norm is defined as kAkF = {tr(AAT )}1/2 for a given matrix A.
NSL
(l)
11
(l)
For any l, 1 ≤ l ≤ m, let λt and ct be the tth diagonal elements of Λl and Cl , respectively.
Conditional
P
T
on event E, for any j and k, 1 ≤ j, k ≤ n, the absolute value of the (j, k)th element in m
Z
Λ
l Zl is
l=2 l
|
(l)
kl
m X
X
l=2 t=1
(l) (l) (l)
λt ztj ztk | ≤ K −1 q α
kl
m X
X
(l)
λt ,
l=2 t=1
(l)
where ztj and ztk are the (t, j)th and (t, k)th elements in Zl , respectively. By Condition 1, uniformly over
(l)
(l)
1 ≤ l ≤ m and 1 ≤ t ≤ kl , λt = O(q αl ct ). Then it follows that
kn−1 q −α1
m
X
l=2
ZTl Λl Zl kF ≤ n−1 q −α1 (nK −1 q α
= O{q −α1 K −1 q α
kl
m X
X
l=2 t=1
Similarly we would get
(l)
q αl ct } = O{K −1 q α
kl
m X
X
(l)
λt )
l=2 t=1
m
X
l=2
kl Cq αl /q α1 }.
kn−1 q −α1 ZTm+1 Λm+1 Zm+1 kF ≤ K −1 q α km+1 C/q α1 .
Together with α < min{∆, αm − 1} by Condition 2 and km+1 < q, we have
kn−1 q −α1
m
X
l=2
ZTl Λl Zl + n−1 q −α1 ZTm+1 Λm+1 Zm+1 kF
m
X
kl q αl + km+1 )K −1 q α C/q α1 } → 0, as q → ∞.
≤ O{(
(A.26)
l=2
By a similar argument, under Condition 1, we have
kn−1 q −α1 ZT1 Λ1 Z1 − n−1 ZT1 C1 Z1 kF = kn−1 ZT1 (q −α1 Λ1 − C1 )Z1 kF
≤ n−1 [nK −1 q α
k1
X
t=1
(1)
(1)
(q −α1 λt − ct )] ≤ k1 K −1 q α · O(q −∆ ) → 0, as q → ∞.
(A.27)
In view of (A.25), it is immediate that
kq −α1 SD − n−1 ZT1 C1 Z1 kF ≤ kn−1 q −α1 ZT1 Λ1 Z1 − n−1 ZT1 C1 Z1 kF
m
X
+ kn−1 q −α1
ZTl Λl Zl + n−1 q −α1 ZTm+1 Λm+1 Zm+1 kF → 0, as q → ∞.
(A.28)
l=2
Further applying [24, Corollary 6.3.8] gives as q → ∞,
max |ϕi (q −α1 SD ) − ϕi (n−1 ZT1 C1 Z1 )| ≤ kq −α1 SD − n−1 ZT1 C1 Z1 kF → 0.
(A.29)
1≤i≤n
1/2
1/2
Note that n−1 ZT1 C1 Z1 shares the same nonzero eigenvalues with its due matrix n−1 C1 Z1 ZT1 C1
dimensionality k1 . It follows from (A.29) that
1/2
1/2
max |ϕi (q −α1 SD ) − ϕi (n−1 C1 Z1 ZT1 C1 )| → 0.
of
(A.30)
i∈J1
Moreover, by part (b) of Condition 2, we have
1/2
1/2
1/2
1/2
max |ϕi (n−1 C1 Z1 ZT1 C1 ) − ϕi (C1 )| ≤ kC1 (n−1 Z1 ZT1 − Ik1 )C1 kF → 0.
i∈J1
Therefore, (A.30) and (A.31) together yield that uniformly over i ∈ J1 ,
bi = ϕi (q −α1 SD ) → ϕi (C1 ) = ci ,
q −α1 λ
(A.31)
12
Z. Zheng, J. Lv and W. Lin
as q → ∞. It completes the proof of Step 2.
Step 3: Convergence of eigenvalues with indices in J2 , . . . , Jm . As the second part of induction, for
bi → ci for any i ∈ Jt , as q → ∞. The basic idea in this step is to
any fixed t, 2 ≤ t ≤ m, we will show q −αt λ
bi , and show that
use Weyl’s inequality (Lemma 3) to get both a lower bound and an upper bound of q −αt λ
they converge to the same limit.
We derive the upper bound first. Divided by q αt on both sides of (A.24) gives
q −αt SD = n−1 q −αt
t−1
X
ZTl Λl Zl + n−1 q −αt
l=1
m+1
X
ZTl Λl Zl .
l=t
Applying Weyl’s inequality, we get
t−1
m+1
X
X
ZTl Λl Zl /nq αt ) + ϕi−Pt−1 kl (n−1 q −αt
ϕi (q −αt SD ) ≤ ϕ1+Pt−1 kl (
ZTl Λl Zl )
l=1
l=1
l=1
l=t
= ϕi−Pt−1 kl (n−1 q −αt
l=1
m+1
X
ZTl Λl Zl ),
(A.32)
l=t
Pt−1
Pt−1
where the first term is indeed zero since n−1 q −αt l=1 ZTl Λl Zl has a rank no more than l=1 kl . It gives
an upper bound of ϕi (q −αt SD ). By the same argument as (A.28) in Step 2, under Conditions 1 and 2, we
have
m+1
X
kn−1 q −αt
ZTl Λl Zl − n−1 ZTt Ct Zt kF → 0, as q → ∞.
l=t
Similar to (A.29), it implies the upper bound of ϕi (q −αt SD ) in (A.32) converges to the same limit as
ϕi−Pt−1 kl (n−1 ZTt Ct Zt ) uniformly over i ∈ Jt as q → ∞.
l=1
On the other hand, by Weyl’s inequality, we also have
ϕi (q −αt SD ) ≥ ϕi (n−1 q −αt
≥ ϕi (n−1 q −αt
t
X
ZTl Λl Zl ) + ϕn (n−1 q −αt
l=1
t
X
m+1
X
ZTl Λl Zl )
l=t+1
ZTl Λl Zl ),
l=1
Pm+1
Pm+1
where the second term vanishes since the eigenvalues of l=t+1 ZTl Λl Zl are non-negative. In fact, n−1 q −αt l=t+1 ZTl Λl Zl
would converge to a zero
Pt matrix in Frobenius norm under Conditions 1 and 2, similarly as in (A.26). For
the term ϕi (n−1 q −αt l=1 ZTl Λl Zl ), we use Weyl’s inequality once more to get
ϕPt
l=1
kl (n
−1 −αt
q
t−1
X
ZTl Λl Zl )
l=1
≤ ϕi (n−1 q −αt
t
X
ZTl Λl Zl ) + ϕ1−i+Ptl=1 kl (−n−1 q −αt ZTt Λt Zt ).
l=1
Note
Pt−1 that the term on the left hand side is indeed zero since the inside matrix has a rank no more than
l=1 kl . It follows that
ϕi (n−1 q −αt
t
X
l=1
=ϕkt −(1−i+
Pt
l=1
ZTt Λt Zt ) ≥ −ϕ1−i+Ptl=1 kl (−n−1 q −αt ZTt Λt Zt )
kl )+1 (n
−1 −αt
q
ZTt Λt Zt ) = ϕi−Pt−1 kl (n−1 q −αt ZTt Λt Zt ),
l=1
where we make use of the fact that ϕi (A) = −ϕn−i+1 (−A) for any n × n real symmetric matrix A, and any
1 ≤ i ≤ n.
NSL
13
Therefore, we get a lower bound ϕi−Pt−1 kl (n−1 q −αt ZTt Λt Zt ) for ϕi (q −αt SD ). In terms of sample eigenl=1
bi , i ∈ Jt , 1 ≤ t ≤ m,
values, the above argument shows that for any λ
bi = ϕi (SD ) ≥ ϕ Pt−1 (n−1 ZT Λt Zt ),
λ
t
kl
i−
(A.33)
l=1
which is useful in proving the convergence properties of the sample score vectors.
Now we show that the two bounds converge to the same limit. Similar to (A.27), as q → ∞, we have
kn−1 q −αt ZTt Λt Zt − n−1 ZTt Ct Zt kF → 0,
which gives
max ϕi−Pt−1 kl (n−1 q −αt ZTt Λt Zt ) − ϕi−Pt−1 kl (n−1 ZTt Ct Zt ) → 0.
i∈Jt
l=1
l=1
SD ) converges to the same limit as ϕi−Pt−1 kl (n−1 ZTt Ct Zt ) uniIt shows that the lower bound of ϕi (q
l=1
formly over i ∈ Jt , so does the upper bound in (A.32). It follows that ϕi (q −αt SD ) would also converge to
the same limit as ϕi−Pt−1 kl (n−1 ZTt Ct Zt ) uniformly over i ∈ Jt . That is, as q → ∞,
−αt
l=1
max |ϕi (q −αt SD ) − ϕi−Pt−1 kl (n−1 ZTt Ct Zt )| → 0.
i∈Jt
l=1
By a similar argument as in (A.30) and (A.31), we then have
ϕi (q −αt SD ) → ϕi−Pt−1 kl (Ct ) = ci ,
l=1
uniformly over i ∈ Jt , as q → ∞. Along with the first step of induction in Step 2, we finish the proof of
Lemma 1.
B.2. Proof of Lemma 2
To prove the probability bound in Lemma 2, we will apply Bonferroni’s inequality and Gaussian tail probb + ε, some important bounds are needed before continuation. First, the
ability bound. Since e
ε = (F − F)γ
inequality kγk1 ≤ KT follows immediately from the fact kγk∞ ≤ T . Moreover, based on the estimation
b in Condition 3, we know that inequality (A.14) holds. These two inequalities yield
error bound of F
c2 p
c2 p
b
b
log n =
log n,
(A.34)
k(F − F)γk
2 ≤ kγk1 · max kfj − fj k2 ≤ KT ·
1≤j≤K
2KT
2
which gives
n
−1
|xTi (F
b
− F)γ|
≤n
−1/2
c2
b
k(F − F)γk
2 ≤
2
r
c2
log n
≤
n
2
r
log p
.
n
(A.35)
p
b
Similarly we have n−1 |fTj (F − F)γ|
≤ 2−1 c2 (log n)/n.
Now we proceed to prove the probability bounds of the two events. Recall that both fj and bfj have
been rescaled to have L2 -norm n1/2 and ε ∼ N (0, σ 2 In ) (Section 2). Given xi and bfj , it follows that
T
n−1 xTi ε ∼ N (0, σ 2 /n) and n−1bfj ε ∼ N (0, σ 2 /n) for any i and j. By Bonferroni’s inequality, the tail
probability of Ee satisfies
P (Eec ) ≤
p
X
i=1
K
X
p
p
T
ε| > c2 (log p)/n .
P |n−1bfj e
P |n−1 xTi e
ε| > c2 (log p)/n +
j=1
By inequality (A.35) and Gaussian tail probability bound, for the first term on the right hand side above,
we have
!
!
r
r
p
p
X
X
|xTi ε|
|xTi e
ε|
log p
log p
−1 T
b
P
≤
> c2
> c2
− n |xi (F − F)γ|
P
n
n
n
n
i=1
i=1
!
r
√
p
p
X
X
c2
4σ
|xTi ε|
c2 log p
2 2σ
2
1 − c22 log2 p
√ e 8σ ≤ √
√
≤
P
≤
>
p1− 8σ2 .
n
2
n
c log p 2π
c2 π log p
j=1 2
i=1
14
Z. Zheng, J. Lv and W. Lin
For the second term, similarly we have
√
K
X
p
2 2σK − c222
−1bT
ε| > c2 (log p)/n ≤ √
P |n fj e
p 8σ .
c
π
log
p
2
j=1
As K is no larger than p, the two bounds above give
P (Eec ) ≤
√
c2
2
4 2σ
√
p1− 8σ2 .
c2 π log p
By a similar argument, the bound on P (Ee0c ) can be derived as
√
c2
2
2 2σs
c
e
P (E0 ) ≤ √
n− 8σ2 .
c2 π log n
Thus, for the intersection event Ee ∩ Ee0 , we have
P {(Ee ∩ Ee0 )c } ≤ P (E˜c ) + P (Ee0c ) ≤
√
√
c2
c2
2 2σs
4 2σ
2
2
√
p1− 8σ2 + √
n− 8σ2 ,
c2 π log p
c2 π log n
√
which converges to zero as n → ∞ for c2 > 2 2σ. It completes the proof of Lemma 2.
| 10math.ST
|
NON-PERIPHERAL IDEAL DECOMPOSITIONS OF ALTERNATING
KNOTS
arXiv:1610.09901v1 [math.GT] 31 Oct 2016
STAVROS GAROUFALIDIS, IAIN MOFFATT, AND DYLAN P. THURSTON
Abstract. An ideal triangulation T of a hyperbolic 3-manifold M with one cusp is nonperipheral if no edge of T is homotopic to a curve in the boundary torus of M . For such a
triangulation, the gluing and completeness equations can be solved to recover the hyperbolic
structure of M . A planar projection of a knot gives four ideal cell decompositions of its
complement (minus 2 balls), two of which are ideal triangulations that use 4 (resp., 5) ideal
tetrahedra per crossing. Our main result is that these ideal triangulations are non-peripheral
for all planar, reduced, alternating projections of hyperbolic knots. Our proof uses the small
cancellation properties of the Dehn presentation of alternating knot groups, and an explicit
solution to their word and conjugacy problems. In particular, we describe a planar complex
that encodes all geodesic words that represent elements of the peripheral subgroup of an
alternating knot group. This gives a polynomial time algorithm for checking if an element
in an alternating knot group is peripheral. Our motivation for this work comes from the
Volume Conjecture for knots.
Contents
1. Introduction
1.1. Motivation: the Volume Conjecture
1.2. Non-peripheral ideal triangulations of alternating knots
1.3. Alternating knots and small cancellation theory
Acknowledgements
2. Small cancellation theory
2.1. The (augmented) Dehn presentation of a knot group
2.2. Square and grid presentations
2.3. The word problem for square presentations
2.4. The peripheral complex
2.5. Some properties of the peripheral complex
2.6. The peripheral word problem
2.7. Proof of Theorem 1.3
2.8. The peripheral complex and the Gauss code of an alternating knot
References
2
2
2
3
4
4
4
6
7
8
10
14
15
17
19
Date: November 1, 2016.
2010 Mathematics Subject Classification. Primary 57N10. Secondary 20F06, 57M25.
Key words and phrases. ideal triangulations, knots, hyperbolic geometry, ideal tetrahedra, small cancellation theory, Dehn presentation, alternating knots, Volume Conjecture.
S.G. and D.T. were supported in part by National Science Foundation Grants DMS-15-07244 and DMS14-06419 respectively.
1
2
STAVROS GAROUFALIDIS, IAIN MOFFATT, AND DYLAN P. THURSTON
1. Introduction
1.1. Motivation: the Volume Conjecture. The motivation of our paper comes from the
Kashaev’s Volume Conjecture for knots in 3-space, which states that for a hyperbolic knot
K in S 3 we have:
1
Vol(K)
lim log |hKiN | =
n→∞ n
2π
where hKiN is the Kashaev invariants of K; see [Kas97, MM01]. This gives a precise
connection between quantum topology and hyperbolic geometry. The Volume Conjecture
has been verified for only a handful hyperbolic knots: initially for the simplest hyperbolic 41
knot and now, due to the work of Ohtsuki [Oht17], and Ohtsuki and Yokota [OY16], for all
hyperbolic knots with at most 6 crossings.
The Volume Conjecture requires a common input for computing both the Kashaev invariant and the hyperbolic volume. Such an input turns out to be a planar projection of a
knot K which allows one to express the Kashaev invariant as a multi-dimensional state sum
whose summand is a ratio of quantum factorials (4 or 5, depending on the model used).
On the other hand, a planar projection gives four ideal cell decompositions of its complement (minus 2 balls), two of which are ideal triangulations that use 4 (resp., 5) ideal tetrahedra per crossing. These ideal triangulations are well-known from the early days of hyperbolic
geometry, and were used by Weeks [Wee05] (in his computer program SnapPy [CDW]), by
the third author [Thu99], Yokota [Yok02, Yok11], Sakuma-Yokota [SY] and others.
An approach to the Volume Conjecture initiated by the third author in [Thu99], and also by
Yokota, Kashaev, Hikami, the first author and others (see [Gar08, KY, Hik01, Yok02]), is to
convert multi-dimensional state-sum formulas for the Kashaev invariant to multi-dimensional
state-integral formulas over suitable cycles, and then to apply a steepest descent method to
study the asymptotic behaviour of the Kashaev invariant. The summand (and hence, the
integrand) depends on the planar projection and the steepest descend method is applied
to a leading term of the integrand, the so-called potential function. The critical points of
the potential function have a geometric meaning, namely they are solutions to the gluing
equations. The latter are a special system of polynomial equations (studied by W. Thurston
and Neumann-Zagier in [Thu77, NZ85]) that are associated to the ideal triangulations of the
knot complement discussed above. A suitable solution to the gluing equations recovers the
hyperbolic structure, and the value of the potential function is the volume of the knot.
The problem is that every planar projection leads to ideal triangulations, hence to gluing
equations, and even if we know that the knot is hyperbolic, it is by no means obvious that
those gluing equations have a suitable solution (or in fact, any solution) that recovers the
complete hyperbolic structure. It turns out that if a knot is hyperbolic, the lack of a suitable
solution occurs only when edges of the ideal triangulation are homotopic to peripheral curves
in the boundary tori.
1.2. Non-peripheral ideal triangulations of alternating knots. Ideal triangulations
of hyperbolic 3-manifolds with cusps were introduced by W. Thurston in his study of Geometrization of 3-manifolds; see [Thu77]. For thorough discussions, see [BP92, CDW, NZ85,
Wee05]. An ideal triangulation T of a hyperbolic 3-manifold M with one cusp is nonperipheral if no edge of T is homotopic to a curve in the boundary torus of M . For such
a triangulation, the gluing and completeness equations of [NZ85] can be solved to recover
NON-PERIPHERAL IDEAL DECOMPOSITIONS OF ALTERNATING KNOTS
3
the hyperbolic structure of M . For a proof, see [Til12, Lem.2.2] and also the discussion in
[DG12, Sec.3].
A planar projection ∆ of a knot gives rise to four ideal cell decompositions of its com◦
◦
plement (namely, T2B (∆), TO◦ (∆), T4T
(∆) and T5T
(∆)), the last two of which are ideal
triangulations that use 4 (resp., 5) ideal tetrahedra per crossing. We will briefly recall these
decompositions here, although their precise definition is not needed for the statement and
proof of Theorem 1.3 below.
• T2B (∆) is a decomposition of the knot complement into one ball above and one ball below
the planar projection. These two balls have a cell-decomposition that matches the planar
projection of the knot, and were originally studied by W. Thurston, and more recently by
Lackenby [Lac04].
• TO◦ (∆) is a decomposition of the knot complement minus two balls into ideal octahedra,
one at each crossing of ∆. This was described by Weeks [Wee05], and also by the third
author [Thu99], and by Yokota [Yok02, Yok11].
• Each ideal octahedron can be subdivided into 4 ideal tetrahedra, or into 5 ideal tetrahedra.
Thus, a subdivision of TO◦ (∆) gives rise to two ideal triangulations of the knot complement
◦
◦
minus two balls, denoted by T4T
(∆) and T5T
(∆).
Theorem 1.1. If ∆ is a prime, reduced, alternating projection of a non-torus knot K, then
◦
◦
the four ideal cell decompositions T2B (∆), TO◦ (∆), T4T
(∆) and T5T
(∆) are non-peripheral.
Consequently, the gluing equations have a solution that recovers the complete hyperbolic structure.
1.3. Alternating knots and small cancellation theory. The above theorem follows from
proving that all edges of the above ideal triangulations are homotopically non-peripheral.
Luckily, we can describe those edges directly in terms of the planar projection of the knot
as follows.
Definition 1.2. Let ∆ ⊂ R2 be a knot diagram with n crossings. Consider the projection
plane R2 as the xy-plane of R3 , and consider the knot K ⊂ S 3 = R3 ∪ {∞} obtained from ∆
by “pulling” the overcrossing arcs above the plane and undercrossing arcs under the plane in
the standard way. Fix a basepoint for π1 (S 3 \ K) in the unbounded region near one strand
of K. We distinguish four kinds of loops in π1 (S 3 \ K).
(1) A Wirtinger arc follows the double of ∆ through k crossings with 1 < k < 2n and
then returns to the basepoint through either the upper or lower half-space.
(2) A Wirtinger loop starts at the basepoint, travels in either the upper (resp. lower)
half-space to pass through a region R of ∆, passes through a region adjacent to R,
and then returns through the upper (resp. lower) half-space to the basepoint. We
forbid the short loop around the strand near the basepoint, which is manifestly a
meridian.
(3) A Dehn arc starts at the basepoint, travels in the upper (resp. lower) half-space
through a region of ∆ and then returns to the basepoint through the lower (resp.
upper) half-space without passing through the projection plane.
(4) A short arc follows the double of ∆ from the basepoint until some crossing, where
it jumps to the other strand in the crossing and then follows the double back to the
basepoint.
There four types of arc are illustrated in Figure 1.
These arcs are denoted by the letters A, B, C and D in [SY].
4
STAVROS GAROUFALIDIS, IAIN MOFFATT, AND DYLAN P. THURSTON
Wirtinger arc
Wirtinger loop
Dehn arc
Short arc
Figure 1. Four types of loop in a knot complement.
Theorem 1.3. If ∆ is a prime, reduced, alternating projection of a non-torus knot K, then
all Wirtinger arcs, Wirtinger loops, Dehn arcs and short arcs are non-peripheral.
Theorem 1.1 immediately follows from Theorem 1.3, since all of the arcs that appear in
any of the decompositions in Theorem 1.1 are of one of the four types in Theorem 1.3.
The proof of Theorem 1.3 uses the small cancellation property of the Dehn presentation of
hyperbolic alternating knots. Curiously, our proof uses an explicit solution to the conjugacy
problem of the Dehn presentation of a prime reduced alternating planar projection ∆. See
Remark 2.17 below.
Acknowledgements. A first draft of this paper was written in 2002 and was completed in
2007, but unfortunately remained unpublished. During a conference in Waseda University
in 2016 in honour of the 20th anniversary of the Volume Conjecture, an alternative proof of
the results of our paper (using cubical complexes) was announced by Sakuma-Yokota [SY],
and with the same motivation as ours. We thank Sakuma-Yokota for their encouragement to
publish our results, and the organisers of the Waseda conference (especially Jun Murakami)
for their hospitality.
2. Small cancellation theory
2.1. The (augmented) Dehn presentation of a knot group. We begin with a discussion
of the augmented Dehn presentation of a knot diagram. As it turns out, the augmented Dehn
presentation (defined below) is a small cancellation group and this structure provides a quick
NON-PERIPHERAL IDEAL DECOMPOSITIONS OF ALTERNATING KNOTS
5
and implementable solution to its word problem. Background on small cancellation groups
and combinatorial group theory can be found in [LS77].
Throughout this paper we implicitly symmetrize all group presentations. This means that
when we write a set of relators R, we actually mean the set of all relators which can be
obtained from R by inversion and cyclic permutation.
Let ∆ be a n crossing planar diagram of a link L. Of the n + 2 regions of the diagram ∆,
exactly n + 1 of these regions are bounded. Assign a unique label 1, 2, . . . , n + 1 to each of
these bounded region and the label 0 to the unbounded region. We identify each region with
its label.
We obtain a group presentation from the labelled diagram ∆ as follows. Take one generator
Xi for each region i = 0, 1, 2, . . . , n + 1 of ∆. Take one relator Ri for each of the n crossings
of ∆ which is read from the diagram thus
a
b
d
c
Xa Xb−1 Xc Xd−1
If we choose a base point above the projection plane, and we choose a point pi in the interior
of each region i. Then the generator Xi can be described geometrically by a loop in the knot
complement which passes from the base point, downwards through the region pi then back
up to the base point through the point p0 which lies in the unbounded region. Dehn showed
that
def
D∆ = h X0 , X1 , . . . , Xn+1 | R1 , R2 , . . . Rn , X0 i
is a presentation for the knot group π1 (S 3 \ L). We call this the Dehn presentation of
π1 (S 3 \ L) read from the diagram ∆. In what follows, we use a minor modification of the
Dehn presentation which has better small cancellation properties.
The augmented Dehn presentation, A∆ , of ∆ is the group presentation
def
A∆ = h X0 , X1 , . . . , Xn+1 | R1 , R2 , . . . Rn i .
The augmented Dehn presentation arises as a Dehn presentation of a link. Given a labelled
link diagram ∆, construct a new labelled link diagram ∆ ∪ O by adding a zero-crossing
component O, which bounds ∆. This is called the augmented link diagram. The augmented
Dehn presentation of ∆ is a presentation for the for the augmented link group π1 (S 3 \(K ∪O)),
i.e.,
(1)
A∆ ∼
= D∆∪O ∼
= π1 (S 3 \ K) ∗ Z .
We will solve the word problem in D∆ by solving it in A∆ . For completeness, let us say a
few words about why it is sufficient to solve the word problem in A∆ . This is a consequence of
some standard facts about group presentations that can be found in, for example, [LS77]. Let
PG = hg1 , . . . , gk | r1 , . . . rj i and PH = hh1 , . . . , hl | s1 , . . . sm i be presentations for groups G
and H respectively. Then the standard presentation, which we denote by PG ∗ PH , for the
free product G ∗ H is
PG ∗ PH = h g1 , . . . , gk , h1 , . . . , hl | r1 , . . . rj , s1 , . . . sm i .
A standard consequence of the normal form for free products (again see [LS77]) is that with
PG , PH and PG ∗ PH as above, if w is a word in the generators g1 , . . . , gk and their inverses,
6
STAVROS GAROUFALIDIS, IAIN MOFFATT, AND DYLAN P. THURSTON
then w =G 1 if and only if w =G∗H 1. Thus, by (1), the word problem in D∆ ∼
= π1 (S 3 \ K)
3
can be solved by the word problem in A∆ ∼
= π1 (S \ K) ∗ Z.
An an explicit isomorphism of the augmented Dehn presentation with a standard presentation for the free product π1 (S 3 \ K) ∗ Z is given by
(2)
φ : A∆ → D∆ ∗ hY | i
where
Y
if i = 0
.
Xi Y −1 otherwise
Geometrically, φ corresponds to isotoping the component O of the augmented link in S 3
away from the subdiagram ∆ so that it bounds a disc in the projection plane.
φ : Xi 7→
Remark 2.1. Let ι : D∆ → D∆ ∗ hY | i denote the natural inclusion. Given a projection
l of a loop ` ∈ π1 (S 3 \ K) in the diagram ∆, we can read off a representative φ−1 (ι(w))
as follows: follow the loop l from its basepoint in the direction of its orientation. When l
“passes downwards” through a region i of ∆ assign a generator Xi ; and whenever l “passes
upwards” through a region i of ∆ assign a generator Xi−1 . The word thus obtained clearly
represents the loop `. Thus, w 6=π1 (S 3 \K) 1 if and only if φ−1 (ι(w)) 6=A∆ 1.
2.2. Square and grid presentations. The augmented Dehn presentation of a prime, reduced, alternating knot diagram has small cancellation properties, as was first observed by
Weinbaum in [Wei71].
Let G = hX|Ri be a symmetrized group presentation. We call a non-empty word r a piece
with respect to R if there exist distinct words s, t ∈ R such that s = ru and t = rv.
Definition 2.2. (a) A symmetrized presentation hX|Ri is called a square presentation if it
satisfies the following two small cancellation conditions:
Condition C 00 (4). All relators have length four and no defining relator is a product of fewer
than four pieces.
Condition T (4) . Let r1 , r2 and r3 be any three defining relators such that no two of the
words are inverses to each other, then one of r1 r2 , r2 r3 or r3 r1 is freely reduced without
cancellation.
(b) A symmetrized presentation hX|Ri is called a grid presentation if it is a square presentation and in addition X is colored by two colors (black or white) and every relator alternates
in the two colors and in taking inverse.
Remark 2.3. There does not appear to be a standard terminology of the above definition. In
[Wei71], Weinbaum calls square presentations C 00 (4) − T (4) presentations. In [Joh97, Joh00],
Johnsgard uses the term parity to denote the black/white coloring of a grid presentation. In
[Wis06, Defn.3.1] and [Wis07, Defn.2.2], Wise uses the terms squared presentations and VH
presentations for our square presentations and grid presentations.
We may depict a relator r of a grid presentation by a Euclidean square as follows:
c
ab−1 cd−1
←→ d
b
a
It is easy to see that in a grid presentation the following holds:
NON-PERIPHERAL IDEAL DECOMPOSITIONS OF ALTERNATING KNOTS
7
• Relator squares have oriented edges, labelled from X. There are two sinks and two
sources in each relator square.
• We call a two letter subword of a relator a pair. The C 00 (4) condition says that a pair
uniquely determines a relator up to cyclic permutation and inversion.
• T (4) says that if ab and b−1 c are pairs then ac is not.
• If a, b and c are letters such that ab and b−1 c are both pairs (with b 6= c), then the
word ac is called a sister-set. By the T (4) condition, no pair is a sister-set.
• The edges of a relator square have an additional coloring: they are vertical or horizontal. Moreover, going around a relator square we alternate between black and
white.
• We can invoke a convention that the black and white colorings correspond to horizontal and vertical line placement in our drawings of relator squares.
• A rotation or reflection of a relator square corresponds to the cyclic permutation or
inversion of a relator.
We can now state Weinbaum’s theorem.
Theorem 2.4. [Wei71] The augmented Dehn presentation of a prime, reduced, alternating
knot diagram is a grid presentation.
In [LS77] Lyndon and Schupp show that square and grid presentations have have solvable
word and conjugacy problems. Since the appearance of that work, polynomial time algorithms have been given for the word (see [Joh97, Sec.7])) and conjugacy problems ([Joh97])
of these groups. We use these more efficient algorithms here.
2.3. The word problem for square presentations. In this section we recall the solution
to the word problem of square presentations. To any group presentation G = hX|Ri we can
associate a standard 2-complex K in the usual way: K consists of one 0-cell, one labelled
1-cell for each generator and one 2-cell for each relator, where the 2-cell Dr representing
the relator r ∈ R is attached to the 1-skeleton, K (1) , by a continuous map which identifies
the boundary ∂Dr with a loop representing r in the 1-skeleton. We impose a piece-wise
Euclidean structure on the standard 2-complex and set all 1-cells to be of unit length.
A word w represents the identity in G if and only if there is a simply connected planar
2-complex ∆, and a map φ : (D, ∂D) → (K, K (1) ) such that the 0-cells are mapped to
0-cells, open i-cells are mapped to open i-cells, for i = 1, 2 and ∂D is mapped to the loop
representing w in K (1) . Such a 2-complex, labelled in the natural way, is called a Dehn
diagram.
Throughout this text we use two concepts of labels of edge-paths of the standard 2complex, peripheral complex (introduced below) or Dehn diagram. The label of an edge-path
is the sequence of letters determined by the edge-path, where travelling along an edge labelled
a contributes the letter a. This is distinct from the word labelling an edge-path, which is the
word in the group determined by the path, where travelling along an edge labelled a against
the orientation contributes the letter a−1 , and travelling with the orientation, the letter a.
A word in a group presentation is said to be geodesic if it contains the least number of letters
over all representatives of the same word, i.e., w is geodesic if |w| = min{|w0 | | w =G w0 }.
A geodesic word represents the identity if and only if it is the empty word. A word in a
group presentation is geodesic if and only if it labels a geodesic edge-path in the standard
two complex of the presentation.
8
STAVROS GAROUFALIDIS, IAIN MOFFATT, AND DYLAN P. THURSTON
A key result of small cancellation theory is the following Geodesic Characterisation Theorem; see [Joh97, Sec.3] and also [Kap97, Lem.3.2].
Theorem 2.5. A word in a square presentation is geodesic if and only if it is freely reduced
and contains no subword x1 . . . xn which is part of a chain:
x2
xn−1
x3
xn
x1
The word x1 . . . xn is called a chain word.
Remark 2.6. Observe that the Geodesic Characterisation Theorem immediately provides a
quadratic time solution for the word problem in square presentations: Given a word w, freely
reduce it to obtain a word w0 . If w0 is the empty word then w =G 1, otherwise search w0 for
a chain word. If w0 does not contain a chain word then w0 6=G 1. If w0 does contain a chain
word, replace it with the shorter word which bounds the “other side” of the chain to obtain
a shorter word w00 . Repeat the above process with the word w00 in place of w.
2.4. The peripheral complex. A reduced, prime, alternating, oriented knot diagram gives
rise to a grid presentation with solvable word problem; see Theorems 2.4 and 2.5. This
grid presentation contains a peripheral Z2 -subgroup generated by the meridian m and the
longitude l of the knot. Of course, a peripheral subgroup does not exist for a general grid
presentation.
Theorem 1.3 requires us to solve the peripheral word problem. Following Johnsgard (see
[Joh97, Sec.7]), we consider the (rather overlooked) peripheral complex, and we discuss how
it solves the peripheral word and conjugacy problem.
Let ∆ be a reduced, prime, alternating, oriented knot diagram with n crossings, and
let A∆ be its augmented Dehn presentation. Each relator of A∆ is a word of length four
whose exponents alternate in sign. We may think of the relators as 1 × 1 Euclidean squares
with directed and labelled edges. For convenience, we impose some conventions upon our
construction. We discuss the effect of these conventions in Remark 2.9 below.
From the base point of ∆ and in the direction of the orientation, walk around the diagram
and label the n crossings of ∆ with c1 , c2 , . . . , c2n in the order we meet them and in such a
way that the label c1 is assigned to the first under crossing we meet. For example, for the
52 knot we have:
0
c5 c10
1
c4
5
c7
c6
4
c1
6
2
c3
c9
c2
3
c8
NON-PERIPHERAL IDEAL DECOMPOSITIONS OF ALTERNATING KNOTS
9
We construct a 2n × 1 rectangle made out of 2n relator squares inductively as follows.
Position the relator square C1 on the Euclidean plane in such a way that the label of the edgepath from (0, 0) to (1, 1) describes a loop which follows the knot through the undercrossing
at c1 (on the left is shown the crossing c1 and on the right is shown the relator square C1 ):
c
a
b
b
d
d
b
a
or
c
c
a
d
Suppose we have placed a relator square Ck (which arises from the crossing ck ). The
relator squares Ck and Ck+1 have exactly two edge-labels in common (since the diagram ∆
is prime and reduced). Identify the right edge of Ck with the unique edge of Ck+1 which has
the same label in a way that preserves the orientation of the edges. This gives a (k + 1) × 1
rectangle. Continue this process until we have added the relator square C2n .
We call such a 2n × 1 rectangle of relator squares a fundamental block of ∆. For example,
the fundamental block of the 52 knot above is
6
5
0
2
1
6
3
6
0
4
0
1
5
6
6
2
0
0
5
1
6
4
6
0
3
0
1
2
6
5
0
Observe that the fundamental block has oriented edges and its vertices are either sinks or
sources. We will often simplify figures by drawing sinks as thickened black vertices. This
determines the orientations of the edges.
Notice in the example above that the word labelling the top edge of the fundamental block
is a cyclic permutation of the word labelling the bottom edge of the fundamental block, and
that the labels and orientations on the left and right edges coincide. This observation holds
in general and it allows us to piece together the fundamental blocks in a way that tiles the
plane.
Lemma 2.7. In the fundamental block of a reduced, prime, alternating, oriented knot diagram ∆,
(1) the label and orientation of the rightmost and leftmost vertical edges of the fundamental block coincide;
(2) the label and orientation on the top of the relator square Ci is the same as the label
and orientation on the bottom of the relator square Ci+1 , where the indices are taken
modulo 2n.
We defer the proof of this lemma until the end of Section 2.5.
Using Lemma 2.12 we can piece together together the fundamental blocks according to
the following pattern,
and tile the whole plane by relator squares.
10
STAVROS GAROUFALIDIS, IAIN MOFFATT, AND DYLAN P. THURSTON
Definition 2.8. We call the resulting 2-dimensional CW complex the peripheral complex.
For example, a portion of the peripheral complex for the 52 knot is given by:
2
3
4
5
2
5
4
3
2
5
1
0
6
0
6
1
0
6
0
6
1
5
2
3
4
5
2
5
4
3
2
6
1
0
6
0
6
1
0
6
0
6
2
5
2
3
4
5
2
5
4
3
0
6
1
0
6
0
6
1
0
6
0
3
2
5
2
3
4
5
2
5
4
6
0
6
1
0
6
0
6
1
0
6
4
3
2
5
2
3
4
5
2
5
0
6
0
6
1
0
6
0
6
1
0
5
4
3
2
5
2
3
4
5
2
1
0
6
0
6
1
0
6
0
6
1
2
5
4
3
2
5
2
3
4
5
6
1
0
6
0
6
1
0
6
0
6
5
2
5
4
3
2
5
2
3
4
0
6
1
0
6
0
6
1
0
6
0
4
5
2
5
4
3
2
5
2
3
6
0
6
1
0
6
0
6
1
0
6
3
4
5
2
5
4
3
2
5
2
0
6
0
6
1
0
6
0
6
1
2
3
4
5
2
5
4
3
2
5
0
Remark 2.9. Several choices and conventions were made in the construction of the peripheral
complex. Namely the choice of base point on ∆, the label c1 was assigned to the first under
crossing we met, and the positioning of the first relator square C1 . It is clear from the
construction of the complex that a different choice of base point (as well as orientation) and
a different placement of C1 on the Euclidean plane would result in a peripheral complex
which is isometric to the one constructed here. We discuss this in more detail in Section 2.8.
All of the arguments presented here can be made with any construction of the peripheral
complex, however the directions specified in the statements of results and proofs in this paper
may change.
2.5. Some properties of the peripheral complex. By construction, the peripheral complex embeds in the standard 2-complex of the augmented Dehn presentation. In fact it
embeds geodesically:
Lemma 2.10. The peripheral complex of ∆ embeds geodesically in the standard 2-complex
of the augmented Dehn presentation A∆ . In particular, the word labelling any geodesic edgepath in the peripheral complex is a geodesic word in the augmented Dehn presentation.
Proof. The proof uses the Geodesic Characterisation Theorem (Theorem 2.5). Since any two
paths in the peripheral complex with common beginning and ending represent the same word
in A∆ , it suffices to show that a path p that goes horizontally q steps and then vertically r
steps in the peripheral complex is geodesic in the standard 2-complex.
Consider an edge-path p in the peripheral complex that goes horizontally q steps and then
vertically r steps. Such a path has at most one pair subword, since a sister-set is never a
pair by the T (4) condition. Thus we see that the label of the edge-path cannot contain a
chain word, as this requires two pairs. (Recall the definitions of pairs and sister sets from
Section 2.2.)
It remains to show that the label of the edge-path is freely reduced. To see why this is
we begin by observing that since ∆ is reduced, four distinct regions of ∆ meet at every
NON-PERIPHERAL IDEAL DECOMPOSITIONS OF ALTERNATING KNOTS
11
crossing and therefore every relator square has four distinct labels. Now suppose that ab is a
subword of the word labelling the edge-path p. If the subword belongs to the horizontal path,
then it labels the bottom of two relator squares Di and Di+1 in the peripheral complex. By
Lemma 2.7, the bottom label of Di+1 is also a label of the top of the relator square Di . This
means that b cannot label the bottom of Di and therefore a 6= b and ab is freely reduced.
If the letter a comes from a horizontal edge and b from a vertical edge of p, then the
subword ab labels two sides of a relator square and is therefore freely reduced.
Finally, If the subword belongs to the vertical path, then it labels the right hand side
of two relator squares Di and Di+1 in the peripheral complex. By the periodicity of the
peripheral complex, the right hand label of Di is also the label of the left hand side of the
relator square Di+1 . This means that b cannot label the right of Di+1 and therefore a 6= b
and ab is freely reduced.
Proof of Lemma 2.7. Since the exponents of the relators of the augmented Dehn presentation
alternate in sign, all of the orientations of the edges of the fundamental block are of the form
required by the lemma.
It remains to show that the edge labels are of the required form. First we show that the
label on the top of the relator square Ci is the same as the label on the bottom of the relator
square Ci+1 for i = 1, . . . , n.
Consider the relator square C1 positioned as
b
d
C1
a
c
By convention the labels a and b also appear in C2 (since the regions a and b of ∆ are
incident with the crossings c1 and c2 ). Therefore C2 has one of the following forms
b
a
C2
or
a
C2
b
These two relators have edge-paths b−1 a and ab−1 respectively. By the small cancellation
conditions, a pair uniquely determines a relator, so the word ab−1 cannot appear in C2 (as
it appears in C1 as (ab−1 )−1 ). Therefore b must be the label on the bottom of C2 .
We proceed inductively. Suppose that we have shown that the label on the top of Ck−1
coincides with the label on the bottom of Ck . Since the crossing ck shares two incident
regions with ck−1 and two with ck+1 , the relator square Ck share two labels with Ck−1 and
two with Ck+1 . By hypothesis, Ck shares the labels on the bottom and left-hand edges with
Ck−1 , so it shares the labels on the top and right-hand edges with Ck+1 .
Suppose the word on the edge-path which follows the right-hand and then top edge of Ck
is rs−1 or r−1 s. We will deal with each case separately.
If the path is rs−1 . Then Ck+1 also has edges labelled r and s and must contain the word
−1
sr or r−1 s. Since a pair determines a relator and Ck 6= Ck+1 , we have that Ck+1 must
contain the word r−1 s (since sr−1 = (rs−1 )−1 ). The only way this can happen is if the letter
s is on the bottom of Ck+1 .
12
STAVROS GAROUFALIDIS, IAIN MOFFATT, AND DYLAN P. THURSTON
Similarly, if the path is r−1 s. Then Ck+1 also has edges labelled r and s and must contain
the word s−1 r or rs−1 . Since a pair determines a relator and Ck 6= Ck+1 , we have that Ck+1
must contain the word rs−1 . The only way this can happen is if the letter s−1 is on the
bottom of Ck+1 .
We have shown that the label on the top of the relator square Ci is the same as the label
on the bottom of the relator square Ci+1 for i = 1, . . . , n.
To complete the proof, consider the relator square C2n in the fundamental block. C2n is
of the form
s
p
r
C2n
q
where the labels p and q are shared with C2n−1 and r and s are shared with C1 (since c1 and
c2n share incident regions of ∆). But again, the small cancellation conditions say that a pair
uniquely determines a relator and C2n 6= C1 , therefore we must have s = c and r = d, where
c and d are the labels of C1 as shown above. This completes the proof of the lemma.
Lemma 2.11. In the fundamental block of a reduced, prime, alternating, oriented knot
diagram ∆,
(1) the label of an edge-path from the bottom-right to top-left corner of C2n describes a
curve homotopic to a meridional loop, or its inverse, of the knot through the base
point of ∆;
(2) the label of an edge-path from the bottom left to top right corner of C2i−1 , i = 1, . . . , n
describes a loop which follows the under-crossing of the knot at c2i−1 ;
Proof. The relator square C2n comes from a crossing of the form
a
b
d
c
and, since b and c are also labels of C1 , it must appear in the fundamental block in one of
the following forms
c
b
a
c
d
or
d
b
a
In either case we see that the edge-path bottom-right to top-left corner describes a meridian
or its inverse. This proves the first statement of the lemma.
We now prove the second statement. The relator square C2i−1 appears in the fundamental
block with orientation
NON-PERIPHERAL IDEAL DECOMPOSITIONS OF ALTERNATING KNOTS
13
and at c2i−1 we travel along an undercrossing of the form
a
b
d
c
which contributes the relator ab−1 cd−1 . Therefore C2i−1 is of one of the four forms
c
d
a
b
a
b
d
d
c
c
b
a
b
a
c
d
(these are all possible ways that the relator can fit the orientation of C2i−1 ). But, by the
construction of the fundamental block, a or d must label the vertical left edge of C2i−1 . This
eliminates two of the four possible labellings of C2i−1 above, and it is easily seen that in the
remaining two possibilities, the label of an edge-path from the bottom left to top right corner
describes a loop which follows the under-crossing of the knot at c2i−1 , as required.
Let n denote the number of crossings of ∆. Further, let λ denote the double of the diagram
∆ determined by the blackboard framing, and based at a point x0 . Let µ be the meridian
of ∆ based at x0 . The orientations of µ and λ are determined by the orientation of ∆. A
peripheral element of the knot group π1 (S 3 −K) is then a product λa µb , a, b ∈ Z. The curves
λ and µ in ∆ also determine canonical elements φ−1 (ι(λ)) and φ−1 (ι(µ)) of the augmented
knot group G = π1 (S 3 − (K ∪ O)). We abuse notation and also denote these elements by λ
and µ respectively. We say that an element of the (augmented) knot group is peripheral if it
represents the element λa µb for some a, b ∈ Z.
Lemma 2.12. Let w be a word which labels an edge-path from the point (0, 0) to the point
(an − b, an + b) in the peripheral complex of ∆, where a, b ∈ Z and n is the number of
crossings of ∆. Then w is peripheral and represents the element λa µb . Conversely, every
peripheral element λa µb has a representative as the label of an edge-path from the point (0, 0)
to the point (an − b, an + b) in the peripheral complex.
Proof. We show that there exists one word la mb labelling the edge-path from (0, 0) to (an −
b, an + b) in the peripheral complex which represents λa µb , for each choice of a and b. Since
the peripheral complex complex embeds in the standard 2-complex of the augmented Dehn
presentation, it follows that any word which labels a edge-path from the point (0, 0) to the
point (an − b, an + b) represents the peripheral element la mb .
Label the crossings of ∆ by c1 , . . . , c2n according to the conventions in Subsection 2.4. We
can find a representative l of λ in the augmented Dehn presentation A∆ as follows: take a
framed double λ of ∆. Begin by taking l to be the empty word. Walk once around λ and
14
STAVROS GAROUFALIDIS, IAIN MOFFATT, AND DYLAN P. THURSTON
concatenate a subword Xa Xb−1 to the right of l whenever we pass under an arc of ∆ from a
region labelled a to a region labelled b. The word l obtained clearly represents λ.
Since the knot is alternating and, by our convention on the labelling of the crossing,
the double λ of ∆ passes under an arc of ∆ at the crossings c2i−1 , for i = 1, . . . , n. By
Lemma 2.11, the two letter subword contributed to l at the crossing c2i−1 is exactly the label
of an edge-path from the bottom left to top right corner of a relator square C2i−1 in the
peripheral complex. Therefore l can be described as an edge-path from the bottom left to
the top right of the following complex:
C2n−1
C3
C1
Clearly, such a complex is embedded in the peripheral complex and the edge-path is a path
from (0, 0) to (n, n). By the periodicity of the peripheral complex, it follows that the word
la is an edge-path in the peripheral complex from (0, 0) to (an, an) for all a ∈ Z.
We have shown that powers of the longitude are contained in the peripheral complex.
We now show that powers of meridians and peripheral elements are also contained in the
peripheral complex.
By Lemma 2.11 the label m of an edge-path from (0, 0) to (−1, 1) represents the meridian
µ or its inverse µ−1 . By the periodicity of the peripheral complex the label mb of an edge
path from (0, 0) to (−b, b), b ∈ Z, represents a power of the meridian and, again by the
periodicity of the peripheral complex (an, an) to (an − b, an + b), b ∈ Z represents a power
of the meridian for each a, b ∈ Z.
Therefore the label of an edge-path from the point (0, 0) to the point (an − b, an + b) in
the peripheral complex is la mb and is peripheral.
2.6. The peripheral word problem. The aim of this section is to solve the peripheral
word problem using the peripheral complex.
Theorem 2.13. Let w be a geodesic word in the augmented Dehn presentation of the n
crossing diagram ∆. Then w is peripheral and represents the element λa µb if and only if it
labels a geodesic edge-path from (0, 0) to (an − b, an + b) in the peripheral complex of ∆ for
some a, b ∈ Z.
To prove the theorem we need the following result from [Joh00] and [Kra94].
Theorem 2.14. Let w be a geodesic word in a square presentation of a group G all of whose
relators are of length four. Then w uniquely determines a tiling of relator squares bounded
by (but not necessarily filling) a rectangle in the Euclidean plane such that:
(1) the tiling embeds in the standard 2-complex of the group, i.e. it is a Dehn diagram;
NON-PERIPHERAL IDEAL DECOMPOSITIONS OF ALTERNATING KNOTS
15
(2) the word labels a geodesic edge-path from one corner of the rectangle to the opposite
corner; and
(3) if w0 is a geodesic word then w0 =G w if and only if w0 labels a geodesic edge-path
from one corner of the rectangle to the opposite corner path homotopic to w.
The tiling produced by the theorem for a geodesic word w is called the geodesic completion
of w.
Proof of Theorem 2.13. Let Rab be the rectangle in the peripheral complex determined by
the points (0, 0) and (an − b, an + b) for some integers a and b, and let wab be the label
of any geodesic edge-path between these two points (for example the edge-path from (0, 0)
to (an − b, 0) to (an − b, an + b) will do). Then since the words labelling geodesic edgepaths in the peripheral complex are geodesic words in the augmented Dehn presentation (by
Lemma 2.10), wab is a geodesic word.
Therefore, wab is a geodesic word in a grid presentation which labels a geodesic edge-path
between two opposite corners of the rectangle Rab . By Theorem 2.14, a geodesic word in
the augmented Dehn presentation represents the word wab if and only if it is the label of
a geodesic edge-path between (0, 0) and (an − b, an + b). So all geodesic representatives of
wab are words labelling geodesic edge-paths from (0, 0) and (an − b, an + b) in the peripheral
complex. Finally, by Lemma 2.12, every peripheral element is presented by a word wab for
some a, b ∈ Z and the result follows.
2.7. Proof of Theorem 1.3. Let ∆ be a prime, reduced alternating projection of a knot
K in S 3 . Theorem 1.3 follows from the following lemma.
Lemma 2.15. (a) The Wirtinger arcs, Wirtinger loops, and Dehn arcs and conjugates of
the short arcs of ∆ have explicit geodesic representatives in the peripheral complex.
(b) The above geodesic representatives of Wirtinger arcs, Wirtinger loops and Dehn arcs are
non-peripheral, and the above geodesic representatives of the short arcs are not conjugate
to a peripheral element.
Proof. We apply the notation and discussion from the last three paragraphs of Subsection 2.1.
Let w be a word representing a Wirtinger arc, Wirtinger loop, short arc or Dehn arc of
∆. Recall the map φ from Equation (2) and the method for reading the a representative
word φ−1 (ι(w)) in the augmented Dehn presentation described in Remark 2.1. We use the
peripheral complex to show that its image φ−1 (ι(w)) is non-peripheral.
We deal with each type of loop separately. Throughout we let l = l1 l2 · · · l2n be a geodesic
representative of λ which was constructed in the proof of Lemma 2.12. It is given by an edgepath following the sequence of relator squares C1 , C2 , . . . , C2n−1 . Each two letter subword
l2i−1 l2i of l labels an edge path on C2i−1 . Also let m = m1 m2 be a geodesic representative of
µ in the augmented Dehn presentation. By Theorem 2.13, l and m are labels of edge paths
from (0, 0) to (n, n) and (0, 0) to (−1, 1) respectively, in the peripheral complex.
Wirtinger arcs: Wirtinger arcs are loops which follow the double λ of ∆ returning to
the base-point after passing through fewer than 2n crossings. Therefore a Wirtinger loop is
represented by a subword l1 l2 · · · l2p , wher 2p < 2n, of l. Moreover, l1 l2 · · · l2p is represented
by the label of any edge-path from (0, 0) to (p, p) for p < n. By Theorem 2.13, it follows that
l1 l2 · · · l2p is non-peripheral as it does not label a path between (0, 0) and (an − b, an + b).
Wirtinger loops: We may move a Wirtinger loop close to some crossing ci . By the way that
the relators of the augmented Dehn presentation are read from ∆ (see Subsection 2.1), we
16
STAVROS GAROUFALIDIS, IAIN MOFFATT, AND DYLAN P. THURSTON
see that the Wirtinger loop can be described by a geodesic edge-path between two opposite
corners of Di (which two opposite corners depends on the given Wirtinger loop). Let the
label of this edge-path be w.
The word w is geodesic of length two. Therefore, if w is peripheral it must represent µ±1 .
However, by Lemma 2.11, the only geodesic words which represent µ±1 arise as a path from
(0, 0) to (∓1, ±1) in the peripheral complex, so w cannot be the label of such an edge-path
since by the definition of Wirtinger loops, w is not a representative of the meridian, which
is described by a path from (0, 0) and (−1, 1).
Dehn arcs: Suppose that a given Dehn arc intersects the bounded region a of ∆. Then
it is represented by Xa X0−1 in the augmented Dehn presentation. The word Xa X0−1 is
geodesic since it is freely reduced (a 6= 0) and clearly does not contain a chain subword (see
Theorem 2.5).
The meridian µ has exactly two geodesic representatives which label the edge-path (0, 0)
to (−1, 1) of C2n . It is easily seen from the definition of Dehn arcs that neither of these
words can be Xa X0−1 .
Short arcs: Short arcs are found by walking around the double λ of ∆, and at some point,
jumping to an adjacent arc of λ and walking back to the base point in one of two ways.
Short arcs are then represented by words of the form
(3)
l1 l2 · · · lk lp · · · l2n
and
(4)
l1 l2 · · · lk lp · · · l1 .
These representatives of short arcs do not necessarily embed as edge-paths in the peripheral
complex. However, since l2 embeds in the complex as a path from (−n, −n) to (n, n), the
conjugates
(5)
lp · · · l2n l1 l2 · · · lk
and
(6)
lp · · · l1 l1 l2 · · · lk
both embed in the peripheral complex as paths from (−p, −p) to (k, k), for (5), and (q, q) to
(r, r) or (r, r) to (q, q), for (6), where q = min{k, p} and r = max{k, p}. These paths travel
along the South-West to North-East axis.
Since these words in (3) and (4) are of length less that 2n, if they are peripheral, then
they must be equal to a power of the meridian mb . This will happen if and only if (3) and
(4) represent
(7)
(l1 · · · lk )−1 mb (l1 · · · lk ),
for some integer b. This conjugate of mb embeds into the peripheral complex as a path
from (k, k) to (0, 0) to (−b, b) to (k − b, k + b) which has a geodesic representative as a path
from (k, k) to (k − b, k + b). But this path travels along the South-East to North-West axis.
Therefore, by Theorem 2.14, the geodesic words in (3) and (4) cannot be equal to the words
of the form in (7) and the short arcs are non-peripheral.
NON-PERIPHERAL IDEAL DECOMPOSITIONS OF ALTERNATING KNOTS
17
Remark 2.16. The argument above showing that short arcs are non-peripheral in fact proves
a stronger result. It shows that any loop in the knot complement that follows the double of ∆
from the basepoint, then at some point jumps (above the projection plane) to any other point
on the double, then follows it back to the basepoint in either direction is non-peripheral.
Remark 2.17. Notice that in the proof for the non-peripherality of short arcs we actually
solved the conjugacy problem. We could also have shown that these elements were nonperipheral by using Johnsgard’s solution to the conjugacy problem [Joh97]: using Johnsgard’s
algorithm, the fact that the peripheral complex contains the geodesic completion of lm , and
the periodicity of the peripheral complex, it is straight-forward to show that a geodesic word
in the augmented Dehn presentation is conjugate to a peripheral element la mb if and only if
it embeds as a geodesic path from (0, k) to (an + k − b, an + b), for some integer k. It is easy
to see that two words in (5) and (6) are not of this form. We can use a similar argument for
Wirtinger loops.
Also note that this characterisation of conjugates of peripheral elements as paths in the
peripheral complex provides a method for solving the peripheral conjugacy problem.
2.8. The peripheral complex and the Gauss code of an alternating knot. In our
proof of Theorem 1.3, the peripheral complex plays a key role, and encodes the peripheral
structure of a prime, reduced, alternating projection of a knot. In this section we discuss
additional properties of the peripheral complex and its relation with the Gauss code.
In Subsection 2.4, we constructed the peripheral complex by placing relator squares of
the augmented Dehn presentation on the plane in a way determined by the oriented knot
diagram. As previously noted, some conventions were used in this construction. There is a
way to construct the peripheral complex directly from the relators of the augmented Dehn
presentation without any reference to the knot diagram:
(1) Choose any relator square from the augmented Dehn presentation of a prime, reduced,
alternating knot diagram, and place it in the Euclidean plane.
(2) Choose two diagonally opposite vertices of this relator square, call them a and b.
Form a “diagonal line” of relator squares by placing copies of the relator square in
such a way that each vertex a is identified with a vertex b and all of the relator
squares are translations of the first.
(3) Complete the tiling by adding relator squares from the augmented Dehn presentation
in a way consistent with the words labelling edge-paths. (Proposition 2.18 tells us
that this can be done in a unique way.)
The construction is indicated in Figure 2.8. Throughout this section we call this the unoriented construction of the peripheral complex, and we refer to the complex constructed in
Subsection 2.4 as the oriented construction. We will also refer to the resulting complexes as
the unoriented and oriented peripheral complexes respectively.
The following proposition tells us that the complex just described exists and is the peripheral complex.
Proposition 2.18. The unoriented construction of the peripheral complex described above,
produces a unique plane tiling of relator squares. Moreover, the resulting complex is isometric
to the oriented peripheral complex constructed in Subsection 2.4.
Proof. First of all, we note that if the complex exists, then it must be unique, since the corners
between pairs of relator squares in the “diagonals” used in the construction are labelled by
pairs and, in a grid presentation, a pair uniquely determines a relator.
18
STAVROS GAROUFALIDIS, IAIN MOFFATT, AND DYLAN P. THURSTON
1
1
1
1
1
1
1
2
2
2
1
1
2
1
2
3
2
3
3
To show existence, let T be the relator square in the plane from the first step of the
unoriented construction above. Suppose also that T has the vertices a and b specified.
Since every relator square and the reflection of every relator square of the augmented Dehn
presentation appears in the fundamental block of the oriented peripheral complex, there is
an isometry taking T to a relator square of the peripheral complex which sends the vertices
a and b to the top-left and bottom-right vertices of that relator square. By uniqueness, this
extends to an isometry of the complexes.
The following proposition tells us that an unoriented knot can be recovered from its
unoriented peripheral complex, and an oriented knot from its oriented peripheral complex.
Proposition 2.19. Let ∆ be a prime, reduced, alternating, oriented knot diagram. The
Gauss code of D can be recovered from the oriented peripheral complex; and the Gauss code
of ∆ or its inverse −∆ can be recovered from the unoriented peripheral complex.
Proof. We first prove the result for the oriented complex. Choose any 2n×1 horizontal block
of the complex. Every relator square appears exactly twice in this block. By the construction
of the complex, this block is a cyclic permutation of a fundamental block, and therefore the
order of the relator squares in the block is precisely the order we meet the crossings as
we travel around the knot in the direction of the orientation from some base point. With
this observation, it is straight-forward to recover the Gauss code: label the relator squares
S1 , S2 , . . . , S2n by reading along the strip from left to right. Assign the number −1 to S1 if
it has orientation
otherwise assign the number +1 to S1 . Suppose you have assigned the number ±j to the
relator square Si . If the relator square Si+1 has not been encountered previously assign the
number ∓(j + 1) to it, if the relator square has been encountered previously and has been
assigned the number ±p, then assign the number ∓p to this square. The resulting sequence
is the Gauss code.
To recover a Gauss code from an unoriented peripheral complex, we can use the same
method. However, since the 2n × 1 horizontal strip of the unoriented complex can be a
reflection of a 2n × 1 horizontal strip of the oriented complex, we are unable to determine if
the Gauss code obtained is that of the knot diagram or its inverse.
NON-PERIPHERAL IDEAL DECOMPOSITIONS OF ALTERNATING KNOTS
19
References
[BP92] Riccardo Benedetti and Carlo Petronio, Lectures on hyperbolic geometry, Universitext, SpringerVerlag, Berlin, 1992.
[CDW] Marc Culler, Nathan Dunfield, and Jeffery Weeks, SnapPy, a computer program for studying the
geometry and topology of 3-manifolds.
[DG12] Nathan Dunfield and Stavros Garoufalidis, Incompressibility criteria for spun-normal surfaces,
Trans. Amer. Math. Soc. 364 (2012), no. 11, 6109–6137.
[Gar08] Stavros Garoufalidis, Chern-Simons theory, analytic continuation and arithmetic, Acta Math. Vietnam. 33 (2008), no. 3, 335–362.
[Hik01] Kazuhiro Hikami, Hyperbolic structure arising from a knot invariant, Internat. J. Modern Phys. A
16 (2001), no. 19, 3309–3333.
[Joh97] Karin Johnsgard, The conjugacy problem for groups of alternating prime tame links is polynomialtime, Trans. Amer. Math. Soc. 349 (1997), no. 3, 857–901.
, Geodesic edge paths in nonpositively curved complexes, J. Pure Appl. Algebra 147 (2000),
[Joh00]
no. 1, 59–93.
[Kap97] Ilya Kapovich, Small cancellation groups and translation numbers, Trans. Amer. Math. Soc. 349
(1997), no. 5, 1851–1875.
[Kas97] Rinat Kashaev, The hyperbolic volume of knots from the quantum dilogarithm, Lett. Math. Phys.
39 (1997), no. 3, 269–275.
[Kra94] Mile Krajcevski, Tilings of the plane and hyperbolic groups, ProQuest LLC, Ann Arbor, MI, 1994,
Thesis (Ph.D.)–State University of New York at Binghamton.
[KY]
Rinat Kashaev and Yoshiyuki Yokota, in preparation.
[Lac04] Marc Lackenby, The volume of hyperbolic alternating link complements, Proc. London Math. Soc.
(3) 88 (2004), no. 1, 204–224, With an appendix by Ian Agol and Dylan Thurston.
[LS77] Roger Lyndon and Paul Schupp, Combinatorial group theory, Springer-Verlag, Berlin, 1977, Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 89.
[MM01] Hitoshi Murakami and Jun Murakami, The colored Jones polynomials and the simplicial volume of
a knot, Acta Math. 186 (2001), no. 1, 85–104.
[NZ85] Walter Neumann and Don Zagier, Volumes of hyperbolic three-manifolds, Topology 24 (1985), no. 3,
307–332.
[Oht17] Tomotada Ohtsuki, On the asymptotic expansion of the Kashaev invariant of the 52 -knot, Quantum
Topol. (2017).
[OY16] Tomotada Ohtsuki and Yoshiyuki Yokota, On the asymptotic expansions of the Kashaev invariant
of the knots with 6 crossings, 2016, Preprint.
[SY]
Makoto Sakuma and Yoshiyuki Yokota, An application of non-positively curved cubings of alternating links, In preparation.
[Thu77] William Thurston, The geometry and topology of 3-manifolds, Universitext, Springer-Verlag, Berlin,
1977, Lecture notes, Princeton.
[Thu99] Dylan Thurston, Hyperbolic volume and the Jones polynomial, Notes from lectures at the Grenoble
summer school “Invariants des nœuds et de variétés de dimension 3”, June 1999, available from
http://pages.iu.edu/~dpthurst/speaking/Grenoble.pdf.
[Til12] Stephan Tillmann, Degenerations of ideal hyperbolic triangulations, Math. Z. 272 (2012), no. 3-4,
793–823.
[Wee05] Jeff Weeks, Computation of hyperbolic structures in knot theory, Handbook of knot theory, Elsevier
B. V., Amsterdam, 2005, pp. 461–480.
[Wei71] Carl Weinbaum, The word and conjugacy problems for the knot group of any tame, prime, alternating knot, Proc. Amer. Math. Soc. 30 (1971), 22–26.
[Wis06] Daniel Wise, Subgroup separability of the figure 8 knot group, Topology 45 (2006), no. 3, 421–463.
[Wis07]
, Complete square complexes, Comment. Math. Helv. 82 (2007), no. 4, 683–724.
[Yok02] Yoshiyuki Yokota, On the potential functions for the hyperbolic structures of a knot complement,
Invariants of knots and 3-manifolds (Kyoto, 2001), Geom. Topol. Monogr., vol. 4, Geom. Topol.
Publ., Coventry, 2002, pp. 303–311 (electronic).
20
[Yok11]
STAVROS GAROUFALIDIS, IAIN MOFFATT, AND DYLAN P. THURSTON
, On the complex volume of hyperbolic knots, J. Knot Theory Ramifications 20 (2011), no. 7,
955–976.
School of Mathematics, Georgia Institute of Technology, Atlanta, GA 30332-0160, USA
http://www.math.gatech.edu/~stavros
E-mail address: [email protected]
Department of Mathematics, Royal Holloway, University of London, Egham, Surrey,
TW20 0EX, United Kingdom
http://www.personal.rhul.ac.uk/uxah/001
E-mail address: [email protected]
Department of Mathematics, Indiana University, Bloomington, IN 47405-7106, USA
http://pages.iu.edu/~dpthurst
E-mail address: [email protected]
| 4math.GR
|
1
The Indoor Mobile Coverage Problem Using
UAVs
arXiv:1705.09771v1 [cs.IT] 27 May 2017
Hazim Shakhatreh, Abdallah Khreishah, and Issa Khalil
Abstract
Unmanned aerial vehicles (UAVs) can be used as aerial wireless base stations when cellular networks
are not operational due to natural disasters. They can also be used to supplement the ground base
station in order to provide better coverage and higher data rates for the users. Prior studies on UAVbased wireless coverage typically consider an Air-to-Ground path loss model, which assumes that the
users are outdoor and located on a 2D plane. In this paper, we propose using UAVs to provide wireless
coverage for indoor users inside a high-rise building. First, we present realistic Outdoor-Indoor path loss
models and describe the tradeoff introduced by these models. Then we study the problem of efficient
placement of a single UAV, where the objective is to minimize the total transmit power required to
cover the entire high-rise building. The formulated problem is non-convex and is generally difficult to
solve. To that end, we consider three cases of practical interest and provide efficient solutions to the
formulated problem under these cases. Then we study the problem of minimizing the number of UAVs
required to provide wireless coverage to high rise buildings and prove that this problem is NP-complete.
Due to the intractability of the problem, we use clustering to minimize the number of UAVs required to
cover the indoor users. We demonstrate through simulations that the method that clusters the building
into regular structures and places the UAVs in each cluster requires 80% more number of UAVs relative
to our clustering algorithm.
Index Terms
Unmanned aerial vehicles, Outdoor-to-Indoor path loss model, gradient descent algorithm, particle
swarm optimization, k-means clustering.
Hazim Shakhatreh and Abdallah Khreishah are with the Department of Electrical and Computer Engineering, New Jersey
Institute of Technology (email: {hms35,abdallah}@njit.edu)
Issa Khalil is with Qatar Computing Research Institute (QCRI), HBKU, Doha, Qatar (email: [email protected])
Part of this work was presented in ICC 2017 [1] and ICICS 2017 [2].
2
I. I NTRODUCTION
UAVs can be used to provide wireless coverage during emergency cases where each UAV
serves as an aerial wireless base station when the cellular network goes down [3]. They can also
be used to supplement the ground base station in order to provide better coverage and higher
data rates for the users [4].
In order to use a UAV as an aerial wireless base station, the authors in [5] presented an Airto-Ground path loss model that helped the academic researchers to formulate many important
UAV-based coverage problems. The authors in [6] utilized this model to evaluate the impact of a
UAV altitude on the downlink ground coverage and to determine the optimal values for altitude
which lead to maximum coverage and minimum required transmit power. In [7], the authors
used the path loss model to propose a power-efficient deployment for UAVs under the constraint
of satisfying the rate requirement for all ground users. The authors in [8] utilized the path loss
model to study the optimal deployment of multiple UAVs equipped with directional antennas,
using circle packing theory. The 3D locations of the UAVs are determined in a way that the total
coverage area is maximized. In [9], the authors used the path loss model to find the minimum
number of UAVs and their 3D locations so that all outdoor ground users are served. However,
it is assumed that all users are outdoor and the location of each user can be represented by an
outdoor 2D point. These assumptions limit the applicability of this model when one needs to
consider indoor users.
Providing good wireless coverage for indoor users is very important. According to Ericsson
report [10], 90% of the time people are indoor and 80% of the mobile Internet access traffic
also happens indoors [11], [12]. To guarantee wireless coverage, service providers are faced
with several key challenges, including providing service to a large number of indoor users and
the ping pong effect due to interference from near-by macro cells [13]–[15]. In this paper, we
propose using UAVs to provide wireless coverage for users inside a high-rise building during
emergency cases and special events (such as concerts, indoor sporting events, etc.), when the
cellular network service is not available or it is unable to serve all indoor users. To the best of
our knowledge, this is the first work that proposes using UAVs to provide wireless coverage for
indoor users. We summarize our main contributions as follows:
•
We utilize an Outdoor-Indoor path loss model for low-SHF band (450 MHz to 6 GHz) [16],
certified by ITU, and an Outdoor-Indoor path loss model for high-SHF band (over 6
3
GHz) [17], then we show the tradeoff introduced by these models.
•
We formulate the problem of efficient placement of a single UAV, where the objective is to
minimize the total transmit power required to cover the entire high-rise building.
•
Since the formulated problem is non-convex and is generally difficult to solve, we consider
three cases of practical interest and provide efficient solutions to the formulated problem
under these cases and for different operating frequencies (low-SHF and high-SHF bands).
In the first case, we aim to find the minimum transmit power such that an indoor user with
the maximum path loss can be covered. In the second case, we assume that the locations of
indoor users are symmetric across the dimensions of each floor (such as office buildings or
hotels), and propose a gradient descent algorithm for finding an efficient location of a UAV.
In the third case, we assume that the locations of indoor users are uniformly distributed
in each floor, and propose a particle swarm optimization algorithm to find an efficient 3D
placement of a UAV that tries to minimize the total transmit power required to cover the
indoor users.
•
Due to the limited transmit power of a UAV, we formulate the problem of minimizing the
number of UAVs required to provide wireless coverage to high rise building and prove that
this problem is NP-complete.
•
Due to the intractability of the problem, we use clustering to minimize the number of UAVs
required to cover indoor users. We demonstrate through simulations that the method that
clusters the building into regular structures and places the UAVs in each cluster requires
80% more number of UAVs relative to our clustering algorithm.
II. S YSTEM M ODEL
A. System Settings
Let (xU AV ,yU AV ,zU AV ) denote the 3D location of the UAV. We assume that all users are
located inside a high-rise building as shown in Figure 1, and use (xi ,yi ,zi ) to denote the location
of user i. The dimensions of the high-rise building, in the shape of a rectangular prism, are
[0, xb ] × [0, yb ] × [0, zb ]. Also, let dout,i be the distance between the UAV and indoor user i, let
θi be the incident angle , and let din,i be the distance between the building wall and indoor user
i.
4
B. Outdoor-Indoor Path Loss Models
The Air-to-Ground path loss model presented in [5] is not appropriate when we consider
wireless coverage for indoor users, because this model assumes that all users are outdoor and
located at 2D points. In this paper, we adopt the Outdoor-Indoor path loss model, certified by
the ITU [16], for low-SHF operating frequency. The path loss is given as follows:
Li = LF + LB + LI = (w log10 dout,i + w log10 fGhz + g1 )
+(g2 + g3 (1 − cos θi )2 ) + (g4 din,i )
where LF is the free space path loss, LB is the building penetration loss, and LI is the indoor
loss. In this model, we also have w=20, g1 =32.4, g2 =14, g3 =15,g4 =0.5 [16] and fGhz is the
carrier frequency.
In [17], the authors clarify the Outdoor-to-Indoor path loss characteristics based on the measurement for 0.8 to 37 GHz frequency band. We adopt this path loss model for high-SHF
operating frequency. The path loss is given as follows:
Li = LF + LB + LI = (α1 + α2 log10 dout,i + α3 log10 fGhz ) +
(β1 +
β2 − β1
) + (γ1 din,i )
1 + exp(−β3 (θi − β4 ))
In this model, we have α1 =31.4, α2 =20, α3 =21.5, β1 =6.8, β2 =21.8, β3 =0.453, β4 =19.7 and
γ1 =0.49.
Note that there is a key tradeoff in the path loss models when the horizontal distance between
the UAV and a user changes. When this horizontal distance increases, the free space path loss
(i.e., LF ) increases as dout,i increases, while the building penetration loss (i.e., LB ) decreases as
the incident angle (i.e., θi ) decreases (Figure 2 shows the penetration loss for high-SHF band).
Similarly, when this horizontal distance decreases, the free space path loss (i.e., LF ) decreases
as dout,i decreases, while the building penetration loss (i.e., LB ) increases as the incident angle
(i.e., θi ) increases.
5
Fig. 1: Parameters of the Fig. 2: Building penetration loss Fig. 3: Transmit power required to
path loss model
for high-SHF
III. P ROVIDING W IRELESS
cover the building
COVERAGE USING A SINGLE
UAV
A. Problem Formulation
Consider a transmission between a UAV located at (xU AV ,yU AV ,zU AV ) and an indoor user i
located at (xi ,yi ,zi ). The rate for user i is given by:
Ci = Blog2 (1 +
Pt,i /Li
)
N
where B is the transmission bandwidth of the UAV, Pt,i is the UAV transmit power to indoor
user i, Li is the path loss between a UAV and an indoor user i and N is the noise power. In
this paper, we do not explicitly model interference, and instead, implicitly model it as noise.
Let us assume that each indoor user has a channel with bandwidth equals B/M, where M
is the number of users inside the building and the rate requirement for each user is v. Then the
minimum power required to satisfy this rate for each user is given by:
Pt,i,min = (2
v.M
B
− 1) ⋆ N ⋆ Li
Our goal is to find the optimal location of UAV such that the total transmit power required to
satisfy the downlink rate requirement of each indoor user is minimized. The objective function
can be represented as:
M
X
v.M
P =
(2 B − 1) ⋆ N ⋆ Li ,
i=1
6
where P is the UAV total transmit power. Since (2
v.M
B
− 1) ⋆ N is constant, our problem can be
formulated as:
min
xU AV ,yU AV ,zU AV
LT otal =
M
X
Li
i=1
subject to
xmin ≤ xU AV ≤ xmax ,
ymin ≤ yU AV ≤ ymax ,
zmin ≤ zU AV ≤ zmax ,
LT otal ≤ Lmax
Here, the first three constraints represent the minimum and maximum allowed values for xU AV ,
yU AV and zU AV . In the fourth constraint, Lmax is the maximum allowable path loss and equals
Pt,max /((2
v.M
B
− 1) ⋆ N), where Pt,max is the maximum transmit power of UAV.
Finding the optimal placement of UAV is generally difficult because the problem is nonconvex. Therefore, in the next subsection, we consider three special cases of practical interest
and derive efficient solutions under these cases.
B. Efficient Placement of a Single UAV
Case 1. The worst location in building: In this case, we find the minimum transmit power
required to cover the building based on the location that has the maximum path loss inside the
building. The location that has the maximum path loss in the building is the location that has
maximum dout,i , maximum θi , and maximum din,i . The locations that have the maximum path
loss are located at the corners of the highest and lowest floors. Since the locations that have
the maximum path loss inside the building are the corners of the highest and lowest floors,
we place the UAV at the middle of the building (yU AV = 0.5yb and zU AV =0.5zb ). Then, given
Outdoor-to-Indoor path loss models for low-SHF and high-SHF bands, we need to find an
efficient horizontal point xU AV for the UAV such that the total transmit power required to cover
the building is minimized.
Now, when the horizontal distance between the UAV and this location increases, the free space
path loss also increases as dout,i increases, while the building penetration loss decreases because
we decrease the incident angle θi . In Figure 3, we demonstrate the minimum transmit power
7
Fig. 4: Transmit power re- Fig. 5: Transmit power re- Fig. 6: Transmit power required to cover the building, quired to cover the building, quired to cover 30 meter buildfc =2 GHz
fc =10 GHz
ing height
required to cover a building of different heights, where the minimum transmit power required
to cover the building is given by:
Pt,min (dB) = Pr,th + Li
Pr,th (dB) = N + γth
Here, Pr,th is the minimum received power, N is the noise power (equals -120dBm), γth is the
threshold SNR (equals 10dB), yb=50 meters , xb =20 meters and the carrier frequency is 2Ghz.
The numerical results show that there is an optimal horizontal point that minimizes the total
transmit power required to cover a building. Also, we note that when the height of the building
increases, the optimal horizontal distance increases. This is to compensate for the increased
building penetration loss due to an increased incident angle.
In Theorem 1, we characterize the optimal incident angle θ for low-SHF band that minimizes
the transmit power required to cover the building. This helps us finding the optimal horizontal
distance between the UAV and the building.
Theorem 1. For the low-SHF operating frequency case, when we place the UAV at the middle
of building , the optimal incident angle θ that minimizes the transmit power required to cover
the building will be equal to 48.654o and the optimal horizontal distance between the UAV and
0.5zb
)2 − (0.5yb )2 )0.5 − xb .
the building will be equal to ((
tan(48.654o)
8
Proof. In order to find the optimal horizontal point, we rewrite the equation that represents the
path loss in terms of the incident angle (θi ) and the altitude difference between the UAV and
the user i (∆hi ):
Li (∆hi , θi ) = w log10
∆hi
+ w log10 fGhz + g1
sin θi
+g2 + g3 (1 − cos θi )2 + g4 din,i
We know that the altitude difference between the UAV and the location that has the maximum
path loss is constant for a given building. Now, when we take the first derivative with respect
to θ and assign it to zero, we get:
−∆h. cos θ
sin2 θ
+ 2g3 sin θ(1 − cos θ) = 0
∆h
sin θ
−w cos θ
dL(θ)
=
+ 2g3 sin θ(1 − cos θ) = 0
dθ
ln10 sin θ
w
cos θ = 2g3 sin2 θ(1 − cos θ)
ln10
w
cos θ = 2g3 (1 − cos2 θ)(1 − cos θ)
ln10
w
+ 2g3 ) cos θ + 2g3 = 0
2g3 cos3 θ − 2g3 cos2 θ − (
ln10
To prove that the function is convex, we take the second derivative and we get:
w
dL(θ)
=
dθ
ln10
d2 L
w
1
+ 2g3 cos θ(1 − cos θ) + 2g3 sin2 θ > 0
=
2
dθ
ln10 sin2 θ
(1)
f or 0 < θ ≤ 90
Ecrf (1) has only one valid solution which is cos θ=0.6606. Therefore, the optimal incident angle
between the UAV and the location that has the maximum path loss inside the building will be
48.654o.
In order to find the optimal horizontal distance between the UAV and the building, we apply
the pythagorean’s theorem. This gives us:
dH = ((
0.5zb
)2 − (0.5yb)2 )0.5
tan(48.654o )
Therefore, the optimal horizontal distance between the UAV and the building is given by:
dopt = ((
0.5zb
)2 − (0.5yb )2 )0.5 − xb
o
tan(48.654 )
9
In Figure 4, we demonstrate the transmit power required to cover the building as a function
of the incident angle, we notice that the optimal angle that we characterize in Theorem 1 gives
us the minimum transmit power.
Now, we find an efficient incident angle θ for high-SHF band that minimizes the transmit
power required to cover the building. In order to find an efficient angle, we rewrite the equation
that represents the path loss in terms of the incident angle (θ) and the altitude difference between
the UAV and location that has the maximum path loss inside the building (∆h), we get:
∆h
+ α3 log10 fGhz ) +
sin θ
β2 − β1
(β1 +
) + (γ1 din,i )
1 + exp(−β3 (θi − β4 ))
L(∆h, θ) = (α1 + α2 log10
By numerically plotting the transmit power required to cover the location that has the maximum
path loss inside the building (see Figure 5 and Figure 6), where yb =50 meters and xb =20 meters,
we show that for different building heights and different operating frequencies there exists only
one global minimum value. As can be seen from the figures, to provide wireless coverage to small
buildings, the UAV transmit power must be very high, due to the high free space path loss, this
demonstrates the need for multiple UAVs to cover the high rise building when we use high-SHF
operating frequency. To find an efficient incident angle that could give us the global minimum
value, we use the ternary search algorithm. A ternary search algorithm is a method for finding
the minimum of a unimodal function, it iteratively splits the domain into three separate regions
and discards the one where the minimum does not belong to. The pseudo code of this algorithm
is shown in Algorithm 1. From our numerical results, we found that the angle that minimizes
the power is always 15o. This is because the building penetration loss will be minimized at this
angle (see Figure 2). The angles less than 15o will also give us minimum building penetration
loss but the free space path loss will increase as the incident angle θi decreases. Note that for
the high-SHF case the incident angle that results in the minimum path loss is smaller than that
for low-SHF case. This is due to the fact that the building penetration loss at high operating
frequency will be higher than that at low operating frequency.
Case 2. The locations of indoor users are symmetric across the xy and xz planes: In
this case, we assume that the locations of indoor users are symmetric across the xy-plane
((0,0,0.5zb ),(xb ,0,0.5zb ) ,(xb ,yb ,0.5zb ),(0,yb ,0.5zb ))) and the xz-plane ((0,0.5yb ,0), (xb ,0.5yb,0),
(xb ,0.5yb,zb ),(0,0.5yb,zb )). First, we prove that zU AV =0.5zb and yU AV =0.5yb when the locations of
indoor users are symmetric across the xy and xz planes and the operating frequency is low-SHF
10
Algorithm 1 Ternary search algorithm
Input:
The interval [a,b] of unimodal function that contains the efficient incident angle.
The absolute precision =µ.
If |b-a| < µ:
Return
(a+b)
2
l = a+ (b−a)
3
r = b− (b−a)
3
If f (l) > f (r)
Return ternary search(f , l, b, µ)
Else
Return ternary search(f , a, r, µ)
(Theorem 2) or high-SHF (Theorem 3). Then we use the gradient descent algorithm to find an
efficient xU AV that minimizes the transmit power required to cover the building.
Theorem 2. For the low-SHF operating frequency case, when the locations of indoor users
are symmetric across the xy and xz planes, the optimal (yU AV ,zU AV ) that minimizes the power
required to cover the indoor users will be equal (0.5yb ,0.5zb ).
The proof is presented in Appendix A. The question now is how to find an efficient horizontal
point xU AV that minimizes the total transmit power. In order to find this point, we use the
gradient descent algorithm [18]:
xU AV,n+1 = xU AV,n − a
dLT otal
dxU AV,n
Where:
((xi − xU AV )2 + (yi − yU AV )2 )0.5
dLT otal X w −(xi − xU AV )
+
2g
.(1
−
=
).
3
dxU AV
ln10
d2out,i
dout,i
i=1
M
(
(xi − xU AV )dout,i ((xi − xU AV )2 + (yi − yU AV )2 )−0.5
−
d2out,i
((xi − xU AV )2 + (yi − yU AV )2 )0.5 (xi − xU AV )d−1
out,i
)
2
dout,i
11
a: the step size.
dout,i =((xi − xU AV )2 + (yi − yU AV )2 + (zi − zU AV )2 )0.5
The pseudo code of this algorithm is shown in Algorithm 2. Now, we prove that zU AV = 0.5zb
and yU AV = 0.5yb when the locations of indoor users are symmetric across the xy and xz planes
and the operating frequency is high-SHF.
Algorithm 2 Efficient xU AV using gradient descent algorithm
Input:
The 3D locations of the users inside the building.
The step size a, the step tolerance ǫ.
The dimensions of the building [0, xb ] × [0, yb ] × [0, zb ].
The maximum number of iterations Nmax .
Initialize xU AV
For n=1,2,..., Nmax
dLT otal
dxU AV,n
− xU AV,n+1 k < ǫ
xU AV,n+1 ← xU AV,n − a
If k xU AV,n
Return: xU AV,opt = xU AV,n+1
End for
Theorem 3. For the high-SHF operating frequency case, when the locations of indoor users
are symmetric across the xy and xz planes, the optimal (yU AV , zU AV ) that minimizes the power
required to cover the indoor users will be equal (0.5yb ,0.5zb ).
The proof is presented in Appendix B. To find an efficient horizontal point xU AV that minimizes
the total transmit power, we use the gradient descent algorithm, where:
i )(xU AV −xi )
3
M
−(β2 − β1 )( √−β
)
)( −(zU AV −z
d3out,i
dLT otal X α2 (xU AV − xi )
1−u2
+
(
=
.
2
−1 u − β )))
dxU AV
ln10
d
(1
+
exp(−β
(sin
3
4
out,i
i=1
exp(−β3 (sin−1 u − β4 ))
)
(1 + exp(−β3 (sin−1 u − β4 )))
dout,i =((xi − xU AV )2 + (yi − yU AV )2 + (zi − zU AV )2 )0.5
(zU AV − zi )
)
u=(
((xU AV − xi )2 + (yU AV − yi )2 + (zU AV − zi )2 )0.5
12
Case 3. The locations of indoor users are uniformly distributed in each floor: In this case,
we propose the Particle Swarm Optimization (PSO) [19] to find an efficient 3D placement of
the UAV, when the locations of indoor users are uniformly distributed in each floor.
The particle swarm optimization algorithm starts with (npop) random solutions and iteratively tries to improve the candidate solutions based on the best experience of each candidate
(particle(i).best.location) and the best global experience (globalbest.location). In each iteration,
the best location for each particle (particle(i).best.location) and the best global location (globalbest.location) are updated and the velocities and locations of the particles are calculated based
on them [9]. The velocity value indicates how much the location can be changed (see ecrf (2)).
The velocity is given by:
particle(i).velocity = w ∗ particle(i).velocity +
c1 ∗ rand(varsize) ∗ (particle(i).best.location − particle(i).location)
+c2 ∗ rand(varsize) ∗ (globalbest.location − particle(i).location)
where w is the inertia weight, c1 and c2 are the personal and global learning coefficients, and
rand(varsize) are random positive numbers. Also, the location of each particle is updated as:
particle(i).location = particle(i).location + particle(i).velocity
(2)
The pseudo code of the PSO algorithm is shown in Algorithm 3. Convergence of the candidate
solutions has been investigated for PSO [20]. This analyses has resulted in guidelines for selecting
a set of coefficients (κ,φ1 ,φ2 ) that are believed to cause convergence to a point and prevent
divergence of the swarms particles. We selected our parameters according to this analysis (see
Table I and Algorithm 3).
IV. PROVIDING WIRELESS COVERAGE USING MULTIPLE UAV S
Providing wireless coverage to High-rise building using a single UAV can be impractical, due
to the limited transmit power of a UAV. The transmit power required to cover the building is
too high. It is in the range of 50dBm to 65dBm (see Figures 3, 5 and 6), which corresponds to
100-3000 watts.
13
Algorithm 3 Efficient UAV placement using PSO algorithm
Input:
The lower and upper bounds of decision variable (varmin,varmax), Construction coefficients
(κ,φ1 ,φ2 ), Maximum number of iterations (maxit), Population size (npop)
Initialiaztion:
φ=φ1 +φ1 , χ = 2κ/|2 − φ − (φ2 − 4φ)0.5 |
w=χ, c1 =χφ1 , c2 =χφ2 , globalbest.cost=inf
for i=1:npop
particle(i).location=unifrnd(varmin, varmax, varsize)
particle(i).velocity=zeros(varsize)
particle(i).cost=costfunction(particle(i).location)
particle(i).best.location=particle(i).location
particle(i).best.cost=particle(i).cost
if particle(i).best.cost < globalbest.cost
globalbest=particle(i).best
end if
end
PSO Loop:
for t=1:maxit
for i=1:npop
particle(i).velocity=w*particle(i).velocity+
c1 *rand(varsize)*(particle(i).best.location- particle(i).location)+
c2 *rand(varsize)*(globalbest.location-particle(i).location)
particle(i).location=particle(i).location+ particle(i).velocity
particle(i).cost=costfunction(particle(i).location)
if particle(i).cost < particle(i).best.cost
particle(i).best.location = particle(i).location
particle(i).best.cost = particle(i).cost
if particle(i).best.cost < globalbest.cost
globalbest=particle(i).best
end if
end if
end
end
14
Our problem can be formulated as:
min |k|
subject to
|k|
X
yij = 1
∀i ∈ m (3.a)
j=1
|m|
X
v.|m|
B
∀j ∈ k
(3.b)
xmin ≤ xj ≤ xmax
∀j ∈ k
(3.c)
ymin ≤ yj ≤ ymax
∀j ∈ k
(3.d)
zmin ≤ zj ≤ zmax
∀j ∈ k
(3.e)
(2
− 1).N.Lij .yij ≤ P
(3)
i=1
where k is a set of fully charged UAVs, m is a set of indoor users, υ is the rate requirement for
each user (constant), N is the noise power (constant), B is the transmission bandwidth (constant),
Lij is the total path loss between UAV j and user i and P is the maximum transmit power of
UAV (constant). We also introduce the binary variable yij that takes the value of 1 if the indoor
user i is connected to the UAV j and equals 0 otherwise. The objective is to minimize the
number of UAVs that are needed to provide a wireless coverage for indoor users. Constraint set
(3.a) ensure that each indoor user should be connected to one UAV. Constraint set (3.b) ensure
that the total power consumed by a UAV should not exceed its maximum power consumption
limit. Constraints (3.c-3.e) represent the minimum and maximum allowed values for xj , yj and
zj .
Theorem 4. The problem represented by (3) is NP-complete.
Proof. The number of constraints is polynomial in terms of the number of indoor users, UAVs
and 3D locations. Given any solution for our problem, we can check the solutions feasibility in
polynomial time, then the problem is NP.
To prove that the problem is NP-hard, we reduce the Bin Packing Problem which is NP-hard [21]
to a special case of our problem. In the Bin Packing Problem, we have a set of items G =
{1, 2, .., N} in which each item has volume zn where n ∈ G. All items must be packed into a
finite number of bins (b1 , b,...,bB ), each of volume V in a way that minimizes the number of
bins used. The reduction steps are: 1) The b-th bin in the Bin Packing Problem is mapped to the
j-th UAV in our problem, where the volume V for each bin is mapped to the maximum transmit
power of the UAV P . 2) The n-th item is mapped to the indoor i-th user, where the volume for
15
each item n is mapped to the power required to cover the i-th indoor user. 3) All UAVs have the
same maximum transmit power P . 4) The power required to cover the i-th indoor user from any
3D location will be constant. If there exists a solution to the bin packing problem with cost C,
then the selected bins will represent the UAVs that are selected and the items in each bin will
represent the indoor users that the UAV must cover and the total cost of our problem is C.
Due to the intractability of the problem, we study clustering indoor users. In the k-means
clustering algorithm [22], we are given a set of points m, and want to group the points into a k
clusters such that each point belongs to the cluster with the nearest mean. The first step in the
algorithm is to choose the number of clusters k. Then, randomly initialize k clusters centroids.
In each iteration, the algorithm will do two things:1) Cluster assignment step. 2) Move centroids
step. In cluster assignment step, the algorithm goes through each point and chooses the closest
centroids and assigns the point to it. In move centroids step, the algorithm calculates the average
for each group and moves the centroids there. The algorithm will repeat these two steps until
it converges. The algorithm will converge when the assignments no longer change. To find the
minimum number of UAVs required to cover the indoor users, we utilize this algorithm to cluster
the indoor users. In our algorithm, we assume that each cluster will be covered by only one UAV.
We start the algorithm with k = 2 and after it finishes clustering the indoor users, it applies the
particle swarm optimization [19] to find the UAV 3D location and UAV transmit power needed
to cover each cluster. Then, it checks if the maximum transmit power is sufficient to cover each
cluster, if not, the number of clusters k is incremented by one and the problem is solved again.
The pseudo code of this algorithm is shown in Algorithm 4.
V. N UMERICAL R ESULTS
A. Simulation results for single UAV
First, we verify our results for the second case, when the locations of indoor users are
symmetric across the xy and xz planes, using different operating frequencies, 2GHz for low-SHF
band and 15GHz for high-SHF. We assume that each floor contains 20 users. Then we apply
the gradient descent (GD) algorithm to find the optimal horizontal point xU AV that minimizes
the transmit power required to cover the indoor users. Table I lists the parameters used in the
numerical analysis for single UAV cases.
16
Fig. 7: UAV optimal placement (upper part) and convergence speed of the GD algorithm (lower
part) for different building heights, fc = 2G Hz
Fig. 8: UAV optimal placement (upper part) and convergence speed of the GD algorithm (lower
part) for different building heights, fc = 15G Hz
In Figures 7 and 8, we find the optimal horizontal points for a building of different heights. In
the upper part of the figures, we find the total path loss at different locations (xU AV ,0.5yb,zU AV )
and the optimal horizontal point xU AV that results in the minimum total path loss using the GD
algorithm. In the lower part of the figures, we show the convergence speed of the GD algorithm.
17
Algorithm 4 Clustering Indoor Users
Input:
The maximum transmit power of UAV (P ).
The 3D locations of m indoor users (xi , yi , zi ).
Number of clusters (|k| = 2).
START:
1: Initialize cluster centroids γ1 , γ2, ..., γk ∈ Rn randomly.
2: Repeat until convergence:
For every indoor user i ∈ m, set
c(i) = arg min ||(xi , yi , zi ) − γj ||2
j∈k
For each cluster
X j ∈ k, set
(xi , yi , zi )
γj =
i∈m,c(i) =j
X
1
i∈m,c(i) =j
3: Using particle swarm optimization algorithm, calculate the UAV efficient 3D location and
the transmit power for each cluster j ∈ k:
X
v.|m|
P (j) =
(2 B − 1) ⋆ N ⋆ Li
i∈m,c(i) =j
4: For j = 1 to |k|
If (P (j) > P )
|k| = |k| + 1
go to START
End
Output:
|k| Clusters.
The transmit Power of each UAV.
The 3D locations of UAVs.
As can be seen from the figures, when the height of the building increases, the optimal horizontal
point xU AV increases. This is to compensate the increased building penetration loss due to an
increased incident angle.
In Figures 9 and 10, we investigate the impact of different building widths (i.e., xb ). We fix
the building height to be 250 meters for low-SHF operating frequency and 25 meters for high-
18
TABLE I: Parameters in numerical analysis for single UAV
Vertical width of building yb
50 meters
Hight of each floor
5 meters
Step size a ”GD algorithm”
0.01
Maximum number of iterations Nmax ”GD algorithm”
500
The carrier frequency fGhz , low-SHF
2Ghz
The carrier frequency fGhz , high-SHF
15Ghz
Number of users in each floor
20 users
(varmin,varmax) ”PSO algorithm”
(0,1000)
(κ,φ1 ,φ2 ) ”PSO algorithm”
(1,2.05,2.05)
Fig. 9: UAV optimal placement (upper part) and convergence speed of the GD algorithm (lower
part) for different building widths, fc = 2G Hz
SHF, then we vary the building width. As can be seen from the figures, when the building width
increases, the optimal horizontal distance decreases. This is to compensate for the increased
indoor path loss due to an increased building width.
Now, we validate the simulation results for low-SHF operating frequency by using the particle
swarm optimization (PSO) algorithm and verify our result for the third case, when the locations
of indoor users are uniformly distributed in each floor, using low-SHF operating frequency. As
can be seen from the simulation results in Table II, both algorithms converge to the same 3D
19
Fig. 10: UAV optimal placement (upper part) and convergence speed of the GD algorithm (lower
part) for different building widths, fc = 15G Hz
Fig. 11: UAV efficient placement (upper part) and convergence speed of the PSO algorithm
(lower part) for different building heights
placement, when the locations of indoor users are symmetric across the xy and xz planes.
After that, we assume that each floor contains 20 users and the locations of these users are
uniformly distributed in each floor. When we apply the GD algorithm, the 3D efficient placements
and the total costs for 200 meter, 250 meter and 300 meter buildings are (24.7254, 25, 100)
20
TABLE II: Simulation Results
Algorithm
Distribution
Building height
Horizontal building
Vertical building
zb
width xb
width yb
Efficient 3D placement
Efficient total
path loss(dB)
GD
symmetric
200
20
50
(20.025, 25, 100)
7.8825 ∗ 104
PSO
symmetric
200
20
50
(20.040, 25.0130, 100.0015)
7.8825 ∗ 104
GD
symmetric
250
20
50
(30.809, 25, 125)
9.9971 ∗ 104
PSO
symmetric
250
20
50
(30.736 , 24.960, 124.956)
9.9971 ∗ 104
GD
symmetric
300
20
50
(40.746, 25, 150)
1.2146 ∗ 105
PSO
symmetric
300
20
50
(40.758, 25.048, 150.054)
1.2146 ∗ 105
GD
uniform
200
20
50
(24.725, 25, 100)
7.8853 ∗ 104
PSO
uniform
200
20
50
(21.799, 37.389, 111.790)
7.8645 ∗ 104
GD
uniform
250
20
50
(33.818, 25, 125)
9.9855 ∗ 104
PSO
uniform
250
20
50
(32.921, 28.712, 124.029)
9.9725 ∗ 104
GD
uniform
300
20
50
(43.117, 25, 150)
1.2154 ∗ 105
PSO
uniform
300
20
50
(46.589, 31.506 ,143.858)
1.2117 ∗ 105
GD
uniform
250
10
50
(38.521, 25, 125)
9.7413 ∗ 104
PSO
uniform
250
10
50
(32.104, 21.017, 129.266)
9.7252 ∗ 104
GD
uniform
250
30
50
(29.393, 25, 125)
1.0275 ∗ 105
PSO
uniform
250
30
50
(25.529, 4.938, 138.765)
1.0211 ∗ 105
GD
uniform
250
50
50
(22.711, 25, 125)
1.0753 ∗ 105
PSO
uniform
250
50
50
(14.548, 17.308 ,131.8940)
1.0696 ∗ 105
(7.8853∗104), (33.8180, 25, 125) (9.9855∗104) and (43.1170, 25, 150)(1.2154∗105), respectively.
UAV efficient placement and the convergence speed of the PSO algorithm for different building
heights is shown in Figure 11. The 3D efficient placements and the total costs for 200 meter,
250 meter and 300 meter buildings are (21.7995, 37.3891, 111.7901) (7.8645 ∗ 104 ), (32.9212,
28.7125, 124.0291) (9.9725 ∗ 104 ) and (46.5898, 31.5061 ,143.8588)(1.2117 ∗ 105 ), respectively.
As can be seen from the simulation results, the PSO algorithm provides better results. It provides
total cost less than the cost that the GD algorithm provides by (37dB-208dB). This is because
21
Fig. 12: UAV efficient placement (upper part) and convergence speed of the PSO algorithm
(lower part) for different building widths
the PSO algorithm is designed for the case in which the locations of indoor users are uniformly
distributed in each floor. On the other hand, the GD algorithm is designed for the case in which
the locations of indoor users are symmetric across the dimensions of each floor.
We also investigate the impact of different building widths (i.e., xb ) using the GD and PSO
algorithms (see Figure 12). We fix the building height to be 250 meters and vary the building
width. As can be seen from the simulation results, the PSO algorithm provides better results. It
provides total cost less than the cost that the GD algorithm provides by (57dB-161dB).
We can notice that the tradeoff in case three is similar to that in case two, when the height of the
building increases, the efficient horizontal point xU AV computed by our algorithm increases. This
is to compensate the increased building penetration loss due to an increased incident angle. Also,
when the building width increases, the efficient horizontal distance computed by our algorithm
decreases. This is to compensate the increased indoor path loss due to an increased building
width.
B. Simulation results for multiple UAVs
In this section, we verify our results for multiple UAVs scenario. First, we assume that a
building will host a special event (such as concert, conference, etc.), the dimensions of the
building are [0, 20] × [0, 50] × [0, 100]. The organizers of the event reserve all floors higher than
22
75 meters and they expect that 200 people will attend the event. Due to interference from near-by
macro cells, the organizers decide to use UAVs to provide wireless coverage to the indoor users.
We assume that 200 indoor users are uniformly distributed in upper part of the building (higher
than 75 meters) and 200 indoor users are uniformly distributed in the lower part (less than 75
meters). Then, we apply the clustering indoor users algorithm to find the minimum number of
UAVs required to cover the indoor users. Table III lists the parameters used in the numerical
analysis for multiple UAVs.
The algorithm starts with k = 2 and after it finishes clustering the indoor users, it applies the
particle swarm optimization to find the UAV 3D location and UAV transmit power needed to
cover each cluster. Then, it checks if the maximum transmit power is sufficient to cover each
cluster, if not, the number of clusters k is incremented by one and the problem is solved again.
As can be seen from the simulation results in Figure 13, we need 5 UAVs to cover the indoor
users. We can notice that an efficient horizontal point xU AV for all UAVs 3D locations is the
same xU AV = 25, the minimum allowed value for xU AV , this is because the tradeoff (shown in
Figure 3) disappears when a UAV covers small height of building.
In Figure 14, we uniformly split the building into k parts and cover it by k UAVs. As can
be seen from the simulation results, we need 9 UAVs to cover the indoor users. The clustering
algorithm provides better results, this is because it utilizes the distribution of indoor users to
divide them into clusters. On the other hand, the uniformly split method is designed for the case
in which the locations of indoor users uniformly distributed in the building.
TABLE III: Parameters in numerical analysis for multiple UAVs
Maximum transmit power of UAV (P )
5 Watt
Operating frequency (f )
2Ghz
Transmission bandwidth (B)
50M Hz
Rate requirement for each user (υ)
2.2Mbps
Noise power (N )
-150 dBm
Min and Max allowed values for xj ,[xmin , xmax ]
[25,1000]
Min and Max allowed values for yj , [ymin , ymax ]
[0,50]
Min and Max allowed values for zj , [zmin , zmax ]
[0,1000]
23
Fig. 13: UAVs efficient placements using clustering algorithm
Fig. 14: UAVs efficient placements using uniform split method
VI. C ONCLUSION
In this paper, we study the problem of providing wireless coverage for users inside a highrise building using UAVs. First, we demonstrate why the Air-to-Ground path loss model is not
appropriate for considering indoor users with 3D locations. Then, we present Outdoor-to-Indoor
path loss models, show the tradeoff in these models, and study the problem of minimizing the
transmit power required to cover the building. Due to the intractability of the problem, we study
an efficient placement of a single UAV under three cases. Due to the limited transmit power
of a UAV, we formulate the problem of minimizing the number of UAVs required to provide
24
wireless coverage to high rise building and prove that this problem is NP-complete. Due to the
intractability of the problem, we use clustering to minimize the number of UAVs required to
cover the indoor users. In order to model more realistic scenarios, we will study the problem of
providing wireless coverage for multiple buildings in our future work.
A PPENDIX A
P ROOF
OF
T HEOREM 2
Consider that m1 represents the users that have altitude lower than the UAV altitude and m2
represents the users that have altitude higher than the UAV altitude, then:
dout,i = ((xU AV − xi )2 + (yU AV − yi )2 + (zU AV − zi )2 )0.5 , ∀zU AV > zi
dout,i = ((xU AV − xi )2 + (yU AV − yi )2 + (zi − zU AV )2 )0.5 , ∀zU AV < zi
Also,
cosθi =
((xU AV
cosθi =
((xU AV − xi )2 + (yU AV − yi )2 )0.5
, ∀zU AV > zi
− xi )2 + (yU AV − yi )2 + (zU AV − zi )2 )0.5
((xU AV − xi )2 + (yU AV − yi )2 )0.5
, ∀zU AV < zi
− xi )2 + (yU AV − yi )2 + (zi − zU AV )2 )0.5
((xU AV
Rewrite the total path loss:
m1
m2
X
X
2
LT otal =
(wlog10 (dout,i ) + g3 (1 − cos θi ) ) +
(wlog10 (dout,i ) + g3 (1 − cos θi )2 ) + K
i=1
i=1
Where:
M
X
K=
(wlog10 fGhz + g1 + g2 + g4 din,i )
i=1
Now, take the derivative with respect to zU AV , we get:
m1
w
dLT otal X
(zU AV − zi )
=
+
2
dzU AV
ln10 ((xU AV − xi ) + (yU AV − yi )2 + (zU AV − zi )2 )
i=1
((xU AV − xi )2 + (yU AV − yi )2 )0.5
).
((xU AV − xi )2 + (yU AV − yi )2 + (zU AV − zi )2 )0.5
((xU AV − xi )2 + (yU AV − yi )2 )0.5 (zU AV − zi )
(
3 ) +
((xU AV − xi )2 + (yU AV − yi )2 + (zU AV − zi )2 ) 2
m2
X
w
−(zi − zU AV )
2g3 .(1 −
i=1
ln10 ((xU AV − xi )2 + (yU AV − yi )2 + (zi − zU AV )2 )
((xU AV − xi )2 + (yU AV − yi )2 )0.5
).
((xU AV − xi )2 + (yU AV − yi )2 + (zi − zU AV )2 )0.5
−((xU AV − xi )2 + (yU AV − yi )2 )0.5 (zi − zU AV )
(
3 )
((xU AV − xi )2 + (yU AV − yi )2 + (zi − zU AV )2 ) 2
+2g3.(1 −
25
Rewrite the
dLT otal
again, we have:
dzU AV
1
((xU AV − xi )2 + (yU AV − yi )2 )0.5
dLT otal X
w (zU AV − zi )
+
2g
.(1
−
=
).
3
2
dzU AV
ln10
d
d
out,i
out,i
i=1
m
2
X
w −(zi − zU AV )
((xU AV − xi )2 + (yU AV − yi )2 )0.5 (zU AV − zi )
)
+
+
(
3
2
dout,i
ln10
d
out,i
i=1
m
2g3 .(1 −
((xU AV − xi )2 + (yU AV − yi )2 )0.5 −((xU AV − xi )2 + (yU AV − yi )2 )0.5 (zi − zU AV )
)
).(
dout,i
d3out,i
The equation above equals zero when the UAV altitude equals the half of the building height,
where the locations of indoor users are symmetric across the xy and xz planes.
A PPENDIX B
P ROOF
OF
T HEOREM 3
Consider that m1 represents the users that have altitude lower than the UAV altitude and m2
represents the users that have altitude higher than the UAV altitude, then:
dout,i = ((xU AV − xi )2 + (yU AV − yi )2 + (zU AV − zi )2 )0.5 ,
∀zU AV > zi
dout,i = ((xU AV − xi )2 + (yU AV − yi )2 + (zi − zU AV )2 )0.5 ,
∀zU AV < zi
Also,
θi = sin−1 (
θi = sin−1 (
((xU AV
((xU AV
(zU AV − zi )
),
− xi )2 + (yU AV − yi )2 + (zU AV − zi )2 )0.5
(zi − zU AV )
),
2
− xi ) + (yU AV − yi )2 + (zi − zU AV )2 )0.5
∀zU AV > zi
∀zU AV < zi
Rewrite the total path loss:
LT otal =
m1
X
α2 log10 (dout,i ) +
+
i=1
m2
X
(β2 − β1 )
(1 + exp(−β3 (sin−1 (u) − β4 )))
α2 log10 (dout,i ) +
(β2 − β1 )
+K
(1 + exp(−β3 (sin−1 (u) − β4 )))
i=1
Where:
u=(
u=(
K=
((xU AV
((xU AV
M
X
i=1
(zU AV − zi )
),
− xi + (yU AV − yi )2 + (zU AV − zi )2 )0.5
(zi − zU AV )
),
2
− xi ) + (yU AV − yi )2 + (zi − zU AV )2 )0.5
)2
(α1 + α3 log10 fGhz + β1 + γ1 din,i )
∀zU AV > zi
∀zU AV < zi
26
Now, take the derivative with respect to zU AV , we get:
1
α2
(zU AV − zi )
dLT otal X
=
2
dzU AV
ln10 ((xU AV − xi ) + (yU AV − yi )2 + (zU AV − zi )2 )
i=1
m
+(
3
−(β2 − β1 )( √−β
)(
1−u2
dout,i −(zU AV −zi )2 d−1
out,i
)
d2out,i
(1 + exp(−β3 (sin−1 u − β4 )))
.
exp(−β3 (sin−1 u − β4 ))
)+
(1 + exp(−β3 (sin−1 u − β4 )))
m2
X
α2
−(zi − zU AV )
2
ln10 ((xU AV − xi ) + (yU AV − yi )2 + (zi − zU AV )2 )
i=1
+(
3
−(β2 − β1 )( √−β
)(
1−u2
−dout,i +(zU AV −zi )2 d−1
out,i
)
d2out,i
(1 + exp(−β3 (sin−1 u − β4 )))
.
exp(−β3 (sin−1 u − β4 ))
)
(1 + exp(−β3 (sin−1 u − β4 )))
The equation above equals zero when the UAV altitude equals the half of the building height,
where the locations of indoor users are symmetric across the xy and xz planes.
R EFERENCES
[1] H. Shakhatreh, A. Khreishah, and B. Ji, “Providing wireless coverage to high-rise buildings using uavs,” in (IEEE
International Conference on Communications, IEEE ICC 2017 (accepted).
IEEE, 2017.
[2] H. Shakhatreh, A. Khreishah, A. Alsarhan, I. Khalil, A. Sawalmeh, and O. Noor Shamsiah, “Efficient 3d placement of
a uav using particle swarm optimization,” in The International Conference on Information and Communication Systems
(ICICS 2017) (accepted).
[3] P. Bupe, R. Haddad, and F. Rios-Gutierrez, “Relief and emergency communication network based on an autonomous
decentralized uav clustering network,” in SoutheastCon 2015.
IEEE, 2015, pp. 1–8.
[4] R. I. Bor-Yaliniz, A. El-Keyi, and H. Yanikomeroglu, “Efficient 3-d placement of an aerial base station in next generation
cellular networks,” in Communications (ICC), 2016 IEEE International Conference on.
IEEE, 2016, pp. 1–5.
[5] A. Al-Hourani, S. Kandeepan, and A. Jamalipour, “Modeling air-to-ground path loss for low altitude platforms in urban
environments,” in 2014 IEEE Global Communications Conference.
IEEE, 2014, pp. 2898–2904.
[6] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Drone small cells in the clouds: Design, deployment and performance
analysis,” in IEEE Global Communications Conference (GLOBECOM), 2015, pp. 1–6.
[7] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Optimal transport theory for power-efficient deployment of unmanned
aerial vehicles,” IEEE International Conference on Communications (ICC), Kuala Lumpur, Malaysia,, 2016.
[8] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Efficient deployment of multiple unmanned aerial vehicles for optimal
wireless coverage,” IEEE Communications Letters, vol. 20, no. 8, pp. 1647–1650, 2016.
[9] E. Kalantari, H. Yanikomeroglu, and A. Yongacoglu, “On the number and 3d placement of drone base stations in wireless
cellular networks,” in IEEE Vehicular Technology Conference, 2016, pp. 18–21.
[10] “Ericsson report optimizing the indoor experience, http://www.ericsson.com/res/docs/2013/real-performance-indoors.pdf,”
2013.
[11] “In-building wireless: One size does not fit all, http://www.alcatel-lucent.com/solutions/in-building/in-building-infographic.”
[12] “Cisco
eration,
service
provider
wi-fi:
A
platform
for
business
innovation
and
revenue
gen-
http://www.cisco.com/c/en/us/solutions/collateral/service-provider/service-provider-wi-
fi/solution overview c22 642482.html.”
27
[13] “Using high-power das in high-rise buildings, http://www.commscope.com/docs/using-high-power-das-in-high-risebuildings-an-318376-ae.pdf.”
[14] “Coverage
solution
for
high-rise
building,
http://www.amplitec.net/products-2-coverage-solution-for-high-rise-
building.html.”
[15] S. Zhang, Z. Zhao, H. Guan, and H. Yang, “Study on mobile data offloading in high rise building scenario,” in Vehicular
Technology Conference (VTC Spring), 2016 IEEE 83rd.
IEEE, 2016, pp. 1–5.
[16] M. Series, “Guidelines for evaluation of radio interface technologies for imt-advanced,” Report ITU, no. 2135-1, 2009.
[17] T. Imai, K. Kitao, N. Tran, N. Omaki, Y. Okumura, and K. Nishimori, “Outdoor-to-indoor path loss modeling for 0.8 to
37 ghz band,” in Antennas and Propagation (EuCAP), 2016 10th European Conference on.
[18] R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction.
IEEE, 2016, pp. 1–4.
MIT press Cambridge, 1998, vol. 1, no. 1.
[19] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Neural Networks, 1995. Proceedings., IEEE International
Conference on, vol. 4. IEEE, 1995, pp. 1942–1948.
[20] M. Clerc and J. Kennedy, “The particle swarm-explosion, stability, and convergence in a multidimensional complex space,”
IEEE transactions on Evolutionary Computation, vol. 6, no. 1, pp. 58–73, 2002.
[21] R. E. Korf, “A new algorithm for optimal bin packing,” in AAAI/IAAI, 2002, pp. 731–736.
[22] A. Ng, “Cs229 lecture notes.”
| 7cs.IT
|
UNIFYING DAGS AND UGS
arXiv:1708.08722v8 [stat.ML] 1 Mar 2018
JOSE M. PEÑA
IDA, LINKÖPING UNIVERSITY, SWEDEN
[email protected]
Abstract. We introduce a new class of graphical models that
generalizes Lauritzen-Wermuth-Frydenberg chain graphs by relaxing the semi-directed acyclity constraint so that only directed cycles are forbidden. Moreover, up to two edges are allowed between
any pair of nodes. Specifically, we present local, pairwise and global
Markov properties for the new graphical models and prove their
equivalence. We also present an equivalent factorization property.
Finally, we present a causal interpretation of the new models.
1. INTRODUCTION
Lauritzen-Wermuth-Frydenberg chain graphs (LWF CGs) are usually described as unifying directed acyclic graphs (DAGs) and undirected graphs (UGs) (Lauritzen, 1996, p. 53). However, this is arguable
because the only constraint that DAGs and UGs jointly impose is the
absence of directed cycles, whereas LWF CGs forbid semi-directed cycles which is a stronger constraint. Moreover, LWF CGs do not allow
more than one edge between any pair of nodes. In this work, we consider graphs with directed and undirected edges but without directed
cycles. The graphs can have up to two different edges between any pair
of nodes. Therefore, our graphs truly unify DAGs and UGs. Hence,
we call them UDAGs.
As we will see, UDAGs generalize LWF CGs. Two other such generalizations that can be found in the literature are reciprocal graphs
(RGs) (Koster, 1996) and acyclic graphs (AGs) (Lauritzen and Sadeghi,
2017). The main differences between UDAGs and these two classes of
graphical models are the following. UDAGs are not a subclass of RGs
because, unlike RGs, they do not constrain the semi-directed cycles allowed. UDAGs are a subclass of AGs. However, Lauritzen and Sadeghi
define a global Markov property for AGs but no local or pairwise
Markov property. We define the three properties for UDAGs. Lauritzen and Sadeghi
do define though a pairwise Markov property for a subclass of AGs
called chain mixed graphs (CMGs), but no local Markov property.
Moreover, UDAGs are not a subclass of CMGs because, unlike CMGs,
they can have semi-directed cycles. In addition to the local, pairwise
Date: 02:52, 02/03/18, unification8.tex.
1
2
and global Markov properties, we also define a factorization property
for UDAGs. Such a property exists for RGs but not yet for AGs or
CMGs. We also note that the algorithm developed by Sonntag et al.
(2015) for learning LWF CGs from data can easily be adapted to learn
UDAGs (see Appendix A). To our knowledge, there is no algorithm
for learning RGs, AGs or CMGs. Finally, it is worth mentioning that
our work complements that by Richardson (2003), where DAGs and
covariance (bidirected) graphs are unified.
The rest of the paper is organized as follows. Section 2 introduces
some notation and definitions. Sections 3 and 4 present the global, local and pairwise Markov properties for UDAGs and prove their equivalence. Section 5 does the same for the factorization property. Section
6 presents a causal interpretation of UDAGs (which inspires a learning
algorithm that can be found in Appendix B). Section 7 closes the paper
with some discussion.
2. PRELIMINARIES
In this section, we introduce some concepts about graphical models.
Unless otherwise stated, all the graphs and probability distributions
in this paper are defined over a finite set of random variables V . The
elements of V are not distinguished from singletons. An UDAG G is a
graph with possibly directed and undirected edges but without directed
cycles, i.e. A → ⋯ → A is forbidden. There may be up to two different
edges between any pair of nodes. Edges between a node and itself are
not allowed. We denote by A ⊸ B that the edge A → B or A − B or
both are in G.
Given an UDAG G, the parents of a set X ⊆ V are pa(X) = {B∣B →
A is in G with A ∈ X}. The children of X are ch(X) = {B∣A → B is
in G with A ∈ X}. The neighbors of X are ne(X) = {B∣A − B is in
G with A ∈ X}. The ancestors of X are an(X) = {B∣B ⊸ ⋯ ⊸ A is
in G with A ∈ X}. Moreover, X is called ancestral set if X = an(X).
The descendants of X are de(X) = {B∣A ⊸ ⋯ ⊸ B is in G with
A ∈ X}. The sets just defined are defined with respect to G. When
they are defined with respect to another UDAG, this is indicated with
a subscript.
Given an UDAG G, the moral graph of G is the UG Gm such that
A−B is in Gm if and only if A ⊸ B, A → C ← B, or A → C−⋯−D ← B is
in G. Given a set W ⊆ V , we let GW denote the subgraph of G induced
by W . Given an UG H, we let H W denote the marginal subgraph of
H over W , i.e. the edge A − B is in H W if and only if A − B is in H
or A − V1 − ⋯ − Vn − B is H with V1 , . . . , Vn ∉ W . A set of nodes of
H is complete if there exists an undirected edge between every pair of
nodes in the set. A clique of H is a maximal complete set of nodes.
The cliques of H are denoted as cl(H).
3
A route between two nodes V1 and Vn of an UDAG G is a sequence
of (not necessarily distinct) edges E1 , . . . , En−1 in G such that Ei links
the nodes Vi and Vi+1 . A route is called a path if the nodes in the
route are all different. An undirected route is a route whose edges
are all undirected. A section of a route ρ is a maximal undirected
subroute of ρ. A section V2 − ⋯ − Vn−1 of ρ is called collider section if
V1 → V2 − . . . − Vn−1 ← Vn is a subroute of ρ. Given a set Z ⊆ V , ρ is said
to be Z-active if (i) every collider section of ρ has a node in Z, and (ii)
every non-collider section of ρ has no node in Z.
Let X, Y , W and Z be disjoint subsets of V . We represent by
X ⊥ p Y ∣Z that X and Y are conditionally independent given Z in a
probability distribution p. Every probability distribution p satisfies
the following four properties: Symmetry X ⊥ p Y ∣Z ⇒ Y ⊥ p X∣Z, decomposition X ⊥ p Y ∪ W ∣Z ⇒ X ⊥ p Y ∣Z, weak union X ⊥ p Y ∪ W ∣Z ⇒ X ⊥
p Y ∣Z ∪ W , and contraction X ⊥ p Y ∣Z ∪ W ∧ X ⊥ p W ∣Z ⇒ X ⊥ p Y ∪ W ∣Z.
If p is strictly positive, then it also satisfies the intersection property
X ⊥ p Y ∣Z ∪ W ∧ X ⊥ p W ∣Z ∪ Y ⇒ X ⊥ p Y ∪ W ∣Z.
3. GLOBAL MARKOV PROPERTY
Given three disjoint sets X, Y, Z ⊆ V , we say that X is separated
from Y given Z in an UDAG G, denoted as X ⊥ Y ∣Z, if every path
in (Gan(X∪Y ∪Z) )m between a node in X and a node in Y has a node
in Z. As the theorem below proves, this is equivalent to saying that
there is no route in G between a node of X and a node of Y that is Zactive. Note that these separation criteria generalize those developed
by Lauritzen (1996) and Studený (1998) for LWF CGs.
Theorem 1. The two separation criteria for UDAGs in the paragraph
above are equivalent.
Proof. Assume that there is a Z-active route ρ in G between A ∈ X and
B ∈ Y . Clearly, every node in a collider section is in an(Z). Moreover,
every node in a non-collider section is ancestor of A, B or a node in a
collider section, which implies that it is in an(A ∪ B ∪ Z). Therefore,
there is a route between A and B in (Gan(X∪Y ∪Z) )m . Moreover, the
route can be modified into a route ̺ that circumvents Z by noting that
there is an edge V1 −Vn in (Gan(X∪Y ∪Z) )m whenever V1 → V2 −⋯−Vn−1 ←
Vn is a subroute of ρ. The route ̺ can be converted into a path by
removing loops.
Conversely, assume that there is a path ρ in (Gan(X∪Y ∪Z) )m between
A ∈ X and B ∈ Y that circumvents Z. Note that ρ can be converted
into a route ̺ in G as follows: If the edge V1 − Vn in ρ was added to
(Gan(X∪Y ∪Z) )m because V1 ⊸ Vn , V1 ← Vn or V1 → V2 − ⋯ − Vn−1 ← Vn
was in Gan(X∪Y ∪Z) , then replace V1 − Vn with V1 ⊸ Vn , V1 ← Vn or
V1 → V2 − ⋯ − Vn−1 ← Vn , respectively. Note that the non-collider
sections of ̺ have no node in Z for ρ to circumvent Z, whereas the
4
A
C
E
D
F
B
H
Figure 1. Example where non-adjacency does not imply separation.
collider sections of ̺ have all their nodes in an(X ∪Y ∪Z) by definition
of (Gan(X∪Y ∪Z) )m .
Note that we can assume without loss of generality that all the collider sections of ̺ have some node in an(Z) because, otherwise, if there
is a collider section with no node in an(Z) but with some node C in
an(X) then there is a route A′ ⊸ ⋯ ⊸ C with A′ ∈ X which can
replace the subroute of ̺ between A and C. Likewise for an(Y ) and
some B ′ ∈ Y .
Finally, note that every collider section V1 → V2 − ⋯ − Vn−1 ← Vn of ̺
that has no node in Z must have a node Vi in an(Z)∖Z with 2 ≤ i ≤ n−1,
which implies that there is a route Vi ⊸ ⋯ ⊸ C where C is the only
node of the route that is in Z. Therefore, we can replace the collider
section with V1 → V2 − ⋯ − Vi ⊸ ⋯ ⊸ C ⊸ ⋯ ⊸ Vi − ⋯ − Vn−1 ← Vn .
Repeating this step results in a Z-active route between a node in X
and a node in Y .
We say that a probability distribution p satisfies the global Markov
property with respect to an UDAG G if X ⊥ p Y ∣Z for all disjoint sets
X, Y, Z ⊆ V such that X ⊥ Y ∣Z. Note that two non-adjacent nodes in
G are not necessarily separated. For example, C ⊥ D∣Z does not hold
for any Z ⊆ {A, B, E, F, H} in the UDAG in Figure 1. This drawback
is shared by AGs. Although this problem cannot be solved for general
AGs (Lauritzen and Sadeghi, 2017, Figure 6), it can be solved for the
subclass of CMGs by adding edges without altering the separations represented (Lauritzen and Sadeghi, 2017, Corollary 3.1). Unfortunately,
a similar solution does not exist for UDAGs. For example, adding the
edge C ⊸ D to the UDAG in Figure 1 makes A ⊥ B∣D cease holding,
whereas adding the edge C ← D makes A ⊥ B∣C ∪ F cease holding.
Adding two edges between C and D does not help either, since one of
them must be C − D. The following lemma characterizes the problematic pairs of nodes.
Lemma 1. Given two non-adjacent nodes V1 and Vn in an UDAG
G, V1 ⊥ Vn ∣Z does not hold for any Z ⊆ V ∖ (V1 ∪ Vn ) if and only
if V1 → V2 − ⋯ − Vn−1 ← Vn is in G, and Vi ∈ an(V1 ∪ Vn ) for some
1 < i < n.1
1In
the terminology of Lauritzen and Sadeghi (2017), this route is a primitive
inducing walk.
5
A
K
B
C
E
F
I
J
L
M
D
N
Figure 2. Example of UDAG without Markov equivalent LWF CG.
Proof. To prove the if part, assume without loss of generality that
Vi ∈ an(V1 ). This together with the route in the lemma imply that G
has a route ρ of the form V1 ⊸ ⋯ ⊸ Vi − ⋯ − Vn−1 ← Vn . If no node
in Z is in ρ, then V1 ⊥ Vn ∣Z does not hold due to ρ. If C ∈ Z is in
the subroute Vi − ⋯ − Vn−1 ← Vn of ρ, then V1 ⊥ Vn ∣Z does not hold
due to the route in the lemma. Finally, if C ∈ Z is in the subroute
V1 ⊸ ⋯ ⊸ Vi of ρ, then V1 ⊥ Vn ∣Z does not hold due to the route
V1 → V2 − ⋯ − Vi ⊸ ⋯ ⊸ C ⊸ ⋯ ⊸ Vi − ⋯ − Vn−1 ← Vn .
To prove the only if part, simply consider Z = ∅ and note that V1
and Vn are adjacent in (Gan(V1 ∪Vn ) )m only if G has a subgraph of the
form described in the lemma.
Finally, we show that the independence models representable with
UDAGs are a proper superset of those representable with LWF CGs. In
particular, we show that there is no LWF CG that is Markov equivalent
to the UDAG in Figure 2, i.e. there is no LWF CG that represents
exactly the independence model represented by the UDAG. Assume
to the contrary that there is a LWF CG H that is Markov equivalent
to the UDAG G in the figure. First, note that A ⊥ B∣∅ and A ⊥
/ B∣E
imply that H must have an induced subgraph A → E ← B. Likewise,
H must have induced subgraphs C → F ← D, K → I ← L, and M →
J ← N. Next, note that A ⊥ I∣{E, J} implies that H cannot have an
edge E ← I. Likewise, H cannot have an edge F → J. Note also that
A ⊥ F ∣{B, C, D, E, J} and D ⊥ E∣{A, B, C, F, J} imply that H cannot
have an edge E ← F or E → F . Likewise, H cannot have an edge I ← J
or I → J. Consequently, H must have a subgraph of the form
A
K
B
C
E
F
I
J
L
M
D
N
6
This implies that A ⊥ N∣{B, C, D, E, F, I, J, K, L, M} holds in G but
not in H, which is a contradiction.
The following lemma shows that the existence of a semi-directed cycle is not sufficient to declare an UDAG non-equivalent to any LWF
CG. Instead, the semi-directed cycle must occur in a particular configuration, e.g. as in Figure 2. For instance, the lemma implies that
the UDAG A → B ← C − B is Markov equivalent to the LWF CG
A − B − C − A.
Lemma 2. Let G denote an UDAG. If (i) W is an ancestral set of
nodes in G of size greater than one, and (ii) W is minimal with respect
to the property (i), then replacing GW with (GW )m in G results in an
UDAG H that is Markov equivalent to G.
Proof. First, note that H is an UDAG because no directed cycle can
be created by replacing GW with (GW )m in G. Now, consider checking
whether a separation X ⊥Y ∣Z holds in G and H. Consider the following
cases.
(1.) Assume that an(X ∪ Y ∪ Z) in G includes no node in W . Then,
(Gan(X∪Y ∪Z) )m = (Han(X∪Y ∪Z) )m and thus X ⊥ Y ∣Z holds in
both G and H or in none.
(2.) Assume that an(X ∪ Y ∪ Z) in G includes exactly one of the
nodes in W , here denoted by A. Then, an(X ∪ Y ∪ Z) in H
includes all the nodes in W because W is connected in G since,
otherwise, it is not minimal which is a contradiction. Moreover,
note that A is the only node shared by (Gan(X∪Y ∪Z) )m and
(GW )m because, otherwise, there must be a second node in W
that is in an(X ∪ Y ∪ Z) in G, which is a contradiction. Then,
(Han(X∪Y ∪Z) )m = (Gan(X∪Y ∪Z) )m ∪ (GW )m and thus X ⊥ Y ∣Z
holds in both G and H or in none.
(3.) Assume that an(X ∪ Y ∪ Z) in G includes more than one of the
nodes in W . Then, (Gan(X∪Y ∪Z) )m includes all the nodes in W
because, otherwise, W is not minimal which is a contradiction.
Then, (Gan(X∪Y ∪Z) )m = (Han(X∪Y ∪Z) )m and thus X ⊥Y ∣Z holds
in both G and H or in none. To see it, note that by construction
(Gan(X∪Y ∪Z) )m and (Han(X∪Y ∪Z) )m differ only if the former has
an edge V1 −Vn that the latter does not have. This occurs only if
Gan(X∪Y ∪Z) has a subgraph of the form V1 → V2 − ⋯ − Vn−1 ← Vn ,
whereas Han(X∪Y ∪Z) has a subgraph of the form V1 − V2 − ⋯ −
Vn−1 ← Vn or V1 − V2 − ⋯ − Vn−1 − Vn . The former case implies
that V1 , . . . , Vn−1 are in W whereas Vn is not. This contradicts
the fact that W is an ancestral set. The latter case implies that
V1 , . . . , Vn are in W , which implies that V1 − Vn is in H, which
is a contradiction.
7
Note that the condition in the lemma is sufficient but not necessary:
The UDAGs A → B → C and A → B − C are Markov equivalent.
3.1. SEPARATION ALGORITHM. Since there may be infinite
many routes in an UDAG G, one may wonder if the separation criterion
based on ruling out Z-active routes that we have presented above is of
any use in practice. The algorithm below shows how to implement it to
check in a finite number of steps whether X ⊥Y ∣Z holds. The algorithm
is a generalization of the one developed by Studený (1998) for LWF
CGs, which was later slightly improved by Sonntag et al. (2015). The
algorithm basically consists in repeatedly executing some rules to build
the sets U1 , U2 , U3 ⊆ V , which can be described as follows.
● B ∈ U1 if and only if there exists a Z-active route between A ∈ X
and B in G which ends with the subroute Vi → Vi+1 −⋯−Vi+k = B
with k ≥ 1.
● B ∈ U2 if and only if there exists a Z-active route between
A ∈ X and B in G which does not end with the subroute Vi →
Vi+1 − ⋯ − Vi+k = B with k ≥ 1.
● B ∈ U3 if and only if there exists a node C ∈ U1 ∪ U2 and a
route C = V1 → V2 − ⋯ − Vk = B in G with k ≥ 2 such that
{V2 , . . . , Vk } ∩ Z ≠ ∅.
The algorithm starts with U1 = U3 = ∅ and U2 = X. The algorithm
executes the following rules until U1 , U2 and U3 cannot be further
enlarged.
● C ∈ U2 , C ⊸ D is in G, and D ∉ Z ⇒ D ∈ U2 .
● C ∈ U1 ∪ U2 , C → D is in G, and D ∉ Z ⇒ D ∈ U1 .
● C ∈ U1 , C − D is in G, and D ∉ Z ⇒ D ∈ U1 .
● C ∈ U1 ∪ U2 , C → D is in G, and D ∈ Z ⇒ D ∈ U3 .
● C ∈ U1 , C − D is in G, and D ∈ Z ⇒ D ∈ U3 .
● C ∈ U3 , and C − D is in G ⇒ D ∈ U3 .
● C ∈ U3 , C ← D is in G, and D ∉ Z ⇒ D ∈ U2 .
One can prove that, when the algorithm halts, there is a Z-active
route between each node in U1 ∪ U2 and some node in X. The proof is
identical to the one for LWF CGs by Studený (1998, Lemma 5.2) and
Sonntag et al. (2015, Proposition 1). Therefore, X ⊥ Y ∣Z if and only
if Y ⊆ V ∖ (U1 ∪ U2 ). In Appendix A, we use this result to develop an
algorithm for learning UDAGs from data.
4. LOCAL AND PAIRWISE MARKOV PROPERTIES
We say that a probability distribution p satisfies the local Markov
property with respect to an UDAG G if for any ancestral set W ,
A⊥ p W ∖ (A ∪ ne(GW )m (A))∣ne(GW )m (A)
for any A ∈ W . Similarly, we say that a probability distribution p
satisfies the pairwise Markov property with respect to G if for any
8
ancestral set W ,
A⊥ p B∣W ∖ (A ∪ B)
for any A, B ∈ W such that B ∉ ne(GW )m (A).
Theorem 2. Given a probability distribution p satisfying the intersection property, p satisfies the local Markov property with respect to an
UDAG G if and only if it satisfies the pairwise Markov property with
respect to G.
Proof. The if part follows by repeated application of the intersection
property. The only if part follows by the weak union property.
Theorem 3. Given a probability distribution p satisfying the intersection property, p satisfies the pairwise Markov property with respect to
an UDAG G if and only if it satisfies the global Markov property with
respect to G.
Proof. The if part is trivial. To prove the only if part, let W = an(X ∪
Y ∪ Z) and note that the pairwise and global Markov properties are
equivalent for UGs (Lauritzen, 1996, Theorem 3.7).
Note that the local Markov property for LWF CGs specifies a single
independence for each node (Lauritzen, 1996, p. 55). However, the
local Markov property for UDAGs specifies many more independences,
specifically an independence for any node and ancestral set containing
the node. All in all, our local Markov property serves its purpose,
namely to identify a subset of the independences specified by the global
Markov property that implies the rest. In the next section, we show
how to reduce this subset.
4.1. REDUCTION. The number of independences specified by the
local Markov property for UDAGs can be reduced by considering only
maximal ancestral sets for any node A, i.e. those ancestral sets W ′ such
that A ∈ W ′ and ne(GW ′ )m (A) ⊂ ne(GW ′′ )m (A) for any ancestral set W ′′
such that W ′ ⊂ W ′′ . Note that there may be several maximal ancestral
sets W ′ for A, each for a different set ne(GW ′ )m (A) as will be shown.
The independences for the non-maximal ancestral sets follow from the
independences for the maximal ancestral sets by the decomposition
property. In other words, for any non-maximal ancestral set W and
A ∈ W,
A⊥ p W ∖ (A ∪ ne(GW )m (A))∣ne(GW )m (A)
follows from
A⊥ p W ′ ∖ (A ∪ ne(GW ′ )m (A))∣ne(GW ′ )m (A)
where W ′ is the maximal ancestral set for A such that ne(GW )m (A) =
ne(GW ′ )m (A). In the UDAG in Figure 3, for instance, W1 = {A, B, C, D},
W2 = {A, B, C, D, E, I, J, K}, and W3 = {A, B, C, D, E, F, H, I, J, K}
are three ancestral sets that contain the node B. However, only W2
9
A
B
C
F
H
I
D
J
E
K
Figure 3. Example where the local Markov property
can be improved by considering only maximal ancestral
sets.
and W3 are maximal for B: W1 is not maximal because W1 ⊂ W2 but
ne(GW1 )m (B) = ne(GW2 )m (B), and W2 is maximal because W2 ⊂ W3 and
ne(GW2 )m (B) ⊂ ne(GW3 )m (B). Note that W1 specifies B ⊥ p D∣{A, C},
and W2 specifies B ⊥ p {D, E, I, J, K}∣{A, C}. Clearly, the latter independence implies the former by the decomposition property. Therefore,
there is no need to specify both independences, as the local Markov
property does. It suffices to specify just the second.
A more convenient characterization of maximal ancestral sets is the
following. An ancestral set W ′ is maximal for A ∈ W ′ if and only if W ′ =
V ∖ [(ch(A) ∪ de(ch(A))) ∖ W ′ ]. To see it, note that B ∈ ne(GW ′ )m (A)
if and only if A ⊸ B, A → B, or A → C − ⋯ − D ← B is in GW ′ .
Note that all the parents and neighbors of A are in W ′ , because W ′
is ancestral. However, if there is some child B of A that is not in W ′ ,
then any ancestral set W ′′ that contains W ′ and B or any node that is
a descendant of B will be such that ne(GW ′ )m (A) ⊂ ne(GW ′′ )m (A).
The number of independences specified by the pairwise Markov property can also be reduced by considering only maximal ancestral sets.
This can be proven in the same way as Theorem 2.
5. FACTORIZATION PROPERTY
Theorem 4. Given a probability distribution p satisfying the intersection property, p satisfies the pairwise Markov property with respect to
an UDAG G if and only if for any ancestral set W ,
p(W ) =
∏
ϕ(K)
K∈cl((GW )m )
where ϕ(K) is a non-negative function.
Proof. It suffices to recall the equivalence between the pairwise Markov
property and the factorization property for UGs (Lauritzen, 1996, Theorem 3.9).
5.1. REDUCTION. The number of factorizations specified by the
factorization property for UDAGs can be reduced by considering only
maximal ancestral sets, i.e. those ancestral sets W ′ such that (GW ′ )m
′
is a proper subgraph of ((GW ′′ )m )W for any ancestral set W ′′ such
that W ′ ⊂ W ′′ . These maximal ancestral sets do not necessarily coincide with the ones defined in Section 4.1. The factorizations for
10
A
B
C
D
E
F
H
I
Figure 4. Example where the factorization property
can be improved by considering only maximal ancestral
sets.
the non-maximal ancestral sets follow from the factorizations for the
maximal ancestral sets. To see it, note that for any non-maximal ancestral set W , the probability distribution p(W ) can be computed
by marginalization from p(W ′ ) where W ′ is any maximal ancestral
set such that ((GW ′ )m )W = (GW )m . Note also that p(W ) factorizes
according to ((GW ′ )m )W and thus according to (GW )m , by Studený
(1997, Lemma 3.1) and Lauritzen (1996, Theorems 3.7 and 3.9). In the
UDAG in Figure 4, for instance, W1 = {A, B}, W2 = {A, B, C, D, E},
and W3 = {A, B, C, D, E, F, H, I} are three ancestral sets. However,
only W1 and W3 are maximal: W2 is not maximal because W2 ⊂ W3
but ((GW3 )m )W2 = (GW2 )m , and W1 is maximal because W1 ⊂ W3 and
(GW1 )m is a proper subgraph of ((GW3 )m )W1 . Note that W3 specifies
p(W3 ) =ϕ3 (A, B)ϕ3 (A, C)ϕ3 (B, E)ϕ3 (C, D)ϕ3 (D, E)
⋅ ϕ3 (A, F )ϕ3 (B, I)ϕ3 (F, H)ϕ3 (H, I)
and W2 specifies
p(W2 ) = ϕ2 (A, B)ϕ2 (A, C)ϕ2 (B, E)ϕ2 (C, D)ϕ2 (D, E).
Clearly, the former factorization implies the latter by taking
ϕ2 (A, B) =ϕ3 (A, B) ∑ ϕ3 (A, F )ϕ3 (B, I)ϕ3 (F, H)ϕ3 (H, I)
F,H,I
ϕ2 (A, C) =ϕ3 (A, C)
ϕ2 (B, E) =ϕ3 (B, E)
ϕ2 (C, D) =ϕ3 (C, D)
ϕ2 (D, E) =ϕ3 (D, E).
Therefore, there is no need to specify both factorizations, as the factorization property does. It suffices to specify just the first.
A more convenient characterization of maximal ancestral sets is the
following. An ancestral set W ′ is maximal if and only if pa(A∪an(A)∖
W ′ ) ∩ W ′ is not a complete set in (GW ′ )m for any node A ∈ V ∖ W ′ .2
To see it, note that any ancestral set W ′′ that contains W ′ ∪ A will
2In
the terminology of Frydenberg (1990a), A ∪ an(A) ∖ W ′ is a non-simplicial
set in (GW ′ )m .
11
also contain an(A) ∖ W ′ . Note also that no node B ∈ A ∪ an(A) ∖ W ′
has a neighbor or child in W ′ because, otherwise, B ∈ W ′ which is
a contradiction. So, any such node B can only have parents in W ′ .
Moreover, since pa(A ∪ an(A) ∖ W ′ ) ∩ W ′ is not a complete set in
(GW ′ )m , there must be two nodes in pa(A ∪ an(A) ∖ W ′ ) ∩ W ′ that are
not adjacent in (GW ′ )m . However, there is a path between these two
nodes in (GW ′′ )m through B, which implies that (GW ′ )m is a proper
′
subset of ((GW ′′ )m )W .
6. CAUSAL INTERPRETATION
In this section, we propose a causal interpretation of UDAGs. We
start by introducing some notation. Given an UDAG G, let W1 , . . . , Wt
denote all the minimal ancestral sets in G. Assume that the sets are
sorted such that if Wi ⊂ Wj then i < j. Moreover, let Ci = Wi ∖
∪j<i Wj . Note that all the edges between a node in Ci and a node
in Cj with i < j are directed edges from Ci to Cj . Note also that
every node in Ci is an ancestor of the rest of the nodes in Ci . Let
bd(Ci ) = pa(Ci )∖Ci . Moreover, let (GCi ∪bd(Ci ) )∗ be the result of adding
undirected edges to (GCi ∪bd(Ci ) )m until bd(Ci ) is a complete set. Note
that for LWF CGs, the sets Ci correspond to the chain components,
bd(Ci ) = pa(Ci ), and (GCi ∪bd(Ci ) )∗ = (GCi ∪bd(Ci ) )m . For instance, in the
UDAG in Figure 1 we have that W1 = {A}, W2 = {B}, W3 = {B, D}
and W4 = {A, B, C, D, E, F, H}, and C1 = {A}, C2 = {B}, C3 = {D}
and C4 = {C, E, F, H}, and bd(C1 ) = ∅, bd(C2 ) = ∅, bd(C3 ) = {B} and
bd(C4 ) = {A, D}.
The following theorem presents a factorization property for UDAGs.
Compared to that in Theorem 4, it is simpler and resembles the factorization property for LWF CGs. However, it is necessary but not
sufficient. It will be instrumental to derive our causal interpretation of
UDAGs.
Theorem 5. Let p be a probability distribution satisfying the intersection property. If p satisfies the pairwise Markov property with respect
to an UDAG G, then
p(V ) = ∏ p(Ci ∣bd(Ci )) = ∏
i
∏
ϕ(K)
i K∈cl((GC ∪bd(C ) )∗ )
i
i
where ϕ(K) is a non-negative function.
Proof. The first equality follows from the fact that Ci ⊥∪j<i Cj ∖bd(Ci )∣bd(Ci ).
To prove the second equality for i = t, note that p satisfies the pairwise Markov property with respect to Gm , because V is an ancestral
set. Then, p satisfies the global Markov property with respect to Gm
by Theorem 3. Now, add undirected edges to Gm until bd(Ct ) is a
complete set, and call the resulting undirected graph H. Note that
p satisfies the global Markov property with respect to H. Note also
12
that (GCt ∪bd(Ct ) )∗ = HCt ∪bd(Ct ) . Then, p(Ct , bd(Ct )) satisfies the global
Markov property with respect to HCt ∪bd(Ct ) (Frydenberg, 1990b, Proposition 2.2). This implies the second equality in the theorem because
(i) p(Ct , bd(Ct )) = ∏K∈cl((GC ∪bd(C ) )∗ ) φ(K) (Lauritzen, 1996, Theorem
i
i
3.9), (ii) p(Ct ∣bd(Ct )) = p(Ct , bd(Ct ))/p(bd(Ct )), and (iii) bd(Ct ) is a
complete set in (GCt ∪bd(Ct ) )∗ . Finally, note that V ∖ Ct is an ancestral
set and, thus, p(V ∖ Ct ) satisfies the pairwise Markov property with
respect to GV ∖Ct . Then, repeating the reasoning above for p(V ∖ Ct )
and GV ∖Ct proves the second equality in the theorem for all i.
For instance, in the UDAG in Figure 1 we have that
p(V ) =p(A)p(B)p(D∣B)p(CEF H∣AD)
=ϕ(A)ϕ(B)ϕ(DB)ϕ(CAD)ϕ(CF A)ϕ(CEF )ϕ(F H)ϕ(HD).
Our causal interpretation of UDAGs is a generalization of the one
proposed by Lauritzen and Richardson (2002) for LWF CGs. Specifically, they show that any probability distribution that satisfies the
globally Markov property with respect to a LWF CG coincides with
the equilibrium distribution of a dynamic system with feed-back loops.
The proof consists in building a Gibbs sampler with the desired equilibrium distribution. The sampler samples the chain components in
the order C1 , . . . , Ct . Sampling a component Ci consists in repeatedly
updating the values of the variables A ∈ Ci in random order according
to the distribution p(A∣bd(Ci ), Ci ∖A) until equilibrium is reached. The
interesting observation is that
p(A∣bd(Ci ), Ci ∖ A) = p(A∣bd(Ci ), ne(A))
and thus the sampling process can be seen as a dynamic process with
feed-back loops, since A is dynamically affected by ne(A) and vice
versa. Thanks to the first equality in Theorem 5, the causal interpretation just discussed can be generalized to UDAGs if
p(A∣bd(Ci ), Ci ∖ A) = p(A∣bd(Ci ), pa(A), ne(A))
that is, if Ci ∖ (A ∪ pa(A) ∪ ne(A)) carry no information about A given
bd(Ci ) ∪ pa(A) ∪ ne(A). One case where this may hold is when the
causal relationships in the domain are constrained in their functional
form, e.g. the effect is a function of the cause plus some noise, also
known as additive noise model (ANM). If the function is non-linear,
then it is unlikely that the cause can be expressed as an ANM of the
effect (Hoyer et al., 2009; Peters et al., 2014). As a consequence, we
expect any other variable to be hardly informative about the effect
given the cause. Specifically, we expect Ci ∖ (A ∪ pa(A) ∪ ne(A)) to
be hardly informative about A given bd(Ci ) ∪ pa(A) ∪ ne(A), because
we interpret the latter variables as the causes of A. So, our causal
interpretation of UDAGs should approximately hold under the ANM
13
assumption. We refer the reader to Appendix B for an algorithm for
learning causal UDAGs and some preliminary experimental results.
7. DISCUSSION
We have introduced UDAGs, a new class of graphical models that
unifies DAGs and UGs since it just forbids directed cycles and it allows up to two edges between any pair of nodes. We have presented
local, pairwise and global Markov properties for UDAGs and proved
their equivalence. We have also presented an equivalent factorization
property. Finally, we have presented a causal interpretation of UDAGs.
We refer the reader to the appendices for two learning algorithms for
UDAGs and preliminary experimental results.
A natural question to tackle in the future is the characterization of
Markov equivalent UDAGs. Although we have shown that UDAGs are
a strict superclass of LWF CGs, it is unclear how much more expressive
they are. Addressing this question is also of much interest. Finally, we
are also interested in studying methods for parameterizing the factorization for UDAGs proposed here, as well as in improving the causal
interpretation of UDAGs given here.
References
A. Biere, M. Heule, H. van Maaren, and T. Walsh. Handbook of Satisfiability. IOS Press, 2009.
M. Frydenberg. Marginalization and Collapsibility in Graphical Interaction Models. Annals of Statistics, 18:790–805, 1990a.
M. Frydenberg. The Chain Graph Markov Property. Scandinavian
Journal of Statistics, 17:333–353, 1990b.
M. Gebser, B. Kaufmann, R. Kaminski, M. Ostrowski, T. Schaub, and
M. Schneider. Potassco: The Potsdam Answer Set Solving Collection. AI Communications, 24:107–124, 2011.
M. Gelfond and V. Lifschitz. The Stable Model Semantics for Logic
Programming. In Proceedings of 5th Logic Programming Symposium,
pages 1070–1080, 1988.
A. Gretton, K. Fukumizu, C. H. Teo, L. Song, B. Schölkopf, and A. J.
Smola. A Kernel Statistical Test of Independence. In Advances in
Neural Information Processing Systems 20, pages 585–592, 2008.
P. O. Hoyer, D. Janzing, J. Mooij, J. Peters, and B. Schölkopf. Nonlinear Causal Discovery with Additive Noise Models. In Advances in
Neural Information Processing Systems 21, pages 689–696, 2009.
J. T. A. Koster. Markov Properties of Nonrecursive Causal Models.
Annals of Statistics, 24:2148–2177, 1996.
S. L. Lauritzen. Graphical Models. Oxford University Press, 1996.
S. L. Lauritzen and T. S. Richardson. Chain Graph Models and Their
Causal Interpretations. Journal of the Royal Statistical Society B,
64:321–348, 2002.
14
S. L. Lauritzen and K. Sadeghi. Unifying Markov Properties for Graphical Models. Annals of Statistics, 2017.
J. M. Mooij, J. Peters, D. Janzing, J. Zscheischler, and B. Schölkopf.
Distinguishing Cause from Effect Using Observational Data: Methods and Benchmarks. Journal of Machine Learning Research, 17:
1–102, 2016.
I. Niemelä. Logic Programs with Stable Model Semantics as a Constraint Programming Paradigm. Annals of Mathematics and Artificial Intelligence, 25:241–273, 1999.
J. Pearl. Causality: Models, Reasoning, and Inference. Cambridge
University Press, 2009.
J. Peters, J. M. Mooij, D. Janzing, and B. Schölkopf. Identifiability of
Causal Graphs using Functional Models. In Proceedings of the 27th
Conference on Uncertainty in Artificial Intelligence, pages 589–598,
2011.
J. Peters, J. M. Mooij, D. Janzing, and B. Schölkopf. Causal Discovery with Continuous Additive Noise Models. Journal of Machine
Learning Research, 15:2009–2053, 2014.
J. Peters, D. Janzing, and B. Schölkopf. Elements of Causal Inference:
Foundations and Learning Algorithms. The MIT Press, 2017.
C. E. Rasmussen and C. K. I. Williams. Gaussian Processes for Machine Learning. The MIT Press, 2005.
T. Richardson. Markov Properties for Acyclic Directed Mixed Graphs.
Scandinavian Journal of Statistics, 30:145–157, 2003.
P. Simons, I. Niemelä, and T. Soininen. Extending and Implementing the Stable Model Semantics. Artificial Intelligence, 138:181–234,
2002.
D. Sonntag, M. Järvisalo, J. M. Peña, and A. Hyttinen. Learning Optimal Chain Graphs with Answer Set Programming. In Proceedings of
the 31st Conference on Uncertainty in Artificial Intelligence, pages
822–831, 2015.
M. Studený. On Marginalization, Collapsibility and Precollapsibility.
In Distributions with Given Marginals and Moment Problems, pages
191–198, 1997.
M. Studený. Bayesian Networks from the Point of View of Chain
Graphs. In Proceedings of the 14th Conference on Uncertainty in
Artificial Intelligence, pages 496–503, 1998.
M. Studený.
Probabilistic Conditional Independence Structures.
Springer, 2005.
15
APPENDIX A: ALGORITHM FOR LEARNING UDAGS
In this appendix, we describe an exact algorithm for learning UDAGs
from data via answer set programming (ASP). The algorithm builds on
the results in Section 3.1 and it is essentially the same as the one developed by Sonntag et al. (2015) for learning LWF CGs. ASP is a declarative constraint satisfaction paradigm that is well-suited for solving
computationally hard combinatorial problems (Gelfond and Lifschitz,
1988; Niemelä, 1999; Simons et al., 2002). ASP represents constraints
in terms of first-order logical rules. Therefore, when using ASP, the first
task is to model the problem at hand in terms of rules so that the set
of solutions implicitly represented by the rules corresponds to the solutions of the original problem. One or multiple solutions to the original
problem can then be obtained by invoking an off-the-shelf ASP solver
on the constraint declaration. Each rule in the constraint declaration
is of the form head :- body. The head contains an atom, i.e. a fact.
The body may contain several literals, i.e. negated and non-negated
atoms. Intuitively, the rule is a justification to derive the head if the
body is true. The body is true if its non-negated atoms can be derived,
and its negated atoms cannot. A rule with only the head is an atom. A
rule without the head is a hard-constraint, meaning that satisfying the
body results in a contradiction. Soft-constraints are encoded as rules
of the form :~ body. [W], meaning that satisfying the body results
in a penalty of W units. The ASP solver returns the solutions that
meet the hard-constraints and minimize the total penalty due to the
soft-constraints. A popular ASP solver is clingo (Gebser et al., 2011),
whose underlying algorithms are based on state-of-the-art Boolean satisfiability solving techniques (Biere et al., 2009).
Table 1 shows the ASP encoding of our learning algorithm. The
input to the algorithm is the set of independences in the probability
distribution at hand, e.g. as determined from some available data.
These are represented as a set of predicates ind(A,B,Z) indicating
that the nodes A and B are independent given the set of nodes Z.
It is known that these pairwise independences (also called elementary
triplets) uniquely identify the rest of independences in the distribution,
or in a semi-graphoid for that matter (Studený, 2005, Lemma 2.2). The
predicates node(A) and set(Z) represent that A is the index of a node
and Z is the index of a set of nodes. The predicates line(A,B) and
arrow(A,B) represent that there is an undirected and directed edge
from the node A to the node B. The rules 4 and 5 encode a nondeterministic guess of the edges, which means that the ASP solver
will implicitly consider all possible UDAGs during the search, hence
the exactness of the search. The rules 6 and 7 enforce the fact that
undirected edges are symmetric and there is at most one directed edge
between two nodes. The predicate ancestor(A,B) represents that the
16
node A is an ancestor of the node B. The rules 8-10 enforce that there
are no directed cycles. The predicates in the rules 11 and 12 represent
whether a node A is or is not in a set of nodes Z. The rules 13-23
encode the separation criterion for UDAGs as it was described in Section 3.1. Specifically, the predicate inU1(A,D,Z) represents that there
is a Z-active route from the node A to the node D that warrants the
inclusion of D in the set U1 . Similarly for the predicates inU2(A,D,Z)
and inU3(A,D,Z). The predicate act(A,B,Z) in the rules 24 and 25
represents that there is a Z-active route between the node A and the
node B. The rule 26 enforces that each dependence in the input must
correspond to an active route. The rules 27 and 28 represent a penalty
of one unit per edge. Other penalty rules can be added similarly. By
calling the ASP solver, the solver will essentially perform an exhaustive search over the space of UDAGs and return the sparsest minimal
independence map.
It is worth noting that the algorithm in Table 1 can easily be modified to learn DAGs and LWF CGs, which demonstrates the versatility of our approach. Specifically, learning DAGs can be performed
by adding :- line(A,B). Learning LWF CGs can be performed by
adding :- line(A,B), arrow(A,B)., :- line(A,B), arrow(B,A).,
and ancestor(A,B) :- line(A,B).
Finally, preliminary experiments indicate that the algorithm in Table
1 runs in acceptable time in a regular computer for up to seven nodes.
To scale the learning process to larger domains, one may need to give
up either the exactness or the assumption free nature of the algorithm
here proposed.
17
Table 1. Algorithm for learning UDAGs.
% input predicate: ind(A,B,Z), the nodes A and B are independent given
%
the set of nodes Z
#const n=7.
node(1..n).
set(0..(2**n)-1).
% edges
{ line(A,B) } :- node(A), node(B), A != B.
{ arrow(A,B) } :- node(A), node(B), A != B.
line(A,B) :- line(B,A).
:- arrow(A,B), arrow(B,A).
% directed acyclity
ancestor(A,B) :- arrow(A,B).
ancestor(A,B) :- ancestor(A,C), ancestor(C,B).
:- ancestor(A,B), arrow(B,A).
% set membership
in(A,Z) :- node(A), set(Z), 2**(A-1) & Z != 0.
out(A,Z) :- node(A), set(Z), 2**(A-1) & Z = 0.
% rules
inU2(A,A,Z)
inU2(A,D,Z)
inU2(A,D,Z)
inU1(A,D,Z)
inU1(A,D,Z)
inU1(A,D,Z)
inU3(A,D,Z)
inU3(A,D,Z)
inU3(A,D,Z)
inU3(A,D,Z)
inU2(A,D,Z)
:::::::::::-
node(A), set(Z), out(A,Z).
inU2(A,C,Z), arrow(D,C), out(D,Z).
inU2(A,C,Z), line(D,C), out(D,Z).
inU1(A,C,Z), arrow(C,D), out(D,Z).
inU2(A,C,Z), arrow(C,D), out(D,Z).
inU1(A,C,Z), line(C,D), out(D,Z).
inU1(A,C,Z), arrow(C,D), in(D,Z).
inU2(A,C,Z), arrow(C,D), in(D,Z).
inU1(A,C,Z), line(C,D), in(D,Z).
inU3(A,C,Z), line(C,D).
inU3(A,C,Z), arrow(D,C), out(D,Z).
% active routes
act(A,B,Z) :- inU1(A,B,Z), A != B.
act(A,B,Z) :- inU2(A,B,Z), A != B.
% satisfy all the dependences
:- not ind(A,B,Z), not act(A,B,Z), node(A), node(B), set(Z), A != B,
out(A,Z), out(B,Z).
% minimize the number of lines/arrows
:~ line(A,B), A < B. [1,A,B,1]
:~ arrow(A,B). [1,A,B,2]
% rule 4
% rule 6
% rule 8
% rule 11
% rule 13
% rule 24
% rule 26
% rule 27
% show results
#show.
#show arrow(A,B) : arrow(A,B).
#show line(A,B) : line(A,B), A < B.
APPENDIX B: ALGORITHM FOR LEARNING CAUSAL
UDAGS
In Section 6, we have proposed a causal interpretation of UDAGs
under the assumption that it approximately holds that
p(A∣bd(Ci ), Ci ∖ A) = p(A∣bd(Ci ), pa(A), ne(A))
18
Table 2. Algorithm for learning causal UDAGs.
Input: A dataset D over V with M observations, and an integer L.
Output: A causal UDAG over V .
1 Let G be a random sample of L UDAGs over V
2 For each G in G
3
For each A ∈ V
4
fˆ(A∣pa(A), ne(A)) = GP (A, pa(A), ne(A), D)
5
For m = 1, . . . , M
6
ÛAm = Am − fˆ(Am ∣pam (A), nem (A))
7
Let DU denote the dataset over U created in the previous line
8
pvalue(G) = HSIC(DU )
9 Return the simplest model in the set arg maxG∈G pvalue(G)
that is, Ci ∖ (A ∪ pa(A) ∪ ne(A)) are hardly informative about A given
bd(Ci ) ∪ pa(A) ∪ ne(A). Note that any (pa(A) ∪ ne(A))-active route
between a node in bd(Ci ) ∖ pa(A) and A reaches A through a node in
(ch(A) ∩ Ci ) ∖ ne(A). Since we have assumed that Ci ∖ (A ∪ pa(A) ∪
ne(A)) are hardly informative about A given bd(Ci ) ∪ pa(A) ∪ ne(A),
so are (ch(A) ∩ Ci ) ∖ ne(A). Then, we expect bd(Ci ) ∖ pa(A) to be
hardly informative about A given pa(A) ∪ ne(A) and, thus, it should
approximately hold that
p(A∣bd(Ci ), Ci ∖ A) = p(A∣pa(A), ne(A)).
So, learning a causal UDAG boils down to learning the causes pa(A)
and ne(A) of each variable A under the ANM constraint. Additive
noise is a rather common assumption in causal discovery (Peters et al.,
2017), mainly because it produces tractable models which are useful
for gaining insight into the system under study. Note also that linear
structural equation models, which have extensively been studied for
causal effect identification (Pearl, 2009), are ANMs.
In this appendix, we propose an algorithm for learning causal UDAGs
under the assumptions discussed above. The algorithm builds on the
ideas by Hoyer et al. (2009) and Peters et al. (2011, 2014), who exploit
the non-linearities in the data to identify the directions of the causal
relationships. Specifically, consider two variables A and B that are
causally related as B = f (A) + UB . Assume that there is no confounding, selection bias or feed-back loop between A and B, which implies
that A and UB are independent. Hoyer et al. (2009) prove that if the
function f is non-linear, then the correct direction of the causal relationship between A and B is generally identifiable from observational
19
data: A and UB are independent for the correct direction, whereas B
and UA are dependent for the incorrect direction. This leads to the
following causal discovery algorithm. If A and B are independent then
they are not causally related because we have assumed no confounding, selection bias or feed-back loop. If they are dependent then first
construct a non-linear regression of B on A to get an estimate fˆ of
f , then compute the error ÛB = B − fˆ(A), and finally test whether A
and ÛB are independent. If they are so then accept the model A → B.
Repeat the procedure for the model B → A. When both models or
no model is accepted, it may be indicative that the assumptions do
not hold. Peters et al. (2011, 2014) generalize this idea to more than
two variables: Given a DAG over some variables, first construct a nonlinear regression of each node on its parents, then compute each node’s
error, and finally test whether these errors are mutually independent.
If they are so then accept the DAG as the true causal model. We
propose to generalize this idea even further: Given an UDAG over
some variables, first construct a non-linear regression of each node on
its parents and neighbors, then compute each node’s error, and finally
test whether these errors are mutually independent. If they do so then
accept the UDAG as the true causal model. This leads to the learning algorithm in Table 2. It receives as input a dataset D with M
observations over the random variables V , and an integer L. The algorithm consists in sampling L random UDAGs over V (line 1), scoring
each of them with respect to D (lines 2-8), and returning the best one
(line 9). Scoring an UDAG G starts in the lines 3-4 pretty much like
the algorithms by Hoyer et al. (2009) and Peters et al. (2011, 2014),
i.e. obtaining an estimate fˆ of f by constructing a non-linear regression of each node A on pa(A) ∪ ne(A) using Gaussian processes (GPs)
(Rasmussen and Williams, 2005). This estimate is used in the lines 5-7
to compute the errors. We use a superscript to indicate the value of
a set of variables in a particular instance of D, i.e. Am and pam (A)
and nem (A) represent the value of A and its parents and neighbors
in the m-th instance of D. Finally, the line 8 scores the whole model
by testing the independence of the errors. The null hypothesis is joint
independence. Specifically, the function HSIC returns the p-value of
the Hilbert Schmidt independence criterion, which is a kernel statistical
test of independence (Gretton et al., 2008). Note that several UDAGs
may score the highest p-value, e.g. every supergraph of an UDAG with
the highest p-value may also receive the highest p-value. Therefore,
the line 9 applies the Occam’s Razor principle and returns the simplest
best UDAG. Using GPs and the HSIC test are choices shared with
Hoyer et al. (2009) and Peters et al. (2011, 2014). Other choices are
also possible. We have implemented our learning algorithm in R. We
use the packages kernlab and dHSIC for the GPs and the HSIC test.
20
Altitude
Sunshine duration
T emperature
P recipitation
Figure 5. Causal UDAG learned from the DWD dataset.
For the GPs, we use the Gaussian kernel with automatic width estimation. For the HSIC test, we use the gamma distribution approximation
to the null distribution of the test statistic.3
The rest of this appendix reports on preliminary results obtained
by running the learning algorithm above on the DWD dataset, which
contains climate data from the German Weather Service and has been
used before for benchmarking causal discovery algorithms (Mooij et al.,
2016; Peters et al., 2014). We use the data provided by the first reference.4 The data consists of 349 instances, each corresponding to a
weather station in Germany. Each instance consists of measurements
for six random variables. We use only four of them so that the number
of UDAGs is manageable and, thus, our learning algorithm has a chance
to test most if not all of them. Specifically, there are 34752 UDAGs
over four nodes (543 DAGs times 64 UGs), and we let the algorithm
sample 50000 in line 1. The four random variables that we consider
are altitude (A), temperature (T ), precipitation (P ), and sunshine duration (S). The last three variables represent annual mean values over
the years 1961-1990. Mooij et al. argue that the causal relationships
A → T , A → P and A → S are true. Their arguments are meteorological, i.e. not based on the data. We confirmed that these decisions
make sense according to Wikipedia (entries for the terms ”rain” and
”precipitation”).
Figure 5 shows the causal UDAG learned, which is in fact a LWF
CG. This UDAG is clearly preferred (p-value = 0.0007) over an UDAG
with only the three ground truth relationships (p-value = 5.4e-40). The
p-values should be interpreted with caution. Their relative values are
informative. However, their absolute values may not, because low pvalues may be the result of the GPs underfitting the data. Of course,
low p-values may also be indicative of the inadequacy of the ANM
assumption. It is also worth mentioning that the best and second best
UDAGs found by our algorithm are Markov equivalent but they receive
different p-values (0.0007 versus 0.0004) since they represent different
causal models.
Now, we argue that the UDAG learned is plausible. The relationships A → T and A → P are confirmed by both Mooij et al. and
3Code available at https://www.dropbox.com/s/fuz4eow8f66omda/UDAGs.R?dl=0.
4Data available at https://www.dropbox.com/s/rs4q8oeutgqfcdn/D1.csv?dl=0.
21
Wikipedia. The relationship S → T seems natural. We can also think
of an intervention where we install new suns. We expect that the more
suns the warmer. The relationship A − S is confirmed by Mooij et al.
due to selection bias: All the mountains in Germany are in the south
and the south is typically sunnier. Note that this selection bias is at
odds with the ANM assumption and, thus, one should not interpret the
relationship A − S as a feed-back loop. The relationship A → S is also
confirmed by Mooij et al. However, our algorithm is unable to learn an
UDAG with a subgraph A → S − A, because removing the edge A → S
results in an UDAG with the same score, which is preferred by our
algorithm because it is simpler. The relationship T − P confirms the
complex (including possibly feed-back) interplay between temperature
and precipitation. According to Wikipedia, rain is produced by the
condensation of atmospheric water vapor. Therefore, increasing water
vapor in the air and/or decreasing the temperature are the main causes
of precipitation. One way water vapor gets added to the air is due to
increased temperature, causing evaporation from the surface of oceans
and other water bodies. Moreover, precipitation typically causes a decrease in temperature, as rain drops form at high altitude where it is
colder. The relationship S → P is unconfirmed.
Finally, we mention some additional experiments that we plan to
carry out. As discussed above, the best UDAG found is plausible.
However, the second best UDAG receives a relatively close score (0.0007
versus 0.0004) but it is less plausible, as it includes the edge P → A.
That the learning algorithm does not discriminate better these two
models may be due to the inadequacy of the ANM assumption, but not
necessarily. It may be that we have to choose carefully the width of
the Gaussian kernel in the GPs, or consider other kernels, or consider
other non-linear regressors, or consider the exact permutation-based
null distribution of the HSIC test, or consider alternative hypothesis
tests. We plan to study all these possibilities. In our experiments, the
UDAG returned is actually a LWF CG. We expect UDAGs to reveal
all their potential in larger and more complex domains. However, our
brute-force learning algorithm does not scale well. That is why we plan
to develop a greedy hill-climbing version of the algorithm that evaluates
all the models that differ from the current one by one edge and then
moves to the best of them. Note that we do not have to compute the
score from scratch for each candidate model to evaluate, as at most
two nodes are affected by a single edge modification.
| 2cs.AI
|
Maximum Likelihood Estimations Based on Upper Record Values for
Probability Density Function and Cumulative Distribution Function in
Exponential Family and Investigating Some of Their Properties
S. D. Gore1, S. Hosseini2, P. Nasiri3
Abstract
In this paper a useful subfamily of the exponential family has been considered. The ML estimation based
on upper record values has been calculated for the parameter, Cumulative Density Function, and
Probability Density Function of the family. Also, the relations between MLE based on record values and a
random sample has been discussed. Additionally, some properties of these estimators has been
investigated. Finally, it has been proven that these estimators have some useful properties for samples
with large size.
Keywords: Exponential family, Record values, ML Estimation, Asymptotically Unbiasedness.
1. Introduction:
An exponential family includes a wide range of statistical distributions in two discrete and continuous
states, that has a large importance in the distribution theory, and with which it might be possible to
integrally investigate many properties of the distribution. Different forms have been presented for this
family in the one-parametric state. Generally, a one-variable exponential family is a group of distributions
that their probability density function is as follows:
f ( x; ) h( x) g ( )exp{ ( )T ( x)}
Here, it is considered an important subfamilies from the exponential family.
F ( x; ) 1 exp{ B( ) A( x)},
in which A is an increasing function and
B( ) 0; x [a, b]; A(a) 0, A(b) ; a, b R
It is called the first type exponential family for ease in this paper. Many one-parametric continuous
distributions can be studied by choosing the above general form. The following tables describe this
subject:
1
Professor at department of statistics, Savirtibai Phule Pune University, India
Correspondent author and lecturer at department of accounting, Cihan University of Erbil, Iraq. Email:
[email protected]
3
Associate professor at Payam Noor University of Tehran
2
Table 1:Instances for the exponential family of the first type
A( x)
B( )
A( x) x
A( x) Log (1 x)
A( x) x
A( x) Ln( x)
B( )
1
Distribution Name
Exponential
Lomax
B( )
Weibull
B( )
Pareto
Other distributions, which belong to the mentioned family, may be found by more investigations.
Therefore, the reason for choosing the above general family lies in its wideness and comprehensiveness.
Probability density function (PDF) diagram for the distributions in Table.1 and their cumulative
distribution function (CDF) diagram are shown in Figures 1.1 and 1.2, respectively.
Figure 1.1
Figure 1.2
Determination and investigation of the properties of the maximum likelihood estimations based on record
values for probability density function and cumulative distribution function in the given exponential
family is the goal of this paper; thus the maximum likelihood estimations for the unknown parameters of
the exponential family, probability density function, and cumulative distribution function of the
exponential family have been generally obtained based on a random sample and also values of upper
records. In addition, validity of maximum likelihood estimators based on record values in different states
has been presented along with a series of examples. Likewise, the criterion of Mean Square Error (MSE)
has been considered in order to investigate and compare them. Additionally, the properties of these
estimators in asymptotic conditions are interested by this paper. Chandler suggested for the first time the
concept of records under a mathematical model (Chandler, 1952). During a very short period after him,
this concept was changed and completed by many statisticians and the theory of records was formed
which has very wide applications in analysis of competitive phenomena such as commerce and
economics. After 21 years from the introduction of Chandler to this discussion, a wide range of researches
and other works were performed in this field, and some interesting conclusions were resulted. The main
conclusion was offered by Shorrock in 1973 when completing the asymptotic theory of records (Shorrock
1973). In addition, Ahsanullah in 1973 covered the main concepts of the theory of records in 1998 in a
paper named “an introduction to record values”. The theory of records, nowadays, is utilized in various
statistical issues like the estimation theory, queuing theory, and etc. Among the works recently done in
the field of estimation of distribution parameters based on records, it might be pointed out to those of
Ahmadi & Doostparast (2006), Ahmadi, Ja’fari Jozani, Eric Marchand & Parsian (2009), Baklizi (2008)
and Ragheb (2007), Nasiri and Hosseini (2012). Ahmadi & Balakrishnan in a paper published in 2004,
did investigate the interval estimations based on record values for quantiles. As well, Nasiri and Hosseini
in 2013 used a statistic based on record values to estimate Bayesian estimation for exponential
distribution.
2. Maximum Likelihood Estimations Based on Random Sample and Upper Record Values for
Unknown Parameter of Exponential Family
Considering an iid random sample of size n from the exponential family of the first type, the joint
probability density function will be obtained as following:
n
n
i 1
i 1
(1)
f ( x1 ,..., xn ; ) B ( ) A' ( xi ) exp{ B( ) A( xi )}
n
The following fact has been used in the above relation:
f ( x; )
dF ( x; )
A' ( x ) B( ) exp{ A( x ) B( )}
dx
Hence, the maximum likelihood function is determined as follows considering the joint probability
density function ( X1 ,..., X n ) (1).
n
n
i 1
i 1
L log f x1 ,, xn ; nlogB logA' ( X i ) B A X i
On the other hand, it is known that the maximum likelihood estimations are determined from solving the
equation
L
0 , therefore:
n
L
nB ' ( )
0
B ' A X i 0 B
B( )
i 1
n
i1A X i
n
.
Now, if B( ) is a one-to-one function from to R, the following expression is simply resulted:
B 1 (
n
i1A X i
n
)
(2)
Since the probability density function f ( x; ) A ( x) B( )exp{ A( x) B( )} with the properties and
'
conditions described for A and B is a continuous function, it is possible to obtain its first and second order
derivatives.
n
2 L nB '' ( ) B( ) n[ B ' ( )]2
''
B
(
)
t
,
(
t
A( xi ))
2
B 2 ( )
i 1
In addition, it is simply observed by substituting (2) in (3) that:
(3)
L
2 B
2
n
i1A X i
n
n
nB '' ( )( ) n[ B ' ( )]2
[ B ' ( )t ]2
t
B '' ( )t
0
n 2
n
( )
t
Considering the negative sign of the likelihood function, it is found out that B (
n
1
i1A X i
likelihood value. With respect to the described issues, ML estimation of the parameter
n
) is a
based on the
random sample ( X1 ,..., X n ) will be attained as following:
ˆMLE ,( X ,..., X ) B 1 (
1
n
n
i1A X i
n
)
(4)
On the other hand, A( X ) is distributed as exponential distribution by the parameter
A( X )
Distribution
Exp(
(
1
) ), because
B( )
P( A( X ) t ) P( X A1 (t )) 1 exp{B( )t}
Besides, since X i is a random sample, A( X i ) is also a random sample and
i:1,..., n
i:1,..., n
A( X i )
Distribution
i:1,..., n
Exp(
1
),
B( )
(5)
so
T i 1 A( X i )
n
Distribution
Gamma(n,
1
).
B( )
Therefore, regarding equations (4) and (5):
ˆMLE ,( X ,..., X
1
Distribution
n)
n
B 1 ( )
T
Distribution
in which T
Gamma(n,
(6)
1
).
B( )
Considering a random sample of size m which has n values of the upper records (n m) , the joint
probability density function of the upper records for this sample is determined as follows (Arnold,
Balakrishnan and Nagaraja, 1998):
n 1
n
i 1
i 1
f r1 ,, rn ; f rn ; h ri ; B n ( ){ A' (ri )}exp{ B A rn }
As a result, the likelihood function based on upper records is obtained like follows:
n
LRe cords Logf r1 ,, rn ; nLogB( ) LogA' (ri ) B( ) A(rn ) .
i 1
Maximum likelihood estimation based on upper record values for parameter
solving the equation
will be obtained from
LRe cords
0 ; hence:
LRe cords
nB ' ( )
n
0
B ' ( ) A(rn ) 0 ˆ B 1 (
).
B( )
A( Rn )
(7)
On the other hand:
2 LRe cords nB '' ( ) B( ) n[ B ' ( )]2
B'' ( )l , (l A(rn ))
2
B 2 ( )
(8)
Substituting (7) in (8) the following relation is obtained:
2 LRe cords
2 B
n
A( rn )
n
nB '' ( )( ) n[ B ' ( )]2
[ B ' ( )l ]2
l
B '' ( )l
0.
n 2
n
( )
l
Which implies that (7) is a maximum likelihood estimation based on upper record values because it
maximizes the likelihood function based on record values. As a result:
ˆMLE ,( R ,..., R ) B 1 (
1
n
n
).
A( Rn )
(9)
On the other hand, since A( Ri ) is distributed as gamma by the parameter (i,
i:1,..., n
distributed as gamma by the parameter (n,
T A( Rn )
Distribution
Gamma (n,
1
) , then A( Rn ) is
B( )
1
) (Arnold, Balakrishnan and Nagaraja, 1998).
B( )
1
).
B( )
Considering (9) and (10) simultaneously leads to:
(10)
ˆMLE ,( R ,..., R )
1
n
Distribution
in which T A( Rn )
n
B 1 ( )
T
Distribution
(11)
Gamma(n,
1
).
B( )
The above-mentioned issues have been summarized in the following theorem:
Theorem 1
In the exponential family of the first type, the maximum likelihood estimator based on a random sample
of size n and the maximum likelihood estimator based on the first n values of the upper records are
identically distributed. In other words, if g is a real function then:
Distribution ˆ
i g (ˆMLE ,(( X1 ,..., X n )) )
g ( MLE ,( R1 ,..., Rn ) )
ii MSE ( g (ˆMLE ,(( X1 ,..., X n )) )) MSE ( g (ˆMLE ,( R1 ,..., Rn ) ))
Proof:
Regarding the described issues and the relations (6) and (11), the proof is obvious.
The second part of the theorem 1 can be regarded as the most primary reason for using record estimators.
Namely, in likelihood estimation of any function of , a random sample of size n and n values of the
upper records have the same errors (if MSE is regarded as the error criterion). Some examples will be
presented in the next sections in which the record estimations are more justifiable. Before this subject,
however, it is useful to define the symbol 𝛼
as follows:
MSE ( g (ˆMLE ,( X1 ,..., X n ) )) MSE ( g (ˆMLE ,( R1 ,..., Rn ) ))
n
n
n
E[{g ( B 1 ( )) g ( )}2 ] E[ g 2 ( B 1 ( ))] 2 g ( ) E[ g ( B 1 ( ))] ( g ( )) 2 (n)
T
T
T
(12)
Result 1
Considering the first part of theorem 1 and relation (12), if a number of m upper record values R 1 ,..., Rm
exist in a random sample X 1 ,..., X n , then the MSE value of ML estimation based on record values in this
sample equals with MSE of ML estimation based on a random sample of size m, i.e.
MSE (ˆMLE ,( R1 ,...,Rm ) ) MSE (ˆMLE based onrandomsampleof sizem( X1 ,..., X m ) ) (m)
Result 2
(13)
The relation (12) and the result (1) provide us a new tool and idea in order to decide by using it that in
which conditions employing the record estimations is better. Two following instances clarify this issue.
Example 1. If A( x) x and B( )
gamma
n
T
( B 1 ( )
1
n
T
, then B 1 ( )
by
the
T
n T
; it is known that B 1 ( ) is distributed as
n
T
n
parameters
( n, )
n
T Distribution
Gamma(n, ) ) . Now, considering the relation (12)
n
n
n
2
n
E[( B 1 ( ))2 ] 2 , E[ B 1 ( )]
T
n
T
therefore
MSE (n )
2
n
.
With respect to the figure, it is readily concluded that for a random sample of size n, if m records of the
upper records are available (m n) , then the maximum likelihood estimation based on record values
have always more errors.
Figure 1.2. Comparison of maximum likelihood estimation
based on random sample and records in an exponential distribution.
As a consequence of Example 1, the maximum likelihood estimations based on record values might be
inefficient and unsuitable in the first looking. It is shown that this belief is not in general the case and
there are circumstances in which estimations based on record values are more efficient (i.e. have less
MSEs). The following example clarifies this subject.
x
k
Example 2. Assuming that A( x) Log ( ) and B( ) , it is wanted to estimate the parameter
g ( ) k .
As a consequence of the relation (12),
n
n
n
1 n
T
B ( ) , B ( ) , g ( B ( )) k
T
T
T
1
1
Thus
n
n
T
n
T
( Ln(k T ))i
k exp{Ln(k )}
, if for example k e then
i!
i 0
n
n i
( Ln(k T ))i ( T )
.
k exp{Ln(k )}
i!
i!
i 0
i 0
n
T
n
T
n i
n
)
( )i
(
n
n
Considering E[{g ( B 1 ( ))} ] E[exp{ }] E[ T ], and T
0
T
T
i!
i!
i 0
(
Each variable
n
T
i!
)i
is greater than zero then, that’s why it is possible to interchange the integration and
summation.
E[
i 0
(
n
)i
n
)i
i i n 1 n
i n i n n i 1
T ] E[ T ] { n t exp{ t} dt}
{ t
exp( t )dt ,
t ii !(n 1)!
i!
i!
i 0
i 0 0
i 0 i !( n 1)! 0
(
n and i are round numbers that’s why
t
0
n i 1
exp( t )dt
( n i )
n i
.
(14)
n i
)
(
n
i ni n n i 1
i ni n (n i )
E[{g ( B 1 ( ))} ] E[ T ]
{ t
exp( t )dt
.
T
i!
n i
i 0
i 0 i !( n 1)! 0
i 0 i !( n 1)!
Having considered the above relation and the fact that Gamma function is defined for positive values, the
MSE is readily resulted as follows:
n 1
MSE ( g ( ))
i 0
n 1
2i ni n (n i )
ni n (n i )
2
e
e2 .
n i
n i
i !(n 1)!
i 0 i !( n 1)!
If we draw MSE ( g ( )) for different values of n, then the following figure is obtained:
Figure 2.2. Different values of MSE in terms of n when estimating g
Consider the figure 2.2; if for instance the number of upper records in a sample of size 10 or 11 equals to
5 or 6, then the maximum likelihood estimation g ( ) based on records does have less amount of error.
For another example, it is seen, with respect to the figure 2.2, that (4) (5) (7) (12) , that is,
if a sample of size 12 has 4 values of the upper records, then the ML estimation of g ( ) based on record
values does have less amount of error, since it is clearly seen in the figure that (4) (11) .
3. Maximum Likelihood Estimations for the PDF and CDF from the Exponential Family of the
First Type Based on Random Sample and Upper Record Values
Suppose that a random sample of size n from the exponential family of the first type ( X1 ,..., X n ) has m
values from the upper record ( R1 ,..., Rm ) . Considering the relation (4) and the point that the maximum
likelihood estimations are stable, the ML estimations for f and F are determined like follows:
ˆ
MLE ,( X1 ,..., X n )
B 1 (
n
A Xi
i 1
n
fˆMLE ,( X1 ,..., X n ) A' ( x) B( B 1 (
nA' ( x)
i1A X i
n
exp(
),
n
nA( x)
n
A( X )
i 1
A Xi
i 1
n
)) exp( A( x) B( B 1 (
n
A Xi
i 1
n
)))
)
i
(15)
Similarly the maximum likelihood estimation for cumulative distribution function based on random
sample is determined as
nA( x)
FˆMLE ,( X1 ,..., X n ) 1 exp( n
) . (16)
A( X i )
i 1
In addition, the maximum likelihood estimators f ( x) and F ( x) based on the upper record values in
this sample (sample of size n) are obtained as follows:
mA' ( x)
mA( x) ˆ
mA( x)
fˆMLE ,( R1 ,..., Rm )
exp(
), FMLE ,( R1 ,..., Rm ) 1 exp(
)
A( Rm )
A( Rm )
A( Rm )
(17)
Theorem 2. If the random sample ( X1 ,..., X n ) has m values from the upper records ( R1 ,..., Rm ) , then:
The ML estimations of the PDF and CDF based on upper record values and based on a random sample of
size m are distributed identically.
Proof: it is easily proven by considering (6), (10), (15), (16), and (17).
Theorem 3. The ML estimations based on random sample and based on the upper record values are
biased estimators and
n 2
(n i 1)
ˆ
i E[ f MLE ,( X1 ,..., X n ) ] [
( 1)i {nB( )}i 1 A' ( x )]{ A( x )}i
i 0 ( n ) (i 1)
m 2
ii E[ fˆMLE ,( R1 ,...,Rm ) ] [
i 0
(m i 1)
( 1)i {mB( )}i 1 A' ( x )]{ A( x )}i
(m) (i 1)
n 1
iii E[ FˆMLE ,( X1 ,..., X n ) ] 1 [
i 0
( n i )
( 1)i {nB( )}i ]{ A( x )}i
(n ) (i 1)
( m i )
iv E[ FˆMLE ,( R1 ,...,Rm ) ] 1 [
( 1)i {mB( )}i ]{ A( x )}i
(
m
)
(
i
1)
i 0
m 1
Proof: Here, the forth part is proven, other parts are proven similarly:
distribution
mA( x )
1
FˆMLE ,( R1 ,...,Rm ) 1 exp(
), T A( Rm )
Gamma(m,
),
A( Rm )
B( )
mA( x )
mA( x )
E[ FˆMLE ,( R1 ,...,Rm ) ] E[1 exp(
)] 1 E[exp(
)]
A( Rm )
A( Rm )
E[exp(
mA( x )
( 1)i mi Ai ( x )
mi Ai ( x )
m i Ai ( x )
)] E[ (
]
E
[
(
)]
E
[
(
T ii !
T ii ! )]
A( Rm )
T ii !
i 0
even is
odd is
Notice that the variables
and summation.
mi Ai ( x )
are positive, that’s why it is possible to interchange the expectations
T ii !
E[
even is
even is 0
mi Ai ( x )
mi Ai ( x )
mi Ai ( x )
mi Ai ( x )
)]
E
[
(
)]
E
(
)
E
(
)
T ii !
T ii !
T ii !
T ii !
odd is
even is
odd is
(
mi Ai ( x ) t m1 B m ( ) exp( B( )t )
dt
tii !
(m)
odd is
mi Ai ( x ) B m ( )
i ! (m)
even is
0
t
using (14) instead of
t mi 1 exp( B( ))dt
m i 1
0
mi Ai ( x ) t m1 B m ( ) exp( B( )t )
dt
tii !
(m)
mi Ai ( x ) B m ( )
i ! (m)
odd is
t
m i 1
exp( B( ))dt ,
0
exp( B( ))dt above expression is resulted as:
0
mi Ai ( x ) B m ( )(m i ) mi Ai ( x ) B m ( ) (m i ) (m i )
( mA( x ) B( ))i ,
m i
m i
i
!
(
m
)
B
(
)
i
!
(
m
)
B
(
)
(
i
1)
(
m
)
even is
odd is
i 0
finally, having considered the above relations and the fact that Gamma function is defined for positive
values
the
expectation
is
readily
resulted
as
follows:
m 1
mA( x )
(m i )
E[ FˆMLE ,( R1 ,...,Rm ) ] 1 E[exp(
)] 1
( mA( x ) B( ))i .
A( Rm )
(
i
1)
(
m
)
i 0
Theorem 4.
i MSE[ fˆMLE ,( X ,..., X ) ] [( B( ) A' ( x ))2 ( 2nB( ) A( x )) ( n i 2) ]
n 3
1
n
i
( n ) (i 1)
i 0
n 2
2nA' ( x )( B( )) 2 exp{ A( x ) B( )} [
i 0
( nA( x ) B( ))i ( n i 1)
] ( A' ( x ) B( )) 2 exp{2 A( x ) B( )}
( n ) (i 1)
m 3
ii MSE[ fˆMLE ,( R ,...,R ) ] [( B( ) A' ( x ))2
1
m
i 0
m 2
2mA' ( x )( B( )) 2 exp{ A( x ) B( )} [
i 0
n 1
iii MSE[ FˆMLE ,( X ,..., X ) ] [
1
n
n 1
2 exp{ A( x ) B( )} [
i 0
i 0
m 1
m
m 1
2 exp{ A( x ) B( )} [
i 0
( mA( x ) B( ))i ( m i 1)
] ( A' ( x ) B( )) 2 exp{2 A( x ) B( )}
( m) (i 1)
( 2nB( ) A( x ))i ( n i )
]
( n ) (i 1)
( nB( ) A( x ))i ( n i )
] exp{2 A( x ) B( )}
( n ) (i 1)
iv MSE[ FˆMLE ,( R ,...,R ) ] [
1
( 2mB( ) A( x ))i ( m i 2)
]
( m) (i 1)
i 0
( 2mB( ) A( x ))i ( m i )
]
( m ) (i 1)
( mB( ) A( x ))i ( m i )
exp{2 A( x ) B( )}
( m) (i 1)
.
Proof: Here the forth part is proven. Other parts are proved in the same way.
MSE ( FˆMLE ,( R1 ,...,Rm ) ) E[ FˆMLE ,( R1 ,...,Rm ) F ( x )]2 E[{1 exp(
E[exp(
mA( x )
)} {1 exp( B( ) A( x ))}]2
A( Rm )
mA( x )
2mA( x )
mA( x )
) exp( B( ) A( x ))]2 E[exp(
)] 2 exp( B( ) A( x )) E[exp(
)]
A( Rm )
A( Rm )
A( Rm )
exp( 2 B( ) A( x ))
From equation number (10) it is clear that
T A( Rm )
distribution
1
Gamma(n,
)
B( )
Considering above relation,
MSE ( FˆMLE ,( R1 ,...,Rm ) ) E (exp(
2mA( x )
mA( x )
)) 2exp( B( ) A( x )) E (
) exp( 2 B( ) A( x)) .
T
T
In order to obtain the above relation we should obtain the following general expectation
mA( x )
W ( ) E (exp(
))
T
0
( mA( x )) t
t ii !
i 0
i
It is clear that
i
exp(
0
mA( x ) t m 1B m ( ) exp( B( )t )
)
T
( m )
B ( ) exp( B( )t )
( m )
( mA( x )) t
t ii !
odd indices
0
m 1
m
m 1
0
B ( ) exp( B( )t )
.
( m )
m
( mA( x ))i
0 then,
t ii !
( mA( x ))i t m 1B m ( ) exp( B( )t )
t ii !
( m )
even indices
0
( mA( x ))i t m 1B n ( ) exp( B( )t )
dt
t ii !
( m )
even indices
even indices 0
( mA( x ))i t m 1B m ( ) exp( B( )t )
dt
0 odd
t ii !
( m )
indices
( mA( x ))i t m 1B m ( ) exp( B( )t )
dt
t ii !
( m )
odd indices
( mA( x )) B ( )
(i 1) ( m)
even indices
m 1
( mA( x )) B ( )
(i 1) ( m)
even indices
i
i
m
m
t
0
t
0
( mA( x ))i t m 1B m ( ) exp( B ( )t )
dt
t ii !
( m )
exp( B( )t )
( mA( x ))i B m ( )
dt
(i 1)(m)
ti
odd indices
m i 1
0
m 1
0
( mA( x )) B ( )
exp( B ( )t )dt
(i 1) ( m)
odd indices
i
m
t
t
exp( B( )t )
dt
ti
m i 1
exp( B( )t )dt
0
( mA( x ))i B m ( ) ( m i )
( mA( x ))i B m ( ) ( m i )
(i 1)(m) B mi ( ) odd
(i 1) ( m) B m i ( )
even indices
indices
( mA( x ))i B m ( ) ( m i ) ( mB( ) A( x ))i ( m i ) n 1 ( mB( ) A( x )) i ( m i )
.
(i 1) ( m) B m i ( ) i 0
(i 1) ( m)
(i 1) ( m)
i 0
i 0
Finally, W ( )
( mB( ) A( x ))i (m i )
.
(i 1)(m)
i 0
m 1
Now the MSE function is obtained easily as follow:
MSE W (2) 2 exp( B( ) A( x ))W (1) exp( 2 B( ) A( x))
m 1
( 2mB( ) A( x ))i ( m i )
( mB( ) A( x))i ( m i )
2
exp(
B
(
)
A
(
x
))
exp( 2 B( ) A( x )).
(i 1)(m)
(i 1)( m)
i 0
i 0
m 1
The next theorem clarifies some advantageous aspects of the above likelihood estimators. That is, the
above-mentioned estimators are always nearly unbiased in large samples.
Theorem 5. The maximum likelihood estimations based on random sample and on the upper record
values for the PDF and CDF in the exponential family of the first type are asymptotically unbiased. In the
other words
i Lim E[ fˆMLE ,( X1 ,..., X n ) ] f ( x)
n
ii Lim E[ fˆMLE ,( R1 ,..., Rm ) ] f ( x)
m
iii Lim E[ FˆMLE ,( X1 ,..., X n ) ] F ( x)
n
iv Lim E[ FˆMLE ,( R1 ,..., Rm ) ] F ( x)
m
The following lemma is needed in order to prove the above theorem.
(n i 1)ni 1
Lemma 1. Lim
1 for the reason that
n
( n)
Lim
n
(n i 1)n i 1
Lim
n
( n )
ni
i 1
(n k )
Lim
n
ni
1
ni
k 1
It must be proven that
(n i 1)ni 1
0, N , s.t , n N :
1
( n)
Now, assuming that is arbitrary, N is chosen as N
i 1
( 1) i i 1 ( 1)
i 1
( 1) 1
. Now
(n i 1)ni 1 (n i 2)!ni 1
ni 1
n
n
n
n
...
...
( n)
(n 1)!
(n 1)(n 2)...(n i 1) n 1
n i 1 n i 1
n i 1
( i 1) times
( i 1) times
On the other hand, it can be simply shown that
n N , and , i 1,2,3,..., N 2 :
n
N
,
n i 1 N i 1
thus
n
n
N
N
...
...
n i 1
n i 1 N i 1
N i 1,
(i 1)times
(i 1)times
and consequently
(n i 1)ni 1
N
N
.
n N :
...
( n)
N i 1
N i 1
( i 1) times
Finally by substituting N
i 1
( 1) i i 1 ( 1)
i 1
( 1) 1
in the last relation, the following result is reached:
(n i 1)ni 1
1
( n)
Proof:
Only the first part is proven; the other parts have similar proofs. First notice that
(n i 1)
Lim[ fˆMLE ,( X1 ,..., X n ) ] [ Lim
(1)i ni 1{B( )}i 1 A' ( x)]{ A( x)}i ]
n
i 0 n ( n)(i 1)
Lim(
i 0
n
(n i 1)ni 1 { B( )}i B( ) A' ( x){ A( x)}i ]
]
( n)
(i 1)
By considering lemma (1)
{ A( x ) B( )}
Lim[ fˆMLE ,( X1 ,..., X n ) ]
B( ) A' ( x ) A '( x ) B( ) exp{ A( x ) B( )}.
n
i
!
i 0
i
The above theorem makes us to pay attention to convergence in probability and to the following theorem.
Theorem 6. If a random sample ( X1 ,..., X n ) has m values from the upper records ( R1 ,..., Rm ) , then the
maximum likelihood estimations for the CDF and PDF based on upper record values are consistent
estimations for the CDF and PDF.
P
i fˆMLE ,( R1 ,...,Rm )
f ( x )
P
ii FˆMLE ,( R1 ,...,Rm )
F ( x )
Proof. The part i is proven; other part has a similar proof. It is considered the following abbreviation just
for easiness.
fˆMLE ,( R1 ,...,Rn ) fˆMLE ,R
Based on the Markov’s theorem
Var ( fˆMLE ,R )
P{ fˆMLE ,R E[ fˆMLE ,R ]}
On the other hand, it is concluded by some mathematical calculations that
E[{ fˆMLE ,R }2 ]
i 0
(m i 2)
{ A '( x )}2 {B( )}2 ( 2)i {nB( )}i { A( x)}i
(i 1)(m)
and consequently
(m i 2)
Var( fˆMLE ,R )
{ A '( x)}2 {B( )}2 ( 2)i {mB( )}i { A( x)}i {E[ fˆMLE ,R ]}2
(
i
1)
(
m
)
i 0
thus
m (m i 2) {2 B( )} { A( x )}
LimVar ( fˆMLE ,R ) Lim[
]
{ A '( x )}2 {B( )}2 { f ( x )}2
m
m
( m )
(i 1)
i 0
i
i
i
{ A '( x )}2 {B( )}2 exp{2 A( x ) B( )} { f ( x )}2 0
Therefore
Lim P{ fˆMLE ,R E[ fˆMLE ,R ] } 0
m
P
fˆMLE ,R
E[ fˆMLE ,R ]
so
or
P
fˆMLE ,R E[ fˆMLE ,R ]
0
P
On the other hand, it is reminded that if Lim am a and X
r then (Billingsley (1995))
m
P
X am
r a
(18)
Considering (18) and respecting the second part of theorem 5 ( Lim E[ fˆMLE ,R ] f ( x ) ), it is obtained
m
that
P
fˆMLE ,R E[ fˆMLE ,R ] E[ fˆMLE ,R ] f
0
,
or equally
P
fˆMLE ,R
f.
The following remembrance is essential before expressing the next theorem.
P
{ X } and D is a continuous function of y {D( y) : R R} , then
Remembrance 1: if { X n }
(Billingsley, 1995)
P
D({X n })
D({X })
Theorem 7.
Suppose that a random sample of size n from the exponential family of the first type is available. Also
suppose that this sample includes m values from the upper record values. The estimators ˆMLE ,( R1 ,...,Rm ) are
consistent for . Namely
P
ˆMLE ,( R ,...,R )
1
Proof.
m
With respect to the continuity of functions B(t ), Log (1 t ), F (t ) 1 exp{ A(t ) B( )}, and the
continuity theorem of composite functions, it is concluded that the function D(t ) B(
is always continuous on
Log (1 F (t ))
)
A( x)
. On the other hand, it is known from the second part of theorem 6 that
P
FˆMLE ,( R1 ,...,Rm )
F ( x ),
or equivalently
P
A( x) B(ˆMLE ,( R1 ,..., Rm ) ) exp{ A( x) B(ˆMLE ,( R1 ,..., Rm ) )}
F ( x) .
Now with respect to remembrance 1,
P
D( FˆMLE ,( R1 ,..., Rm ) )
D( F ( x))
which is equivalent to
P
ˆMLE ,( R ,..., R )
1
m
4. Conclusions
In the beginning, considering two general subfamilies from the exponential family, it was tried to
determine the maximum likelihood estimation for these two families at the first step and general state;
then, the maximum likelihood estimations based on upper were calculated for the unknown parameter of
these two families; it was engaged via some theorems to description of this problem that the record
estimations can be acceptable in general state; then the estimations of maximum likelihood and of
maximum likelihood based on records were calculated for the cumulative distribution function and for the
probability density function of these families, and their corresponding errors were described and studied
by using some theorems; in addition, we engaged to proof of this problem that at asymptotical states,
these estimations are unbiased estimators. Subsequently, several theorems associated with convergence in
probability were expressed for the resulted estimators.
References
1- Chandler, K. N. (1952). The distribution and frequency of record values. Journal of the Royal
234-
5-
Statistical Society. Series B (Methodological), 220-228.
Shorrock, R. W. (1972). A limit theorem for inter-record times. Journal of Applied
Probability, 9(1), 219-223.
Ahsanullah, M. (1988). Introduction to record values.
Ahmadi, J., Doostparast, M., & Parsian, A. (2005). Estimation and prediction in a two-parameter
exponential distribution based on k-record values under LINEX loss function. Communications in
Statistics–Theory and Methods, 34(4), 795-805.
Ahmadi, J., & Doostparast, M. (2006). Bayesian estimation and prediction for some life
distributions based on record values. Statistical Papers, 47(3), 373-392.
6- Ahmadi, J., Jozani, M. J., Marchand, É., & Parsian, A. (2009). Bayes estimation based on krecord data from a general class of distributions under balanced type loss functions. Journal of
Statistical Planning and Inference, 139(3), 1180-1189.
7- Baklizi, A. (2008). Likelihood and Bayesian estimation of using lower record values from the
generalized exponential distribution. Computational Statistics & Data Analysis, 52(7), 34683473.
8- Raqab, M. Z., Ahmadi, J., & Doostparast, M. (2007). Statistical inference based on record data
from Pareto model. Statistics, 41(2), 105-118.
9- Nasiri, P., & Hosseini, S. (2012). Statistical inferences for Lomax distribution based on record
values (Bayesian and classical). Journal of Modern Applied Statistical Methods, 11(1), 15.
10- Ahmadi, J., & Balakrishnan, N. (2004). Confidence intervals for quantiles in terms of record
range. Statistics & probability letters, 68(4), 395-405.
11- Nasiri, P., Hosseini, S., Yarmohammadi, M., & Hatami, F. (2013). Bayesian inference for
exponential distribution based on upper record range. Arabian Journal of Mathematics, 2(4), 349364.
12- B. C. Arnold, N. Balakrishnan, H.N. Nagaraja,"records,"John Wiley and Sons,Canada1998.
13- Billingsley, P. (1995). Probability and Measure, ser. Probability and Mathematical Statistics.
New York: Wiley.
| 10math.ST
|
arXiv:1710.02316v1 [cs.CV] 6 Oct 2017
A Multiscale Patch Based Convolutional Network
for Brain Tumor Segmentation.
Jean Stawiaski
Stryker Corporation, Freiburg, Germany.
September 22, 2017.
Abstract
This article presents a multiscale patch based convolutional neural network for the automatic segmentation of brain tumors in multimodality 3D MR images. We use multiscale deep supervision and
inputs to train a convolutional network. We evaluate the effectiveness
of the proposed approach on the BRATS 2017 segmentation challenge
[1, 2, 3, 4] where we obtained dice scores of 0.755, 0.900, 0.782 and 95%
Hausdorff distance of 3.63mm, 4.10mm, and 6.81mm for enhanced tumor core, whole tumor and tumor core respectively.
Keywords. Brain tumor, convolutional neural network, image segmentation.
1
Introduction
Brain Gliomas represent 80% of all malignant brain tumors. Gliomas can
be categorized according to their grade which is determined by a pathologic
evaluation of the tumor:
• Low-grade gliomas exhibit benign tendencies and indicate thus a better
prognosis for the patient. However, they also have a uniform rate of
recurrence and increase in grade over time.
• High-grade gliomas are anaplastic; these are malignant and carry a
worse prognosis for the patient.
Brain gliomas can be well detected using magnetic resonance imaging.
The whole tumor is visible in T2-FLAIR, the tumor core is visible in T2 and
1
the enhancing tumor structures as well as the necrotic parts can be visualized using contrast enhanced T1 scans. An example is illustrated in figure 1.
Figure 1: Example of images from the BRATS 2017 dataset. From left
to right: T1 image, T2 image: the whole tumor and its core are visible,
T2 FLAIR image: discarding the cerebrospinal fluid signal from the T2
image highlights the tumor region only, T1ce: contrast injection permits
to visualize the enhancing part of the tumor as well as the necrotic part.
Finally the expected segmentation result is overlaid on the T1ce image. The
edema is shown in red, the enhancing part in white and the necrotic part of
the tumor is shown in blue.
Automatic segmentation of brain tumor structures is particularly important in order to quantitatively assess the tumor geometry. It has also
a great potential for surgical planning and intraoperative surgical resection
guidance. Automatic segmentation still poses many challenges because of
the variability of appearances and sizes of the tumors. Moreover the differences in the image acquisition protocols, the inhomogeneity of the magnetic
field and partial volume effects have also a great impact on the image quality
obtained from routinely acquired 3D MR images.
In the recent years, deep neural networks have shown to provide stateof-the-art performance for various challenging image segmentation and classification problems [6, 7, 9, 10, 8]. Medical image segmentation problems
have also been successfully tackled by such approaches [11, 12, 16, 17, 18].
Inspired by these works, we present here a relatively simple architecture that
produces competitive results for the BRATS 2017 dataset [1, 2, 3, 4]. We
propose a variant of the well known U-net [11], fed with multiscale inputs
[14], having residual connections [13], and being trained in a multiscale deep
supervised manner [15].
2
2
Multiscale Patch Based Convolutional Network
This section details our network architecture, the loss function used to train
the network end-to-end as well as the training data preparation.
2.1
Network Architecture
Our architecture is illustrated in figure 2. The network processes patches of
643 voxels and takes multiscale version of these patches as input. We detail
here some important properties of the network:
• each sample image y is normalised to have zero mean and unit variance
for voxels inside the brain:
x − mbr
y=
,
(1)
σbr
where mbr and σbr is the mean and the variance of voxels inside the
brain (non zero voxels of a given 3D image),
• batch normalisation is performed after each convolutional layer using
a running mean and standard deviation computed on 5000 samples:
by =
(y − mb )
×γ+c,
(σb + )
(2)
where mb and σb is the mean and variance of the minibatches and γ
and c are learnable parameters,
• each layer is composed of residual connections as illustrated in figure
3,
• different layers of the network are combined using (1x1) convolution
kernels as illustrated in figure 4,
• the activation function is an exponential linear unit,
• convolution kernels are (3x3x3) kernels,
• convolutions are computed using reflective border padding,
• downsampling is performed by decimating a smooth version of the
layer:
dy = (y ∗ Gσ ) ↓2 ,
(3)
where Gσ is a gaussian kernel,
• upsampling is performed by nearest neighbor interpolation.
3
Figure 2: Network architecture: multiscale convolutional neural network.
2.2
Scale-wise Loss Function
We define a loss function for each scale of the network allowing a faster
training and a better model convergence. Using this deep supervision, gradients are efficiently injected at all scales of the network during the training
process. Downsampled ground truth segmentation images are used to compute the loss associated for each scale. The loss function is defined as the
combination of the mean cross entropy (mce) and the Dice coefficients (dce)
between the ground truth class probability and the network estimates:
ce =
X −1 X
k
n
yik log(pki )
(4)
i
where yik and pki represent respectively the ground truth probability and the
network estimate for the class k at location i.
4
Figure 3: Residual connections in a convolutional layer.
Figure 4: Layer combination using (1x1) convolution kernels.
dce =
X
k6=0
P
2 i pki yik
1
P k 2 P k 2
.
1.0 −
n
i (pi ) +
i (yi )
(5)
Note that we exclude the background class for the computation of the
dice coefficient.
2.3
Training Data Preparation
We used the BRATS 2017 training and validation sets for our experiments
[1, 2, 3, 4]. The training set contains 285 patients (210 high grade gliomas
and 75 low grade gliomas). The BRATS 2017 validation set contains 46
patients with brain tumors of unknown grade with unknown ground truth
segmentations. Each patient contains four modalities: T1, T1 with contrast
enhancement, T2 and T2 FLAIR. The aim of this experiment is to segment
5
automatically the tumor necrotic part, the tumor edema and the tumor enhancing part.
The segmentation ground truth provided with the BRATS 2017 dataset
presents however some imperfections. The segmentation is relatively noisy
and does not present a strong 3D coherence as illustrated in figure 5. We
have thus decided to manually smooth each ground truth segmentation map
independently such that:
y k = (y k ∗ Gσ ) ,
(6)
where y k is the probability map associated with the class k, Gσ is a normalised gaussian kernel. Note that this process still ensures that yk is a
probability map:
X
yik = 1 .
(7)
k
Figure 5: Noisy segmentation ground truth. Example of class wise probability smoothing. (Necrotic parts is shown in dark gray, edema in light gray
and enhancing tumor in white).
In order to deal with the class imbalance, patches are sampled so that
at least 0.01 % of the voxels contain one of the tumor classes.
2.4
Implementation
The network is implemented using Microsoft CNTK 1 . We use stochastic
gradient descent with momentum to train the network. The network is
trained using 3 Nvidia GTX 1080 receiving a different subset of the training
data. The inference of the network is done on one graphic card and takes
approximatively 5 seconds to process an image by analyzing non overlapping
sub volumes of 643 voxels.
1
https://www.microsoft.com/en-us/cognitive-toolkit/
6
3
Results
Due to our training data preparation (class wise segmentation smoothing)
and due to our data augmentation method (additive noise), our segmentation results tends to be smoother than the ground truth segmentation. This
effect is illustrated in figure 6.
Figure 6: Left: segmentation obtained on a image from the training data.
Middle: obtained segmentation result. Right: Ground truth segmentation.
The edema is shown in red, the enhancing part in white and the necrotic
part of the tumor is shown in blue. Our results tend to be smoother than
the ground truth delineation.
We uploaded our segmentation results to the BRATS 2017 server 2 which
evaluates the segmentation and provides quantitative measurements in terms
of Dice scores, sensitivity, specificity and Hausdorff distances of enhanced
tumor core, whole tumor, and tumor core. Results of the BRATS 2017 validation phase are presented in Table 1. The table summarizes the scores as
they appeared on the leaderboard the 22 September 2017. We observe that
the proposed method does not perform as well as the other best methods
in terms of dice coefficients. On the other side our method produces very
competitive distances metrics.
Team
UCL-TIG
biomedia1
Zhouch
MIC DKFZ
stryker
2
Dice ET
0.785
0.757
0.760
0.731
0.755
Dice WT
0.904
0.901
0.903
0.896
0.900
Dice TC
0.837
0.820
0.824
0.797
0.782
Dist. ET
3.28
4.22
3.71
4.54
3.63
https://www.cbica.upenn.edu/BraTS17/lboardValidation.html
7
Dist. WT
3.89
4.55
4.87
6.97
4.10
Dist. TC
6.47
6.10
6.74
9.47
6.81
Table 1: BRATS 2017 Validation scores, dice coefficients and the 95% Hausdorff distances. Our results corresponds to the team name ”stryker”.
Different segmentation results are illustrated in figure 7. The proposed
network tends to produce smooth and compact segmentation results which
are often very close in terms of Euclidean distance to the ground truth segmentation. We have consciously chosen to privilege this effect by smoothing
the ground truth segmentation and augmenting data with additive noise.
Different approaches may be better suited for other kind of quality metrics.
Figure 7: Segmentation results obtained on images from the validation data.
(Top: good results, Bottom: incorrect detection of necrotic parts.) The
edema is shown in red, the enhancing part in white and the necrotic part of
the tumor is shown in blue.
4
Conclusion
We have presented a relatively simple but efficient approach for automatic
brain tumor segmentation using a convolutional network. We obtained competitive scores on the BRATS 2017 segmentation challenge 3 . Future work
3
https://www.cbica.upenn.edu/BraTS17/lboardValidation.html
8
will concentrate on making the network more compact and more robust in
order to be used clinically in an intraoperative setup. A possible improvement of the presented method could consist in adding semantic constraints
by using a hierarchical approach such as the one presented in [5].
References
[1] Menze BH, Jakab A, Bauer S, Kalpathy-Cramer J, Farahani K, Kirby J,
Burren Y, Porz N, Slotboom J, Wiest R, Lanczi L, Gerstner E, Weber
MA, Arbel T, Avants BB, Ayache N, Buendia P, Collins DL, Cordier
N, Corso JJ, Criminisi A, Das T, Delingette H, Demiralp , Durst CR,
Dojat M, Doyle S, Festa J, Forbes F, Geremia E, Glocker B, Golland P,
Guo X, Hamamci A, Iftekharuddin KM, Jena R, John NM, Konukoglu
E, Lashkari D, Mariz JA, Meier R, Pereira S, Precup D, Price SJ, Raviv
TR, Reza SM, Ryan M, Sarikaya D, Schwartz L, Shin HC, Shotton J,
Silva CA, Sousa N, Subbanna NK, Szekely G, Taylor TJ, Thomas OM,
Tustison NJ, Unal G, Vasseur F, Wintermark M, Ye DH, Zhao L, Zhao B,
Zikic D, Prastawa M, Reyes M, Van Leemput K. ”The Multimodal Brain
Tumor Image Segmentation Benchmark (BRATS)”, IEEE Transactions
on Medical Imaging 34(10), 1993-2024 (2015)
[2] Bakas S, Akbari H, Sotiras A, Bilello M, Rozycki M, Kirby JS, Freymann
JB, Farahani K, Davatzikos C. ”Advancing The Cancer Genome Atlas
glioma MRI collections with expert segmentation labels and radiomic
features”, Nature Scientific Data, 2017.
[3] Bakas S, Akbari H, Sotiras A, Bilello M, Rozycki M, Kirby J, Freymann
J, Farahani K, Davatzikos C. ”Segmentation Labels and Radiomic Features for the Pre-operative Scans of the TCGA-GBM collection”, The
Cancer Imaging Archive, 2017.
[4] Bakas S, Akbari H, Sotiras A, Bilello M, Rozycki M, Kirby J, Freymann
J, Farahani K, Davatzikos C. ”Segmentation Labels and Radiomic Features for the Pre-operative Scans of the TCGA-LGG collection”, The
Cancer Imaging Archive, 2017.
[5] Guotai Wang, Wenqi Li, Sebastien Ourselin, Tom Vercauteren, Automatic Brain Tumor Segmentation using Cascaded Anisotropic Convolutional Neural Networks, arXiv:1709.00382, https://arxiv.org/abs/
1709.00382, 2016.
9
[6] J. Long, E. Shelhamer, and T. Darrell, Fully convolutional networks
for semantic segmentation, arXiv:1605.06211, https://arxiv.org/abs/
1605.06211, 2016.
[7] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L.Yuille,
Semantic image segmentation with deep convolutional nets and fully
connected CRF, arXiv:1412.7062, https://arxiv.org/abs/1412.7062,
2014.
[8] Hyeonwoo Noh, Seunghoon Hong, Bohyung Han, Learning Deconvolution Network for Semantic Segmentation, arXiv:1505.04366, https:
//arxiv.org/abs/1505.04366, 2015.
[9] Vijay Badrinarayanan, Alex Kendall, Roberto Cipolla, SegNet: A Deep
Convolutional Encoder-Decoder Architecture for Image Segmentation,
arXiv:1511.00561, https://arxiv.org/abs/1511.00561, 2015.
[10] Fisher Yu, Vladlen Koltun, Multi-Scale Context Aggregation by Dilated Convolutions, arXiv:1511.07122, https://arxiv.org/abs/1511.
07122, 2016.
[11] Olaf Ronneberger, Philipp Fischer, Thomas Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation, arXiv:1505.04597,
https://arxiv.org/abs/1505.04597, 2015.
[12] Fausto Milletari, Nassir Navab, Seyed-Ahmad Ahmadi, V-Net: Fully
Convolutional Neural Networks for Volumetric Medical Image Segmentation, arXiv:1606.04797, https://arxiv.org/abs/1606.04797, 2016.
[13] Zifeng Wu, Chunhua Shen, and Anton van den Hengel, Wider or
Deeper: Revisiting the ResNet Model for Visual Recognition, https:
//arxiv.org/abs/1611.10080, 2016.
[14] Iasonas Kokkinos, Pushing the Boundaries of Boundary Detection using Deep Learning, arXiv:1611.10080, https://arxiv.org/abs/1511.
07386, 2015.
[15] Chen-Yu Lee, Saining Xie, Patrick Gallagher, Zhengyou Zhang,
Zhuowen Tu, Deeply-Supervised Nets. arXiv:1409.5185, https://
arxiv.org/abs/1409.5185, 2014.
[16] Qikui Zhu, Bo Du, Baris Turkbey, Peter L . Choyke, Pingkun Yan,
Deeply-Supervised CNN for Prostate Segmentation. arXiv:1703.07523,
https://arxiv.org/abs/1703.07523, 2017.
10
[17] Lequan Yu, Xin Yang, Hao Chen, Jing Qin, Pheng-Ann Heng, Volumetric ConvNets with Mixed Residual Connections for Automated
Prostate Segmentation from 3D MR Images. Proceedings of the ThirtyFirst AAAI Conference on Artificial Intelligence (AAAI-17). 2017.
[18] Christ P.F., Elshaer M.E.A., Ettlinger F., Tatavarty S., Bickel M.,
Bilic P., Remper M., Armbruster M., Hofmann F., Anastasi M.D.,
Sommer W.H., Ahmadi S.a., Menze B.H., Automatic Liver and Lesion Segmentation in CT Using Cascaded Fully Convolutional Neural Networks and 3D Conditional Random Fields. arXiv:1610.02177,
https://arxiv.org/abs/1610.02177, 2016.
11
| 1cs.CV
|
JOURNAL OF COMPUTER SCIENCE AND ENGINEERING, VOLUME 1, ISSUE 1, MAY 2010
68
Algorithm for Predicting Protein Secondary
Structure
K. K Senapati, G. Sahoo and D. Bhaumik
Abstract—Predicting protein structure from amino acid sequence is one of the most important unsolved problems of molecular biology and
biophysics.Not only would a successful prediction algorithm be a tremendous advance in the understanding of the biochemical mechanisms
of proteins , but , since such an algorithm could conceivably be used to design proteins to carry out specific functions.Prediction of the secondary structure of a protein (alpha-helix, beta-sheet, coil) is an important step towards elucidating its three dimensional structure as well as
its function. In this research, we use different Hidden Markov models for protein secondary structure prediction. In this paper we have proposed an algorithm for predicting protein secondary structure. We have used Hidden Markov model with sliding window for secondary structure prediction.The secondary structure has three regular forms, for each secondary structural element we are using one Hidden Markov
Model.
Index Terms—Bioinformatics, Hidden Markov Model, Homologous modeling, Protein folding, Structure Prediction.
—————————— ——————————
1 INTRODUCTION
P
ROTEINS, the fundamental molecules of all organisms have three dimensional structures that are fully
specified by sequence of amino acids.
Normally the amino acids are specified by a unique
one-letter code. The sequence of amino acids in a given
protein is called the primary structure and it is believed
that the 3D-structure of most proteins is derived from
their primary structures. Determining protein structure
from its amino acid sequence would greatly help in understanding the structure-function relationship. For instance, by determining the structures of viral proteins it
would enable researchers to design drugs for specific viruses. At present, 100% accurate protein structures are
determined experimentally using X-ray crystallographic
or Nuclear Magnetic Resonance (NMR) techniques. However these methods are not feasible because they are tedious and time consuming, taking months or even years
to complete. In addition, large-scale sequencing projects
(such as the Human Genome Project) produce protein
sequences at a very fast pace. As a result, the gap between
the number of known protein sequences and the number
of known structures is getting larger. Protein structure
prediction aims at reducing this sequence structure gap.
Until now, however, the protein structure cannot be theoretically predicted 100% accurately. This is due to the fact
that there are 20 different amino acids mixed with water
and carbohydrates to produce primary protein structure
(Amino Acid Sequence).Thus there are too many ways in
which similar structures can be generated in proteins by
different amino acid sequences. The secondary structure
has three regular forms: Helical (alpha (α) helices), Extended (beta (β) sheets) and Loops (also called reverse
turns or coils). In the problem of the protein secondary
structure prediction, the inputs are the amino acid sequences while the output is the predicted structure also
called conformation, which is the combination of alpha
helices, beta sheet and loops. A typical protein sequence
and its conformation class are shown below:
Protein Sequence:
ABABABABCCQQFFFAAAQQAQQA
Conformation Class:
HHHHCCCCCEEEECCCHHHHHHC
H means Helical, E means Extended, and C’s are the
Coiled conformations.
2 RELATED WORK
A lot of work has been done on predicting secondary
structures, and over the last 10 to 20 years the methods
have gradually improved in accuracy. Some of the first
work on the secondary structure prediction was based on
statistical methods in which the likelihood of each amino
acid being in one of the three types of secondary structures was estimated from known protein structures.
These probabilistic were then averaged in some way over
a small window to obtain the prediction. Around 1988 the
————————————————
first attempts were made to use neural networks to preK.K.Senapati is with the Department of Computer Science and Engineerdict protein secondary structures. The accuracy of the
ing, Birla Insyitute of Technology, Mesra, Ranchi, Jharkhand 835215, Inpredictions made by Qian Sjnowskiseemed is better than
dia.
G. Sahoo is with the Department of Information Technology, Birla Institute those obtained by previous methods and was reported to
of Technology, Mesra, Ranchi, Jharkhand 835215, India.
be in the range of 62.7-64.4%. Rost and Sander have deD. Bhaumik is the student of Department of Computer Science and Engiveloped the prediction mail server called PHD with a
neering, Birla Institute of Technology, Mesra, Ranchi, Jharkhand 835215,
prediction accuracy of 71.6% was reported.
India.
69
In our proposed algorithm instead of constructing One
Hidden Markov model for three states of secondary structural elements. We propose three separate Hidden Markov Models i.e. for every secondary structural element we
construct One Hidden Markov Model. Every Hidden
Markov Model will give a probability. Out of these three
probabilities which Hidden Markov Model is giving maximum probability that becomes the secondary structure of
that particular amino acid.
3 SECONDARY STRUCTURE ELEMENTS
INTERPRETATION
We follow a three state identification of secondary structures, namely Helix (H), Coil(C), Extended (β-sheet)(E)
because it provides a consistent set of structure assignment. We have utilized a reduced version of DSSP (Database of Secondary Structure in Proteins) classification that
uses eight types of secondary structure assignment’s(αhelix), E(extended β-strand), G(310 helix), I(∏-helix),
B(bridge, a single residue β strand), T(β-turn), S(bend),
and C(coil).
DSSP alphabet can be defined [11]:
Helices(H) = {H,G,I}, Strand(E)={E,B}, Coil(C)={T, S, C}.
4 SSP_H ALGORITHM
In our proposed algorithm we have taken the input as
amino acid sequence, where n is the length of the sequence. Initially we assume probability for amino acid
sequence as input to the Hidden Markov Model.
The algorithm is described as SSP_H (Secondary Structure Prediction using Hidden Markov Model).
SSP_H(S, n)
{
S=Sequence of Amino Acid
n = The length of the Amino Acid
H=Secondary structure sequence
// Input
//Input
//Initially it is null
For i=5 to n-5
{
B [1]=Helix_hmm(S[i-5,i+5])
B [2]=Coil_hmm(S[i-5,i+5])
B[3]=Strand_hmm(S[i-5,i+5])
c= max (B) // B is an array
If c=B[1]
H*i+ =’H’
//helix
else if c=B[2]
H*i+ =’C’
//coil
else
H*i+ =’N’
//strand
}
Print array H // Output
//This is the secondary structure of the given Amino Acid
sequence.
}
In the above algorithm the functions Helix_hmm(),
Coil_hmm(), and Strand_hmm() take a window sequence
of amino acids as input but return the maximum probability of the ith amino acid having Helical structure, Coil
structure and Strand structure respectively.
4.1 Flowchart for Protein Secondary Structure
Algorithm
70
5 EXAMPLE
In the given example S1 is the initial state and S2 is the
next state. The inputs of Amino Acid sequence are feed to
the SSP_H () to get the desired output.
Fig. 3. Hidden Markov Model for Helix (H)
Fig. 1. Hidden Markov Model for Coil (H)
5.1 Amino Acid Sequence
A typical Amino Acid sequence is as follows:
For predicting the secondary structure of amino acid “C”,
we take 2 of its left neighbors as well as 2 of its right
neighbors.Now this window of amino acids<A, C, C, T,
G> is giving to 3 Hidden Markov models. In every Hidden Markov Model we find the optimal path for the window of amino acids by using VITERBI Algorithm. Finally
for all of these 3 optimal paths we calculate 3 respective
probabilities. The Hidden Markov model that gives the
maximum probability will become the secondary structure of that particular amino acid.
H
E
C
Max_probability=Max {0.3, 0.4, 0.6}
= 0.6.
Where, H=Helix, E=Strand, C=Coil.
So Secondary Structure for the amino acid “C” is Coil.
6 CONCLUSION
Fig. 2. Hidden Markov Model for Strand (E)
Prediction accuracy of this model is based on the following factors. Weights of each Hidden Markov Model, i.e.
Emission and transition matrices of all Hidden Markov
Models. Training of every Hidden Markov Model with
existed data, and similarity between all the training sequences is also effects on the prediction accuracy. On the
71
base of number of states used in every Hidden Markov
model, this factor is also affects the prediction accuracy.
When we are constructing a Hidden Markov Model all
these factors we have to consider. And also different combinations of all these factors will produce different accuracy measures. In this approach we achieved 64% accuracy for helical structure. Currently we are working to
achive the same accuracy in beta-sheet as well as coil.
K. K. Senapati recived his M.Sc and M.Tech degree in Computer
Science and Engineering from UTKAL university, Orissa, India. He
has worked in a project Task Scheduling in Multiprocessor at Central
University, Hydrerabad. He has obtained CCNA from IIIT Hydrebad
as well as CCI from Asian Schol of Cyber Law, Pune. His research is
focused on Bioinformatics and Pattern Matching. He is associated
with Birla Institute of Technology, Mesra, Ranchi, India since 2005.
He is currenly working as a Senior Lecturer in the department of
Computer Science and Engineering.
G. Sahoo received his P.G. degree from Utkal University in the
7 REFERENCES
[1]
H. Cheng, T. Z. Sen, A. Kloczkowski, D. Margaritis and R. L.
Jernigan, “Prediction of Protein Secandary Structure by mining
Structural Fragment Database.”, vol.46, no.12, pp.4314-4321.
[2] J. Ruana, K. Wanga, J. Yanga, L. A. Kurganb, K. Cioscde, “Highly Accurate and Consistent Method for Prediction of Helix and
Strand Content from Primary Protein Sequences.”, vol.35, no.1,
pp.19-35, 2005.
[3] L. R. Rabiner, “A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition.”, Proc. IEEE, vol. 77,
no. 2, pp.257-286, 1989.
[4] J. Martin, J. F. Gilbart, and F. Rodolphe, “Choosing the optimal
Hidden Markov Model for Secondary Protein Structure Prediction”, French National Institute of Agriculture Research.
[5] A. Krogh, “Hidden Markov Models for Labeled Sequences,”
Proc. 12th Int’l Conf. Pattern Recognition (ICPR 94), IEEE CS Press,
pp. 140-144, 1994.
[6] S. L. Salzberg, D. B. Searls and S. Kasif “In Computational Methods in Molecular Biology”, pages 45-63. Elsevier, 1998.
[7] S. Kwong, C. Chau, k. Man, and K. Tang, “Optimization Hmm
Topology and Its Model Parameters by Genetic Algorithms”,
Pattern Recognition, vol. 34, pp.509-522, 2001.
[8] Rost B, Sander C, “Prediction of Protein Secondary Structure at
Better Than 70% Accuracy”, Journal of Molecular Biology, 232,
pp. 584-599, 1993.
[9] Rost B, Sander C”Combining Evolutionary Information and
Neural Networks to Predict Protein Secondary Structure”,
PROTEINS: Structure, Function and Genetics, 19, pp. 55-72,
1994.
[10] W. H. Press, B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling,
Numerical Recipes in C: The Art of Scientific Computing, Cambridge University Press, New York, 1992.
[11] Chandonia J M, Karplus M, “The Importance of Large Data Sets
For Protein Secondary Structure Prediction with Neural Networks”, Protein Science, 5, pp. 768-774, 1996.
[12] Qian N and Sejnowski TJ. 1998. Predicting the Secondary
Structure of Globular Proteins using Neural Network Models.
Journal of Molecular Biology. Vol. 202. Pages:865-884.
year 1980 and Ph.D degree in the area of Computational Mathematics from Indian Institute of Technology, Kharagpur in the year 1987.
He is associated with Birla Institute of Technology, Mesra, Ranchi,
India since 1988. He is currently working as a professor and heading
the Department of Information Technology. His research interest
includes theoretical computer science, parallel and distributed computing, data security, image processing and pattern recognition.
.
D. Bhaumik recived his B.Tech degree in Information Technology
from West Bengal University of Technology, West Bengal, India. He
has obtained OCA certification from SQL Star International Ltd.,
Kolkata. He is pursuing his ME degree in Softwre Engineering at
Birla Institute of Technology, Mesra, Ranchi, India.
| 5cs.CE
|
2
e
Dynamic Bridge-Finding in O(log
n) Amortized Time
Jacob Holm∗, Eva Rotenberg, and Mikkel Thorup∗
arXiv:1707.06311v2 [cs.DS] 22 Aug 2017
University of Copenhagen (DIKU),
[email protected], [email protected], [email protected]
August 23, 2017
Abstract
We present a deterministic fully-dynamic data structure for maintaining information about
e
the bridges in a graph. We support updates in O((log
n)2 ) amortized time, and can find a
bridge in the component of any given vertex, or a bridge separating any two given vertices,
in O(log n/ log log n) worst case time. Our bounds match the current best for bounds for
deterministic fully-dynamic connectivity up to log log n factors.
e
The previous best dynamic bridge finding was an O((log
n)3 ) amortized time algorithm by
Thorup [STOC2000], which was a bittrick-based improvement on the O((log n)4 ) amortized time
algorithm by Holm et al.[STOC98, JACM2001].
Our approach is based on a different and purely combinatorial improvement of the algorithm
e
of Holm et al., which by itself gives a new combinatorial O((log
n)3 ) amortized time algorithm.
e
Combining it with Thorup’s bittrick, we get down to the claimed O((log
n)2 ) amortized time.
Essentially the same new trick can be applied to the biconnectivity data structure from
e
[STOC98, JACM2001], improving the amortized update time to O((log
n)3 ).
We also offer improvements in space. We describe a general trick which applies to both of our
new algorithms, and to the old ones, to get down to linear space, where the previous best use
O(m + n log n log log n). Finally, we show how to obtain O(log n/ log log n) query time, matching
the optimal trade-off between update and query time.
Our result yields an improved running time for deciding whether a unique perfect matching
exists in a static graph.
∗
This research is supported by Mikkel Thorup’s Advanced Grant DFF-0602-02499B from the Danish Council for
Independent Research under the Sapere Aude research career programme.
1
Introduction
In graphs and networks, connectivity between vertices is a fundamental property. In real life, we
often encounter networks that change over time, subject to insertion and deletion of edges. We call
such a graph fully dynamic. Dynamic graphs call for dynamic data structures that maintain just
enough information about the graph in its current state to be able to promptly answer queries.
Vertices of a graph are said to be connected if there exists a path between them, and k-edge
connected if no sequence of k − 1 edge deletions can disconnect them. A bridge is an edge whose
deletion would disconnect the graph. In other words, a pair of connected vertices are 2-edge
connected if they are not separated by a bridge. By Menger’s Theorem [17], this is equivalent to
saying that a pair of connected vertices are two-edge connected if there exist two edge-disjoint paths
between them. By edge-disjoint is meant that no edge appears in both paths.
For dynamic graphs, the first and most fundamental property to be studied was that of dynamic
connectivity. In general, we can assume the graph has a fixed set of n vertices, and we let m denote
√
the current number of edges in the graph. The first data structure with sublinear O( n) update
time is due to Frederickson [5] and Eppstein et al. [4]. Later, Frederickson [6] and Eppstein et
√
al. [4] gave a data structure with O( n) update time for two-edge connectivity. Henzinger and
King achieved poly-logarithmic expected amortized time [8], that is, an expected amortized update
time of O((log n)3 ), and O(log n/ log log n) query time for connectivity. And in [9], O((log n)5 )
expected amortized update time and O(log n) worst case query time for 2-edge connectivity. The
first polylogarithmic deterministic result was by Holm et al in [10]; an amortized deterministic
update time of O((log n)2 ) for connectivity, and O((log n)4 ) for 2-edge connectivity. The update
time for deterministic dynamic connectivity has later been improved to O((log n)2 / log log n) by
Wulff-Nilsen [21]. Sacrificing determinism, an O(log n(log log n)3 ) structure for connectivity was
presented by Thorup [20], and later improved to O(log n(log log n)2 ) by Huang et al. [12]. In the
same paper, Thorup obtains an update time of O((log n)3 log log n) for deterministic two-edge
connectivity. Interestingly, Kapron et al. [13] gave a Monte Carlo-style randomized data structure
with polylogarithmic worst case update time for dynamic connectivity, namely, O((log n)4 ) per edge
insertion, O((log n)5 ) per edge deletion, and O(log n/ log log n) per query. We know of no similar
result for bridge finding. The best lower bound known is by Pǎtraşcu et al. [18], which shows a
t
trade-off between update time tu and query time tq of tq lg ttuq = Ω(lg n) and tu lg tuq = Ω(lg n).
1.1
Our results
We obtain an update time of O((log n)2 (log log n)2 ) and a query time of O(log n/ log log n) for the
bridge finding problem:
Theorem 1. There exists a deterministic data structure for dynamic multigraphs in the word RAM
model with Ω(log n) word size, that uses O(m + n) space, and can handle the following updates, and
queries for arbitrary vertices v or arbitrary connected vertices v, u:
• insert and delete edges in O((log n)2 (log log n)2 ) amortized time,
• find a bridge in v’s connected component or determine that none exists, or find a bridge that
separates u from v or determine that none exists. Both in O(log n/ log log n) worst-case time
for the first bridge, or O(log n/ log log n + k) worst case time for the first k bridges.
1
• find the size of v’s connected component in O(log n/ log log n) worst-case time, or the size of
its 2-edge connected component in O(log n(log log n)2 ) worst-case time.
Since a pair of connected vertices are two-edge connected exactly when there is no bridge
separating them, we have the following corollary:
Corollary 2. There exists a data structure for dynamic multigraphs in the word RAM model with
Ω(log n) word size, that can answer two-edge connectivity queries in O(log n/ log log n) worst case
time and handle insertion and deletion of edges in O((log n)2 (log log n)2 ) amortized time, with space
consumption O(m + n).
Note that the query time is optimal with respect to the trade-off by Pǎtraşcu et al. [18]
As a stepping stone on the way to our main theorem, we show the following:
Theorem 3. There exists a combinatorial deterministic data structure for dynamic multigraphs
on the pointer-machine without the use of bit-tricks, that uses O(m + n) space, and can handle
insertions and deletions of edges in O((log n)3 log log n) amortized time, find bridges and determine
connected component sizes in O(log n) worst-case time, and find 2-edge connected component sizes
in O((log n)2 log log n) worst-case time.
Our results are based on modifications to the 2-edge connectivity data structure from [11].
Applying the analoguous modification to the biconnectivity data structure from the same paper
yields a structure with O((log n)3 (log log n)2 ) amortized update time and O((log n)2 (log log n)2 )
worst case query time. The details of this modification are beyond the scope of this paper.
1.2
Applications
While dynamic graphs are interesting in their own right, many algorithms and theorem proofs for
static graphs rely on decremental or incremental graphs. Take for example the problem of whether
or not a graph has a unique perfect matching? The following theorem by Kotzig immediately yields
a near-linear algorithm if implemented together with a decremental two-edge connectivity data
structure with poly-logarithmic update time:
Theorem 4 (A. Kotzig ’59 [16]). Let G be a connected graph with a unique perfect matching M .
Then G has a bridge that belongs to M .
The near-linear algorithm for finding a unique perfect matching by Gabow, Kaplan, and Tarjan [7]
is straight-forward: Find a bridge and delete it. If deleting it yields connected components of odd
size, it must belong to the matching, and all edges incident to its endpoints may be deleted—if the
components have even size, the bridge cannot belong to the matching. Recurse on the components.
Thus, to implement Kotzig’s Theorem, one has to implement three operations: One that finds a
bridge, a second that deletes an edge, and a third returning the size of a connected component.
Another example is Petersen’s theorem [19] which states that any cubic, two-edge connected
graph contains a perfect matching. An algorithm by Biedl et al. [2] finds a perfect matching in
such graphs in O(n log4 n) time, by using the Holm et al two-edge connectivity data structure as a
subroutine. In fact, one may implement their algorithm and obtain running time O(nf (n)), by using
as subroutine a data structure for amortized decremental two-edge connectivity with update-time
f (n). Here, we thus improve the running time from O(n(log n)3 log log n) to O(n(log n)2 (log log n)2 ).
2
In 2010, Diks and Stanczyk [3] improved Biedl et al.’s algorithm for perfect matchings in
two-edge connected cubic graphs, by having it rely only on dynamic connectivity, not two-edge
connectivity, and thus obtaining a running time of O(n(log n)2 / log log n) for the deterministic
version, or O(n log n(log log n)2 ) expected running time for the randomized version. However, our
data structure still yields a direct improvement to the original algorithm by Biedl et al.
Note that all applications to static graphs have in common that it is no disadvantage that our
running time is amortized.
1.3
Techniques
As with the previous algorithms, our result is based on top trees [1] which is a hierarchical tree
structure used to represent information about a dynamic tree — in this case, a certain spanning
tree of the dynamic graph. The original O((log n)4 ) algorithm of Holm et al. [11] stores O((log n)2 )
counters with each top tree node, where each counter represent the size of a certain subgraph. Our
new O((log n)3 ) algorithm applies top trees the same way, representing the same O((log n)2 ) sizes
with each top tree node, but with a much more efficient implicit representation of the sizes.
Reanalyzing the algorithm of Holm et al. [11], we show that many of the sizes represented in
the top nodes are identical, which implies that that they can be represented more efficiently as
a list of actual differences. We then need additional data structures to provide the desired sizes,
and we have to be very careful when we move information around as the the top tree changes, but
overall, we gain almost a log-factor in the amortized time bound, and the algorithm remains purely
combinatorial.
Our combinatorial improvement can be composed with the bittrick improvement of Thorup [20].
Thorup represents the same sizes as the original algorithm of Holm et al., but observes that we
don’t need the exact sizes, but just a constant factor approximation. Each approximate size can
be represented with only O(log log n) bits, and we can therefore pack Ω(log n/ log log n) of them
together in a single Ω(log n)-bit word. This can be used to reduce the cost of adding two O(log n)dimensional vectors of approximate sizes from O(log n) time to O(log log n) time. It may not be
obvious from the current presentation, but it was a significant technical difficulty when developing
our O((log n)3 log log n) algorithm to make sure we could apply this technique and get the associated
speedup to O((log n)2 (log log n)2 ).
The “natural” query time of our algorithm is the same as its update time. In order to reduce
the query time, we observe that we can augment the main algorithm to maintain a secondary
structure that can answer queries much faster. This can be used to reduce the query time for the
combinatorial algorithm to O(log n), and for the full algorithm to the optimal O(log n/ log log n).
The secondary structure needed for the optimal O(log n/ log log n) query time uses top trees of
degree O(log n/ log log n). While the use of non-binary trees is nothing new, we believe we are the
first to show that such top trees can be maintained in the “natural” time.
Finally, we show a general technique for getting down to linear space, using top trees whose
base clusters have size Θ(logc n).
1.4
Article outline
In Section 2, we recall how [11] fundamentally solves two-edge connectivity via a reduction to a
certain set of operations on a dynamic forest. In Section 3, we recall how top trees can be used to
maintain information in a dynamic forest, as shown in [1]. In Sections 4, 5, and 6, we describe how
3
to support the operations on a dynamic tree needed to make a combinatorial O((log n)3 log log n)
algorithm for bridge finding, as stated in Theorem 3. Then, in Section 7, we show how to use
Approximate Counting to get down to O((log n)2 (log log n)2 ) update time, thus, reaching the update
time of Theorem 1. We then revisit top trees in Section 8, and introduce the notion of B-ary top
trees, as well as a general trick to save space in complex top tree applications. We proceed to show
how to obtain the optimal Θ(log n/ log log n) query time in Section 9. Finally, in Section 10, we
show how to achieve optimal space, by only storing cluster information with large clusters, and
otherwise calculating it from scratch when needed.
2
Reduction to operations on dynamic trees
In [11], two-edge connectivity was maintained via operations on dynamic trees, as follows. For each
edge e of the graph, the algorithm explicitly maintains a level, `(e), between 0 and `max = blog2 nc
such that the edges at level `max form a spanning forest T , and such that the
2-edge-connected
components in the subgraph induced by edges at level at least i have at most n/2i vertices. For
each edge e in the spanning forest, define the cover level, c(e), as the maximum level of an edge
crossing the cut defined by removing e from T , or −1 if no such edge exists. The cover levels are
only maintained implicitly, because each edge insertion and deletion can change the cover levels of
Ω(n) edges. Note that the bridges are exactly the edges in the spanning forest with cover level −1.
The algorithm explicitly maintains the spanning forest T using a dynamic tree structure supporting
the following operations:
1. Link(v, w). Add the edge (v, w) to the dynamic tree, implicitly setting its cover level to −1.
2. Cut(v, w). Remove the edge (v, w) from the dynamic tree.
3. Connected(v, w). Returns true if v and w are in the same tree, false otherwise.
4. Cover(v, w, i). For each edge e on the tree path from v to w whose cover level is less than i,
implicitly set the cover level to i.
5. Uncover(v, w, i). For each edge e on the tree path from v to w whose cover level is at most i,
implicitly set the cover level to −1.
6. CoverLevel(v). Return the minimal cover level of any edge in the tree containing v.
7. CoverLevel(v, w). Return the minimal cover level of an edge on the path from v to w. If
v = w, we define CoverLevel(v, w) = `max .
8. MinCoveredEdge(v). Return any edge in the tree containing v with minimal cover level.
9. MinCoveredEdge(v, w). Returns a tree-edge on the path from v to w whose cover level is
CoverLevel(v, w).
10. AddLabel(v, l, i). Associate the user label l to the vertex v at level i.
11. RemoveLabel(l). Remove the user label l from its vertex vertex(l).
12. FindFirstLabel(v, w, i). Find a user label at level i such that the associated vertex u has
CoverLevel(u, meet(u, v, w)) ≥ i and minimizes the distance from v to meet(u, v, w).
13. FindSize(v, w, i). Find the number of vertices u such that CoverLevel(u, meet(u, v, w)) ≥ i.
Note that FindSize(v, v, −1) is just the number of vertices in the tree containing v.
4
Lemma 5 (Essentially the high level algorithm from [11]). There exists a deterministic reduction
for dynamic graphs with n nodes, that, when starting with an empty graph, supports any sequence
of m Insert or Delete operations using:
• O(m) calls to Link, Cut, Uncover, and CoverLevel.
• O(m log n) calls to Connected, Cover, AddLabel, RemoveLabel, FindFirstLabel, and FindSize.
And that can answer FindBridge queries using a constant number of calls to Connected, CoverLevel,
and MinCoveredEdge.
Proof. See Appendix A for a proof and pseudocode.
#
Operation
1
2
3
4
5
6
7
8
9
10
11
12
Link(v, w, e)
Cut(e)
Connected(v, w)
Cover(v, w, i)
Uncover(v, w, i)
CoverLevel(v)
CoverLevel(v, w)
MinCoveredEdge(v)
MinCoveredEdge(v, w)
AddLabel(v, l, i)
RemoveLabel(l)
FindFirstLabel(v, w, i)
FindSize(v, w, i)
FindSize(v, v, −1)
13
4
Asymptotic worst case time per call, using structure in section
5
6
7
9
f (n) log n
log f (n)
log n
(log n)2 log log n
log n log log n
log n(log log n)2
log n
log f (n)
-
-
log n log log n
-
-
log n
(log n)2 log log n
log n
log n
log n(log log n)2
log n
log n
log f (n)
-
Table 1: Overview of the worst case times achieved for each tree operation by the data structures
presented in this paper. In the last column, f (n) ∈ O( logloglogn n ) can be chosen arbitrarily.
The algorithm in [11] used a dynamic tree structure supporting all the operations in O((log n)3 )
time, leading to an O((log n)4 ) algorithm for bridge finding. Thorup [20] showed how to improve
the time for the dynamic tree structure to O((log n)2 log log n) leading to an O((log n)3 log log n)
algorithm for bridge finding.
Throughout this paper, we will show a number of data structures for dynamic trees, implementing
various subsets of these operations while ignoring the rest (See Table 1). Define a CoverLevel
structure to be one that implements operations 1–9, and a FindSize structure to be a CoverLevel
structure that additionally implements the FindSize operation. Finally, we define a FindFirstLabel
structure to be one that implements operations 1–12 (all except for FindSize).
The point is that we can get different trade-offs between the operation costs in the different
structures, and that we can combine them into a single structure supporting all the operations using
the following
Lemma 6 (Folklore). Given two data structures S and S 0 for the same problem consisting of a set
U of update operations and a set Q of query operations. If the respective update times are fu (n) and
fu0 (n) for u ∈ U , and the query times are gq (n) and gq0 (n) for q ∈ Q, we can create a combined data
5
structure running in O(fu (n) + fu0 (n)) time for update operation u ∈ U , and O(min gq (n), gq0 (n) )
time for query operation q ∈ Q.
Proof. Simply maintain both structures in parallel. Call all update operations on both structures,
and call only the fastest structure for each query.
Proof of Theorem 3. Use the CoverLevel structure from Section 4, the FindSize structure from
Section 5, and the FindFirstLabel structure from Section 6, and combine them into a single structure
using Lemma 6. Then the reduction from Lemma 5 gives the correct running times but uses
O(m + n log n) space. To get linear space, modify the FindSize and FindFirstLabel structures as
described in Section 10.
Proof of Theorem 1. Use the CoverLevel structure from Section 9, the FindSize structure from
Section 5, as modified in Section 7 and 10, and the FindFirstLabel structure from Section 6, and
combine them into a single structure using Lemma 6. Then the reduction from Lemma 5 gives the
required bounds.
3
Top trees
A top tree is a data structure for maintaining information about a dynamic forest. Given a tree T ,
a top tree T is a rooted tree over subtrees of T , such that each non-leaf node is the union of its
children. The root of T is T , its leaves are the edges of T , and its nodes are clusters, which we will
define in two steps. For any subgraph H of a graph G, the boundary ∂H consists of the vertices of
H that have a neighbour in G \ H. A cluster is a connected subgraph with a boundary of size no
larger than 2. We denote them by point clusters if the boundary has size ≤ 1, and path clusters
otherwise. For a path cluster C with boundary ∂C = {u, v}, denote by π(C) the tree path between
u and v, also denoted the cluster path of C. Similarly, for a point cluster C with boundary vertex v,
π(C) is the trivial path consisting solely of v. The top forest supports dynamic changes to the forest:
insertion (link) or deletion (cut) of edges. Furthermore, it supports the expose operation: expose(v),
or expose(v1 , v2 ), returns a top tree where v, or v1 , v2 , are considered boundary vertices of every
cluster containing them, including the root cluster. All operations are supported by performing
a series of destroy, create, split, and merge operations: split destroys a node of the top tree and
replaces it with its two children, while merge creates a parent as a union of its children. Destroy and
create are the base cases for split and merge, respectively. Note that clusters can only be merged if
their union has a boundary of size at most 2.
A top tree is binary if each node has at most two children. We call a non-leaf node heterogeneous
if it has both a point cluster and a path cluster among its children, and homogeneous otherwise.
Theorem 7 (Alstrup, Holm, de Lichtenberg, Thorup [1]). For a dynamic forest on n vertices we
can maintain binary top trees of height O(log n) supporting each link, cut or expose with a sequence
of O(1) calls to create or destroy, and O(log n) calls to merge or split. These top tree modifications
are identified in O(log n) time. The space usage of the top trees is linear in the size of the dynamic
forest.
6
4
A CoverLevel structure
In this section we show how to maintain a top tree supporting the CoverLevel operations. This part
is is essentially the same as in [10,11] (with minor corrections), but is included here for completeness
because the rest of the paper builds on it. Pseudocode for maintaining this structure is given in
Appendix B.
For each cluster C we want to maintain the following two integers and up to two edges:
coverC := min {c(e) | e ∈ π(C)} ∪ {`max }
globalcoverC := min {c(e) | e ∈ C \ π(C)} ∪ {`max }
minpathedgeC := arg min c(e) if |∂C| = 2, and nil otherwise
e∈π(C)
minglobaledgeC := arg min c(e) if C 6= π(C), and nil otherwise
e∈C\π(C)
Then
CoverLevel(v) = globalcoverC
)
where C is the point cluster returned by Expose(v)
MinCoveredEdge(v) = minglobaledgeC
CoverLevel(v, w) = coverC
MinCoveredEdge(v, w) = minpathedgeC
)
where C is the path cluster returned by Expose(v, w)
The problem is that when handling Cover or Uncover we cannot afford to propagate the
information all the way down to the edges. When these operations are called on a path cluster
C, we instead implement them directly in C, and then store “lazy information” in C about what
should be propagated down in case we want to look at the descendants of C. The exact additional
information we store for a path cluster C is
cover−
C := max level of a pending Uncover, or −1
cover+
C := max level of a pending Cover, or −1
−
+
We maintain the invariant that coverC ≥ cover+
C , and if coverC ≤ coverC then coverC = coverC .
This allows us to implement Cover(v, w, i) by first calling Expose(v, w), and then updating the
returned path cluster C as follows:
+
coverC = max {coverC , i}
cover+
C = max coverC , i
Similarly, we can implement Uncover(v, w, i) by first calling Expose(v, w), and then updating the
returned path cluster C as follows if coverC ≤ i:
−
coverC = −1
cover+
cover−
C = −1
C = max coverC , i
−
+
Together, cover
C and coverC represent the fact that for each path descendant D of C, if
+ 1
+
coverD ≤ max cover−
, we need to set cover
whenever a path
C , coverC
D = coverC . −In particular
cluster C is split, for each path child D of C, if max coverD , coverD ≤ cover−
we
need to set
C
−
cover−
D = coverC
1
In [10, 11] this condition is erroneously stated as coverD ≤ cover−
C.
7
+
Furthermore, if coverD ≤ max cover−
C , coverC we need to set
coverD = cover+
C
+
cover+
D = coverC
Note that only coverD is affected. None of globalcoverD , minpathedgeD , or minglobaledgeD depend
directly on the lazy information.
Now suppose we have k clusters2 A1 , . . . , Ak that we want to merge into a single new cluster C.
For 1 ≤ i ≤ k define
(
globalcoverAi
if ∂Ai ⊆ π(C) or globalcoverAi ≤ coverAi
0
globalcoverC,Ai :=
coverAi
otherwise
(
minglobaledgeAi if ∂Ai ⊆ π(C) or globalcoverAi ≤ coverAi
minglobaledge0C,Ai :=
minpathedgeAi
otherwise
Note that for a point-cluster Ai , globalcoverAi is always ≤ coverAi = lmax .
We then have the following relations between the data of the parent and the data of its children:
coverC = `max if |∂C| < 2, otherwise
min
1≤i<k,∂Ai ⊆π(C)
coverAi
minpathedgeC = nil if |∂C| < 2, otherwise minpathedgeAj where j =
arg min
1≤i<k,∂Ai ⊆π(C)
coverAi
globalcoverC = min globalcover0C,Ai
1≤i<k
minglobaledgeC = minglobaledge0C,Aj
cover−
C
cover+
C
where j = arg min globalcover0C,Ai
1≤i<k
= −1
= −1
Analysis For any constant-degree top tree, Merge and Split with this information takes constant
time, and thus, all operations in the CoverLevel structure in this section take O(log n) time. Each
cluster uses O(1) space, so the total space used is O(n).
Note that we can extend this so for each cluster C, if all the least-covered edges (on or off the
cluster path) lie in the same child of C, we have a pointer to the closest descendant D of C that is
either a base cluster or has more than one child containing least-covered edges. We can use this
structure to find the first k bridges in O(log n + k) time.
5
A FindSize Structure
We now proceed to show how to extend the CoverLevel structure from Section 4 to support FindSize
in O(log n log log n) time per Merge and Split. Later, in Section 7 we will show how to reduce this
to O((log log n)2 ) time per Merge and Split. See Appendix C for pseudocode.
We will use the idea of having a single vertex label for each vertex, which is a point cluster with
no edges, having that vertex as boundary vertex and containing all relevant information about the
vertex. The advantage of this is that it simplifies handling of the common boundary vertex during a
merge by making sure it is uniquely assigned to (and accounted for by) one of the children.
2
k = 2 for now, but we will reuse this in section 9 with a higher-degree top tree.
8
Let C be a cluster in T , let v be a vertex in C, and let 0 ≤ i < `max . Define
pointsizeC,v,i := {u ∈ C | CoverLevel(u, v) ≥ i}
For convenience, we will combine all the O(log n) levels together into a single vector3
pointsizeC,v := pointsizeC,v,i {0≤i<` }
max
Let (Cv ){v∈π(C)} be the point clusters that would result from deleting the edges of π(C) from C.
Then we can define the vector
X
sizeC :=
pointsizeCm ,m
m∈π(C)
Note that with this definition, if ∂C = {v} then pointsizeC,v = sizeC so even when v = w we have
FindSize(v, w, i) = sizeC,i
where C = Expose(v, w)
So for any cluster C, the sizeC vector is what we want to maintain.
The main difficulty turns out be computing the sizeC vector for the heterogeneous point clusters.
To help with that we will for each cluster C and boundary vertex v ∈ ∂C additionally maintain the
following two size vectors for each −1 ≤ i ≤ `max :
X
partsizeC,v,i :=
pointsizeCm ,m
diagsizeC,v,i := M(i) · partsizeC,v,i
m∈π(C)
CoverLevel(v,m)=i
Where M(i) is a diagonal matrix whose entries are defined (using Iverson brackets, see [14]) by
M(i)jj = [i ≥ j]
+
Note that these vectors are independent of cover−
C and coverC as defined in Section 4. The
corresponding “clean” vectors are not explicitly stored, but computed when needed as follows
if i > `
C,v,i
partsize
P`
0
+
partsizeC,v,i =
partsize
if
i
=
cover
C,v,j
j=−1
C
0
otherwise
+
where ` = max cover−
C , coverC
if i > `
C,v,i
diagsize
P`
0
+
diagsizeC,v,i = M(i) · j=−1 partsizeC,v,j if i = coverC
0
otherwise
The point of these definitions is that each path cluster inherits most of its partsize and diagsize
vectors from its children, and we can use this fact to get an O(`max / log `max ) = O(log n/ log log n)
speedup compared to [11].
3
All vectors and matrices in this section have indices ranging from 0 to `max − 1.
9
Merging along a path (the general case) Let A, B be clusters that we want to merge into a
new cluster C, and suppose ∂A ∪ ∂B ⊆ π(C). This covers all types of merge in a normal binary top
tree, except for the heterogeneous point clusters. Let ∂A ∩ ∂B = {c}. If |∂C| = 1, let a = b = c,
otherwise let ∂C = {a, b} with a ∈ ∂A, b ∈ ∂B. Then
sizeC = sizeA + sizeB
0
partsizeA,a,i P
max
partsizeC,a,i = partsize0A,a,i + `j=i
partsize0B,c,j
partsize0B,c,i
0
diagsizeA,a,i
P max
diagsizeC,a,i = diagsize0A,a,i +M(i) · `j=i
partsize0B,c,j
diagsize0B,c,i
if i > coverA
if i = coverA
if i < coverA
if i > coverA
if i = coverA
if i < coverA
Merging off the path (heterogeneous point clusters) Now let A be a path cluster with
∂A = {a, b}, let B be a point cluster with ∂B = {b}, and suppose we want to merge A, B into a
new point cluster C with ∂C = {a}. Then
!
`X
max
0
sizeC =
diagsizeA,a,i + M(coverA ) · sizeB
i=−1
partsizeC,a,i
(
sizeC
=
0
if i = `max
otherwise
diagsizeC,a,i = partsizeC,a,i
Analysis The advantage of our new approach is that each merge or split is a constant number of
splits, concatenations, searches, and sums over O(`max )-length lists of `max -dimensional vectors. By
representing each list as an augmented balanced binary search tree (see e.g. [15, pp. 471–475]), we
can implement each of these operations in O(`max log `max ) time, and using O(`max ) space per cluster,
as follows. Let C be a cluster and let v ∈ ∂C. The tree has one node for each key i, −1 ≤ i ≤ `max
such that partsizeC,v,i is nonzero, augmented with the following additional information:
key := i
partsize := partsizeC,v,i
diagsize := diagsizeC,v,i
X
partsizesum :=
partsizeC,v,j
j descendant of i
diagsizesum :=
X
partsizeC,v,j
j descendant of i
Each split, concatenate, search, or sum operation can be implemented such that it touches O(log `max )
nodes, and the time for each node update is dominated by the time it takes to add two `max dimensional vectors, which is O(`max ). The total time for each Cover, Uncover, Link, Cut, or
FindSize is therefore O(log n · `max · log `max ) = O((log n)2 log log n), and the total space used for
the structure is O(n · `max ) = O(n log n).
10
Comparison to previous algorithms For any path cluster C and vertex v ∈ ∂C, let SC,v be
the matrix whose jth column 0 ≤ j < `max is defined by
T
(SC,v
)j :=
`X
max
partsize0C,v,k
k=j
Then SC,v is essentially the size matrix maintained for path clusters in [10, 11, 20]. Notice that
diag(SC,v ) =
`X
max
diagsize0C,v,k
k=−1
which explains our choice of the “diag” prefix.
6
A FindFirstLabel Structure
We will show how to maintain information that allows us to implement FindFirstLabel; the function
that allows us to inspect the replacement edge candidates at a given level. The implementation uses
a “destructive binary search, with undo” strategy, similar to the non-local search introduced in [1].
The idea is to maintain enough information in each cluster to determine if there is a result.
Then we can start by using Expose(v, w), and repeatedly split the root containing the answer until
we arrive at the correct label. After that, we simply undo the splits (using the appropriate merges),
and finally undo the Expose.
Just as in the FindSize structure, we will use vertex labels to store all the information pertinent
to a vertex. We store all the added user labels for each vertex in the label object for that vertex in
the base level of the top tree. For each level where the vertex has an associated user label, we keep
a doubly linked list of those labels, and we keep a singly-linked list of these nonempty lists. Thus,
FindFirstLabel(v, w, i) boils down to finding the first vertex label that has an associated user label
at the right level. Once we have that vertex label, the desired user label can be found in O(`max )
time.
Let C be a cluster in T , and let v ∈ ∂C. Define bit vectors4
"
#!
CoverLevel(v, vertex(l)) = i
pointincidentC,v :=
∃label l ∈ C :
∧ `(l) = i
{0≤i<`max }
_
incidentC :=
pointincidentCm ,m
m∈π(C)
Maintaining the incidentC bit vectors, and the corresponding partincidentC,v and diagincidentC,v
bit vectors, can be done completely analogous to the way we maintain the size vectors used for
FindSize, with the minor change that we use bitwise OR on bit vectors instead of vector addition.
Updating the vertex label cluster C in the top tree during AddLabel(v, l, i), or a RemoveLabel(l)
where v = vertex(l) and `(l) = i can be done by first calling detach(C), then updating the linked
4
(
1
Here, [P ] =
0
if P is true
is the Iverson Bracket (see [14]), and ∨ denotes bitwise OR.
otherwise
11
lists containing the user labels and setting
incidentC = ([v has associated labels at level j]){0≤j<`max }
(
incidentC if i = `max
partincidentC,vertex(l),i =
0
otherwise
diagincidentC,vertex(l) = partincidentC
and then reattaching C. Finally FindFirstLabel(v,w,i) can be implemented in the way already
described, by examining pointincidentC,v,i for each cluster. Note that even though we don’t explicitly
maintain it, for any cluster C and any v ∈ ∂C we can easily compute
pointincidentC,v =
`_
max
diagincident0C,v,i
i=−1
=
`_
max
!
diagincidentC,v,i
+
M(cover+
C)
·
`
_
!
partincidentC,v,i
i=−1
i=`+1
+
Where ` := max cover−
C , coverC
S
In general, let A1 , . . . Ak be the clusters resulting from an expose or split, let v, w ∈ ki=1 ∂Ai
(not necessarily distinct). Then we can define
(
Ax
if Ax is a label
FindFirstLabel((A1 , · · · , Ak ); v, w, i) =
FindFirstLabel(Split(Ax ); vx , wx , i) otherwise
where for 1 ≤ j ≤ k
vj = arg min dist(v, u)
u∈∂Aj
wj = arg max dist(u, w)
u∈∂Aj
and
(
I=
1≤j≤k
CoverLevel(v, vj ) ≥ i
∧
)
pointincidentAj ,vj ,i = 1
x = arg min (3 · dist(v, meet(vj , v, w)) + |∂Aj ∩ v · · · w|)
j∈I
FindFirstLabel(v, w, i) = FindFirstLabel(Expose(v, w); v, w, i)
Analysis By the method described in this section, AddLabel, RemoveLabel, and FindFirstLabel
are maintained in O(log n · `max · log `max ) = O((log n)2 log log n) worst-case time.
This can be reduced to O(log n · log `max ) = O(log n log log n) by realizing that each `max dimensional bit vector fits into O(1) words, and that each bitwise OR therefore only takes constant
time.
The total space used for a FindFirstLabel structure with n vertices and m labels is O(m + n)
plus the space for O(n) bit vectors. If we assume a word size of Ω(log n), this is just O(m + n) in
total. If we disallow bit packing tricks, we may have to use O(m + n · `max ) = O(m + n log n) space.
12
7
Approximate counting
As noted in [20], we don’t need to use the exact component sizes at each level. If s is the actual
correct size, it is sufficient to store an approximate value s0 such that s0 ≤ s ≤ e s0 , for some constant
0 < < ln 2. Then we are no longer guaranteed that component sizes drop by a factor of 12 at each
level, but rather get a factor of e2 . This increases the number of levels to `max = bln n/(ln 2 − )c
(which is still O(log n)), but leaves the algorithm otherwise unchanged. Suppose we represent each
size as a floating point value with a b-bit mantissa, for some b to be determined later. For each
addition of such numbers the relative error increases. The relative error at the root of a tree of
−b
additions of height h is (1 + 2−b )h ≤ e2 h , thus to get the required precision it is sufficient to set
b = log2 h . In our algorithm(s) the depth of calculation is clearly upper bounded by h ≤ h(n) · `max ,
where h(n) = O(log n) is the height of the top tree. It follows that some b ∈ O(log log n) is sufficient.
Since the maximum size of a component is n, the exponent has size at most dlog2 ne, and can be
represented in dlog2 dlog2 nee bits. Thus storing the sizes as O(log log n) bit floating point values is
sufficient to get the required precision. Assuming a word size of Ω(log n) this lets us store O logloglogn n
sizes in a single word, and to add them in parallel in constant time.
Analysis We will show how this applies to our FindSize structure from Section 5. The bottlenecks
in the algorithm all have to do with operations on `max -dimensional size vectors. In particular, the
amortized update time is dominated by the time to do O(log n · log `max ) vector additions, and
O(log n) multiplications of a vector by the M(i) matrix. With approximate counting, the vector
additions each take O(log log n) time. Multiplying a size vector x by M(i) we get:
(
xj if i ≥ j
(M(i) · x)j =
0 otherwise
And clearly this operation can also be done on O logloglogn n sizes in parallel when they are packed
into a single word. With approximate counting, each multiplication by M(i) therefore also takes
O(log log n) time. Thus the time per operation is reduced to O(log n(log log n)2 ).
The space consumption of the data structure is O(n) plus the space needed to store O(n) of the
`max -dimensional size vectors. With approximate counting that drops to O(log log n) per vector, or
O(n log log n) in total.
Comparison to previous algorithms Combining the modified FindSize structure with the
CoverLevel structure from Section 4 and the FindFirstLabel structure from Section 6 gives us
the first bridge-finding structure with O((log n)2 (log log n)2 ) amortized update time. This structure uses O(m + n log log n) space, and uses O(log n) time for FindBridge and Size queries, and
O(log n(log log n)2 ) for 2-size queries.
For comparison, applying this trick in the obvious way to the basic O((log n)4 ) time and
O(m + n(log n)2 ) algorithm from [10,11] gives the O((log n)3 log n) time and O(m + n log n log log n)
algorithm briefly mentioned in [20].
8
Top trees revisited
We can combine the tree data structures presented so far to build a data structure for bridge-finding
that has update time O((log n)2 (log log n)2 ), query time O(log n), and uses O(m + n log log n) space.
13
In order to get faster queries and linear space, we need to use top-trees in an even smarter way.
For this, we need the full generality of the top trees described in [1].
8.1
Level-based top trees, labels, and fat-bottomed trees
As described in [1], we may associate a level with each cluster, such that the leaves of the top tree
have level 0, and such that the parent of a level i cluster is on level i + 1. As observed in Alstrup et
al. [1, Theorem 5.1], one may also associate one or more labels with each vertex. For any vertex,
v, we may handle the label(s) of v as point clusters with v as their boundary vertex and no edges.
Furthermore, as described in [1], we need not have single edges on the bottom most level. We may
generalize this to instead have clusters of size Q as the leaves of the top tree.
Theorem 8 (Alstrup, Holm, de Lichtenberg, Thorup [1]). Consider a fully dynamic forest and let
Q be a positive integer parameter. For the trees in the forest, we can maintain levelled top trees
whose
base clusters are
of size at most Q and such that if a tree has size s, it has height h = O(log s)
i
and O(s/(Q(1 + ε) )) clusters on level i ≤ h. Here, ε is a positive constant. Each link, cut, attach,
detach, or expose operation is supported with O(1) creates and destroys, and O(1) joins and splits on
each positive level. If the involved trees have total size s, this involves O(log s) top tree modifications,
all of which are identified in O(Q + log s) time. For a composite sequence of k updates, each of the
above bounds are multiplied by k. As a variant, if we have parameter S bounding the size of each
underlying tree, then we can choose to let all top roots be on the same level H = O(log S).
8.2
High degree top trees
Top trees of degree two are well described and often used. However, it turns out to be useful to also
consider top trees of higher degree B, especially for B ∈ ω(1).
Lemma 9. Given any B ≥ 2, one can maintain top trees of degree B and height O(log n/ log B).
Each expose, link, or cut is handled by O(1) calls to create or destroy and O(log n/ log B) calls to
split or merge. The operations are identified in O(B(log n/ log B)) time.
Proof. Given a binary levelled top tree T2 of height h, we can create a B-ary levelled top tree TB ,
where the leaves of TB are the leaves of T2 , and where the clusters on level i of TB are the clusters
on level i · blog2 Bc of T2 . Edges in TB correspond to paths of length blog2 Bc in T2 . Thus, given a
binary top tree, we may create a B-ary top tree bottom-up in linear time.
We may implement link, cut and expose by running the corresponding operation in T2 . Each
cut, link or expose operation will affect clusters on a constant number of root-paths in T2 . There
are thus only O(log n/ log B) calls to split or merge of a cluster on a level divisible by blog2 Bc.
Thus, since each split or merge in TB corresponds to a split or merge of a cluster in T2 whose level
is divisible by blog2 Bc, we have only O(log n/ log B) calls to split and merge in TB .
However, since there are O(B) clusters whose parent pointers need to be updated after a merge,
the total running time becomes O(B(log n/ log B)).
8.3
Saving space with fat-bottomed top trees
In this section we present a general technique for reducing the space usage of a top tree based data
structure to linear. The properties of the technique are captured in the following:
14
Lemma 10. Given a top tree data structure of height h(n) ∈ O(log n) that uses s(n) space per
cluster, and t(n) worst case time per merge or split.
Suppose that the complete information for a cluster of size q, including information that is
shared with its children, has total size s0 (q, n) and can be computed directly in time t0 (q, n). Suppose
further that there exists a function q of n such that s(n) < s0 (q(n), n) ∈ O(q(n)).
Then there exists a top tree data structure, maintaining the same information, that uses linear
space in total and has O(t(n) · h(n) + t0 (q(n), n)) update time for link, cut, and expose.
Proof. This follows directly from Theorem 8 by setting Q = q(n). Then the top tree will have
O(n/q(n)) clusters of size at most s0 (q(n), n) = O(q(n)) so the total size is linear. The time per
update follows because the top tree uses O(h(n)) merges of split and O(1) create and destroy per
link cut and expose. These take t(n) and t0 (q(n), n) time respectively.
9
A Faster CoverLevel Structure
If we allow ourselves to use bit tricks, we can improve the CoverLevel data structure from Section 4.
The main idea is, for some 0 < < 1, to use top trees of degree b(n) = (log n) ∈ O(w/ log `max ).
log n
Such top trees have height h(n) ∈ O( log
log n ), and finding the sequence of merges and splits for a
1+
n)
1+ ) time.
given link, cut or expose takes O(b(n) · h(n)) ∈ O( (log
log log n ) ⊆ o((log n)
The high-level algorithm makes at most a constant number of calls to link and cut for each insert
or delete, so we are fine with the time for these operations. However, we can no longer use Expose
to implement Cover, Uncover, CoverLevel and MinCoveredEdge, as that would take too long.
In this section, we will show how to overcome this limitation by working directly with the
underlying tree.
The data The basic idea is to maintain a buffer with all the cover, cover− , cover+ and globalcover
values one level up in the tree, in the parent cluster. Since the degree is O(w/ log `max ), and each
value uses at most O(log `max ) bits, these fit into a constant number of words, and so we can use
bit tricks to operate on the values for all children of a node in parallel.
Let C be a cluster with children A1 , . . . , Ak . Since k ≤ w/ log `max , we can define the following
vectors that each fit into a constant number of words.
packedcoverC := (coverAi ){1≤i≤k}
−
packedcover−
C := (coverAi ){1≤i≤k}
+
packedcover+
C := (coverAi ){1≤i≤k}
packedglobalcoverC := (globalcoverAi ){1≤i≤k}
The description of Split and Merge from Section 4 still apply, if we think of the “packed” values
as a separate layer of degree 1 clusters between each pair of “real” clusters.
For concreteness, let C be a cluster with children A1 , . . . , Ak , and define operations
• CleanToBuffer(C).
For each 1 ≤ i ≤ k: If Ai is a path child of C and
−
max packedcoverC,i , packedcover−
C,i ≤ coverC , set:
−
packedcover−
C,i = coverC
15
+
Then if packedcoverC,i ≤ max cover−
C , coverC set
packedcoverC,i = cover+
C
+
packedcover+
C,i = coverC
+
After updating all k children, set cover−
C = coverC = −1. Note that this can be done in
parallel for all 1 ≤ i ≤ k in constant time using bit tricks.
−
• CleanToChild(C, i). If Ai is a path child of C and max coverAi , cover−
Ai ≤ packedcoverC,i ,
set
−
cover−
Ai = packedcoverC,i
+
Then if coverAi ≤ max packedcover−
C,i , packedcoverC,i set
coverAi = packedcover+
C,i
+
cover+
Ai = packedcoverC,i
+
Finally set packedcover−
C,i = packedcoverC,i = −1. Again, note that this takes constant time.
• ComputeFromChild(C, i). Set
packedcoverC,i = coverAi
packedcover−
C,i = −1
packedcover+
C,i = −1
packedglobalcoverC,i = globalcoverAi
• ComputeFromBuffer(C). For 1 ≤ i ≤ k define
packedglobalcoverC,i
0
packedglobalcoverC,i =
packedcoverC,i
minglobaledgeAi
0
minglobaledgeC,i =
minpathedgeAi
16
if ∂Ai ⊆ π(C)
or packedglobalcoverC,i ≤ packedcoverC,i
otherwise
if ∂Ai ⊆ π(C)
or globalcoverAi ≤ coverAi
otherwise
We can then compute the data for C from the buffer as follows:
min packedcoverC,i
1≤i<k
coverC = ∂Ai ⊆π(C)
`
max
minpathedgeAj
where j = arg min packedcover
C,i
minpathedgeC =
1≤i<k
∂Ai ⊆π(C)
nil
if |∂C| = 2
otherwise
if |∂C| = 2
otherwise
globalcoverC = min packedglobalcover0C,i
1≤i<k
minglobaledgeC = minglobaledge0C,j
where j = arg min packedglobalcover0C,i
1≤i<k
cover−
C
cover+
C
= −1
= −1
This can be computed in constant time, because (packedglobalcover0C,i ){1≤i≤k} fits into a
constant number of words that can be computed in constant time using bit tricks, and thus
each “min” or “arg min” is taken over values packed into a constant number of words.
Then Split(C) can be implemented by first calling CleanToBuffer(C), and then for each 1 ≤ i ≤ k
calling CleanToChild(C, i). This ensures that all the lazy cover information is propagated down correctly. Similarly, Merge(C; A1 , . . . , Ak ) can be implemented by first calling ComputeFromChild(C, i)
for each 1 ≤ i ≤ k, and then calling ComputeFromBuffer(C). Thus Split and Merge each take
O(b(n)) time.
Computing CoverLevel(v) and MinCoveredEdge(v) With the data described in the previous
section, we can now answer the “global” queries as follows
CoverLevel(v) = globalcoverC
MinCoveredEdge(v) = minglobaledgeC
where C is the point cluster returned by root(v)
Note that, for simplicity, we assume the top tree always has a single vertex exposed. This can easily
be arranged by a constant number of calls to Expose after each link or cut, without affecting the
asymptotic running time. Computing CoverLevel(v) or MinCoveredEdge(v) therefore takes O(h(n))
worst case time.
Computing CoverLevel(v, w) and MinCoveredEdge(v, w) Since we can no longer use Expose to
implement Cover and Uncover, we need a little more machinery.
What saves us is that all the information we need to find CoverLevel(v, w) is stored in the
O(h(n)) clusters that have v or v as internal vertices, and that once we have that, we can find a
single child X of one of these clusters such that MinCoveredEdge(v, w) = minpathedgeX .
17
Before we get there, we have to deal with the complication of cover− and cover+ . Fortunately,
all we need to do is make O(h(n)) calls to CleanToBuffer and CleanToChild, starting from the root
and going down towards v and w. Since each of these calls take constant time, we use only O(h(n))
time on cleaning.
Now, the path v · · · w consists of O(h(n)) edge-disjoint fragments, such that:
• Each fragment f is associated with, and contained in, a single cluster Cf .
• For each fragment f , the endpoints are either in {v, w} (and then Cf is a base cluster) or are
boundary vertices of children of Cf .
We can find the fragments in O(h(n)) time, and for each fragment f , we can in constant time
find its cover level by examining packedcoverCf .
Let f1 , . . . , fk be the fragments of the path, and for 1 ≤ i ≤ k let vi , wi be the endpoints of the
fragment closest to v, w respectively. Then
CoverLevel(v, w) = min CoverLevel(vi , wi )
1≤i≤k
MinCoveredEdge(v, w) = MinCoveredEdge(vj , wj )
where j = arg min CoverLevel(vi , wi )
1≤i≤k
MinCoveredEdge(vj , wj ) = minpathedgeX
where X =
arg min
coverY
Y path child of Cfj
So computing CoverLevel(v, w) or MinCoveredEdge(v, w) takes O(h(n)) worst case time.
Cover and Uncover We are now ready to handle Cover(v, w, i) and Uncover(v, w, i). First we
make O(h(n)) calls to CleanToBuffer and CleanToChild. Then let f1 , . . . , fk be the fragments of the
v · · · w path, and for 1 ≤ i ≤ k let vi , wi be the endpoints of the fragment closest to v, w respectively.
Then for each f ∈ f1 , . . . , fk , and each path child Aj of Cf , Cover(v, w, i) needs to set
n
o
packedcoverCf ,j = max packedcoverCf ,j , i
n
o
packedcover+
=
max
packedcover
,
i
Cf ,j
Cf ,j
Similarly, for each f ∈ f1 , . . . , fk , and for each path child Aj of Cf , if packedcoverCf ,j ≤ i,
Uncover(v, w, i) needs to set
packedcoverCf ,j = −1
packedcover+
Cf ,j = −1
n
o
−
packedcover−
=
max
packedcover
,
i
Cf ,j
Cf ,j
In each case, we can use bit tricks to make this take constant time per fragment. Finally, we need
to update all the O(h(n)) ancestors to the clusters we just changed. We can do this bottom-up
using O(h(n)) calls to ComputeFromChild and ComputeFromBuffer.
We conclude that Cover(v, w, i) and Uncover(v, w, i) each take worst case O(h(n)) time.
18
Analysis
log n
Choosing any b(n) ∈ O(w/ log `max ) we get height h(n) ∈ O( log
b(n) ), so Link and
log n
Cut take worst case O( b(n)
log b(n) ) time with this CoverLevel structure. The remaining operations,
log n
Connected, Cover, Uncover, CoverLevel and MinCoveredEdge all take O( log
b(n) ) worst case time.
√
For the purpose of our main result, choosing b(n) ∈ Θ( log n) is sufficient. Each cluster uses O(1)
space, so the total space used is O(n).
Note that the pointers that allow us to find the first k least-covered edges can still be maintained
in O(h) time per update, and allows us to find the first k least-covered edges in O(h + k) time.
10
Saving Space
We now apply the space-saving trick from Lemma 10 to the FindSize structures from Section 5
and 7. Let D be the number of words used for each size vector in our FindSize structure. This is
O(log n) for the purely combinatorial version, and O(log log n) in the version using approximate
counting. As shown previously these use s(n) = O(D) space per cluster and t(n) = O(log n · D)
worst case time per merge and split.
Lemma 11. The complete information for a cluster of size q in the FindSize structure, including
information that would be shared with its children, has total size s0 (q, n) = O(q + `max · D).
Proof. The complete information for a cluster C with |C| = q consists of
• c(e) for all e ∈ C.
+
• coverC , cover−
C , coverC , globalcoverC , sizeC .
• partsizeC,v,i and diagsizeC,v,i for v ∈ ∂C and −1 ≤ i ≤ `max .
The total size for all of these is s0 (q, n) = O(q + `max · D)
Note that when keeping n fixed, this is clearly O(q). In particular, we can choose q(n) ∈
Θ(`max · D) such that s(n) < s0 (q(n), n) ∈ O(q(n)).
Lemma 12. The complete information for a cluster of size q in the FindSize structure, including
information that would be shared with its children, can be computed directly in time t0 (q, n) =
O(q log q + `max · D).
Proof. Let C be the cluster of size |C| = q. For each v ∈ ∂C, we can in O(q) time find and partition
the cluster path into the at most `max parts such that in part i, each vertex m on the cluster path
have CoverLevel(v, m) = i. For each part i, run the following algorithm:
1: Vector x ← 0
2: Initialize empty max-queue Q
3: j ← `max
4: for w ← each vertex in the fragment that is on π(C) do
5:
Mark w as visited
6:
xj ← xj + 1
7:
for e ← each edge incident to w that is not on π(C) do
8:
if c(e) ≥ 0 then
9:
Add e to Q with key c(e)
19
while Q is not empty do
e ← extract-max(Q)
12:
while c(e) < j do
13:
xj−1 = xj
14:
j ←j−1
15:
w ← the unvisited vertex at the end of e
16:
Mark w as visited
17:
xj ← xj + 1
18:
for e ← each edge incident to w that has an unvisited end do
19:
if c(e) ≥ 0 then
20:
Add e to Q with key c(e)
21: partsizeC,v,i ← x
22: diagsizeC,v,i ← M(i) · x
If the ith part has size qi than it can be processed this way in O(qi log qi + D) time. Summing over
all O(`max ) parts gives the desired result.
10:
11:
Analysis Applying Lemma 10 with the s(n), t(n), s0 (q, n), t0 (q, n) and q(n) derived in this section
immediately gives a FindSize structure with O(log n · D · log `max ) worst case time per operation and
using O(n) space. A completely analogous argument shows that we can convert the bitpacking-free
version of the FindFirstLabel structure from O(log n · `max · log `max ) time and O(m + n · `max ) space
to one using linear space. (If bitpacking is allowed the structure already used linear space). In either
case is the same time per operation as the original versions, so using the modified version here does
not affect the overall running time, but reduces the total space of each bridge-finding structure to
O(m + n).
Note that we can explicitly store lists with all the least-covered edges for these large base clusters,
so this does not change the time to report the first k least-covered edges.
References
[1] Stephen Alstrup, Jacob Holm, Kristian De Lichtenberg, and Mikkel Thorup. Maintaining
information in fully dynamic trees with top trees. ACM Trans. Algorithms, 1(2):243–264,
October 2005.
[2] Therese C. Biedl, Prosenjit Bose, Erik D. Demaine, and Anna Lubiw. Efficient algorithms for
petersen’s matching theorem. Journal of Algorithms, 38(1):110 – 134, 2001.
[3] Krzysztof Diks and Piotr Stanczyk. Perfect Matching for Biconnected Cubic Graphs in
O(n log2 n) Time, pages 321–333. Springer Berlin Heidelberg, Berlin, Heidelberg, 2010.
[4] David Eppstein, Zvi Galil, and Giuseppe F. Italiano. Improved sparsification. Technical report,
1993.
[5] Greg N. Frederickson. Data structures for on-line updating of minimum spanning trees, with
applications. SIAM Journal on Computing, 14(4):781–798, 1985.
20
[6] Greg N. Frederickson. Ambivalent data structures for dynamic 2-edge-connectivity and k
smallest spanning trees. SIAM J. Comput., 26(2):484–538, 1997.
[7] Harold N. Gabow, Haim Kaplan, and Robert Endre Tarjan. Unique maximum matching
algorithms. J. Algorithms, 40(2):159–183, 2001. Announced at STOC ’99.
[8] Monika R. Henzinger and Valerie King. Maintaining minimum spanning trees in dynamic
graphs, pages 594–604. Springer Berlin Heidelberg, Berlin, Heidelberg, 1997.
[9] Monika Rauch Henzinger and Valerie King. Fully dynamic 2-edge connectivity algorithm in
polylogarithmic time per operation, 1997.
[10] Jacob Holm, Kristian de Lichtenberg, and Mikkel Thorup. Poly-logarithmic deterministic
fully-dynamic algorithms for connectivity, minimum spanning tree, 2-edge, and biconnectivity.
In Proceedings of the Thirtieth Annual ACM Symposium on Theory of Computing, STOC ’98,
pages 79–89, New York, NY, USA, 1998. ACM.
[11] Jacob Holm, Kristian de Lichtenberg, and Mikkel Thorup. Poly-logarithmic deterministic
fully-dynamic algorithms for connectivity, minimum spanning tree, 2-edge, and biconnectivity.
J. ACM, 48(4):723–760, July 2001.
[12] Shang-En Huang, Dawei Huang, Tsvi Kopelowitz, and Seth Pettie. Fully dynamic connectivity
in o(log n(log log n)2) amortized expected time. In Proceedings of the Twenty-Eighth Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA ’17, pages 510–520, Philadelphia, PA,
USA, 2017. Society for Industrial and Applied Mathematics.
[13] Bruce M. Kapron, Valerie King, and Ben Mountjoy. Dynamic graph connectivity in polylogarithmic worst case time. In Proceedings of the Twenty-fourth Annual ACM-SIAM Symposium
on Discrete Algorithms, SODA ’13, pages 1131–1142, Philadelphia, PA, USA, 2013. Society for
Industrial and Applied Mathematics.
[14] Donald E. Knuth. Two notes on notation. The American Mathematical Monthly, 99(5):403–422,
1992.
[15] Donald E. Knuth. The Art of Computer Programming, Volume 3: (2nd Ed.) Sorting and
Searching. Addison Wesley Longman Publishing Co., Inc., Redwood City, CA, USA, 1998.
[16] Anton Kotzig. On the theory of finite graphs with a linear factor II. 1959.
[17] Karl Menger. Zur allgemeinen Kurventheorie. Fundamenta Mathematicae, 10, 1927.
[18] Mihai Patrascu and Erik D Demaine. Logarithmic lower bounds in the cell-probe model. SIAM
Journal on Computing, 35(4):932–963, 2006.
[19] Julius Petersen. Die Theorie der regulären graphs. Acta Math., 15:193–220, 1891.
[20] Mikkel Thorup. Near-optimal fully-dynamic graph connectivity. In Proceedings of the Thirtysecond Annual ACM Symposium on Theory of Computing, STOC ’00, pages 343–350, New
York, NY, USA, 2000. ACM.
[21] Christian Wulff-Nilsen. Faster deterministic fully-dynamic graph connectivity. In Encyclopedia
of Algorithms, pages 738–741. 2016.
21
A
Details of the high level algorithm
Lemma 5 (Essentially the high level algorithm from [11]). There exists a deterministic reduction
for dynamic graphs with n nodes, that, when starting with an empty graph, supports any sequence
of m Insert or Delete operations using:
• O(m) calls to Link, Cut, Uncover, and CoverLevel.
• O(m log n) calls to Connected, Cover, AddLabel, RemoveLabel, FindFirstLabel, and FindSize.
And that can answer FindBridge queries using a constant number of calls to Connected, CoverLevel,
and MinCoveredEdge.
Proof. The only part of the high level algorithm from [11] that does not directly and trivially
translate into a call of the required dynamic tree operations (see pseudocode below) is in the Swap
method where given a tree edge e = (v, w) we need to find a nontree edge e0 covering e with
`(e0 ) = i = CoverLevel(e). We can find this e0 by using FindFirstLabel and increasing the level of
each non-tree edge we examine that does not cover e. For at least one side of (v, w), all non-tree
edges at level i incident to that side will either cover e or can safely have their level increased
without violating the size invariant. So we can simply search the side where the level i component
is smallest until we find the required edge (which must exist since e was covered on level i). The
amortized cost of all operations remain unchanged with this implementation. Counting the number
of operations (see Table 2) gives the desired bound.
#
Operation
1
2
3
4
5
6
7
8
9
10
11
12
Link(v, w, e)
Cut(e)
Connected(v, w)
Cover(v, w, i)
Uncover(v, w, i)
CoverLevel(v)
CoverLevel(v, w)
MinCoveredEdge(v)
MinCoveredEdge(v, w)
AddLabel(v, l, i)
RemoveLabel(l)
FindFirstLabel(v, w, i)
FindSize(v, w, i)
FindSize(v, v, −1)
13
Insert+Delete
1
1
log n
log n
1
0
1
0
0
log n
log n
log n
log n
0
#Calls during
FindBridge(v) FindBridge(v, w)
0
0
0
0
0
1
0
0
0
0
1
0
0
1
1
0
0
1
0
0
0
0
0
0
0
0
0
0
Size(v)
0
0
0
0
0
0
0
0
0
0
0
0
0
1
2-Size(v)
0
0
0
0
0
0
0
0
0
0
0
0
1
0
Table 2: Overview of how many times each tree operation is called for each graph operation, ignoring
constant factors. The “Insert+Delete” column is amortized over any sequence starting with an
empty set of edges. The remaining columns are worst case.
function 2-edge-connected(v, w)
2:
return T.Connected(v, w) ∧ T.CoverLevel(v, w)≥ 0
3: function FindBridge(v)
4:
if T.CoverLevel(v)= −1 then
1:
22
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
45:
46:
47:
48:
return T.MinCoveredEdge(v)
else
return nil
function FindBridge(v, w)
if T.CoverLevel(v, w)= −1 then
return T.MinCoveredEdge(v, w)
else
return nil
function Size(v)
return T.FindSize(v,v,−1)
function 2-Size(v)
return T.FindSize(v,v,0)
function Insert(v, w, e)
if ¬T.Connected(v, w) then
T.Link(v, w, e)
`(e) ← `max
else
T.AddLabel(v, e.label1, 0)
T.AddLabel(w, e.label2, 0)
`(e) ← 0
T.Cover(v, w, 0)
function Delete(e)
(v, w) ← e
α ← `(e)
if α = `max then
α ← T.CoverLevel(v, w)
if α = −1 then
T.Cut(e)
return
Swap(e)
T.RemoveLabel(e.label1)
T.RemoveLabel(e.label2)
T.Uncover(v, w, α)
for i ← α, . . . , 0 do
Recover(w,v,i)
function Swap(e)
(v, w) ← e
α ← T.CoverLevel(v, w)
T.Cut(e)
e0 ←FindReplacement(v,w,α)
(x, y) ← e0
T.RemoveLabel(e0 .label1)
T.RemoveLabel(e0 .label2)
T.Link(x, y, e0 )
23
49:
50:
51:
52:
53:
54:
55:
56:
57:
58:
59:
60:
61:
62:
63:
64:
65:
66:
67:
68:
69:
70:
71:
72:
73:
74:
75:
76:
77:
78:
79:
80:
81:
82:
83:
B
`(e0 ) ← `max
T.AddLabel(v,e.label1, α)
T.AddLabel(w,e.label2, α)
`(e) ← α
T.Cover(v,w,α)
function FindReplacement(v,w,i)
sv ← T.FindSize(v, v, i)
sw ← T.FindSize(w, w, i)
if sv ≤ sw then
return RecoverPhase(v, v, i, sv )
else
return RecoverPhase(w, w, i, sw )
function Recover(v,w,i)
s ← bT.FindSize(v, w, i)/2c
RecoverPhase(v,w,i,s)
RecoverPhase(w,v,i,s)
function RecoverPhase(v, w, i, s)
l ← T.FindFirstLabel(v, w, i)
while l 6= nil do
e ← l.edge
(q, r) ← e
if ¬T.Connected(q, r) then
return e
if T.FindSize(q, r, i + 1) ≤ s then
T.RemoveLabel(e.label1)
T.RemoveLabel(e.label2)
T.AddLabel(q, e.label1, i + 1)
T.AddLabel(r, e.label2, i + 1)
`(e) = i + 1
T.Cover(q,r,i + 1)
else
T.Cover(q,r,i)
return nil
l ← T.FindFirstLabel(v, w, i)
return nil
Pseudocode for the CoverLevel structure
function CL.Cover(v,w,i)
C ← TopTree.Expose(v, w)
3:
coverC ← max {cover
C , i}
+
4:
cover+
←
max
cover
C
C, i
1:
2:
function CL.Uncover(v,w,i)
6:
C ← TopTree.Expose(v, w)
5:
24
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
45:
46:
47:
48:
49:
coverC ← −1
cover+
C ← −1
−
cover−
C ← max coverC , i
function CL.CoverLevel(v)
C ← TopTree.Expose(v)
return globalcoverC
function CL.CoverLevel(v, w)
C ← TopTree.Expose(v, w)
return coverC
function CL.MinCoveredEdge(v)
C ← TopTree.Expose(v)
return minglobaledgeC
function CL.MinCoveredEdge(v, w)
C ← TopTree.Expose(v, w)
return minpathedgeC
function CL.Split(C)
for each path
child D of C do
−
if max coverD , cover−
D ≤ coverC then
−
−
coverD ← coverC
+
if coverD ≤ max cover−
D , coverD then
+
coverD ← coverC
+
cover+
D ← coverC
function CL.Merge(C; A1 , . . . , Ak )
coverC ← `max
minpathedgeC ← nil
globalcoverC ← `max
minglobaledgeC ← nil
for i ← 1, . . . , k do
if ∂Ai ⊆ π(C) then
if coverAi < coverC then
coverC ← coverAi
minpathedgeC ← minpathedgeAi
else
if coverAi < globalcoverC then
globalcoverC ← coverAi
minglobaledgeC ← minpathedgeAi
if globalcoverAi < globalcoverC then
globalcoverC ← globalcoverAi
minglobaledgeC ← minglobaledgeAi
cover−
C ← −1
cover+
C ← −1
function CL.Create(C; edge e)
coverC ← −1
25
50:
51:
52:
53:
54:
55:
56:
57:
58:
C
globalcoverC ← −1
if C is a point cluster then
minpathedgeC ← nil
minglobaledgeC ← e
else
minpathedgeC ← e
minglobaledgeC ← nil
cover−
C ← −1
cover+
C ← −1
Pseudocode for the FindSize structure
In the following, we use the notation
[key : partsize, diagsize]
to denote the root of a new tree consisting of a single node with the given values. And for a given
tree root and given x, y
(tree{x≤i≤y} )
is the root of the subtree consisting of all nodes whose keys are in the given range. Similarly, for
any given i, let
(treei )
denote the node in the tree having the given key.
1: function FS.FindSize(v, w, i)
2:
C ← TopTree.Expose(v, w)
3:
return sizeC,i
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
function FS.Merge(C; A, B)
{c} ← ∂A ∩ ∂B
if c ∈ π(C) then
if |∂C| <= 1 then
a ← c, b ← c
else
{a, b} ← ∂C with a ∈ ∂A and b ∈ ∂B.
sizeC ← sizeA + sizeB
for (x, X) ← (a, A), (b, B) do
if x = c then
tree0X,x ← treeX,x , undo0X,x ← nil
else
for v ← x, c do
+
` ← max cover−
X , coverX
s ← (treeX,v ). partsizesum
d ← M(cover+
X) ∗ s
26
. Merge along path
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
45:
46:
47:
48:
49:
50:
51:
52:
53:
54:
55:
56:
57:
58:
59:
60:
61:
tree0X,v ← treeX,v,{i>`} , undo0X,v ← treeX,v,{i≤`}
tree0X,v ← tree0X,v +[cover+
X : s, d]
for (x, X, y, Y ) ← (a, A, b, B), (b, B, a, A) do
s ← (tree0Y,c,{coverX ≤i≤`max } ). partsizesum
p ← (tree0X,x,coverX ). partsize +s
d ← (tree0X,x,coverX ). diagsize +M(coverX ) ∗ s
if x = c then
tree00X,x ← [`max : sizeX , sizeX ], undo00X,x ← nil
else
tree00X,x ← tree0X,x,{i>coverX } , undo00X,x ← tree0X,x,{i≤coverX }
if y = c then
000
tree000
Y,c ← nil, undoY,c ← [`max : sizeY , sizeY ]
else
000
0
0
tree000
Y,c ← treeY,c,{i<coverX } , undoY,c ← treeY,c,{i≥coverX }
treeC,x ← tree00X,x +[coverX : p, d] + tree000
Y,c
else
{a} ← ∂C \ {c}
if a 6∈ ∂A then
Swap A and B
+
` ← max cover−
A , coverA
d ← (treeA,a,{`<i≤`max } ). diagsizesum
p ← (treeA,a,{−1≤i≤`} ). partsizesum
sizeC ← d + M(cover+
A ) ∗ p + M(coverA ) ∗ sizeB
treeC,a ← [`max : sizeC , sizeC ]
function FS.Split(C)
A, B ← the children of C
{c} ← ∂A ∩ ∂B
if c ∈ π(C) then
if |∂C| <= 1 then
a ← c, b ← c
else
{a, b} ← ∂C with a ∈ ∂A and b ∈ ∂B.
for (x, X, y, Y ) ← (a, A, b, B), (b, B, a, A) do
tree00X,x ← treeC,x,{i>coverX } , tree000
Y,c ← treeC,x,{i<coverX }
if y 6= c then
000
tree0Y,c ← tree000
Y,c + undoY,c
if x 6= c then
tree0X,x ← tree00X,x + undo00X,x
for (x, X) ← (a, A), (b, B) do
if x 6= c then
for v ← x, c do
treeX,v ← tree0X,v, i>cover+ + undo0X,v
{
X}
27
. Merge off path
. Split along path
function FS.Create(C; edge e)
sizeC ← 0
64:
for v ∈ ∂C do
65:
treeC,v ← [`max : 0, 0]
62:
63:
function FS.Create(C; vertex label l)
sizeC ← (1){0≤i<`max }
68:
for v ∈ ∂C do
69:
treeC,v = [`max : sizeC , sizeC ]
66:
67:
28
| 8cs.DS
|
arXiv:1002.1422v1 [cs.PL] 7 Feb 2010
Integrating Interval Constraints
into Logic Programming
M.H. van Emden
Department of Computer Science
University of Victoria
Research report DCS-133-IR
Abstract
The clp scheme uses Horn clauses and sld resolution to generate multiple constraint satisfaction problems (csps). The possible csps include
rational trees (giving Prolog) and numerical algorithms for solving linear
equations and linear programs (giving clp(r)). In this paper we develop
a form of csp for interval constraints. In this way one obtains a logic semantics for the efficient floating-point hardware that is available on most
computers.
The need for the method arises because in the practice of scheduling
and engineering design it is not enough to solve a single csp. Ideally one
should be able to consider thousands of csps and efficiently solve them or
show them to be unsolvable. This is what clp/ncsp, the new subscheme
of clp described in this paper is designed to do.
1
Introduction
Floating-point arithmetic is marvelously cheap, and it works most of the
time. Many textbooks on numerical analysis contain examples of how
spectacularly, or insidiously, it can go wrong when it does not work. It
would seem that in a mature computing technology there is only place
for reliable techniques. Yet floating-point arithmetic is not to be lightly
dismissed: it is one of the main beneficiaries of the enormous increase in
processor performance of the last few decades. In combination with the
insatiable demand for more computationally intensive mathematical modeling, this gives every motivation to use interval arithmetic [21] as a way
to safely use the dangerous technology that is floating-point arithmetic.
Interval arithmetic ensures that, in spite of the errors inherent in
floating-point arithmetic, a computation can be interpreted as a proof
that the real-valued result is contained in the (interval) result of the computation. However, the correctness provided by interval arithmetic is
limited to the evaluation of a single expression; it does not extend to the
algorithm in which such evaluations take place. To ensure correctness of
the way an algorithm combines expression evaluations one could of course
1
use verification methods for imperative programs, such as Floyd’s assertions. In this paper we consider the alternative of replacing the algorithms
by logic programs, thus allowing programs to be read as specifications.
Logic programming is more than just an alternative to Floyd’s assertions. The logic framework suggests a relational form for interval computation. Such a relational form is provided by interval constraints [7, 6],
an improvement to interval arithmetic itself. Incorporating interval constraints into logic programming has the added advantage that the result
goes beyond the constraint processing paradigm by yielding programs that
generate multiple constraint satisfaction problems in addition to solving
them. In scheduling and in engineering design it is typically the case that
one has an entire search space of such problems. clp/ncsp, the integrated
system described in this paper, generates such search spaces. Solving is
not only used for obtaining results, but also for pruning the search spaces
by inducing early failure.
In Section 4 we start at the logic end with a review of the clp scheme
[17, 18]. We use Clark’s method [9] for the semantics of logic programming
schemes. As this method uses a mild form of algebraic logic, it needs
some introduction; this happens in Section 3. In Section 5 we start at
the opposite end with a suitably modified version of the main features
of the Constraint Processing framework (csp). To bring together the two
established constraint approaches of the literature we develop in Section 6
what we call here the dc subscheme of the clp scheme. The integration
of interval constraints (reviewed in Section 7) into logic programming is
described in Section 8.
2
Related work
The pioneering work in constraint logic programming is [15], implemented
as chip [13]. Prolog has sld resolution as sole inference rule; [15] added
Forward Checking, Look-Ahead, and Partial Look-Ahead as additional
inference rules, to be applied to goals, depending how they are declared.
chip was restricted to finite domains. ichip [20] proposed extending
chip to include floating-point intervals as domains for real-valued variables. Descendants of chip such as the Eclipse system (see [4] for a
recent description), implemented floating-point intervals.
The earliest design for integrating interval arithmetic into Prolog is
Cleary’s [10], which served as basis for bnr-Prolog [8, 7]. Cleary’s proposal
of a “logical arithmetic” for Prolog described an implementation, but not
a logical semantics. His paper and [12] are the first to describe relational,
rather than functional, interval arithmetic. It remains to be seen whether
the mathematical model given by Older and Vellino [22] can be connected
to logic. bnr Prolog and Prolog IV [11] are mentioned here because of
their connection with Prolog, but not because of connection with logic
programming.
The clp scheme [17] gives a logical semantics that combines pure Prolog with constraint solving. This scheme supersedes chip and its descendants as it is both simpler and more general.
The clp scheme served as the basis for the clp(r) system [19]. It
2
uses the scheme to generate answers to numerical problems in the form of
“active constraints”. In the derivation of these, floating-point arithmetic
is used without due precaution, so that the validity of answers is lost
through rounding errors.
In [23] it was shown that the clp scheme is general enough to accommodate both interval and finite-domain constraints. It does this by
introducing “value constraints” without suggesting any way of interfacing
these with intervals. This is done in this paper by means by means of
clp/dc, the dc subscheme of the clp scheme. In this way we obtain
clp/ncsp, the first logic programming language (as distinct from extension of Prolog) with real variables in which only the precision, but not the
validity, of answers is affected by rounding errors.
3
3.1
Logic Preliminaries
Relations
Relations play a central role in the integration of interval constraints into
logic programming: both constraints and the meanings of logic predicates
are relations. Here we do not attempt to define relations as generally as
possible: we only strive for adequacy for the purpose of this paper. For a
more drastic generalization of the usual notion of relation, see [24].
As usually defined, a relation is a subset of a Cartesian product S1 ×
· · · × Sk . That is, it consists of tuples ha1 , . . . , ak i with ai ∈ Si for i =
1, . . . , k. Such tuples are indexed by the integers 1, . . . , k. In the following
we will need such relations, as for example the ternary relation sum =
{hx, y, zi ∈ R3 | x + y = z}, indexed by the set {1, 2, 3}.
But we will also need relations consisting of tuples indexed by variables
instead of integers. For example, the constraint written as sum(x2 , x2 , x1 )
is intended to be a relation distinct for the sum relation just mentioned.
As another example, sum(x2 , x2 , x1 ) ∧ sum(x3 , x4 , x1 ) is intended to be a
relation. If so, which set of tuples? How indexed?
In this section we introduce the suitable type of relation; in Section 3.4
we define how they arise as meaning of the constraint expressions just
shown.
Definition 1 Given a set X = {x1 , . . . , xN } of variables, a relation ρ ⊂
S1 × · · · × Sn consisting of tuples indexed by {1, . . . , n} and a sequence
hv1 , . . . , vn i of variables (not necessarily distinct), the relation ρ on v is
the set of tuples τ indexed by the set of the k ≤ n variables in v such that
τ (vi ) = ti for a tuple t ∈ ρ, for all i = 1, . . . , n.
Example 1 Let ρ be the ternary sum relation over the set N of natural
numbers, X = {x1 , . . . , x100 }, and v = hx2 , x2 , x1 i. Then we have as
example of a tuple τ in the relation ρ on v:
τ (v1 ) =
τ (x2 )
= t1
τ (v2 ) =
τ (x2 )
= t2
τ (v3 ) =
τ (x1 )
= t3
3
τ
x1
0
2
4
6
8
10
...
x2
0
1
2
3
4
5
...
ρ1 ✶ ρ2
x1
0
2
2
2
4
4
...
x2
0
1
1
1
2
2
...
x3
0
0
1
2
0
1
...
x4
0
2
1
0
4
3
...
Table 1: On the left, tabular form of the relation sum on hx2 , x2 , x1 , i where
sum = {hx, y, zi ∈ R3 | x + y = z}. On the right, tabular form of ρ1 ✶ ρ2 from
Example 2.
For such a τ to exist, only tuples t ∈ ρ qualify where the first two elements
are equal to each other. τ consists of tuples indexed by the set {x1 , x2 }.
The tabular form of τ is as shown in Table 1.
Definition 2 Let ρ1 (ρ2 ) be a relation in which the tuples are indexed by
a set X1 (X2 ) of variables. The join of ρ1 and ρ2 is written as ρ1 ✶ ρ2 and
is defined as the relation in which the tuples τ are indexed by X1 ∪ X2 and
are such that there exists, for each tuple τ ∈ (ρ1 ✶ ρ2 ), tuples τ1 ∈ ρ1 and
τ2 ∈ ρ2 exist such that τ (x) = τ1 (x) if x ∈ X1 and τ (x) = τ2 (x) if x ∈ X2 .
(Note that this implies that τ1 and τ2 have to be such that τ1 (x) = τ2 (x)
for all x such that x ∈ X 1 ∩ X2 .)
Example 2 Let ρ be the ternary sum relation over the set N of natural
numbers, X = {x1 , . . . , x100 }, v1 = hx2 , x2 , x1 i, and v2 = hx3 , x4 , x1 i. Let
ρ1 be ρ on v1 and ρ2 be ρ on v2 . Then ρ1 ✶ ρ2 is a relation of which
some tuples are shown in Table 1.
3.2
Language
The vocabulary of logic is formalized as a signature Σ = hP, F, V i, a tuple
of disjoint, countably infinite, sets of predicates, functors, and variables.
P is partitioned according to whether it may occur in a constraint or in
a program. Thus we have “constraint predicates” and “program predicates”. The constraint predicates include the nullary true and false and
the binary =.
A term is a variable or an expression of the form f (t0 , . . . , tk−1 ), where
f ∈ F and t0 , . . . , tk−1 are terms. If k = 0, then the term is a constant.
An atom (or atomic formula) is an expression of the form p(t0 , . . . , tk−1 ),
where p ∈ P is a predicate and t0 , . . . , tk−1 are terms. If p is a program
(constraint) predicate, then an atom with p as predicate is a program
(constraint) atom.
A goal statement is a conjunction of program atoms or constraint
atoms. A constraint is a conjunction of constraint atoms.
4
3.3
Interpretations
Interpretations depend on a language’s signature. They are formalized as
Σ-structures I = hD, P, Fi where
• D is a non-empty set called the domain of the interpretation.
• P is a function mapping every k-ary predicate in P to a subset of
Dk . P maps true to true = {hi}, false to false = {}, and = to
{ha, ai | a ∈ D}.
• F is a function mapping every functor f in F to a function mapping
each k-ary functor in F to a k-adic function in Dk → D.
3.4
Denotations
An interpretation hD, F, Pi determines a function M mapping variablefree terms to their denotations, as follows:
• M(t) = F(t) ∈ D if t is a constant
• M(f (t0 , . . . , tk−1 )) = (F(f ))(M(t0 ), . . . , M(tk−1 )) for k > 0
• A ground atom p(t0 , . . . , tk−1 ) is true in an interpretation iff
hM(t0 ), . . . , M(tk−1 )i ∈ P(p) We give denotations of non-atomic
formulas later, via relations.
We now consider denotations of terms and atoms that contain variables. Let A be an assignment, which is a function in V → D, assigning
an individual in D to every variable. (In other words, A is a tuple of elements of D indexed by V ). As denotations of formulas with free variables
depend on A , we write MA .
• MA (t) = A(t) if t ∈ V
• MA (f (t0 , . . . , tk−1 )) =
(F(f ))(MA (t0 ), . . . , MA (tk−1 )) for k > 0.
• An atom p(t0 , . . . , tk−1 ) is true in an interpretation iff
hMA (t0 ), . . . , MA (tk−1 )i ∈ P(p)
We give denotations of non-atomic formulas later, via relations.
The existential closure ∃x0 , . . . , xn−1 of a set C of atoms is true in an
interpretation iff there is an assignment A such that MA (A) = true for
every atom A ∈ C.
Definition 3 Let XC ⊂ V be the set of the free variables in formula C.
R(C), the relation denoted by C, given the interpretation determining
MA , is defined as
R(C) = {A ↓ XC | A is an assignment and MA (C) = true}.
By A ↓ XC we mean the function A : V → D restricted to arguments in
XC ⊂ V .
Thus R(C) consists of tuples indexed by variables. R allows us to
translate between algebraic expressions in terms of relations and formulas of logic. This is useful because the results in constraint satisfaction
5
problems are expressed in terms of relations, whereas the constraint logic
programming scheme is expressed in terms of first-order predicate logic.
We have of course R(true) = M(true) = true; also R(false) = M(false) =
false. More interestingly, we may have R(C1 ∧ C2 ) = R(C1 ) ∩ R(C2 ) and
R(C1 ∨ C2 ) = R(C1 ) ∪ R(C2 ). But these hold only when C1 and C2 have
the same set of variables. As this is not always the case, we also need to
define R(Z, C), where Z is a set {z1 , . . . zN } of variables containing XC ,
the set of the free variables of C:
Definition 4
R(Z, C) = {A ↓ Z | A is an assignment and MA (C) = true}.
Definitions (3) and (4) were suggested by a similar device first brought to
our attention by [9]. The version here is modified to allow translations of
a wider class of formulas. Their advantage is that of simplicity compared
to other systems of algebraic logic such as [14].
R(XC , C)
=
R(C)
R(XC1 ∪ XC2 , C1 ∧ C2 )
=
R(XC1 ∪ XC2 , C1 ) ∩ R(XC1 ∪ XC2 , C2 )
R(XC1 ∪ XC2 , C1 ∨ C2 )
=
R(XC1 ∪ XC2 , C1 ) ∪ R(XC1 ∪ XC2 , C2 ).
To be able to interface the CLP scheme, expressed in terms of predicate
logic formulas with CSPs, expressed in terms of relations, we will use the
following lemma.
Lemma 1 R(XC1 ∪ XC2 , C1 ∧ C2 ) = R(C1 ) ✶ R(C2 ).
3.5
Logical implication
In the usual formulation of first-order predicate logic we find the notation
T |= S for the sentence S being logically implied by sentence T , where
“sentence” means closed formula. The meaning of the implication is that
S is true in all models of T . The denotations just defined allow logical
implication to be generalized to apply to formulas that have free variables
[9]:
Definition 5 Let S and T be formulas and let Z be the set of variables
occurring in them. Then we write T |= S to mean that in all interpretations R(Z, S) ⊂ R(Z, T ). Likewise, T |= R(Z, S) ⊂ R(Z, T ) means that
R(Z, S) ⊂ R(Z, T ) holds in all models of T .
4
Review of the CLP scheme
The clp scheme is based on the observation that in logic programming the
Herbrand base can be replaced by any of many other semantic domains.
Hence the scheme has as parameter a tuple hΣ, I, L, T i, where Σ is a
signature, I is a Σ-structure, L is a class of Σ-formulas, and T is a firstorder Σ-theory. These components play the following roles. Σ determines
the relations and functions that can occur in constraints. I is the structure
6
over which computations are performed 1 . L is the class of constraints
that can be expressed. Finally, T axiomatizes properties of I.
Derivations in the clp scheme are defined by means of transitions
between states. A state is defined as a tuple hG, A, P i where the goal
statement G is a set of atoms and constraints and A and P are sets of
constraints2 . Together A and P form the constraint store. The constraints
in A are called the active constraints; those in P the passive constraints.
The query Q corresponds to the initial state hQ, ∅, ∅i. A successful
derivation is one that ends in a state of the form h∅, A, P i.
The role of A and P in this formula is to describe the answer to the
query Q. A ∧ P is clp’s generalization of Prolog’s answer substitution.
It describes an answer, if consistent. Such an answer may not be useful,
as P may still represent a difficult computational problem. All that the
derivation has done is to reduce the program atoms to constraint atoms,
directly or indirectly via program atoms. Derivations also transfer as
much as possible the computational burden of the passive constraints P
to the easily solvable active constraints A.
4.1
Operational semantics
A derivation is a sequence of states such that each next state is obtained
from the previous one by a transition. There are four transitions, →r ,
→c , →i , and →s :
The resolution transition →r : hG ∪ {a}, A, P i →r hG ∪
B, A, P ∪ {s1 = t1 , . . . , sn = tn }i if a is the atom selected out of G ∪ {a}
by the computation rule, h ← B is a rule of P, renamed to new variables,
and h = p(t1 , . . . , tn ) and a = p(s1 , . . . , sn ).
hG ∪ {a}, A, P i →r fail is the transition that applies if a is the atom
selected by the computation rule, and, for every rule h ← B in P, h and
a have different predicate symbols.
1.
2. The constraint transfer transition →c : hG ∪ {c}, A, P i →c
hG, A, P ∪ {c}i if constraint c is selected by the computation rule.
3. The constraint store management transition →i : hG, A, P i →i
hG, A′ , P ′ i
if hA′ , P ′ i = infer(A, P ).
4. The consistency test transition →s : hG, A, P i →s hG, A, P i
if A is consistent;
hG, A, P i →s fail otherwise.
1 I is a structure consisting of a set D of values (the carrier of the structure) together
with relations and functions over D as specified by the signature Σ. For example, there is a
complete ordered field that has R, the set of real numbers, as carrier.
2 We will often regard A and P as formulas. Then they are the conjunctions of the atoms
they contain.
7
4.2
Logic semantics
For the logic semantics of the clp scheme we follow [9].
Theorem 1 (soundness) Whenever we have a successful derivation from
query Q resulting in P and A as passive and active constraints we have
P, T |= R(∃(P ∧ A)) ⊂ R(Q), where the quantification is over the free
variables in P ∧ A that do not occur free in Q. Note Definition 5 for “|=”.
Theorem 2 (completeness) Let Q be a query with variables XQ . If
P, T |= R(XQ , Γ) ⊂ R(Q) for a constraint atom Γ, then there are k
successful derivations from Q with answer constraints Γ1 , . . . , Γk such that
T |= R(XQ , Γ) ⊂ R(XQ , Γ1 ) ∪ · · · ∪ R(XQ , Γk ).
For credits see [9].
5
Constraint Satisfaction Problems
Constraint Satisfaction Problems (csps) can be defined as a framework
to cover a variety of specific situations, each exploiting an algorithmic opportunity. For example, the csp framework can be instantiated to graphcolouring problems exploiting an efficient algorithm for the all-different
constraint based on matching in bipartite graphs. It can also be instantiated to the solution of arithmetical constraints over real-valued variables
using efficient algorithms and hardware for floating-point intervals. It is
for this latter instantiation that we are interested in csps. But before
describing it, first the general framework.
5.1
CSPs according to Apt
K. Apt was early in recognizing [2] that csps can be defined rigorously,
yet in such a way as to be widely applicable. The following definition is
distilled from [2, 5], and uses his notation.
Definition 6 A csp hX , D, Ci consists of a sequence X = hx1 , . . . , xn i
of variables, a sequence D = hD1 , . . . , Dn i of sets called domains, and a
set C = {c1 , . . . , ck } of constraints. Each constraint is a constraint on a
subsequence of X . An n-tuple hd1 , . . . , dn i ∈ D1 × · · · × Dn is a solution
to hX , D, Ci iff for every c ∈ C on a sequence of variables hxi1 , . . . , xim i
from X we have hdi1 , . . . , dim i ∈ c.
In this definition X is probably intended to consist of n different variables.
Once that condition is assumed, X need not be a sequence, but can be a
set without further qualification.
Example 3 To see Apt’s definition at work, consider the following example. X = hx1 , x2 , x3 , x4 i, D = hN , N , N , N i, and C = {c1 , c2 }. Constraints c1 and c2 are on hx2 , x2 , x1 i and hx3 , x4 , x1 i, respectively. To
determine some of the solutions we construct Table 2.
The table for c1 is constructed according to the rule x2 +x2 = x1 ; for c2
the rule is x3 + x4 = x1 . In Definition 6 a constraint remains a black box:
there is no opportunity to specify a rule according to which the tuples
8
c1
c2
solutions
x1
0
2
4
6
...
0
1
1
2
2
2
0
2
...
x2
0
1
2
3
...
...
0
1
...
x3
x4
0
0
1
0
1
2
...
0
0
...
0
1
0
2
1
0
...
0
2
...
Table 2: Table for Example 3: illustrating solution according to Apt’s definition.
are constructed. This omission can be a disadvantage, as is seen in the
important type of discrete csp that can be viewed as a graph-colouring
problem. In practical applications such csps have a small domain, consisting of the “colours”. At the same time they have a large number of
variables and a large number of constraints, both numbers running in the
thousands. Yet all these constraints have an important property in common: they derive from the “all different” constraint that requires that no
two of their arguments have the same value.
The remedy for this problem was prepared by Definition 1, which is
used in our alternative Definition 7 for csp. If the definition of csp included a language for expressing constraints, then these expressions would
clarify the connection between c1 and c2 . For example, sum(x2 , x2 , x1 )
would be a good expression for c1 and sum(x3 , x4 , x1 ) for c2 .
Predicate logic is a potential candidate for a formal constraint language. To realize this potential we modify Apt’s definition to obtain the
definition given in the following section. To be able to interface the solving algorithm for csps with the clp/dc scheme, we modify the algorithm
also. In the section after that we define how predicate logic can be used
as the constraint language.
5.2
A modified definition of CSPs
Definition 7 A Constraint Satisfaction Problem (csp) consists of a finite set X = {x1 , . . . , xn } of variables, a finite set C = {c1 , . . . , cm } of
constraints, each of which is a relation over a sequence of elements of X
in the sense of Definition 1. With each variable xi is associated a universe
Di , which is the set of values that xi can assume. A solution of a csp
is an assignment to each variable xi of an element of Di such that each
constraint in C is satisfied.
9
Apparently, the solution set of a csp with set X of variables is a
relation on X in the sense of Definition 1. A compact characterization of
the solution set can be given as follows.
Lemma 2 The solution set equals c1 ✶ · · · ✶ cm where ✶ is as in Definition 2.
For certain csps it is practical to enumerate the solutions. In other
cases the solution set, though finite, is too large to be enumerated. And
it may be the case that the solution set is uncountable; moreover its
individual solution tuples may consist of reals that are not computerrepresentable.
Thus it is often necessary to approximate the solution set. A convenient form is that of a Cartesian product D1 × · · · × Dn that is contained
in D1 × · · · × Dn . Such an approximation has the property that xi 6∈ Di
for any i ∈ {1, . . . , n} ensures that hx1 , . . . , xn i is not a solution. Making
D1 , . . . , Dn as small as possible gives us as much information about the
solution set as is possible for approximations of this form.
Di is called the domain for xi for i ∈ {1, . . . , n}. We need to ensure
that the subsets Di of Di are computer-representable. This may not be
a restriction when Di is finite and small. It is when Di = R. In general
we require that the subsets of Di that are allowed as Di include Di itself
and are closed under intersection. We call such subsets a domain system.
Lemma 3 Given sets D1 , . . . , Dn , each with a domain system and S ⊂
D1 ×· · ·×Dn . There is a unique least Cartesian product of domain system
elements containing S.
Definition 8 Given sets D1 , . . . , Dn , each with a domain system and
S ⊂ D1 ×· · ·×Dn . The least Cartesian product of domain system elements
containing S, which exists according to Lemma 3, is denoted ✷S.
With each constraint there is associated a domain reduction operation
(dro), which is intended to reduce the domains of one or more variables
occurring in the constraint.
Definition 9 Given a csp and a relation ρ ⊂ Di1 × · · · × Dik . Let
constraint c be a relation ρ on hxi1 , . . . , xik i A domain reduction operation
(dro ) for c is a function that maps Cartesian products Di1 × · · · × Dik ⊂
Di1 × · · · × Dik to Cartesian products of the same type. The map of the
function is given by Di1 ×· · ·×Dik 7→ Di′1 ×· · ·×Di′k where Di′1 ×· · ·×Di′k
satisfies
✷((Di1 × · · · × Dik ) ∩ ρ) ⊂ Di′1 × · · · × Di′k ⊂ Di1 × · · · × Dik .
If the left inclusion is equality, then we call the dro a strong one.
This operation was introduced by [7] under the name “narrowing”.
The intended application had intervals for the domains, hence the name.
Note that domains are reduced only by removing non-solutions. As one
can see, dros are contracting: if they do not succeed in removing anything,
they leave the domains unchanged. Strong dros are idempotent: multiple
successive applications of the same dro have the same effect as a single
application.
Success of the constraint satisfaction method of solving problems depends on finding efficiently executable strong dros.
10
5.2.1
Constraint Propagation
Definition 10 A computation state of a csp is D1 × · · · × Dn where
Di ⊆ Di is a domain and is associated with xi , for i = 1, . . . , n.
A computation of a csp is a sequence of computation states in which each
(after the initial one) is obtained from the previous one by applying the
dro of one of the constraints.
The limit of a computation is the intersection of its states.
A fair computation of a csp is a computation in which each of the constraints is represented by its dro infinitely many times.
Fair computations have infinite length. However, no change occurs from
a certain point onward (domain systems have a finite number of sets).
By the idempotence of strong dros, this is detectable by algorithms that
generate fair computations, so that they can terminate accordingly.
Theorem 3 [3] The limit of a fair computation of a csp is equal to the
intersection of the initial state of the computation with the greatest fixpoint
common to all dros.
For a given csp the intersection of the states of any fair computation
only depends on the initial state. It is therefore independent of the computation itself. Apparently the csp maps the set of Cartesian products to
itself. It is a contracting, idempotent mapping.
Lemma 4 Let D be the initial state of a fair computation of a csp. Then
the limit of the fair computation contains the intersection of D with the
solution set.
Definition 11 The transition from the initial state of a computation to
the limit of that computation is called constraint propagation.
The reason for the name is that the effect of a dro application on a
domain may cause a subsequent dro applications to reduce other domains.
5.2.2
Enumeration
Constraint propagation only goes part way toward solving a csp: it results
in a single Cartesian product containing all solutions. In general this single
Cartesian product needs to be split up to give more information about
any solutions that might be contained in it. This is what enumeration
does.
Before a more precise definition, let us sketch the solving process by
means of the csp arising from a graph-colouring problem. In case constraint propagation yields an empty domain in the computation state, the
solving process is over: absence of solutions has been proved. Suppose
the resulting computation state does not have an empty domain. We only
know that any solutions that may exist are elements of the Cartesian product of the domains. If all domains are singletons, then the corresponding
tuple is a solution. If not, one enumerates a domain with more than one
element (say, the smallest such). In turn, for each element in that domain, one assumes it as the value of the variable concerned and leaves the
other domains unchanged. To the smaller CSP thus obtained, one applies
11
constraint propagation. This may, in turn, require enumeration; and so
on.
To make the idea applicable to the case where there are infinite domains, we split a domain instead of enumerating it. Then it works as
above if the domains are countable.
To split an uncountable domain, then we need the property that the
domain system is finite. Splitting is restricted to producing results that
belong to the domain system. This implies that only a finite number of
splits are possible. In case of an uncountable domain it is not in general
possible to identify solutions.
Enumeration yields tuples consisting of domains that are as small as
the domain system allows that together contain all solutions, if any exist.
And of course solving the CSP results in eliminating almost all of the
Cartesian product of the initial domains as not containing any solutions.
Enumeration algorithm
To enumerate computation state S:
If a domain is empty, then halt.
If one of the domains is a singleton,
then substitute the element as value of the corresponding variable
and construct the computation state S ′ with that variable eliminated.
Enumerate S ′ .
else split a domain d into domain system elements d0 and d1 .
Construct computation states Si by replacing d in S by di , for i = 0, 1.
Enumerate S0 ; Enumerate S1 .
Often too many enumeration results are generated. Sometimes the
domain system comes with a suitable notion of adjacency so that adjacent enumeration results can be consolidated into a single one. Such a
consolidation may trigger further consolidations.
6 The Domain Constraint subscheme of
the CLP Scheme
The clp scheme is open-ended: it is basically a scheme for using Hornclause rules to generate a multitude of constraint-satisfaction problems.
The parameters of the scheme allow a great variety of useful algorithms
and of data-types for these to act on. A first step in reducing the vast
variety of options is clp/dc, the domain-constraint subscheme of the clp
scheme. We define clp/dc by visiting first the parameters hΣ, I, L, T i,
and then the transitions of the clp scheme.
6.1
The parameters
Σ: Some domains are such that individual elements may not be representable in a computer, if only because there are infinitely many of them.
Satisfactory results can still be obtained by designating a finite set of
subsets of the domain that are computer-representable. To accommodate
12
these the signature Σ includes a unary representability predicate for each
of these subsets.
I: the domain component D of the Σ-structure I has to admit a
domain system: a finite set of subsets of D that includes D and is closed
under intersection.
L: the language of constraints consists of conjunctions of atomic formulas.
T : the theory giving the semantics of the constraints links unary representability predicates to representable subsets of D. This is done in
part by clauses describing the effects of the dros. In Definition 9 let the
constraint c be r(xi1 , . . . , xik ). Then the clauses describing the dro of c
are
(1)
d′j (xij ) ← d1 (xi1 ), . . . , dk (xik ), r(xi1 , . . . , xik )
for j = 1, . . . , k. Further details depend on the instance concerned of
clp/dc. The idea of expressing the action of a dro in the form of an
inference rule is due to [2]. This is closely related to the inclusion of an
implication like the one above in a theory.
6.2
The transitions
The →r and →c transitions: These only serve to transform goal atoms
into constraint atoms, and are needed unchanged in the clp/dc subscheme.
The →i transition: In the clp scheme this transition is intended to
accommodate any inference that transfers the burden of constraint from
the passive constraints P to the efficiently solvable active constraints A.
In the clp/dc subscheme such inference is restricted to those forms that
leave P unchanged: the information contained in them is only used to
strengthen the active constraints A. Moreover, A is restricted to the
form {d1 (x1 ), . . . , dn (xn )} where each variable in the passive constraint
P occurs exactly once and where {d1 , . . . , dn } are unary representability
predicates.
As P is the unchanging conjunction of the constraints, we refer to
it as C in the clp/dc subscheme. As A = {d1 (X1 ), . . . , dn (Xn )} only
states of each of the variables that it belongs to a certain domain we refer to it as D in the clp/dc subscheme. As a result of these renamings
we have a close relationship between csp and clp/dc: C and D in csp
and in clp/dc are counterparts of each other. As a result of these restrictions and renamings, the constraint store management transition becomes
hG, D, Ci →i hG, D′ , Ci if hD′ , Ci = infer(D, C).
The infer operation is performed by setting up a csp with an initial
state and determining the limit of the fair computations from the initial
state. This limit is then the D′ in hD′ , Ci = infer(D, C).
The csp that implements infer in this way has the following components.
1. The variables are those that occur in the passive constraint C.
2. The universes over which the variables range are equal to each other
and to D.
13
3. If a constraint atom cj of clp/dc is r(xi1 , . . . , xikj ), then the corresponding constraint of the csp is ρ, the meaning of r, on hxi1 , . . . , xikj i,
with “on” as in Definition 1.
4. If the active constraint is {d1 (x1 ), . . . , dn (xn )}, then the initial computation state in the csp is D1 ×· · ·×Dn with Di = {x ∈ D | di (x)},
for i = 1, . . . , n.
In the csp thus obtained a fair computation is constructed with limit
D1′ × · · · × Dn′ . These domains are then used to determine the active
constraint D′ = {d′1 (x1 ), . . . , d′n (xn )}, where the di are obtained from
Di′ = {x ∈ D | d′i (x)}, for i = 1, . . . , n.
In this way csp computations can be used in clp/dc.
The →s transition: In the clp scheme this transition checks as best
as it can whether P ∧ A is consistent. In the clp/dc subscheme no
attempt is made to check C for consistency. It does this only for D =
d1 (x1 ) ∧ · · · ∧ dn (x) and this is simply a check whether any of the di is the
predicate for the empty subset of D.
Lemma 5 The existence of a successful derivation implies that C, T |=
R(C) ⊂ R(Q).
Proof 1 By theorem 1 we have C, T |= R(C ∧ D) ⊂ R(Q) and we have
C, T |= R(C) ⊂ R(D).
7
Interval constraints
We have used in Section 4 the clp scheme as starting point. To establish
the direction in which to proceed, we identified in Section 5 a desirable
point outside of logic programming: the csp paradigm. Here are to be
found useful algorithms for computational tasks of interest. These range
from the “most discrete” such as graph colouring to the “most continuous”
such as solving non-linear equalities and inequalities over the reals.
After thus establishing a line along which to travel, we went back in
Section 6 to establish a subscheme of the clp scheme, that of the domain
constraints, to emulate within logic the main features of the csp paradigm.
It is now time to declare our main interest: real valued variables rather
than discrete ones. It so happens that there is a real-variable specialization
of the csp paradigm, interval constraints, and it will be useful to take an
excursion from logic again and review this next.
We are interested in csps with the following characteristics. The variables range over the reals; that is, all universes D1 , . . . , Dn are equal to
the set R of reals. The domain system is that of the floating-point intervals. The constraints include the binary ≤ and the ternary sum and
prod. The reason is that these have strong dros that are efficiently computable. Strong dros are also available for =, max, abs, and for rational
powers. For the constraints corresponding to the transcendental functions
dros are available that are idempotent, but not strong. The definition of
“strong” requires them to be the least floating-point box containing the
intersection of the relation with the argument box. That the dro is not
14
strong has to do with the difficulty of bounding these function values between adjacent floating-point numbers. But dros closely approximating
this ideal are used in some systems [16].
Let us consider an example of a dro for use with real-valued variables
constrained by the relation
sum = {hx, y, zi ∈ R | x + y = z}.
Suppose the domains for x, y, and z are [0, 2], [0, 2], and [3, 5]. Clearly,
neither x nor y can be close to 0, nor can z be close to 5. Accordingly,
when this dro is applied, these intervals are reduced to [1, 2], [1, 2], and
[3, 4].
The numbers 1 and 4 arise by computing 3 − 2 and 2 + 2. Here no
rounding errors were made. This is exceptional. Let us now consider the
case in which the initial intervals are scaled down by a factor of ten to
[0.0, 0.2+ ], [0.0, 0.2+ ], and [0.3− , 0.5+ ]. Here 0.2+ is the least floatingpoint number not less than 0.2, and similarly for the other superscripts.
Now the corresponding operations 0.3− − 0.2+ and 0.2+ + 0.2+ do incur
rounding errors. 0.3− − 0.2+ is evaluated to a floating-point number we
shall name 0.1−− ; similarly, 0.2+ + 0.2+ is evaluated to 0.4++ , so that
the dro gives the intervals [0.1−− , 0.2+ ], [0.1−− , 0.2+ ], and [0.3− , 0.4++ ]
for x, y, and z, respectively. Here 0.1−− may equal 0.1− or (0.1− )− . The
decimal equivalents of the binary floating-point numbers computed here
are so lengthy that users are neither willing to write nor to read them,
so that further containment precautions are called for on (decimal) input
and output.
In this way single arithmetic operations find their counterpart in interval constraints. To give an idea of how the arbitrarily complex arithmetic
expressions in nonlinear equalities and inequalities are translated to interval constraints consider the equation 1/x + 1/y = 1/z relating the
resistance z of two resistors in parallel with resistances x and y. Constraint processing is not directly applicable when, as we assume here, we
only have dros for sum and inv. We therefore convert the equation to
the equivalent form
∃u, v, w ∈ R. inv(x, u) ∧ inv(y, v) ∧ inv(z, w) ∧ sum(u, v, w).
Accordingly, the equation is translated to a CSP with X = hx, y, z, u, v, wi
and C = {inv(x, u), inv(y, v), inv(z, w), sum(u, v, w)}.
In numerical CSPs we can conclude, according to Theorem 4, that
the solution set is empty when the limit of the computation is empty.
However, a nonempty limit can still coexist with an empty solution set.
It is possible to develop dros for complex expressions such as 1/x +
1/y = 1/z [6]. It is useful to know that this paper is antedated by the
technical report version of [7].
8 CLP/NCSP: the CLP/DC subscheme
with a numerical CSP
In Section 6 we described how the open-ended clp scheme is narrowed
down to the domain-constraints subscheme clp/dc. In this section we
15
take a step further in this direction to obtain a subscheme suitable for
numerical computation. We do this by following the specification in Section 6.
8.1
The hierarchy of theories
Σ: The signature contains the language elements needed for the usual
theory of the real numbers: constants including 0 and 1; the unary function symbol −; the binary function symbols +, −, ∗, and /; the binary
predicates ≤ and ≥. To these we add:
• A unary representability predicate da,b for every floating-point interval in a given floating-point number system. For the IEEE-standard
double-length floating-point numbers this means in the order of 2127
unary predicates. Not a mathematically elegant signature, but a
finite one.
• Ternary predicates sum and prod.
I: the domain component D of the Σ-structure I is the set R of
real numbers. The domain system consists of the floating-point intervals, which are sets of reals. The floating-point intervals include R itself
and are closed under intersection, so include the empty interval.
T : to the axioms of the usual theory of the reals we add:
∀x. [d−∞,b (x)
↔
x ≤ b] for every floating-point number b
∀x. [da,b (x)
↔
a ≤ x, x ≤ b] for every pair of flpt numbers such that a ≤ b
∀x. [da,+∞ (x)
↔
a ≤ x] for every floating-point number a
∀x, y, z. [sum(x, y, z)
↔
x + y = z]
We refer to the resulting theory as T1 . The only difference with the usual
axiomatization of the reals is that meanings are established for the newly
introduced predicates.
The effect of the dro of a constraint is described in clauses as in
equation 1 in Section 6. For each of the atomic constraints in the passive
constraint C this causes clauses to be added to T1 . We call the resulting
theory T2 .
Theorem 4 Let C be the passive constraint, let D be the initial active
′
constraint, and let D1′ , . . . , Dm
be the active constraints corresponding to
the results of a csp enumeration starting with initial constraint corresponding to D and constraints corresponding to C. Then T2 |= R(C ∧D) ⊂
′
[R(D1′ ) ∪ · · · ∪ R(Dm
)].
It would be more convincing if we could assert that T |= R(C ∧ D) ⊂
′
)], as T is the usual theory of the reals, without
[R(D1′ ) ∪ · · · ∪ R(Dm
computer-related artifacts. This is not possible, as R(C) and R(D) contain constraint predicates and these do not occur in T . However, all
axioms that are in T2 and not in T are logical consequences of T .
Proof 2 Every application of a dro corresponds to an inference with one
of the rules in T2 of the form of Equation (1).
16
It is now time to look at examples of what we can do with the tools
developed so far. The first two examples concern a polynomial in a single
real variable and represent it by a term p in the variable x. In these
examples the problem is stated in a single constraint, so only uses a part
of the clp paradigm. The third example is a toy design problem. Here
the clp paradigm is fully exercised: multiple derivations are generated,
each of which is potentially a significant numerical csp.
8.2
Semantics of solving numerical inequalities
Consider the problem of determining where the given polynomial is nonpositive. This corresponds to the constraint p ≤ 0. In T2 we can translate
p ≤ 0 to a set C of constraints. For example, if p is x ∗ (x − 2) we have in
T2
∀x[x ∗ (x − 2) = 0 ↔ ∃v, w. sum(v, 2, x) ∧ prod(x, v, w) ∧ w ≤ 0]
so that we have the constraint Γ equal to {sum(v, 2, x)∧prod(x, v, w)∧w ≤
0}. A highly complex p will give rise to a C with many atoms and many
variables.
Soundness (theorem 4) implies that the active constraints in the answer
constraint for this problem clp/ncsp contains all intervals in which p
is zero or negative. Completeness implies that whenever we have for a
constraint Γ that
T2 |= R(Γ) ⊂ R(p ≤ 0)
(2)
there are m > 0 derivations ending in anwer constraints Γ1 , . . . , Γm such
that T2 |= R(Γ) ⊂ R(Γ1 ) ∪ · · · ∪ R(Γm ). We cannot replace Equation (2)
by T2 |= R(p ≤ 0) 6= ∅. This would be reducible to the problem of deciding
equality between two reals, a problem shown to be unsolvable [1].
8.3
Semantics of equation solving
A well-known numerical problem that can present computational difficulties is the one of determining R(p = 0).
Theorem 4 shows that the active constraints in the answer constraint
for this problem clp/ncsp contain all zeroes of the polynomial. It also
shows that in case of finite failure the polynomial has no zeroes. The
possibility remains that finite failure does not occur, yet there are no
zeroes. This is unavoidable. The problem of deciding whether T2 |=
R(p = 0) = ∅ reduces again to the problem of deciding equality between
two reals. The best we can hope for is attained here: showing emptiness
or finding small intervals in which all solutions, if any, are contained.
Completeness (Theorem 2) has nothing to say about this problem: it
is rare for a polynomial p to make T2 |= R(Γ) ⊂ R(p = 0) true for nonempty R(Γ). With respect to T2 the set R(p = 0) is a finite set of reals,
and it is rare for these to be a floating-point number. For most p, the
least R(Γ) containing containing any root of it is an interval of positive
width.
17
Conventional numerical computation produces single floating-point
numbers that are intended to be near a solution, and mostly are. Sometimes they are not, and one cannot tell from the program’s output. Interval arithmetic and numerical csps improve on this by returning intervals
that contain the solutions, if any, and by failing to return any intervals
in which it is certain that no solutions exist. clp/ncsp improves on this
by giving a logic semantics, of which Theorem 4 is an example. However,
interval arithmetic and interval constraints are limited in that they only
solve a single csp. A more important advantage of clp/ncsp is that, in
addition to solving csps, it automates the generation of the multiple csps
that are often required in scheduling and in engineering design. We close
by giving an example of this mode of operation.
8.4
A toy example in CLP/NCSP
Consider an electrical network in which resistors are connected to each
other. The network as a whole has a certain resistance. We have available
twelve resistors; three each of 100, 150, 250, and 500 ohms. From this
inventory we are to build a network that has a specified resistance so that
it can function as part in a larger apparatus. Fortunately there is a certain
latitude: the resistance of the resulting network has to lie between 115
and 120 ohms. The structure of the network is not given. This is a design
problem in addition to being a computational problem.
Even with the dozen components given in this problem there is a large
number of ways in which they can be connected. We can nest parallel
networks inside a series network, or the other way around, to several
levels deep. Evaluation of each such combination requires a non-negligible
amount of computation involving real-valued variables. The search space
is sizable, hence the importance of constraint propagation to eliminate
most of it.
Let us imagine for clp/ncsp a Prolog-like syntax. Please do not be
misled by the type writer-like font into believing in an implementation:
none exists. The figures given in the example are for illustration only and
are chosen to be merely plausible.
According to clp bodies of clauses contain both constraint atoms and
program atoms. We separate them with a semicolon: the constraints,
if any, come first. Instead of writing da,b (x) for the domain constraints,
we write for ease of typing <a|X|b> in the style of Dirac’s bra and ket
notation. When a is infinite, we write -inf; this is a single mnemonic
identifier, denoting that particular floating-point value. Similarly for b
and inf or +inf. We omit constraints like <-inf|X|+inf>, which do not
constrain their argument.
The predicate netw(A,N,B,R,PL) asserts that network represented by
N connects terminals A and B, has resistance R, and has parts list PL. The
term N can be at(X) for an atomic network, which is in this case a single
resistor; it can be ser(N1,N2), for two networks in series, or par(N1,N2),
for two networks in parallel.
1: netw(A,at(R),B,R,(r150:1).nil)
:- <149.9|R|150.1>;.
18
% Similarly for 100, 250, and 500 ohms.
2: netw(A,ser(N1,N2),C,R,PL)
:- sum(R1,R2,R);
netw(A,N1,B,R1,PL1), netw(B,N2,C,R2,PL2),
merge(PL1,PL2,PL).
3: netw(A,par(N1,N2),B,R,PL)
:- inv(R,RR),inv(R1,RR1),inv(R2,RR2), sum(RR1,RR2,RR);
netw(A,N1,B,R1,PL1), netw(A,N2,B,R2,PL2),
merge(PL1,PL2,PL).
Clause 1 says that the network can be atomic, consisting of a single
resistor with resistance R, represented by the term at(R). Its parts list is
a list consisting of a single item r150:1, being a resistor of nominal value
150 ohms in quantity 1. The condition of clause 1 states that the actual
resistance R, a real variable, belongs to the interval [149.9, 150.1], which
expresses the tolerance. There are similar clauses to represent the other
sizes of resistor that are available.
Clause 2 says the network can be ser(N1,N2), the series composition
of two networks N1 and N2 of unspecified structure, with resistance R,
that satisfies the constraint for the resistance of a serial composition of
networks: sum(R1,R2,R), which means that R1 + R2 = R.
In clause 3 the constraint means 1/R1 + 1/R2 = 1/R, which is the
constraint for resistances R1 and R2 in parallel giving resistance R.
The predicate merge is left as a black box. Suffice it to know that the
goal merge(PL1,PL2,PL) merges parts lists PL1 and PL2, which satisfy the
inventory restrictions, into parts list PL unless the latter does not satisfy
the inventory restrictions, in which case the goal fails.
The query
:- <149.9|R150|150.1>, ...;
netw(a, par(at(R150), ser(at(R500), par(at(R100), at(R250)))), b, R, PL).
succeeds without search to an answer that could include something like
<117.1|R|119.3>. The program looks like it has been written with such
queries in mind. However, as explained below, it also succeeds, though
with some search, to answer
:- <115.0|R|120.0>; netw(A,N,B,R,PL).
with
N = par(at(R150), ser(at(R500), par(at(R100), at(R250))))
and
<117.1|R|119.3>, <149.9|R150|150.1>, ....
In response to the latter query clp/ncsp has synthesized a suitable
network, thereby solving the design problem. It traversed a search space
consisting of multiple csps that was generated by clp derivations. Many
of these derivations were cut short by failing csps.
As in the first two examples, the soundness of Theorem 1 guarantees for
this problem that all networks that are found have a resistance contained
in the required interval. We noticed that in the case of polynomial roots
completeness has no interesting consequence. This was true because the
problem had the form of a single constraint with equality. In the design of
19
a resistor network there is a goal statement with a program atom. It gives
rise to mutiple derivations. Completeness (theorem 2) implies that in case
a solution exists, derivation are generated to cover the given interval for
the network’s resistance: all solutions are found.
9
Concluding remarks
clp/ncsp only incorporates numerical csps into the clp scheme. Other
types of csp such as those dealing with finite domains, can be incorporated
in the same way. In fact, the clp scheme is not restricted to including
special-purpose computation into the logic framework: it has remedies for
those difficulties that prevented Prolog from being a logic programming
language.
Pure Prolog held out the promise of a programming language with
logical-implication semantics. The impracticality of the occurs check in
unification and of symbolic implementation of numerical computation
caused standard Prolog to compromise semantics. In this paper we described a method for including the power of hardware floating-point arithmetic without semantical compromise. We should not lose sight of the fact
that the clp scheme also has a remedy for the other blemish of standard
Prolog: compromised unification. As Clark [9] showed, the Herbrand
Equality Theory, which requires the occurs check, is only one possible
unification theory for the clp scheme. It can be replaced by Colmerauer’s
Rational Tree Equality Theory, so that we have the prospect of a fully
practical programming language with logic-implication semantics.
References
[1] Aberth, O. 1998. Precise Numerical Methods Using C++. Academic
Press.
[2] Apt, K. 1998. A proof theoretic view of constraint programming.
Fundamenta Informaticae 34, 295 – 321.
[3] Apt, K. 1999. The essence of constraint propagation. Theoretical
Computer Science 221(1-2), 179–210.
[4] Apt, K. and Wallace, M. 2006. Constraint Logic Programming
Using ECLiPSe. Cambridge University Press.
[5] Apt, K. R. 2003. Principles of Constraint Programming. Cambridge
University Press.
[6] Benhamou, F., McAllester, D., and Hentenryck, P. V. 1994.
CLP(Intervals) revisited. In Logic Programming: Proc. 1994 International Symposium. 124–138.
[7] Benhamou, F. and Older, W. J. 1997. Applying interval arithmetic
to real, integer, and Boolean constraints. Journal of Logic Programming 32, 1–24.
[8] BNR. 1988. BNR Prolog user guide and reference manual.
20
[9] Clark, K. L. 1991. Logic-programming schemes and their implementations. In Computational Logic, J.-L. Lassez and G. Plotkin, Eds. MIT
Press, 487–541.
[10] Cleary, J. 1987. Logical arithmetic. Future Computing Systems 2,
125–149.
[11] Colmerauer, A. 1996. Les bases de Prolog IV. Tech. rep., PrologIA,
Marseilles, France. Chapter 2 of “Le Manuel de Prolog IV”.
[12] Davis, E. 1987. Constraint propagation with interval labels. Artificial Intelligence 32, 281–331.
[13] Dincbas, M., Hentenryck, P. V., Simonis, H., Aggoun, A.,
Graf, T., and Berthier, F. 1988. The constraint programming
language CHIP. In Proc. Int. Conf. on Fifth Generation Computer
Systems.
[14] Henkin, L., Monk, J. D., and Tarski, A. 1985. Cylindric Algebras,
Parts I, II. Studies in Logic and the Foundations of Mathematics.
North-Holland.
[15] Hentenryck, P. V. 1989. Constraint Satisfaction in Logic Programming. MIT Press.
[16] Hentenryck, P. V., Michel, L., and Deville, Y. 1997. Numerica:
A Modeling Language for Global Optimization. MIT Press.
[17] Jaffar, J. and Lassez, J.-L. 1987. Constraint logic programming.
In Proceedings of the 14th ACM SIGACT-SIGPLAN symposium on
Principles of Programming Languages. ACM, 111 – 119.
[18] Jaffar, J. and Maher, M. J. 1994. Constraint logic programming:
A survey. Journal of Logic Programming 19/20, 503–582.
[19] Jaffar, J., Michaylov, S., Stuckey, P. J., and Yap, R. H. 1992.
The CLP(R) language and system. ACM Transactions of Programming
Languages and Systems 14, 339–395.
[20] Lee, J. and van Emden, M. 1993. Interval computation as deduction in CHIP. Journal of Logic Programming 16, 255–276.
[21] Moore, R. E., Kearfott, R. B., and Cloud, M. J. 2009. Introduction to Interval Analysis. SIAM.
[22] Older, W. and Vellino, A. 1993. Constraint arithmetic on
real intervals. In Constraint Logic Programming, F. Benhamou and
A. Colmerauer, Eds. MIT Press, 175–195.
[23] van Emden, M. 1997. Value constraints in the CLP Scheme. Constraints 2, 163–183.
[24] van Emden, M. 2006. Set-theoretic preliminaries for computer scientists. Tech. Rep. DCS-304-IR, Department of Computer Science,
University of Victoria. Paper cs.DM/0607039 in Computing Research
Repository (CoRR) 2006.
21
| 6cs.PL
|
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014
1
Uncertainty Quantification of Oscillation
Suppression during DBS in a Coupled Finite
Element and Network Model
arXiv:1705.10496v1 [q-bio.NC] 30 May 2017
Christian Schmidt1 , Eleanor Dunn2 , Madeleine Lowery2 , Member, IEEE and Ursula van Rienen3 , Member, IEEE
Abstract—Models of the cortico-basal ganglia network and
volume conductor models of the brain can provide insight into
the mechanisms of action of deep brain stimulation (DBS). In
this study, the coupling of a network model, under parkinsonian
conditions, to the extracellular field distribution obtained from
a three dimensional finite element model of a rodent’s brain
during DBS is presented. This coupled model is used to investigate
the influence of uncertainty in the electrical properties of brain
tissue and encapsulation tissue, formed around the electrode after
implantation, on the suppression of oscillatory neural activity
during DBS. The resulting uncertainty in this effect of DBS
on the network activity is quantified using a computationally
efficient and non-intrusive stochastic approach based on the
generalized Polynomial Chaos. The results suggest that variations
in the electrical properties of brain tissue may have a substantial
influence on the level of suppression of oscillatory activity during
DBS. Applying a global sensitivity analysis on the suppression
of the simulated oscillatory activity showed that the influence of
uncertainty in the electrical properties of the encapsulation tissue
had only a minor influence, in agreement with previous experimental and computational studies investigating the mechanisms
of current-controlled DBS in the literature.
Index Terms—Deep brain stimulation, Uncertainty quantification, Finite element method, Neural network models.
I. I NTRODUCTION
D
EEP BRAIN STIMULATION (DBS) is a neurosurgical
therapy used to treat symptoms of neurodegenerative disorders including Parkinson’s disease (PD). Despite its approval
for the treatment of PD and its widely acknowledged success,
the mechanisms of action of DBS remain uncertain. Electrophysiological studies in patients and in animal models have
shown that PD is associated with excessively synchronized
oscillatory neural activity in the cortico-basal ganglia network
accompanied by an increased power in the beta frequency
range (15 Hz-35 Hz) [1], [2], [3], [4].
For almost two decades, computational models of DBS
have been developed to investigate its mechanisms of action.
Among these, biological models of the cortico-basal ganglia
network, based on the excitatory and inhibitory connections
between the different deep brain nuclei, have been used to
investigate the effect of square-wave DBS signals on the
The work was supported by the DFG (German Science Foundation) Research Training Group 1505 ”welisa” and ERC Consolidator grant (646923).
Asterisk indicates corresponding author.
1∗ C. Schmidt and U. van Rienen are with the Institute of General
Electrical Engineering, University of Rostock, 18059 Rostock, Germany
cschmidt18057 at gmail.com
2 E. Dunn and M. Lowery are with the School of Electrical, Electronic and
Communications Engineering, University College Dublin, Dublin 4, Ireland
network dynamics [5], [6], [7]. In these network models, the
DBS signal was applied as an intracellular current to the
neurons of the target nuclei, typically the subthalamic nucleus
(STN) or globus pallidus interna (GPi) for PD DBS. This
modeling approach does not capture the reduction in amplitude
of the extracellular potential with increasing distance from the
electrode or the effect of the extracellular field distribution
along the neuron axons. The distribution of the electric field
is determined by the electrode geometry, the electrode-tissueinterface, and the heterogeneity of the brain tissue as well as
its electrical properties.
Computational studies based on volume conductor models
of the brain have shown that neural activation in the proximity
of the stimulation electrode depends strongly on the field
distribution caused by the DBS signal and, therefore, on the
aforementioned factors [8], [9], [10]. The electrical properties
of biological tissue such as brain tissue and encapsulation
tissue are subject to uncertainty due to wide variations in the
values reported in the literature, most of which is based upon
measurement ex vivo [11]. There is considerable variation
among reported conductivity and relative permittivity values
at lower frequencies in particular. This is due largely to the
limitations of available techniques for measuring dielectric
properties of biological tissues within the low frequency range,
including electrode polarization [11].
Simulation results suggest that the uncertainty in the electrical properties of brain tissue can have a substantial influence
on the field distribution and stimulation amplitude required to
activate the target nuclei [12]. The impact that the volume conductor effects and material property uncertainty consequently
have on the network dynamics and the level of suppression
of oscillatory neural activity for a given stimulus, however, is
not clear.
To investigate the influence of parametric uncertainties in
the electrical properties of brain tissue on the suppression
of beta activity in the cortico-basal ganglia network, it is
necessary to quantify the uncertainty in the time-dependent
extracellular potential distribution in the brain as well as in
the firing patterns of the neuronal network. These kind of
stochastic problems can be solved with standard brute-force
approaches such as Monte-Carlo simulation This approach,
however, requires a large number of subsequent evaluations
of the computational model. Alternatively, computationally
efficient approaches such as that based on the generalized
Polynomial Chaos (gPC) technique may be used. This approach substantially reduces the number of model evaluations
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014
required by generating a surrogate model of the quantities of
interest that may be more easily evaluated. The reduction of
the number of required model evaluations for generating this
surrogate model is carried out by a multi-variate interpolation,
which requires the model to be evaluated only for specific
samples of the brain tissue electrical properties to compute
the multi-variate interpolation polynomials.
The aim of this study was to explore how uncertainty in
the electrical properties of brain tissue can effect predicted
levels of suppression of pathological neural oscillations in
the cortico-basal ganglia network during DBS. To address
this, a coupled model of volume conduction within a rodent’s
brain and neural activity within the cortico-basal ganglia
network during DBS was developed. The gPC method was
then applied to quantify the influence of uncertainty in the
electrical properties of brain tissue and encapsulation tissue.
The network model is based on a model of the cortico-basal
ganglia, which was previously used to investigate the relative
effects of antidromic and orthodromic activation of corticoSTN afferents [13]. To enable coupling between the field
distribution in the rodent’s brain and the network model, the
nodes of the cortico-STN afferent axons were first identified to
allow application of the corresponding extracellular potential.
The coupled model was then expanded to a stochastic model
with the electrical properties of the brain and encapsulation
tissue modeled as random variables. This expanded model was
used to investigate the influence of varying tissue electrical
properties on the suppression of beta-band oscillatory activity
in the network. A preliminary version of this study has been
reported in [14].
II. M ETHODS
A. Cortico-Basal Ganglia Network Model
The network model is based on a previous cortico-basal
ganglia model in which the stimulus was applied to corticoSTN afferent axons using an analytical, purely resistive point
source to model the field distribution during DBS [13]. In
this network model, the neurons in the STN, and globus
pallidus external (GPe) are modeled as single compartment,
conductance based Hodgkin-Huxley type models, with threshold crossing models representing the interneurons within the
cortex. The STN was based on a parkinsonian rat model which
captures the generation of plateau potentials [15]. The cortex
was composed of a threshold crossing model representing
cortical interneurons [16] and a multi-compartment model of
cortical neurons consisting of a soma, an axon initial segment,
a myelinated main axon, comprising five nodes of Ranvier
and four internodes, and an axon collateral which connects
to the STN. The soma was based on the regular spiking
neuron model developed by Pospischill et al [17], while the
axon initial segment, axon, and collateral were based on a
mouse model developed by Foust et al [18]. Each nucleus
and neuron type consisted of a population of 30 neurons.
The nuclei were connected by excitatory (glutamergic) and
inhibitory (GABA) synapses (Fig. 1). The type and direction of
connections between nuclei of the cortico-basal ganglia were
based on data available for the network in the diseased state
2
DBSgFieldgDistribution
(VolumegConductorgModel)
Cortex
CorticalgNeuron
Soma
-
+
Interneuron
AxongInitialgSegment
Axon
Collateral
+
GPe
+
STN
-
Fig. 1. Schematic illustration of the cortico-basal ganglia network with the
DBS applied to the cortical axons and branching collaterals to the STN.
Excitatory pathways and inhibitory pathways are indicated by ”+” and ”-”
signs.
[19], however, it is more difficult to ascertain the exact number
of connections between individual neurons in different nuclei.
The STN neurons received inhibitory input from the GPe and
excitatory input from three randomly chosen cortical neurons.
The PD state of the network was simulated by applying a
DC current to the cortical neurons, as the STN receives substantial direct cortical input [20]. The model connections and
parameters were further designed to match antidromic firing
rate results from a parkinsonian rat model [21], which showed
an increasing frequency of antidromic spikes as the frequency
of stimulation was increased, until reaching a peak at 125 Hz
stimulation frequency. and increasing the synaptic gains within
the network as described in [22], leading to oscillatory activity
in the STN neurons within the beta frequency range. The
network model was implemented in NEURON 7.1 [23]. A time
step of 1 µs was used and the random seed was reused to ensure
same connections between the nuclei and synaptic noises for
each computation. The model was initialized for 5 seconds
to reach the PD steady-state before the application of DBS.
DBS was then modeled by applying the computed extracellular
potential for 2 seconds to both the axon collateral and the
main axon of the cortical neurons. Pulse trains representing the
firing times of all STN neurons were summed and the power
spectrum of the resulting composite pulse train was computed.
The level of beta activity within the STN was then estimated
as the integral of the power spectrum of the composite spike
train in the frequency range between 15 Hz and 30 Hz [24].
The power spectra and level of beta activity in the GPe and
cortical neurons were similarly estimated.
B. Extracellular Field Distribution in the Rodent’s Brain
The volume conductor model of the rodent’s brain is
based on a three-dimensional digital atlas database of the
adult C57BL/6J mouse brain. The segmented brain atlas data
comprises an average from magnetic resonance microimages
of 10 adult males and has an isotropic resolution of 47 µm
[25]. The brain surface was parametrized using non-uniform
rational B-splines with the software Geomagic Studio R
(http://www.geomagic.com). The finite element model of the
rodent’s brain was generated using Comsol MultiphysicsTM
(http://www.comsol.com) by importing the parametrized brain
surface model and a model of a DBS microelectrode (diameter:
125 µm, electrode contact height: 320 µm [26]), as illustrated
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014
3
Cortical Neuron Model
3D Surface Model
Soma
-0.7
-0.6
Finite Element Model
OF THE UNIFORMLY DISTRIBUTED CONDUCTIVITIES OF BRAIN TISSUE
AND ENCAPSULATION TISSUE .
-0.5
μm
500
400
300
200
100
0
.
.
.
-0.4
(D: 25 μm,
L: 35 μm)
TABLE I
L OWER AND UPPER BOUNDARY, MEAN VALUE , AND RELATIVE DEVIATION
Tissue type
Brain tissue
Encapsulation tissue
Internode
(D: 1.4 μm,
L: 500 μm)
Field Distribution
During DBS Stimulus
Node
(D: 1.2 μm,
L: 2 μm)
Collateral
μm
500
(D: 0.5 μm,
L: 500 μm)
400
300
200
100
0
-0.8
-0.9
-1.0
Extracellular Potential During Stimulus [mV]
Fig. 2. Schematic illustration of the coupling between the volume conductor
model and the cortical neurons. The collaterals are positioned randomly
in the proximity of the electrode contact. The time-dependent extracellular
potential, shown here for a stimulation amplitude of 150 µA is applied along
the collateral and the nodes and internodes along the axon.
in Fig. 2. The microelectrode was positioned in the STN
area ventral and lateral to the thalamus by comparing the
image data with the Allen Brain Atlas of the mouse brain
(http://atlas.brain-map.org). The electrical ground of the model
was set to the area of the cerebellum to model the placement of
the ground electrode in the neck of the mouse. A square-wave
current-controlled DBS signal with a frequency of 130 Hz
and a pulse duration of 60 µs was applied to the electrode
contact via a Dirichlet boundary condition [27]. The timedependent field distribution was computed using the Fourier
Finite Element Method [8] by solving the Laplace equation
∇κ(r)∇ϕ(r) = 0
(1)
where κ(r) is the electric conductivity and ϕ(r) the electric
potential at location r in the computational domain Ω. Due
to reactions of the body to the implant, scar tissue forms
an encapsulation layer around the electrode. The electrical
properties of this highly resistive layer as well as brain tissue
vary considerably within the literature [9], [28]. To capture
these variations, their electrical properties were modeled as
uniformly distributed random variables with a standard deviation of 10 %, which corresponds to a relative deviation with
respect of the 97.5 percentile and 2.5 percentile of 16.5 %
(Table I). The mean value of brain tissue conductivity was
chosen according to the value of white matter at a frequency
of 2 kHz [29], which constitutes a good approximation of
Conductivity σ [mS/m]
Min
Mean
Max
64.0
32.0
77.4
38.7
90.8
45.4
Rel. Deviation
16.5 %
16.5 %
the dispersive nature of the electrical properties in volume
conductor models of the brain [10]. The mean value of the
encapsulation tissue was set to half the mean value of the
brain tissue [12].
DBS was applied to the cortical collaterals and axons,
following the results of an optogenetic study in 6-OHDA
rats that found that theraputic effects of STN DBS could be
accounted for by direct selective stimlation of afferent fibers
projecting from the cortex to the STN [30]. The collaterals
of the cortical neurons branching to the STN were randomly
distributed within a distance of 250 µm from the center of the
electrode contact using uniformly distributed random variables
for their Cartesian coordinates (Fig. 2). The collaterals were
orientated perpendicular to the main axis of the electrode,
with the longitudinal axes of each collateral centered at the
electrode. The axons of the cortical neurons, from which
the collaterals branched, were orientated perpendicular to the
collaterals and parallel to the electrode. To ensure a sufficient
resolution of the finite element mesh with respect to the size
of the collaterals and the axons, manual mesh refinement was
applied by setting a maximum mesh size of 25 µm in the area
of the collaterals, 50 µm in the area of the axons, and 10 µm
at the electrode contact surface. The final mesh consisted of
approximately 1.2 million elements resulting in approximately
1.7 million degrees of freedom using quadratic basis functions.
The linear system was solved using generalized minimal
residual method with a relative tolerance of 1 · 10−6 .
C. Coupling of the Extracellular Field Distribution to the
Network Model
The spatial distribution of the cortical neurons resulted in
differences in the extracellular potential along the collaterals
and main axons of each neuron. To incorporate the effect of
variations in extracellular field distribution on the dynamics of
the neural network, the extracellular field was coupled to the
nodes of the target axons in the discretized network model.
Following the approach of previous computational studies to
couple the field distribution from a finite element model to
neuron models [8], [31], the time-dependent potential of the
finite element model was computed at the location of each
segment of the cortical axons and collaterals in the neural
network model and was applied as a discrete time-dependent
extracellular potential.
In a previous study, the extracellular potential during DBS
was assumed to preferentially activate the branching collaterals
and the extracellular potential was, therefore, applied solely at
the collateral [13]. To capture potential effects of the extracellular potential on both the cortical axons and collaterals,
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014
in the present study the time-dependent extracellular potential
computed using the volume conductor model was applied to
each segment along both collaterals and axons at a total of
10 equidistant nodes along each collateral and 55 nodes along
each axon.
D. Uncertainty Quantification Using Polynomial Chaos
During the application of DBS, the firing times of neurons
within the cortex and STN will vary with the time-dependent
extracellular potential. The beta-band activity of the neural
network model will thus depend on the time-dependent extracellular potential distribution computed using the volume
conductor model. The field distribution, in turn, depends on
the conductivity of the encapsulation tissue and brain tissue.
In volume conductor models, the isotropic conductivity, and
permittivity if considered, of the biological tissues are typically
modeled using a single value. There is, however, considerable
uncertainty in the true values of the dielectric properties, due
to challenges in the measurement of the electrical properties of
biological tissue, especially in the low Hz and kHz frequency
regime [11]. To capture the uncertainty arising from this
variation, in this study, the conductivity of each tissue type was
described using a random variable to model its uncertainty. To
compute the statistics of the quantities of interest, standard
stochastic methods such as Monte-Carlo simulation could
be used. To provide a sufficient accuracy in the stochastic
measures, Monte-Carlo simulation requires a large number
of random samples of the model parameters [32]. For each
sample, the whole computational model must be solved from
which a corresponding sample of each quantity of interest can
be computed. If the deterministic model is computationally
expensive, as it is the case in the present study, this approach
becomes computationally inefficient.
To reduce the computational effort for determining the
stochastic measures of the quantities of interest, a nonintrusive variant of the gPC technqiue is applied. This technique determines a surrogate model of each quantity of interest
by expanding them in a series of multi-variate orthogonal polynomials and computing the model solution at predefined nodes
in the parameter space, which is spanned in this study by the
uncertain electrical properties of brain tissue [33]. These nodes
are determined by combining the one-dimensional ClenshawCurtis quadrature rule with a combinatorical algorithm to
generate a multi-dimensional sparse grid representation of
the parameter space (sparse-pseudospectral approach). These
model solutions at the grid nodes are used to compute the
coefficients of the multi-variate polynomial series expansion
with multi-dimensional numerical integration. The number of
grid nodes depend on the polynomial degree p of the gPC
expansion as well as on the number of model parameters
M [34]. In general, gPC is computationally advantageous
over Monte-Carlo simulation for a small number of model
parameters (e.g. M = 6 and p = 3 requires 389 evaluations of
the model). In addition, the series expansions of the quantities
of interest provided by the gPC approach allow for a simple
global sensitivity analysis by Sobol’ indices. These indices
determine the sensitivity of these quantities with respect to
4
each uncertain parameter as well as their stochastic interactions by evaluating their conditional variances [35]. In this
relation, the term conditional variance describes the variance
of the quantity of interest with respect to each uncertain
parameter, the uncertain electrical properties of brain tissue,
or any combination of them.
The gPC and sparse-pseudo-spectral approach has previously been successfully applied to other bio-electrical applications including optimizing the stimulation protocol in a multielectrode transcranial direct current stimulation setup as well
as for an electro-stimulative hip-revision system [36], [37].
The method was implemented in Matlab as a function package, which determines the deterministic parameter samples
on the chosen sparse grid based on the predefined properties
of the random model parameters and the polynomial degree
p, as described in [36]. The list of parameter samples was
used to subsequently initialize the solution process of the
volume conductor model with COMSOL and of the network
model with NEURON. From the resulting list of solutions, the
corresponding sample list of the desired quantities of interest
including the extracellular potential and the level of beta
suppression was derived, from which the respective series expansion, as mentioned above, were computed. The evaluation
of these series expansions of the quantities of interest for any
samples in the parameter space require only the subsequent
evaluation of the product of basis polynomials, which provides
a suitable model for Monte Carlo simulation [34]. Therefore,
the determined surrogate models of the quantities of interest
allow for the application of Monte Carlo simulation with
a large number of random samples to provide sufficiently
accurate stochastic measures.
III. R ESULTS
Under simulated parkinsonian conditions, STN neurons in
the model exhibited synchronized activity within the beta
frequency range (Fig. 3). To investigate the response of the
network model to DBS of increasing stimulation intensity,
the level of beta suppression in the STN neurons within the
network model was investigated for stimulation amplitudes
between 50 µA and 250 µA. At each stimulation amplitude,
the corresponding extracellular field distribution was computed
using the mean conductivity values of encapsulation and
brain tissue, as given in Table I. The beta-band power of
the point process representing the firing times of the STN
neurons progressively decreased as the amplitude of the DBS
signal was increased (Fig. 4(d)). As the DBS amplitude was
increased, an increase in the number of activated collaterals
and axons was observed, leading to an increase in the region
of directly stimulated tissue. (Fig. 3). The collateral and axonal
activation orthodromically activated the STN leading to a
gradual entrainment of the STN neurons to the stimulus pulse
and subsequent reduction of the level of beta-band oscillations
(Fig. 4(a)), which is also noticeable for the GPe neurons (Fig.
4(b)). The cortical neurons and inhibitory interneurons were
simultaneously antidromically activated, reducing beta-band
activity from the cortex to the STN (Fig. 4(c)). For stimulation
amplitudes of 150 µA and higher, a slight increase in the
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014
5
Fig. 3. Firing patterns of the STN, GPe, and Cortical Neurons in the PD state with no DBS applied, and for DBS stimulation amplitudes of 130 µA, 140 µA,
and 250 µA. The patterns were recorded for the cell bodies of each neuron.
(a)
(b)
(c)
(d)
Fig. 4. Power spectral densities, normalized with respect to the DBS Off condition, of the (a) STN, (b) GPe, and (c) cortical neurons for different stimulation
amplitudes. The average beta suppression is shown for all three nuclei in (d).
amplitude of the beta-band high-frequencies between 30 Hz
and 40 Hz can be seen. The increased activity in this frequency
range is accompanied by a more dense spiking pattern at
250 µA compared to that at 150 µA (Fig. 3). Nevertheless, a
clear suppression of most of the power spectrum due to an
almost complete entrainment of all STN, GPe, and cortical
neurons, is observed for all stimulation amplitudes between
140 µA and 250 µA (Fig. 4(d)).
Following evaluation of the network response with the tissue
conductivities in the volume conductor model set to their
respective mean values, the gPC method with a polynomial
order of p = 4 was used to investigate the influence of
uncertainty in the tissue conductivity on the quantities of
interest. Using the sparse-pseudo spectral approach and the
multi-dimensional implementation of the numerical integration
method based on the Clenshaw-Curtis rule, the computation
of the corresponding gPC series expansions required the
evaluation of the computational model for 29 deterministic
samples of the tissue conductivity for the computation of
the uncertain extracellular potential. The uncertainty of the
extracellular potential within the proximity of the stimulation
electrode was approximately 15 % and 18 % with respect to
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014
TABLE II
M EAN VALUE , LOWER QUANTILE (2.5 PERCENTILE ), AND UPPER
QUANTILE (97.5 PERCENTILE ) OF THE UNCERTAIN BULK TISSUE
IMPEDANCE DETERMINED FROM THE VOLUME CONDUCTOR MODEL WITH
UNCERTAIN TISSUE PROPERTIES .
Stochastic Measure
Lower quantile
Mean value
Upper quantile
Impedance [kΩ]
Rel. Deviation
15.9
13.8
12.0
15.4 %
0.0
13.1 %
the upper quantile (97.5 percentile) and the lower quantile
(2.5 percentile), respectively. Exemplarily results are presented
for a reference cortical collateral positioned perpendicular
to the electrode at a radial distance of 100 µm and for a
stimulation amplitude of 250 µA in Figure 5(a). The slightly
larger deviation with respect to the lower quantile compared
to the upper quantile arises from the inversely proportional
relationship between conductivity and potential distribution for
current-controlled stimulation, which results in an asymmetric
probability density function and, therefore, unequal differences
in the mean value [12]. The level of uncertainty in the
time-dependent extracellular potential at a specific location
was similar at the other locations of the collateral and axon
segments (Fig. 5(b,c)). Comparing the average relative deviation of approximately 16.3 % in the extracellular potential to
the uncertainty in the tissue conductivities of approximately
16.5 % (Table I), an almost 1:1 relation between the level
of their uncertainty can be noted. A similar relationship was
observed for the uncertainty in the bulk tissue impedance of
the volume conductor model (Table II). An increase of the
conductivity of brain tissue as well as encapsulation tissue
resulted in a decrease in the bulk tissue impedance as well
as the extracellular potential distribution, and vice versa.
Regarding the accuracy of the stochastic measures in these
quantities of the volume conductor model, a doubling in the
polynomial order of p = 8, which required the computation
of additional 36 model evaluations, resulted in a difference of
below 9 · 10−5 % and 0.03 % in their mean values and relative
deviations, respectively. Since the extracellular potential distribution within the volume conductor model scales linearly
with respect to the applied stimulation amplitude due to the
linear material properties and the Laplace equation (1), the
solution for different stimulation amplitudes was obtained by
simply multiplying the reference solution by the corresponding
stimulation amplitude.
While this linear property of the volume conductor model
allowed for a reuse of the reference solution, which only had to
be computed once, the deterministic solution of the network
model had to be reevaluated for each stimulation amplitude
due to the nonlinearities in the dynamics of the network model
and individual neurons [13]. This nonlinearity in the response
of the network model with respect to the stimulation amplitude
is apparent in the suppression-amplitude relation shown in
Figure 4. Recent studies have demonstrated suppression of
beta band activity witihn the basal ganglia local field potential
(LFP) during DBS, accompanied by a reduction in the parkinsonian motor symptoms of rigidity and bradykensia [38]. The
6
presence of beta band LFP oscillations are interpreted as being
indicative of pathological synchronization between neurons
and have attracted substantial interest as a potential biomarker
of the parkinsonian state [39]. The influence of uncertainty
in the tissue electrical properties on the level of suppression
of beta frequency neural oscillations in the STN was thus
chosen for investigation. The effect on uncertainty in the
material properties of the brain and encapsulation tissue on
uncertainty in the level of suppression of synchronous beta
band activity among the STN neurons is shown in Figure 6.
For lower conductivity values, as in the case of the lower
quantile values, a decrease in the STN beta-band activity occurred at lower stimulation amplitudes as the potential within
the bulk tissue for a given stimulus amplitude was higher.
Higher stimulation amplitudes were correspondingly required
to elicit comparable levels of beta suppression when higher
conductivity values, as in the case of the upper quantile values,
were assumed within the model. Therefore, the uncertainty in
the tissue conductivities introduced a large uncertain area in
the window of effective stimulation spanned by the lower and
upper quantiles (Fig. 6(a)). Stimulation amplitudes below this
window were too low to sufficiently modify the membrane
potential along the cortical axons and collaterals to have an
effect on the spiking pattern in the STN neurons. Conversely,
stimulation amplitudes above this window seem to have no
additional effect on the spiking pattern. This effect can be
explained by investigating the membrane potential along the
cortical neurons. These neurons were already entrained by
the DBS stimulus and, therefore, a further increase in the
stimulation amplitude did not influence the dynamics of the
already stimulated cortical axons (Fig. 3). The resulting level
of uncertainty was greatest for a stimulation amplitude of
120 µA with an average value of approximately 260 % (Fig.
6). The asymmetry in the relative deviation with respect to
the upper and lower quantile of the extracellular potential and
bulk tissue impedance (Fig. 5 and Table II) is also noticeable
in the uncertainty of the level of beta suppression. In general,
the uncertainty in the level of beta suppression in the network
model was substantially larger than the predefined uncertainty
in the tissue conductivities, which can be ascribed to the
nonlinearity in the mathematical formulation of the network
model. This nonlinear behavior affects also the accuracy of the
stochastic measures with respect to the polynomial order of
the gPC. The uncertain beta supression of beta activity in the
network model was computed using gPC with a polynomial
order of p = 8, which required the evaluation of 65 network
models for each stimulation amplitude. A doubling of the
polynomial order of p = 16 resulted in a difference of less than
0.1 % and 9.1 % in the mean value and relative deviations of
the level of beta suppression, respectively, evaluated at a stimulation amplitude of 120 µA, which is substantially larger than
for the stochastic measures of the volume conductor model.
Nevertheless, the accuracy in the stochastic measures of the
network model is sufficiently large for an investigation of the
overall trend of its uncertainty. Regarding the required number
of random samples to determine the stochastic measures of the
surrogate models obtained by the gPC method, a number of
100.000 random samples was sufficient to provide an accuracy
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014
UpperpQuantile
MeanpValue
LowerpQuantile
-5
14.5p4
18.0p4
-10
0
50
100
Timep[µs]
PotentialpAlongpCollateral
8b(
150
ExtracellularpPotentialp[mV]
ExtracellularpPotentialp[mV]
0
0
UpperpQuantile
MeanpValue
LowerpQuantile
-5
-10
-200
PotentialpAlongpAxon
8c(
ExtracellularpPotentialp[mV]
PotentialpatpCollateralpSegmentp5
8a(
7
-100
0
100
200
CollateralpSegmentsp[µm]
0
-5
-10
0
UpperpQuantile
MeanpValue
LowerpQuantile
500 1000 1500 2000 2500
AxonpSegmentsp[µm]
Fig. 5. Mean value, lower (2.5 percentile), and upper quantile (97.5 percentile) of the uncertain extracellular potential (a) at the 5th segment of a reference
cortical collateral positioned perpendicular to the electrode contact at a radial distance of 100 µm, (b) along the cortical collateral at at time 50 µs, (c) and
along the corresponding cortical axon.
TABLE III
S ENSITIVITY OF THE BULK TISSUE IMPEDANCE AND EXTRACELLULAR
POTENTIAL WITH RESPECT TO THE UNCERTAINTY IN BRAIN σbrain AND
ENCAPSULATION TISSUE σencap AS WELL AS THEIR INTERACTION .
Quantities of Interest
σbrain
Bulk Tissue Impedance
Extracellular Potential
Fig. 6. (top) Mean value, lower quantile (2.5 percentile), and upper quantile
(97.5 percentile) of the beta-band suppression across all STN neurons with
increasing stimulation amplitude. (bottom) Relative deviation between the
mean value of the beta-band suppression and the corresponding lower quantile
(blue) and upper quantile (red).
91.1
99.9
Sensitivity [%]
σencap
(σbrain , σencap )
8.9
0.1
0.0
0.0
Since the dynamics of the network model depend on the
extracellular potential along the axons and collaterals of the
cortical neurons distributed in the volume conductor model,
a similar behavior would be assumed for the sensitivity in
the level of beta suppression. Investigating this sensitivity for
stimulation amplitudes between 120 µA and 170 µA, this is
confirmed (Fig. 7). Nevertheless, the influence of uncertainty
in encapsulation tissue conductivity as well as its interaction
with the uncertainty in brain tissue conductivity seem to
increase for stimulation amplitudes below and above that
range, which is assumed to be caused by dynamic effects
related to the random noise in the network model and high
sensitivity of the neurons to the extracellular potential when
close to the threshold for excitation.
IV. D ISCUSSION
Fig. 7. Sensitivity of the beta-band suppression to the uncertainty in brain
tissue, encapsulation tissue, and the interaction of the uncertainty in both
tissue types.
of approximately 0.1 %.
The investigation of the sensitivity of the quantities of
interest by evaluating the Sobol’ indices revealed that the
major influence on the uncertainty in the extracellular potential
and the bulk tissue impedance is due to the uncertainty in
brain tissue conductivity (Table III). As expected, due to the
linearity of the volume conductor model, no interaction effects
of the uncertainty in both tissue conductivities is noticeable.
The coupling of a cortico-basal ganglia network model with
the extracellular field distribution in the target area of DBS
allows for an investigation of the mechanisms of action of
DBS in a simulation model that captures both network and
volume conductor effects. Analytic models of the extracellular
field distribution add the spatial location and extracellular
stimulation of the neurons to the network model, but are
unable to consider effects of the geometry of the electrode, the
brain anatomy, and the electrode-tissue interface. Therefore,
in this study a 3D finite element model of a rodent brain was
coupled to a network model to investigate the influence of
uncertainty in the conductivity profiles of biological tissue on
beta-band suppression in the network model during DBS. The
application of the gPC method allowed for a computationally
efficient quantification of the uncertainty in the quantities
of interest of the volume conductor model and the network
model on a common workstation, which would be infeasible
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014
if standard stochastic methods such as Monte Carlo simulation
were used. The computation time and memory consumption
for one realization was approximately 2 hours and 3.5 Gb on
a 2.4 Ghz core of a common workstation, with over 95 %
of the computation time utilized by the network model. The
computation of the realizations for all tissue conductivity
samples and stimulation amplitudes required less than 10 days
on 12×2.4 Ghz cores of the workstation. A Monte Carlo simulation with comparable accuracy, which would have required
a large number of several thousand random samples, would
take an estimated computational time of several months and a
substantially larger amount of storage. Further, the surrogate
models determined by the gPC method of the investigated
quantities of interest consist of a series expansion of multivariate polynomials, which are evaluable for a large number
of random samples by Monte Carlo simulation in a comparably
small amount of time.
The window of stimulation amplitudes, at which the level
of beta suppression is sensitive to uncertainty in the electrical
properties of brain and encapsulation tissue, ranges from
110 µA to 185 µA (Fig. 6) for the assumed uncertainty in the
tissue conductivities of the volume conductor model (Table
I). The stimulation amplitude required to effectively modify
the spiking patterns in the neurons of the network model,
and subsequently the level of beta suppression, depends not
only on the tissue conductivity, but also on various factors
such as the exact location of the neurons in the proximity of
the electrode, the properties of the target neuron axons, the
electrode geometry, and the modeling of the encapsulation
layer. Despite these factors, similar stimulation amplitudes
to suppress PD symptoms in experimental studies in rodents
have been reported with values between 30 and 100 µA for a
130 Hz DBS stimulus effective [40], [41]. In [41], an effective
stimulation radius of 250 µm for a stimulation amplitude of
100 µA was reported. The axon collaterals in the present study
were positioned within a similar radius around the stimulation
electrode. The magnitude of the beta-band oscillatory activity
within the network was substantially reduced at stimulation
amplitudes for which not all STN neurons were directly
entrained by the DBS (Fig. 4). This result suggests that the
estimation of the “volume of tissue activated” alone to predict
neural activation during DBS [8], [9], [12] may also overestimate the required stimulus amplitude to obtain therapeutic effects of DBS, as the propagation of stimulation effects through
the connections within the cortico-basal ganglia network is not
considered.
The volume conductor model of a rodent’s brain used in
this study consists of a DBS electrode, which is surrounded
by an encapsulation tissue layer and is placed within the
region of the STN, and the bulk brain tissue medium. The
model allows for computing the field distribution around the
electrode with respect to the influence of the encapsulation
and brain tissue. Given that this medium is homogeneous,
the extracellular potential around the cylindrical electrode
resembles an almost spherical field distribution with increasing distance to the electrode with the exterior boundary of
the rodent brain having most likely only a minor influence.
Nevertheless, neural elements can be located very close to
8
the encapsulation layer, as it is the case in this study, which
requires a sufficiently detailed model of the electrode and
the neighboring areas. In addition, the generated 3D rodent
brain model allows arbitrary orientations and locations for the
electrode and neuronal elements inside the brain to be chosen
and the brain structures to be incorporated. In the future, the
model will be extended with models of basal ganglia brain
structures, which will result in heterogeneous and anisotropic
electrical properties of the corresponding tissue types within
the model and, therefore, result in a complex, non-spherical
field distribution.
The network model used in this study focuses on the STNGPe loop and cortical inputs to the STN. The application of
DBS to the model is based upon the hypothesis that DBS
exerts its effects, at least partially, through the activation of
STN afferent fibers projecting from the cortex [42]. In this
model, DBS activates corticofugal fibre afferent projecting
from the cortex to the STN. The resulting activity antidromically modulates the firing patterns of the cortical neurons while
also orthodromically activating the single compartment STN
neurons. Previous computational studies have simulated the
effect of extracellular potential on isolated neurons [43], [44]
or used a simplified point source approximation to apply the
potential to the network [13]. Though the present study is
the first to combine a detailed volume conductor model with
a network model, it is likely that DBS may also effect other
neurons and network circuits, not included here, either directly
or indirectly [45]. In particular, the question as to whether DBS
results in excitation or inhibition of STN neurons remains a
controversial topic with both excitatory and inhibitory effects
reported [46], alongside the possibility that DBS may inhibit
activity of the STN soma while simultaneously activating the
afferent output of the STN [44]. There is greater agreement
regarding the effects of DBS on the afferent output of the
STN, with it generally accepted that STN DBS results in
stimulation of the STN output time-locked to the stimulus
[47]. Previous models of the corticobasal ganglia network have
stimulated STN DBS through the intracellular stimulation of
single compartment STN neurons [7], [48], [6], [5]. The single
compartment representation of STN neurons in the present
study approximates the “input-output” characteristics of the
STN neurons and does not rule out the possibility that the cellbody of the STN neurons may be inhibited while the nerve
axons are simultaneously stimulated. Multi-compartment models of these nuclei, especially of the STN, would increase the
physiological accuracy of the model. However, this would also
substantially increase the computational time and the use of
computational resources. For example, the model used in this
study considered in total 1, 950 neural segments along the cortical neurons for extracellular stimulation and required for one
realization approximately 3.5 Gb memory. Replacing the STN
neurons in this study by multi-compartment models, as used
for example in [8], would require extracellular stimulation of
additional 6, 630 neural segments, which would result in a
substantial increase of the required computational resources
and computational time. The number of neurons in each
nucleus of the network model, while consistent with what has
been used in previous computational studies [5], [49], is lower
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014
than it is physiologically realistic. The relatively low numbers
of neurons, which are highly interconnected throughout the
basal ganglia-network, make the model sensitive to changes
in electric potential as even small changes in potential can
quickly increase the percentage of neurons activated.
The development of realistic volume conductor models
coupled to network models of the basal ganglia move towards the prediction of required stimulation amplitudes to
obtain therapeutic effects of DBS, based on realistic brain
anatomy and electrical properties as well as the electrodetissue-interface and neural dynamics. For this problem class
and computational models, the results of this study are of
crucial importance, since they show the substantial influence
of brain tissue uncertainty on pathological neural oscillations
which are one possible objective function of the corresponding optimization problem. Therefore, an objective function
omitting this uncertainty may not sufficiently render patientspecific stimulation profiles, but only represent one possible
solution.
Regarding the minor influence of the uncertainty in encapsulation tissue conductivity on the level of beta suppression
and the extracellular potential, the results of this study are in
agreement with other computational and experimental studies
concerning current-controlled DBS [8], [9], [50]. For voltagecontrolled stimulation, a more precise knowledge of the encapsulation layer properties in the post-operative state may be
crucial for modeling the mechanisms of action of DBS in the
rodent’s brain [12].
V. C ONCLUSION
Uncertainty in the electrical properties of biological tissues,
including grey and white matter and the encapsulation tissue
which forms around the electrode, represents a significant
limitation of existing models of the electric field in the vicinity
of the DBS electrode. The potential impact of this uncertainty
on the dynamics of neural firing and pathological neural oscillations within the cortico-basal ganglia network during DBS
is unclear. The methodology presented in this paper offers a
computationally efficient method with which to quantify these
effects by coupling a finite element model of the electric field
with a physiologically-based network model, and applying
a stochastic approach based on the generalized Polynomial
Chaos. The results of this study indicate that modest variations
in the electrical conductivity of brain tissue of 13 − 15 %,
can result in substantial deviations, of up to 300 % about the
mean, in the level of beta suppression among STN neurons
during DBS. The methods proposed offer an efficient method
with which to incorporate known uncertainties in parameters
of interest in models of DBS.
R EFERENCES
[1] V. Litvak, A. Jha, A. Eusebio, R. Oostenveld, T. Foltynie, P. Limousin,
L. Zrinzo, M. Hariz, K. Friston, and P. Brown, “Resting oscillatory
cortico-subthalamic connectivity in patients with parkinson’s disease,”
Brain, vol. 134, pp. 359–374, Feb 2011.
[2] A. Eusebio, C. C. Chen, C. S. Lu, S. T. Lee, C. H. Tsai, P. Limousin,
M. Hariz, and P. Brown, “Effects of low-frequency stimulation of the
subthalamic nucleus on movement in parkinson’s disease,” Exp Neurol,
vol. 209, pp. 125–130, Jan 2008.
9
[3] A. A. Kühn, F. Kempf, C. Brücke, L. Gaynor Doyle, I. Martinez-Torres,
A. Pogosyan, T. Trottenbrg, A. Kupsch, G. H. Schneider, M. I. Hariz,
W. Vandenberghe, B. Nuttin, and P. Brown, “High-frequency stimulation of the subthalamic nucleus suppresses oscillatory beta activity in
patients with Parkinson’s disease in parallel with improvement in motor
performance,” J Neurosci, vol. 28, pp. 6165–6173, Jun 2008.
[4] H. Bronte-Stewart, C. Barberini, M. M. Koop, B. C. Hill, J. M.
Henderson, and B. Wingeier, “The STN beta-band profile in Parkinson’s
disease is stationary and shows prolonged attenuation after deep brain
stimulation,” Exp Neurol, vol. 215, pp. 20–28, Jan 2009.
[5] J. Rubin and D. Terman, “High frequency stimulation of the subthalamic
nucleus eliminates pathological thalamic rhythmicity in a computational
model,” J Comput Neurosci, vol. 16, pp. 211–235, May 2004.
[6] P. J. Hahn and C. C. McIntyre, “Modeling shifts in the rate and pattern
of subthalamopallidal network activity during deep brain stimulation,”
J Comput Neurosci, vol. 28, pp. 425–441, Jun 2010.
[7] G. Kang and M. M. Lowery, “Interaction of Oscillations, and Their
Suppression via Deep Brain Stimulation, in a Model of the CorticoBasal Ganglia Network,” IEEE T Neur Sys Reh, vol. 21, pp. 244–253,
Mar 2013.
[8] C. R. Butson and C. C. McIntyre, “Tissue and electrode capacitance
reduce neural activation volumes during deep brain stimulation,” Clin
Neurophysiol, vol. 116, pp. 2490–2500, Oct 2005.
[9] P. F. Grant and M. M. Lowery, “Effect of Dispersive Conductivity and
Permittivity in Volume Conductor Models of Deep Brain Stimulation,”
IEEE T Bio-med Eng, vol. 57, pp. 2386–2393, Oct 2010.
[10] C. Schmidt and U. van Rienen, “Single Frequency Approximation of
Volume Conductor Models for Deep Brain Stimulation Using Equivalent
Circuits,” Proceedings of the International Conference on Bio-inspired
Systems and Signal Processing, pp. 38–47, Feb 2014.
[11] C. Gabriel, A. Peyman, and E. H. Grant, “Electrical conductivity of
tissue at frequencies below 1 MHz,” Phys Med Biol, vol. 54, pp. 4863–
4878, Aug 2009.
[12] C. Schmidt, P. Grant, M. Lowery, and U. van Rienen, “Influence
of Uncertainties in the Material Properties of Brain Tissue on the
Probabilistic Volume of Tissue Activated,” IEEE Trans Biomed Eng,
vol. 60, pp. 1378–1387, May 2013.
[13] G. Kang and M. M. Lowery, “Effects of antidromic and orthodromic
activation of STN afferent axons during DBS in Parkinson’s disease:
a simulation study,” Front Comput Neurosci, vol. 57. article 32, Mar
2014.
[14] C. Schmidt, E. Dunn, M. Lowery, and U. van Rienen, “Simulating
the Therapeutic Effects of Deep Brain Stimulation in Rodents Using
a Cortico-Basal Ganglia Network and Volume Conductor Model,” Proceedings of the 7th International IEEE EMBS Conference on Neural
Engineering (NER), pp. 852–855, Jun 2015.
[15] T. Otsuka, T. Abe, T. Tsukagawa, and W. J. Song, “Conductance-based
model of the voltage-dependent generation of a plateau potential in
subthalamic neurons,” J Neurophysiol, vol. 92, pp. 255–264, Jul 2004.
[16] E. M. Izhikevich, “Simple Model of Spiking Neurons,” IEEE T Neural
Networks, vol. 14, pp. 1569–1572, 2003.
[17] M. Pospischil, M. Toledo-Rodriguez, C. Monier, Z. Piwkowska, T. Bal,
Y. Fregnac, H. Markram, and A. Destexhe, “Minimal Hodgkin-Huxley
type models for different classes of cortical and thalamic neurons,” Biol
Cybern, vol. 99, pp. 427–441, Nov 2008.
[18] A. J. Foust, Y. Yu, M. Popovic, D. Zecevic, and D. A. McCormick,
“Somatic membrane potential and Kv1 channels control spike repolarization in cortical axon collaterals and presynaptic boutons,” J Neurosci,
vol. 26, pp. 15 490–15 498, Oct 2011.
[19] G. M. G. Shepherd, “Corticostriatal connectivity and its role in disease,”
Nat Rev Neurosci, vol. 14, pp. 1030–1041, Jul 2014.
[20] R. L. Albin, A. B. Young, and J. B. Penney, “The functional anatomy
of basal ganglia disorders,” Trends Neurosci, vol. 12, pp. 366–375, Oct
1989.
[21] Q. Li, Y. Ke, D. C. Chan, Z. M. Qian, K. K. Yung, H. Ko, G. W.
Arbuthnott, and W. H. Yung, “Therapeutic deep brain stimulation in
Parkinsonian rats directly influences motor cortex,” Neuron, vol. 76, pp.
1030–1041, Dec 2012.
[22] E. M. Dunn and M. M. Lowery, “Simulation of PID control schemes
for closed-loop deep brain stimulation,” 6th International IEEE/EMBS
Conference on Neural Engineering (NER), pp. 1182–1185, Nov 2013.
[23] M. L. Hines and N. T. Carnevale, “The NEURON simulation environment,” Neural Comput, vol. 9, pp. 1179–1209, Aug 1997.
[24] P. Brown, “Cortical drives to human muscle: the Piper and related
rhythms,” Prog Neurobiol, vol. 60, pp. 97–108, Jan 2000.
[25] Y. Ma, P. R. Hof, S. C. Grant, S. J. Blackband, R. Bennet, L. Slatest,
M. D. McGuigan, and H. Benveniste, “A three-dimensional digital atlas
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
database of the adult C57BL/6J mouse brian by magnetic resonance
microscopy,” Neuroscience, vol. 135, pp. 1203–1215, Sep 2005.
G. G. de Haas, “Towards the neurobiology of compulsive rituals,” PhD
Thesis. BOXPress Oisterwijk, p. 81, 2012.
J. Gimsa, B. Habel, U. Schreiber, U. van Rienen, U. Strauss, and
U. Gimsa, “Choosing electrodes for deep brain stimulation experiments
- electrochemical considerations,” J Neurosci Meth, vol. 142, pp. 251–
265, Mar 2005.
N. Yousif and X. Liu, “Investigating the depth electrode-brain interface
in deep brain stimulation using finite element models with graded
complexity in structure and solution,” J Neurosci Meth, vol. 184, no. 1,
pp. 142–151, Oct 2009.
S. Gabriel, R. W. Lau, and C. Gabriel, “The dielectric properties of
biological tissues: III Parametric models for the dielectric spectrum of
tissues,” Phys Med Biol, vol. 41, pp. 2271–2293, Nov 1996.
V. Gradinaru, M. Mogri, K. R. Thompson, J. M. Henderson, and
K. Deisseroth, “Optical deconstruction of parkinsonian neural circuitry,”
Science, vol. 17, pp. 354–359, Apr 2009.
M. Åström, E. Diczfalusy, H. Martens, and K. Wårdell, “Relationship
between Neural Activation and Electric Field Distribution during Deep
Brain Stimulation,” IEEE Trans Bio Med Eng, vol. 62, pp. 664–672, Feb
2015.
D. Xiu, Numerical Methods for Stochastic Computations: A Spectral
Method Approach. Princeton University Press, 2010.
P. G. Constantine, M. S. Eldred, and E. T. Phipps, “Sparse pseudospectral
approximation method,” Comput Method Appl M, vol. 229232, pp. 1 –
12, July 2012.
F. Nobile, R. Tempone, and C. G. Webster, “A sparse grid stochastic
collocation method for partial differential equations with random input
data,” SIAM J Numer Anal, vol. 46, pp. 2309–2345, May 2008.
I. M. Sobol, “Global sensitivity indices for nonlinear mathematical
models and their Monte Carlo estimates,” Ma Comput Sci Eng, vol. 55,
pp. 271–280, Feb 2001.
C. Schmidt, S. Wagner, M. Burger, U. van Rienen, and C. H. Wolters,
“Impact of Uncertain Head Tissue Conductivity in the Optimization of
Trancranial Direct Current Stimulation for an Auditory Target,” J Neural
Eng, vol. 12. art. no. 046028, Aug 2015.
C. Schmidt, U. Zimmermann, and U. van Rienen, “Modeling of an Optimized Electro-Stimulative Hip Revision System Under Consideration of
Uncertainty in the Conductivity of Bone Tissue,” IEEE J Biomed Heal
Inf, vol. 19, pp. 1321–1330, Jun 2015.
A. A. Kühn, A. Tsui, T. Aziz, N. Ray, C. Brücke, A. Kupsch,
G. H. Schneider, and P. Brown, “Pathological synchronisation in the
subthalamic nucleus of patients with Parkinson’s disease relates to both
bradykinesia and rigidity,” Exp Neurol, vol. 215, pp. 380–387, Feb 2009.
C. Hammond, H. Bergman, and P. Brown, “Pathological synchronization in Parkinson’s disease: networks, models and treatments,” Trends
Neurosci, vol. 30, pp. 357–364, Jul 2007.
R. G. So, G. C. McConnell, A. T. August, and W. M. Grill, “Characterizing Effects of Subthalamic Nucleus Deep Brain Stimulation
on Methamphetamine-Induced Circling Behavior in Hemi-Parkinsonian
Rats,” IEEE Trans Neural Syst Rehabil Eng, vol. 20, pp. 626–635, Sep
2012.
A. L. Spieles-Engemann, T. J. Collier, and C. E. Sortwell, “A functionally relevant and long-term model of deep brain stimulation of the rat
subthalamic nucleus: Advantages and considerations,” Eur J Neurosci,
vol. 32, pp. 1092–1099, Oct 2010.
V. Gardinaru, M. Mogri, K. R. Thompson, J. M. Henderson, and
K. Deisseroth, “Optical deconstruction of parkinsonian neural circuitry,”
Science, vol. 324, pp. 354–359, Apr 2009.
W. M. Grill, A. N. Snyder, and S. Miocinovic, “Deep brain stimulation
creates an informational lesion of the stimulated nucleus,” Neuroreport,
vol. 15, pp. 1137–1140, May 2004.
C. C. McIntyre, W. M. Grill, D. L. Sherman, and N. V. Thakor, “Cellular
effects of deep brain stimulation: model-based analysis of activation and
inhibition,” J Neurophysiol, vol. 91, pp. 1457–1469, Apr 2004.
B. N. Jávor-Duray, M. Vinck, M. van der Roest, A. B. Mulder, C. J.
Stam, H. W. Berendse, and P. Voorn, “Early-onset cortico-cortical
synchronization in the hemiparkinsonian rat model,” J Neurophysiol,
vol. 113, pp. 925–936, Feb 2015.
C. Hammond, R. Ammari, B. Bioulac, and L. Garcia, “Latest view on
the mechanism of action of deep brain stimulation,” Mov Disord, vol. 23,
pp. 2111–2121, Nov 2008.
J. M. Deniau, B. Degos, C. Bosch, and N. Maurice, “Deep brain stimulation mechanisms: beyond the concept of local functional inhibition,”
Eur J Neurosci, vol. 32, pp. 1080–1091, Oct 2010.
10
[48] K. Kumaravelu, D. T. Brocker, and W. M. Grill, “A biophysical model
of the cortex-basal ganglia-thalamus network in the 6-OHDA lesioned
rat model of Parkinson’s disease,” J Comput Neurosci, vol. 40, pp. 207–
229, Apr 2016.
[49] D. Terman, J. E. Rubin, A. C. Yew, and J. Wilson, “Activity Patterns
in a Model for the Subthalamopallidal Network of the Basal Ganglia,”
J Neurosci, vol. 22, pp. 2963–2976, Apr 2002.
[50] S. F. Lempka, M. D. Johnson, S. Miocinovic, J. L. Vitek, and C. C.
McIntyre, “Current-controlled deep brain stimulation reduces in vivo
voltage fluctuations observed during voltage-controlled stimulation,”
Clin Neurophysiol, vol. 12, pp. 2128–2133, Dec 2010.
Christian Schmidt (M’12) received his PhD from
the Faculty of Computer Science and Electrical
Engineering at the University of Rostock in 2013.
His field of research comprises computational engineering for several applications ranging from neuroscience and electro-stimulative implants to the
design of superconducting radio frequency cavities
as used in accelerator facilities. His research interests
include also the application and development of
stochastical methods for uncertainty quantification.
Eleanor Dunn received her B.S. degree in biomedical engineering from Rensselaer Polytechnic Institute, Troy, New York, United States in 2010.
She is currently working toward a Ph.D degree
in biomedical engineering from University College
Dublin, Dublin, Ireland. Her research interests include computational modelling of neural circuits and
deep brain stimulation.
Madeleine Lowery (M’00) is currently an Associate
Professor in the School of Electrical and Electronic
Engineering, University College Dublin. Her research is focused on using engineering methods to
understand the human nervous system as it relates to
movement in health and disease, and to design therapies and technologies to improve impaired motor
function. Her research interests include electromyography, myoelectric prosthetic control, bioelectromagnetics, electrical stimulation, deep brain stimulation
and neural control of movement.
Ursula van Rienen (M’01) received the venia legendi for the fields ”Electromagnetic Field Theory”
and ”Scientific Computing” at the Technische Universität Darmstadt. Since 1997, she holds the chair
in ”Electromagnetic Field Theory” at the University
of Rostock. Her research work is focused on computational electromagnetics with various applications,
ranging from biomedical engineering to accelerator
physics.
| 5cs.CE
|
An Oracle Inequality for Quasi-Bayesian
Non-Negative Matrix Factorization
arXiv:1601.01345v3 [stat.ML] 24 Aug 2016
Pierre Alquier∗& Benjamin Guedj†
August 25, 2016
Abstract
The aim of this paper is to provide some theoretical understanding of
Bayesian non-negative matrix factorization methods. We derive an oracle inequality for a quasi-Bayesian estimator. This result holds for a
very general class of prior distributions and shows how the prior affects the rate of convergence. We illustrate our theoretical results with
a short numerical study along with a discussion on existing implementations.
1
Introduction
Non-negative matrix factorization (NMF) is a set of algorithms in highdimensional data analysis which aims at factorizing a large matrix M with
non-negative entries. If M is an m 1 × m 2 matrix, NMF consists in decomposing it as a product of two matrices of smaller dimensions: M ' UV T where
U is m 1 × K , V is m 2 × K , K ¿ m 1 ∧ m 2 and both U and V have non-negative
entries. Interpreting the columns M·, j of M as (non-negative) signals, NMF
amounts to decompose (exactly, or approximately) each signal as a combination of the “elementary” signals U·,1 , . . . ,U·,K :
M ·, j '
K
X
`=1
∗
V j,`U·,` .
(1)
CREST, ENSAE, Université Paris Saclay, [email protected]. This author gratefully acknowledges financial support from the research programme New Challenges for New
Data from LCL and GENES, hosted by the Fondation du Risque, from Labex ECODEC
(ANR - 11-LABEX-0047) and from Labex CEMPI (ANR-11-LABX-0007-01).
†
Modal project-team, Inria, [email protected].
1
Since the seminal paper from Lee and Seung (1999), NMF was successfully
applied to various fields such as image processing and face classification
(Guillamet and Vitria, 2002), separation of sources in audio and video processing (Ozerov and Févotte, 2010), collaborative filtering and recommender
systems on the Web (Koren et al., 2009), document clustering (Xu et al.,
2003; Shahnaz et al., 2006), medical image processing (Allen et al., 2014) or
topics extraction in texts (Paisley et al., 2015). In all these applications, it
has been pointed out that NMF provides a decomposition which is usually
interpretable. Donoho and Stodden (2003) have given a theoretical foundation to this interpretatibility by exhibiting conditions under which the decomposition M ' UV T is unique. However, let us stress that even when this
is not the case, the results provided by NMF are still sensibly interpreted by
practitioners.
Since a prior knowledge on the shape and/or magnitude of the signal is available in many settings, a Bayesian strategy seems appropriate. Bayesian
tools have extensively been used for (general) matrix factorization (Corander and Villani, 2004; Lim and Teh, 2007; Salakhutdinov and Mnih, 2008;
Lawrence and Urtasun, 2009; Zhou et al., 2010) and have been adapted for
the Bayesian NMF problem (Moussaoui et al., 2006; Cemgil, 2009; Févotte
et al., 2009; Schmidt et al., 2009; Tan and Févotte, 2009; Zhong and Girolami, 2009, among others).
The aim of this paper is to provide some theoretical analysis on the performance of Bayesian NMF. We propose a general Bayesian estimator for
which we derive an oracle inequality. The message of this theoretical bound
is that our procedure is able to adapt to the unknown rank of M . This result holds even for noisy observations, with no parametric assumption on
the noise. That is, the likelihood used to build the Bayesian estimator does
not have to be well-specified (it is usually referred to as a quasi-likelihood).
To this regard, our procedure may be called quasi-Bayesian. The use of
quasi-likelihoods in Bayesian estimation is advocated by Bissiri et al. (2013)
using decision-theoretic arguments. This methodology is also popular in machine learning, and various authors developed a theoretical framework to
analyze it (Shawe-Taylor and Williamson, 1997; McAllester, 1998; Catoni,
2003, 2004, 2007; Dalalyan and Tsybakov, 2008, this is known as the PACBayesian theory).
The paper is organized as follows. Notation for the Bayesian NMF framework and the definition of our quasi-Bayesian estimator are given in Section 2. The oracle inequality, which is our main contribution, is given in
Section 3 and its proof is postponed to Appendix A. We illustrate this theoretical result by a short numerical study. To that matter, we discuss two
2
implementations (MCMC and optimization) of the quasi-Bayesian estimator in Section 4. Section 5 contains our numerical experiments on synthetic
and real data. The main message of these experiments is that the choice of
the prior appear to have limited impact on the accuracy of the reconstruction of M . However, a fine tuning of the prior may be crucial if the goal is to
enforce shrinkage of many terms in (1) to 0.
2
Notation
For any p × q matrix A we denote by A i, j its ( i, j )-th entry, A i,· its i -th row
and A ·, j its j -th column. For any p × q matrix B we define
〈 A, B〉F = Tr( AB> ) =
p X
q
X
A i, j B i, j .
i =1 j =1
We define the Frobenius norm k A kF of A by k A k2F = 〈 A, A 〉F . Let A − i,· denote
the matrix A where the i -th column is removed. In the same way, for a vector
v ∈ R p , v− i ∈ R p−1 is the vector v with its i -th coordinate removed. Finally,
let Diag(v) denote the p × p diagonal matrix given by [Diag(v)] i,i = v i .
2.1
Model
The object of interest is an m 1 × m 2 target matrix M possibly polluted with
some noise E. So we actually observe
Y = M + E,
(2)
and we assume that E is random with E(E) = 0. The objective is to approximate M by a matrix UV T where U is m 1 × K , V is m 2 × K for some
K ¿ m 1 ∧ m 2 , and where U , V and M all have non-negative entries. Note
that, under (2), depending on the distribution of E, Y might have some negative entries (the non-negativity assumption is on M rather than on Y ). Our
theoretical analysis only requires the following assumption on E.
C1. The entries E i, j of E are i.i.d. with E(ε i, j ) = 0. With the notation m( x) =
E[ε i, j 1(ε i, j ≤ x) ] and F ( x) = P(ε i, j ≤ x), assume that there exists a non-negative
and bounded function g (put σ2 := k gk∞ ) such that
Z v
Z v
m( x)d x =
g( x)dF ( x).
u
u
3
The introduction of this rather involved condition is due to the technical
analysis of our estimator which is based on Theorem 2 in Appendix A. Theorem 2 has first been proved by Dalalyan and Tsybakov (2007) using Stein’s
formula with a Gaussian noise. However, Dalalyan and Tsybakov (2008)
have shown that C1 is actually sufficient to prove Theorem 2. For the sake
of understanding, note that C1 is fulfilled when the noise is Gaussian (ε i, j ∼
N(0, s2 ) and σ2 := s2 ) or uniform (ε i, j ∼ U[−b, b] and σ2 := b2 /2).
2.2
Prior
We are going to define a prior π(U, V ), where U is m 1 × K and V is m 2 × K ,
for a fixed K . Regarding the choice of K , we prove in Section 5 that our
quasi-Bayesian estimator is adaptive, in the sense that if K is chosen much
larger than the actual rank of M , the prior will put very little mass on many
columns of U and V , automatically shrinking them to 0. This seems to advocate for setting a large K prior to the analysis, say K = m 1 ∧ m 2 . However,
keep in mind that the algorithms discussed below have a computational cost
growing with K . Anyhow, the following theoretical analysis only requires
2 ≤ K ≤ m1 ∧ m2 .
With respect to the Lebesgue measure on R+ , let us fix a density f such that
Z ∞
S f :=
x2 f ( x)d x < +∞.
0
For any α, x > 0, let
g α ( x ) :=
1 ³x´
f
.
α α
We define the prior on U and V by
U i,` , Vi,` indep. ∼ g γ` (·)
where
γ` indep. ∼ h(·)
and h is a density on R+ . With the notation γ = (γ1 , . . . , γK ), define π by
Ã
!Ã
!
m1
m2
K Y
Y
Y
π(U, V , γ) =
g γ` (U i,` )
g γ` (V j,` ) h(γ` )
`=1 i =1
and
π(U, V ) =
j =1
Z
K
R+
π(U, V , γ)dγ.
4
(3)
The idea behind this prior is that under h, many γ` should be small and lead
to non-significant columns U·,` and V·,` . In order to do so, we must assume
that a non-negligible proportion of the mass of h is located around 0. This is
the meaning of the following assumption.
σ2
,
2S f K 2
C2. There exist constants 0 < α < 1 and β ≥ 0 such that for any 0 < ε ≤ p
ε
Z
0
h( x)d x ≥ αεβ .
Finally, the following assumption on f is required to prove our main result.
C3. There exist a non-increasing density fe w.r.t. Lebesgue measure on R+
and a constant C f > 0 such that for any x > 0
f ( x) ≥ C f fe( x).
As shown in Theorem 1, the heavier the tails of fe( x), the better the performance of Bayesian NMF.
Note that the general form of (3) encompasses as special cases almost all
the priors used in the papers mentioned in the introduction. We end this
subsection with classical examples of functions f and h.
1. Exponential prior f ( x) = exp(− x) with fe = f , C f = 1 and S f = 2. This
is the choice made by Schmidt et al. (2009). A generalization of the
exponential prior is the gamma prior used in Cemgil (2009).
2. Truncated Gaussian prior f ( x) ∝ exp(2ax − x2 ) with a ∈ R.
3. Heavy-tailed prior f ( x) ∝ (1+1x)ζ with ζ > 1.
¡ ¢
a
For h, the inverse gamma prior h( x) = Γb(a) xa1+1 exp − bx is classical in the
literature (see for example Salakhutdinov and Mnih, 2008; Alquier, 2013).
Alquier et al. (2014) chose h as the density of the Γ( m 1 + m 2 − 1/2, b) distribution for b > 0. Both lead to explicit conditional posteriors for γ.
2.3
Quasi-posterior and estimator
We define the quasi-likelihood as
£
¤
b (U, V ) = exp −λkY − UV > k2
L
F
for some fixed parameter λ > 0. Note that under the assumption that ε i, j ∼
N(0, 1/2λ), this would be the actual likelihood up to a multiplicative constant. As we pointed out, the use of quasi-likelihoods to define quasi-posteriors
5
is becoming rather popular in Bayesian statistics and machine learning literatures. Here, the Frobenius norm is to be seen as a fitting criterion rather
than as a ground truth. Note that other criterion were used in the literature: the Poisson likelihood (Lee and Seung, 1999), or the Itakura-Saito
divergence (Févotte et al., 2009).
Definition 1. We define the quasi-posterior as
1b
L(U, V )π(U, V , γ)
Z
£
¤
1
= exp −λkY − UV > k2F π(U, V , γ),
Z
ρbλ (U, V , γ) =
where
Z
Z :=
£
¤
exp −λkY − UV > k2F π(U, V , γ)d(U, V , γ)
is a normalization constant. The posterior mean will be denoted by
Z
c
Mλ = UV T ρbλ (U, V , γ)d(U, V , γ).
cλ . A short
Section 3 is devoted to the study the theoretical properties of M
discussion on the implementation will be provided in Section 4.
3
An oracle inequality
Most likely, the rank of M is unknown in practice. So, as recommended
above, we usually choose K much larger than the expected order for the
rank, with the hope that many columns of U and V will be shrinked to 0.
The following set of matrices is introduced to formalize this idea. For any
r ∈ {1, . . . , K }, let Mr be the set of pairs of matrices (U 0 , V 0 ) with non-negative
entries such that
0
0
0
0
U11
. . . U1r
0 ... 0
V11
. . . V1r
0 ... 0
.
..
.. . . .. 0 ..
..
.. . . ..
..
..
U 0 = ..
.
. . ,V = .
.
. . .
.
.
.
.
0
0
Um
. . . Um
0 ... 0
1r
11
0
0
Vm
. . . Vm
0 ... 0
2r
21
We also define Mr (L) as the set of matrices (U 0 , V 0 ) ∈ Mr such that, for any
( i, j, `), U i,0` , V j,0` ≤ L.
We are now in a position to state our main theorem, in the form of the following sharp oracle inequality.
6
Theorem 1. Fix λ = 4σ1 2 . Under assumptions C1, C2 and C3,
(
¢
cλ − M k2 ≤ inf
E kM
F
¡
inf
1≤ r ≤K (U 0 ,V 0 )∈Mr
kU 0 V 0> − M k2F
p
¡ 0
¢2
0
2( m 1 ∨ m 2 ) kU kF + kV kF + σK r
+ 8σ2 ( m 1 ∨ m 2 ) r log
r
σC f
Ã
!
!
Ã
X
X
1
1
log
+ 4σ2
+ 4σ2
log
p
p
0
e
σ
)
f
(
U
+
fe(V j0` + σ)
1≤ j ≤ m 2
1≤ i ≤ m 1
i`
s
1≤`≤ r
1≤`≤ r
Ã
+ 4σ2 βK log
2S f
p
p
¡
¢2 !
m 1 m 2 kU 0 kF + kV 0 kF + σK r
rσ
)
µ ¶
1
+ 8σ2 r + 4σ2 K log
+ 4σ2 log(4) .
α
We remind the reader that the proof is given in Appendix A. The main
cλ is as close to M as would be an estimator
message of the theorem is that M
cλ is adaptive to r ),
designed with the actual knowledge of its rank (i.e., M
up to remainder terms. These terms might be difficult to read. In order to
explicit the rate of convergence, we now provide a weaker version, where we
cλ to the best factorization in Mr (L).
only compare M
Corollary 1. Fix λ = 4σ1 2 . Under assumptions C1, C2 and C3,
½
inf
kU 0 V 0> − M k2F
1≤ r ≤K (U 0 ,V 0 )∈Mr
!
Ã
2
3
2(
L
+
σ
)(
m
m
)
1
2
+ 8σ2 ( m 1 ∨ m 2 ) r log
p
σC f f˜(L + σ)
Ã
!
µ ¶
¾
2
3
2
S
(
L
+
σ
)(
m
m
)
1
1
2
f
2
2
2
2
+ 4σ βK log
+ 8σ r + 4σ K log
+ 4σ log(4) .
σ
α
¢
cλ − M k2 ≤ inf
E kM
F
¡
First, note that when L2 = O(σ), up to log terms, the magnitude of the error
bound is
σ2 ( m 1 ∨ m 2 ) r,
which is roughly the variance multiplied by the number of parameters to be
estimated in any (U 0 , V 0 ) ∈ Mr (L). Alternatively, when M ∈ Mr (L) only for
huge L, the log term in
µ
¶
L2 + σ
2
8σ ( m 1 ∨ m 2 ) r log
p
σ f˜(L + σ)
7
becomes significant. Indeed, in the case of the truncated Gaussian prior
f ( x) ∝ exp(2ax − x2 ), the previous quantity is in
8σ2 ( m 1 ∨ m 2 ) rL2
which is terrible for large L. On the contrary, with the heavy-tailed prior
f ( x) ∝ (1 + x)−ζ (as in Dalalyan and Tsybakov, 2008), the leading term is
8σ2 ( m 1 ∨ m 2 ) r (ζ + 2) log(L)
which is way more satisfactory. Still, this prior has not received much attention from practitioners, since its implementation seems less straightforward.
4
Algorithms for Bayesian NMF
The method of choice for computing Bayesian estimators for complex models is Monte-Carlo Markov Chain (MCMC). In the case of Bayesian matrix
factorization, the Gibbs sampler was considered in the literature: see for
example Salakhutdinov and Mnih (2008), Alquier et al. (2014) for the general case and Moussaoui et al. (2006), Schmidt et al. (2009) and Zhong and
Girolami (2009) for NMF.
However, it is well known by practitioners that the computational cost of
MCMC-based methods becomes prohibitive in very high dimensional models. Indeed, the optimization algorithms used in non-Bayesian NMF are
much faster to converge in practice. Many of these algorithms share the
characteristic to minimize iteratively the criterion kY − UV > k2F with respect
to U , then V , an approach known as block coordinate descent (Bertsekas,
1999). This minimization may be achieved by the original multiplicative
algorithm Lee and Seung (1999, 2001) or projected gradient descent (Lin,
2007; Guan et al., 2012). This approach is studied in full generality in Xu
and Yin (2013). Other methods include second order schemes (Kim et al.,
2008), linear progamming (Bittorf et al., 2012) or ADMM (alternative direction method of multipliers Boyd et al., 2011; Xu et al., 2012).
In order to enjoy the desirable computational properties of the aforementioned algorithms in Bayesian statistics, some authors proposed to use optimization tools to compute an approximation of the posterior. This method
is known as Variational Bayes (Jordan et al., 1999; MacKay, 2002; Bishop,
2006). The theoretical properties of this approximation are studied in Alquier
et al. (2015). It was used for Bayesian matrix factorization (Lim and Teh,
2007; Alquier et al., 2014) and more recently in Bayesian NMF (Paisley
et al., 2015).
8
Finally, another option is to use optimization algorithms to compute the
mode of the posterior, also known as maximum a posteriori (MAP). While
this estimator is in general different from the posterior mean, it still leads
to optimal rates of convergence in some complex models (see Abramovich
and Lahav, 2015, in the context of additive regression).
Still, we would like to point out that the function to be maximized is not
concave with respect to (U, V ), which makes the optimization problem hard
to solve. In particular, note that no rates of convergence for optimization
algorithms are known for NMF in general. We do not intend to solve this
problem here as it goes far beyond the scope of this paper.
In the sequel, we give a pseudo-code for the Gibbs sampler, and block coordinate descent to compute the MAP.
4.1
General form of the conditional posteriors for the
Gibbs sampler
The Gibbs sampler (described in its general form in Bishop, 2006, for example), is given by Algorithm 1.
Algorithm 1 Gibbs sampler.
Input Y , λ.
Initialization U (0) , V (0) , γ(0) .
For k = 1, . . . , N :
For i = 1, . . . , m 1 : draw U i,(k)
∼ ρbλ (U i,· |V (k−1) , γ(k−1) , Y ).
·
For j = 1, . . . , m 2 : draw V j,(k)
∼ ρbλ (V j,· |U (k) , γ(k−1) , Y ).
·
For ` = 1, . . . , K : draw γ(k)
∼ ρbλ (γ` |U (k) , V (k) , Y ).
`
We now explicit the conditional posteriors. We first remind the quasi-posterior
formula
1b
L(U, V )π(U, V , γ)
Z
"
#
m1
m2
K
Y
Y
¡
¢Y
1
> 2
= exp −λkY − UV kF
h(γ` )
g γ` (U i` )
g γ` (V j` ) .
Z
`=1
i =1
j =1
ρbλ (U, V , γ) =
9
As a function of U i,· ,
K
¡
¢Y
b (U, V )π(U, V , γ) ∝ exp −λkY − UV > k2
L
g γ` (U i,` )
F
`=1
K
¡
¢Y
> 2
g γ` (U i,` ).
∝ exp −λkYi,· − U i,· V k
`=1
b i = Yi,· V (V T V )−1 and ΣU = (V T V )−1 . This yields
Let U
ρbλ (U i,· |U− i,· , V , γ, Y ) = ρbλ (U i,· |V , γ, Y )
³
´Y
K
b i − U i,· )(ΣU )−1 (U
b i − U i,· )T
∝ exp −λ(U
g γ` (U i,` ).
`=1
bj = Y T U (U T U )−1 and ΣV = (U T U )−1 and we
In the same way, we define V
·, j
have
ρbλ (V j,· |V− j,· ,U, γ, Y ) = ρbλ (V j,· |U, γ, Y )
³
´Y
K
bj − V j,· )(ΣV )−1 (V
bj − V j,· )T
g γ` (V j,` ).
∝ exp −λ(V
`=1
Finally
ρbλ (γ` |U, V , γ−` , Y ) = ρbλ (γ` |U, V , Y )
m1
m2
Y
Y
∝ h(γ` )
g γ` (U i` )
g γ` (V j` ).
i =1
j =1
In all generality, sampling from ρbλ (U i,· |V , γ, Y ) might require considerable
effort. In such cases, a Metropolis-within-Gibbs approach is often the best
choice, sadly at the cost of quite substantial computational power. In the
high-dimensional context of NMF, this choice appeared unrealistic to us.
However, it appears that when the prior f is exponential or truncated Gaussian, sampling from ρbλ (U i,· |V , γ, Y ) becomes straightforward. The detailed
algorithms are provided in Appendix B.
4.2
Optimization through block coordinate descent
In this section, we discuss an algorithm for the implementation of the MAP
estimator
eλ , V
eλ , γ
eλ ) = arg max ρbλ (U, V , γ)
(U
U,V ,γ
10
(
λkY − UV > k2F − log π(U, V , γ)
= arg min
U,V ,γ
−
m1 X
K
X
i =1 `=1
)
m2 X
K
K
¡
¢
¡
¢ X
¢ X
log h(γ` ) .
log g γ` (V j,` ) −
log g γ` (U i,` ) −
¡
`=1
j =1 `=1
The block coordinate descent approach is used in practice with reasonnable
results. This algorithm seems to be relatively standard in NMF as discussed
above and is described in Algorithm 2.
Algorithm 2 Pseudo-algorithm for block coordinate descent.
Input Y , λ.
Initialization U (0) , V (0) , γ(0) .
While not converged, k := k + 1:
(
U (k) := arg min λkY − U (V (k−1) )> k2F −
U
(
V (k) := arg min λkY − U (k) V > k2F −
V
(k)
γ
:= arg min
γ
K
X
`=1
m1 X
K
X
i =1 `=1
m2 X
K
X
j =1 `=1
)
log[ g γ(k−1) (U i,` )]
`
)
log[ g γ(k−1) (V j,` )]
`
½ m1
¾
m2
X
X
(k)
(k)
− log[ g γ` (U i,` )] −
log[ g γ` (V j,` )] − log[ h(γ` )]
i =1
j =1
Note that when the functions f and h are not conveniently chosen, this optimization problem can be very cumbersome, if tractable at all. In the examples mentioned in Section 2, that is, when f is exponential or truncated
Gaussian, the optimization problems in U and V are quadratic problems
with a non-negativity constraint, that can be solved thanks to the various
approaches mentioned above. We derive explicit forms of Algorithm 2 using
a projected gradient algorithm when the priors are exponential or truncated
Gaussian in Appendix C.
5
Numerical experiments
Note that NMF has attracted a great deal of interest and the number of
available algorithms is massive. The objective of this section is not to convince the reader that Bayesian NMF would uniformly be the best possible
11
Figure 1: Results of the simulations with K = 5: MSEs and vectors γ obtained for each value of b.
b
MSE
100
0.000762
101
0.002750
102
0.001089
103
0.000712
104
0.000774
105
0.000711
106
0.002456
107
0.009929
108
0.146539
109
0.632924
γ
b
MSE
γ
method. Instead, our main objective is to illustrate the influence of the
choice of the prior hyperparameters. To do so, we performed a numerical
study of the performances of the Bayesian MAP estimator coupled with the
exponential prior with parameter 1 on (U, V ) and with the gamma prior
Γ( m 1 + m 1 − 1/2, b) on the coefficients γ j . We approximated this estimator
thanks to Algorithm 2.
To assess the impact of the hyperparameter b on the quality of the factorization, let us consider the following exponential grid: b ∈ {100 , 101 , . . . , 109 }.
cλ k2 . It is also worth
We define the mean square error: MSE = m11m2 k M − M
F
mentioning the possibility to infer an optimal size for the dictionary U . In
theory, note that our estimator will report a dictionary U with maximal size
K . However in practice, we can expect that many γ` s will be shrunk to 0
so that thresholding their values will not affect the performance of the reconstruction, thus setting many U·,` s to 0. As highlighted below, this will
typically occur for large values of b.
In a first experiment, we simulate Y = M + E = UV T + E with m 1 = m 2 = 100
and U , V two 100 × 2 matrices with entries drawn independently from a
uniform U([0, 3]) distribution. In a first time choose K = 5. We simulate the
entries E i, j of E independenly from a Gaussian N(0, σ2 ) with σ2 = 0.01. We
report in Figure 1 the MSEs and vectors γ obtained for each value of b.
Clearly, for any value of b between 1 and 106 the MSE’s are of similar magnitude (between 0.0007 and 0.002). While this is satisfactory, our method fails
12
Figure 2: Results of the simulations = with K = 20: MSEs and vectors γ
obtained for each value of b.
b
MSE
100
0.013176
101
0.010170
102
0.009517
103
0.009934
104
0.006730
105
0.006475
106
0.001572
107
0.007313
108
0.607693
109
0.640797
γ
b
MSE
γ
to identify the minimal support of γ. For larger b, a minimal support of γ
is identified. This remark is in accordance with the fact that model identification and estimation objectives are often incompatible in high-dimensional
statistics (as pointed out by Yang, 2005, among others). The overall good
results are obviously related to the rather strong prior knowledge K = 5:
we repeated the same design with the more relaxed assumption K = 20 (reported in Figure 2).
The optimal value in terms of MSE seems to lie around b = 106 . About the
estimation of the size of the dictionary, the vector γ is sparse for b = 106
but we only identify the true sparsity for b = 107 . The rank identification
problem seems more sensitive to a proper tuning of b than the estimation
problem, i.e., minimization of the MSE.
These findings are confirmed by an additional experiment on the famous
USPS database (Le Cun et al., 1990). Each image is a 16×16 pixels and
thus can be represented by a vector in R256 . We store all the images of zeros
and ones in a 2199×256 matrix Y . We then run our algorithm with K = 6.
We still choose f as the exponential prior and h as the gamma prior with
different values of b. We plot the images corresponding to all the U·,` for
` ∈ {1, . . . , 6} in Figure 3. For small values of b, we identify various shapes
of zeros. When b increases, we shrink the dictionary, and this leads to a
smaller set of shapes of zeros. A huge value for b leads to a uniformly null
decomposition.
13
Figure 3: Experiments on the USPS dataset with f as the exponential prior
and h as the gamma prior with parameter b. We represent the images U·,`
obtained for various values of b. A uniformly grey image means that U·,` = 0.
b
U
b
103
109
108
1010
U
Acknowledgements
The authors are grateful to Jialia Mei and Yohann de Castro (Université
Paris-Sud) for insightful discussions and for providing many references on
NMF.
References
F. Abramovich and T. Lahav. Sparse additive regression on a regular lattice.
Journal of the Royal Statistical Society: Series B (Statistical Methodology),
77(2):443–459, 2015. 9
G. I. Allen, L. Grosenick, and J. Taylor. A generalized least-square matrix
decomposition. Journal of the American Statistical Association, 109(505):
145–159, 2014. doi: 10.1080/01621459.2013.852978. 2
P. Alquier. Bayesian methods for low-rank matrix estimation: short survey
and theoretical study. In Algorithmic Learning Theory 2013, pages 309–
323. Springer, 2013. 5, 19
P. Alquier, V. Cottet, N. Chopin, and J. Rousseau. Bayesian matrix completion: prior specification. Preprint arXiv:1406.1440, 2014. 5, 8, 26
14
P. Alquier, J. Ridgway, and N. Chopin. On the properties of variational approximations of Gibbs posteriors. Preprint arXiv:1506.04091, 2015. 8
D. P. Bertsekas. Nonlinear programming. Athena Scientific, 1999. 8
C. M. Bishop. Pattern Recognition and Machine Learning, chapter 10.
Springer, 2006. 8, 9
P. Bissiri, C. Holmes, and S. Walker. A general framework for updating belief
distributions. Preprint arXiv:1306.6430, to appear in the Journal of the
Royal Statistical Society Series B, 2013. 2
V. Bittorf, B. Recht, C. Re, and J. Tropp. Factoring nonnegative matrices with
linear programs. In Advances in Neural Information Processing Systems,
pages 1214–1222, 2012. 8
S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein. Distributed optimization and statistical learning via the alternating direction method of
multipliers. Foundations and Trends in Machine Learning, 3(1):1–122,
2011. 8
O. Catoni. A PAC-Bayesian approach to adaptive classification. Preprint
Laboratoire de Probabilités et Modèles Aléatoires, PMA-840, 2003. 2, 19
O. Catoni. Statistical Learning Theory and Stochastic Optimization. SaintFlour Summer School on Probability Theory 2001 (Jean Picard ed.), Lecture Notes in Mathematics. Springer, 2004. 2, 19
O. Catoni. PAC-Bayesian supervised classification: the thermodynamics of
statistical learning. Institute of Mathematical Statistics Lecture Notes—
Monograph Series, 56. Institute of Mathematical Statistics, Beachwood,
OH, 2007. 2, 19
A. T. Cemgil. Bayesian inference for nonnegative matrix factorisation models. Computational Intelligence and Neuroscience, 2009, 2009. 2, 5
J. Corander and M. Villani. Bayesian assessment of dimensionality in reduced rank regression. Statistica Neerlandica, 58:255–270, 2004. 2
A. Dalalyan and A. B. Tsybakov. Aggregation by exponential weighting,
sharp PAC-Bayesian bounds and sparsity. Machine Learning, 72(1-2):39–
61, 2008. 2, 4, 8, 19, 20
15
A. S. Dalalyan and A. B. Tsybakov. Aggregation by exponential weighting and sharp oracle inequalities. In N. Bshouty and C. Gentile, editors, Learning Theory, volume 4539 of Lecture Notes in Computer Science,
pages 97–111. Springer Berlin Heidelberg, 2007. 4
D. Donoho and V. Stodden. When does non-negative matrix factorization give
a correct decomposition into parts? In Advances in neural information
processing systems, 2003. 2
C. Févotte, N. Bertin, and J.-L. Durrieu. Nonnegative matrix factorization
with the Itakura-Saito divergence: With application to music analysis.
Neural computation, 21(3):793–830, 2009. 2, 6
I. Giulini. PAC-Bayesian bounds for Principal Component Analysis in
Hilbert spaces. Preprint arXiv:1511.06263, 2015. 19
N. Guan, D. Tao, Z. Luo, and B. Yuan. NeNMF: an optimal gradient method
for nonnegative matrix factorization. IEEE Transactions on Signal Processing, 60(6):2882–2898, 2012. 8
B. Guedj and P. Alquier. PAC-Bayesian Estimation and Prevision in Sparse
Additive Models. Electronic Journal of Statistics, 7:264–291, 2013. 19
B. Guedj and S. Robbiano. PAC-Bayesian High Dimensional Bipartite Ranking. Preprint arXiv:1511.02729, 2015. 19
D. Guillamet and J. Vitria. Classifying faces with nonnegative matrix factorization. In Proc. 5th Catalan conference for artificial intelligence, pages
24–31, 2002. 2
M. I. Jordan, Z. Ghahramani, T. S. Jaakkola, and L. K. Saul. An introduction
to variational methods for graphical models. Machine Learning, 37:183–
233, 1999. 8
D. Kim, S. Sra, and I. S. Dhillon. Fast projection-based methods for the least
squares nonnegative matrix approximation problem. Statistical Analysis
and Data Mining, 1(1):38–51, 2008. 8
Y. Koren, R. Bell, and C. Volinsky. Matrix factorization techniques for recommender systems. Computer, 42(8):30–37, 2009. 2
N. D. Lawrence and R. Urtasun. Non-linear matrix factorization with Gaussian processes. In Proceedings of the 26th Annual International Conference on Machine Learning, pages 601–608. ACM, 2009. 2, 26
16
Y. Le Cun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard,
and L. D. Jackel. Handwritten digit recognition with a back-propagation
network. In Advances in neural information processing systems. Citeseer,
1990. 13
D. D. Lee and H. S. Seung. Learning the parts of objects by non-negative
matrix factorization. Nature, 401(6755):788–791, 1999. 2, 6, 8
D. D. Lee and H. S. Seung. Algorithms for non-negative matrix factorization. In Advances in neural information processing systems, pages 556–
562, 2001. 8
G. Leung and A. R. Barron. Information theory and mixing least-squares
regressions. IEEE Transactions on Information Theory, 52(8):3396–3410,
2006. 19
L. Li, B. Guedj, and S. Loustau. PAC-Bayesian online clustering. arXiv
preprint arXiv:1602.00522, 2016. 19
Y. J. Lim and Y. W. Teh. Variational Bayesian approach to movie rating
prediction. In Proceedings of KDD Cup and Workshop, volume 7, pages
15–21, 2007. 2, 8, 26
C.-J. Lin. Projected gradient methods for nonnegative matrix factorization.
Neural computation, 19(10):2756–2779, 2007. 8
D. J. C. MacKay. Information Theory, Inference and Learning Algorithms.
Cambridge University Press, 2002. 8
T. T. Mai and P. Alquier. A Bayesian approach for matrix completion: optimal rates under general sampling distributions. Electronic Journal of
Statistics, 9:823–841, 2015. 19, 26
D. McAllester. Some PAC-Bayesian theorems. In Proceedings of the Eleventh
Annual Conference on Computational Learning Theory, pages 230–234,
New York, 1998. ACM. 2, 19
S. Moussaoui, D. Brie, A. Mohammad-Djafari, and C. Carteret. Separation of
non-negative mixture of non-negative sources using a Bayesian approach
and MCMC sampling. IEEE Transactions on Signal Processing, 54(11):
4133–4145, 2006. 2, 8
A. Ozerov and C. Févotte. Multichannel nonnegative matrix factorization in
convolutive mixtures for audio source separation. IEEE Transactions on
Audio, Speech, and Language Processing, 18(3):550–563, 2010. 2
17
J. Paisley, D. Blei, and M. I. Jordan. Bayesian nonnegative matrix factorization with stochastic variational inference, volume Handbook of Mixed
Membership Models and Their Applications, chapter 11. Chapman and
Hall/CRC, 2015. 2, 8
R. Salakhutdinov and A. Mnih. Bayesian probabilistic matrix factorization
using Markov chain Monte Carlo. In Proceedings of the 25th international
conference on Machine learning, pages 880–887. ACM, 2008. 2, 5, 8, 26
M. N. Schmidt, O. Winther, and L. K. Hansen. Bayesian non-negative matrix
factorization. In Independent Component Analysis and Signal Separation,
pages 540–547. Springer, 2009. 2, 5, 8
F. Shahnaz, M. W. Berry, V. P. Pauca, and R. J. Plemmons. Document clustering using nonnegative matrix factorization. Information Processing &
Management, 42(2):373–386, 2006. 2
J. Shawe-Taylor and R. Williamson. A PAC analysis of a Bayes estimator. In
Proceedings of the Tenth Annual Conference on Computational Learning
Theory, pages 2–9, New York, 1997. ACM. 2, 19
T. Suzuki. Convergence rate of Bayesian tensor estimatior and its minimax
optimality. In Proceedings of the 32nd International Conference on Machine Learning (Lille, 2015), pages 1273–1282, 2015. 19
V. Y. Tan and C. Févotte. Automatic relevance determination in nonnegative matrix factorization. In SPARS’09-Signal Processing with Adaptive
Sparse Structured Representations, 2009. 2
S. Wilhelm. tmvtnorm: Truncated Multivariate Normal and Student t Distribution, 2015. URL http://CRAN.R-project.org/package=tmvtnorm.
R package version 1.4-10. 26
W. Xu, X. Liu, and Y. Gong. Document clustering based on non-negative
matrix factorization. In Proceedings of the 26th annual international ACM
SIGIR conference on research and development in information retrieval,
pages 267–273. ACM, 2003. 2
Y. Xu and W. Yin. A block coordinate descent method for regularized multiconvex optimization with applications to nonnegative tensor factorization
and completion. SIAM Journal on imaging sciences, 6(3):1758–1789, 2013.
8
18
Y. Xu, W. Yin, Z. Wen, and Y. Zhang. An alternating direction algorithm for
matrix completion with nonnegative factors. Frontiers of Mathematics in
China, 7(2):365–384, 2012. 8
Y. Yang. Can the strengths of AIC and BIC be shared? A conflict between
model indentification and regression estimation. Biometrika, 92(4):937–
950, 2005. 13
M. Zhong and M. Girolami. Reversible jump MCMC for non-negative matrix
factorization. In International Conference on Artificial Intelligence and
Statistics, pages 663–670, 2009. 2, 8
M. Zhou, C. Wang, M. Chen, J. Paisley, D. Dunson, and L. Carin. Nonparametric Bayesian Matrix Completion. In Proc. IEEE SAM, 2010. 2, 26
A
Proofs
This appendix contains the proof to the main theoretical claim of the paper
(Theorem 1).
A.1
A PAC-Bayesian bound from Dalalyan and Tsybakov
(2008)
The analysis of quasi-Bayesian estimators with PAC bounds started with
Shawe-Taylor and Williamson (1997). McAllester improved on the initial
method and introduced the name “PAC-Bayesian bounds” (McAllester, 1998).
Catoni also improved these results to derive sharp oracle inequalities (Catoni,
2003, 2004, 2007). This methods were used in various complex models of statistical learning (Guedj and Alquier, 2013; Alquier, 2013; Suzuki, 2015; Mai
and Alquier, 2015; Guedj and Robbiano, 2015; Giulini, 2015; Li et al., 2016).
Dalalyan and Tsybakov (2008) proved a different PAC-Bayesian bound based
on the idea of unbiased risk estimation (see Leung and Barron, 2006). We
first recall its form in the context of matrix factorization.
Theorem 2. Under C1, as soon as λ ≤ 4σ1 2 ,
¾
½Z
K( ρ , π )
2
>
2
c
Ek Mλ − M kF ≤ inf
,
kUV − M kF ρ (U, V , γ)d(U, V , γ) +
ρ
λ
where the infimum is taken over all probability measures ρ absolutely continuous with respect to π, and K(µ, ν) denotes the Kullback-Leibler divergence
between two measures µ and ν.
19
We let the reader check that the proof in Dalalyan and Tsybakov (2008),
stated for vectors, is still valid for matrices.
The end of the proof of Theorem 1 is organized as follows. First, we define in
Section A.2 a parametric family of probability distributions ρ :
©
ª
ρ r,U 0 ,V 0 ,c : c > 0, 1 ≤ r ≤ K, (U 0 , V 0 ) ∈ Mr .
We then upper bound the infimum over all ρ by the infimum over this parametric family. So, we have to calculate, or upper bound
Z
kUV > − M k2F ρ r,U 0 ,V 0 ,c (U, V , γ)d(U, V , γ)
and
K(ρ r,U 0 ,V 0 ,c , π).
This is done in two lemmas in Section A.3 and Section A.4 respectively. We
finally gather all the pieces together in Section A.5, and optimize with respect to c.
A.2
A parametric family of factorizations
0
0
We define, for
p any r ∈ {1, . . . , K } and any pair of matrices (U , V ) ∈ Mr , for
any 0 < c ≤ σK r , the density
ρ r,U 0 ,V 0 ,c (U, V , γ) =
A.3
1{kU −U 0 kF ≤ c,kV −V 0 kF ≤ c} π(U, V , γ)
¡©
ª¢ .
π kU − U 0 kF ≤ c, kV − V 0 kF ≤ c
Upper bound for the integral part
Lemma A.1.
Z
kUV > − M k2F ρ r,U 0 ,V 0 ,c (U, V , γ)d(U, V , γ)
³
´2
p
≤ kU 0 V 0> − M k2F + 4 c2 kU 0 kF + kV 0 kF + σK r .
Proof. Note that (U, V ) belonging to the support of ρ r,U 0 ,V 0 ,c implies that
kUV > − U 0 V 0> kF = kU (V > − V 0> ) + (U − U 0 )V 0> kF
≤ kU (V > − V 0> )kF + k(U − U 0 )V 0> kF
≤ kU kF kV − V 0 kF + kU − U 0 kF kV 0 kF
≤ (kU 0 kF + c) c + ckV 0 kF
20
¡
¢
= c kU 0 kF + kV 0 kF + c .
Now, let Π be the orthogonal projection on the set
© 0
¡
¢ª
M : k M 0 − U 0 V 0> kF ≤ c kU 0 kF + kV 0 kF + c
with respect to the Frobenius norm. Note that
kUV > − M k2F ≤ kUV > − Π( M )k2F + kΠ( M ) − M k2F
£ ¡
¢¤2
≤ 2 c kU 0 kF + kV 0 kF + c + kU 0 V 0> − M k2F .
p
σK r to get the result.
Integrate with respect to ρ r,U 0 ,V 0 ,c and use c ≤
A.4
Upper bound for the Kullback-Leibler divergence
Lemma A.2. Under C2 and C3,
Ã
K(ρ r,U 0 ,V 0 ,c , π) ≤ 2( m 1 ∨ m 2 )r log
Ã
+
X
log
1≤ i ≤ m 1
1≤`≤ r
s
2
Cf
1
2n
m1 ∧ m2
!
Ã
!
+
g(U i0` + 1)
X
log
1≤ j ≤ m 2
1≤`≤ r
Ã
+ βK log
2S f n
s
r
1
!
g(V j0` + 1)
!
µ ¶
2K
1
+ K log
+ log(4).
m1 m2
α
Proof. By definition
K(ρ r,U 0 ,V 0 ,c , π) =
Z
ρ r,U 0 ,V 0 ,c (U, V , γ) log
Ã
= log R
µ
ρ r,U 0 ,V 0 ,c (U, V , γ)
π(U, V , γ)
1
¶
d(U, V , γ)
!
1{kU −U 0 kF ≤ c,kV −V 0 kF ≤ c} π(U, V , γ)d(U, V , γ)
.
Then, note that
Z
1{kU −U 0 kF ≤ c,kV −V 0 kF ≤ c} π(U, V , γ)d(U, V , γ)
¶
Z µZ
=
1{kU −U 0 kF ≤ c,kV −V 0 kF ≤ c} π(U, V |γ)d(U, V ) π(γ)dγ
¶
¶
Z µZ
Z µZ
1{kV −V 0 kF ≤ c π(V |γ)dV π(γ)dγ .
=
1{kU −U 0 kF ≤ c π(U |γ)dU π(γ)dγ
|
{z
}|
{z
}
=:I 1
=:I 2
21
So we have to lower bound I 1 and I 2 . We deal only with I 1 , as the method to
lower bound I 2 is exactly the same. We define the set E ⊂ RK as
(
Ã
#)
c
K
E = γ ∈ R : γ1 , . . . , γr ∈ (0, 1] and γr+1 , . . . , γK ∈ 0,
.
p
2S f 2K m 1
Then
Z µZ
¶
1{kU −U 0 kF ≤ c π(U |γ)dU π(γ)dγ ≥
¶
1{kU −U 0 kF ≤ c π(U |γ)dU π(γ)dγ
{z
}
|
Z µZ
E
=:I 3
and we first focus on a lower-bound for I 3 when γ ∈ E .
¯
¯
¯
X
¯
I3 = π
(U i,` − U i,0` )2 ≤ c2 ¯γ
¯
1≤ i ≤ m 1
¯
1≤`≤K
¯
¯
¯
X
X
¯
= π
(U i,` − U i,0` )2 +
U i,2` ≤ c2 ¯γ
¯
1≤ i ≤ m 1
1≤ i ≤ m 1
¯
1≤`≤ r
r +1≤`≤K
¯
¯
¯
¯
2¯
2¯
X
c
c
¯
¯
X
(U i,` − U i,0` )2 ≤ ¯γ
≥ π
U i,2` ≤ ¯γ π
2 ¯¯
2 ¯¯
1≤ i ≤ m 1
1≤ i ≤ m 1
1≤`≤ r
r +1≤`≤K
¯
¯
¯ ¶
µ
2¯
c2 ¯¯
c
X
¯ Y
2
0 2
≥ π
U i,` ≤ ¯γ
γ .
π (U i,` − U i,` ) ≤
2 ¯¯ 1≤ i≤m1
2m1 r ¯
1≤ i ≤ m 1
| r+1≤`≤K {z
} 1≤`≤r
=:I 4
Now, using Markov’s inequality,
¯
¯
c ¯
X
2
1 − I4 = π
U i,` ≥ ¯γ
2 ¯¯
1≤ i ≤ m 1
r +1≤`≤K
¯ ¶
µ
¯
P
2 ¯
Eπ
1≤ i ≤ m 1 U i,` ¯γ
r +1≤`≤K
≤2
c2
P
2 2
1≤ i ≤ m 1 γ j S f
= 2 r+1≤`≤2K
c
1
≤ ,
2
2¯
22
and as on E , for ` ≥ r + 1, γ j ≤ c/(2S f
p
2K m 1 ). So
1
I4 ≥ .
2
Next, we remark that
µ³
´2
π U i,` − U i,0` ≤
¯ ¶ Z U0 + p c
µ ¶
i,`
c2 ¯¯
u
2m1 r 1
f
du
γ ≥
¯
0
2m1 r
γj
γj
U i,`
µ ¶
Z U0 + p c
i,`
u
2m1 r C f
e
≥
f
d u.
γj
γj
U i,0`
e
Elementary
¡
¢ calculus shows that, as f is non-negative and non-increasing,
γ j 7→ fe u/γ j /γ j is non-increasing. As such, when γ ∈ E and j ≤ r , γ j ≤ 1,
µ³
´2
π U i,` − U i,0` ≤
¯ ¶
µ
¶
cC f
c2 ¯¯
c
0
e
γ ≥p
f U i,` + p
2m1 r ¯
2m1 r
2m1 r
³
cC f
p ´
≥p
fe U i,0` + σ
2m1 r
p
p
as c ≤ σK r ≤ σ m 1 r . We plug this result and the lower-bound I 4 ≥ 1/2 into
the expression of I 3 to get
µ
¶m1 r
³
´
cC f
1
Y e 0 p
f U i,` + σ .
I3 ≥
p
2
2m1 r
1≤ i ≤ m 1
1≤`≤ r
So
Z
I1 ≥
E
I 3 π(γ)dγ
cC f
1
p
2
2m1 r
¶m1 r
cC f
1
=
p
2
2m1 r
¶m1 r
cC f
1
≥
p
2
2m1 r
¶m1 r
µ
=
µ
µ
Y
1≤ i ≤ m 1
1≤`≤ r
Y
1≤ i ≤ m 1
1≤`≤ r
Y
1≤ i ≤ m 1
1≤`≤ r
Z
³
´
p
0
fe U i,` + σ π(γ)dγ
E
µZ
´
³
p
0
e
f U i,` + σ
¶r µZ
1
h( x)d x
0
Ã
´
³
p
fe U i,0` + σ αK
23
2S f
2K m 1
0
c
p
2S f 2K m 1
¶K − r
pc
!β(K −r)
h( x)d x
cC f
1
≥
p
2
2m1 r
¶m1 r
µ
Y
1≤ i ≤ m 1
1≤`≤ r
Ã
³
´
p
fe U i,0` + σ αK
c
p
2S f 2K m 1
!β K
,
using C2. Proceeding exactly in the same way,
!β K
Ã
µ
¶m2 r
³
´
cC f
1
c
Y e 0 p K
f V j,` + σ α
.
I2 ≥
p
p
2
2m2 r
2S f 2K m 2
1≤ j ≤ m 2
1≤`≤ r
We combine these inequalities, and we use trivia between m 1 , m 2 , m 1 ∨ m 2
and m 1 + m 2 to obtain
à p
!
2 2σ ( m 1 ∨ m 2 ) r
K(ρ r,U 0 ,V 0 ,c , π) ≤ 2( m 1 ∨ m 2 )r log
cC f
!
!
Ã
Ã
X
X
1
1
+
log
+
log
p
p
fe(U i0` + σ)
fe(V j0` + σ)
1≤ j ≤ m 2
1≤ i ≤ m 1
1≤`≤ r
1≤`≤ r
Ã
!
p
µ ¶
2S f 2σK m 1 m 2
1
+ βK log
+ K log
+ log(4).
c
α
This ends the proof of the lemma.
A.5
Conclusion
We now plug Lemma A.1 and Lemma A.2 into Theorem 2. We obtain, under
C1, C2 and C3,
(
¢
cλ − M k2 ≤ inf
E kM
F
¡
inf
inf
p
1≤ r ≤K (U 0 ,V 0 )∈Mr 0< c≤ σK r
kU 0 V 0> − M k2F
à p
!
2 2σ( m 1 ∨ m 2 ) r
2( m 1 ∨ m 2 ) r
+
log
λ
cC f
Ã
Ã
!
!
1 X
1
1 X
1
+
log
log
+
p
p
λ 1≤ i≤m1
λ 1≤ j≤m2
fe(U 0 + σ)
fe(V 0 + σ)
i`
1≤`≤ r
+
βK
λ
Ã
log
2S f
p
2σ K m 1 m 2
c
24
1≤`≤ r
!
j`
µ ¶
K
1
1
+ log(4)
+ log
λ
α
λ
)
³
´2
p
.
+ 4 c kU 0 kF + kV 0 kF + σK r
Remind that we fixed λ = 4σ1 2 . We finally choose
p
2σ2 r
2 σ
c= p
≤p
=
p
K
σσK r
σ(kU 0 kF + kV 0 kF + σK r )2
p
and so the condition c ≤ σK r is always satisfied as we imposed K ≥ 2. The
inequality becomes
2σ2 r
(
¢
cλ − M k2 ≤ inf
E kM
F
¡
inf
1≤ r ≤K (U 0 ,V 0 )∈Mr
kU 0 V 0> − M k2F
p
¡ 0
¢2
0
2( m 1 ∨ m 2 ) kU kF + kV kF + σK r
+ 8σ2 ( m 1 ∨ m 2 ) r log
r
σC f
Ã
!
Ã
!
X
X
1
1
+ 4σ2
+ 4σ2
log
log
p
p
0
e
f
(
U
+
σ
)
fe(V j0` + σ)
1≤ i ≤ m 1
1≤ j ≤ m 2
i`
s
1≤`≤ r
1≤`≤ r
Ã
+ 4σ2 βK log
2S f
p
p
¢2 !
¡
m 1 m 2 kU 0 kF + kV 0 kF + σK r
rσ
)
µ ¶
1
+ 8σ2 r + 4σ2 K log
+ 4σ2 log(4) ,
α
which ends the proof.
B
B.1
Explicit formulas for the Gibbs sampler with
exponential and Gaussian priors
Gibbs Sampler with an exponential prior f
¡ ¢
Here f ( x) = exp(− x) and g α ( x) = α1 exp − αx . So
ρbλ (U i,· |V , γ, Y )
Ã
b i − U i,· )(ΣU )−1 (U
b i − U i,· )T −
∝ exp −λ(U
³
K U
X
i,`
`=1
γ`
´
!
b i − U i,· )(ΣU )−1 (U
b i − U i,· )T − U i γ−1
= exp −λ(U
à ·µ
¶
¸
·µ
¶
¸T !
1
1
bi −
bi −
= exp −λ U
ΣU γ−1 − U i,· (ΣU )−1 U
ΣU γ−1 − U i,·
,
2λ
2λ
25
where we use the abusive notation γ−1 = (1/γ1 , . . . , 1/γK ). So ρbλ (U i,· |V , γ, Y )
amounts to a truncated Gaussian distribution
µ
¶
1
−1 2
N Û i − ΣU γ , ΣU 1R+K
2λ
λ
restricted to vectors with non-negative entries. Sampling from it can be done
using the R package tmvtnorm from Wilhelm (2015) (as in Mai and Alquier,
2015). Computation of ρbλ (V j,· |U, γ, Y ) is similar.
Note that
ρbλ (γ` |U, V , γ−` , Y ) ∝ h(γ` )
m1
Y
i =1
g γ` (U i` )
m2
Y
j =1
g γ` (V j` )
à Pm1
−(m 1 + m 2 )
= h(γ` )γ`
i =1
exp −
U i,` +
Pm2
j =1
V j,`
γ`
!
,
providing an incentive to consider the inverse gamma prior IΓ(a, b), i.e.,
a
h( x) = Γb(a) x−a+1 exp(− b/ x). This leads to the conditional quasi-posterior
Ã
IΓ a + m 1 + m 2 , b +
m1
X
U i,` +
i =1
m2
X
!
V j,` ,
j =1
which is a classical choice in the Bayesian literature (see Lim and Teh, 2007;
Salakhutdinov and Mnih, 2008; Lawrence and Urtasun, 2009; Zhou et al.,
2010; Alquier et al., 2014). However, as pointed out in Alquier et al. (2014),
another conjugate choice is the gamma prior Γ(a, b) for a = m 1 + m 2 − 1/2.
a
Actually, when h( x) = Γb(a) xa−1 exp(− bx),
à Pm1
−(m 1 + m 2 )+a−1
ρbλ (γ` |U, V , γ−` , Y ) ∝ γ`
exp −
i =1
U i,` +
γ`
Pm2
j =1
V j,`
!
− bγ` .
Thus, choosing the prior Γ( m 1 + m 2 − 1/2, b) yields the conditional quasiposterior
s P
P 2
Ã
!
m1
m1
m2
U + m
V
X
X
i =1 i,`
j =1 j,`
IG
,2
U i,` +
V j,` ,
b
i =1
j =1
where IG(µ, ν) denotes the
Gaussian distribution, whose density is
³
´´
³ inverse
1
ν x
−3/2
proportional to x
exp − 2 µ2 + x . Alquier et al. (2014) contains numerical experiments to assess that this prior is less sensitive than the inverse
gamma prior to a misspecification of K .
26
B.2
Gibbs sampler with a truncated Gaussian prior f
Here, f ( x) ∝ exp(2ax − x2 ). So
ρbλ (U i,· |V , γ, Y )
Ã
b i − U i,· )(ΣU )−1 (U
b i − U i,· )T + 2a
∝ exp −λ(U
K U
X
i,`
γ`
`=1
K U2
X
i,`
−
!
γ2`
`=1
³
´
b i − U i,· )(ΣU )−1 (U
b i − U i,· )T + 2aU i,· γ−1 − U i,· Diag(γ)−2U T ,
= exp −λ(U
i,·
which is the density of the truncated Gaussian distribution
µµ
¶
µ
¶¶
¡ −1
¢
1
1
2
−1 b >
2
N
ΣU + Diag(γ) aγ + λΣU U i , 2 ΣU + Diag(γ) 1RK .
+
λ
λ
Computation of ρbλ (V j,· |U, γ, Y ) is similar. Next,
ρbλ (γ` |U, V , Y )
Pm1
Ã
−(m 1 + m 2 )
∝ h(γ` )γ`
exp 2a
i =1
U i,` +
Pm2
j =1
V j,`
γ`
Pm1
i =1
−
U i,2` +
Pm2
j =1
V j,2`
γ2`
!
.
Clearly, in all generality we cannot hope to recover an inverse gamma nor
an inverse Gaussian distribution. However, when a = 0,
ρbλ (γ` |U, V , Y ) ∝ h(γ` )
¡
¢−
γ2`
m1 +m2
2
à Pm1
exp −
i =1
U i,2` +
Pm2
j =1
V j,2`
γ2`
!
,
and in that case, considering a more convenient prior IΓ(a, b) for γ2` instead
of γ` , we have
!
Ã
m1
m2
X
X
m
+
m
1
2
V j,2` .
γ2` |U, V , Y ∼ IΓ a +
, U i,2` +
2
i =1
j =1
Alternatively, with the prior γ2` ∼ Γ(( m 1 + m 2 − 1)/2, b),
s P
γ2` |U, V , Y ∼ IG
P 2 2
m1
U2 + m
V
i =1 i,`
j =1 j,`
b
"
,2
m1
X
i =1
U i,2` +
m2
X
j =1
#
V j,2` ,
allowing for efficient sampling of our quasi-Bayesian estimator.
27
C
Explicit optimization algorithms with exponential and Gaussian priors
We optimize with respect to U and V by projected gradient descent. Since
the modes
p of IΓ(α, β) and of IG(µ, ν) are known to be respectively β/(α + 1)
and µ[ 1 + (9µ2 )/(4ν2 ) − (3µ)(2ν)], we provide an explicit optimization with
respect to γ. The algorithm for the exponential prior is detailed in Algorithm 3, whereas Algorithm 4 is adapted to the truncated Gaussian prior.
Algorithm 3 Block coordinate descent - exponential prior for U, V .
Input Y , λ.
Initialization U (0) , V (0) , γ(0) and a decreasing sequence (α` ).
While not converged, k = k + 1:
` := 0 and U (k,0) := U (k−1)
While not converged, ` = ` + 1:
³
³
´´
U (k,`) := P U (k,`−1) + α` 2λ[Y − U (k,`−1) (V (k−1) )> ]V (k−1) − (γ(k−1) )−1
U (k) := U (k,`)
` := 0 and V (k,0) := V (k−1)
While not converged, ` = ` + 1:
³
³
´´
V (k,`) := P V (k,`−1) + α` 2λ[Y > − V (k,`−1)U (k)> ]U (k) − (γ(k−1) )−1
V (k) := V (k,`)
For ` = 1, . . . , K :
If Inverse gamma prior IΓ(a, b),
γ(k)
`
:=
Pm
Pm
b+ i=11 U i,(k`) + j=21 V j,(k`)
a+ m 1 + m 2 +1
rP
Pm
m1
U (k) + j=21 V j,(k`)
i =1 i,`
γ(k)
:
=
`
b
If Gamma prior Γ( m 1 + m 2 − 1/2, b),
r
9
3
³P
´− r ³
1+
P
´
m1
m
P m 1 ( k) P m 2 ( k)
16b i=1 U i,(k`) + j=21 V j,(k`)
4 b
U +
V
i =1
28
i,`
j =1
j,`
×
Algorithm 4 Block coordinate descent - truncated Gaussian prior for U, V .
Input Y , λ.
Initialization U (0) , V (0) , γ(0) and a decreasing sequence (α` ).
While not converged, k = k + 1:
` := 0 and U (k,0) := U (k−1)
While not converged, ` = ` + 1:
³
³
U (k,`) := P U (k,`−1) + αk 2λ[Y − U (k,`−1) (V (k−1) )> ]V (k)
¶¶
³
´−2
(k,`−1) (k) −1
(k,`−1)
(k)
(k,`−1)>
+2aU
(γ ) − U
Diag γ
U
U (k) := U (k,`)
` := 0 and V (k,0) := V (k−1)
While not converged, ` = ` + 1:
V
(k,`)
³
³
(k,`−1)
:= P U
+ αk 2λ[Y > − V (k,`−1)U (k)> ]U (k)
¶¶
³
´−2
(k,`−1) (k) −1
(k,`−1)
(k)
(k,`−1)>
+2aV
(γ ) − V
Diag γ
V
V (k) := V (k,`)
For ` = 1, . . . , K :
If Inverse gamma prior IΓ(a, b),
γ(k)
:=
`
b+
Pm1 ³ (k) ´2 Pm2 ³ (k) ´2
U i,` + j=1 V j,`
i =1
a + m1 +2 m2 + 1
r
Pm1 ³ (k) ´2 Pm2 ³ (k) ´2
U i,` + j=1 V j,`
i =1
If Gamma
prior Γ( m 1 + m 2 − 1/2, b), γ(k)
:=
b
`
×
s
9
3
µ
¶
s
1+
−
³
´
³
´
µ
¶
2 Pm
2
Pm
´2 P ³
´2
Pm ³
m
16b i=11 U i,(k`) + j=21 V j,(k`)
4 b i=11 U i,(k`) + j=21 V j,(k`)
29
| 10math.ST
|
Clubs-based
Particle Swarm Optimization
Wesam Elshamy, Hassan M. Emara and A. Bahgat
Department of Electrical Power and Machines,
Faculty of Engineering, Cairo University, Egypt
{wesamelshamy, hmrashad, ahmed.bahgat}@ieee.org
Abstract – This paper introduces a new dynamic neighborhood network for particle swarm optimization. In the proposed
Clubs-based Particle Swarm Optimization (C-PSO) algorithm,
each particle initially joins a default number of what we call
‘clubs’. Each particle is affected by its own experience and the
experience of the best performing member of the clubs it is a
member of. Clubs membership is dynamic, where the worst
performing particles socialize more by joining more clubs to
learn from other particles and the best performing particles are
made to socialize less by leaving clubs to reduce their strong
influence on other members. Particles return gradually to default membership level when they stop showing extreme performance. Inertia weights of swarm members are made random
within a predefined range. This proposed dynamic neighborhood algorithm is compared with other two algorithms having
static neighborhood topologies on a set of classic benchmark
problems. The results showed superior performance for C-PSO
regarding escaping local optima and convergence speed.
I. INTRODUCTION
The relatively recent invention of computers with its increasing computational power allowed researchers to implement myriads of paradigms and algorithms and verify their
efficiency. Researchers developed many algorithms inspired
from natural phenomena such as the annealing in metallurgy
[1], biological processes such as in genetics [2] and the immune system [3], animals and insects behaviors such as ants
[4] and birds [5] and even cultures [6].
Particle Swarm Optimization (PSO) [5] is among these nature inspired algorithms. It is inspired by the ability of birds
flocking to find food that they have no previous knowledge
of its location. Every member of the swarm is affected by its
own experience and its neighbors’ experiences. Although the
idea behind PSO is simple and can be implemented by two
lines of programming code, the emergent behavior is complex and hard to completely understand [7].
In this paper we propose and study a new dynamic social
network for PSO, where the social networks of the best performing particles shrink to reduce their influence on other
particles, while the social networks of the worst performing
particles expand to allow them to learn from more particles.
The organization of this paper is as follows; in part II we
present the basic version of PSO and some of its variants and
related work, followed by an explanation of the proposed
Clubs-based PSO (C-PSO) algorithm in section III, while
section IV presents the experiments to be conducted. The
experiments’ results are presented in section V, then a conclusion follows in section VI.
II. PARTICLE SWARM OPTIMIZATION
Particle Swarm Optimization was inspired by the ability of
a flock of birds or a school of fish to capitalize on their collective knowledge in finding food or avoiding predators.
Each swarm member or particle has a small memory that
enables it to remember the best position it found so far and its
goodness. Particles are affected by their own experience (best
found position) and their neighbors’ experiences (best found
position by the neighbors). The behavior of particles is described by (1) and (2).
v id (t + 1) = w × v id (t ) + lrn 1 × rand 1 × ( p id (t ) − x id (t ) )
+ lrn 2 × rand 2 × ( p gd (t ) − x id (t ) )
(1)
x id ( t + 1) = x id ( t ) + v id ( t + 1)
(2)
In (1), vid is the speed of particle i in dimension d. The first
right hand side term corresponds to the inertia force that
pushes the particle in its old direction, where w is the weight
value that controls this inertia force. The second term corresponds to the cognitive or personal experience component. It
attracts the particle from its current position xid to its best
found position so far in that dimension pid affected by a learning weight lrn1 and a uniformly distributed random variable
rand1 in the range (0, 1). The third term corresponds to the
social influence of the neighbors on the particle. It affects the
particle by attracting it from its current position xid to the best
position found by its neighbors pgd and this influence is controlled by a learning weight lrn2 and another independent
random variable rand2 uniformly distributed in the range
(0, 1). For each time step, as described by (2), each particle
moves by a step of value vid in the d th dimension.
The PSO algorithm itself has evolved. The weight parameter w was not included in the basic algorithm. It was added
later and researchers examined the effect of varying its value
[8]. A speed limit for the particles was introduced to prevent
the explosion of speed values.
PSO operates in three spaces, the social network space, the
parameter space of problem variables and the evaluative
space [7] where estimates for the goodness of solutions are
defined. Various social networks have been proposed and
investigated by researchers [9]. In the original PSO
algorithm, the social network connects every particle to all
other particles and it is only influenced by the one that has
the best experience compared to all particles. We will refer to
this algorithm as PSO-g where ‘g’ stands for global. Though
this algorithm converges rapidly, it could get easily trapped
in local minima. After the particles are initialized, the first
best found position by all particles attracts them all and as
long as the particles experience improving performance in
their new positions while heading for this first found minimum, they get more strongly attracted towards this local
minimum.
A variant of the simple PSO has a ring social network. In
this algorithm the particles are arranged in an imaginary ring
and every particle is connected to its immediately preceding
and succeeding particles in this ring. We will refer to this
algorithm as PSO-l where ‘l’ stands for local. This algorithm
converges slower than PSO-g but it is less susceptible to local
minima and enjoys a higher degree of particles diversity. The
influence of each particle in the swarm is limited to its two
immediate neighbors. This influence limitation helps the particles to explore the search space with different points of attraction instead of a single best found point in the PSO-g algorithm. On the other hand, it may lead to excessive wandering for the particles leading to slow convergence even in easy
problems having single optimum.
Related Work:
Both PSO-g and PSO-l are based on a static neighborhood
network. Because the first stages of the search for the global
best position require exploration of possible solutions, which
PSO-l can do better. While later stages require exploitation of
the best found candidate solutions by early stages of the
search, which PSO-g is clever at. Researchers suggested using a dynamic neighborhood.
In [10], the neighborhood of each swarm member expands
from an initial network that connects each particle to itself at
early stages of the search, to a network that fully connects it
to all other particles. This algorithm transforms gradually
from acting like PSO-l in early stages of the search, to behave
more like PSO-g at late stages. Two network expanding procedures have been introduced. Both of them depend on the
current position of the particle to search for nearby particles
to add to its neighborhood list.
In [11], a Fitness-Distance-Ratio based PSO (FDR-PSO)
algorithm is introduced. In this algorithm, each particle is
affected by three components; the cognitive, social and the
FDR components. The third component corresponds to the
influence of the particle that maximizes the FDR. The higher
the fitness of the neighbor and the closer its distance to the
original particle, the more likely it will influence this particle.
A new learning factor is introduced for the FDR component.
In [12], a randomly generated directed graphs are used to
define neighborhood where graph links are unidirectional, so
a link from a to b means that a considers b as a neighbor, but
not vice versa. Two methods for modifying the neighborhood
structure are tested. The ‘random edge migration’ method
disconnects one side of an edge and connects it to another
neighbor, while the ‘neighborhood re-structuring’ method
totally re-initializes the structure after it is kept fixed for a
period of time.
In [13], a Hierarchical PSO (H-PSO) version is introduced.
In this algorithm, particles are arranged in a hierarchy structure and the best performing particles ascend the tree to influence more particles, replacing relatively worse performing
particles which descend the tree. A variant of this algorithm
where the structure of the tree itself is made dynamic is presented and tested.
III. CLUBS-BASED PSO
PSO first models were confined to perceive the swarm as a
flock of birds that fly in the search space. The picture of flying birds has limited the imagination of researchers somehow
for sometime. Recently, a more broad perception of the
swarm as a group of particles, whether birds, humans, or any
socializing group of particles began to emerge.
In our proposed C-PSO algorithm, we create ‘clubs’ for
particles analogous to our clubs where we meet and socialize.
In our model, every particle can join more than one club, and
each club can accommodate any number of particles. Vacant
clubs are allowed.
After randomly initializing the particles position and speed
in the initialization range, each particle joins a predefined
number of clubs, which is known as its ‘default membership
level’, and the choice of these clubs is made random. Then,
current values of particles are evaluated and the best local
position for each particle is updated accordingly. While updating the particles’ speeds, each particle is influenced by its
best found position and the best found position by all its
neighbors, where its neighborhood is the set of all clubs it is a
member of. After speed and position update, the particles’
new positions are evaluated and the cycle is repeated.
While searching for the global optima, if a particle shows
superior performance compared to other particles in its
neighborhood, the spread of the strong influence by this particle is reduced by reducing its membership level and forcing
it to leave one club at random to avoid premature convergence of the swarm. On the other hand, if a particle shows
poor performance, that it was the worst performing particle in
its neighborhood, it joins one more club selected at random to
widen its social network and increase the chance of learning
from better particles.
The cycle of joining and leaving clubs is repeated every
time step, so if a particle continues to show the worst performance in its neighborhood, it will join more clubs one
after the other until it reaches the maximum allowed membership level. While the one that continues to show superior
performance in every club it is a member of will shrink its
membership level and leave clubs one by one till it reaches
the minimum allowed membership level.
During this cycle of joining and leaving clubs, particles
which no longer show extreme performance in its neighborhood, either by being the best or the worst, go back gradually
to default membership level. The speed of going back to default membership level is made slower than that of diverting
from it due to extreme performance. The slower speed of
regaining default membership level allows the particle to
linger, and adds some stability and smoothness to the performance of the algorithm. A check is made every rr (retention ration) iterations to find the particles that have membership levels above or below the default level, and take them
back one step towards the default membership level if they
do not show extreme performance.
We replace the static inertia weight which controls the
momentum of the particle by a uniformly distributed random
number in the range (0, w). A pseudo code explaining the
algorithm is shown below.
1
2
1
Initialize particles and clubs
while (termination condition = false)
do
evaluate particles fitness: f(x)
update P
for (i = 1 to number of particles)
gi = best of neighborsi
for d = 1 to number of dimensions
vid = w×rand1×vid + lrn1×rand2×(pid – xid)
+ lrn2×rand3×(gid -xid)
xid = xid + vid
next d
next i
update neighbors
for j = 1 to number of particles
if (xj is best of neighborsj) and
(|membershipj| > min_membership)
leave random club
end if
if (xj is worst of neighborsj) and
(|membershipj| < max_membership)
join random club
end if
if (|membershipj| ≠ default_membership)
and (remainder(iteration/rr) = 0)
update membershipj
end if
next j
end do
iteration = iteration + 1
evaluate termination condition
end while
2
3
3
3
2
2
1
4
4
1
5
7
5
6
6
5
7
2
8
best particle
in club
begin
3
8
6
4
8
6
worst particle
in club
5
7
particle
Fig. 1. A snapshot of clubs during a simulation of
the C-PSO algorithm
Where P is local best position, neighborsi is the set of
particle i neighbors, membershipi, |membershipi|
are the set of clubs that particle i is a member of and the size
of this set respectively. rand1,2,3 are three independent uniformly distributed random numbers in the range (0, 1).
Fig. 1 shows a snapshot of the clubs during an execution of
the C-PSO algorithm. In this example, the swarm consists of
8 particles, and there are 6 clubs available for them to join.
Given the previous pseudo code, and that the minimum, default and maximum membership levels are 2, 3 and 5 respectively, the following changes in membership will happen to
particles in Fig. 1 for the next iteration which is a multiple of
rr:
1. Particle3 will leave club1, 2 or 3 because it is the best
particle in its neighborhood.
2. Particle5 will join club1, 2 or 4 because it is the worst
particle in its neighborhood.
3. Particle2 will leave club1, 2, 3 or 4, while particle4 will
join club2, 3, 4, or 6 to go one step towards default
membership level because they do not show extreme performance in their neighborhood.
Flow of influence:
The flow of influence or how the effect of the best performing particles spreads and affects other particles in the
swarm is critical to the performance of all PSO algorithms. If
the influence spreads quickly through the swarm, they get
strongly attracted to the first optimum they find, which is a
local optimum in most cases. On the other hand, if the influence spreads slowly, the particles will go wandering in the
search space and will converge very slowly to the global or a
local optimum.
In order to study the effect of different default membership
levels on the flow of influence we do the following experiment. We create a swarm of 20 particles. Clubs membership
is assigned randomly but every particle joins exactly m of
total 100 clubs. The membership level m is kept fixed for
every single run, so best and worst performing particles do
not leave or join clubs. All the particles are initialized to
random initial positions in the range [1000 2000]n except for
one particle which is initialized to [0]n. The value of each
Average Value
1500
4 clubs
5 clubs
7 clubs
10 clubs
15 clubs
80 clubs
1000
500
0
0
20
40
60
80
100
Iteration
Fig. 2. Effect of different default membership levels on the speed of
the flow of influence
particle which we want to minimize is simply the sum of its
coordinate position values.
The flow of influence for some default membership levels
is shown in Fig. 2. The average value of all particles is shown
against search progress. The average value of particles decreases because they are influenced by the best performing
particle which has a value of ‘0’. So, rapid decrease of the
average value indicates faster flow of influence speed. It is
clear that the flow of influence speed monotonically decreases with decreasing default membership levels.
IV. EXPERIMENTS
The purpose of this paper is to test and analyze the effect
of the dynamic social network employed in the proposed
C-PSO algorithm on its performance and compare it with the
performance of other PSO algorithms which have static social networks. We use five well known benchmark problems
presented in TABLE I. The first two functions are simple
unimodal functions. They test the ability of the optimizers to
deal with smooth landscapes. The next three functions are
multimodal functions containing a considerable number of
local minima where the algorithm may fall into, so these
functions test the ability of the algorithm to escape these
traps.
We compare the performance of the different optimizers
using two criteria which were used in [13]. The first one is
the ability to escape local minima, and is measured by the
degree of closeness to the global optimum the optimizer
achieves after a long number of iterations. The second one is
the convergence speed, which is measured by the required
number of iterations to achieve a certain degree of closeness
to the global optimum in the evaluation space.
Using these metrics on the five benchmark functions, we
compare three versions of the C-PSO with different default
membership levels of 10, 15 and 20 from a total number of
100 clubs, along with PSO-g and PSO-l.
The three default membership levels are chosen based on
initial empirical results. It was found that lower membership
levels decrease the speed of flow of influence, as shown in
the previous section, which was reflected on slow conver-
TABLE I
BENCHMARK FUNCTIONS
n
Sphere
(unimodal)
f 1 ( x ) = ∑ xi2
Rosenbrock
(unimodal)
f2 (x) =
∑
Rastrigin (multimodal)
f 3 (x ) =
∑ [x
Schaffer f6
(multimodal)
f 4 (x ) = 0 .5 +
Achley
(multimodal)
⎛
1 n 2 ⎞⎟
f 5 ( x ) = − 20 exp ⎜ − 0 . 2
∑ xi ⎟
⎜
n i =1
⎠
⎝
⎞
⎛1 n
− exp ⎜ ∑ cos (2 π x i )⎟ + 20 + e
⎠
⎝ n i =1
i =1
n −1
i =1
[100 (x
n
i =1
2
i
− x i2
i +1
)
2
+ (x i − 1 )
2
− 10 cos (2 π x i ) + 10
sin
(
2
)
]
]
x 2 + y 2 − 0 .5
(1 + 0 .001 × (x
2
+ y2
))
2
TABLE II
PARAMETERS FOR BENCHMARK FUNCTIONS
w
Function
Dim.
Init. range
V
max
Sphere
Rosenbrock
30
30
n
[-100; 100]
n
[-30; 30]
n
C-PSO
100
1.2
30
1.2
Rastrigin
30
[-5.12; 5.12]
5.12
1.4
Schaffer’s f6
2
[-100; 100]n
100
1.65
32
1.36
Achley
30
n
[-32; 32]
gence. While higher membership levels cause premature
convergence.
For all simulation runs we use lrn1 = 1.494, lrn2 = 1.494,
which were used in [13] and suggested by [14]. For PSO-g
and PSO-l we use w = 0.729 as in [13] and [14], while the
value of w for C-PSO which reflects the range of the random
inertia weight is presented in TABLE II for each problem.
The minimum and maximum allowed membership levels
we use are 5 and 33 respectively, while rr = 2. We use a
swarm of 20 particles for all simulation runs.
The particles position and speed are randomly initialized in
the ranges shown in TABLE II depending on the benchmark
problem used. The absolute speed values for particles are
kept within the Vmax limit for all dimensions during simulation. On the other hand, particles movements are not restricted by boundaries, so particles may go beyond the initialization range and take any value.
Every simulation run was allowed to go for 10000 iterations, and each simulation has been repeated 50 times. All
simulation runs were executed using MATLAB® R2006a.
V. RESULTS
Each graph presented in this section represents the average
of the 50 independent simulation runs for all optimizers
unless otherwise stated.
100
-3
C-PSO (10)
C-PSO (15)
C-PSO (20)
PSO-g
PSO-l
0
10
-100
10
6
Distance to Global Optimum
Distance to Global Optimum
10
-200
10
0
2000
4000
6000
Iteration
5
4
C-PSO (10)
C-PSO (15)
C-PSO (20)
PSO-g
PSO-l
3
2
1
0
8000
x 10
0
2000
Fig. 3. Sphere⎯Closeness to global optimum for
C-PSO (10, 15, 20), PSO-g and PSO-l
C-PSO (10)
C-PSO (15)
C-PSO (20)
PSO-g
PSO-l
40
20
10
10
10
10
10
6
0
2000
4000
6000
Iteration
8000
Fig. 4. Rosenbrock⎯Closeness to global optimum for
C-PSO (10, 15, 20), PSO-g and PSO-l
Distance to Global Optimum
80
70
60
C-PSO (10)
C-PSO (15)
C-PSO (20)
PSO-g
PSO-l
50
40
30
0
2000
4000
6000
8000
Iteration
Fig. 5. Rastrigin⎯Closeness to global optimum for
C-PSO (10, 15, 20), PSO-g and PSO-l
A. Escaping Local Minima:
We begin with the first criterion which is the ability of the
algorithm to escape local minima. The Sphere and Rosenbrock problems have the lowest number of local minima.
8000
Fig. 6. Schaffer’s f6⎯Closeness to global optimum for
C-PSO (10, 15, 20), PSO-g and PSO-l
Distance to Global Optimum
Distance to Global Optimum
100
4000
6000
Iteration
10
2
C-PSO (10)
C-PSO (15)
C-PSO (20)
PSO-g
PSO-l
1
0
-1
-2
0
2000
4000
6000
8000
Iteration
Fig. 7. Ackley⎯Closeness to global optimum for
C-PSO (10, 15, 20), PSO-g and PSO-l
Their unique minimum makes them the easiest of the five
benchmark problems in finding the global minimum.
For the Sphere problem as shown in Fig. 3, all C-PSO versions managed to finish closer to the unique minimum than
PSO-g and PSO-l, and the lower the membership level the
faster the algorithm converges. The PSO-l algorithm was the
worst performer followed by PSO-g.
For the Rosenbrock problem presented in Fig. 4, C-PSO
(10, 20) and PSO-g show very close performance, though
PSO-g is little behind them. C-PSO (15) follows them by a
short distance, while PSO-l is the worst of all, lagging behind
by a relatively long distance.
As shown in Fig. 5 and Fig. 6, we can see that all C-PSO
versions perform better than both PSO-g and PSO-l for the
Rastrigin and Schaffer’s f6 test problems respectively. In
both of them, PSO-g gives the worst performance and converges prematurely in the Rastrigin problem, followed by
PSO-l as the second worst. All C-PSO versions give similar
performance for the Rastrigin problem as we can hardly distinguish them.
As shown in Fig. 7 for the Ackley problem. C-PSO (10)
outperforms all the other algorithms followed by PSO-l,
C-PSO (20, 15) respectively, while PSO-g suffers premature
Sphere
Rosen.
Rastrigin
Schaffer
Ackley
PSO-l
7.4e-77
21.08
60.37
0.00077
0.06
PSO-g
4.2e-93
6.88
75.32
0.00466
4.45
C-PSO (20)
1.3e-107
5.92
35.10
0
0.10
C-PSO (15)
5.4e-137
9.11
34.34
0
0.20
C-PSO (10)
1.1e-152
6.07
34.30
0.00019
0.03
convergence again and falls by a long distance behind. The
distances to global optima after 10000 iterations of the optimizers are shown in TABLE III.
As expected, the performance of PSO-g and PSO-l depend
on the problem they optimize. For the first two problems
which have a single optimum, PSO-g performs better than
PSO-l, as all the particles get strongly attracted to the unique
optimum due to the fully connected social network in PSO-g.
On the other hand, PSO-l goes wandering and converges
slowly.
For the last three problems, which have many local optima, PSO-l outperforms PSO-g. The partially connected social network of PSO-l creates many points of attraction for
the particles in the swarm that help them escape some local
optima compared to PSO-g.
Unlike PSO-g and PSO-l, C-PSO performance is much
less problem dependent. C-PSO (10) outperformed both
PSO-g and PSO-l for all problems. The results obtained for
the Sphere problem were unexpected. The unique minimum
and the non-deceptive landscape of the problem make a perfect match with PSO-g. The fully connected social network
should do a better job in attracting the particles to the unique
global minimum than any other social network.
These results necessitated further investigation into the behavior of the optimizers and specially the flow of influence
through the swarm in unimodal and multimodal problems. So
we do the following experiment.
B. Further Investigation of Optimizers’ Behaviors:
We run three optimizers, C-PSO (10), PSO-l and PSO-g on
the unimodal Rosenbrock test problem and the multimodal
Rastrigin problem. During the simulation run we record the
index of the best performing particle in the swarm for each
iteration. We use the same parameters used previously for the
first criterion.
Fig. 8 and Fig. 9 show the best performing particles in
C-PSO (10) (top), PSO-l and PSO-g (bottom) for the Rosenbrock and Rastrigin problems respectively. For each plot, the
index of particles (20 particles) is drawn against the number
of iterations elapsed. A dot at (13, 5000) indicates that
‘particle 13’ has the global best value in the swarm during
‘iteration number 5000’.
First, considering Rosenbrock problem shown in Fig. 8.
The status of being the best performing particle in the case of
C-PSO is almost uniformly distributed over all particles, once
a particle finds a good solution, another particle finds a better
one.
Particle (PSO-g) Particle (PSO-l) Particle (C-PSO)
Algorithm
20
10
0
20
10
0
20
10
0
0
1000 2000 3000 4000 5000 6000 7000 8000 9000
Iteration
Fig. 8. Best particle in the swarm for Rosenbrock problem using
C-PSO(10) (top), PSO-l, and PSO-g (bottom)
Particle (PSO-g) Particle (PSO-l) Particle (C-PSO)
TABLE III
DISTANCES TO GLOBAL OPTIMA AFTER 10000 ITERATIONS
20
10
0
20
10
0
20
10
0
0
1000 2000 3000 4000 5000 6000 7000 8000 9000
Iteration
Fig. 9. Best particle in the swarm for Rastrigin problem using
C-PSO(10) (top), PSO-l, and PSO-g (bottom)
A reason for this behavior is that once a particle finds the
good solution it becomes the best particle in the swarm, making it the best in its neighborhood as well. The particle
shrinks its membership level one by one and reduces its influence on other particles accordingly. Neighbors of this superior particle will carry its influence to other clubs they are
member of, so other particles are still indirectly guided by it,
but are more free to find a steeper way down the hill to the
global optimum. Once a particle finds it, it becomes the new
best particle and continues or starts shrinking its membership
level (because it may become the best in its neighborhood
before becoming the global best). The particles which are no
longer the best in their neighborhood regain their default
membership level to increase their chance in learning from
better particles to become the new global best, and the cycle
continues.
On the other hand, the best particle status in PSO-l goes
bouncing between two particles on the ring (note that particle
1 is connected to particle 20) as shown in Fig. 8 (middle).
The particles outside this arc search totally inefficient regions
of the search space. This is clear from the fact that none of
them shows up even once as the best particle in the swarm for
the last 7500 iterations. This clustering mechanism may help
the algorithm to overcome local optima in multimodal problems, but in unimodal problems it has detrimental effect.
The reason that PSO-g algorithm came second to C-PSO in
both unimodal problems is clear in Fig. 8 (bottom). After
around 2600 iterations, 12 particles acted as guides for the
other 8 particles and literally dragged them behind. None of
the 8 particles showed superior performance till the end of
the 10000 iterations.
Second, we consider Rastrigin problem presented in Fig. 9.
This multimodal problem requires diversity in the swarm and
a clever social network to overcome local optima. The property of fully connected social network in PSO-g provokes all
the particles to jump to the best found position by all particles
in the swarm.
This makes the first 1500 iterations for PSO-g look almost
the same for both unimodal and multimodal problems. But
after these 1500 iterations the algorithm prematurely converges in the case of the multimodal Rastrigin problem as
shown in Fig. 9 (bottom).
On the other hand, PSO-l algorithm presented in Fig. 9
(middle) maintains its diversity for a longer period than
PSO-g does. Along with its clustering property explained
earlier, it manages to escape local optima to some extent and
get closer to global optimum than PSO-g can get.
Finally for the C-PSO optimizer as shown in Fig. 9 (top),
the algorithm maintains diversity longer than PSO-g and
PSO-l do. Moreover, the best performing particle status is
distributed over the particles, unlike PSO-l, and the particles
do not jump over the best particle once it emerges. This can
be seen as the particles create more clusters than in the case
of PSO-l and PSO-g. These clusters represent local optima
found by the particles. The most interesting result found is
the ability of the C-PSO to explore new regions after a period
of stagnation. We can see how C-PSO finds better regions at
around iteration 5200 after it has stagnated for nearly 2000
iterations.
An explanation for this behavior is that the best performing
particles in their neighborhood create different points of attraction for the particles. The particles are grouped according
to their clubs’ membership and search the space around these
points of attraction. At the same time, the worst particles on
their neighborhood expand their membership and bridge the
influence between different groups of searching particles. If a
searching group finds a better solution, its influence is transmitted over the bridge acting particles to other groups and
diverts them from searching inefficient regions indefinitely.
They start searching for other optima which could be better
than the best one found and create different points of attraction, and the cycle goes on.
C. Convergence Speed:
The second criterion to be considered is the convergence
speed of the algorithms. As explained earlier, it is being
measured by the number of iterations the algorithm takes to
reach a certain degree of closeness to the global optimum.
TABLE IV. NUMBER OF ITERATIONS NEEDED TO REACH A
CERTAIN DEGREE OF CLOSENESS TO GLOBAL OPTIMUM FOR
THE FIVE OPTIMIZERS. (BEST VALUES ARE BOLD FACED)
Algorithm
Avg.
Med.
Max.
Min.
Suc.%
PSO-l
1030.8
Sphere – (Closeness = 0.0001)
1036
1103
965
PSO-g
684.88
672
1012
489
100
C-PSO (20)
611.68
571
1057
421
100
C-PSO (15)
528.18
506.5
711
417
100
518.14
513.5
652
443
100
PSO-l
1429.6
Rosenbrock – (Closeness = 100)
907
7465
604
PSO-g
874.3
425
6749
251
100
C-PSO (20)
697.3
424
4537
240
100
C-PSO (15)
569
473
1605
218
98
C-PSO (10)
725.8
376
6016
226
98
PSO-l
1695.7
Rastrigin – (Closeness = 50)
1068
8015
500
26
PSO-ga
250
221
C-PSO (10)
100
98
313
216
6
C-PSO (20)
813.9
702
3396
254
88
C-PSO (15)
695.4
597.5
1829
262
88
C-PSO (10)
753.3
667
1932
299
96
PSO-l
1076.2
Schaffer’s f6 – (Closeness = 0.001)
422
7021
84
92
PSO-g
791.1
279.5
4276
60
52
C-PSO (20)
1138.3
524
8462
80
100
C-PSO (15)
1120.1
432
4966
88
100
C-PSO (10)
945.6
401
9668
48
98
PSO-l
968.2
Ackley – (Closeness = 0.01)
954.5
1531
827
96
PSO-ga
499
499
499
499
2
C-PSO (20)
831.1
806
1148
610
92
C-PSO (15)
800.6
793.5
1141
570
84
C-PSO (10)
863.7
841
1151
672
98
a
Not considered in comparison due to its very low success rate
This number of iterations should be small enough to reflect
the ability of the algorithm to converge rapidly, and not its
ability to escape local optima and achieve better values at
later stages of the run. On the other hand, the closeness value
chosen should lie close enough to the global optimum to be
efficient in practical applications. We choose closeness values for the five benchmark problems that are satisfied by
most algorithms around the range of [500, 1000] iterations.
These closeness values are shown in TABLE IV next to problem names.
The figures presented in TABLE IV are compiled from the
same results data set collected for the first criterion. They
represent the Average, Median, Maximum, Minimum and
Success rate of 50 independent simulation runs for the five
optimizers. Only data of successful runs were used to evaluate these values, so the sample number is not the same for all
figures.
We can see from TABLE IV that C-PSO (15) achieves the
overall best results. For the Sphere problem, all the algorithms achieve the desired closeness in every single run,
though C-PSO (10, 15) come ahead of them. The situation is
similar in the second unimodal Rosenbrock problem, however the success rate is lower for PSO-l and C-PSO (10, 15).
Moving to multimodal problems, we notice how PSO-g
shows poor performance in reaching the closeness values.
For Rastrigin and Ackley problems, it only succeeds in six
and two percent of the runs respectively, compared to much
higher success rates in all C-PSO versions.
C-PSO (15) outperform all the other algorithms for
Rastrigin and Ackley problems, except for the Rastrigin
problem where it comes second to C-PSO (20) regarding the
minimum number of iterations in the 50 samples. We do not
consider PSO-g in our comparison for Rastrigin and Ackley
problems due to its very low success rate.
Finally for Shaffer’s f6 problem, PSO-g achieved the best
results for the mean, median and maximum number of iterations. It should be noted however that it has a low success
rate of 52% which is almost half the success rate for all
C-PSO versions. This low success rate makes it unreliable in
practical applications.
VI. CONCLUSION
Particle swarm optimizers are very sensitive to the shape
of their social network. Both PSO-g and PSO-l lack the ability of adapting their social network to the landscape of the
problem they optimize.
The proposed C-PSO algorithm overcomes this problem.
The dynamic social network of the optimizer shrinks the
membership level of superior particles to reduce their influence on other particles, while expanding the membership
level for the worst particles to increase their chance in learning from better particles.
C-PSO versions achieved better results than PSO-l and
PSO-g either in escaping local optima or in convergence
speed to global optima for almost all benchmark problems we
considered.
Further investigations have shown that the dynamic social
network allowed particles to be guided indirectly by the superior particles, while searching for better solutions more freely
than the case of PSO-g. It was shown using empirical results
that C-PSO is able to explore and find better regions in the
search space during periods of stagnation, making it attractive for use in multimodal problems.
REFERENCES
[1] S. Kirkpatrick, C. D. Gelatt and M. P. Vecchi, “Optimization by simulated annealing,” Science, vol. 220, pp. 671–680, May 1983.
[2] J. H. Holland, Adaptation in Natural and Artificial Systems, Ann Arbor: University of Michigan Press, 1975.
[3] J. D. Farmer, N. Packard and A. Perelson, “The immune system, adaptation and machine learning,” Physica D, vol. 22, pp. 187–204,
1986.
[4] A. Colorni, M. Dorigo and V. Maniezzo, “Distributed optimization by
ant colonies,” in Proc. First European Conference on Artificial Life,
1991, pp. 134–142.
[5] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Proc.
IEEE Int. Conf. on Neural Networks, vol. 4, 1995, pp. 1942–1947.
[6] R. G. Reynolds, “An introduction to cultural algorithms,” in Proc.
Evolutionary Programming III, 1994, pp. 131–139.
[7] J. Kennedy, “Particle swarms: optimization based on sociocognition,”
in Recent Development in Biologically Inspired Computing, L. N. de
Castro and F. J. Von Zuben, Ed. Idea Group, 2005, pp. 235–269.
[8] Y. Shi and R. C. Eberhart, “Parameter selection in particle swarm optimization,” in Proc. Evolutionary Programming VII, vol. 1447, 1998,
pp. 591–600.
[9] J. Kennedy, “Small worlds and mega-minds: effects of neighborhood
topology on particle swarm performance,” in Proc. Cong. on Evolutionary Computation (CEC 1999), vol. 3, 1999, pp. 1931–1938.
[10] P. N. Suganthan, “Particle swarm optimization with a neighborhood
operator,” in Proc. Cong. on Evolutionary Computation (CEC 1999),
vol. 3, 1999, pp. 1958–1962.
[11] K. Veeramachaneni, T. Peram, C. Mohan and L. A. Osadciw, “Optimization using particle swarms with near neighbor interactions,” in
Proc. Genetic and Evolutionary Computation (GECCO 2003), vol.
2723, 2003, pp. 110–121.
[12] A. S. Mohais, R. Mendes, C. Ward and C. Posthoff, “Neighborhood
re-structuring in particle swarm optimization,” in Proc. Australian
Conference on Artificial Intelligence, vol. 3809, 2005, pp. 776–785.
[13] S. Janson and M. Middendorf, “A hierarchical particle swarm optimizer and its adaptive variant,” IEEE Systems, Man and Cybernetics Part B, vol. 35, no. 6, pp. 1272–1282, Dec. 2005.
[14] I. C. Trelea, “The particle swarm optimization algorithm: convergence
analysis and parameter selection,” Inform. Process. Lett., vol. 85, pp.
317–325, 2003.
| 9cs.NE
|
Resource Allocation for D2D Communications with
Partial Channel State Information
arXiv:1711.01820v2 [cs.IT] 8 Nov 2017
Anushree Neogi, Prasanna Chaporkar and Abhay Karandikar
links in order to determine an optimal resource allocation
algorithm that is centralized. The channel gains from a CU
and a D2D transmitter to the BS are known because every user
transmits its channel quality indicator (CQI) to the BS either
periodically or aperiodically [4]. However, the BS cannot
determine the CQI between a D2D transmitter-receiver pair
and also that between a CU and a D2D receiver. For a fast
fading channel, even if these channel gains are known through
some technique, conveying them to the BS in every time
slot requires significant control overhead and power. Thus,
partial CSI is a limiting factor to the BS’s resource allocation
capability which makes this a challenging problem.
With partial CSI, resource allocation problems are generally
modeled game theoretically and distributed resource allocation
algorithms are designed. In such systems the decisions that are
taken by the BS when perfect CSI is available are now taken
by the user equipments (UEs) through a learning algorithm.
This reduces the computational load at the BS considerably.
Such systems are generally modeled as non-cooperative, in
which players take actions to enhance their own utilities. The
resulting Nash equilibrium solution to the resource allocation
problem fails to achieve a socially optimum solution since the
best response of every player may not maximize the overall
social utility. The inefficiency of the Nash equilibrium is
highlighted by the well acknowledged econometric principle,
“Tragedy of the Commons”, named after the celebrated work of
Hardin [5] and can be quantified through the price of anarchy.
Abstract—Enhancement of system capacity is one of the
objectives of the fifth generation (5G) networks in which deviceto-device (D2D) communications is anticipated to play a crucial
role. This can be achieved by devising efficient resource allocation
strategies for the D2D users. While most of the works in
resource allocation assume full knowledge of the channel states,
transmitting it in every time slot reduces the system throughput
due to extra control overhead and leads to wastage of power. In
this paper, we address the problem of D2D resource allocation
with partial channel state information (CSI) at the base station
(BS) and ensure that the interference from the D2D users do
not jeopardize the communications of cellular users (CUs). With
partial CSI, existing algorithms determine the Nash equilibrium
in a distributed manner, whose inefficiency in maximizing the
social utility is well known as the players try to maximize their
own utilities. This is the first work in the D2D resource allocation
field in which within a game theoretic framework, an optimal
D2D resource allocation algorithm is proposed which maximizes
the social utility of the D2D players such that a social optimum
is attained. Each D2D player with the help of the BS learns
to select the optimal action. We consider the channel to exhibit
path loss. Next, we consider both a slow and fast fading with
CU mobility and propose two heuristic algorithms. We validate
the performance of our proposed algorithms through simulation
results.
I. I NTRODUCTION
Device-to-device (D2D) communication is envisioned to be
a promising component of the fifth generation (5G) wireless
networks. It explores the possibility of short range communication between two D2D users in an underlay Long Term
Evolution (LTE) network so that they can reuse the radio
resources already allocated to cellular users (CUs) through
efficient resource allocation algorithms. The CUs are treated as
primary users whose quality of service should not get affected
due to the transmissions of D2D users. Since the network is
able to accommodate more users, the network capacity gets
enhanced.
B. Related Work
We give a brief overview of some of the research works
pertaining to D2D resource allocation, assuming perfect CSI,
in a game-theoretic framework. In [6], the authors employ a
Stackelberg game theoretic framework for joint power control
and channel allocation. The existence and uniqueness of the
Stackelberg equilibrium (subgame perfect Nash equilibrium)
is analyzed. The Stackelberg model is also used in [7] in which
the authors state that striving to reach a global maximum
is a difficult mathematical feat and is also computationally
complex. Thus, they propose a low complexity two-stage
algorithm that addresses the power and resource allocation
problems separately. However, the proposed algorithm is suboptimal. A reverse iterative combinatorial auction framework
is proposed in [8] but the iterative nature of the algorithm
increases the communication overhead.
With partial CSI, the application of game theory to D2D
resource allocation problems is limited. In [9], the authors
have proposed a distributed power allocation scheme which is
A. Motivation
Most of the previous works on D2D resource allocation
assume perfect channel state information (CSI) [1] - [3]. This
implies that the base station (BS) as a central resource allocation unit has knowledge of all the channel gains among the
different communicating entities. From this knowledge, the BS
can determine the achievable rates of all the communication
The authors are with Indian Institute of Technology Bombay, Mumbai, Maharashtra, 400076, India. Email: (anushreen,chaporkar,karandi)@ee.iitb.ac.in.
This work is funded by the Department of Electronics and Information
Technology (DeitY), Government of India through a grant on “5G Research
and Building Next Gen Solutions for Indian Market”.
1
11
00
00
11
modeled as a non-cooperative game. They have shown that a
unique Nash equilibrium exists. Though their proposed method
does not require global CSI, it requires the knowledge of
local CSI. With local exchange of information, the signaling
overhead increases considerably. Moreover, for multiple D2D
pairs, they have not addressed the existence of the Nash
equilibrium. In [10], based on the Stackelberg game model,
a distributed iterative scheme of resource allocation is proposed. However, due to the iterative nature of their proposed
algorithm, signaling overhead is substantial. The authors of
[11] propose a centralized graph based resource allocation
scheme and model the power control problem as an exact
potential game. They show that although a pure strategy Nash
equilibrium maximizes the potential function but it does not
necessarily imply that the social utility of the D2D players is
also maximized.
gcB
c
gdB
gcd
gd
Fig. 1: Resource reuse between a CU and a D2D pair.
fast fading with mobility of CUs. With this randomness
into consideration, we have proposed two heuristic utility
calculation methods which achieve good performance. In
the field of D2D communications, ours is the first resource
allocation algorithm that aims at maximizing the social
utility of all the D2D players with partial CSI.
C. Contributions
In this paper, we take a departure from the usual notion
of selfish players and employ game theoretic principles such
that every player adjusts its action so as to obtain an allocation
strategy that maximizes the social utility or the sum throughput
of the D2D players. Moreover, it is also ensured that the
signal-to-interference-plus-noise-ratio (SINR) requirements of
the CUs are satisfied. We assume that the CUs have already
been allocated resources by the BS. The D2D players take their
actions in a distributed manner and transmit it to the BS, based
on which the BS decides on the final allocation that maximizes
the social utility of the D2D players. This allocation is nothing
but an assignment or mapping of CUs’ resources to the D2D
players. A feasible allocation is one which does not hamper
CU communications. Our work is motivated by [12], in which
the authors consider the optimal association problem of users
to base stations. They consider a game theoretic model and
propose a distributed algorithm to maximize the social utility
of all the users.
The main contributions of our work are as follows:
• Our work differs from [12] in that our algorithm is not
a completely distributed algorithm. In [12], due to a lack
of supervision, multiple players choose the same BS. This
results in a decrease in the network throughput. However,
this problem does not arise in our case because of the
intervention of the BS which allocates orthogonal resources
to the D2D players. Thus, multiple players cannot share
the same CU’s resources, otherwise it results in interference
among the D2D players which would decrease the D2D network’s throughput. Moreover, since the BS assigns orthogonal resources, the search space for the optimal strategy gets
restricted. Hence, the rate of convergence of the algorithm
to the optimal strategy decreases in comparison to [12].
• In [12], the utility of a player is the rate that it observes
when it transmits using the resources of a CU. However,
in our paper, we propose a different way of designing the
utility which is based on a frame structure.
• Further, the authors of [12] have not considered any kind
of randomness in the system model, which is a challenging
problem. We have considered both slow fading as well as
The organization of this paper is as follows. In Section II
we present the system model which comprises of an underlay
D2D network and the game theoretic modeling of the D2D
resource allocation problem. In Section III, we propose an
optimal resource allocation algorithm for the D2D players.
We next discuss how the system state process is a finite state
discrete time Markov chain (DTMC) in Section IV. In Section
V, we prove the optimality of our proposed algorithm. We
present an example to demonstrate its optimality in Section
VI. Next, in Section VII we consider both slow fading and fast
fading with CU mobility and propose two heuristic algorithms.
We verify the performance of our proposed algorithms through
simulations in Section VIII. In Section IX, we summarize our
findings.
II. S YSTEM M ODEL
A. Network Model
We consider a single cell scenario with a BS and NC CUs in
uplink transmission mode. We assume that the BS has already
allocated resources to these CUs. Consider ND D2D pairs
which have to be allocated the resources of these CUs. We
denote the BS by B, a CU by c and a D2D pair by d. As
shown in Fig. 1, the channel gain between a CU c to the BS
B is denoted by gcB and the channel gain between the D2D
transmitter d to the BS B is given by gdB . We assume that both
these channel gains are known at the BS. Similarly, for a D2D
pair d, the channel gain between its transmitter and receiver
is given by gd . The channel gain between a CU c and the
D2D receiver is given by gcd . We assume that a D2D receiver
does not know the channel gains gd and gcd . We consider the
channel to exhibit pathloss.
Let a CU’s transmit power be PC and that of a D2D user’s
be PD . Given that a D2D user d is allocated the resources of
a CU c, the SINR of the CU c at the BS is given by,
γc =
2
PC gcB
,
PD gdB + N0
(1)
where N0 is the average noise power. In case a D2D transmitter’s interference power at the BS is very high when it is
reusing the resources of the CU c and its SINR γc becomes
less than a threshold γtgt , then it is not allocated this CU’s
resources. Then, this allocation is infeasible.
III. R ESOURCE A LLOCATION A LGORITHM
The resource allocation algorithm that we propose is a
learning algorithm that guarantees convergence to an optimal
action profile provided interdependence is ensured in the game.
Interdependence implies that the utility of every player is
affected by the choice of actions of other players. This in
turn means that the change of utility of a player affects the
utility of every other player.
B. Game Theoretic Model
Let G be a strategic form game with ND D2D players
constituting a set ND = {d1 , d2 , ... , dND }. Each D2D player
d can choose an action from a set of finite actions Ad . Let
D
the joint action set be A = AN
and the utility function be
d
Ud : A → R.
As per the LTE standard, time is divided into subframes
of 1 ms duration. We define a frame to be a time window
consisting of ND successive subframes. A subframe is indexed
by ñ while a frame by n. At the start of a frame n, the
action taken by every D2D player d is to select a list ld (n)
and transmit it to the BS. The list contains K possible CUs
drawn from a set NC = {c1 , c2 , ... , cNC } of available CUs
to form a K-tuple. Hence, the total number of lists that a
D2D player can generate is NCPK = L. This set of L lists
constitutes the action set Ad of a D2D player. Let the action
profile be l = (l1 , l2 , ... , lND ) ∈ A. We denote l−d to be
the action profile of the D2D players other than the D2D
player d. Therefore, the action profile can also be written as
(ld , l−d ). The utility of a D2D player can therefore be written
as Ud (l) = Ud (ld , l−d ).
In each subframe ñ of a frame, as per the D2D players’ lists,
the BS decides on which CU’s resources are to be allocated
to the D2D players. Using these resources, each D2D player
then starts transmitting. A D2D player d then observes a rate
of rd (ñ). We define the utility of a D2D player as follows.
Definition 3. (Interdependence) A game G is interdependent
if for every action profile l ∈ A and for every proper subset of
players N ⊂ ND , there exists a player d ∈
/ N and a choice
|N |
0
0
of actions lN
∈ Ad such that rd (lN
, l−N ) 6= rd (lN , l−N ).
We consider the internal state variables of a D2D player
to be its list ld (n), its utility ud (n) and another variable
called its mood md (n), based on which a D2D player decides
on its action in every frame n. The mood can take two
values: content (C) or discontent (D). The resource allocation
algorithm which we will discuss next consists of the following
stages: 1) list selection of a D2D player, 2) the BS’s resource
allocation rule 3) utility calculation for a D2D player and 4) its
mood calculation method (refer to Algorithm 1).
A. List Selection
The list selection for every D2D player is done on a frame
by frame basis. The mood of a D2D player in the previous
frame helps it to select its list in the present frame. Let be
the exploration/experimentation rate, such that > 0 and k is
a constant such that k > ND [13]. A D2D player whose mood
is content can either decide to retain (exploit) its previous list
with probability 1 − k or select (explore) a new list with
probability k /(L − 1). It is more likely to retain its previous
list rather than exploring other lists if the exploration rate k is
less than 1 − 1/L. Thus, k is an important design parameter
that decides on the frequency of exploring versus exploiting
lists. On the other hand, a discontent D2D player explores all
the L possible lists with an equal probability of 1/L. Steps
1-6 of Algorithm 1 demonstrate the list selection method for
a D2D player d. Once this is over for all the D2D players,
every D2D player transmits its list to the BS. The BS then
allocates the resources of the CUs to the D2D players as per
the following allocation algorithm.
Definition 1. (Utility) The utility rd (n) of a D2D player d,
calculated at the end of a frame n is the average of the rates
obtained by it over all the sub-frames of a frame and is given
by,
rd (n) =
ND
1 X
rd (ñ).
ND
(2)
ñ=1
We normalize the utility rd (n) such that it is strictly bounded
between 0 and 1, [13].
Definition 2. (Social Utility) The social utility is the sum of
utilities of the D2DP
players obtained in every frame n and is
given by Wl (n) = d∈ND rd (n).
B. Allocation Algorithm at the BS
At the beginning of every frame, the lists of all the D2D
players are available at the BS. The BS follows two round
robin (RR) sequences to prioritize the D2D players in every
subframe and also across a frame. It first orders the D2D
players in a RR sequence, RR seq = d1 , d2 , ... , dND . This is
followed in the first subframe of a frame as shown in Fig. 2.
In the next subframe, the RR sequence starts from 2 such that
RR seq = d2 , ... , dND , d1 and so on till the last subframe
ND when the RR sequence starts from dND , as seen in Fig.
2. Thus, the selection of the first player changes as per this
RR sequence in every subframe of a frame.
In every subframe, the BS ensures orthogonal allocation of
Our objective is to maximize the social utility (sum throughput) of the D2D players in every frame n, over all possible
action profiles l, to get the optimal action profile l∗ (n), subject
to a rate constraint of rtgt ,
X
l∗ (n) = arg max
rd (n),
l
s.t.
rd (n) ≥ rtgt ,
d∈ND
∀d ∈ ND .
(3)
Thus, the optimal action profile l∗ (n) is the set of socially
optimal actions (Pareto efficient) taken by the D2D players
such that the sum throughput of the D2D players is maximized.
3
Algorithm 1 Resource Allocation Algorithm
Frame n
RR in subframes
ñ
d1
d2
dND
d2
d3
d3
d4
d1
dND
d1
d2
d1
d2
1:
2:
3:
4:
5:
dND −1 dND
RR across a frame
l1 (n)
c5
lND (n)
c8
l2 (n)
c5
l3 (n)
c1
c4
c1
c5
c3
c6
c2
c3
c7
6:
27:
28:
29:
Allocation Algorithm at the BS (l(n))
Initialize RR seq ← [d1 d2 ... dND ]
for all ñ = 1 to ND do
for all j = 1 to ND do
index ← 1
d ← RR seq[index]
Select D2D player d’s list
list index ← 1
while list index 6= list length + 1 do
c ← ld (n, list index)
if c’s resources are not allocated to players before
d in ñ then
Assign c’s resources to d
Break
else
list index ← list index + 1
end if
end while
index ← index + 1
if γc < γtgt then
c’s resources are not assigned to d
end if
end for
Left shift RR seq by 1
end for
30:
Utility Calculation
PND (c, rd (ñ))
rd (ñ)
rd (n) ← N1D ñ=1
7:
8:
9:
CU selection
Fig. 2: RR sequencing and the selection of CUs from the lists
of all the D2D players in the first subframe of a frame.
10:
11:
12:
13:
14:
15:
CU resources as follows. It selects the first element (a CU)
of a D2D player’s list and allocates the CU’s resources to it
only if it is not allocated to the other players prior to it in the
RR sequence followed in that subframe (see Fig. 2). If this
element has already been assigned to another D2D player, the
BS selects the next element of the D2D player’s list till it
reaches the end of the list. In case all the elements of its list
are already assigned to the other players, it is not allocated
any resource.
The RR across the subframes and way in which the utility
is calculated enforces interdependence among the players as
the priority of picking up a D2D player’s list shifts from one
D2D player to the next in each subframe. We explain this
by considering the following example. As seen from Fig. 2,
consider the first D2D player d1 of the first subframe, which
gets the first priority. Its selection of a CU, c5 from its list
l1 (n) and in turn its observed rate is not affected by the choice
of CUs (or lists) of other players because it gets the first
precedence over others in the RR sequence followed in that
subframe. Thus, its rate does not depend on the actions (lists)
of others. As per Definition 3, interdependence should ensure
that a player’s utility depends on the actions of others. Now,
with RR across the frame, in the second subframe, the priority
shifts to the next D2D player d2 . Then the priority of d1
becomes ND . Its CU selection and hence its rate gets affected
by the choice of CUs (or lists) of other players appearing
before it in the RR sequence. Thus, when its utility is finally
calculated as an average of the rates obtained by it in every
subframe, the utility depends on the rate it achieves in every
subframe and since the rate depends on the lists of other
players, the utility depends on the actions of all the players.
This is true for the other players also.
List Selection (md (n − 1), ld (n − 1))
if md (n − 1) == C then
k
, ∀ i ∈ Ad \ld (n − 1)
ld (n) ← i, w.p. L−1
ld (n) ← ld (n − 1), w.p. 1 − k
else
ld (n) ← i, w.p. L1 , ∀ i ∈ Ad
end if
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
31:
32:
33:
34:
35:
36:
37:
38:
Mood Calculation (sd (n − 1), ld (n), rd (n))
if md (n − 1) == C & [ld (n − 1), rd (n − 1)] ==
[ld (n), rd (n)] then
md (n) ← C
else if rd (n) ≥ rtgt then
md (n) ← C, w.p. 1−rd (n)
md (n) ← D, w.p. 1 − 1−rd (n)
else
md (n) ← D
end if
C. Allocation Feasibility Test
In order to ensure that CU communications are not hampered, in each subframe, the BS checks whether every CU’s
SINR decreases below γtgt or not, due to the interference from
the D2D transmitters which have been allocated the resources
of these CUs. For a CU whose SINR goes down below γtgt ,
the BS does not allocate its resources to the D2D player and
the D2D player’s rate rd (n) is zero in that subframe. After
this feasibility test is over for all the CUs, the BS conveys to
each D2D player which CU’s resources are allocated to it. The
mapping of CUs’ resources to the D2D players is termed as
the allocation profile. For example, from Fig. 2 if we assume
that the elements c5 , c1 , c3 , ... , c8 selected by the D2D players
from its lists l1 (n) to lND (n) pass the feasibility test, then the
allocation profile is (c5 , c1 , c3 , ... , c8 ).
4
prove that the system state space S is a subset of S 0 .
Let the cardinality of the space of l(n), r(n) and m(n)
be |A|, |U| and |M|. Since each of the ND D2D players can
choose any of the L possible lists, |A| = LND . We know
from the notion of interdependence that the rate rd (n) of a
D2D player d is a function of the action profile l(n), that
is, rd (n) = Ud (l). Since multiple action profiles (lists) can
generate the same allocation profile and thus the same utility
profile r(n), Ud (·) is a many to one function. Thus, |U| is less
than |A|. Since the mood of every D2D player can take two
possible values, |M| = 2ND . Moreover, those system states
in which at least one D2D player d’s rate rd (n) is less than
rtgt and yet its mood md (n) is content, cannot be part of the
state space S. Therefore, S is finite and is a subset of S 0 .
3) Since self transitions are possible from every system state,
the DTMC is aperiodic.
4) As the system states communicate with each other, they
form a single recurrent communication class. Thus, the DTMC
is irreducible.
D. Utility Calculation
In a subframe ñ, a D2D player transmits using the resources
of the CU allocated to it by the BS and observes a certain rate
rd (ñ). At the end of a frame, it calculates its utility rd (n) as
per Eqn. 2.
E. Mood Calculation
The mood of a D2D player is determined (refer to lines 3138 of Algorithm 1) at the end of each frame from its previous
state, its present list and utility. A content D2D player will
remain content, if its present configuration does not not change
with respect to its previous one, that is, (ld (n), rd (n)) =
(ld (n − 1), rd (n − 1)). However this condition is violated if
ld (n) 6= ld (n − 1) or rd (n) 6= rd (n − 1) or both. We next
explain these cases. Suppose, in the present frame n, a content
D2D player d decides to explore other lists. Then, its list
ld (n) changes and ld (n) 6= ld (n − 1). Thus, its configuration
changes. If this player decides to retain its previous list
ld (n) = ld (n − 1), its present utility rd (n) can still change
with respect to its previous utility rd (n−1). This is so because
if some other player changes its list, then it would change the
utility of this player because of interdependence. Therefore, as
its configuration changes, the chances of becoming discontent
or content depends on its present utility rd (n). If rd (n) is
greater than the minimum required target rate rtgt , it becomes
content with probability 1−rd (n) . The chances of becoming
content are more than that of becoming discontent, if rd (n)
is sufficiently high. If rd (n) is less than rtgt , then it becomes
discontent with probability (w.p.) one.
On the other hand, for a discontent D2D player d to become
content, it has to explore all the lists uniformly randomly. If
its rate rd (n) is greater than rtgt , it becomes content with
probability 1−rd (n) or it will remain discontent.
Since the DTMC is aperiodic and irreducible, it has a
unique stationary distribution π on the state space S. Let us
denote the transition probability matrix by P . We consider
this DTMC P , to be the perturbed version of the process P 0
with equal to zero and is a regular perturbation of P 0 [13],
[14]. The transitions in P 0 occur in P with high probabilities
but with a very small probability some transitions in P occur
which would not have occurred in P 0 . For all s1 , s2 ∈ S,
the state transitions Ps1 s2 approach Ps01 s2 as tends to zero,
i.e, lim→0 Ps1 s2 = Ps01 s2 . We are interested in finding those
states of P called the stochastically stable states in which
the system spends a large fraction of its time that maximizes
the social utility, with the D2D players content and their rate
constraints satisfied.
Theorem 1. The stochastically stable states of a regular
perturbed DTMC are the states s∗ ∈ S, which satisfy the
following conditions:
1) TheP
action profile l(n) should maximize the social utility
Wl = d∈ND rd (l) while satisfying the rate constraints of all
the D2D players.
2) For each D2D player d, its rate rd (n) should be aligned to
the action profile l(n) of all the D2D players in the system,
that is, rd (n) = Ud (l).
3) In a stochastically stable state s∗ , all the D2D players must
be content.
IV. S YSTEM S TATE P ROCESS
We define the present state sd (n) of a D2D player d as
a 3-D tuple, sd (n) = (ld (n), rd (n), md (n)). The collection
of states of all the D2D players constitute the system state
s(n) = (s1 (n), ... , sND (n)). Let the action profile in every
frame n be l(n) = (l1 (n), ... , lND (n)), the utility profile be
r(n) = (r1 (n), ... , rND (n)) and the mood profile be m(n) =
(m1 (n), ... , mND (n)).
Lemma 1. The system state process s(n) is a finite state,
aperiodic and irreducible DTMC.
Proof. 1) A D2D player’s utility rd (n) depends on the lists of
the other players. The selection of the present list ld (n) of a
D2D player in turn depends on its previous mood md (n). Its
present mood md (n) is also determined from its previous state.
Therefore, its present state sd (n) = (ld (n), rd (n), md (n))
depends on its previous state. Hence, the system state process
s(n) is a DTMC.
2) Let us rearrange s(n) in terms of the action, utility and
mood profiles as s(n) = (l(n), r(n), m(n)). Consider the
sets A, U and M to contain all possible values of l(n), r(n)
and m(n). Let the set S 0 = A × U × M be the state space of
all possible combinations of l(n), r(n) and m(n). We next
V. P ROOF OF T HEOREM 1
In order to determine the stochastically stable states we
define a few terms as follows.
Definition 4. (Resistance of Transition) Let Ps1 s2 be the
transition probability from the system state s1 to s2 . If
Ps1 s2 > 0 for some , then there exists a unique real number
rs1 s2 ≥ 0 called the resistance
of transition from s1 to s2
Ps s
1 2
such that 0 < lim→0 rs1 s2 < ∞. The resistance rs1 s2 is
zero if Ps1 s2 > 0.
5
D0
Definition 5. (i-Tree) In a directed graph G, an i-tree Ti is
a spanning tree such that there is exactly one directed path
from every vertex j to i, such that j 6= i.
x
Definition 6. (Resistance of an i-Tree) The resistance rTi of
an i-tree TP
i is the sum of resistances of its edges and is given
Ti
.
by rTi = j rji
y
Definition 7. (Stochastic Potential) In a directed graph G,
the stochastic potential γi of vertex i is the minimum i-tree
resistance among the resistances of all possible i-trees.
C
z
0
C0
Fig. 3: Graph G to determine γC 0 and γD0 .
Definition 8. (Stochastically Stable States) The stochastically
stable states of a perturbed DTMC P are the states which are
contained in the recurrence classes of P 0 with the minimum
stochastic potential [14].
process P and determine the stochastic potential of these recurrence classes. The transitions among the recurrence classes
arise due to the perturbation . Let the stochastic potential
of C 0 and D0 be γC 0 and γD0 respectively. The minimum
resistance of transition from C 0 → D0 , D0 → C 0 and
C 0 → C 0 can be determined as follows.
1) The transition from C 0 to a state in D0 can occur when
at least one content player decides to explore with probability
k /(L−1) and becomes discontent. Thus, content D2D players
become discontent. Therefore, the minimum resistance of
transition from C 0 to D0 is k.
2) Similarly, the transition from a system state in D0 to
C 0 occurs when each D2D player explores lists uniformly
randomly and becomes content with a probability of 1−rd (n) .
Thus,
the minimum resistance of transition from D0 to C 0 is
P
d∈ND (1 − rd (n)).
3) The transition from a content state to another content state
in C 0 , can occur when a content D2D player explores and its
utility rd (n) is greater than rtgt . Thus, the minimum resistance
of transition in this case is k + mind∈ND (1 − rd (n)).
Let us consider the system state space S to consist of three
recurrence classes: D0 and two singleton classes in the set C 0 .
We construct a directed graph G with these recurrence classes
as its vertices and denote them by x, y and z. Fig. 3 shows the
transitions or edges between the different recurrence classes.
We now identify the different recurrence classes of P 0 .
Lemma 2. The recurrence classes of P 0 are: 1) the class D0
consisting of all the system states in which every D2D player
is discontent and 2) the set of all singleton classes C 0 = {C 0 },
where in each system state C 0 every D2D player is content.
Proof. 1) In a system state of D0 , all the D2D players are
discontent and they start exploring lists uniformly randomly.
As their actions change, the action profile also changes.
This results in a state transition. Since is equal to zero
(refer Algorithm 1), irrespective of the fact whether its utility
rd (n) ≥ rtgt , a discontent D2D player remains discontent with
probability one. Thus, when the process enters a system state
in D0 , it can only transition to one of the states in D0 as it
cannot transition to any of the content states in C 0 . Hence, D0
is a recurrence class of P 0 .
2) Let us consider a state in C 0 in which every D2D player
is content. Referring to Algorithm 1, if is zero, the D2D
players retain their previous lists in the present frame with
probability one. The action profile of the system therefore
remains the same over frames. Due to RR allocation at the BS,
the set of allocation profiles of a frame also remain the same
over all the frames and thus the utility profile also remains
the same. As there is no change in any of the D2D player’s
configuration, that is, (ld (n−1), rd (n−1)) == (ld (n), rd (n)),
every content D2D player remains content. Thus, a system
state in C 0 maintains its state because the action profile, the
utility profile and the mood profile remains the same. These
are therefore absorbing states and are therefore recurrent.
Next, let us consider a system state s ∈ S in which some of
the D2D players are content while others are discontent. Every
discontent D2D player selects its list uniformly randomly and
still remains discontent with probability one since is zero.
Due to interdependence enforced by the BS, the utility of a
content D2D player gets affected due to the change of actions
of the discontent players. Thus, the content D2D player’s
utility changes, even though its list remains the same. Since its
configuration changes, it becomes discontent with probability
one. Gradually, all the content D2D players become discontent.
Thus, the state s0 transitions to one of the discontent states of
D0 . Hence, all such system states s0 ∈ S are transient.
0
Lemma 3. The stochastic potential
P of a recurrence class C
0
is given by γC 0 = k(|C | − 1) + d∈ND (1 − rd (n)).
Proof. From graph G of Fig. 3, there are three possible i-trees,
T1 , T2 and T3 that are rooted at vertex y. We represent them by
a set of directed edges as follows: T1 = {(x, y), (z, y)}, T2 =
{(x, z), (z, y)} and T3 = {(z, x), (x, y)}. The minimum resistances ofP
these three i-trees are rT1 = rT2 =
Pk+mind∈ND (1−
rd (n))+ d∈ND (1−rd (n)) and rT3 = k+ d∈ND (1−rd (n)).
Thus, the minimum resistance i-tree is T3 with resistance rT3 ,
which is the stochastic potential of C 0 . If the total number of
0
states in C 0 is |C 0 |, then the
P stochastic potential of C is given
0
by γC 0 = k(|C | − 1) + d∈ND (1 − rd (n)).
Lemma 4. The stochastic potential of the recurrence class
D0 is given by γD0 = k|C 0 |.
Proof. From graph G, we note that there are three possible i-trees rooted at the vertex x, T1 = {(y, x), (z, x)},
T2 = {(z, y), (y, x)} and T3 = {(y, z), (z, x)}. The minimum
resistance of these i-trees are, rT1 = 2k and rT2 = rT3 =
With perturbation , we now consider the regular perturbed
6
dR 2
c2
dT1
00
11
00
11
00
11
dT3
d
1
0
0
c3 1
dR 3
Sum throughput of D2D players in bps/Hz
d
dT2
1
0
0
1
c1
d
dR 1
Fig. 4: Example demonstrating the optimal mapping of resources between the CUs and the D2D pairs.
1.4
ǫ=0.5, k=11
1.2
X: 1256
Y: 1.224
1
0.8
0.6
0.4
0.2
0
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
Frames
Fig. 5: Sum throughput of the D2D players over frames for
the concept illustration example.
2k + mind∈Nd (1 − rd (n)). Thus, T1 is the i-tree with the
minimum resistance rT1 , which is the stochastic potential of
D0 . When the number of states in C 0 is |C 0 |, the stochastic
potential of D0 is given by γD0 = k|C 0 |.
BS with respect to its receiver. The distance between a D2D
transmitter and receiver can be calculated from Fig. 4 and is
17.43 m. Let PD = PC be 10 mW, the path loss exponent be
2 and the average noise power N0 be 0.1 mW. Let γtgt be
equal to 0 dB.
3) Analysis: Since the distances of the three CUs to the BS are
R1 , the powers received from them at the BS are the same.
The interference power received from the D2D transmitters
at the BS is the same for all the D2D players because they
are all equidistant from the BS. Hence, the SINRs of all the
CUs are the same and we assume these to be higher than γtgt .
Therefore, irrespective of the mapping of the resources of CUs
to the D2D players, every allocation profile is feasible.
Now the best CU for a D2D player is one which is
farthest from a D2D player’s receiver. If CU c1 ’s resources
are allocated to D2D pair d3 , then CU c2 ’s resources can be
allocated to either d1 or d2 . If c2 ’s resources are allocated
to d2 , then d2 faces the worst case interference from c2 .
Thus, c3 ’s resources have to be allocated to d1 . Clearly, this
allocation is not socially optimal because while d1 and d3 get
their best choices of CUs, d2 gets the worst CU. Thus, c3 ’s resources should be allocated to d2 and c2 ’s resources should be
allocated to d1 . Therefore, a1 = {(d1 , c2 ), (d2 , c3 ), (d3 , c1 )} is
an optimal allocation profile. Note that the allocation profile
a2 = {(d1 , c3 ), (d2 , c1 ), (d3 , c2 )} is also optimal. With allocation profile a1 = {(d1 , c2 ), (d2 , c3 ), (d3 , c1 )}, the rates of the
D2D players are calculated as rd1 = 0.4076, rd2 = 0.4076
and rd3 = 0.4089 bps/Hz. Thus, the utility profile r(n) is
(0.4076, 0.4076, 0.4089). Therefore, the maximum value of
the sum throughput or the social utility is the sum of these
rates which is 1.224 bps/Hz.
4) Verification: We simulate our proposed algorithm with the
above topology and parameter setting. We set = 0.5, k = 11
and obtain a plot of the sum throughput of the D2D players
versus the number of frames as shown in Fig. 5. We observe
that the maximum value of the sum throughput of the D2D
players is 1.224 bps/Hz with the D2D players content. This
matches our theoritical calculation. Two of the optimal action
profiles are as follows: l1∗ (n) = ([2 3]T , [3 1]T , [1 2]T ) and
l2∗ (n) = ([3 1]T , [1 3]T , [2 1]T ). When these lists are transmitted to the BS at the start of the frame, as per RR it determines
The stochastic potential of C 0 is therefore less than that of
D0 , i.e., γC 0 < γD0 . Thus, the stochastically stable states
are present in the recurrence classes of C 0 which have the
minimum stochastic potential.
The optimal action profile l∗ (n) is the one that minimizes
γC 0 and is given by,
X
l∗ (n) ∈ arg min k(|C 0 | − 1) +
(1 − rd (n)),
l∈A
d∈ND
which is equivalent to maximizing the sum throughput,
X
l∗ (n) ∈ arg max
rd (n).
l∈A
d∈ND
This completes the proof.
VI. C ONCEPT I LLUSTRATION
We demonstrate through the following example how the
social optimum maximizes the social utility of the D2D players
where the social optimum is the optimum allocation profile and
the social utility is the sum throughput of the D2D players.
The social optimum maximizes the social utility such that at
least one of the player’s utility is the best with none of the
other players’ utilities any worse off either. We consider a
simple topology to illustrate how the optimal allocation profile
can be determined. In order to theoretically calculate it we
assume that all the channel gains are known, from which we
can determine the value of the maximum sum throughput of
the D2D players. Next, we show through simulations that our
proposed algorithm also gives the same value. We are then able
to identify the actions/lists that the D2D players select which
gives this maximum value of the sum throughput. From this
information, we can then determine the stochastically stable
states of the system.
1) Topology: Consider a topology with three CUs c1 , c2 and
c3 positioned in a semi-circle of radius R1 as shown in Fig.
4. Three D2D pairs d1 , d2 and d3 are positioned in an outer
semi-circle of radius R2 .
2) Parameter Setting: We set R1 = 50 m and R2 = 100 m.
Every D2D transmitter subtends an angle of 10 degree at the
7
Algorithm 2 Threshold Based Utility Calculation
for all ñ = 1 to ND do
ad (ñ) ← c
end for
4: nd (c) ←
nd (c) + 1
5: ra (c) ← 1 − n 1(c) ra (c) +
d
resources. In a continuum of subframes, a CU c’s resource can
be allocated several times to a D2D player d, though it may not
be allocated exactly in two consecutive subframes. However,
every time that it is allocated this CU, it observes a change in
its rate as a result of the CU’s mobility and the channel. Let
us say that in the first frame of length ND of a superframe,
a D2D player is allocated the resources of different CUs. It
stores this sequence of CUs in a vector ad of length ND . In a
superframe, every time a D2D player d observes a certain rate
with the CU c, it calculates a running average of all these rates
using Monte Carlo (MC) averaging and stores it in a vector
ra of length NC at its cth index. It also stores the number
of times that the CU c’s resources are allocated to the D2D
player d in a vector nd of length NC at its cth index. This
is required for the MC averaging of the rates that it observes
with this CU over time. This averaging is also done for all the
other CUs whose resources are allocated to it. It also maintains
a vector ud of length NC which stores the average rates of
ra in it in the first superframe (initially). Then in the next
superframe as ra gets populated, an absolute difference of the
elements of both the vectors ud and ra is calculated. If any of
the resultants obtained exceed a threshold ∆, then its indices
are mapped to the utility vector ud which is now updated with
the corresponding values of ra . Thus, the decision to update
the entries of the utility vector ud with that of ra depends on
∆.
At the end of every superframe, a D2D player retrieves the
rates from the vector ud corresponding to the indices which
are the contents of the vector ad . It then sums up these rates
and divides it by ND which is the utility rd (n̂) of the D2D
player d at the end of every superframe n̂. The mood of the
D2D player is then determined as per the mood calculation
method of Algorithm 1 from its utility, the list selected by it
in the present superframe and its previous state.
1:
2:
3:
1
nd (c) rd (ñ)
2
if ñ == mND
then
for all c = 1 to NC do
if ud (c) == 0 or abs(ud (c) − ra (c)) > ∆ then
ud (c) ← ra (c)
end if
end for
rd (n̂) ← sum(ud (ad ))/ND
13: end if
6:
7:
8:
9:
10:
11:
12:
the optimal allocation profiles as {(d1 , c2 ), (d2 , c3 ), (d3 , c1 )}
and {(d1 , c3 ), (d2 , c1 ), (d3 , c2 )}. Note that these allocation
profiles match our theoretically obtained ones. Thus, the
optimal action profile, the utility profile and the mood profile
constitute two stochastically stable states.
VII. FADING AND CU M OBILITY
When the CUs are mobile and their location information is
unavailable, devising efficient resource allocation algorithms
for the D2D players is a challenging task. The mobility of a
CU is characterized by its velocity and direction of movement.
Therefore, the distance of a mobile CU to the D2D player’s
receiver changes with time. We consider slow fading due to
shadowing as well as fast fading due to multi-path propagation.
Thus, the D2D player’s SINR and consequently its rate varies
over subframes when it is allocated this CU’s resources. Now
with the same allocation profile or the same action profile,
an infinite set of utility profiles are generated. The DTMC
becomes infinite and all the states become transient. Therefore
the steady state distribution π does not exist. Thus, Algorithm
1 fails to converge. We devise two novel utility calculation
methods and modify Algorithm 1 such that on an average
the system performance does not degrade. The main concept
behind these methods is to detect a significant change in the
network topology or the channel variation that results in a
change in the allocation profile of the D2D players. We next
discuss these two methods.
B. Intercept-Slope Based Utility Calculation
This method is non-parameterized unlike the threshold
based algorithm which depends on ∆ and is more robust
(refer Algorithm 3). As in Algorithm 1, the list selection for
every D2D player occurs at the beginning of each frame while
its mood calculation occurs at the end of it. As discussed
in the previous method, a D2D player can be allocated a
CU’s resources multiple times over subframes. A D2D player
stores its previous rates observed with every CU and the
corresponding time instants in two vectors rp and tp of length
NC each. In a subframe, if it is allocated a CU c, then it
retrieves the value of the previous rate that it has observed with
this CU and the previous time stamp from the vectors rp and
tp respectively, from their cth index. From these values and the
rate rd (ñ) that it observes with this CU and the time stamp
ñ, it calculates the intercept-by-slope ratio corresponding to
the cth CU. This is stored in the cth index of another vector
ba of length NC . This computation is done in every subframe
corresponding to the CU whose resources are allocated to it.
It then sorts this vector ba containing these intercept-by-slope
ratios in descending order and stores it in z2 . This is compared
with the sorted order obtained in the previous subframe, stored
A. Threshold Based Utility Calculation
In this method, we modify Algorithm 1 so that a frame is
repeated ND times to form a superframe which is indexed by
n̂. The list selection by every D2D player and their mood calculation occurs at epochs which are multiples of a superframe
instead of a frame. As per Algorithm 1 the list generation
method for every D2D player and the allocation method at
the BS remains the same. The utility calculation method will
now differ, which we explain next (refer Algorithm 2).
In any subframe ñ, two pieces of information are available
to the D2D player d: 1) the cth CU whose resources are
allocated to it and 2) its present rate rd (ñ) using this CU’s
8
1:
2:
3:
4:
5:
6:
7:
8:
9:
Sum throughput of D2D players in Mbps
Algorithm 3 Intercept-Slope Based Utility Calculation
(b, a) ← Linear regression (rp (c), tp (c), ñ, rd (ñ))
ba (c) ← b/a
z2 (ñ) ← sort(ba )
rp (c) ← rd (ñ)
tp (c) ← ñ
if ud (c) == 0 or z2 (ñ) 6= z1 (ñ − 1) then
ud (c) ← rd (ñ)
end if
1
wd (ñ) ← 1 − ñ wd (ñ − 1) + ñ1 ud (c)
if ñ == ND then
rd (n) ← wd (ñ)
12: end if
Sum throughput of D2D players in Mbps
10:
11:
ǫ=0.5, k=11
ǫ=0.6, k=13
40
30
20
10
0
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
is PC = 250 mW and that of a D2D player is PD = 1 mW. We
consider the LTE pathloss model, P L = 128.1 + 37.6 log (d).
The UE and BS noise figures are 9 dB and 5 dB respectively.
We assume that each CU is allocated one physical resource
block (PRB) of 180 kHz bandwidth in each subframe. The
SINR target for the CUs is γtgt = 0 dB. The minimum rate
rtgt for each D2D player is 50 % of the rates observed by them
in their initial positions. The utilities of all the D2D players
are normalized by a factor such that these lie in the range of
0 to 1. We set the length of the lists, K to 3.
30
25
20
15
10
5
0
1000
50
Fig. 7: Sum throughput of the D2D players over superframes
with the threshold based utility calculation method.
35
500
ǫ=0.5, k=11, ∆=0.1
ǫ=0.6, k=13, ∆=0.1
Superframes
40
0
60
1500
2000
2500
3000
3500
4000
4500
5000
Frames
Fig. 6: Sum throughput of the D2D players over frames for
Algorithm 1.
A. Without Fading and Mobility of CUs
Fig. 6 demonstrates the performance of Algorithm 1 with
two different values of and k which are initially set to 0.5
and 11. As we increase them to 0.6 and 13, the exploration
rate k increases and the D2D players explore other lists more
often. This results in their individual utilities to change more
often due to a change in the action profile. This is reflected in
the fluctuations observed in the sum throughput of the D2D
players. The average sum throughput of the D2D players is
approximately 18 Mbps.
in another vector z1 .
A change in the sorting orders indicate that the D2D
player’s observed rate with the CU c’s resources in the present
subframe has either increased or decreased with respect to the
rates of the other CUs. This implies that the topology of the
network or the channel could have changed significantly. Then
the D2D player’s present rate rd (ñ) with the CU c is updated
in the utility vector ud of length NC at its cth index. If the
sorting order remains the same, it discards the present rate.
The vector z2 is then updated with the new sorting order. It
then retrieves from the cth index of the vector ud , the rate
corresponding to the CU c. Over subsequent subframes of the
frame, it continues to retrieve values from the indices of ud
corresponding to the CUs allocated to it in every subframe. It
uses these values to calculate a running average (MC) wd (ñ)
till the end of the frame n, which is the utility rd (n).
B. With Fading and Mobility of CUs
We model slow fading with lognormal random variables of
standard deviation 8 dB and for fast fading we model the
channel gains as independent and exponentially distributed
random variables with mean 1. The CU mobility is modeled
as per the Random Way Point (RWP) model. The random
direction in which a CU travels is uniformly distributed in the
interval [0 2π]. We assume a constant pedestrian speed of 1
m/s for all the CUs.
When we incorporate the threshold based utility calculation
method in Algorithm 1, we observe the following results with
∆ equal to 0.1. As shown in Fig. 7, from a typical sample
path realization of the sum throughput of the D2D players over
superframes we observe that the average sum throughput of the
D2D players is approximately 18 Mbps which is comparable
to the optimal algorithm.
With the intercept-slope based method, which does not
require any predefined parameter unlike the threshold based
VIII. R ESULTS
In this section, we verify the performances of the our proposed algorithms. The simulation parameters are as follows.
The macro cell radius is 250 m in which the CUs (NC = 10)
and the D2D transmitters (ND = 10) are uniformly distributed.
We assume the range of D2D communications to be 50
m. A D2D receiver is uniformly distributed around a D2D
transmitter within this range. Note that the same topology is
used for all the simulation results. The transmit power of a CU
9
Sum throughput of D2D players in Mbps
60
ǫ=0.5, k=11
ǫ=0.6, k=13
50
40
30
20
10
0
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
Frames
Fig. 8: Sum throughput of the D2D players over frames with
the intercept-slope based utility calculation method.
method, we observe from Fig. 8 that the algorithm achieves a
more robust performance as compared to the performance of
the threshold based method. It shows less fluctuations before
settling to a constant sum throughput. As we increase and
k from 0.5 and 11 respectively to 0.6 and 13, we observe
that the sum throughput of the D2D players fluctuates, as they
explore other lists more frequently. Further, the average sum
throughput of the D2D players is also around 18 Mbps.
IX. C ONCLUSIONS
We have proposed an optimal resource allocation algorithm
for D2D players in a game theoretic framework that ensures
that the social utility is maximized, while ensuring that the
communications of CUs are not hampered. We explain the
notion of stochastically stable states and prove that it is in
these system states that a D2D network’s sum throughput is
maximized. Our algorithm results in fast convergence to these
states because the BS allocates resources orthogonally to the
D2D players. We also propose two novel learning algorithms
that show good performance in a fast fading channel with CU
mobility.
R EFERENCES
[1] J. Wang, D. Zhu, C. Zhao, J. C. F. Li and M. Lei, “Resource Sharing of
Underlaying Device-to-Device and Uplink Cellular Communications,”
IEEE Commun. Lett., vol. 17, no. 6, pp. 1148-1151, June 2013.
[2] D. Feng, L. Lu, Y. Y. Wu, G. Y. Li, G. Feng and S. Li, “Device-toDevice Communications Underlaying Cellular Networks,” IEEE Trans.
Commun., vol. 61, no. 8, pp. 3541-3551, Aug. 2013.
[3] W. Zhao and S. Wang, “Resource Allocation for Device-to-Device
Communication Underlaying Cellular Networks: An Alternating Optimization Method,” IEEE Commun. Lett., vol. 19, no. 8, pp. 1398-1401,
Aug. 2015.
[4] S. Sesia, I. Toufik and M. Baker, “LTE-The UMTS Long Term Evolution:
From Theory to Practice,” 2nd ed., New York, NY, USA: Wiley, 2011.
[5] G. Hardin, “The Tragedy of the Commons”, Science, vol. 162, no. 3859,
pp. 1243-1248, Dec. 1968.
[6] F. Wang, L. Song, Z. Han, Q. Zhao and X. Wang, “Joint Scheduling and
Resource Allocation for Device-to-Device Underlay Communication,” in
Proc. WCNC, pp. 134-139, Apr. 2013.
[7] X. Chen, R. Q. Hu and Y. Qian, “Distributed Resource and Power
Allocation for Device-to-Device Communications Underlaying Cellular
Network,” in Proc. Globecom, pp. 4947-4952, Dec 2014.
[8] C. X. L. Song, Z. Han, Q. Zhao, X. Wang, X. Cheng and B. Jiao,
“Efficiency Resource Allocation for Device-to-Device Underlay Communication Systems: A Reverse Iterative Combinatorial Auction Based
Approach,” IEEE J. Sel. Areas. Commun., vol. 31, no. 9, pp. 348-358,
Sept. 2013.
10
[9] H. H. Nguyen, M. Hasegawa and W. J. Hwang, “Distributed Resource
Allocation for D2D Communications Underlay Cellular Networks,”
IEEE Commun. Lett., vol. 20, no. 5, pp. 942-945, May 2016.
[10] R. Yin, C. Zhong, G. Yu, Z. Zhang, K. K. Wong and X. Chen, “Joint
Spectrum and Power Allocation for D2D Communications Underlaying
Cellular Networks,” IEEE Trans. Veh. Tech., vol. 65, no. 4, pp. 21822195, Apr. 2016.
[11] S. Maghsudi and S. Stanczak, “Hybrid Centralized-Distributed Resource
Allocation for Device-to-Device Communication Underlaying Cellular
Networks,” IEEE Trans. Veh. Tech., vol. 65, no. 4, pp. 2481-2495, Apr.
2016.
[12] M. Singh and P. Chaporkar, “An Efficient and Decentralised User Association Scheme for Multiple Technology Networks,” in Proc. WiOpt,
pp. 460-467, May 2013.
[13] J. R. Marden, H. P. Young and L. Y. Pao, “Achieving Pareto Optimality
through Distributed Learning,” in Proc. CDC, pp. 7419-7424, Dec. 2012.
[14] H. P. Young, “The Evolution of Conventions,” Econometrica, vol. 61,
no. 1, pp. 57-84, Jan. 1993.
| 7cs.IT
|
ANNALS OF THE UNIVERSITY OF ORADEA
FASCICLE OF TEXTILES, LEATHERWORK
SEAM PUCKERING OBJECTIVE EVALUATION METHOD FOR
SEWING PROCESS
BRAD Raluca1, HĂLOIU Eugen2, BRAD Remus2
1
Lucian Blaga University of Sibiu, Romania, Department of Industrial Machinery and Equipment, Faculty of Engineering,
B-dul Victoriei 10, 550024 Sibiu, Romania, E-Mail: [email protected]
2
Lucian Blaga University of Sibiu, Romania, Department of Computer Science and Electrical Engineering, Faculty of
Engineering, B-dul Victoriei 10, 550024 Sibiu, Romania, E-Mail: [email protected], [email protected]
Corresponding author: Brad, Raluca, E-mail: [email protected]
Abstract: The paper presents an automated method for the assessment and classification of puckering defects
detected during the preproduction control stage of the sewing machine or product inspection. In this respect, we
have presented the possible causes and remedies of the wrinkle nonconformities. Subjective factors related to the
control environment and operators during the seams evaluation can be reduced using an automated system
whose operation is based on image processing. Our implementation involves spectral image analysis using
Fourier transform and an unsupervised neural network, the Kohonen Map, employed to classify material
specimens, the input images, into five discrete degrees of quality, from grade 5 (best) to grade 1 (the worst).
Key words: Seams, puckering, image processing, neural network, Discrete Fourier Transform.
1. QUALITY AND AUTOMATION IN THE TEXTILE INDUSTRY
The textile industry is one of the traditional and dynamic sectors where the customer quality
requirements are constantly changing as a result of trends in fashion and the development of
production tools. In order to satisfy clients’ demands, the variables that affect product quality must be
kept under control during the production cycle: design, manufacturing, delivery and maintenance.
The evaluation process of a sewn product relating to appearance and performance have to rely
on a holistic perspective that includes both fabrics and sewing threads assessment, but also consider
their interactions during sewing, wearing and maintenance of the product. Throughout the
manufacturing process, the woven, non-woven or knitted fabrics are controlled from two to more than
six times in order to detect the defects which may occur, followed by their classification and if
possible, remedying. A good compatibility between sewing thread and materials will influence the
product quality and productivity. Otherwise, during the sewing process, the fabric is damaged or the
machine stops at unanticipated time intervals [1].
After execution, a correct seam need to be smooth and flat, without puckering, tuck
developing or seam damage, having an appropriate behavior during pressing and cleaning. There are
several standard test methods for evaluating the interactions between the threads and fabrics after the
execution of the seams and stitching. Some procedures assess seam strength, slippage, failure, damage,
pucker and jamming before and after cleaning [2]. ISO 7770, AATCC 88 B and AATCC 143 [3,4,5]
standards use sets of images and rating scales in order to evaluate the appearance of seams, using
grades from 1 (worse) to 5 (best quality seam). Inspectors should compare the stitching samples with
the standard images, in different environments, which cause subjective results (figure 1).
Fig. 1: Standard images for the subjective assessment of seam pucker [4].
Due to long reaction time and fatigue of the human operator, an automatic inspection would
be able to verify and classify the seams with a much higher speed and would eliminate the subjective
factor. The system can be used both in the pre-production stage, for sewing machines adjustment, and
also in product inspection. The ability to recognize flaws and stop production immediately after the
occurrence of the defect is important for clothing manufacturers.
The automatic control system may use different technologies for image acquisition, containing
mechanical components, computer software, video cameras, lighting and video equipment. In
particular, an automatic defect detection is be based on an electro-optical device for fabric surface
inspection using a two-dimensional scanner of the warp and weft directions, or on a complex of video
camera and uniform illumination source connected to video acquisition system [6].
2. THE SEAM PUCKERING DEFECT: CAUSES AND REMEDIES
The seam puckering phenomenon is defined as a local defect of a clothing item in the form of
large ridges of material beside the seam and is considered one of the most serious defects in garment
manufacturing [7]. The complete elimination of this flaw during pressing operations is almost
impossible, and therefore in practice, it is often to accept a lower grade as normal. Consequently, the
objective assessment of seam puckering is essential, such that the final product will be acceptable to
the client [8].
Seam pucker may also be described as a differential shrinkage that arises throughout the seam
line and is caused by the instability of the seam. Usually, wrinkling appears due to improper selection
of stitching parameters and material properties, leading to an inequality of fabrics lengths that are
sewn together and affecting the appearance characteristics. In serious cases, puckering can appear as a
wave coming from the seams. Since sewing operation is subjected to excessive tensions, it produces a
stretching of sewing threads, leading to an extension over the whole surfaces of the fabrics.
Although sewing threads have usually a controlled elasticity, they are overstretched when
large tensions are implied in process. After sewing, the threads tend to relax, trying to return to the
original length. As the stitches shrink, wrinkles appear in the material and can not be detected
immediately, but in a later stage. The threads used in sewn products must also have a good stability to
washing and ironing, as differential shrinkage between sewing thread and fabrics may cause
puckering.
Other fabrics characteristics which affect seam stability and puckering are fabric density and
structure. By stitching, threads snap the fiber material into a new position, inserting into material
structure and tending to change it. This phenomenon is more obvious when the fabric is made of fine,
dense and low resistance to compression yarns. In the case of differences in fiber composition, fabrics
structure, extensibility and stability, puckers may occur due to feeding failure. In order to avoid this
situation, it is necessary to adjust the presser foot pressure to a minimum value. The feeding systems
used in stitching control are: a positive or negative differential conveyor, consisting of two teeth
elements in front and behind the needle with adjustable amplitude, and a simple oriented tooth feed.
In order to reduce puckering, sewing machine and stitching parameters are adjusted. In the
case of the sewing machine, adjustments are made on the conveyor mechanism, tension control and
needle selection. Regarding the processing parameters, the stitching step should be as small as
possible, while the value of cutting and sewing angle must be correlated with yarn and fabric structure.
Using a similar fabric fiber composition thread with thermo stability, low elongation and recovery,
puckers can be decreasead or avoided.
2
ANNALS OF THE UNIVERSITY OF ORADEA
FASCICLE OF TEXTILES, LEATHERWORK
3. AN OBJECTIVE SEAM PUCKERING EVALUATION METHOD
Subjective evaluation methods have the disadvantage a higher assessment time, differences
between appraisals, partiality towards certain colors or models, and training needs. In the attempt of
objectification, two SP synthetic indicators may be used for seam pucker description: related to width
and to the length. [9] The hypothesis assumes that following wrinkled defect occurrence, the thickness
of two layers sewn assembly increases, while the length decreases, comparing to the initial length of
unraveled material. In these conditions, the following formula can be used:
t − 2t
l − ls
SP = s
*100 (%) or SP =
*100 (%)
(1)
ls
2t
where ts = seam thickness, t = fabric thickness, l = length of unraveled fabric, ls = length of sewn
assembly.
However, incoherence and time-consuming are noticed for these assessments.
Image processing techniques have been already applied in the textile industry. Research was
carried out to investigate the cross section of fibers, yarn structure, yarn thickness [10], texture fault
detection, seam pucker etc. Image processing is sometimes combined with a classification/recognition
step achieved using neural networks. A large review paper on this field, including many textile
applications, has been published by [11].
The assessment of seams is one of the research topics in the aim of textile industry
automation. The foundations of this field start with a geometric modeling of puckering and a review of
the methods and techniques available at that time, in the view of measurement [12]. In [13], the use of
a k nearest neighbor classifier achieved an 81% rate of successful classification rate compared with
human experts. This implementation is trying to improve a previous one from the same authors,
presented in [14]. A wavelet based detectors of surface smoothness or wrinkles and puckering defects
have been used by [15], joined with a 3D scanning system.
Our implementation was derived from [16] and involves a spectral image analysis using
Fourier transform, and an unsupervised neural network, Kohonen Map to classify material specimens,
which are the input images into five discrete degrees of quality, to grade 5 (best) to grade 1 (the
worst). A similar approach was proposed using fractal theory [17]. The learning and testing stages are
depicted in figure 2. The next paragraphs will present the basics of our processing scheme.
3.1. The Otsu Algorithm for image binarization
In image processing, the Otsu's method is used to perform image binarization, in order to
separate objects from background. The algorithm assumes that the image to be segmented contains
two classes of pixels (bimodal histogram) and calculates the optimum threshold to separate the classes.
The frequencies of grey levels are established and the probabilities for each possible threshold level
are computed. The variance of the pixels levels on either side of the threshold will be estimated, both
for the object and the background region. Among all possible values, the threshold that minimizes the
inter-class variance will be selected, being defined as a weighted sum of the two classes’ variances
[18]:
σ w2 (t ) = w1 (t )σ 12 (t ) + w2 (t )σ 22 (t )
(2)
where w1,2 represents the probabilities of the two classes separated by threshold t, and σ 12, 2 symbolizes
the variance of the two classes. The average of each class is given by the weighted average of
frequency intensities, with L the number of grey levels:
t
L
ip(i)
ip(i )
µ1 (t ) =
and
µ 2 (t ) =
(3)
i =1 w1 (t )
i = t +1 w2 (t )
The individual variances of the classes are:
t
L
p (i)
p(i )
σ 12 (t ) = [i − µ1 (t )]2
and
σ 22 (t ) =
[i − µ 2 (t )]2
(4)
w1 (t )
w2 (t )
i =1
i =t +1
∑
∑
∑
∑
image database
image acquisition
grayscale transform
noise reduction
grayscale transform
noise reduction
Otsu binarization
Otsu binarization
Discrete Fourier
Transform
Discrete Fourier
Transform
image characteristics
image characteristics
Kohonen network
training
Kohonen network
recognition
puckering
classification
last
training
image?
no
yes
training
completed
Fig. 2: The proposed image processing framework.
3.2. The Fourier transform
The Fourier Transform is an important tool in image processing, representing the input spatial
domain image in the frequency field and therefore, each point in the output representing a particular
frequency contained in the image. It is used in a wide range of applications, such as image analysis,
image filtering, image reconstruction and image compression. In the context of the present paper, the
Discrete Fourier Transform (DFT) will be used. The DFT is actually a sampled Fourier transform and
will not contain all frequencies being present in the image, but only a set of samples large enough to
describe the spatial domain image, as the number of frequencies corresponds to the number of pixels
[19].
For an NxN image, the DFT is given by:
km ln
−i 2π
+
N N
1 N −1 N −1
F (k , l ) = 2
f ( m.n)e
(5)
N m =0 n = 0
where f(m, n) is the spatial domain image. Each pixel value is multiplied with an exponential term,
called the basis function, and summed over the domain. F(0, 0) represents the continuous component
of the image (ie the mean gray level), while F(N-1, N-1) characterizes the larger frequency. In the
same way, the DFT image can be retransformed in the spatial domain as an inverse transform:
∑∑
4
ANNALS OF THE UNIVERSITY OF ORADEA
FASCICLE OF TEXTILES, LEATHERWORK
km ln
−i 2 π
+
N N
1 N −1 N −1
f (m, n) = 2
F ( k .l )e
(6)
N k =0 l =0
1
where 2 is a normalization term.
N
The double sum in equation 1 must be computed for each point (k, l). However, due to the
separabilty of the Fourier transform, the above formula can be expressed in two N series of onedimensional transforms, decreasing the number of computations:
∑∑
1
F (k , l ) =
N
where
N −1
∑ P(k .n)e
−i 2π
ln
N
(7)
n= 0
−i 2 π
km
N
1 N −1
P( k , n) =
f ( m, n)e
(8)
N m =0
The DFT produces an output which can be displayed using two images: either the real and
imaginary part or the magnitude and phase. In image processing, the amplitude values are usually
employed since information about the geometrical structure of the image space is carried out.
Nevertheless, in order to compute the inverse DFT after processing the frequency domain, both
amplitude and phase array of values must be preserved and stored on wide representation variables
[19].
∑
3.3. The Kohonen Map
The Kohonen networks are also known as “self-organizing maps”, a special type of artificial
neural networks trained using unsupervised learning in order to produce a discrete representation of
the input space. Self-organizing maps are different from other neural networks as they use a function
to preserve neighborhoods topological properties of the input space. Therefore, Kohonen maps are
used to approximate the distribution of input vectors, the dimensionality reduction while maintaining
data in the vicinity or for clustering.
Kohonen maps are organized in two layers. The first level of the network is the input layer,
while the second is the competitive level, organized as an array. The two layers are fully
interconnected; each input node is connected to all nodes in the competitive layer, as in figure 3 [20].
Fig. 3: The basic structure of the Kohonen network [20].
Each connection has an associated weight. In the initial state, the network weights have
random values in the range [0, 1]. The input pattern is an n dimensional vector. As a result, the input
models are uniformly distributed over a square. The first step in the operation of Kohonen maps is to
compute a matching value for each node in the competitive layer. This value measures the extent to
which the weight of each node corresponds to the input node. The matching value is the distance
between X (inputs) and W (weights) vectors:
X − Wi =
∑(x
j
− wij ) 2
(9)
j
The node with the best match wins the competition. This node is set as follows:
X − Wc = min i { X − Wi }
(10)
where c is the best node.
After the winning node is identified, the next step is to identify its neighborhood, as those
nodes in a square centered on the winning node. The weights are updated for all neurons that are in the
neighborhood of the winning node. The update equation is:
α (e j − u ij ), if node i is in the neigborhoo d
(11)
∆uij =
0 else
where α is the learning step, initially (α0) in the range [0.2, 0.5]. During the learning process, the value
is decreased over iterations until 0, using:
t
α i = α 0 1 −
(12)
T
where t is the current iteration and T the total number of iterations.
Besides the learning step, the neighborhood size must also be updated over the iterations in the
learning process, as in equation 13.
c−d < x <c+d
(13)
c−d < y <c+d
with c representing the winning node, d the distance from c to the edge of the neighborhood, and x, y
the node coordinates. Usually, the initial value of d is chosen as half to three quarters the size of the
competitive layer and updated using:
t
d i = d 0 1 −
(14)
T
where t is the current iteration and T, the total number of iterations.
In summary, the basic rules of the Kohonen network can be described as follows [20]:
§ Locate the unit in the competitive layer whose weight fits best with the input,
§ Update the weights of the selected unit and its neighbors, thus increasing the level of
matching,
§ Gradually decreases the neighborhood size and adjust the weights during the iterations
of the learning process
4. RESULTS
The input image is represented by a sewn sample image with puckering. At the beginning of
the learning or testing framework, a preprocessing block was introduced, due to his main role in
filtering. As in seam defects detection, color is not necessary; the acquired images are transformed to
grayscale (figure 4).
Fig. 4: Grayscale image of a sewn presenting puckering.
6
ANNALS OF THE UNIVERSITY OF ORADEA
FASCICLE OF TEXTILES, LEATHERWORK
The result of a DFT is represented by the amplitude and phase of frequency components of the
input image. The amplitude shows how much of a particular frequency is being presented in the
image, while the phase represents how the signal is offset from the origin, or particularly, how much
the sinus wave is shifted to the left or the right, as shown in figure 5.
a)
b)
Fig. 5: The amplitude a) and phase b) result of the above image DFT.
Our classification approach of puckering is based on series of samples that have been
categorized accordance to some human experts. In this sense, the puckering features present in the
learning and test images have been pre-classified using the seam puckering quality standard. In the
case of the proposed framework, the image features are extracted using spectral analysis by Fourier
transform and the results stored in an array. We have down-sampled the corresponding array, in order
to fit into the low resolution Kohonen Map, of a 100x100 size. The network training stage will consist
in presenting five input vectors (derived from the down-sampled arrays), representing the puckering
grades. The network learning stops when the error falls below a value that is very close to zero. The
puckering classification consists in providing an input vector derived from the image supposed to be
classified. A scalar product between the input values vectors and the weighted training images is
computed. The result will be assigned to one of the five classes of which the input image belongs.
This method requires the image acquisition of samples in special conditions. An oblique light
is applied on the material and, due to the wrinkled material appearance, the shadows will highlight the
nonconformities of the fabric surface.
Fig. 6: The training set.
The investigation was performed on 26 samples of 100% polyester plain woven fabric, with
21 cm-1 warp and weft density, and 170 grams/sqm specific weight. Specimens were cut into 10x10
cm sizes and midst sewn in pairs using 301 type lockstitch and a 4 cm-1 seam density. Two colors of
sewing thread were used: white to the needle and red to the bobbin, with 80 Nm sewing thread count,
100% PES.
The network was trained using the features images from figure 6, extracted using the DFT. In
order to test the functionality of the network another set of 21 test images was used (figure 7). These
images were originally classified subjectively using visual information. The classification results using
the Kohonen network are as follows:
• 71.42% of the images were classified correctly.
• 28.58% of the images were misclassified.
Fig. 7: The testing set.
5. CONCLUSIONS
The applicability domain of the presented framework is in the textile industry, namely seams
quality control and sewn assemblies classification in terms of visual quality. Various algorithms have
been applied on the acquired images in order to improve their processability. Currently, defect
detection is done using input images containing only horizontal seams. For further development of the
application, fault detection will be completed on several types of stitches and seam shapes, for
example, circular.
The image classification is done using visual information, based on subjective standard
images. We have created five quality classes using puckering images, used in the training of the neural
network. By processing a large number of samples, classification could be redefined and improved.
The detection of seams defects will be investigated using three images categories, containing
horizontal seams from clothing manufacturing, airbag assembly and automotive upholstery.
REFERENCES
[1] S. J. Kadolph. (2007). “Quality Assurance for Textiles and Apparel”, Fairchild
Publications, New York
[2] ASTM D1683 / D1683M - 11a, “Standard Test Method for Failure in Sewn Seams of
Woven Apparel Fabric”
[3] ISO 7770:2009, “Textiles - Test method for assessing the smoothness appearance of seams
in fabrics after cleansing”
8
ANNALS OF THE UNIVERSITY OF ORADEA
FASCICLE OF TEXTILES, LEATHERWORK
[4] AATCC 88B-2011, “Smoothness of Seams in Fabrics after Repeated Home Laundering”
[5] AATCC 143-2011, “Appearance of Apparel and Other Textile End Products after
Repeated Home Laundering”
[6] M. J. Ferreira, A. Fernandes, J.A. Campos Neves, J. L. Monteiro, "Automatic Visual
Inspection on Shoe and Textile Industries", IASTED CGIM'98, Halifax, Canada, 1998. Available
http://www.dei.uminho.pt/~aparicio/pdf/Apa-CGIM98-s.pdf
[7] The Shorter Oxford Dictionary. (1967), Oxford Univ. Press, U.K. vol.II, 3rd ed, pp. 1614.
[8] K. L. Mak and Wei Li, “Objective Evaluation of Seam Pucker on Textiles by Using SelfOrganizing Map”, IAENG International Journal of Computer Science, 35:1, IJCS_35_1_07. Available
http://www.iaeng.org/IJCS/issues_v35/issue_1/IJCS_35_1_07.pdf
[9] S. Hati and B.R. Das. „Seam Pucker in Apparels: A Critical Review of Evaluation
Methods”, Asian Journal of Textile, vol. 1, pp. 60-73, 2011
[10] B. K. Behera. (2004). “Image-Processing in Textiles”, Textile Progress, vol. 35, 2-4
[11] R. Chattopadhyay and A. Guha. (2004). “Artificial Neural Networks: Applications to
Textiles”, Textile Progress, vol. 35:1, pp. 1-46
[12] Chang Kyu Park and Tae Jin Kang, “Objective Evaluation of Seam Pucker Using
Artificial Intelligence: Part I: Geometric Modeling of Seam Pucker”, Textile Research Journal, vol.
69, pp. 735-742, October 1999.
[13] I. G. Mariolis & E. S. Dermatas, “Automated assessment of textile seam quality based on
surface roughness estimation”, Journal of Textile Institute, vol. 101:7, pp. 653-659, 2010
[14] I. G. Mariolis & E. S. Dermatas, “Automated quality control of textile seams based on
puckering evaluation”, 37th Int. Symp. on Nov. in Text., Ljubljana, Slovenia, June 2006.
[15] Tae Jin Kang, Soo Chang Kim, In Hwan Sul, Jae Ryoun Youn, Kwansoo Chung, “Fabric
Surface Roughness Evaluation Using Wavelet-Fractal Method: Part I: Wrinkle, Smoothness and Seam
Pucker”, Textile Research Journal, vol. 75, pp. 751-760, November 2005
[16] C. Bahlmann, G. Heidemann, H. Ritter, “Artificial Neural Networks for Automated
Quality Control of Textile Seams”, Pattern Recognition, vol. 32(6), pp. 1049-1060, 1999.
[17] T. Aibara, T. Mabuchi, O. Kenji, “Automatic evaluation of the appearance of seam
puckers on suits”, Proc. SPIE, vol. 3652: 110, 1999.
[18] http://en.wikipedia.org/wiki/Otsu's_method
[19] http://homepages.inf.ed.ac.uk/rbf/HIPR2/fourier.htm.
[20] S. Grossberg, “Neural Networks and Natural Intelligence”, Cambridge, MA: MIT Press,
1988.
| 5cs.CE
|
Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action
Recognition
Sijie Yan, Yuanjun Xiong, Dahua Lin
arXiv:1801.07455v2 [cs.CV] 25 Jan 2018
Department of Information Engineering, The Chinese University of Hong Kong
{ys016, dhlin}@ie.cuhk.edu.hk, [email protected]
Abstract
Dynamics of human body skeletons convey significant information for human action recognition. Conventional approaches for modeling skeletons usually rely on hand-crafted
parts or traversal rules, thus resulting in limited expressive
power and difficulties of generalization. In this work, we
propose a novel model of dynamic skeletons called SpatialTemporal Graph Convolutional Networks (ST-GCN), which
moves beyond the limitations of previous methods by automatically learning both the spatial and temporal patterns
from data. This formulation not only leads to greater expressive power but also stronger generalization capability. On two
large datasets, Kinetics and NTU-RGBD, it achieves substantial improvements over mainstream methods.
Figure 1: The spatial temporal graph of a skeleton sequence used
1
Introduction
Human action recognition has become an active research
area in recent years, as it plays a significant role in video
understanding. In general, human action can be recognized
from multiple modalities(Simonyan and Zisserman 2014;
Tran et al. 2015; Wang, Qiao, and Tang 2015; Wang et al.
2016; Zhao et al. 2017), such as appearance, depth, optical flows, and body skeletons (Du, Wang, and Wang 2015;
Liu et al. 2016). Among these modalities, dynamic human
skeletons usually convey significant information that is complementary to others. However, the modeling of dynamic
skeletons has received relatively less attention than that of
appearance and optical flows. In this work, we systematically study this modality, with an aim to develop a principled and effective method to model dynamic skeletons and
leverage them for action recognition.
The dynamic skeleton modality can be naturally represented by a time series of human joint locations, in the form
of 2D or 3D coordinates. Human actions can then be recognized by analyzing the motion patterns thereof. Earlier
methods of using skeletons for action recognition simply
employ the joint coordinates at individual time steps to form
feature vectors, and apply temporal analysis thereon (Wang
et al. 2012; Fernando et al. 2015). The capability of these
methods is limited as they do not explicitly exploit the spatial relationships among the joints, which are crucial for
understanding human actions. Recently, new methods that
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
in this work where the proposed ST-GCN operate on. Blue dots
denote the body joints. The intra-body edges between body joints
are defined based on the natural connections in human bodies.
The inter-frame edges connect the same joints between consecutive frames. Joint coordinates are used as inputs to the ST-GCN.
attempt to leverage the natural connections between joints
have been developed (Shahroudy et al. 2016; Du, Wang, and
Wang 2015). These methods show encouraging improvement, which suggests the significance of the connectivity.
Yet, most existing methods rely on hand-crafted parts or
rules to analyze the spatial patterns. As a result, the models
devised for a specific application are difficult to be generalized to others.
To move beyond such limitations, we need a new method
that can automatically capture the patterns embedded in the
spatial configuration of the joints as well as their temporal dynamics. This is the strength of deep neural networks.
However, as mentioned, the skeletons are in the form of
graphs instead of a 2D or 3D grids, which makes it difficult to use proven models like convolutional networks. Recently, Graph Neural networks (GCNs), which generalize
convolutional neural networks (CNNs) to graphs of arbitrary structures, have received increasing attention and successfully been adopted in a number of applications, such as
image classification (Bruna et al. 2014), document classification (Defferrard, Bresson, and Vandergheynst 2016), and
semi-supervised learning (Kipf and Welling 2017). However, much of the prior work along this line assumes a fixed
graph as input. The application of GCNs to model dynamic
graphs over large-scale datasets, e.g. human skeleton sequences, is yet to be explored.
In this paper, we propose to design a generic representation of skeleton sequences for action recognition by extending graph neural networks to a spatial-temporal graph
model, called Spatial-Temporal Graph Convolutional Networks (ST-GCN). As illustrated in Figure 1 this model is
formulated on top of a sequence of skeleton graphs, where
each node corresponds to a joint of the human body. There
are two types of edges, namely the spatial edges that conform to the natural connectivity of joints and the temporal
edges that connect the same joints across consecutive time
steps. Multiple layers of spatial temporal graph convolution
are constructed thereon, which allow information to be integrated along both the spatial and the temporal dimension.
The hierarchical nature of ST-GCN eliminates the need of
hand-crafted part assignment or traversal rules. This not only
leads to greater expressive power and thus higher performance (as shown in our experiments), but also makes it easy
to generalize to different contexts. Upon the generic GCN
formulation, we also study new strategies to design graph
convolution kernels, with inspirations from image models.
The major contributions of this work lie in three aspects:
1) We propose ST-GCN, a generic graph-based formulation
for modeling dynamic skeletons, which is the first that applies graph-based neural networks for this task. 2) We propose several principles in designing convolution kernels in
ST-GCN to meet the specific demands in skeleton modeling. 3) On two large scale datasets for skeleton-based action
recognition, the proposed model achieves superior performance as compared to previous methods using hand-crafted
parts or traversal rules, with considerably less effort in manual design. The code and models of ST-GCN are made publicly available1 .
construct the CNN filters on the spatial domain, by limiting
the application of each filter to the 1-neighbor of each node.
Skeleton Based Action Recognition. Skeleton and joint
trajectories of human bodies are robust to illumination
change and scene variation, and they are easy to obtain owing to the highly accurate depth sensors or pose estimation
algorithms (Shotton et al. 2011; Cao et al. 2017a). There
is thus a broad array of skeleton based action recognition
approaches. The approaches can be categorized into handcrafted feature based methods and deep learning methods. The first type of approaches design several handcrafted
features to capture the dynamics of joint motion. These
could be covariance matrices of joint trajectories (Hussein
et al. 2013), relative positions of joints (Wang et al. 2012),
or rotations and translations between body parts (Vemulapalli, Arrate, and Chellappa 2014). The recent success
of deep learning has lead to the surge of deep learning
based skeleton modeling methods. These works have been
using recurrent neural networks (Shahroudy et al. 2016;
Zhu et al. 2016; Liu et al. 2016; Zhang, Liu, and Xiao
2017) and temporal CNNs (Li et al. 2017; Ke et al. 2017;
Kim and Reiter 2017) to learn action recognition models in
an end-to-end manner. Among these approaches, many have
emphasized the importance of modeling the joints within
parts of human bodies. But these parts are usually explicitly assigned using domain knowledge. Our ST-GCN is the
first to apply graph CNNs to the task of skeleton based action recognition. It differentiates from previous approaches
in that it can learn the part information implicitly by harnessing locality of graph convolution together with the temporal
dynamics. By eliminating the need for manual part assignment, the model is easier to design and potent to learn better
action representations.
3
2
Related work
Neural Networks on Graphs. Generalizing neural networks to data with graph structures is an emerging topic
in deep learning research. The discussed neural network
architectures include both recurrent neural networks (Tai,
Socher, and Manning 2015; Van Oord, Kalchbrenner, and
Kavukcuoglu 2016) and convolutional neural networks
(CNNs) (Bruna et al. 2014; Henaff, Bruna, and LeCun 2015;
Duvenaud et al. 2015; Li et al. 2016; Defferrard, Bresson,
and Vandergheynst 2016). This work is more related to the
generalization of CNNs, or graph convolutional networks
(GCNs). The principle of constructing GCNs on graph generally follows two streams: 1) the spectral perspective,
where the locality of the graph convolution is considered
in the form of spectral analysis (Henaff, Bruna, and LeCun 2015; Duvenaud et al. 2015; Li et al. 2016; Kipf and
Welling 2017); 2) the spatial perspective, where the convolution filters are applied directly on the graph nodes and their
neighbors (Bruna et al. 2014; Niepert, Ahmed, and Kutzkov
2016). This work follows the spirit of the second stream. We
1
https://github.com/yysijie/st-gcn
Spatial Temporal Graph ConvNet
When performing activities, human joints move in small local groups, known as “body parts”. Existing approaches for
skeleton based action recognition have verified the effectiveness of introducing body parts in the modeling (Shahroudy
et al. 2016; Liu et al. 2016; Zhang, Liu, and Xiao 2017).
We argue that the improvement is largely due to that parts
restrict the modeling of joints trajectories within “local regions” compared with the whole skeleton, thus forming a hierarchical representation of the skeleton sequences. In tasks
such as image object recognition, the hierarchical representation and locality are usually achieved by the intrinsic properties of convolutional neural networks (Krizhevsky,
Sutskever, and Hinton 2012), rather than manually assigning object parts. It motivates us to introduce the appealing
property of CNNs to skeleton based action recognition. The
result of this attempt is the ST-GCN model.
3.1
Pipeline Overview
Skeleton based data can be obtained from motion-capture
devices or pose estimation algorithms from videos. Usually
the data is a sequence of frames, each frame will have a set
of joint coordinates. Given the sequences of body joints in
...
Action
Classification
Pose
Estimation
Input Video
Running
Class Score
ST-GCNs
Figure 2: We perform pose estimation on videos and construct spatial temporal graph on skeleton sequences. Multiple layers
of spatial-temporal graph convolution (ST-GCN) will be applied and gradually generate higher-level feature maps on the graph.
It will then be classified by the standard Softmax classifier to the corresponding action category.
the form of 2D or 3D coordinates, we construct a spatial
temporal graph with the joints as graph nodes and natural
connectivities in both human body structures and time as
graph edges. The input to the ST-GCN is therefore the joint
coordinate vectors on the graph nodes. This can be considered as an analog to image based CNNs where the input is
formed by pixel intensity vectors residing on the 2D image
grid. Multiple layers of spatial-temporal graph convolution
operations will be applied on the input data and generating
higher-level feature maps on the graph. It will then be classified by the standard SoftMax classifier to the corresponding
action category. The whole model is trained in an end-toend manner with backpropagation. We will now go over the
components in the ST-GCN model.
3.2
Skeleton Graph Construction
A skeleton sequence is usually represented by 2D or 3D coordinates of each human joint in each frame. Previous work
using convolution for skeleton action recognition (Kim and
Reiter 2017) concatenates coordinate vectors of all joints
to form a single feature vector per frame. In our work, we
utilize the spatial temporal graph to form hierarchical representation of the skeleton sequences. Particularly, we construct an undirected spatial temporal graph G = (V, E) on
a skeleton sequence with N joints and T frames featuring
both intra-body and inter-frame connection.
In this graph, the node set V = {vti |t = 1, . . . , T, i =
1, . . . , N } includes the all the joints in a skeleton sequence.
As ST-GCN’s input, the feature vector on a node F (vti ) consists of coordinate vectors, as well as estimation confidence,
of the i-th joint on frame t. We construct the spatial temporal graph on the skeleton sequences in two steps. First, the
joints within one frame are connected with edges according to the connectivity of human body structure, which is
illustrated in Fig. 1. Then each joint will be connected to
the same joint in the consecutive frame. The connections
in this setup are thus naturally defined without the manual
part assignment. This also enables the network architecture
to work on datasets with different number of joints or joint
connectivities. For example, on the Kinetics dataset, we use
the 2D pose estimation results from the OpenPose (Cao et al.
2017b) toolbox which outputs 18 joints, while on the NTURGB+D dataset (Shahroudy et al. 2016) we use 3D joint
tracking results as input, which produces 25 joints. The STGCN can operate in both situations and provide consistent
superior performance. An example of the constructed spatial temporal graph is illustrated in Fig. 1.
Formally, the edge set E is composed of two subsets,
the first subset depicts the intra-skeleton connection at each
frame, denoted as ES = {vti vtj |(i, j) ∈ H}, where H is
the set of naturally connected human body joints. The second subset contains the inter-frame edges, which connect the
same joints in consecutive frames as EF = {vti v(t+1)i }.
Therefore all edges in EF for one particular joint i will represent its trajectory over time.
3.3
Spatial Graph Convolutional Neural Network
Before we dive into the full-fledged ST-GCN, we first look
at the graph CNN model within one single frame. In this
case, on a single frame at time τ , there will be N joint nodes
Vt , along with the skeleton edges ES (τ ) = {vti vtj |t =
τ, (i, j) ∈ H}. Recall the definition of convolution operation on the 2D natural images or feature maps, which can
be both treated as 2D grids. The output feature map of a
convolution operation is again a 2D grid. With stride 1 and
appropriate padding, the output feature maps can have the
same size as the input feature maps. We will assume this
condition in the following discussion. Given a convolution
operator with the kernel size of K × K, and an input feature
map fin with the number of channels c. The output value for
a single channel at the spatial location x can be written as
fout (x) =
K X
K
X
fin (p(x, h, w)) · w(h, w),
(1)
h=1 w=1
where the sampling function p : Z 2 × Z 2 → Z 2 enumerates the neighbors of location x. In the case of image
convolution, it can also be represented as p(x, h, w) =
x + p0 (h, w). The weight function w : Z 2 → Rc provides
a weight vector in c-dimension real space for computing the
inner product with the sampled input feature vectors of dimension c. Note that the weight function is irrelevant to the
input location x. Thus the filter weights are shared everywhere on the input image. Standard convolution on the image domain is therefore achieved by encoding a rectangular
grid in p(x). More detailed explanation and other applications of this formulation can be found in (Dai et al. 2017).
The convolution operation on graphs is then defined by
extending the formulation above to the cases where the input features map resides on a spatial graph Vt . That is, the
t
feature map fin
: Vt → Rc has a vector on each node of
the graph. The next step of the extension is to redefine the
sampling function p and the weight function w.
Sampling function. On images, the sampling function
p(h, w) is defined on the neighboring pixels with respect
to the center location x. On graphs, we can similarly define the sampling function on the neighbor set B(vti ) =
{vtj |d(vtj , vti ) ≤ D} of a node vti . Here d(vtj , vti ) denotes the minimum length of any path from vtj to vti . Thus
the sampling function p : B(vti ) → V can be written as
p(vti , vtj ) = vtj .
(2)
In this work we use D = 1 for all cases, that is, the 1neighbor set of joint nodes. The higher number of D is left
for future works.
Weight function. Compared with the sampling function,
the weight function is trickier to define. In 2D convolution,
a rigid grid naturally exists around the center location. So
pixels within the neighbor can have a fixed spatial order. The
weight function can then be implemented by indexing a tensor of (c, K, K) dimensions according to the spatial order.
For general graphs like the one we just constructed, there is
no such implicit arrangement. The solution to this problem
is first investigated in (Niepert, Ahmed, and Kutzkov 2016),
where the order is defined by a graph labeling process in the
neighbor graph around the root node. We follow this idea to
construct our weight function. Instead of giving every neighbor node a unique labeling, we simplify the process by partitioning the neighbor set B(vti ) of a joint node vti into a fixed
number of K subsets, where each subset has a numeric label.
Thus we can have a mapping lti : B(vti ) → {0, . . . , K − 1}
which maps a node in the neighborhood to its subset label.
The weight function w(vti , vtj ) : B(vti ) → Rc can be implemented by indexing a tensor of (c, K) dimension or
w(vti , vtj ) = w0 (lti (vtj )).
(3)
subsets to the output. Substituting Eq. 2 and Eq. 3 into Eq. 4,
we arrive at
X
1
fout (vti ) =
fin (vtj ) · w(lti (vtj )). (5)
Zti (vtj )
vtj ∈B(vti )
It is worth noting this formulation can resemble the standard
2D convolution if we treat a image as a regular 2D grid. For
example, to resemble a 3 × 3 convolution operation, we have
a neighbor of 9 pixels in the 3 × 3 grid centered on a pixel.
The neighbor set should then be partitioned into 9 subsets,
each having one pixel.
Spatial Temporal Modeling. Having formulated spatial
graph CNN, we now advance to the task of modeling the
spatial temporal dynamics within skeleton sequence. Recall
that in the construction of the graph, the temporal aspect of
the graph is constructed by connecting the same joints across
consecutive frames. This enable us to define a very simple
strategy to extend the spatial graph CNN to the spatial temporal domain. That is, we extend the concept of neighborhood to also include temporally connected joints as
B(vti ) = {vqj |d(vtj , vti ) ≤ K, |q − t| ≤ bΓ/2c}. (6)
The parameter Γ controls the temporal range to be included
in the neighbor graph and can thus be called the temporal
kernel size. To complete the convolution operation on the
spatial temporal graph, we also need the sampling function,
which is the same as the spatial only case, and the weight
function, or in particular, the labeling map lST . Because the
temporal axis is well-ordered, we directly modify the label
map lST for a spatial temporal neighborhood rooted at vti to
be
lST (vqj ) = lti (vtj ) + (q − t + bΓ/2c) × K,
(7)
where lti (vtj ) is the label map for the single frame case at
vti . In this way, we have a well-defined convolution operation on the constructed spatial temporal graphs.
3.4
Partition Strategies.
Given the high-level formulation of spatial temporal graph
convolution, it is important to design a partitioning strategy
to implement the label map l. In this work we explore several
partition strategies. For simplicity, we only discuss the cases
in a single frame because they can be naturally extended to
the spatial-temporal domain using Eq. 7.
We will discuss several partitioning strategies in Sec. 3.4.
Spatial Graph Convolution. With the refined sampling
function and weight function, we now rewrite Eq. 1 in terms
of graph convolution as
fout (vti ) =
X
vtj ∈B(vti )
1
fin (p(vti , vtj )) · w(vti , vtj ),
Zti (vtj )
(4)
where the normalizing term Zti (vtj ) =| {vtk |lti (vtk ) =
lti (vtj )} | equals the cardinality of the corresponding subset.
This term is added to balance the contributions of different
Uni-labeling. The simplest and most straight forward partition strategy is to have subset, which is the whole neighbor
set itself. In this strategy, feature vectors on every neighboring node will have a inner product with the same weight
vector. Actually, this strategy resembles the propagation rule
introduced in (Kipf and Welling 2017). It has an obvious
drawback that in the single frame case, using this strategy
is equivalent to computing the inner product between the
weight vector and the average feature vector of all neighboring nodes. This is suboptimal for skeleton sequence classification as the local differential properties could be lost in
this operation. Formally, we have K = 1 and lti (vtj ) =
0, ∀i, j ∈ V .
(a)
(b)
(c)
(d)
0
1
1
1
1
0
1
Figure 3: The proposed partitioning strategies for constructing convolution operations. From left to right: (a) An example frame
of input skeleton. Body joints are drawn with blue dots. The receptive fields of a filter with D = 1 are drawn with red dashed
circles. (b) Uni-labeling partitioning strategy, where all nodes in a neighborhood has the same label (green). (c) Distance
partitioning. The two subsets are the root node itself with distance 0 (green) and other neighboring points with distance 1.
(blue). (d) Spatial configuration partitioning. The nodes are labeled according to their distances to the skeleton gravity center
(black cross) compared with that of the root node (green). Centripetal nodes have shorter distances (blue), while centrifugal
nodes have longer distances (yellow) than the root node.
Distance partitioning. Another natural partitioning strategy is to partition the neighbor set according to the nodes’
distance d(·, vti ) to the root node vti . In this work, because
we set D = 1, the neighbor set will then be separated into
two subsets, where d = 0 refers to the root node itself and
remaining neighbor nodes are in the d = 1 subset. Thus we
will have two different weight vectors and they are capable
of modeling local differential properties such as the relative
translation between joints. Formally, we have K = 2 and
lti (vtj ) = d(vtj , vti ) .
modeling the dynamics of these parts. In this sense, we add
a learnable mask M on every layer of spatial temporal graph
convolution. The mask will scale the contribution of a node’s
feature to its neighboring nodes based on the learned importance weight of each spatial graph edge in ES . Empirically
we find adding this mask can further improve the recognition performance of ST-GCN. It is also possible to have a
data dependent attention map for this sake. We leave this to
future works.
3.6
Spatial configuration partitioning. Since the body skeleton is spatially localized, we can still utilize this specific spatial configuration in the partitioning process. We design a
strategy to divide the neighbor set into three subsets: 1) the
root node itself; 2)centripetal group: the neighboring nodes
that are closer to the gravity center of the skeleton than the
root node; 3) otherwise the centrifugal group. Here the average coordinate of all joints in the skeleton at a frame is
treated as its gravity center. This strategy is inspired by the
fact that motions of body parts can be broadly categorized
as concentric and eccentric motions. Formally, we have
0 if rj = ri
lti (vt j) = 1 if rj < ri
(8)
2 if rj > ri
where ri is the average distance from gravity center to joint
i over all frames in the training set.
Visualization of the three partitioning strategies is shown
in Fig. 3. We will empirically examine the proposed partioning strategies on skeleton based action recognition experiments. It is expected that a more advanced partitioning
strategy will lead to better modeling capacity and recognition performance.
3.5
Learnable edge importance weighting.
Although joints move in groups when people are performing
actions, one joint could appear in multiple body parts. These
appearances, however, should have different importance in
Implementing ST-GCN
The implementation of graph-based convolution is not as
straightforward as 2D or 3D convolution. Here we provide
details on implementing ST-GCN for skeleton based action
recognition.
We adopt a similar implementation of graph convolution
as in (Kipf and Welling 2017). The intra-body connections
of joints within a single frame are represented by an adjacency matrix A and an identity matrix I representing selfconnections. In the single frame case, ST-GCN with the first
partitioning strategy can be implemented with the following
formula (Kipf and Welling 2017)
1
1
fout = Λ− 2 (A + I)Λ− 2 fin W,
(9)
P
ij
ij
where Λii =
j (A + I ). Here the weight vectors of
multiple output channels are stacked to form the weight matrix W. In practice, under the spatial temporal cases, we can
represent the input feature map as a tensor of (C, V, T ) dimensions. The graph convolution is implemented by performing a 1 × Γ standard 2D convolution and multiplies
the resulting tensor with the normalized adjacency matrix
1
1
Λ− 2 (A + I)Λ− 2 on the second dimension.
For partitioning strategies with multiple subsets, i.e., distance partitioning and spatial configuration partitioning, we
again utilize this implementation. But note now the adjacency matrix
P is dismantled into several matrixes Aj where
A + I = j Aj . For example in the distance partitioning
strategy, A0 = I and A1 = A. The Eq. 9 is transformed
into
fout =
X
− 12
− 21
Λj Aj Λj fin Wj ,
(10)
j
P
ik
where similarly Λii
j =
k (Aj )+α. Here we set α = 0.001
to avoid empty rows in Aj .
It is straightforward to implement the learnable edge importance weighting. For each adjacency matrix, we accompany it with a learnable weight matrix M. And we substitute the matrix A + I in Eq. 9 and Aj in Aj in Eq. 10 with
(A + I) ⊗ M and Aj ⊗ M, respectively. Here ⊗ denotes
element-wise product between two matrixes. The mask M
is initialized as an all-one matrix.
Network architecture and training. Since the ST-GCN
share weights on different nodes, it is important to keep the
scale of input data consistent on different joints. In our experiments, we first feed input skeletons to a batch normalization layer to normalize data. The ST-GCN model is composed of 9 layers of spatial temporal graph convolution operators (ST-GCN units). The first three layers have 64 channels for output. The follow three layers have 128 channels
for output. And the last three layers have 256 channels for
output. These layers have 9 temporal kernel size. The Resnet
mechanism is applied on each ST-GCN unit. And we randomly dropout the features at 0.5 probability after each STGCN unit to avoid overfitting. The strides of the 4-th and
the 7-th temporal convolution layers are set to 2 as pooling
layer. After that, a global pooling was performed on the resulting tensor to get a 256 dimension feature vector for each
sequence. Finally, we feed them to a SoftMax classifier. The
models are learned using stochastic gradient descent with a
learning rate of 0.01. We decay the learning rate by 0.1 after
every 10 epochs. To avoid overfitting, we perform two kinds
of augmentation to replace dropout layers when training on
the Kinetics dataset (Kay et al. 2017). First, to simulate the
camera movement, we perform random affine transformations on the skeleton sequences of all frames. Particularly,
from the first frame to the last frame, we select a few fixed
angle, translation and scaling factors as candidates and then
randomly sampled two combinations of three factors to generate an affine transformation. This transformation is interpolated for intermediate frames to generate a effect as if we
smoothly move the view point during playback. We name
this augmentation as random moving. Second, we randomly
sample fragments from the original skeleton sequences in
training and use all frames in the test. Global pooling at the
top of the network enables the network to handle the input
sequences with indefinite length.
4
Experiments
In this section we evaluate the performance of ST-GCN
in skeleton based action recognition experiments. We experiment on two large-scale action recognition datasets
with vastly different properties: Kinetics human action dataset (Kinetics) (Kay et al. 2017) is by far the
largest unconstrained action recognition dataset, and NTURGB+D (Shahroudy et al. 2016) the largest in-house cap-
tured action recognition dataset. In particular, we first perform detailed ablation study on the Kinetics dataset to examine the contributions of the proposed model components
to the recognition performance. Then we compare the recognition results of ST-GCN with other state-of-the-art methods and other input modalities. To verify whether the experience we gained on in the unconstrained setting is universal, we experiment with the constraint setting on NTURGB+D and compare ST-GCN with other state-of-the-art
approaches. All experiments were conducted on PyTorch
deep learning framework with 8 TITANX GPUs.
4.1
Dataset & Evaluation Metrics
Kinetics. Deepmind Kinetics human action dataset (Kay
et al. 2017) contains around 300, 000 video clips retrieved
from YouTube. The videos cover as many as 400 human action classes, ranging from daily activities, sports scenes, to
complex actions with interactions. Each clip in Kinetics lasts
around 10 seconds.
This Kinetics dataset provides only raw video clips without skeleton data. In this work we are focusing on skeleton
based action recognition, so we use the estimated joint locations in the pixel coordinate system as our input and discard
the raw RGB frames. To obtain the joint locations, we first
resize all videos to the resolution of 340 × 256 and convert
the frame rate to 30 FPS. Then we use the public available
OpenPose (Cao et al. 2017b) toolbox to estimate the location
of 18 joints on every frame of the clips. The toolbox gives
2D coordinates (X, Y ) in the pixel coordinate system and
confidence scores C for the 18 human joints. We thus represent each joint with a tuple of (X, Y, C) and a skeleton frame
is recorded as an array of 18 tuples. For the multi-person
cases, we select 2 people with the highest average joint confidence in each clip. In this way, one clip with T frames
is transformed into a skeleton sequence of these tuples. In
practice, we represent the clips with tensors of (3, T, 18, 2)
dimensions. For simplicity, we pad every clip by replaying
the sequence from the start to have T = 300. We will release
the estimated joint locations on Kinetics for reproducing the
results.
We evaluate the recognition performance by top-1 and
top-5 classification accuracy as recommended by the dataset
authors (Kay et al. 2017). The dataset provides a training set
of 240, 000 clips and a validation set of 20, 000. We train the
compared models on the training set and report the accuracies on the validation set.
NTU-RGB+D: NTU-RGB+D (Shahroudy et al. 2016) is
currently the largest dataset with 3D joints annotations for
human action recognition task. This dataset contains 56, 000
action clips in 60 action classes. These clips are all performed by 40 volunteers captured in a constrained lab environment, with three camera views recorded simultaneously.
The provided annotations give 3D joint locations (X, Y, Z)
in the camera coordinate system, detected by the Kinect
depth sensors. There are 25 joints for each subject in the
skeleton sequences. Each clip is guaranteed to have at most
2 subjects.
The authors of this dataset recommend two benchmarks:
1) cross-subject (X-Sub) benchmark with 40, 320 and
Baseline TCN
Local Convolution
Uni-labeling
Distance partitioning*
Distance Partitioning
Spatial Configuration
ST-GCN + Imp.
Top-1
20.3%
22.0%
19.3%
23.9%
29.1%
29.9%
30.7%
Top-5
40.0%
43.2%
37.4%
44.9%
51.3%
52.2%
52.8%
Table 1: Ablation study on the Kinetics dataset. The “STGCN+Imp.” is used in comparison with other state-of-theart methods. For meaning of each setting please refer to
Sec.4.2.
16, 560 clips for training and evaluation. In this setting the
training clips come from one subset of actors and the models
are evaluated on clips from the remaining actors; 2) crossview(X-View) benchmark 37, 920 and 18, 960 clips. Training clips in this setting come from the camera views 2 and
3, and the evaluation clips are all from the camera view 1.
We follow this convention and report the top-1 recognition
accuracy on both benchmarks.
4.2
Ablation Study
We examine the effectiveness of the proposed components
in ST-GCN in this section by action recognition experiments
on the Kinetics dataset (Kay et al. 2017).
Spatial temporal graph convolution. First, we evaluate
the necessity of using spatial temporal graph convolution
operation. We use a baseline network architecture (Kim and
Reiter 2017) where all spatial temporal convolutions are replaced by only temporal convolution. That is, we concatenate all input joint locations to form the input features at
each frame t. The temporal convolution will then operate on
this input and convolves over time. We call this model “baseline TCN”. This kind of recognition models is known to
work well on constraint dataset such as NTU-RGB+D (Kim
and Reiter 2017). Seen from Table 1, models with spatial temporal graph convolution, with reasonable partitioning strategies, consistently outperform the baseline model on
Kinetics. Actually, this temporal convolution is equivalent to
spatial temporal graph convolution with unshared weights
on a fully connected joint graph. So the major difference between the baseline model and ST-GCN models are the sparse
natural connections and shared weights in convolution operation. Additionally, we evaluate an intermediate model between the baseline model and ST-GCN, referred as “local
convolution”. In this model we use the sparse joint graph as
ST-GCN, but use convolution filters with unshared weights.
We believe the better performance of ST-GCN based models
could justify the power of the spatial temporal graph convolution in skeleton based action recognition.
Partition strategies In this work we present three partitioning strategies: 1) uni-labeling; 2) distance partitioning;
and 3) spatial configuration partitioning. We evaluate the
performance of ST-GCN with these partitioning strategies.
The results are summarized in Table 1. We observe that partitioning with multiple subsets is generally much better than
uni-labeling. This is in accordance with the obvious problem of uni-labeling that it is equivalent to simply averaging features before the convolution operation. Given this observation, we experiment with an intermediate between the
distance partitioning and uni-labeling, referred to as “distance partitioning*”. In this setting we bind the weights of
the two subsets in distance partitioning to be different only
by a scaling factor −1, or w0 = −w1 . This setting still
achieves better performance than uni-labeling, which again
demonstrate the importance of the partitioning with multiple
subsets. Among multi-subset partitioning strategies, the spatial configuration partitioning achieves better performance.
This corroborates our motivation in designing this strategy,
which takes into consideration the concentric and eccentric
motion patterns. Based on these observations, we use the
spatial configuration partitioning strategy in the following
experiments.
Learnable edge importance weighting. Another component in ST-GCN is the learnable edge importance weighting. We experiment with adding this component on the STGCN model with spatial configuration partitioning. This is
referred to as “ST-GCN+Imp.” in Table 1. Given the high
performing vanilla ST-GCN, this component is still able to
raise the recognition performance by more than 1 percent.
Recall that this component is inspired by the fact that joints
in different parts have different importances. It is verified
that the ST-GCN model can now learn to express the joint
importance and improve the recognition performance. Based
on this observation, we always use this component with STGCN in comparison with other state-of-the-art models.
4.3
Comparison with State of the Arts
To verify the performance of ST-GCN in both unconstrained and constraint environment, we perform experiments on Kinetics dataset (Kay et al. 2017) and NTURGB+D dataset(Shahroudy et al. 2016), respectively.
Kinetics. On Kinetics, we compare with three characteristic approaches for skeleton based action recognition. The
first is the feature encoding approach on hand-crafted features (Fernando et al. 2015), referred to as “Feature Encoding” in Table 2. We also implemented two deep learning
based approaches on Kinetics, i.e. Deep LSTM (Shahroudy
et al. 2016) and Temporal ConvNet (Kim and Reiter 2017).
We compare the approaches’ recognition performance in
terms of top-1 and top-5 accuracies. In Table 2, ST-GCN is
able to outperform previous representative approaches. For
references, we list the performance of using RGB frames
and optical flow for recognition as reported in (Kay et al.
2017).
NTU-RGB+D. The NTU-RGB+D dataset is captured in a
constraint environment, which allows for methods that require well stabilized skeleton sequences to work well. We
RGB(Kay et al. 2017)
Optical Flow (Kay et al. 2017)
Feature Enc. (Fernando et al. 2015)
Deep LSTM (Shahroudy et al. 2016)
Temporal Conv. (Kim and Reiter 2017)
ST-GCN
Top-1
57.0%
49.5%
14.9%
16.4%
20.3%
30.7%
Top-5
77.3%
71.9%
25.8%
35.3%
40.0%
52.8%
Table 2: Action recognition performance for skeleton based
models on the Kinetics dataset. On top of the table we list
the performance of frame based methods.
Lie Group (Veeriah, Zhuang, and Qi 2015)
H-RNN (Du, Wang, and Wang 2015)
Deep LSTM (Shahroudy et al. 2016)
PA-LSTM (Shahroudy et al. 2016)
ST-LSTM+TS (Liu et al. 2016)
Temporal Conv (Kim and Reiter 2017).
C-CNN + MTLN (Ke et al. 2017)
ST-GCN
X-Sub
50.1%
59.1%
60.7%
62.9%
69.2%
74.3%
79.6%
81.5%
X-View
52.8%
64.0%
67.3%
70.3%
77.7%
83.1%
84.8%
88.3%
Table 3: Skeleton based action recognition performance on
NTU-RGB+D datasets. We report the accuracies on both
the cross-subject (X-Sub) and cross-view (X-View) benchmarks.
also compare our ST-GCN model with the previous state-ofthe-art methods on this dataset. Due to the constraint nature
of this dataset, we do not use any data augmentation when
training ST-GCN models. We follow the standard practice in literature to report cross-subject (X-Sub) and crossview (X-View) recognition performance in terms of top1 classification accuracies. The compared methods include
Lie Group (Veeriah, Zhuang, and Qi 2015), Hierarchical
RNN (Du, Wang, and Wang 2015), Deep LSTM (Shahroudy
et al. 2016), Part-Aware LSTM (PA-LSTM) (Shahroudy et
al. 2016), Spatial Temporal LSTM with Trust Gates (STLSTM+TS) (Liu et al. 2016), Temporal Convolutional Neural Networks (Temporal Conv.) (Kim and Reiter 2017), and
Clips CNN + Multi-task learning (C-CNN+MTLN) (Ke et
al. 2017). Our ST-GCN model, with rather simple architecture and no data augmentation as used in (Kim and Reiter
2017; Ke et al. 2017), is able to outperform previous stateof-the-art approaches on this dataset.
Discussion. The two datasets in experiments have very
different natures. On Kinetics the input is 2D skeletons detected with deep neural networks (Cao et al. 2017a), while
on NTU-RGB+D the input is from Kinect depth sensor. On
NTU-RGB+D the cameras are fixed, while on Kinetics the
videos are usually shot by hand-held devices, leading to
large camera motion. The fact that the proposed ST-GCN
can work well on both datasets demonstrates the effectiveness of the proposed spatial temporal graph convolution operation and the resultant ST-GCN model.
We also notice that on Kinetics the accuracies of skeleton based methods are inferior to video frame based models (Kay et al. 2017). We argue that this is due to a lot of ac-
Method
Accuracy
RGB CNN
70.4%
Flow CNN
72.8%
ST-GCN
72.4%
Table 4: Mean class accuracies on the “Kinetics Motion”
subset of the Kinetics dataset. This subset contains 30 action classes in Kinetics which are strongly related to body
motions.
Single
Model
RGB TSN
X
Flow TSN
ST-GCN
X
X
Ensemble
Model
X
X
X
X
X
X
X
Acc(%)
70.3
51.0
30.7
71.1
71.2
71.7
Table 5: Class accuracies on the Kinects dataset without
ImageNet pretraining. Although our skeleton based model
ST-GCN can not achieve the accuracy of the state of the
art model performed on RGB and optical flow modalities,
it can provide stronger complementary information than optical flow based model.
tion classes in Kinetics requires recognizing the objects and
scenes that the actors are interacting with. To verify this, we
select a subset of 30 classes strongly related with body motions, named as “Kinetics-Motion” and list the mean class
accuracies of skeleton and frame based models (Kay et al.
2017) on this subset in Table 4. We can see that on this subset the performance gap is much smaller. We also explore
using ST-GCN to capture motion information in two-stream
style action recognition. As shown as in Fig. 5, our skeleton based model ST-GCN can also provide complementary
information to RGB and optical flow models. We train the
standard TSN (Wang et al. 2016) models from scratches on
Kinetics with RGB and optical flow models. Adding STGCN to the RGB model leads to 0.9% increase, even better than optical flows (0.8%). Combining RGB, optical flow,
and ST-GCN further raises the performance to 71.7%. These
results clearly show that the skeletons can provide complementary information when leveraged effectively (e.g. using
ST-GCN).
5
Conclusion
In this paper, we present a novel model for skeleton based
action recognition, the spatial temporal graph convolutional
networks (ST-GCN). The model constructs a set of spatial
temporal graph convolutions on the skeleton sequences. On
two challenging large-scale datasets, the proposed ST-GCN
outperforms the previous state-of-the-art skeleton based
model. In addition, ST-GCN can capture motion information
in dynamic skeleton sequences which is complementary to
RGB modality. The combination of skeleton based model
and frame based model further improves the performance
in action recognition. The flexibility of ST-GCN model also
opens up many possible directions for future works. For example, how to incorporate contextual information, such as
scenes, objects, and interactions into ST-GCN becomes a
natural question.
Acknowledgement This work is partially supported by
the Big Data Collaboration Research grant from SenseTime
Group (CUHK Agreement No. TS1610626), and the Early
Career Scheme (ECS) of Hong Kong (No. 24204215).
References
[Bruna et al. 2014] Bruna, J.; Zaremba, W.; Szlam, A.; and
Lecun, Y. 2014. Spectral networks and locally connected
networks on graphs. In ICLR.
[Cao et al. 2017a] Cao, Z.; Simon, T.; Wei, S.-E.; and
Sheikh, Y. 2017a. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR.
[Cao et al. 2017b] Cao, Z.; Simon, T.; Wei, S.-E.; and
Sheikh, Y. 2017b. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR.
[Dai et al. 2017] Dai, J.; Qi, H.; Xiong, Y.; Li, Y.; Zhang, G.;
Hu, H.; and Wei, Y. 2017. Deformable convolutional networks. In arXiv:1703.06211.
[Defferrard, Bresson, and Vandergheynst 2016] Defferrard,
M.; Bresson, X.; and Vandergheynst, P. 2016. Convolutional neural networks on graphs with fast localized spectral
filtering. In NIPS.
[Du, Wang, and Wang 2015] Du, Y.; Wang, W.; and Wang,
L. 2015. Hierarchical recurrent neural network for skeleton
based action recognition. In CVPR, 1110–1118.
[Duvenaud et al. 2015] Duvenaud, D. K.; Maclaurin, D.;
Iparraguirre, J.; Bombarell, R.; Hirzel, T.; Aspuru-Guzik,
A.; and Adams, R. P. 2015. Convolutional networks on
graphs for learning molecular fingerprints. In NIPS.
[Fernando et al. 2015] Fernando, B.; Gavves, E.; Oramas,
J. M.; Ghodrati, A.; and Tuytelaars, T. 2015. Modeling
video evolution for action recognition. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, 5378–5387.
[Henaff, Bruna, and LeCun 2015] Henaff, M.; Bruna, J.; and
LeCun, Y. 2015. Deep convolutional networks on graphstructured data. In arXiv:1506.05163.
[Hussein et al. 2013] Hussein, M. E.; Torki, M.; Gowayyed,
M. A.; and El-Saban, M. 2013. Human action recognition
using a temporal hierarchy of covariance descriptors on 3d
joint locations. In IJCAI.
[Kay et al. 2017] Kay, W.; Carreira, J.; Simonyan, K.; Zhang,
B.; Hillier, C.; Vijayanarasimhan, S.; Viola, F.; Green, T.;
Back, T.; Natsev, P.; et al. 2017. The kinetics human action
video dataset. In arXiv:1705.06950.
[Ke et al. 2017] Ke, Q.; Bennamoun, M.; An, S.; Sohel, F.;
and Boussaid, F. 2017. A new representation of skeleton
sequences for 3d action recognition. In CVPR.
[Kim and Reiter 2017] Kim, T. S., and Reiter, A. 2017. Interpretable 3d human action analysis with temporal convolutional networks. In BNMW CVPRW.
[Kipf and Welling 2017] Kipf, T. N., and Welling, M. 2017.
Semi-supervised classification with graph convolutional networks. In ICLR 2017.
[Krizhevsky, Sutskever, and Hinton 2012] Krizhevsky, A.;
Sutskever, I.; and Hinton, G. E. 2012. Imagenet classification with deep convolutional neural networks. In
NIPS.
[Li et al. 2016] Li, Y.; Zemel, R.; Brockschmidt, M.; and Tarlow, D. 2016. Gated graph sequence neural networks. In
ICLR.
[Li et al. 2017] Li, C.; Zhong, Q.; Xie, D.; and Pu, S. 2017.
Skeleton-based action recognition with convolutional neural
networks. In arXiv:1704.07595.
[Liu et al. 2016] Liu, J.; Shahroudy, A.; Xu, D.; and Wang,
G. 2016. Spatio-temporal lstm with trust gates for 3d human
action recognition. In ECCV, 816–833. Springer.
[Niepert, Ahmed, and Kutzkov 2016] Niepert, M.; Ahmed,
M.; and Kutzkov, K. 2016. Learning convolutional neural
networks for graphs. In International Conference on Machine Learning.
[Shahroudy et al. 2016] Shahroudy, A.; Liu, J.; Ng, T.-T.;
and Wang, G. 2016. Ntu rgb+ d: A large scale dataset for 3d
human activity analysis. In CVPR, 1010–1019.
[Shotton et al. 2011] Shotton, J.; Sharp, T.; Kipman, A.;
Fitzgibbon, A.; Finocchio, M.; Blake, A.; Cook, M.; and
Moore, R. 2011. Real-time human pose recognition in parts
from single depth images. In CVPR.
[Simonyan and Zisserman 2014] Simonyan, K., and Zisserman, A. 2014. Two-stream convolutional networks for action recognition in videos. In Advances in neural information processing systems, 568–576.
[Tai, Socher, and Manning 2015] Tai, K. S.; Socher, R.; and
Manning, C. D. 2015. Improved semantic representations
from tree-structured long short-term memory networks. In
ACL.
[Tran et al. 2015] Tran, D.; Bourdev, L.; Fergus, R.; Torresani, L.; and Paluri, M. 2015. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the
IEEE international conference on computer vision, 4489–
4497.
[Van Oord, Kalchbrenner, and Kavukcuoglu 2016]
Van Oord, A.; Kalchbrenner, N.; and Kavukcuoglu, K.
2016. Pixel recurrent neural networks. In ICML.
[Veeriah, Zhuang, and Qi 2015] Veeriah, V.; Zhuang, N.; and
Qi, G.-J. 2015. Differential recurrent neural networks for
action recognition. In CVPR, 4041–4049.
[Vemulapalli, Arrate, and Chellappa 2014] Vemulapalli, R.;
Arrate, F.; and Chellappa, R. 2014. Human action recognition by representing 3d skeletons as points in a lie group.
In CVPR, 588–595.
[Wang et al. 2012] Wang, J.; Liu, Z.; Wu, Y.; and Yuan, J.
2012. Mining actionlet ensemble for action recognition with
depth cameras. In CVPR. IEEE.
[Wang et al. 2016] Wang, L.; Xiong, Y.; Wang, Z.; Qiao, Y.;
Lin, D.; Tang, X.; and Val Gool, L. 2016. Temporal segment
networks: Towards good practices for deep action recognition. In ECCV.
[Wang, Qiao, and Tang 2015] Wang, L.; Qiao, Y.; and Tang,
X. 2015. Action recognition with trajectory-pooled deep-
convolutional descriptors. In Proceedings of the IEEE conference on computer vision and pattern recognition, 4305–
4314.
[Zhang, Liu, and Xiao 2017] Zhang, S.; Liu, X.; and Xiao,
J. 2017. On geometric features for skeleton-based action recognition using multilayer lstm networks. In WACV.
IEEE.
[Zhao et al. 2017] Zhao, Y.; Xiong, Y.; Wang, L.; Wu, Z.;
Tang, X.; and Lin, D. 2017. Temporal action detection with
structured segment networks. In ICCV.
[Zhu et al. 2016] Zhu, W.; Lan, C.; Xing, J.; Zeng, W.; Li, Y.;
Shen, L.; Xie, X.; et al. 2016. Co-occurrence feature learning for skeleton based action recognition using regularized
deep lstm networks. In AAAI.
| 1cs.CV
|
A Generalization of the Borkar-Meyn Theorem
for Stochastic Recursive Inclusions
arXiv:1502.01953v3 [cs.SY] 27 Sep 2016
Arunselvan Ramaswamy
1
and Shalabh Bhatnagar
2
1
[email protected]
2
[email protected]
1,2
Department of Computer Science and Automation, Indian
Institute of Science, Bangalore - 560012, India.
September 28, 2016
Abstract
In this paper the stability theorem of Borkar and Meyn is extended
to include the case when the mean field is a set-valued map. Two different sets of sufficient conditions are presented that guarantee the ‘stability
and convergence’ of stochastic recursive inclusions. Our work builds on
the works of Benaı̈m, Hofbauer and Sorin as well as Borkar and
Meyn. As a corollary to one of the main theorems, a natural generalization of the Borkar and Meyn Theorem follows. In addition, the original
theorem of Borkar and Meyn is shown to hold under slightly relaxed assumptions. As an application to one of the main theorems we discuss
a solution to the ‘approximate drift problem’. Finally, we analyze the
stochastic gradient algorithm with “constant error gradient estimators”
as yet another application of our main result.
1
Introduction
Consider the following recursion in Rd (d ≥ 1):
xn+1 = xn + a(n) [h(xn ) + Mn+1 ] , f or n ≥ 0, where
(1)
(i) h : Rd → Rd is a Lipschitz continuous function.
(ii) a(n) P
> 0, for all n, is the step-size sequence satisfying
∞
and n=0 a(n)2 < ∞.
P∞
n=0
a(n) = ∞
(iii) Mn , n ≥ 1, is a sequence of martingale difference terms that constitute
the noise.
The stochastic recursion given by (1) is often referred to as a stochastic
recursive equation (SRE). A powerful method to analyze the limiting behavior
of (1) is the ODE (Ordinary Differential Equation) method. Here the limiting
1
behavior of the algorithm is described in terms of the asymptotics of the solution
to the ODE
ẋ(t) = h(x(t)).
This method was introduced by Ljung [12] in 1977. For a detailed exposition
on the subject and a survey of results, the reader is referred to Kushner and
Yin [11] as well as Borkar [10].
In 1996, Benaı̈m [4] showed that the asymptotic behavior of a stochastic
recursive equation can be studied by analyzing the asymptotic behavior of the
associated o.d.e. However no assumptions were made on the dynamics of the
o.d.e. Specifically, he developed sufficient conditions which guarantee that limit
sets of the continuously interpolated stochastic iterates are compact, connected,
internally chain transitive and invariant sets of the associated o.d.e. The results
found in [4] are generalized in [5]; further studies were made by Benaı̈m and
Hirsch in [6]. The assumptions made in [4] are sometimes referred to as the
‘classical assumptions’. One of the key assumptions used by Benaı̈m to prove
this convergence theorem is the almost sure boundedness of the iterates i.e.,
stability of the iterates. In 1999, Borkar and Meyn [13] developed sufficient
conditions which guarantee both the stability and convergence of stochastic
recursive equations. These assumptions were consistent with those developed in
[4]. In this paper we refer to the main result of Borkar and Meyn colloquially
as the Borkar-Meyn Theorem. In the same paper [13], several applications to
problems from reinforcement learning have also been discussed. Another set of
sufficient conditions for SRE’s were developed by Andrieu, Moulines and
Priouret [1] using global Lyapunov functions that guarantee the stability and
convergence of the iterates.
In 2005, Benaı̈m, Hofbauer and Sorin [7] showed that the dynamical
systems approach can be extended to the situation where the mean fields are
set-valued. The algorithms considered were of the form:
xn+1 = xn + a(n) [yn + Mn+1 ] , f or n ≥ 0, where
(2)
(i) yn ∈ h(xn ) and h : Rd → {subsets of Rd } is a Marchaud map. For the
definition of Marchaud maps the reader is referred to section 2.1.
P∞
(ii) a(n) P
> 0, for all n ≥ 0, is the step-size sequence satisfying n=0 a(n) = ∞
∞
and n=0 a(n)2 < ∞.
(iii) Mn , n ≥ 1, is a sequence of martingale difference terms.
A recursion such as (2) is also called stochastic recursive inclusion (SRI). Since
a differential equation can be seen as a special case of a differential inclusion
wherein h(x) is a cardinality one set for all x ∈ Rd , SRE (1) can be seen as a
special case of SRI (2).
The main aim of this paper is to extend the original Borkar-Meyn theorem
to the case of stochastic recursive inclusions. We present two overlapping yet
different sets of assumptions, in Sections 2.2 and 3.3 respectively, that guarantee
the stability and convergence of a SRI given by (2). As a consequence of our
2
main results, Theorems 2 and 3, we present a couple of interesting extensions
to the original theorem of Borkar and Meyn in Section 4. Using the frameworks
presented herein we provide a solution to the problem of approximate drift in
Section 5.1. For more details on the approximate drift problem the reader is
referred to Borkar [10]. In Section 6 we discuss the generality, ease of verifiability
and we also try to explain why the assumptions are “natural” in some sense.
Stochastic gradient descent (SGD) is an important method to find minima of
(continuously) differentiable functions. When implementing the corresponding
approximation algorithm (See (13) in Section 5.2) using gradient estimators, an
error is made at each step in calculating the gradient of the objective function.
Lets call this error the “approximation error”. This is the case when using gradient estimators such as Kiefer-Wolfowitz, simultaneous perturbation stochastic
approximation (SPSA) and smoothed functional (SF) schemes, see [9]. Suppose
the perturbation parameters of the aforementioned estimators are kept constant,
then the “approximation error” is bounded by a constant that depends on the
size of the perturbation parameters. We call such estimators constant-error
gradient estimators. In Section 5.2 we analyze the stochastic gradient approximation algorithm that uses a constant-error gradient estimator. Using Theorem 3 we show that the iterates are stable and converge to a δ-neighborhood
of the minimum set, for a specified δ(> 0). Essentially, our framework gives
a threshold ǫ(δ) for the “approximation error” so that the stochastic gradient
approximation algorithm is stable and converges to a δ-neighborhood of the
minimum set.
It is worth noting that prior to this paper one could only claim that an SGD
using constant-error gradient estimators will only converge to some neighborhood
of the minimum set with high probability. On the other hand, our framework
guarantees almost sure convergence to a small neighborhood of the minimum
set.
2
Preliminaries and Assumptions
2.1
Definitions and Notations
The definitions and notations used in this paper are similar to those in Benaı̈m
et. al. [7], Aubin et. al. [2], [3] and Borkar [10]. In this section, we present a
few for easy reference.
A set-valued map h : Rn → {subsets of Rm } is called a Marchaud map if it
satisfies the following properties:
(i) For each x ∈ Rn , h(x) is convex and compact.
(ii) (point-wise boundedness) For each x ∈ Rn , sup kwk < K (1 + kxk) for
w∈h(x)
some K > 0.
(iii) h is an upper-semicontinuous map. We say that h is upper-semicontinuous,
if given sequences {xn }n≥1 (in Rn ) and {yn }n≥1 (in Rm ) with xn → x,
3
yn → y and yn ∈ h(xn ), n ≥ 1, implies that y ∈ h(x). In other words the
graph of h, {(x, y) : y ∈ h(x), x ∈ Rn }, is closed in Rn × Rm .
Let H be a Marchaud map on Rd . The differential inclusion (DI) given by
ẋ ∈ H(x)
(3)
is guaranteed to have at least one solution that is absolutely
P continuous. The
reader is referred to [2] for more details. We say that x ∈
if x is an absolutely
continuous map that satisfies (3). The set-valued semiflow Φ associated with
d
(3) is defined on [0, +∞)
P × R as:
Φt (x) = {x(t) | x ∈ , x(0) = x}. Let B × M ⊂ [0, +∞) × Rk and define
[
ΦB (M ) =
Φt (x).
t∈B, x∈M
T
Let M ⊆ Rd , the ω − limit set be defined by ωΦ (M ) = t≥0 Φ[t,+∞) (M ).
T
Similarly the limit set of a solution x is given by L(x) = t≥0 x([t, +∞)).
M ⊆ Rd is invariant if for every x ∈ MPthere exists a trajectory, x, entirely
in M with x(0) = x. In other words, x ∈
with x(t) ∈ M , for all t ≥ 0.
Let x ∈ Rd and A ⊆ Rd , then d(x, A) := inf{ka − yk | y ∈ A}. We define
the δ-open neighborhood of A by N δ (A) := {x | d(x, A) < δ}. The δ-closed
neighborhood of A is defined by N δ (A) := {x | d(x, A) ≤ δ}. The open ball
of radius r around the origin is represented by Br (0), while the closed ball is
represented by B r (0).
Internally Chain Transitive Set : M ⊂ Rd is said to be internally chain
transitive if M is compact and for every x, y ∈ M , ǫ > 0 and T > 0 we have the
following: There exist Φ1 , . . . , Φn that are n solutions to the differential inclusion
ẋ(t) ∈ h(x(t)), a sequence x1 (= x), . . . , xn+1 (= y) ⊂ M and n real numbers
t1 , t2 , . . . , tn greater than T such that: Φiti (xi ) ∈ N ǫ (xi+1 ) and Φi[0,ti ] (xi ) ⊂ M
for 1 ≤ i ≤ n. The sequence (x1 (= x), . . . , xn+1 (= y)) is called an (ǫ, T ) chain
in M from x to y.
A ⊆ Rd is an attracting set if it is compact and there exists a neighborhood
U such that for any ǫ > 0, ∃ T (ǫ) ≥ 0 with Φ[T (ǫ),+∞) (U ) ⊂ N ǫ (A). Such a
U is called the fundamental neighborhood of A. In addition to being compact
if the attracting set is also invariant then it is called an attractor. The basin of
attraction of A is given by B(A) = {x | ωΦ (x) ⊂ A}. It is called Lyapunov
stable if for all δ > 0, ∃ ǫ > 0 such that Φ[0,+∞) (N ǫ (A)) ⊆ N δ (A). We use T (ǫ)
and Tǫ interchangeably to denote the dependence of T on ǫ.
We define the lower and upper limits of sequences of sets. Let {Kn }n≥1 be
a sequence of sets in Rd .
1. The lower limit of {Kn }n≥1 is given by, Liminfn→∞Kn := {x | lim d(x, Kn ) =
n→∞
0}.
2. The upper-limit of {Kn }n≥1 is given by, Limsupn→∞Kn := {y | lim d(y, Kn ) =
n→∞
0}.
We may interpret that the lower-limit collects the limit points of {Kn }n≥1
while the upper-limit collects its accumulation points.
4
2.2
The assumptions
Recall that we have the following recursion in Rd :
xn+1 = xn + a(n) [yn + Mn+1 ] , where yn ∈ h(xn ).
We state our assumptions below:
(A1) h : Rd → {subsets of Rd } is a Marchaud map.
P
a(n) = ∞ and
(A2) {a(n)}n≥0 is a scalar sequence such that: a(n) > 0 ∀n,
n≥0
P
a(n)2 < ∞. Without loss of generality we let sup a(n) ≤ 1.
n
n≥0
(A3)
{Mn }n≥1 is a martingale difference sequence with respect to the filtration
Fn := σ (x0 , M1 , . . . , Mn ), n ≥ 0.
(i) {Mn }n≥1 is a square integrable sequence.
(ii) E[kMn+1 k2 |Fn ] ≤ K 1 + kxn k2 , for n ≥ 0 and some constant K >
0. Without loss of generality assume that the same constant, K,
works for both the point-wise boundedness condition of (A1) (see
condition (ii) in the definition of Marchaud map in Section 2.1) and
(A3).
For c ≥ 1 and x ∈ Rd , define hc (x) = {y | cy ∈ h(cx)}. Further, for each
x ∈ Rd , define h∞ (x) := Liminfc→∞ hc (x) i.e. the closure of the lower-limit
of {hc (x)}c≥1 .
(A4) h∞ (x) is non-empty for all x ∈ Rd . Further, the differential inclusion
ẋ(t) ∈ h∞ (x(t)) has an attracting set, A, with B 1 (0) as a subset of its
fundamental neighborhood. This attracting set is such that A ⊆ B1 (0).
(A5) Let cn ≥ 1 be an increasing sequence of integers such that cn ↑ ∞ as
n → ∞. Further, let xn → x and yn → y as n → ∞, such that
yn ∈ hcn (xn ), ∀n, then y ∈ h∞ (x).
Since the attracting set, A ⊆ B1 (0), is compact we conclude that sup kxk < 1.
x∈A
To see this, for all x ∈ A define δ(x) :=
sup
kyk, where ǫ(x) > 0 and
y∈B ǫ(x) (x)
B ǫ(x) (x) ⊆ B1 (0). For all x ∈ A we have δ(x) < 1. Further, {Bǫ(x) (x) | x ∈ A}
is an open cover of A. Let {Bǫ(xi ) (xi ) | 1 ≤ i ≤ n} be a finite sub-cover and
δ := max δ(xi ). Clearly, it follows that sup kxk ≤ δ < 1. Define δ1 := sup kxk
1≤i≤n
x∈A
x∈A
and pick real numbers δ2 , δ3 and δ4 such that sup kxk = δ1 < δ2 < δ3 < δ4 < 1.
x∈A
We shall use this sequence later on.
5
Assumptions (A1) − (A3) are the same as in Benaı̈m [7]. However, the assumption on the stability of the iterates is replaced by (A4) and (A5). We show
that (A4) and (A5) are sufficient conditions to ensure stability of iterates. We
start by observing that hc and h∞ are Marchaud maps, where c ≥ 1. Further,
we show that the constant associated with the point-wise boundedness property
is K of (A1) and (A3).
Proposition 1. h∞ and hc , c ≥ 1, are Marchaud maps.
Proof. Fix c ≥ 1 and x ∈ Rd . To prove that hc (x) is compact, we show that it
is closed and bounded. For n ≥ 1, let yn ∈ hc (x) and let lim yn = y. It follows
n→∞
that cyn ∈ h(cx) for each n ≥ 1 and lim cyn = cy. Since h(cx) is closed, we
n→∞
have that cy ∈ h(cx) and y ∈ hc (x). If we show that hc is point-wise bounded
then we can conclude that hc (x) is compact. To prove the aforementioned, let
y ∈ hc (x), then cy ∈ h(cx). Since h satisfies (A1)(ii), we have that
ckyk ≤ K (1 + kcxk) , hence
1
+ kxk .
kyk ≤ K
c
Since c(≥ 1) and x is arbitrarily chosen, hc is point-wise bounded and the
compactness of hc (x) follows. The set hc (x) = {z/c | z ∈ h(cx)} is convex since
h(cx) is convex and hc (x) is obtained by scaling it by 1c .
Next, we show that hc (x) is upper-semicontinuous. Let lim xn = x, lim
n→∞
n→∞
yn = y and yn ∈ hc (xn ), ∀ n ≥ 1. We need to show that y ∈ hc (x). We have
that cyn ∈ h(cxn ) for each n ≥ 1. Since lim cxn = cx and lim cyn = cy, we
n→∞
n→∞
conclude that cy ∈ h(cx) since h is assumed to be upper-semicontinuous.
It is left to show that h∞ (x), x ∈ Rd is a Marchaud map. To prove that kzk
≤ K (1 + kxk) for all z ∈ h∞ (x), it is enough to prove that kyk ≤ K (1 + kxk)
for all y ∈ Liminfc→∞ hc (x). Fix some y ∈ Liminfc→∞ hc (x) then there exist
zn ∈ hn (x), n ≥ 1, such that lim ky − zn k = 0. We have that
n→∞
kyk ≤ ky − zn k + kzn k.
Since hc , c ≥ 1, is point-wise bounded (the constant associated is independent
of c and equals K), the above inequality becomes
kyk ≤ ky − zn k + K (1 + kxk).
Letting n → ∞ in the above inequality, we obtain kyk ≤ K (1 + kxk). Recall
that h∞ (x) = Liminfc→∞ hc (x), hence it is compact.
Again, to show that h∞ (x) is convex, for each x ∈ Rd , we start by proving
that Liminfc→∞ hc (x) is convex. Let u, v ∈ Liminfc→∞ hc (x) and 0 ≤ t ≤ 1.
We need to show that tu + (1 − t)v ∈ Liminfc→∞ hc (x). Consider an arbitrary
sequence {cn }n≥1 such that cn → ∞, then there exist un , vn ∈ hcn (x) such that
6
kun − uk and kvn − vk → 0 as cn → ∞. Since hcn (x) is convex, it follows that
tun + (1 − t)vn ∈ hcn (x), further
lim ( tun + (1 − t)vn ) = tu + (1 − t)v.
cn →∞
Since we started with an arbitrary sequence cn → ∞, it follows that tu +
(1 − t)v ∈ Liminfc→∞ hc (x). Now we can prove that h∞ (x) is convex. Let
u, v ∈ h∞ (x). Then ∃ {un }n≥1 and {vn }n≥1 ⊆ Liminfc→∞ hc (x) such that
un → u and vn → v as n → ∞. We need to show that tu + (1 − t)v ∈ h∞ (x),
for 0 ≤ t ≤ 1. Since tun + (1 − t)vn ∈ Liminfc→∞ hc (x), the desired result is
obtained by letting n → ∞ in tun + (1 − t)vn .
Finally, we show that h∞ is upper-semicontinuous. Let lim xn = x, lim
n→∞
n→∞
yn = y and yn ∈ h∞ (xn ), ∀ n ≥ 1. We need to show that y ∈ h∞ (x). Since
yn ∈ h∞ (xn ), ∃ zn ∈ Liminfc→∞ hc (xn ) such that kzn − yn k < n1 . Since
zn ∈ Liminfc→∞ hc (xn ), n ≥ 1, it follows that there exist cn such that for all
c ≥ cn , d (zn , hc (xn )) < n1 . In particular, ∃ un ∈ hcn (xn ) such that kzn − un k
< n1 . We choose the sequence {cn }n≥1 such that cn+1 > cn for each n ≥ 1. We
now have the following: lim un = y, un ∈ hcn (xn ) ∀ n and lim xn = x. It
n→∞
n→∞
follows directly from assumption (A5) that y ∈ h∞ (x).
7
3
Stability and convergence of stochastic recursive inclusions
We begin by providing a brief outline of our approach to prove the stability of a SRI under assumptions (A1) − (A5). First we divide the time line,
[0, ∞), approximately into intervals of length T . We shall explain later how we
choose and fix T . Then we construct a linearly interpolated trajectory from the
given stochastic recursive inclusion; the construction is explained in the next
paragraph. A sequence of ‘rescaled’ trajectories of length T is constructed as
follows: At the beginning of each T -length interval we observe the trajectory
to see if it is outside the unit ball, if so we scale it back to the boundary of
the unit ball. This scaling factor is then used to scale the ‘rest of the T -length
trajectory’.
To show that the iterates are bounded almost surely we need to show that the
linearly interpolated trajectory does not ‘run off’ to infinity. To do so we assume
that this is not true and show that there exists a subsequence of the rescaled
T -length trajectories that has a solution to ẋ(t) ∈ h∞ (x(t)) as a limit point in
C([0, T ], Rd ). We choose and fix T such that any solution to ẋ(t) ∈ h∞ (x(t))
with an initial value inside the unit ball is close to the origin at the end of time
T . In this paper we choose T = T (δ2 − δ1 ) + 1. We then argue that the linearly
interpolated trajectory is forced to make arbitrarily large ‘jumps’ within time
T . The Gronwall inequality is then used to show that this is not possible.
Once we prove stability of the recursion we invoke Theorem 3.6 & Lemma
3.8 from Benaı̈m, Hofbauer and Sorin [7] to conclude that the limit set is a
closed, connected, internally chain transitive and invariant set associated with
ẋ(t) ∈ h∞ (x(t)).
We construct the linearly interpolated trajectory x(t), for t ∈ [0, ∞), from
Pn−1
the sequence {xn } as follows: Define t(0) := 0, t(n) :=
i=0 a(i). Let
x(t(n)) := xn and for t ∈ (t(n), t(n + 1)), let
t(n + 1) − t
t − t(n)
x(t) :=
x(t(n)) +
x(t(n + 1)).
t(n + 1) − t(n)
t(n + 1) − t(n)
We define a piecewise constant trajectory using the sequence {yn }n≥0 as follows:
y(t) := yn for t ∈ [t(n), t(n + 1)), n ≥ 0.
We know that the DI given by ẋ(t) ∈ h∞ (x(t)) has an attractor set A such
that δ1 := sup kxk < 1. Let us fix T := T (δ2 − δ1 ) + 1, where T (δ2 − δ1 ) is
x∈A
as defined in section 2.1. Then, kx(t)k < δ2 , for all t ≥ T (δ2 − δ1 ), where
{x(t) : t ∈ [0, ∞)} is a solution to ẋ(t) ∈ h∞ (x(t)) with an initial value inside
the unit ball around the origin.
Define T0 := 0 and Tn := min{t(m) : t(m) ≥ Tn−1 + T }, n ≥ 1. Observe
that there exists a subsequence {m(n)}n≥0 of {n} such that Tn = t(m(n))
∀ n ≥ 0. We construct the rescaled trajectory, x̂(t), t ≥ 0, as follows: Let
x(t)
, where r(n) = kx(Tn )k ∨ 1.
t ∈ [Tn , Tn+1 ) for some n ≥ 0, then x̂(t) := r(n)
−
Also, let x̂(Tn+1 ) := lim x̂(t), t ∈ [Tn , Tn+1 ). The corresponding ‘rescaled y
t↑Tn+1
8
iterates’ are given by ŷ(t) :=
M̂k+1 :=
Mk+1
r(n) ,
y(t)
r(n)
and the rescaled martingale noise terms by
t(k) ∈ [Tn , Tn+1 ), n ≥ 0.
Consider the recursion at hand, i.e.,
x(t(k + 1)) = x(t(k)) + a(k) (y(t(k)) + Mk+1 ) ,
such that t(k), t(k + 1) ∈ [Tn , Tn+1 ). Multiplying both sides by 1/r(n) we get
the rescaled recursion:
x̂(t(k + 1)) = x̂(t(k)) + a(k) ŷ(t(k)) + M̂k+1 .
hr(n) (x̂(t(k))). It is worth
Since y(t(k)) ∈h h (x(t(k))), it
i follows that ŷ(t(k)) ∈
2
2
noting that E kM̂k+1 k |Fk ≤ K 1 + kx̂(t(k))k .
9
3.1
Characterizing limits, in C([0, T ], Rd ), of the rescaled
trajectories
We do not provide proofs for the first three lemmas since they can be found
in Borkar [10] or Benaı̈m, Hofbauer and Sorin [7]. The first two lemmas
essentially state that the rescaled martingale noise converges almost surely.
Lemma 1. sup Ekx̂(t)k2 < ∞.
t∈[0,T ]
Lemma 2. The rescaled sequence {ζ̂n }n≥1 , where ζ̂n =
convergent almost surely.
Pn−1
k=0
a(k)M̂k+1 , is
The rest of the lemmas are needed to prove the stability theorem, Theorem 1.
We begin by showing that the rescaled trajectories are bounded almost surely.
Lemma 3.
sup kx̂(t)k < ∞ a.s.
t∈[0,∞)
As stated earlier we omit the proof of the above stated lemma and establish a
couple of notations used later. Let A = {ω | {ζ̂n (ω)}n≥1 converges}. Since ζ̂n ,
n ≥ 1, converges
Pk−1 on A, there exists Mω < ∞, possibly sample path dependent,
such that k l=0 a(m(n) + l)M̂m(n)+l+1 k ≤ Mw , where Mω is independent of n
and k. Also, let supkx̂(t)k ≤ Kω , where Kω := (1 + Mω + (T + 1)K) eK(T +1)
t≥0
is also a constant that is sample path dependent.
Let xn (t), t ∈ [0, T ] be the solution (upto time T ) to ẋn (t) = ŷ(Tn + t), with
the initial condition xn (0) = x̂(Tn ). Clearly, we have
xn (t) = x̂(Tn ) +
Z
t
ŷ(Tn + z) dz.
(4)
0
The following two lemmas are inspired by ideas from Benaı̈m, Hofbauer
and Sorin [7] as well as Borkar [10]. In the lemma that follows we show that
the limit sets of {xn (· ) | n ≥ 0} and {x̂(Tn +· ) | n ≥ 0} coincide. We seek limits
in C([0, T ], Rd ).
Lemma 4. lim
sup
n→∞ t∈[T ,T +T ]
n
n
kxn (t) − x̂(t)k = 0 a.s.
Proof. Let t ∈ [t(m(n) + k), t(m(n) + k + 1)) and t(m(n) + k + 1) ≤ Tn+1 . We
first assume that
t(m(n) + k + 1) < Tn+1 . We have the following:
x̂(t) =
t(m(n) + k + 1) − t
a(m(n) + k)
x̂(t(m(n)+k))+
t − t(m(n) + k)
a(m(n) + k)
x̂(t(m(n)+k+1)).
Substituting for x̂(t(m(n) + k + 1)) in the above equation we get:
x̂(t) =
t(m(n) + k + 1) − t
t − t(m(n) + k)
x̂(t(m(n) + k)) +
a(m(n) + k)
a(m(n) + k)
x̂(t(m(n) + k)) + a(m(n) + k) ŷ(t(m(n) + k)) + M̂m(n)+k+1 ,
10
hence,
x̂(t) = x̂(t(m(n) + k)) + (t − t(m(n) + k)) ŷ(t(m(n) + k)) + M̂m(n)+k+1 .
Unfolding x̂(t(m(n) + k)) over k we get,
x̂(t) = x̂(Tn ) +
k−1
X
a(m(n) + l) ŷ(t(m(n) + l)) + M̂m(n)+l+1 +
l=0
(t − t(m(n) + k)) ŷ(t(m(n) + k)) + M̂m(n)+k+1 . (5)
Now, we consider xn (t), i.e.,
xn (t) = x̂(Tn ) +
Z
t
ŷ(Tn + z) dz.
0
Splitting the above integral, we get
xn (t) = x̂(Tn ) +
k−1
X Z t(m(n)+l+1)
l=0
t(m(n)+l)
ŷ(z) dz +
Z
t
ŷ(z) dz.
t(m(n)+k)
Thus,
xn (t) = x̂(Tn )+
k−1
X
a(m(n)+l)ŷ(t(m(n)+l))+(t − t(m(n) + k)) ŷ(t(m(n)+k)).
l=0
(6)
From (5) and (6), it follows that
kxn (t)−x̂(t)k ≤
k−1
X
a(m(n) + l)M̂m(n)+l+1 + (t − t(m(n) + k)) M̂m(n)+k+1
l=0
and hence,
kxn (t) − x̂(t)k ≤ kζ̂m(n)+k − ζ̂m(n) k + kζ̂m(n)+k+1 − ζ̂m(n)+k k.
If t(m(n) + k + 1) = Tn+1 then in the proof we may replace x̂(t(m(n) + k + 1))
−
with x̂(Tn+1
). The arguments remain the same. Since ζ̂n , n ≥ 1, converges
almost surely, the desired result follows.
The sets {xn (t), t ∈ [0, T ] | n ≥ 0} and {x̂(Tn + t), t ∈ [0, T ] | n ≥ 0} can be
viewed as subsets of C([0, T ], Rd ). It can be shown that {xn (t), t ∈ [0, T ] | n ≥
0} is equi-continuous and point-wise bounded. Thus from the Arzela-Ascoli
theorem, {xn (t), t ∈ [0, T ] | n ≥ 0} is relatively compact. It follows from
Lemma 4 that the set {x̂(Tn + t), t ∈ [0, T ] | n ≥ 0} is also relatively compact
in C([0, T ], Rd ).
Lemma 5. Let r(n) ↑ ∞, then any limit point of {x̂(Tn + t), t ∈ [0, T ] : n ≥ 0}
Rt
is of the form x(t) = x(0) + 0 y(s) ds, where y : [0, T ] → Rd is a measurable
function and y(t) ∈ h∞ (x(t)), t ∈ [0, T ].
11
,
Proof. We define the notation [t] := max{t(k) | t(k) ≤ t}, t ≥ 0. Let t ∈
[Tn , Tn+1 ), then ŷ(t) ∈ hr(n) (x̂([t])) and kŷ(t)k ≤ K (1 + kx̂([t])k) since hr(n) is
a Marchaud map (K is the constant associated with the point-wise boundedness
property). It follows from Lemma 3 that sup kŷ(t)k < ∞ a.s. Using obsert∈[0,∞)
vations made earlier, we can deduce that there exists a sub-sequence of N, say
d
{l} ⊆ {n}, such that x̂(T
and ŷ(m(l)+· ) → y(· )
l + t) → x(t) in C [0, T ], R
d
weakly in L2 [0, T ], R . From Lemma 4 it follows that xl (· ) → x(· ) in
C [0, T ], Rd . Letting r(l) ↑ ∞ in
xl (t) = xl (0) +
Z
t
ŷ(t(m(l) + z)) dz, t ∈ [0, T ],
0
we get x(t) = x(0) +
kx(0)k ≤ 1.
Rt
0
y(z)dz for t ∈ [0, T ]. Since kx̂(Tn )k ≤ 1 we have
Since ŷ(Tl + · ) → y(· ) weakly in L2 [0, T ], Rd , there exists {l(k)} ⊆ {l} such
that
N
1 X
ŷ(Tl(k) + · ) → y(· ) strongly in L2 [0, T ], Rd .
N
k=1
Further, there exists {N (m)} ⊆ {N } such that
N (m)
X
1
ŷ(Tl(k) + · ) → y(· ) a.e. on [0, T ].
N (m)
k=1
Let us fix t0 ∈ {t |
1
N (m)
PN (m)
k=1
ŷ(Tl(k) + t) → y(t), t ∈ [0, T ]}, then
N (m)
X
1
lim
ŷ(Tl(k) + t0 ) = y(t0 ).
N (m)→∞ N (m)
k=1
Since h∞ (x(t0 )) is convex and compact (Proposition 1), to show that y(t0 ) ∈
h∞ (x(t0 )) it is enough to prove that lim d ŷ(Tl(k) + t0 ), h∞ (x(t0 )) = 0. If
l(k)→∞
not, ∃ ǫ > 0 and {n(k)} ⊆ {l(k)} such that d ŷ(Tn(k) + t0 ), h∞ (x(t0 )) > ǫ.
Since {ŷ(Tn(k) + t0 )}k≥1 is norm bounded, it follows that there is a convergent
sub-sequence. For the sake of convenience we assume that lim ŷ(Tn(k) + t0 ) =
k→∞
y, for some y ∈ Rd . Since ŷ(Tn(k) + t0 ) ∈ hr(n(k)) (x̂([Tn(k) + t0 ])) and lim
k→∞
x̂([Tn(k) + t0 ]) = x(t0 ), it follows from assumption (A5) that y ∈ h∞ (x(t0 )).
This leads to a contradiction.
Note that in the statement of Lemma 5 we can replace ‘r(n) ↑ ∞’ by ‘r(l) ↑
∞’, where {r(l))} is a subsequence of {r(n)}. Specifically we can conclude that
any limit point of {x̂(Tk + t), t ∈ [0, T ]}{k}⊆{n} in C([0, T ], Rd), conditioned on
Rt
r(k) ↑ ∞, is of the form x(t) = x(0) + 0 y(z) dz, where y(t) ∈ h∞ (x(t)) for
t ∈ [0, T ]. It should be noted that y(· ) may be sample path dependent. The
following is an immediate consequence of Lemma 5.
12
Corollary 1. ∃ 1 < R0 < ∞ such that ∀ r(l) > R0 kx̂(Tl +· ) − x(· )k < δ3 − δ2 ,
where {l} ⊆ N and x(· ) is a solution (up to time T ) of ẋ(t) ∈ h∞ (x(t)) such
that kx(0)k ≤ 1. The form of x(· ) is as given by Lemma 5.
Proof. Assume to the contrary that ∃ r(l) ↑ ∞ such that x̂(Tl +· ) is at least
δ3 − δ2 away from any solution to the DI. It follows from Lemma 5 that there
exists a subsequence of {x̂(Tl + t), 0 ≤ t ≤ T : l ⊆ N} guaranteed to converge,
in C([0, T ], Rd ), to a solution of ẋ(t) ∈ h∞ (x(t)) such that kx(0)k ≤ 1. This is
a contradiction.
It is worth noting that R0 may be sample path dependent. Since T = T (δ2 −
δ1 ) + 1 we get kx̂([Tl + T ])k < δ3 for all Tl such that kx(Tl )k(= r(l)) > R0 .
3.2
Stability theorem
We are now ready to prove the stability of a SRI given by (2) under the
assumptions (A1) − (A5). If sup r(n) < ∞, then the iterates are stable and
n
there is nothing to prove. If on the other hand sup r(n) = ∞, there exists
n
{l} ⊆ {n} such that r(l) ↑ ∞. It follows from Lemma 5 that any limit point of
Rt
{x̂(Tl + t), t ∈ [0, T ] : {l} ⊆ {n}} is of the form x(t) = x(0) + 0 y(s) ds, where
y(t) ∈ h∞ (x(t)) for t ∈ [0, T ]. From assumption (A4), we have that kx(T )k <
δ2 . Since the time intervals are roughly T apart, for large values of r(n) we
−
−
can conclude that kx̂ Tn+1
k < δ3 , where x̂(Tn+1
) = limt↑t(m(n+1)) x̂(t), t ∈
[Tn , Tn+1 ).
Theorem 1 (Stability Theorem for DI). Under assumptions (A1) − (A5),
supkxn k < ∞ a.s.
n
Proof. As explained earlier it is sufficient to consider the case when sup r(n) =
n
∞. Let {l} ⊆ {n} such that r(l) ↑ ∞. Recall that Tl = t(m(l)) and that
[Tl + T ] = max{t(k) | t(k) ≤ Tl + T }.
We have kx(T )k < δ2 since x(t) is a solution, up to time T , to the DI given
by ẋ(t) ∈ h∞ (x(t)) and we have fixed T = T (δ2 − δ1 ) + 1. From Lemma 5 we
conclude that there exists N such that all of the following happen:
(i) m(l) ≥ N =⇒ kx̂([Tl + T ])k < δ3 .
(ii) n ≥ N =⇒ a(n) <
δ4 −δ3
[K(1+Kω )+Mω ] .
(iii) n > m ≥ N =⇒ kζ̂n − ζ̂m k < Mω .
(iv) m(l) ≥ N =⇒ r(l) > R0 .
In the above, R0 is defined in the statement of Corollary 1 and Kω , Mω are
explained in Lemma 3.
13
Recall that we chose sup kxk = δ1 < δ2 < δ3 < δ4 < 1 in Section 2.2. Let
x∈A
m(l) ≥ N and t(m(l + 1)) = t(m(l) + k + 1) for some k ≥ 0. Clearly from the
manner in which the Tn sequence is defined, we have t(m(l) + k) = [Tl + T ].
−
As defined earlier x̂(Tn+1
) = limt↑t(m(n+1)) x̂(t), t ∈ [Tn , Tn+1 ) and n ≥ 0.
Consider the equation
−
x̂(Tl+1
) = x̂(t(m(l) + k)) + a(m(l) + k) ŷ(t(m(l) + k)) + M̂m(l)+k+1 .
Taking norms on both sides we get,
−
kx̂(Tl+1
)k ≤ kx̂(t(m(l)+k))k + a(m(l)+k)kŷ(t(m(l)+k))k + a(m(l)+k)kM̂m(l)+k+1 k.
From the way we have chosen N we conclude that:
kŷ(t(m(l) + k))k ≤ K (1 + kx̂(t(m(l) + k)k) ≤ K (1 + Kω ) and that
kM̂m(l)+k+1 k = kζ̂m(l)+k+1 − ζ̂m(l)+k k ≤ Mω .
Thus we have that,
−
kx̂(Tl+1
)k ≤ kx̂(t(m(l) + k))k + a(m(l) + k) (K(1 + Kω ) + Mω ) .
−
Finally we have that kx̂(Tl+1
)k < δ4 and
−
kx̂(Tl+1
)k
kx(Tl+1 )k
=
< δ4 < 1.
kx(Tl )k
kx̂(Tl )k
(7)
It follows from (7) that kx(Tn+1 )k < δ4 kx(Tn )k if kx(Tn )k > R0 . From
Corollary 1 and the aforementioned we get that the trajectory falls at an exponential rate till it enters B R0 (0). Let t ≤ Tl , t ∈ [Tn , Tn+1 ) and n + 1 ≤ l, be
the last time that x(t) jumps from B R0 (0) to the outside of the ball. It follows
that kx(Tn+1 )k ≥ kx(Tl )k. Since r(l) ↑ ∞, x(t) would be forced to make larger
and larger jumps within an interval of T + 1. This leads to a contradiction since
the maximum jump within any fixed time interval can be bounded using the
Gronwall inequality.
We now state one of the main theorems of this paper.
Theorem 2. Under assumptions (A1) − (A5), almost surely, the sequence
{xn }n≥0 generated by the stochastic recursive inclusion, given by (2), is bounded
and converges to a closed, connected, internally chain transitive and invariant
set of ẋ(t) ∈ h(x(t)).
Proof. The stability of the iterates is shown in Theorem 1. The convergence
can be proved under assumptions (A1) − (A3) and the stability of the iterates
in exactly the same manner as in Theorem 3.6 & Lemma 3.8 of Benaı̈m,
Hofbauer and Sorin [7].
We have thus far shown that under assumptions (A1) − (A5) the SRI given
by (2) is stable and converges to a closed, connected, internally chain transitive
and invariant set.
14
3.3
Stability theorem under modified assumptions
In (A4) we assumed that Liminfc→∞hc (x) is nonempty for all x ∈ Rd . In
this section we shall develop a stability criterion for the case when we no longer
make such an assumption. In other words, we work with a modified version of
assumption (A4) that we call (A6).
Modification of Assumption (A4)
Recall the following SRI :
xn+1 = xn + a(n) [yn + Mn+1 ] , for n ≥ 0.
(8)
Since hc is point-wise bounded for each c ≥ 1, we have sup kyk ≤ K(1+kxk),
y∈hc (x)
where x ∈ Rd (see Proposition 1). This implies that {yc }c≥1 , where yc ∈ hc (x),
has at least one convergent subsequence. It follows from the definition of upperlimit of a sequence of sets (see Section 2.1) that Limsupc→∞ hc (x) is non-empty
for every x ∈ Rd . It is worth noting that Liminfc→∞hc (x) ⊆ Limsupc→∞ hc (x)
for every x ∈ Rd . Another important point to consider is that the lower-limits of
sequences of sets are harder to compute than their upper-limits, see Aubin [3]
for more details.
Recall that hc (x) = {y | cy ∈ h(cx)}, where x ∈ Rd and c ≥ 1. Clearly the
upper-limit, Limsupc→∞ hc (x) = {y | lim d(y, hc (x)) = 0} is nonempty for
c→∞
every x ∈ Rd . For A ⊆ Rd , co(A) denotes the closure of the convex hull of A,
i.e., the closure of the smallest convex set containing A.
Define h∞ (x) := co ( Limsupc→∞ hc (x)) .
Below we state the modification of assumption (A4) that we call (A6).
(A6) The differential inclusion ẋ(t) ∈ h∞ (x(t)) has an attracting set A ⊂
B1 (0) and B 1 (0) is a subset of some fundamental neighborhood of A.
Note that in (A4), h∞ (x) := Liminfc→∞ hc (x) while in (A6), h∞ (x) :=
co ( Limsupc→∞ hc (x)). In this section we shall work with this new definition
of h∞ .
Proposition 2. h∞ is a Marchaud map.
Proof. From the definition of h∞ it follows that h∞ (x) is convex, compact for
all x ∈ Rd and h∞ is point-wise bounded. It is left to prove that h∞ is an
upper-semicontinuous map.
Let xn → x, yn → y and yn ∈ h∞ (xn ), for all n ≥ 1. We need to show
that y ∈ h∞ (x). We present a proof by contradiction. Since h∞ (x) is convex
and compact, y ∈
/ h∞ (x) implies that there exists a linear functional on Rd ,
say f , such that sup f (z) ≤ α − ǫ and f (y) ≥ α + ǫ, for some α ∈ R
z∈h∞ (x)
15
and ǫ > 0. Since yn → y, there exists N > 0 such that for all n ≥ N ,
f (yn ) ≥ α + 2ǫ . In other words, h∞ (x) ∩ [f ≥ α + 2ǫ ] 6= φ for all n ≥ N .
We use the notation [f ≥ a] to denote the set {x | f (x) ≥ a}. For the sake of
convenience let us denote the set Limsupc→∞hc (x) by A(x), where x ∈ Rd . We
claim that A(xn ) ∩ [f ≥ α + 2ǫ ] 6= φ for all n ≥ N . We prove this claim later, for
now we assume that the claim is true and proceed. Pick zn ∈ A(xn )∩[f ≥ α+ 2ǫ ]
for each n ≥ N . It can be shown that {zn }n≥N is norm bounded and hence
contains a convergent subsequence, {zn(k) }k≥1 ⊆ {zn }n≥N . Let lim zn(k) = z.
k→∞
Since zn(k) ∈ Limsupc→∞(hc (xn(k) )), ∃ cn(k) ∈ N such that kwn(k) − zn(k) k <
1
n(k) , where wn(k) ∈ hcn(k) (xn(k) ). We choose the sequence {cn(k) }k≥1 such that
cn(k+1) > cn(k) for each k ≥ 1.
We have the following: cn(k) ↑ ∞, xn(k) → x, wn(k) → z and wn(k) ∈
hcn(k) (xn(k) ), for all k ≥ 1. It follows from assumption (A5) that z ∈ h∞ (x).
Since zn(k) → z and f (zn(k) ) ≥ α + 2ǫ for each k ≥ 1, we have that f (z) ≥ α + 2ǫ .
This contradicts the earlier conclusion that sup f (z) ≤ α − ǫ.
z∈h∞ (x)
It remains to prove that A(xn ) ∩ [f ≥ α + 2ǫ ] 6= φ for all n ≥ N . If this were
not true, then ∃{m(k)}k≥1 ⊆ {n ≥ N } such that A(xm(k) ) ⊆ [f < α + 2ǫ ] for all
k. It follows that
h∞ (xm(k) ) = co(A(xm(k) )) ⊆ [f ≤ α + 2ǫ ] for each k ≥ 1. Since yn(k) → y, ∃N1
such that for all n(k) ≥ N1 , f (yn(k) ) ≥ α + 3ǫ
4 . This is a contradiction.
We are now ready to state the second stability theorem for an SRI given by (8)
under a modified set of assumptions. We retain assumptions (A1)−(A3), replace
(A4) by (A6) and finally in (A5) we let h∞ (x) := co ( Limsupc→∞ hc (x)). We
state the theorem under these updated set of assumptions.
Theorem 3 (Stability Theorem for DI #2). Under assumptions (A1) − (A3),
(A5) (with h∞ (x) := co(Limsupc→∞ hc (x))) and (A6), almost surely the sequence {xn }n≥0 generated by the stochastic recursive inclusion, given by (8) is
bounded and converges to a closed, connected internally chain transitive invariant set of ẋ(t) ∈ h(x(t)).
Proof. The statements of Lemmas 1−5 hold true even when h∞ := co ( Limsupc→∞ hc (x))
and (A5) is interpreted as explained earlier. The stability of the iterates can
be proven in an identical manner to the proof of Theorem 1. Next, we invoke
Theorem 3.6 & Lemma 3.8 of Benaı̈m, Hofbauer and Sorin [7] to conclude
that the iterates converge to a closed, connected, internally chain transitive and
invariant set of ẋ(t) ∈ h(x(t)).
Remark 1. Assumptions (A4) and (A6) required that ẋ(t) ∈ h∞ (x(t)) have an
attractor set inside B1 (0) (the open unit ball). Further, it required B 1 (0) to be in
its fundamental neighborhood. Note that h∞ (x) is defined as Liminfc→∞ hc (x)
when using (A4) and it is defined as co ( Limsupc→∞ hc (x)) when using (A6).
Consider the following generalization of (A4)/(A6).
(A4)′ /(A6)′ : ẋ(t) ∈ h∞ (x(t)) has an attractor set A such that A ⊆ Ba (0) and B a (0) is
a subset of its fundamental neighborhood, where 0 ≤ a < ∞.
16
Note that a could be greater than 1, further since A is compact by definition, a is
finite. A sufficient condition for (A4)′ /(A6)′ is when A is a globally attracting,
Lyapunov stable set associated with ẋ(t) ∈ h∞ (x(t)). In this case any compact
set is a fundamental neighborhood of A.
At the beginning of Section 3 we constructed the rescaled trajectory by projecting onto the unit ball around the origin. In order to use (A4)′ /(A6)′ we
build the rescaled trajectory by projecting onto B a (0) instead. We can modify
the proofs such that the statements of Theorems 2 and 3 remain true under
assumptions (A1) − (A3), (A4)′ /(A6)′ and (A5).
Remark 2. The advantage of using (A4)′ /(A6)′ is that one can conclude the
stability of the iterates by merely possessing the knowledge that the associated
DI of the infinity system has a global attractor set. Consider the following
trivial example of a stochastic gradient descent algorithm with linear gradient
function of the from −(Ax+b). The corresponding infinity system, ẋ(t) = −Ax,
is clearly “related” to the associated o.d.e. ẋ(t) = −(Ax + b). Specifically, if
there was a unique global minimizer then both the aforementioned o.d.e.’s have
a global attractor which in turn implies the stability of the iterates as discussed
before. This trivial example also illustrates a finer point that h∞ and h could
be related, hence information about h could help us ascertain if (A4)′ /(A6)′ is
satisfied. Whenever possible one could also construct Lyapunov functions to
ascertain the same. While we did not consider Lyapunov-type conditions for
stability, it would be interesting to extend the Lyapunov-type stability conditions
developed for SRE’s by Andrieu, Priouret and Moulines [1] to include
SRI’s.
17
4
Extensions to the stability theorem of Borkar
and Meyn
We begin this section by listing the assumptions (See Section 2 of [13]) and
statement of the Borkar-Meyn Theorem (See Section 2.1 of [13]). The notations
used are consistent with those of equation (1).
(BM1) (i) The function h : Rd → Rd is Lipschitz continuous, with Lipschitz constant L. There exists a function h∞ : Rd → Rd such that lim h(cx)
=
c
c→∞
h∞ (x), for each x ∈ Rd .
(ii) hc → h∞ uniformly on compacts, as c → ∞.
(iii) The o.d.e. ẋ(t) = h∞ (x(t)) has the origin as the unique globally
asymptotically stable equilibrium.
P
a(n) = ∞ and
(BM2) {a(n)}n≥0 is a scalar sequence such that: a(n) ≥ 0,
n≥0
P
a(n)2 < ∞. Without loss of generality, we also let sup a(n) ≤ 1.
n
n≥0
(BM3) {Mn }n≥1 is a martingale difference sequence with respect to the filtration
Fn := σ (x0 , M1 , . . . , Mn ), n ≥ 0. Thus, E [Mn+1 |Fn ] = 0 a.s., ∀ n≥ 0.
{Mn } is also square integrable with E[kMn+1 k2 |Fn ] ≤ L 1 + kxn k2 , for
some constant L > 0. Without loss of generality, assume that the same
constant, L, works for both (BM 1)(i) and (BM 3).
Theorem 4 (Borkar-Meyn Theorem). Suppose (BM1)-(BM3) hold. Then supkxn k <
n
∞ almost surely. Further, the sequence {xn } converges almost surely to a (possibly sample path dependent) compact connected internally chain transitive invariant set of ẋ(t) = h(x(t)).
In what follows we illustrate a weakening of (BM 1) − (BM 3) stated above
using Theorems 2 & 3. Note that (BM 2) is the standard step-size assumption
while (BM 3) is the assumption on the martingale difference noise; we endeavor
to weaken (BM 1).
4.1
Superfluity of (BM1)(ii) as a consequence of Theorem 2
In this section we discuss in brief how the Borkar-Meyn Theorem (Theorem 4)
can be proven under (BM 1)(i), (iii), (BM 2) and (BM 3). In other words, we
show that (BM 1)(ii) is superfluous. In this direction we begin by showing the
following: A recursion given by (1) satisfies (BM 1)(i), (iii), (BM 2) and (BM 3)
⇒ (1) satisfies (A1) − (A5) of Section 2.2. The following implications are
straightforward: (BM 1)(i), (iii) ⇒ (A1) & (A4); (BM 2) ⇒ (A2); (BM 3) ⇒
(A3). We now show (BM 1)(i), (iii) ⇒ (A5). Given xn → x, cn ↑ ∞ and
hcn (xn ) → y we need to show y = h∞ (x). We have the following:
khcn (xn ) − h∞ (x)k ≤ khcn (xn ) − hcn (x)k + khcn (x) − h∞ (x)k.
If h is Lipschitz with constant L then it can be shown that hc (hc : x 7→ h(cx)
c ,
x ∈ Rd ) is Lipschitz, for every c ≥ 1, with the same constant. Further, hcn (x) →
18
h∞ (x) as cn ↑ ∞. Taking limits (cn ↑ ∞) on both sides of the above equation
gives lim hcn (xn ) = h∞ (x) as required. Since (A1) − (A5) are satisfied it
cn ↑∞
follows from Theorem 2 that a SRE satisfying (BM 1)(i), (iii), (BM 2), (BM 3)
is stable and converges to a closed, connected, internally chain transitive and
invariant set of ẋ(t) = h(x(t)) (Theorem 4).
We discuss in brief how we work around using (BM 1)(ii) in proving the
Borkar-Meyn Theorem. The notations used herein are consistent with those
found in Chapter 3 of Borkar [10]. We list a few below for easy reference.
1. φn (· , x) denotes the solution to ẋ(t) ∈ hr(n) (x(t)) with initial value x.
2. φ∞ (· , x) denotes the solution to ẋ(t) ∈ h∞ (x(t)) with initial value x.
3. xn (t), t ∈ [0, T ] denotes the solution to ẋn (t) = hr(n) (x̂(Tn + t)) with
initial value xn (0) = x̂(Tn ). Then xn (t) = φn (t, x̂(Tn )), t ∈ [0, T ].
In proving the Borkar-Meyn Theorem as outlined in [13] (BM 1)(ii) is used to
show that for large values of r(n), φn (t, x̂(Tn )) is ‘close’ to φ∞ (t, x̂(Tn )), t ∈
[0, T ]. In this paper we deviate from [13] in the definition of xn (t), t ∈ [0, T ], here
xn (· ) denotes the solution up to time T of ẋn (t) = ŷ(Tn + t) = hr(n) (x̂([Tn + t]))
with xn (0) = x̂(Tn ), where [· ] is defined in Lemma 5. In other words, we have
the following:
Z t
k−1
X Z t(m(n)+l+1)
ŷ(z) dz +
ŷ(z) dz.
xn (t) = x̂(Tn ) +
l=0
t(m(n)+l)
t(m(n)+k)
For t ∈ [tn , tn+1 ), ŷ(t) is a constant and equals ŷ(tn ). We get the following:
xn (t) = x̂(Tn )+
k−1
X
a(m(n)+l)hr(n) (x̂([t(m(n) + l)]))+(t − t(m(n) + k)) hr(n) (x̂([t(m(n) + k)])) .
l=0
The proof now proceeds along the lines of Section 3.2 i.e., Lemmas 1 - 5 and
Theorem 1; we essentially show the following: If r(n) ↑ ∞ then the T -length
trajectories given by {xn (· )}n≥0 have φ∞ (x, t), t ∈ [0, T ], as the limit point in
C([0, T ], Rd ), where x ∈ B 1 (0). This is proven in Lemmas 4 and 5, the proofs
of which do not require (BM 1)(ii).
4.2
Further weakening of (BM1) as a consequence of Theorem 3
In this section we use the second stability theorem (Theorem 3) to answer
the following question: If lim hc (x) does not exist for all x ∈ Rd , then what are
c→∞
the sufficient conditions for the stability and convergence of the algorithm?
Taking our cue from assumption (A6), we replace (BM 1) with the following
assumption, call it (BM 4).
(BM4)(i) The function h : Rd → Rd is Lipschitz continuous, with Lipschitz constant
L. Define the set-valued map, h∞ (x) := co (Limsupc→∞{hc (x)}), where
x ∈ Rd .
Note that Limsupc→∞{hc (x)} = {y | lim khc (x) − yk = 0}.
c→∞
19
(BM4)(ii) ẋ(t) ∈ h∞ (x(t)) has an attracting set, A, with B 1 (0) as a subset of its
fundamental neighborhood. This attracting set is such that A ⊆ B1 (0).
Observe that Limsupc→∞{hc (x)} = lim hc (x) when lim hc (x) exists. Recall
c→∞
c→∞
the definition of Limsup, the upper-limit of a sequence of sets, from Section 2.1.
It can be shown that if a recursion given by (1) satisfies assumptions (BM 1)(i)
and (BM 1)(iii) then it also satisfies (BM 4)(i), (ii). Assumption (BM 4) unifies
the two possible cases: when the limit of hc , as c → ∞, exists for each x ∈ Rd
and when it does not.
We claim that a recursion given by (1), satisfying assumptions (BM 2), (BM 3)
and (BM 4) will also satisfy (A1) − (A3), (A6) and (A5) (see section 3.3). From
Theorem 3 it follows that the iterates are stable and converge to a closed, connected, internally chain transitive and invariant set of ẋ(t) = h(x(t)). The
following generalization of the Borkar-Meyn Theorem is a direct consequence of
Theorem 3.
Corollary 2 (Generalized Borkar-Meyn Theorem). Under assumptions (BM 2),
(BM 3) and (BM 4), almost surely the sequence {xn }n≥0 generated by the stochastic recursive equation (1), is bounded and converges to a closed, connected, internally chain transitive and invariant set of ẋ(t) = h(x(t)).
Proof. Assumptions (A1) − (A3) and (A6) follow directly from (BM 2), (BM 3)
and (BM 4). We show that (A5) is also satisfied. Let cn ↑ ∞, xn → x, yn →
y and yn ∈ hcn (xn ) (here yn = hcn (xn )), ∀ n ≥ 1. It can be shown that
khcn (xn ) − hcn (x)k ≤ Lkxn − xk. Hence we get that hcn (x) → y. In other
words, lim khc (x) − yk = 0. Hence we have y ∈ h∞ (x). The claim now follows
c→∞
from Theorem 3.
5
Applications: The problem of approximate drifts
& stochastic gradient descent
5.1
The problem of approximate drifts
Let us recall the standard SRE:
xn+1 = xn + a(n) (h(xn ) + Mn+1 ) ,
(9)
where h : Rd −→ Rd is Lipschitz continuous, {a(n)}n≥0 is the step-size sequence
and {Mn }n≥1 is the noise sequence.
The function h is colloquially referred to as the drift. In many applications
the drift function cannot be calculated accurately. This is referred to as the
approximate drift problem. For more details the reader is referred to Chapter
5.3 of Borkar [10]. Suppose the room for error is at most ǫ(> 0) then such an
algorithm can be characterized by the following stochastic recursive inclusion:
xn+1 = xn + a(n) (yn + Mn+1 ) ,
20
(10)
where yn ∈ h(xn ) + B ǫ (0) is an estimate of h(xn ) and B ǫ (0) is the closed
ball of radius ǫ around the origin. We define a new set-valued map called the
approximate drift by H(x) := h(x) + B ǫ (0) for each x ∈ Rd . In the following
discussion we assume that ǫ ≥ 0. When ǫ = 0, the approximate drift algorithm
described by (10) is really the SRE given by (9).
In this section we show the following: If (9) satisfies (BM 2), (BM 3) and
(BM 4) then the corresponding approximate drift version given by (10) satisfies
(A1) − (A5). For details on (BM 2) and (BM 3) see Section 4.1; see Section 4.2
for (BM 4). We then invoke Theorem 3 to conclude that the iterates converge
to a closed, connected, internally chain transitive and invariant set associated
with ẋ(t) ∈ h(x(t)) + B ǫ (0)(= H(x(t))).
For the remainder of this section it is assumed that (9) satisfies
(BM 2), (BM 3) and (BM 4).
Proposition 3. H(x) = h(x) + B ǫ (0) is a Marchaud map. Further, recursion
(10) satisfies (A1), (A2) and (A3).
Proof. Since B ǫ (0) is convex and compact, it follows that H(x) is convex and
compact for each x ∈ Rd . Fix x ∈ Rd and y ∈ H(x), then kyk ≤ kh(x)k + ǫ
and kyk ≤ kh(0)k + Lkx − 0k + ǫ since h is Lipschitz continuous with Lipschitz
constant L. If we set K := (kh(0)k + ǫ) ∨ L, then we get kyk ≤ K (1 + kxk).
This shows that H is point-wise bounded. To show the upper semi-continuity
of H assume lim xn = x, lim yn = y and yn ∈ H(xn ) for each n ≥ 1. For all
n→∞
n→∞
n ≥ 1, yn = h(xn ) + zn for some zn ∈ B ǫ (0). Further, h(xn ) → h(x) as xn → x.
Since both {yn }n≥1 and {h(xn )}n≥1 are convergent sequences, {zn }n≥1 is also
convergent. Let z := lim zn ; z is such that z ∈ B ǫ (0) since B ǫ (0) is compact.
n→∞
Taking limits on both sides of yn = h(xn ) + zn , we get y = h(x) + z. Thus
y ∈ H(x).
Since (10) is assumed to satisfy (BM 2) and (BM 3) it trivially follows that it
satisfies (A2) and (A3).
Before showing that (10) satisfies (A4), we construct the following family of
set-valued maps:
y
h(cx)
(11)
+
| y ∈ B ǫ (0) ,
Hc (x) :=
c
c
where c ≥ 1 and x ∈ Rd . In other words, Hc (x) = hc (x) + B ǫ/c (0) for each
x ∈ Rd .
Proposition 4. (10) satisfies (A6).
Proof. To prove this it is enough to show that H∞ (x) = h∞ (x), where H∞ (x) :=
Limsupc→∞Hc (x) and h∞ (x) := Limsupc→∞hc (x). Since ẋ(t) ∈ h∞ (x(t))
satisfies (BM 4)(ii) it trivially follows that (A6) is satisfied by (10). Note that
(BM 4)(ii) and (A6) essentially say the same thing.
21
First we show h∞ (x) ⊆ H∞ (x) for every x ∈ Rd . Let y ∈ h∞ (x), ∃cn ↑ ∞
such that hcn → y as cn ↑ ∞. Since hcn (x) ∈ Hcn (x) it follows from the
definition of Limsup that y ∈ H∞ (x). To show H∞ (x) ⊆ h∞ (x) we start by
assuming the negation i.e., for some x ∈ Rd ∃y ∈ H∞ (x) such that y ∈
/ h∞ (x).
Let cn ↑ ∞ and yn ∈ Hcn (xn ) such that lim yn = y. Since kyn − hcn (xn )k ≤ cǫn
cn ↑∞
we have lim hcn (xn ) = y. We have the following:
cn ↑∞
ky − hcn (x)k ≤ ky − hcn (xn )k + khcn (xn ) − hcn (x)k.
Taking limits on both sides we get that ky − hcn (x)k → 0 i.e., y ∈ h∞ (x). This
is a contradiction.
Proposition 5. (10) satisfies (A5).
Proof. Given cn ↑ ∞, xn → x, yn → y and yn ∈ Hcn (xn ) ∀n, we need to show
that y ∈ H∞ (x). As in the proof of Proposition 4 we have lim hcn (xn ) = y.
cn ↑∞
Since khcn (xn )− hcn (x)k ≤ Lkxn − xk we have that lim khcn (xn )− hcn (x)k = 0
cn ↑∞
and lim hcn (x) = y. In other words, y ∈ h∞ (x). In Proposition 3 we have
cn ↑∞
shown that h∞ ≡ H∞ therefore y ∈ H∞ (x).
Corollary 3. If a SRE, given by (9), satisfies (BM 2), (BM 3) and (BM 4)(i), (ii)
then the corresponding approximate drift version given by (10) is stable almost
surely. In addition, it converges to a closed, connected, invariant and internally
chain transitive set of ẋ(t) ∈ H(x(t)), where H(x) = h(x) + B ǫ (0).
Proof. In Propositions 3, 4 and 5 we have shown that (9) satisfies (A1) −
(A3), (A5), (A6); the statement now follows directly from Theorem 3.
Remark 3. In the context of (9), we have that h is Lipschitz and hc : x 7→ h(cx)
c .
Supposing lim hc (x) exists for every x ∈ Rd (see (BM 1)(i) in Section 4) then
c→∞
lim hc (x) = Limsupc→∞ {hc (x)}. Further, Limsupc→∞{hc (x)} is non-empty
c→∞
for every x ∈ Rd (since hc (x) ≤ K(1 + kxk), c ≥ 1), even if lim hc (x) does not
c→∞
exist for some x ∈ Rd . Hence the analysis of the approximate drift problem in
this section is all encompassing. The aforementioned is also the reason why in
Section 4.2 we define h∞ (x) := co(Limsupc→∞ {hc (x)}). It may be noted that
we use Limsupc→∞ {hc (x)} instead of Limsupc→∞hc (x) since Limsup acts on
sets and h (in this context) is a function that is not set-valued. Finally, in
Corollary 3 if we let ǫ = 0 then we may derive Corollary 2.
22
5.2
Stochastic gradient descent
Stochastic gradient descent is a gradient descent optimization technique to
find the minimum set of a (continuously) differentiable function. Suppose we
want to find the minimum of F : Rd → R for which we can run the following
SRE:
xn+1 = xn − a(n)[∇F (xn ) + Mn+1 ],
(12)
where ∇F : Rd → Rd is upper-semicontinuous and k∇F (x)k ≤ K(1 + kxk)
∀x ∈ Rd (point-wise bounded). {a(n)}n≥0 is the given step size sequence and
{Mn+1 }n≥0 is the martingale difference noise sequence. If the assumptions of
Benaı̈m, Hofbauer and Sorin [7] are satisfied by (12) then the iterates converge
to a closed, connected, internally chain transitive and invariant set of ẋ(t) =
−∇F (x(t)) which is also the minimum set of F . In this section we shall not
distinguish between the asymptotic attracting set of ẋ(t) = −∇F (x(t)) and the
minimum set of F .
As explained in Section 1, while implementing (12) one can only hope to calculate an approximate value of the gradient at each step. However, one has control
over the “approximation error”. This is typical when gradient estimators with
fixed perturbation parameters are used, it could also be a consequence of the
inherent computational capability of the computer used to run the algorithm.
In reality one is running the following SRI:
xn+1 = xn + a(n)[yn + Mn+1 ],
(13)
where yn ∈ −∇F (xn ) + B ǫ (0) and ǫ > 0 is the “approximation error”. The
following questions are natural:
1. Are the iterates stable?
2. If so, where do they converge?
Define the following set valued map, H : x 7→ −∇F (x) + B ǫ (0). As in (11)
(cx)
+ B ǫ/c (0) and H∞ (x) := Limsupc→∞Hc (x) =
we define Hcn(x) := −∇F
oc
−∇F (cx)
Limsupc→∞
. Recall the definition of Limsup from Section 2.1.
c
Proposition 6. (13) satisfies (A1) i.e., H is a marchaud map.
Proof. Given xn → x, yn → y and yn ∈ H(xn ) ∀n, we need to show that
y ∈ H(x). For each n we have yn = −∇F (xn ) + zn , where zn ∈ B ǫ (0). Since
∇F is point-wise bounded, it follows that {−∇F (xn )} is a bounded sequence.
Let {n(m)} ⊆ N such that ∇F (xn(m) ) → ∇F (x), yn(m) → y. The subsequence
zn(m) → z for some z ∈ B ǫ (0) i.e.,
−∇F (xn(m) ) + zn(m) → (−∇F (x) + z) ∈ H(x).
If in addition to (A1), equation (13) also satisfies (A2), (A3), (A5) and (A6)
then it follows from Theorem 3 that the iterates are stable and converge to
a closed, connected, internally chain transitive and invariant set of ẋ(t) ∈
−∇F (x(t)) + B ǫ (0) .
23
Suppose F has the quadratic form xT Ax+Bx+c, where A is a positive
definite matrix, B is some matrix and c is some vector. Then it can
be shown that (A1), (A2), (A3), (A5) and (A6) are satisfied by (13) and
the iterates are stable and converge to a closed, connected, internally
chain transitive and invariant set of ẋ(t) ∈ −(Ax(t) + B) + B ǫ (0). If the
comments in Remark 1 are incorporated i.e., we use (A6)′ instead of
(A6) then matrix A need not be positive definite anymore.
For the purpose of this discussion assume that ∇F is Lipschitz continuous. The graph of a set-valued map H : Rd → {subsets of Rd } is given by
Graph(H) = {(x, y) | x ∈ Rd , y ∈ H(x)}. It is easy to see that Graph(−∇F +
B ǫ (0)) ⊆ N 2ǫ (Graph(−∇F )). Let us also assume that A is the global attractor
(minimum set of F ) of ẋ(t) = −∇F (x(t)) then every compact subset of Rd is
its fundamental neighborhood. It follows from the stability of the iterates that
they will remain within a compact subset, say U, that may be sample path
dependent. It follows from Theorem 2.1 of Benaı̈m, Hofbauer and Sorin [8] that
for all δ > 0 there exists ǫ > 0 such that Aδ ⊆ N δ (A) is the attractor set of
ẋ(t) ∈ −∇F (x(t)) + B ǫ (0). Further, the fundamental neighborhood of Aδ is U
itself. In other words, suppose we want to ensure convergence of the iterates
to a δ − neighborhood of the minimum set A then the “approximation error”
should be at most ǫ (ǫ is dependent on δ).
6
Final discussion on the generality of our framework
As explained in Section 3, we run a projective scheme to show stability. In
other words, time is divided into intervals of length T ; the iterates are checked
at the beginning of each time interval to see if they are outside the unit ball; all
the iterates corresponding to [Tn , Tn+1 ) are scaled by r(n) = kx(Tn )k ∨ 1 i.e.,
the iterates are projected onto the unit ball around the origin. For t(m(n)) =
Tn ≤ t(m(n) + k) < Tn+1 we have the following re-scaled iterate:
k−1
x(t(m(n) + k))
x(t(m(n))) X
y(t(m(n) + j)) Mm(n)+j+1
a(m(n)+j)
.
=
+
+
r(n)
r(n)
r(n)
r(n)
j=0
In the above,
y(t(m(n)+j))
r(n)
∈
hr(n)
x(t(m(n)+j))
r(n)
. Since we have to worry
about r(n) running off to infinity it is natural to define h∞ (x) to include all
accumulation points of {hc (x) | c ≥ 1, c → ∞}. This is precisely what the
Limsup function (see Section 2.1) allows us to do. In Lemma 5 it was shown
that the scaled iterates track a solution to ẋ(t) ∈ h∞ (x(t)) provided the original
iterates are unstable i.e., sup r(n) = ∞. Assumptions (A4)/(A6) were never
n
used up to this point. At this stage it seems natural to impose restrictions on
x(t) ∈ h∞ (x(t)) to elicit the stability of the original iterates.
As explained in Section 3.3 Limsupc→∞hc (x) is non-empty for every x ∈ Rd
since h is point-wise bounded. Further, h∞ ≡ co (Limsupc→∞hc ) is shown to
be Marchaud and the DI ẋ(t) ∈ h∞ (x(t)) has at least one solution. Assumption
24
(A6) is the restriction referred to in the previous paragraph that is imposed to
elicit the stability of the original iterates. On a related note, if Liminfc→∞ hc
were non-empty, then we define h∞ ≡ Liminfc→∞hc and check if (A4) is satisfied.
If the DI ẋ(t) ∈ h∞ (x(t)) has global attractor inside B1 (0), then this is a
sufficient condition for (A6) to hold, it then follows from Theorem 3 that the
original iterates are stable and converge to a closed connected internally chain
transitive set associated with ẋ(t) ∈ h(x(t)). More generally, in lieu of Remark 1
it is sufficient that the DI has some global attractor, not necessarily inside the
unit ball, since (A6)′ will then hold. This in turn implies stability.
In case of the original Borkar-Meyn assumptions, (BM 1)(i), (ii) (see Section 4) needed to be checked even before we could define h∞ while in our case
we do not need any extra assumptions to define h∞ . As explained before, constructing a global Lyapunov function for h∞ is one of many sufficient conditions
that guarantee (A4)′ /(A6)′ . In case of Lyapunov-type conditions for stability,
additional properties of the constructed global Lyapunov function need to be
verified before we get stability, see [1] for more details. However, to the best of
our knowledge, there are no Lyapunov-type conditions that guarantee stability
of stochastic approximation algorithms with set-valued mean fields (SRI), the
class of algorithms dealt with in this paper. Hence our assumptions are general
and relatively easy to verify.
7
Conclusions
An extension was presented to the theorem of Borkar and Meyn to include
approximation algorithms with set-valued mean fields. Two different sets of
sufficient conditions were presented that guarantee the ‘stability and convergence’ of stochastic recursive inclusions. As a consequence of Theorems 2 &
3, the original Borkar-Meyn theorem is shown to hold under weaker requirements. Further, as a consequence of Theorem 3, we obtained a solution to
the “approximate drift” problem. Prior to this paper, there was no proof of
stability of stochastic gradient descent algorithms that use constant-error gradient estimators. Hence we could only conclude that the iterates converge to a
small neighborhood, say N , of the minimum set with very high probability. In
Section 5.2 we used our framework to show the stability of the aforementioned
algorithm which in turn allowed us to conclude an almost sure convergence to
N.
An important future direction would be to extend these results to the case
when the set-valued drift is governed by a Markov process in addition to the
iterate sequence. For the case of stochastic approximations, such a situation
has been considered in [ [10], Chapter 6], where the Markov ‘noise’ is tackled
using the ‘natural timescale averaging’ properties of stochastic approximation.
Finally, it would be interesting to develop Lyapunov-type assumptions for stability of stochastic algorithms with set-valued mean fields.
25
8
Acknowledgments
This work was partly supported by Robert Bosch Centre for Cyber-Physical
Systems, IISc.
References
[1] C. Andrieu, E. Moulines, and P. Priouret. Stability of stochastic approximation under verifiable conditions. SIAM J. Control Optim., 44(1):283–312,
2005.
[2] J. Aubin and A. Cellina. Differential Inclusions: Set-Valued Maps and
Viability Theory. Springer, 1984.
[3] J. Aubin and H. Frankowska. Set-Valued Analysis. Birkhäuser, 1990.
[4] M. Benaı̈m. A dynamical system approach to stochastic approximations.
SIAM J. Control Optim., 34(2):437–472, 1996.
[5] M. Benaı̈m. Dynamics of stochastic approximation algorithms. Seminaire
de probabilites XXXIII, pages 1–68, 1999.
[6] M. Benaı̈m and M. W. Hirsch. Asymptotic pseudotrajectories and chain
recurrent flows, with applications. J. Dynam. Differential Equations, 8:141–
176, 1996.
[7] M. Benaı̈m, S. Sorin, and J. Hofbauer. Stochastic approximations and
differential inclusions. SIAM Journal on Control and Optimization, pages
328–348, 2005.
[8] M. Benaı̈m, S. Sorin, and J. Hofbauer. Perturbations of set-valued dynamical systems, with applications to game theory. Dynamic Games and
Applications, 2(2):195–205, 2012.
[9] S. Bhatnagar, H.L. Prasad, and L.A. Prashanth. Stochastic recursive algorithms for optimization: simultaneous perturbation methods, volume 434.
Springer, 2012.
[10] V. S. Borkar. Stochastic Approximation: A Dynamical Systems Viewpoint.
Cambridge University Press, 2008.
[11] H. Kushner and G.G. Yin. Stochastic Approximation and Recursive Algorithms and Applications. Springer, 2003.
[12] L. Ljung. Analysis of recursive stochastic algorithms. Automatic Control,
IEEE Transactions on, 22(4):551–575, 1977.
[13] S.P. Meyn V.S. Borkar. The O.D.E. method for convergence of stochastic approximation and reinforcement learning. SIAM J. Control Optim,
38:447–469, 1999.
26
| 3cs.SY
|
arXiv:1602.04484v5 [cs.LG] 19 Apr 2017
Surprising properties of dropout
in deep networks
David P. Helmbold
Department of Computer Science
University of California, Santa Cruz
Santa Cruz, CA 95064, USA
[email protected]
Philip M. Long
Google
[email protected]
April 21, 2017
Abstract
We analyze dropout in deep networks with rectified linear units and the quadratic loss. Our results
expose surprising differences between the behavior of dropout and more traditional regularizers like weight
decay. For example, on some simple data sets dropout training produces negative weights even though the
output is the sum of the inputs. This provides a counterpoint to the suggestion that dropout discourages
co-adaptation of weights. We also show that the dropout penalty can grow exponentially in the depth of
the network while the weight-decay penalty remains essentially linear, and that dropout is insensitive to
various re-scalings of the input features, outputs, and network weights. This last insensitivity implies that
there are no isolated local minima of the dropout training criterion. Our work uncovers new properties
of dropout, extends our understanding of why dropout succeeds, and lays the foundation for further
progress.
1
Introduction
The 2012 ImageNet Large Scale Visual Recognition challenge was won by the University of Toronto team by
a surprisingly large margin. In an invited talk at NIPS, Hinton [15] credited the dropout training technique
for much of their success. Dropout training is a variant of stochastic gradient descent (SGD) where, as each
example is processed, the network is temporarily perturbed by randomly “dropping out” nodes of the network.
The gradient calculation and weight updates are performed on the reduced network, and the dropped out
nodes are then restored before the next SGD iteration. Since the ImageNet competition, dropout has been
successfully applied to a variety of domains [8, 10, 9, 17, 7], and is widely used [21, 13, 23]; for example, it
is incorporated into popular packages such at Torch [25], Caffe [6] and TensorFlow [24]. It is intriguing that
crippling the network during training often leads to such dramatically improved results, and dropout has
also sparked substantial research on related methods (e.g. [12, 29]).
In this work, we examine the effect of dropout on the inductive bias of the learning algorithm. A match
between dropout’s inductive bias and some important applications could explain the success of dropout, and
its popularity also motivates the study of its unusual properties.
Weight decay training optimizes the empirical error plus an L2 regularization term, λ2 ||w||22 , so we call
λ
2
2 ||w||2 the L2 penalty of w since it is the difference between training criterion evaluated at w and the
empirical loss of w. By analogy, we define the dropout penalty of w to be the difference between the dropout
training criterion and the empirical loss of w (see Section 2). Dropout penalties measure how much dropout
discriminates against weight vectors, so they are key to understanding dropout’s inductive bias.
Even in one-layer networks, conclusions drawn from (typically quadratic) approximations of the dropout
penalty can be misleading [14]. Therefore we focus on exact formal analysis of dropout in multi-layer
1
networks. Theoretical analysis of deep networks is notoriously difficult, so we might expect that a thorough
understanding of dropout in deep networks must be achieved in stages. In this paper we further the process
by exposing some of the surprising ways that the inductive bias of dropout differs from L2 and other standard
regularizers. These include the following:
• We show that dropout training can lead to negative weights even when the output is a positive multiple
of the the inputs. Arguably, such use of negative weights constitutes co-adaptation – this adds a
counterpoint to previous analyses showing that dropout discourages co-adaptation [22, 14].
• Unlike weight decay and other p-norm regularizers, dropout training is insensitive to the rescaling of
input features, and largely insensitive to rescaling of the outputs; this may play a role in dropout’s
practical success. Dropout is also unaffected if the weights in one layer are scaled up by a constant
c, and the weights of another layer are scaled down by c; this implies that dropout training does not
have isolated local minima.
• The dropout penalty grows exponentially in the depth of the network in cases where the L2 regularizer
grows linearly. This may enable dropout to penalize the complexity of the network in a way that more
meaningfully reflects the richness of the network’s behaviors. (The exponential growth with d of the
dropout penalty is reminiscent of some regularizers for deep networks studied by Neyshabur, et al [20].)
• Dropout in deep networks has a variety of other behaviors different from standard regularizers. In particular: the dropout penalty for a set of weights can be negative; the dropout penalty of a set of weights
depends on both the training instances and the labels; and although the dropout probability intuitively
measures the strength of dropout regularization, the dropout penalties are often non-monotonic in the
dropout probability. In contrast, Wager, et al [28] show that when dropout is applied to generalized
linear models, the dropout penalty is always non-negative and does not depend on the labels.
Our analysis is for multilayer neural networks with the square loss at the output node. The hidden layers
use the popular rectified linear units [19] outputting σ(a) = max(0, a) where a is the node’s activation (the
weighted sum of its inputs). We study the minimizers of a criterion that may be viewed as the objective
function when using dropout. This abstracts away sampling and optimization issues to focus on the inductive
bias, as in [5, 30, 4, 18, 14]. See Section 2 for a complete explanation.
Related work
A number of possible explanations have been suggested for dropout’s success. Hinton, et al [16] suggest
that dropout controls network complexity by restricting the ability to co-adapt weights and illustrate how
it appears to learn simpler functions at the second layer. Others [3, 1] view dropout as an ensemble method
combining the different network topologies resulting from the random deletion of nodes. Wager, et al [27]
observe that in 1-layer networks dropout essentially forces learning on a more challenging distribution akin
to ‘altitude training’ of athletes.
Most formal analysis of the inductive bias of dropout has concentrated on the single-layer setting, where
a single neuron combines the (potentially dropped-out) inputs. Wager, et al [28] considered the case that
the distribution of label y given feature vector x is a member of the exponential family, and the log-loss
is used to evaluate models. They pointed out that, in this situation, the criterion optimized by dropout
can be decomposed into the original loss and a term that does not depend on the labels. They then gave
approximations to this dropout regularizer and discussed its relationship with other regularizers. As we have
seen, many aspects of the behavior of dropout and its relationship to other regularizers are qualitatively
different when there are hidden units.
Wager, et al [27] considered dropout for learning topics modeled by a Poisson generative process. They
exploited the conditional independence assumptions of the generative process to show that the excess risk of
dropout training due to training set variation has a term that decays more rapidly than the straightforward
empirical risk minimization, but also has a second additive term related to document length. They also
discussed situations where the model learned by dropout has small bias.
2
Baldi and Sadowski [2] analyzed dropout in linear networks, and showed how dropout can be approximated by normalized geometric means of subnetworks in the nonlinear case. Gal and Ghahramani [11]
described an interpretation of dropout as an approximation to a deep Gaussian process.
The impact of dropout (and its relative dropconnect) on generalization (roughly, how much dropout
restricts the search space of the learner) was studied in [29].
In the on-line learning with experts setting, Van Erven, et al [26] showed that applying dropout in on-line
trials leads to algorithms that automatically adapt to the input sequence without requiring doubling or other
parameter-tuning techniques.
The rest of the paper is organized as follows. Section 2 introduces our notation and formally defines
the dropout model. We prove that dropout enjoys several scaling invariances that weight-decay doesn’t in
Section 3, and that dropout requires negative weights even in very simple situations in Section 4. Section 5
uncovers various properties of the dropout penalty function. Section 6 describes some simulation experiments.
We provide some concluding remarks in Section 7.
2
Preliminaries
Throughout, we will analyze fully connected layered networks with K inputs, one output, d layers (counting
the output, but not the inputs), and n nodes in each hidden layer. We assume that n is a positive multiple
of K and that K is an even perfect square and a power of two to avoid unilluminating floor/ceiling clutter
in the analysis. We will call this the standard architecture. We use W to denote a particular setting of the
weights and biases in the network and W(x) to denote the network’s output on input x using W. The hidden
nodes are ReLUs, and the output node is linear. W can be decomposed as (W1 , b1 , ..., Wd−1 , bd−1 , w, b),
where each Wj is the matrix of weights on connections from the j − 1st into the jth hidden layer, each bj
is the vector of bias inputs into the jth hidden layer, w are the weights into the output node, and b is the
bias into the output node.
We will refer to a joint probability distribution over examples (x, y) as an example distribution. We focus
on square loss, so the loss of W on example (x, y) is (W(x) − y)2 . The risk is the expected loss with respect
def
to an example distribution P , we denote the risk of W as RP (W) = E(x,y)∼P (W(x) − y)2 . The subscript
will often be omitted when P is clear from the context.
The goal of L2 training is to find weights and biases minimizing the L2 criterion with regularization
def
strength λ: J2 (W) = R(W) + λ2 ||W||2 . Here and throughout, we use ||W||2 to denote the sum of the squares
of the weights of W. (As usual, the biases are not penalized.) We use WL2 to denote a minimizer of this
criterion. The L2 penalty, λ2 ||W||2 , is non-negative. This is useful, for example, to bound the risk of a
minimizer WL2 of J2 , since R(W) ≤ J2 (W).
Dropout training independently removes nodes in the network. In our analysis each non-output node is
dropped out with the same probability q, so p = 1 − q is the probability that a node is kept. (The output
node is always kept; dropping it out has the effect of cancelling the training iteration.) When a node is
dropped out, the node’s output is set to 0. To compensate for this reduction, the values of the kept nodes
are multiplied by 1/p. With this compensation, the dropout can be viewed as injecting zero-mean additive
noise at each non-output node [28]. 1
The dropout process is the collection of random choices, for each node in the network, of whether the
node is kept or dropped out. A realization of the dropout process is a dropout pattern, which is a boolean
vector indicating the kept nodes. For a network W, an input x, and dropout pattern R, let D(W, x, R)
be the output of W when nodes are dropped out or not following R (including the 1/p rescaling of kept
nodes’ outputs). The goal of dropout training on an example distribution P is to find weights and biases
minimizing the dropout criterion for a given dropout probability:
def
JD (W) = ER E(x,y)∼P (D(W, x, R) − y)2 .
1 Some authors use a similar adjustment where the weights are scaled down at prediction time instead of inflating the kept
nodes’ outputs at training time.
3
output
1
1
1
1
1
1
1
-1
x1
x2
Figure 1: A network where the dropout penalty is negative.
This criterion is equivalent to the expected risk of the dropout-modified network, and we use WD to denote
a minimizer of it. Since the selection of dropout pattern and example from P are independent, the order of
the two expectations can be swapped, yielding
(1)
JD (W) = E(x,y)∼P ER (D(W, x, R) − y)2 .
Equation (1) is a key property of the dropout criterion. It indicates when something is true about the dropout
criterion for a family of distributions concentrated on single examples, then (usually) the same thing will be
true for any mixture of these single-example distributions.
Consider now the example network in Figure 1. The weight parameters W1 and w are all-1’s, and all of
the biases are 0. W(1, −1) = 0 as each hidden node computes 0. Each dropout pattern indicates the subset
of the four lower nodes to be kept, and when q = p = 1/2 each subset is equally likely to be kept. If R is
the dropout pattern where input x2 is dropped and the other nodes are kept, then the network computes
D(W, (1, −1), R) = 8 (recall that when p = 1/2 the values of non-dropped out nodes are doubled). Only
three dropout patterns produce a non-zero output, so if P is concentrated on the example x = (1, −1), y = 8
the dropout criterion is:
2
13
1
(8 − 8)2 + (4 − 8)2 + (0 − 8)2 = 54.
(2)
16
16
16
As mentioned in the introduction, the dropout penalty of a weight vector for a given example distribution
and dropout probability is the amount that the dropout criterion exceeds the risk, JD (W) − R(W). Wager,
et al [28] show that for 1-layer generalized linear models, the dropout penalty is non-negative.
Since W(1, −1) = 0, we have R(W) = 64, and the dropout penalty is negative in our example. This is
because the variance in the output due to dropout causes the network to better fit the data (on average) than
the network’s non-dropout evaluation. In Section 5.2, we give a necessary condition for this variance to be
beneficial. As with the dropout criterion, the dropout penalty decomposes into an expectation of penalties
over single examples:
JD (W) − R(W) = E(x,y)∼P ER (D(W, x, R) − y)2 − (W(x) − y)2 .
JD (W) =
Definition 1 Define P(x,y) as the distribution with half of its weight on example (x, y) and half of its weight
on example (0, 0).
Unless indicated otherwise, we assume p = q = 1/2 for simplicity, although this is not crucial for our
results.
3
3.1
Scaling inputs, weights and outputs
Dropout is scale-free
Here we prove that dropout regularizes deep networks in a manner that is independent of the scale of the
input features. In other words, training under dropout regularization does not penalize the use of large
4
weights when needed to compensate for small input values.
Definition 2 For any example distribution P , define the dropout aversion with P to be the maximum, over
minimizers WD of the dropout risk JD (WD ), of RP (WD ) − inf W RP (W).
The dropout aversion of P measures the extent to which P is incompatible with the inductive bias
of dropout, measured by the risk gap between the true risk minimizer and the optimizers of the dropout
criterion.
Definition 3 For example distribution P and square matrix A, denote by A ◦ P the distribution obtained by
sampling (x, y) from P , and outputting (Ax, y).
When A is diagonal and has full rank, then A ◦ P is a rescaling of the inputs, like changing one input
from minutes to seconds and another from feet to meters.
Theorem 4 For any example distribution P , and any diagonal full-rank K × K matrix A, the dropout
aversion of P equals the dropout aversion of A ◦ P .
Proof: Choose a network W = (W1 , b1 , ..., Wd−1 , bd−1 , w, b). Let W ′ = (W1 A−1 , b1 ..., Wd−1 , bd−1 , w, b).
For any x, W(x) = W ′ (Ax), as A−1 undoes the effect of A before it gets to the rest of the network, which
is unchanged. Furthermore, for any dropout pattern R, we have D(W, x, R) = D(W ′ , Ax, R). Once again
A−1 undoes the effects of A on kept nodes (since A is diagonal), and the rest of the network W ′ is modified
by R an a manner paralleling W. Thus, there is bijection between networks W and networks W ′ with
JD (W ′ ) = JD (W) and R(W ′ ) = R(W), yielding the theorem.
Theorem 4 indicates that some common normalizations of the input features (e.g. to have unit variance)
do not affect the quality of the dropout criterion minimizers, but normalization might change the speed
of convergence and which minimizer is reached. Centering the features has slightly different properties.
Although it is easy to use the biases to define a W ′ that “undoes” the centering in the non-dropout computation, different W ′ appear to be required for different dropout patterns, breaking the bijection exploited in
Theorem 4.
As we will see in Section 3.4, weight decay does not enjoy such scale-free status.
3.2
Dropout’s invariance to parameter scaling
Next, we describe an equivalence relation among parameterizations for dropout networks of depth d ≥
2. Basically, scaling the parameters at a level creates a corresponding scaling of the output. (A similar
observation was made in a somewhat different context by Neyshabur, et al [20].)
Theorem 5 For any input x, dropout pattern R, any network W = (W1 , b1 , ..., Wd−1 , bd−1 , w, b), and any
positive c1 , ..., cd , if
d−1
d
Y
Y
W ′ = c1 W1 , c1 b1 , c2 W2 , c1 c2 b2 , ..., cd−1 Wd−1 ,
cj bd−1 , cd w,
cj b ,
(3)
j=1
then D(W ′ , x, R) =
Q
d
j=1 cj
′
D(W, x, R). In particular, if
Qd
j=1 cj
j=1
= 1, then for any example distribution
P , networks W and W have the same dropout criterion, dropout penalty, and expected loss.
Note that the re-scaling of the biases at layer j depends not only on the rescaling of the connection
weights at layer j, but also the re-scalings at lower layers.
Proof: Choose an input x and a dropout pattern R. Define W ′ as in (3). For each hidden layer j,
let (hj1 , ..., hjn ) be the jth hidden layer when applying W to x with
QR, and let (h̃j1 , ..., h̃jn ) be the jth
′
hidden layer when applying W instead. By induction, for all i, h̃ji =
ℓ≤j cℓ hji ; the key step is that the
5
pre-rectified value used to compute h̃ji has the same sign as for hji , since
Q rescaling by cj preserves
Qthe sign.
Thus the same units are zeroed out in W and W ′ and D(W, x, R) = ( j cj )D(W ′ , x, R). When j cj = 1,
this implies D(W, x, R) = D(W ′ , x, R). Since this is true for all x and R, we have JD (W) = JD (W ′ ). Since,
similarly, W(x) = W ′ (x) for all x, so R(W) = R(W ′ ), the dropout penalities for W and W ′ are also the
same.
Theorem 5 implies that the dropout criterion never has isolated minimizers, since one can continuously
up-scale the weights on one layer with a compensating down-scaling at another layer to get a contiguous
family of networks computing the same function and having the same dropout criterion. It may be possible to
exploit the parameterization equivalence of Theorem 5 in training by using canonical forms for the equivalent
networks or switching to an equivalent networks whose gradients have better properties. We leave this
question for future work.
3.3
Output scaling with dropout
Scaling the output values of an example distribution P does affect the aversion, but in a very simple and
natural way.
Theorem 6 For any example distribution P , if P ′ is obtained from P by scaling the outputs of P by a
positive constant c, the dropout aversion of P ′ is c2 times the dropout aversion of P .
Proof: If a network W minimizes the dropout criterion for P , then the network W ′ obtained by scaling up
the weights and bias for the output unit by c minimizes the dropout criterion for P ′ , and for any x, y, and
dropout pattern R, (D(W, x, R) − y)2 = (D(W ′ , x, R) − cy)2 /c2 .
3.4
Scaling properties of weight decay
Weight decay does not have the same scaling properties as dropout. Define the weight-decay aversion
analogously to the dropout aversion.
We analyze the L2 criterion for depth-2 networks in Appendix B, resulting in the following theorem. Our
proof shows that, despite the non-convexity of the L2 criterion, it is still possible to identify a closed form
for one of its optimizers.
Theorem 7 Choose an arbitrary number n of hidden nodes, and λ > 0.
λ2
).
The weight-decay aversion of P(x,1) is min( 41 , x·x
Theorem 7 shows that, unlike dropout, the weight decay aversion does depend on the scaling of the input
features.
√
Furthermore, when 2λ > x · x, the weight-decay criterion for P(x,1) has only a single isolated optimum
weight setting2 – all weights set to zero and bias 1/2 at the output node. This means that weight-decay
in 2-layer networks can completely regularize away significant signal in the sample even when λ is finite,
contrasting starkly with weight-decay’s behavior in 1-layer networks.
The “vertical” flexibility to rescale weights between layers enjoyed by dropout (Theorem 5) does not hold
for L2 : one can always drive the L2 penalty to infinity by scaling one layer up by a large enough positive
c, even while scaling another down by c. On the other hand, the proof of Theorem 7 shows that the L2
criterion has an alternative “horizontal” flexibility involving the rescaling of weights across nodes on the
hidden layer (under the theorem’s assumptions). Lemma 32 shows that at the optimizers each hidden node’s
contributions to the output are a constant (depending on the input) times their contribution to the the L2
penalty. Shifting the magnitudes of these contributions between hidden nodes leads to alternative weights
2 Since the output node puts weight 0 on each hidden node and the biases are unregularized, this optimum actually represents
a class of networks differing only in the irrelevant biases at the hidden nodes. One can easily construct other cases when weightdecay has isolated minima in this sense, for example when n = 2 and there is equal probability on x and −x, both with label
1.
6
that compute the same value and have the same weight decay penalty. This is a more general observation
than the permutation symmetry between hidden nodes because any portion of a hidden node’s contribution
can be shifted to another hidden node.
4
Negative weights for monotone functions
If the weights of a unit are non-negative, then the unit computes a monotone function, in the sense that
increasing any input while keeping the others fixed increases the output. The bias does not affect a node’s
monotonicity. A network of monotone units is also monotone.
We first present our theoretical results for many features (Section 4.1) and few features (Section 4.2),
and then discuss the implication of these results in Section 4.3.
4.1
The basic case – many features
In this section, we analyze the simple distribution P(1,1) that assigns probability 1/2 to the example
(0, ..., 0), 0, and probability 1/2 to the example (1, ..., 1), 1. This is arguably the simplest monotone function. Nevertheless, we prove that dropout uses negative weights to fit this data.
The key intuition is that optimizing the dropout criterion requires controlling the variance. Negative
weights at the hidden nodes can be used to control the variance due to dropout at the input layer. When
there are enough hidden nodes this becomes so beneficial that every minimizer of the dropout criterion uses
such negative weights.
Theorem 8 For the standard architecture, if K > 18 and n is large enough relative to K and d, every
optimizer of the dropout criterion for P(1,1) uses at least one negative weight.
To prove Theorem 8, we first calculate JD (Wneg ) for a network Wneg that uses negative weights, and
then prove a lower bound greater than this value that holds for all networks using only non-negative weights.
All of the biases in Wneg are 0.
A key building block in the definition of Wneg is a block of hidden units that we call the first-one gadget.
Each such block has K hidden nodes, and takes its input from the K input nodes. The ith hidden node
in the block takes the value 1 if the Kth input node is 1, and all inputs xi′ for i′ < i are 0. This can be
accomplished with a weight vector w with wi′ = −1 for i′ < i, with wi = 1, and with wi′ = 0 for i′ > i. The
first hidden layer of Wneg comprises n/K copies of the first-one gadget.
Informally, this construction removes most of the variance in the number of 1′ s in the input, as recorded
in the following lemma.
Lemma 9 On any input x ∈ {0, 1}n except (0, 0, ..., 0), the sum of the values on the first hidden layer of
Wneg is exactly n/K.
The weights into the remaining hidden layers of Wneg are all 1, and all the weights into the output
def
layer take a value c =
K2
,
1 d−2
2nd−1 (1+ K
n )(1+ n )
following lemma analyzes Wneg .
1
Lemma 10 JD (Wneg ) = 2 1 −
chosen to minimize the dropout criterion for the network. The
(1−2−K )
d−2
(1+ Kn )(1+ n1 )
.
d−2
1 + n1
denominator in Lemma 10 approaches 1, so
When n is large relative to K and d, the 1 + K
n
JD (Wneg ) approaches 2−K /2 in this case. Lemma 16 below gives a larger lower bound for any network with
all positive weights. In the concrete case when d = 2 and n = K 3 , then Lemma 10 implies JD (Wneg ) < 1/K 2 .
Proof (of Lemma 10): Consider a computation of Wneg (1, 1, ..., 1) under dropout and let ŷ be the
(random) output. Let k0 be the number of input nodes kept, and, for each j ≥ 2, let kj be the number of
7
nodes in the jth hidden layer kept. Call the node in each first-one gadget that computes 1 a key node, and
if no node in the gadget computes 1 because the input is all dropped, arbitrarily make the gadget’s first
hidden node the key node. This ensures there is exactly one key node per gadget, and every non-key node
computes 0. Let k1 be the number
kept key nodes on the first hidden layer. If k0 = 0, the output ŷ of the
Qof
d−1
network is 0. Otherwise, ŷ = c2d j=1 kj .
Note that k0 is zero with probability 2−K . Whenever k0 ≥ 1, k1 is distributed as B(n/K, 1/2). Each
other kj is distributed as B(n, 1/2), and k1 , k2 , ..., kd−1 are independent of one another.
E(ŷ) = Pr(k0 ≥ 1)c2d E[k1 |k0 ≥ 1]
d−1
Y
j=2
E[kj ] = (1 − 2−K )c2d
n n d−2 2c
= (1 − 2−K )nd−1 .
2K
2
K
Using the value of the second moment of the binomial, we get
2
d−1
n n
Y
kj = 4c2 (1 − 2−K )
E(ŷ 2 ) = E 1k0 ≥1 c2d
+ 1 nd−2 (n + 1)d−2
K
K
j=1
=
d−2
4c2 (1 − 2−K ) 2(d−1)
1
K
1
+
1
+
.
n
K2
n
n
Thus,
1
1 − 2E(ŷ) + E(ŷ 2 )
2
d−2 !
2
−K
4c
4c
(1
−
2
)
1
1
K
1 − (1 − 2−K )nd−1 +
n2(d−1) 1 +
=
1+
2
K
K2
n
n
!
(1 − 2−K )
1
1−
=
d−2 ,
2
1+ K 1+ 1
JD (Wneg ) =
n
since c =
K
d−2
1
2nd−1 (1+ K
n )(1+ n )
n
, completing the proof.
Next we prove a lower bound on JD for networks with nonnegative weights. Let W be an arbitrary such
network.
Our lower bound will use a property of the function computed by W that we now define.
Definition 11 A function φ : RK → R is supermodular if for all x, δ 1 , δ 2 ∈ RK where δ 1 , δ 2 ≥ 0
φ(x) + φ(x + δ 1 + δ 2 ) ≥ φ(x + δ 1 ) + φ(x + δ 2 ),
or equivalently:
φ(x + δ 1 + δ 2 ) − φ(x + δ 2 ) ≥ φ(x + δ 1 ) − φ(x)
The latter indicates that adding δ 1 to the bigger input x + δ 2 has at least as large an effect as adding it to
the smaller input x.
Since W has all non-negative weights, it computes a supermodular function of its inputs. (This fact may
be of independent interest.)
Lemma 12 If a network has non-negative weights and its activation functions σ(·) are convex, continuous,
non-decreasing, and differentiable except on a finite set, then the network computes a supermodular function
of its input x.
8
Proof: We will prove by induction over the layers that, for any unit h in the network, if h(x) is the output
of unit h when x is the input to W, then h(·) is a supermodular function of its input.
The base case holds since each input node h outputs the corresponding component of the input, and
(x + δ 1 ) − x = (x + δ 1 + δ 2 ) − (x + δ 2 ).
Now, for the inductive step, let w be the weight vector for node h, let b be its bias, and σ(·) be its
activation function. Let I(x), I(x + δ 1 ), I(x + δ 2 ), and I(x + δ 1 + δ 2 ) be the inputs to node h when the
inputs to the network are x, x + δ 1 , x + δ 2 and x + δ 1 + δ 2 respectively.
By induction, these inputs to node h (componentwise) satisfy
I(x + δ 1 ) − I(x) ≤ I(x + δ 1 + δ 2 ) − I(x + δ 2 ).
Therefore, since w, δ 1 , and δ 2 are non-negative, the interval [w · I(x + δ 2 ) + b, w · I(x + δ 1 + δ 2 ) + b] is at
least as long and starts at least as high as the interval [w · I(x) + b, w · I(x + δ 1 ) + b]. Since σ is continuous
and differentiable except on a finite set, we have
Z w·I(x+δ1 )+b
h(I(x + δ 1 )) − h(I(x)) =
σ ′ (z)dz
w·I(x)+b
≤
Z
w·I(x+δ 1 +δ 2 )+b
σ ′ (z)dz (since σ ′ is non-decreasing)
w·I(x+δ 2 )+b
= h(I(x + δ 1 + δ 2 )) − h(I(x + δ 2 )).
Definition 13 Let r0 ∈ {0, 1}K be the dropout pattern concerning the input layer, and let R′ be the dropout
pattern concerning the rest of the network, so that the dropout pattern R = (r0 , R′ ).
For each ℓ ∈ {0, ..., K}, let ψW (ℓ) be the average output of W under dropout when ℓ of the inputs are
kept: i.e.,
X
r0j = ℓ .
ψW (ℓ) = E D(W, 1K , (r0 , R′ ))
j
Lemma 14 For any ℓ ∈ {1, ..., K − 1, },
ψW (ℓ + 1) − ψW (ℓ) ≥ ψW (ℓ) − ψW (ℓ − 1).
Proof: Generate u, i and j randomly by, first choosing u uniformly at random from among bit vectors with
ℓ ones, then choosing i uniformly from the 0-components of u, and j uniformly from the 1-components of u.
By Lemma 12,
W(u + ei ) − W(u) ≥ W(u − ej + ei ) − W(u − ej )
(4)
always holds. Furthermore, u + ei is uniformly distributed among bit vectors with ℓ + 1 ones, u − ej is
uniformly distributed among bit vectors with ℓ − 1 ones, and u + ei − ej is uniformly distributed among bit
vectors with ℓ ones. This is true for W, but it is also true for any network obtained by dropping out some
of the hidden nodes of W. Thus
ψW (ℓ + 1) − ψW (ℓ)
= E(D(W, 1K , (r0 , R′ )|
= E(D(W, r0 , (1K , R′ )|
≥ E(D(W, r0 , (1K , R′ )|
= ψW (ℓ) − ψW (ℓ − 1),
X
j
X
j
X
j
r0j = ℓ + 1)) − E(D(W, 1K , (r0 , R′ )|
r0j = ℓ + 1)) − E(D(W, r0 , (1K , R′ )|
r0j = ℓ)) − E(D(W, r0 , (1K , R′ )|
9
X
j
X
r0j = ℓ))
j
X
r0j = ℓ))
j
r0j = ℓ − 1)) (by (4))
completing the proof.
We will use the following lower bound on the tail of the binomial. (Many similar lower bounds are
known.)
Lemma 15 If X is distributed according to Bin(n, 1/2), then
√
√
Pr(X < n/2 − n/4) = Pr(X > n/2 + n/4) ≥ 1/4.
√
√
Proof: Using √
the fact that, for any i, Pr(X = i) ≤ 1/ n, we get Pr(|X − n/2| < n/4) ≤ 1/2, so
Pr(X < n/2 − n/4) ≥ 1/4.
Now we are ready for the lower bound on JD (W).
1
Lemma 16 If K > 18 and the weights in W are non-negative, then JD (W) ≥ 36K
.
q
1
1
Proof: Assume to the contrary that JD (W) < 36K
. First, note that ψW (0) ≤ 18K
, or else the contribution
1
. Applying Lemma 15, we have
to JD (W) due to the (0, 0), 0 example is at least 36K
r
r
√
√
2
2
ψW (K/2 − K/4) > 1 −
and ψW (K/2 + K/4) < 1 +
9K
9K
(5)
1
for the (1, . . . , 1), 1 example.
as otherwise the contribution of one of the tails to JD (W) will be at least 36K
√
We will contradict this small variation of ψW (ℓ) around K/2. The bounds on ψW (0) and ψ(K/2 − K/4)
and Lemma 14 imply that ψW (ℓ) grows rapidly when ℓ is around K/2, in particular:
q
q
2
1
1 − 9K
− 18K
√
√
1
√
ψ(K/2 − K/4 + 1) − ψ(K/2 − K/4) ≥
> p
,
K/2 − K/4
9/32K
since K > 18. Now using Lemma 14 repeatedly shows that
√
√
ψ(K/2 + K/4) − ψ(K/2 − K/4) >
which contradicts (5), completing the proof.
√
1
K
×p
=2
2
9/32K
r
2
,
9K
Putting together Lemmas 10 and 16 immediately proves Theorem 8, since for K > 18 and large enough
n, the criterion for Wneg must be less than the criterion for any network with all non-negative weights.
4.2
The case when K = 2
Theorem 8 uses the assumption that K > 18 and n is large enough; is the lower bound on K really necessary?
Here we show that it is not, by treating the case that K = 2.
Theorem 17 For the standard architecture, if K = 2, for any fixed d and large enough n, every optimizer
of the dropout criterion for P(1,1),1 uses negative weights.
Proof: Define Wneg as in the proof of Lemma 10, except that the output layer has a bias of 1/5.
We claim that
lim JD (Wneg ) = 1/10.
n→∞
To see this, consider the joint input/label distribution under dropout:
Pr((0, 0), 0) = 1/2
Pr((0, 0), 1) = 1/8
Pr((2, 0), 1) = 1/8
Pr((0, 2), 1) = 1/8
Pr((2, 2), 1) = 1/8.
10
(6)
Due to the bias of 1/5 on the output, Wneg (0, 0) = 1/5. Thus, contribution to JD from examples with
x = (0, 0) in this joint distribution is 1/2 × (1/5)2 + 1/8 × (4/5)2 = 1/10.
Now, choose x 6= (0, 0). If, after dropout, the input is x, each node in the hidden layer closest to the
input computes 1. Arguing exactly as in the proof of Lemma 10, in such cases,
E((ŷ − 1)2 ) = 1 −
1+
2
n
1
1+
1 d−2
n
.
This proves (6).
Now, let W be an arbitrary network with non-negative weights.
For our distribution,
ER (D(W, (0, 0), R) − 0)2 + ER (D(W, (1, 1), R) − 1)2
JD (W) =
.
2
Let V00 = E(W(0, 0)), V22 = E(W(2, 2)), V20 = E(W(2, 0)), V02 = E(W(0, 2)) where the expectations are
taken with respect to the dropout patterns at the hidden nodes (with no dropout at the inputs). Since each
dropout pattern over the hidden nodes defines a particular network, and Lemma 12 holds for all of them,
the relationships also hold for the expectations, so
V22 ≥ V20 + V02 − V0 .
Using this V notation, handling the dropout at the input explicitly, and the bias-variance decomposition
keeping just the bias terms we get:
(V00 − 0)2 + (V00 − 1)2 + (V22 − 1)2 + (V20 − 1)2 + (V02 − 1)2 /4
2
8JD (W) ≥ 4(V00 − 0)2 + (V00 − 1)2 + (V22 − 1)2 + (V20 − 1)2 + (V02 − 1)2 .
JD (W) ≥
(7)
(8)
We will continue lower bounding the RHS. We can re-write V22 as V20 + V02 − V00 + ǫ where ǫ ≥ 0. This
is convex and symmetric in V02 and V20 so they both take the same value at the minimizer of the RHS, so
we proceed using V20 for this common minimizing value.
2
8JD (W) ≥ (2V20 − V00 + ǫ − 1)2 + 2(V20 − 1)2 + (V00 − 1)2 + 4V00
.
Differentiating with respect to V20 , we see that the RHS is minimized when V20 = (2 + V00 − ǫ)/3, giving
8JD (W) ≥
(V00 − ǫ − 1)2
2
+ (V00 − 1)2 + 4V00
.
3
If V00 ≥ 1, then JD ≥ 1/2 (just from the (0,0),0 example) and when V00 < 1 the RHS of the above is
minimized for non-negative ǫ when ǫ = 0. Using this substitution, the minimizing value of V00 is 1/4 giving
8JD (W) ≥ 1
JD (W) ≥ 1/8.
Combining this with (6) completes the proof.
11
4.3
More general distributions and implications
In the previous sub-section we analyzed the distribution P over the 2-feature examples (1, 1), 1 and (0, 0), 0.
However, these two examples can be embedded in a larger feature space by using any fixed vector of additional
feature values, creating, for instance, a distribution over (0, 1, 0, 0, 0, 1/2, 1), 0 and (0, 1, 1, 1, 0, 1/2, 1), 1 (with
the additional features underlined) The results of Section 4.2 still apply to distribution over these extended
examples after defining Wneg network to have zero weight on the additional features, and noticing that any
weight on the additional features in the positive-weight network W can be simulated using the biases at the
hidden nodes.
It is particularly interesting when the additional features all take the value 0, we call these zero-embeddings.
Every network W with non-negative weights has JD (W) ≥ 1/8 on each of these zero-embeddings of P . On
the other hand, a single Wneg network with n/K copies of the K-input first-one gadget has JD (Wneg ) ≈ 1/10
simultaneously for all of these zero-embeddings of P (when n >> K).
Any source distribution over {0, 1}K × {0, 1} that puts probability 1/2 on the the 0, 0 example and
distributes the other 1/2 probability over examples where exactly two inputs are one is a mixture of zeroembeddings of P , and thus JD (W) ≥ 1/8 while JD (Wneg ) ≈ 1/10 for this mixture and optimizing the
dropout criterion requires negative weights.
In our analysis the negative weights used by dropout are counterintuitive for fitting monotone behavior,
but are needed to control the variance due to dropout. This suggests that dropout may be less effective
when layers with sparse activation patterns are fed into wider layers, as dropout training can hijack part of
the expressiveness of the wide layer to control the artificial variance due to dropout rather than fitting the
underlying patterns in the data.
5
5.1
Properties of the dropout penalty
Growth of the dropout penalty as a function of d
Weight-decay penalizes large weights, while Theorem 5 shows that compensating rescaling of the weights
does not affect the dropout penalty or criterion. On the other hand, dropout can be more sensitive to the
calculation of large outputs than weight decay, and large outputs can be produced in deep networks using
only small weights. We make this observation concrete by exhibiting a family of networks where the depth
and desired output are linked while the size of individual weights remains constant. For this family, the
dropout penalty grows exponentially in the depth d (as opposed to linearly for weight-decay), suggesting
that dropout training is less willing to fit the data in this kind of situation.
Theorem 18 If x = (1, 1, ..., 1) and 0 ≤ y ≤ Knd−1 , for Px,y there are weights W for the standard
y2
,
architecture with R(W) = 0 such that (a) every weight has magnitude at most one, but (b) JD (W) ≥ K+1
whereas (c) J2 (W) ≤
λy 2/d
2 (Kn
+ n2 (d − 2) + n).
1/d
Proof: Let W be the network whose weights are all c = K 1/dyn(d−1)/d and biases are all 0, so that the L2
penalty is the number of weights times λc2 /2. It is a simple induction to show that, for these weights and
input (1, 1, ..., 1), the value computed at each hidden node on level j is cj Knj−1 , so the the network outputs
cd Knd−1 , and has zero square loss (since W(x) = cd Knd−1 = y).
Consider now dropout on this network. This is equivalent to changing all of the weights from c to 2c and,
independently with probability 1/2, replacing the value of each node with 0. For a fixed dropout pattern,
each node on a given layer has the same weights, and receives the same (kept) inputs. Thus, the value
computed at every node on the same layer is the same. For each j, let Hj be the value computed by the
units in the jth hidden layer.
If k0 is the number of input nodes kept under dropout, and, for each j ∈ {1, ..., d − 1}, kj is the number of
Qℓ−1
hidden nodes kept in layer j, a straightforward induction shows that, for all ℓ, we have Hℓ = (2c)ℓ j=0 kj ,
Qd−1
so that the output ŷ of the network is (2c)d j=0 kj .
12
Using a bias-variance decomposition, E((ŷ − y)2 ) = (E[ŷ] − y)2 + Var(ŷ). Since each kj is binomially
distributed, and k0 , ..., kd−1 are independent, we have E(ŷ) = (2c)d (K/2)(n/2)d−1 = cd Knd−1 = y, so
E((ŷ − y)2 ) = Var(ŷ). Since E(ŷ 2 ) = (2c)2d (K(K + 1)/4)(n(n + 1)/4)d−1 = y 2 (1 + 1/K)(1 + 1/n)d−1 , we
have Var(ŷ) = E(ŷ 2 ) − E(ŷ)2 = y 2 ((1 + 1/K)(1 + 1/n)d−1 − 1) ≥ y 2 /K, completing the proof.
If y = exp(Θ(d)), the dropout penalty grows exponentially in d, whereas the L2 penalty grows polynomially.
5.2
A necessary condition for negative dropout penalty
Section 2 contains an example where the dropout penalty is negative. The following theorem provides a
necessary condition.
Theorem 19 The dropout penalty can be negative. For all example distributions, a necessary condition for
this in rectified linear networks is that either a weight, input, or bias is negative.
Proof: [2] show that for networks of linear units (as opposed to the non-linear rectified linear units we
focus on) the network’s output without dropout equals the expected output over dropout patterns, so in
our notation: W(x) equals ER (D(W, x, R)). Assume for the moment that the network consists of linear
units and the example distribution is concentrated on the single example (x, y). Using the bias-variance
decomposition for square loss and this property of linear networks,
JD (W) = ER (D(W, x, R)−y)2 = (ER (D(W, x, R)−y)2 + VarR (D(W, x, R)) ≥ (W(x) − y)2
and the dropout penalty is again non-negative. Since the same calculations go through when averaging
over multiple examples, we see that the dropout penalty is always non-negative for networks of linear nodes.
When all the weights, biases and inputs in a network of rectified linear units are positive, then the rectified
linear units behave as linear units, so the dropout penalty will again be non-negative.
5.3
Multi-layer dropout penalty does depend on labels
In contrast with its behavior on a variety of linear models including logistic regression [28], the dropout
penalty can depend on the value of the response variable in deep networks with ReLUs and the quadratic
loss. Thus in a fundamental and important respect, dropout differs from traditional regularizers like weightdecay or an L1 penalty.
Theorem 20 There are joint distributions P and Q, and weights W such that, for all dropout probabilities
q ∈ (0, 1), (a) the marginals of P and Q on the input variables are equal, but (b) the dropout penalties of W
with respect to P and Q are different.
We will prove Theorem 20 by describing a general, somewhat technical, condition that implies that P
and Q are witnesses to Theorem 20.
For each input x and dropout pattern R, let H(W, x, R) be the values presented to the output node with
dropout. As before, let w ∈ Rn be those weights of W on connections directly into the output node and let
b be the bias at the output node. Let r ∈ {0, 1}n be the indicator variables for whether the various nodes
connecting to the output node are kept.
Proof (of Theorem 20): Suppose that P is concentrated on a single (x, y) pair. We will then get Q by
modifying y.
Let h be the values coming into the output node in the non-dropped out network. Therefore the output
of the non-dropout network is w · h + b while the output of the network with dropout is w · H(W, x, R) + b.
We now examine the dropout penalty, which is the expected dropout loss minus the non-dropout loss. We
13
will use δ as a shorthand for w · (H(W, x, R) − h).
dropout penalty = E (w · H(W, x, R) + b − y)2 − (w · h + b − y)2
= E (w · H(W, x, R) + b − w · h + w · h − y)2 − (w · h + b − y)2
= E δ 2 + 2(w · h + b − y)E(δ)
which depends on the label y unless E(δ) = 0.
Typically E(δ) 6= 0. To prove the theorem, consider the case where x = (1, −2) and there are two hidden
nodes with weights 1 on their connections to the output node. The value at the hidden nodes without
dropout is 0, but with dropout the hidden nodes never negative and computes positive values when only the
negative input is dropped, so the expectation of δ is positive.
6
Experiments
We did simulation experiments using Torch [25]. The code is accessible at
https://www.dropbox.com/sh/6s2lcfrq17zshmp/AAAQ06uDa4gOAuAnw2MAghEMa?dl=0
6.1
Negative Weights
We trained networks with the standard architecture with K = 5 inputs, depth 3, and width 50 on the training
data studied in Section 4: one input with x = (0, 0, ..., 0) and y = 0, and one input with x = (1, 1, ..., 1) and
y = 1. We used the nn.StochasticGradient function from Torch with a maximum of 10000 iterations, and
0.1
a learning rate of 1+0.1×t
on iteration t.
With the above parameters, we repeated the following experiment 1000 times.
• Initialize the parameters of a network WD .
• Clone this network, to produce a copy W with the same initialization.
• Train WD with dropout probability 1/2 at all nodes, and train W without dropout (without using
weight decay for either one).
• Compare the number of negative weights neg(WD ) in WD with neg(W).
Counts of the outcomes are shown in the following table.
neg(WD ) > neg(W)
584
neg(WD ) < neg(W)
338
neg(WD ) = neg(W)
78
Recall that Theorem 8 shows that a global optimizer of the dropout criterion definitely uses negative
weights when trained on data sources similar to P(1,1) , despite the fact that that they are consistent with a
monotone function. This experiment shows that even when optimization is done in a standard way, using
dropout tends to create models with more negative weights for P(1,1) , and we attribute this to the variance
reduction effects associated with Theorem 8.
The experiment also indicates that training with dropout sometimes produces fewer negative weights. Due
to the random initialization, a hidden ReLU unit with negative weights can evaluate to 0 on both inputs,
so its weights will never be updated without dropout. On the other hand, the extra variance from dropout
could cause updates to these negative weights. Another way the non-dropout training could produce more
negative weights is if a hidden node whose output is too large has many small weights that turn negative
with a standard gradient descent step. However, with dropout only about half the small weights will be
updated. Theorem 8 focusses on the effect of dropout at the global minimum, and abstracts away these
kinds of initialization and optimization effects.
14
1.1
Dropout
Weight Decay
1
Training error
0.9
0.8
0.7
0.6
0.5
0.4
0.6
0.8
1
1.2
1.4
1.6
Input Scale
Figure 2: Training error as a function of the scale of the inputs for Dropout and Weight Decay in the
experiment of Section 6.2.
6.2
Scale (in)sensitivity
Our experiments regarding the sensitivity with respect to the scale of the input used the standard architecture
with K = 5, d = 2, n = 5. We used stochastic gradient using the optim package for Torch, with learning
0.01
and momentum of 0.5, and a maximum of 100000 iterations.
rate 1+0.00001t
We performed 10 sets of training runs. In each run:
• Ten training examples were generated uniformly at random from [−1, 1]K .
Q
• Target outputs were assigned using y = i sign(xi ).
• Five training sets S1 , ..., S5 with ten examples each were obtained by rescaling the inputs by {0.5, 0.75, 1, 1.25, 1.5}
and leaving the outputs unchanged.
• The weights of a network Winit were initialized using the default initialization from Torch.
• For each Si :
– Winit was cloned three times to produce WD , W2 and Wnone with identical starting parameters.
– WD was trained with dropout probability 1/2 and no weight decay.
– W2 was trained with weight decay with λ = 1/2 and no dropout.
– Wnone was trained without any regularization.
The average training losses of WD and W2 , over the 10 runs, are shown in Figure 2. (The average training
loss of Wnone was less than 0.05 at all scales.)
The theoretical insensitivity of dropout to the scale of the inputs described in Theorem 4 is also seen
here, along with the contrast with weight decay analyzed in Theorem 7.
The scale of the inputs also affects the dynamics of stochastic gradient descent. With very small inputs,
convergence is very slow, and with very large inputs, SGD is unstable. The effects of the scale of the inputs
on inductive bias analyzed in this paper are visible at the scales where optimization can be done effectively.
7
Conclusions
The reasons behind dropout’s surprisingly good performance in training deep networks across a variety of
applications are somewhat mysterious and there is relatively little existing formal analysis. A variety of
15
explanations have been offered (e.g. [1, 2, 3, 11, 27]), including the possibility that dropout reduces the
amount of coadaptation in a network’s weights [16].
The dropout criterion is an expected loss over dropout patterns, and the variance in the output values
over dropout patterns contributes to this expected loss. Therefore dropout may co-adapt weights in order
to reduce this (artificial) variance. We prove that this happens even in very simple situations where nothing
in the training data justifies negative weights (Theorem 8). This indicates that the relationship between
dropout and co-adaption is not a simple one.
The effects of dropout in deep neural networks are rather complicated, and approximations can be
misleading since the dropout penalty is very non-convex even in 1-layer networks [14]. In Section 3 we show
that dropout does enjoy several scale-invariance properties that are not shared by weight-decay. A perhaps
surprising consequence of these invariances is that there are never isolated local minima when learning a
deep network with dropout. Further exploration of these scale invariance properties is warranted to see
if they are a contributor to dropout’s empirical success or can be exploited to facilitate training. While
contrasting dropout to weight-decay in simple situations, we found that a degenerate all-zero network results
(Theorem 7) when the L2 regularization parameter is above a threshold. This is in dramatic contrast to our
previous intuition from the 1-layer case.
In [28], dropout was viewed as a regularization method, adding a data dependent penalty to the empirical
loss of (presumably) undesirable solutions. Section 5 shows that, unlike the generalized linear models case
analyzed there, the dropout penalty in deeper networks can be negative and depends on the labels in the
training data, and thus behaves unlike most regularizers. On the other hand, the dropout penalty can grow
exponentially in the depth of the network, and thus may better reflect the complexity of the underlying
model space than L2 regularization.
This paper uncovers a number of dropout’s interesting fundamental properties using formal analysis of
simple cases. However, the effects of using dropout training in deep networks are subtle and complex, and
we hope that this paper lays a foundation to promote further formal analysis of dropout’s properties and
behavior.
Acknowledgments
We are very grateful to Peter Bartlett, Seshadhri Comandur, and anonymous reviewers for valuable communications.
References
[1] P. Bachman, O. Alsharif, and D. Precup. Learning with pseudo-ensembles. NIPS, 2014.
[2] P. Baldi and P. Sadowski. The dropout learning algorithm. Artificial intelligence, 210:78–122, 2014.
[3] Pierre Baldi and Peter J Sadowski. Understanding dropout. In Advances in Neural Information Processing
Systems, pages 2814–2822, 2013.
[4] P. L. Bartlett, M. I. Jordan, and J. D. McAuliffe. Convexity, classification, and risk bounds. Journal of the
American Statistical Association, 101(473):138–156, 2006.
[5] L. Breiman. Some infinity theory for predictor ensembles. Annals of Statistics, 32(1):1–11, 2004.
[6] Caffe, 2016. http://caffe.berkeleyvision.edu.
[7] Danqi Chen and Christopher D Manning. A fast and accurate dependency parser using neural networks. In
EMNLP, pages 740–750, 2014.
[8] G. E. Dahl. Deep learning how I did it: Merck 1st place interview, 2012. http://blog.kaggle.com.
[9] G. E. Dahl, T. N. Sainath, and G. E. Hinton. Improving deep neural networks for LVCSR using rectified linear
units and dropout. ICASSP, 2013.
[10] L. Deng, J. Li, J. Huang, K. Yao, D. Yu, F. Seide, M. L. Seltzer, G. Zweig, X. He, J. Williams, Y. Gong, and
A. Acero. Recent advances in deep learning for speech research at microsoft. ICASSP, 2013.
16
[11] Yarin Gal and Zoubin Ghahramani. Dropout as a Bayesian approximation: Representing model uncertainty in
deep learning. arXiv:1506.02142, 2015.
[12] Ian J. Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron C. Courville, and Yoshua Bengio. Maxout networks.
In ICML, pages 1319–1327, 2013.
[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level
performance on imagenet classification. In ICCV, pages 1026–1034, 2015.
[14] D. P. Helmbold and P. M. Long. On the inductive bias of dropout. JMLR, 16:3403–3454, 2015.
[15] G. E. Hinton. Dropout: a simple and effective way to improve neural networks, 2012. videolectures.net.
[16] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. R. Salakhutdinov. Improving neural networks
by preventing co-adaptation of feature detectors, 2012. Arxiv, arXiv:1207.0580v1.
[17] Nal Kalchbrenner, Edward Grefenstette, and Phil Blunsom. A convolutional neural network for modelling
sentences. In ACL, pages 655–665, 2014.
[18] P. M. Long and R. A. Servedio. Random classification noise defeats all convex potential boosters. Machine
Learning, 78(3):287–304, 2010.
[19] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In ICML,
pages 807–814, 2010.
[20] Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. Norm-based capacity control in neural networks. In
COLT, pages 1376–1401, 2015.
[21] Jürgen Schmidhuber. Deep learning in neural networks: An overview. Neural Networks, 61:85–117, 2015.
[22] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: A simple way to prevent
neural networks from overfitting. JMLR, 15:1929–1958, 2014.
[23] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan,
Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. CVPR, 2015.
[24] Tensorflow, 2016. https://www.tensorflow.org.
[25] Torch, 2016. http://torch.ch.
[26] T. Van Erven, W. Kotłowski, and M. K. Warmuth. Follow the leader with dropout perturbations. COLT, pages
949–974, 2014.
[27] S. Wager, W. Fithian, S. Wang, and P. S. Liang. Altitude training: Strong bounds for single-layer dropout.
NIPS, 2014.
[28] S. Wager, S. Wang, and P. Liang. Dropout training as adaptive regularization. NIPS, 2013.
[29] L. Wan, M. Zeiler, S. Zhang, Y. Le Cun, and R. Fergus. Regularization of neural networks using dropconnect.
In ICML, pages 1058–1066, 2013.
[30] T. Zhang. Statistical behavior and consistency of classification methods based on convex risk minimization.
Annals of Statistics, 32(1):56–85, 2004.
17
A
Table of Notation
Notation
1set
(x, y)
σ(·)
W
w, v
W(x)
P
Px,y
RP (W)
q, p
R
r, s
D(W, x, R)
JD (W)
J2 (W)
λ
WD
WL2
n, d
K
B
Meaning
indicator function for “set”
an example with feature vector x and label y
the rectified linear unit computing max(0, ·)
an arbitrary weight setting for the network
specific weights, often subscripted
the output value produced by weight setting W on input x
an arrelbitrary source distribution over (x, y) pairs
the source distribution concentrated on the single example (x, y)
the risk (expected square loss) of W under source P
probabilities that a node is dropped out (q) or kept (p) by the dropout process
a dropout pattern, indicates the kept nodes
dropout patterns on subsets of the nodes
Output of dropout with network weights W, input x, and dropout pattern R
the dropout criterion
the L2 criterion
the L2 regularization strength parameter
an optimizer of the dropout criterion
an optimizer of the L2 criterion
the network width and depth
the number of input nodes
Proof of Theorem 7
Here we prove Theorem 7, showing that the weight-decay aversion depends on the values of the inputs and
the number of input nodes K. Furthermore, unlike the single-layer case, the L2 regularization strength has
a threshold where the minimizer of the L2 criterion degenerates to the all-zero network.
We will focus on the standard architecture with depth d = 2. Recall that we are analyzing the distribution
P(x,1) that assigns probability 1/2 to (x, 1) and probability 1/2 to (0, 0).
Also, recall that, for P(x,1) , the weight-decay aversion is the maximum risk incurred by a minimizer of
J2 . This motivates the following definition.
Definition 21 An aversion witness is a minimizer of J2 that is also a maximizer of the risk R, from among
minimizers of J2 .
The proof of Theorem 7 involves a series of lemmas.
We first show that there is a aversion witness WL2 with a special form, and then relate any hidden node’s
effect on the output to the regularization penalty on the weights in and out of that node. This will allow us
to treat optimizing the L2 criterion as a one-dimensional problem, whose solution yields the theorem.
For some minimizer WL2 of J2 , let v∗j denote the vector of weights into hidden node j and wj∗ denote the
weight from j to the output node. Let a∗j be the bias for hidden node j and let b∗ be the bias for the output
node. Let hj be the function computed by hidden node j.
Note that, if there is an aversion witness with a certain property, then we may assume without loss of
generality that WL2 has that property.
Lemma 22 We may assume without loss of generality that for each hidden node j, there is an input x̃ ∈
{0, x} such that hj (x̃) = 0.
Proof: Suppose neither of hj (0) or hj (x) was 0. If wj∗ = 0, then replacing both with 0 does not affect
the output of WL2 , and does not increase the penalty. If wj∗ 6= 0, then subtracting min{hj (0), hj (x)}
18
from a∗j and adding
min{hj (0),hj (x)}
wj∗
min{hj (0), hj (x)} = 0.
to b does not affect WL2 or the penalty, but, after this transformation,
Lemma 23 We may assume without loss of generality that for each hidden node j, we have |v∗j · x| =
max{hj (0), hj (x)}.
Proof: If hj (x) ≥ hj (0) = 0, then a∗j = 0, and hj (x) = v∗j · x.
If hj (0) > hj (x) = 0. Then, a∗j > 0, and v∗j · x ≤ −a∗j . If needed, the magnitude of v∗j can be decreased
to make v∗j · x = −a∗j . This decrease does not affect WL2 (x) or WL2 (0), and can only reduce the L2 penalty.
Lemma 24 We may assume without loss of generality that for each hidden node j, we have hj (0) = 0.
Proof: Suppose hj (0) > 0. Let z be this old value of hj (0). Then hj (x) = 0 and z = hj (0) = −v∗j · x. If
we negate v∗ and set aj = 0, then Lemma 23 implies that we swap the values of hj (x) and hj (0).
Then, by adding zwj∗ to b∗ and negating wj∗ , we correct for this swap at the output node and do not
affect the function computed by WL2 or the penalty.
Note that Lemma 24 implies that a∗1 = ... = a∗K = 0.
Lemma 25 For all j, v∗j · x ≥ 0.
Proof: Since a∗j = 0, if v∗j · x < 0, we could make WL2 compute the same function with a smaller penalty
by replacing v∗j with 0.
Lemma 25 implies that the optimal WL2 computes the linear function,
WL2 (x̃) = (w∗ )T V ∗ x̃ + b∗ .
(9)
Later we will call (w∗ )T V ∗ x̃ the activation at the output node.
Lemma 26 b∗ =
1−(w∗ )T V ∗ x
.
2
Proof: Minimize J2 (wrt distribution P(x,1) ) as a function of b using Calculus.
Now, we have
WL2 (x̃) =
1
+ (w∗ )T V ∗ (x̃ − x/2)
2
(10)
which immediately implies
WL2 (0) = 1 − WL2 (x).
(11)
Lemma 27 Each v∗j is a rescaling of x.
Proof: Projecting v∗j onto the span of x does not affect hj , and cannot increase the penalty.
Lemma 28 WL2 (x) ≤ 1.
Proof: By (11), if WL2 (x) > 1 then WL2 (0) < 0 and the loss and the penalty would both be reduced by
scaling down w∗ .
Lemma 29 WL2 maximizes W(x) over those weight vectors W that have the same penalty as WL2 and
compute a function of the form W(x̃) = 12 + (w)T V (x̃ − x/2) (i.e. Equation 10).
19
Proof: Let W maximize W(x) over the networks considered. If WL2 (x) < W(x) ≤ 1, then W would have
the same penalty as WL2 but smaller error, contradicting the optimality of WL2 .
f obtained by scaling down the weights in the output layer so that
If W(x) > 1, then the network W
f
W(x) = 1 has a smaller penalty than WL2 and smaller error, again contradicting WL2 ’s optimality.
Informally, Lemmas 28 and 29 engender a view of the learner straining against the yolk of the L2 penalty
to produce a large enough output on x. This motivates us to ask how large W(x) can be, for a given value
of ||W||22 (recall that Lemma 25 allows us to assume that the biases at the hidden nodes are all 0).
Definition 30 For each hidden node j, let αj be the constant such v∗j = αj x, so that hj (x) = αj x · x.
Recall that the activation at the output node on input x is the weighted sum of the hidden-node outputs,
(w∗ )T V ∗ x.
Definition 31 The contribution to the activation at the output due to hidden node j is wj∗ hj (x) = wj∗ αj x · x
and the contribution to the L2 penalty from these weights is λ2 (wj∗ )2 + α2j x · x .
We now bound the contribution to the activation in terms of the contribution to the L2 penalty. Note
that as the L2 “budget” increases, so does the the maximum possible contribution to the output node’s
activation.
Lemma 32 If B is hidden node j’s weight-decay contribution, (wj∗ )2 + α2j x · x, then hidden node k’s contriq
q
B
, where it achieves the
bution to the output node’s activation is maximized when wj∗ = B2 and αj = 2x·x
√
value B x · x/2
q
Proof: Since α2j x · x + (wj∗ )2 = B, we have wj∗ = B − α2j x · x, so the contribution to the activation
q
can be re-written as αj x · x B − α2j x · x. Taking the derivative with respect to αj , and solving, we get
q
q
B
B
αj = ± 2x·x
and we want the positive solution (otherwise the node outputs 0). When αj = 2x·x
we have
q
wj∗ = B2 and thus the node’s maximum contribution to the activation is
r
B
2
r
√
B x·x
B
(x · x) =
.
2x · x
2
Lemma 33 The minimum sum-squared weights for a network W (without biases at the hidden nodes) that
has an activation A at the output node on input x is √2A
.
x·x
√
Proof: When maximized, the contribution of each hidden node to the activation at the output is x · x/2
times the hidden node’s contribution to the sum of squared-weights. Since each weight in W is used in
exactly one hidden node’s contribution to the output node’s activation, this completes the proof.
Note that this bound is independent of n, the number of hidden units, but does depend on the input x.
Proof (of Theorem 7): Let A ≥ 0 be the activation at the output node for WL2 on input x. From
Lemma 26 we get that b∗ = 1−A
2 . Combining Lemmas 28 , 29 and 33, we can re-write the J2 criterion for
WL2 and distribution P(x,1) in terms of A as follows.
λ
1
1 ∗
2
2
(b − 0) + (WL2 (x) − 1) + ||W||22
2
2
2
!
2
2
1−A
1
1+A
2A
+
=
− 1 + λ√
.
2
2
2
x·x
J2 (WL2 ) =
20
(12)
Differentiating with respect to A, we see that the criterion is minimized when
A=1− √
2λ
x·x
2λ
≤1
x·x
2λ
>1
when √
x·x
when √
A=0
since we assumed A ≥ 0; when A = 0 then WL2 has all zero weights with a bias of 1/2 at the output.
The risk part of (12) simplifies to
λ2
(1 − A)2
=
,
4
x·x
1 λ2
so the overall the risk of an aversion witness is min
.
,
4 x·x
21
| 10math.ST
|
On Encoding Temporal Evolution for Real-time Action Prediction
1
Fahimeh Rezazadegan1,2, Sareh Shirazi1,2, Mahsa Baktashmotlagh1, Larry S. Davis3
Queensland University of Technology, 2Australian Centre for Robotic Vision, 3University of Maryland
[email protected]
Abstract
Training Video Samples
Rank Pooling Training Dynamic images
V130
Anticipating future actions is a key component of
intelligence, specifically when it applies to realtime systems, such as robots or autonomous cars.
While recent works have addressed prediction of
raw RGB pixel values, we focus on anticipating the
motion evolution in future video frames. To this
end, we construct dynamic images (DIs) by summarising moving pixels through a sequence of future frames. We train a convolutional LSTMs to
predict the next DIs based on an unsupervised
learning process, and then recognise the activity associated with the predicted DI. We demonstrate the
effectiveness of our approach on 3 benchmark action datasets showing that despite running on videos with complex activities, our approach is able to
anticipate the next human action with high accuracy and obtain better results than the state-of-the-art
methods.
1
Introduction
While recent computer vision systems can recognize objects, scenes and actions with reasonable accuracy
(Russakovsky et al. 2014; Zhou et al. 2014; Wang et al.
2017), anticipating future activities still remains as an open
problem which requires modeling human activities, their
temporal evolution, and recognizing the relationships between objects and human in the scene.
Predicting human action has a variety of applications
from human-robot collaboration and autonomous robot navigation to exploring abnormal situations in surveillance videos and activity-aware service algorithms for personal or
health care purposes. As an example, in autonomous
healthcare services, consider an agent- monitoring a patient’s activities, trying to predict if the patient is losing
her/his balance. If the agent is capable of predicting the next
action, it could determine whether s/he might fall and take
an action to attempt to prevent it. Another example of activity prediction in the robotics area is intelligent human-robot
collaboration. For robots and humans to be cooperative
partners who share tasks naturally and intuitively, it is essential that the robot understands actions of the human and
anticipates the human’s needs e.g. the need for tools and
V11
1
V
15
V215
V230
V21
Vk1
Vk15
Vk30
Prediction
Model
Placing side
panel is the
next Action
Test Dynamic images
Figure 1: The overview of our human action prediction system.
parts in assembly tasks.
Prior research focused on future frame prediction with the
goal of predicting as many future frames as possible (Lotter,
Kreiman and Cox 2016), or predicting trajectories of people,
either holistic or of their parts, and then applying activity
recognition to those predicted trajectories (Bennewitz et al.
2005; Kitani et al. 2012).
To overcome the limitation of the current approaches on
predicting single frame or trajectory of people resulting in
inaccurate prediction, our framework introduces an LSTM
model to predict semantics of future human activity i.e. human pose and objects’ shapes by constructing dynamic images (DIs1) in its input layer (Bilen et al. 2016).
Inspired by the previous work in next-frame video prediction (Lotter, Kreiman and Cox 2016), our LSTM model
continually predicts future images, and leverages deep recurrent convolutional networks with both bottom-up and
top-down connections. Our framework capitalises on the
temporal structure of unlabeled videos to learn to anticipate
both actions and objects in the future. Our proposed framework is experimentally validated on several benchmark da-
1 A DI is a representation incorporating cues of human poses
and object shapes in a sequence of frames, originally introduced as
a representation for video activity recognition.
A Frame Sequence
…
…
…
Figure 2: Generating Dynamic images from video frames of 50
Salads dataset, Cooking activities dataset and IKEA dataset, respectively. Here, we show dynamic images generated by using a
window size of 30 and a stride of 5.
tasets of MPII Cooking dataset (Rohrbach et al. 2012), 50
Salads dataset (Stein and McKenna 2013) and IKEA assembly dataset (Eich, Shirazi and Wyeth 2016), which contain
complex activities composed of several actions. The main
contributions of this work are as follows:
• For the first time in the action prediction literature, we
predict a sequence of (up to) 150 frames in a form of
dynamic image (DI), performing 5 times longer prediction than the existing works.
• We summarise short video clip in a single image by
constructing DIs which gives the potential of going further in the future. We show that predicting a representation like DIs more directly encodes activity than either
future frames or future optical flow fields, while gives
us critical cues about the performed action resulting in
more accurate and long-term activity prediction compared to existing approaches.
• The DI-based Semantic representation of an action in
our approach helps in accurately predicting the action
before its occurrence. This is different from the existing
works on early action recognition which rely on the information of observing the first part of the action for
prediction.
2 Related Work
Action prediction can refer to either early recognition of
human activity from a partially observed video containing a
single action or, more generally, online prediction of the
next activity in a multi task activity video.
There have been a few previous approaches for predicting
future actions based on predicting human trajectories in 2D
from visual data (Bennewitz et al. 2005; Kitani et al. 2012).
For example, Jiang et al. predicted the action movements a
person may take in a human-robot interaction scenario using
RGB-D sensors and an anticipatory temporal CRF model
(Jiang and Saxena 2014). Such works not only enhance the
quality of interactions between the human and robot, but
also are beneficial to reduce risk assessment expenses
(Rezazadegan, Gengb, et al. 2015). However, their accuracy
significantly drops for a long anticipation horizon.
A related problem to action prediction is early detection
of an action -detect/classify an incoming temporal sequence
as early as possible (Ryoo 2011; Lan, Chen and Savarese
2014; Soomro, Idrees and Shah 2016; Sadegh Aliakbarian et
al. 2017).
Another group of researchers approach action prediction
by leveraging the association between humans and scene
elements (Vu et al. 2014), which was shown to be beneficial
for action recognition as well (Rezazadegan, Shirazi, et al.
2015; Rezazadegan et al. 2017).
To the best of our knowledge, Li et. Al. is the only work
that addressed long-duration activity prediction using sequential pattern mining to incorporate context into prediction (Li and Fu 2016). The temporal predictions called action-lets are based on motion velocity peaks that are difficult
to obtain in real-world scenarios.
The first algorithm for video frame prediction inspired by
language modeling was presented in Ranzato et al. (Ranzato
et al. 2014). Later on, other algorithms were introduced using LSTM, ConvNets, generative adversarial training and
predictive coding architectures (Srivastava, Mansimov and
Salakhudinov 2015; Mathieu, Couprie and LeCun 2015;
Lotter, Kreiman and Cox 2016; Finn, Goodfellow and
Levine 2016; Vondrick, Pirsiavash and Torralba 2015).
A convolutional LSTM architecture was employed in
some recent papers on future frame prediction (Finn,
Goodfellow and Levine 2016; Lotter, Kreiman and Cox
2016; Neverova et al. 2017; Luo et al. 2017). For instance,
Finn et al. developed an action-conditioned video prediction model that incorporates appearance information in previous frames with motion predicted by the model and is able
to generate frames 10 time steps into the future.
Among all the above-mentioned approaches for future
frame prediction, Lotter et al. achieves good results for predicting 5 future time steps at 10 fps, i.e. 0.5 seconds, and
Finn et al. predict 10 future time steps, at 10 fps, i.e. 1 second. However long-duration prediction has not been addressed, although it is critical for action prediction, particularly in real-time robotics applications.
Our work targets longer-duration action prediction by
summarising a sequence of future frames in a compact representation rather than predicting individual frames. Our
extensive experiments on the benchmark action prediction
datasets demonstrate the effectiveness of our approach on
achieving reasonably accurate predictions to time horizons
of 5 seconds (150 frames). We employ the capability of
deep networks to predict DIs of future video clips.
3 Overview of the Proposed Approach
To make a human action prediction system useful for real- time applications such as robotics, we need to reduce the
algorithm latency as much as possible and predict longer in
the future. In this work, we predict a compact representation
of the sequence of future frames in the forms of DIs to reduce latency in prediction. Our DI-based representation
helps in suppressing background information and encoding
Action3: Placing Back Panel
…
…
…
…
…
…
Softmax
FC 5
Average Pooling
1×1 Conv, 2048
3×3 Conv, 512
1×1 Conv, 512
1×1 Conv, 1024
3×3 Conv, 256
1×1 Conv, 256
1×1 Conv, 512
3×3 Conv, 128
1×1 Conv, 128
1×1 Conv, 256
3×3 Conv, 64
1×1 Conv, 64
Dynamic images
3×3 Max pool, /2
Sequences of Predicted
7×7 Conv, 64, /2
Network Architecture for Action Recognition; ResNet152
×3
×8
×36
×3
Figure 3: Training a network for action recognition from predicted dynamic images as inputs.
motion of actors and objects involved in the actions. Being
compact (i.e., a few seconds of video are encoded in a single
frame) and fast to compute, action prediction from DIs is
computationally efficient. Figure 1 shows an overview of
our approach to human action prediction.
3.1 Constructing DIs from Video
A dynamic image (DI) is based on rank pooling and is
obtained through the parameters of a ranking machine that
encodes the temporal evolution of frames of video. More
precisely, DIs focus on those areas of images which contain
actors and the objects they interact with, whereas they suppress background pixels and motion patterns which can be
considered as noise for action recognition and prediction. As
a result, the energy in DIs is concentrated in the salient actors’ motion. DIs are constructed by identifying a function ψ
that maps a sequence of video frames (I1,…,IT) into a video
descriptor d*=ρ(I1,…,IT;ψ) , which contains enough information to rank all the frames in the video where ψ(It) stacks
the RGB components of each pixel in image into a large
vector. The ranking function, ρ, associates to each time t a
score S=<d,Vt> where d is a vector of parameters and V is
the time average of ψ:
Vt
1 t
(IT )
t T 1
(1)
Equation 2 reflects the process of learning d and constructing d* from a sequence of video frames, which is called rank
pooling. Note that d is learned so that the scores in later
frames are associated with larger scores.
1
d arg min ( d
2
2
2
max (0,1 S (q | d ) S (t | d )))
T (T 1) q t
(2)
3.2 Predicting the Next Action
To train a deep model for predicting the next dynamic
image of a video, we first construct DIs for 30 frame video
clips and use a stride of 5 frames as we pass the one second
window through a long video. Then we train a deep model
for predicting the next dynamic image of a video. Figure 2
shows examples of 3 datasets and the generated DIs using
this pipeline for 1 second clips. The model for predicting
dynamic images (DIs) is based on recurrent neural networks
and convolutional long short-term memories, extending the
architecture of PredNet (Lotter, Kreiman and Cox 2016).
The model consists of stacked modules which generate local
predictions of the input, and then the logarithmic error between the local predictions and the actual input is flowed to
the next layer of the network. Each module of the network
contains four layers - an input convolutional layer, Al, a
recurrent representation layer, Rl, a prediction layer, Âl, and
an error representation, El (Eq. 3).
l 0
x t
A lt
t
M AX POOL (R ELU (Conv (E l 1 ) )) l 0
t
ˆ
A R ELU (Conv (R t ) )
l
l
E lt [R ELU (A lt Aˆlt ); R ELU (Aˆlt A lt )]
(3)
R lt ConvLSTM (E lt 1 , R lt 1 ,Upsample (R lt 1 ))
The input layer, prediction layer and error representation
layer have the standard shape of deep ConvNets, whereas
recurrent representation layers follow the rules of generative
deconvolutional networks. Prediction of a DI is based on the
last 10 DIs constructed from the previously observed portion of the video. Since each of the DIs summarises a one
second video clip, the recurrent network integrates information over 75 observed frames to construct a compact prediction of the next 30.
In principle, one could predict DIs for either longer future
time intervals or multiple shorter intervals, although we do
not investigate these alternative prediction structures here.
However, we consider extrapolating 5 steps further in the
future by predicting 5 future DIs. The prediction accuracies
for 5 time steps, i.e. 5 seconds into the future, on 3 datasets
are presented in Section 5.1, Figure 5.
To predict the next human action, a feature extraction
system and a classifier are required to recognize action classes from the predicted DIs. Therefore, this part of the system is based on supervised learning, in contrast to the DI
prediction pipeline (Section 3.2) which was unsupervised.
Note that typically there are gaps of inactivity between consecutive sub-actions in a video. For example, in one video
after the actor completes the "using screw driver sub-action"
there is a 3 second gap before the actor initiates the "placing
back panel" sub-action. In general, we label the gap with the
activity label of the next sub- activity as it is the desired
(a) IKEA dataset
(b) 50 Salads dataset
(c) MPII cooking activities dataset
Figure 4: Predicted DIs versus actual ones on (a) IKEA; (b) 50 Salads, (c) MPII cooking datasets; FPS=30, window size=30 and stride=5.
sub-activity we wish to predict. Figure 3 reflects the training
process on predicted DIs of IKEA samples and the corresponding action labels.
4 Experimental Setup
In our experimental evaluations, we apply our method to
predict the next DI on 3 public datasets (Rohrbach et al.
2012; Stein and McKenna 2013; Eich, Shirazi and Wyeth
2016). We compare the MSE of predicted DI individually
with a simple baseline in which we copy the previous DI.
Then, we recognize the action labels of predicted DIs
through a ConvNet framework which is the principal objective of our research and compare it to 3 other baselines and
two state-of-the-art-models for action prediction. Finally,
we conduct an experiment to determine how early we can
predict the next action compared to other methods.
4.1 Datasets
Since existing methods for action prediction mostly focused on early action recognition, they used datasets containing a single action per video. Differently, we aim to predict the next action by seeing a part of the current one.
Therefore, we consider 3 public datasets; each contains multiple actions per video.
The first dataset is MPII Cooking activities dataset which
is suitable for evaluating fine-grained action classification
(Rohrbach et al. 2012). It contains 44 videos, composed of
65 fine-grained cooking actions such as cut slices, pour,
spice and so on, performed in the same kitchen setting. This
dataset covers gender, subject and recipe diversity for making 14 different dishes.
The second dataset is 50 Salads dataset, which is composed of 50 sequences of a mixed salad preparation task
with two sequences per subject (Stein and McKenna 2013).
There are gender and age diversity in videos, and a different
task-ordering for each sequence that makes it challenging.
The third dataset is the IKEA dataset which is specifically
designed for collaborative robotics challenges, contains 50
videos of different people using different setups for the task
of assembling an IKEA drawer (Eich, Shirazi and Wyeth
2016). All actors follow the same instructions, but vary the
order of sub-actions, the duration of each sub-action and the
locations of assembly tools. We consider 5 action labels in
total for the task.
4.2 Implementation Details
For the first experiment, we initially extracted video
frames of all datasets at 30 fps. We allocated 40 videos as
the training set, 2 as the validation set and 2 as the test set
for MPII cooking dataset. For the 50 Salads dataset, training
set, validation set, and test set contain 46 videos, 2 videos
and 2 videos, respectively. For the IKEA dataset, training
set contains 41 videos, each has around 1000 frames, while
we had 6 videos for testing and 3 videos for validation.
Then, we constructed dynamic images (DIs) following
the strategy outlined in Section 3.1 using a window size of
30 and stride of 5. Note that the DIs for clips from different
classes are likely to have many common background pixels
because the actors tend to be centered in the field of view of
the camera. This biases the error signal during training and
results in insufficiently accurate prediction of DIs. We
found that adding low variance white noise to the inputs
during training, and computing the error based on the noisy
predictions, yielded a more accurate prediction network.
Therefore, sequences of 10 noisy DIs are sampled from datasets, center-cropped and downsampled to 128x160 pixels
to be fed into the 4-layer architecture with 3by3 convolutions and layer channel sizes of (3,48,96,192) for the prediction process. For unsupervised training, we used the Adam
(a)
(b)
Figure 5: Accuracy of prediction over 5 time steps (5 seconds) into the future using a) ours on 3 datasets; b) Soomro et al. on 2 datasets.
optimization algorithm using a loss computed only on the
lowest layer (L0). We initially set Adam parameters to
α=0.001, β1=0.9, β2=0.999, in which the learning rate, α,
will be decreased to 0.0001 halfway through training. We
also replaced L1 error with logarithmic error in El formula,
stated in Eq. 3 as it improved the prediction results.
In order to create a model to recognize the action class of
predicted DIs for unseen test videos, we then finetune the
ConvNet on our predicted DIs from Section 3.2, which are
resized to 224×224. Note that we use a model that we finetuned on Sport-1M dataset, for finetuning. This network
contains 152 layers including convolutional layers and skip
connections ending with a global average pooling layer and
a fully connected layer with softmax same as (He et al.
2016), originally proposed for object recognition. The parameters of the network were optimized using stochastic
gradient descent with a learning rate of 0.0001, reduced by
the factor of 0.5 after every 5k iterations, a momentum of
0.9, a weight decay of 0.0005, and mini-batches of size 16.
We present final action prediction results for 3 datasets as
the accuracy of recognized actions from predicted DIs.
5 Experimental Results
In this section, we first present results of the prediction
algorithm for 3 above-mentioned datasets for 1 time step
(Figure 4). Then, we extend the predictions up to 5 time
steps (Figure 5). Finally, we explore how early we can predict actions for IKEA test videos in average (Figure 6).
Figure 4 shows examples of predicted dynamic images
(DIs) versus actual DIs for all 3 datasets. In addition to these
qualitative results, to evaluate how accurate the predicted
DIs are compared to their ground truth, we calculated Model
MSE, and compared with the Prev. MSE in Table 1, to have
a fair comparison with the employed method in (Lotter,
Kreiman and Cox 2016). Model MSE is the mean squared
error between the predicted DI and the actual DI for the next
second of video. Prev. MSE is mean squared error between
the DI for the last second of observed video and the actual
DI of the next second. Table 1, shows about a 50% improvement in Model MSE versus Prev. MSE, for 3 datasets.
To evaluate how accurately we can predict actions from
predicted DIs, we report the recognition accuracy of the
predicted DIs for unseen test videos of 3 datasets, using our
recognition model which we trained based on ResNet152
architecture (He et al. 2016). Table 2 shows the average of
prediction accuracies for all actions of the test videos for all
datasets, and compares our method to two state-of-the-art
methods (Soomro, Idrees and Shah 2016; Sadegh
Aliakbarian et al. 2017).
Table 1: Model MSE compared to MSE between previous Dynamic images (DIs), on 3 public datasets.
Datasets
50 Salads
MPII Cooking
IKEA
Model MSE
0.003696
0.002562
0.001131
Prev. MSE
0.005319
0.003472
0.002321
For (Soomro, Idrees and Shah 2016) since their approach
estimated human pose, we were unable to repeat the experiment for 50 Salads dataset which does not contain the human body. Due to unavailability of the code for (Sadegh
Aliakbarian et al. 2017), we report the accuracy only for
IKEA dataset which the author has kindly provided us.
Table 2: Our final prediction mean accuracy of one second ahead,
on all actions of 3 datasets.
Ours
Soomro et al. Sadegh et al.
Datasets
MPII Cooking
60.36%
40.51%
N/A
50 Salads
57.48%
N/A
N/A
IKEA
84.94%
61.42%
76.67%
We reported the accuracy of predicting the action of one
future time step, i.e. one second into the future, for all baselines in Table 3. Our method using DIs (the 1st column of
Table 3) outperformed all 3 baselines. It verifies the strong
ability of our predicting model, simultaneously exploiting
the temporal and pose information hidden in the DIs as inputs. There are some incorrect labels for sub-actions that are
caused by appearance similarity of sub-actions in datasets.
Table 3: A comparison of mean accuracy of our approach on next
action prediction for all actions of test videos in 3 datasets, with 3
baselines: A) Using the previous DI as prediction; B) Labelling the
original DIs with the class of next action and training a CNN on
these differently labelled DIs for action recognition; C) Training a
LSTM to predict the future action.
Datasets
MPII Cooking
50 Salads
IKEA
Ours
60.36%
57.48%
84.94%
A
48.52%
46.82%
75.76%
B
28.14%
27.5%
41.49%
C
32.25%
31.14%
46.06%
(a)
t=-72
t=-48
t=-24
t=0
Placing Back Panel
Placing Back Panel
Placing Back Panel
Placing Back Panel
t=-32
t=-22
t=-11
t=0
Placing Back Panel
Placing Back Panel
Placing Back Panel
Placing Back Panel
t=-39
t=-26
t=-13
t=0
Placing Back Panel
Placing Back Panel
Placing Back Panel
Placing Back Panel
(b)
(c)
Figure 6: Reflecting how early the next action can be predicted on a sample video, using a) our method; b) Soomro et al.; c) Sadegh et al.
For example, the first row shows that our method predicted that “Placing Back Panel” is going to happen, 72 frames before (t=-72) than
it occurs (t=0); the prediction was not change from the first prediction point to its occurrence time.
Note that there is a gap between the time that the network
predicts the correct label for the next action and the time
that the human starts the action (for example, the time when
a required tool is picked up for the next action). This is consistent with our final goal - passing tools by a robot before
the human needs them.
5.1 Further time step prediction
We consider extrapolating 5 steps further in the future by
finetuning our prediction network by feeding back predictions as the inputs and recursively iterating. More precisely,
we train the model with the loss over 15 time steps, starting
from the trained weights. In fact, we feed the original dynamic images (DIs) into the network for the first 10 and
then we use the model’s predictions for the remaining 5
inputs. This covers predictions for the next 5 time steps (5
seconds). Figure 5 shows a bar graph representing the trend
of degrading prediction accuracy compared to the state-ofthe-art method (Soomro, Idrees and Shah 2016). Note that
since (Soomro, Idrees and Shah 2016) is based on estimating pose and human model, we excluded the 50 Salads dataset in which we do not have the human body, but only
hands. The proposed method in (Sadegh Aliakbarian et al.
2017) is not capable of predicting up to 5 seconds and excluded from this experiment. Despite eventual blurriness
particularly for the fourth and fifth future predicted DIs,
which can be expected due to uncertainty, the fine-tuned
model captures some key structures in its extrapolations
which is still useful for predicting actions.
5.2 Measuring temporal distance
In order to investigate our method’s ability in predicting
future actions at different stages and various actions, we
evaluated the performance in terms of the temporal distance
(frames) between the current image and the starting frame of
the next action, for IKEA dataset. Table 4 reports the average of these temporal distances (AvgOfTD) on 6 test videos,
for all actions placing side panels, placing back panel, placing rear panel, using screw driver and fixing bolts, using our
method compared with other methods. Note that we used
one time step (1 second ahead) prediction here. We also
showed predicted actions using our method compared with
two other methods at different temporal stages for a test
video sample (run42, action3), in Figure 6.
Table 4. A comparison between AvgOfTDs obtained from our
method and other state-of-the-art methods.
AvgOfTD
Action1
Action2
Action3
Action4
Action5
Ours
Soomro et. al
Sadegh et. al
33.6
13.2
51.5
30
13.1
21.8
63.3
28.7
23.5
42
17.6
41.8
30
11.3
36.3
6 Conclusion
We proposed an action prediction framework based on a
convolutional LSTM model to predict a future DI that is the
summary of 30 future RGB frames. We showed the ability
of our method to predict up to 5 dynamic images (DIs)
whose corresponding actions can be recognized by a ConvNet. We attained 85%, 60% and 57% accuracy for predicting a human action one second before it happens on IKEA,
MPII cooking and 50Salads datasets, respectively, which
degrades to 58%, 46% and 40% for the fifth time step (5
seconds ahead). Experiments verified the superiority of our
model on 3 datasets- for motion encoding and predictive
learning compared to 6 baselines and two state-of-the-art
methods for action prediction.
References
M. Bennewitz, W. Burgard, G. Cielniak and S. Thrun. 2005.
"Learning motion patterns of people for compliant robot
motion." The International Journal of Robotics Research
24 (1): 31-48.
H. Bilen, B. Fernando, E. Gavves, A. Vedaldi and S. Gould.
2016. "Dynamic image networks for action recognition."
In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, 3034-3042.
M. Eich, S. Shirazi and G. Wyeth. 2016. "ARTiS:
Appearance-based Action Recognition in Task Space."
arXiv preprint arXiv:1610.05432.
C. Finn, I. Goodfellow and S. Levine. 2016. "Unsupervised
learning for physical interaction through video
prediction." In Advances in Neural Information
Processing Systems, 64-72.
K. He, X. Zhang, S. Ren and J. Sun. 2016. "Deep residual
learning for image recognition." In Proceedings of the
IEEE conference on computer vision and pattern
recognition, 770-778.
Y. Jiang and A. Saxena. 2014. "Modeling HighDimensional Humans for Activity Anticipation using
Gaussian Process Latent CRFs." In Robotics: Science
and Systems, 1-8.
K. Kitani, B. Ziebart, J. Bagnell and M. Hebert. 2012.
"Activity forecasting." In European Conference on
Computer Vision, 201-214.
T. Lan, T. Chen and S. Savarese. 2014. "A hierarchical
representation for future action prediction." In European
Conference on Computer Vision, 689-704.
K. Li, and Y. Fu. 2016. "Actionlets and Activity
Prediction." In Human Activity Recognition and
Prediction, 123-151.
W. Lotter, G. Kreiman and D. Cox. 2016. "Deep predictive
coding networks for video prediction and unsupervised
learning." arXiv preprint arXiv:1605.08104.
Z. Luo, B. Peng, D. Huang, A. Alahi and L. Fei-Fei. 2017.
"Unsupervised learning of long-term motion dynamics
for videos." arXiv preprint arXiv:1701.01821.
M. Mathieu, C. Couprie and Y. LeCun. 2015. "Deep multiscale video prediction beyond mean square error." arXiv
preprint arXiv:1511.05440.
N. Neverova, P. Luc, C. Couprie, J. Verbeek and Y. LeCun.
2017. "Predicting Deeper into the Future of Semantic
Segmentation." arXiv preprint arXiv:1703.07684.
M. Ranzato, A. Szlam, J. Bruna, M. Mathieu, R. Collobert
and S. Chopra. 2014. "Video (language) modeling: a
baseline for generative models of natural videos." arXiv
preprint arXiv:1412.6604.
F. Rezazadegan, S. Shirazi, B. Upcrofit and M. Milford.
2017. "Action recognition: From static datasets to
moving robots." In IEEE International Conference on
Robotics and Automation, 3185-3191.
F. Rezazadegan, J. Gengb, M. Ghirardi, G. Menga, S.
Murèb, G. Camuncolib and M. Demichelac. 2015.
"Risked-based design for the physical human-robot
interaction
(pHRI):
An
overview."
Chemical
Engineering Transactions 43: 1249-1254.
F. Rezazadegan, S. Shirazi, N. Sunderhauf, M. Milford and
B. Upcroft. 2015. "Enhancing human action recognition
with region proposals." In Proceedings of the
Australasian Conference on Robotics and Automation.
M. Rohrbach, S. Amin, M. Andriluka and B. Schiele. 2012.
"A database for fine grained activity detection of
cooking activities." In IEEE Conference on Computer
Vision and Pattern Recognition, 1194-1201.
O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S.
Ma, Z. Huang, A. Karpathy, A. Khosla and M.
Bernstein. 2014. "Imagenet large scale visual recognition
challenge." arXiv preprint arXiv:1409.0575.
M. S. Ryoo, 2011. "Human activity prediction: Early
recognition of ongoing activities from streaming videos."
In IEEE International Conference on Computer Vision,
1036-1043.
M. Sadegh Aliakbarian,, F. S. Saleh, M. Salzmann, B.
Fernando, L. Petersson and L. Andersson. 2017.
"Encouraging LSTMs to Anticipate Actions Very
Early." In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, 280-289.
K. Soomro, H. Idrees and M. Shah. 2016.
where and what of actors and actions
action localization." In Proceedings
Conference on Computer Vision
Recognition, 2648-2657.
"Predicting the
through online
of the IEEE
and Pattern
N. Srivastava, E. Mansimov and R. Salakhudinov. 2015.
"Unsupervised learning of video representations using
lstms." In International Conference on Machine
Learning, 843-852.
S. Stein, and S. J McKenna. 2013. "Combining embedded
accelerometers with computer vision for recognizing
food preparation activities." In Proceedings of the ACM
international joint conference on Pervasive and
ubiquitous computing, 729-738.
C. Vondrick, H. Pirsiavash and A. Torralba. 2015.
"Anticipating the future by watching unlabeled video."
arXiv preprint arXiv:1504.08023.
T. Vu, C. Olsson, I. Laptev, A. Oliva and J. Sivic. 2014.
"Predicting actions from static scenes." In European
Conference on Computer Vision, 421-436.
L. Wang, Y. Xiong, Z. Wang, Y. Qiao, D. Lin, X. Tang and
L. Van Gool. 2017. "Temporal Segment Networks for
Action Recognition in Videos." arXiv preprint
arXiv:1705.02953.
B. Zhou, A. Lapedriza, J. Xiao, A. Torralba and A. Oliva.
2014. "Learning deep features for scene recognition
using places database." In Advances in neural
information processing systems, 487-495.
| 1cs.CV
|
arXiv:1710.00194v1 [math.OC] 30 Sep 2017
Where computer vision can aid physics: dynamic
cloud motion forecasting from satellite images
Sergiy Zhuk
Hendrik Hamann
Tigran Tchrakian
Albert Akhriev
Siyuan Lu
Abstract
This paper describes a new algorithm for solar energy forecasting from a sequence of
Cloud Optical Depth (COD) images. The algorithm is based on the following simple observation: the dynamics of clouds represented by COD images resembles the motion (transport)
of a density in a fluid flow. This suggests that, to forecast the motion of COD images, it is
sufficient to forecast the flow. The latter, in turn, can be accomplished by fitting a parametric
model of the fluid flow to the COD images observed in the past. Namely, the learning phase
of the algorithm is composed of the following steps: (i) given a sequence of COD images,
the snapshots of the optical flow are estimated from two consecutive COD images; (ii) these
snapshots are then assimilated into a Navier-Stokes Equation (NSE), i.e. an initial velocity
field for NSE is selected so that the corresponding NSE’ solution is as close as possible
to the optical flow snapshots. The prediction phase consists of utilizing a linear transport
equation, which describes the propagation of COD images in the fluid flow predicted by NSE,
to estimate the future motion of the COD images. The algorithm has been tested on COD
images provided by two geostationary operational environmental satellites from NOAA serving
the west-hemisphere.
I. I NTRODUCTION AND PROBLEM STATEMENT
Motivation. Solar energy is the most abundant form of renewable energy resources
and its contribution towards the total energy mix is rapidly increasing [9]. However,
integration of high penetration level of solar energy in the electric grid poses significant
challenge and cost [4]. The cloud movement, formation, dissipation and associated
variable shading of solar panels may result in steep ramps of solar power being injected
into the grid. The variability and uncertainty of solar power often forces the system
operators to hold extra reserves of conventional power generation which adds cost.
Ongoing research as well as previous experience of wind power integration shows that
accurate solar forecasting plays a key role in the reliable and cost-effective integration of
solar power [5]. For accurate short-term forecasting (a minute- to hour-ahead) of large
geographical areas (continental scale), forecasting models using geostationary operational
environmental satellite (GOES) imagery [11] are more effective than numerical weather
IBM Research, Dublin, Ireland, {sergiy.zhuk,tigran,albert}@ie.ibm.com
IBM T.J. Watson Research Center, Yorktown, U.S., {lus,hendrikh}@us.ibm.com
Fig. 1: Cloud optical depth and wind-field from the weather model on long-lat domain
[−140, −124]×[39, 51] at four consecutive times, 01.09.2013. A clear counter-clockwise
rotation of the COD field is largely reflected in the wind-field. However, the wind-field
suggests a strong upward flow towards the right of the domain which does not appear
to be reflected in the COD field.
prediction models as the latter typically takes too long (several hours) to ramp up. The
GOES satellite imagery only provides information on the current distribution of cloud,
including cloud optical depth (COD) and top/bottom altitude which can be converted to
the current solar irradiance at the earths surface using radiative transfer modeling. An
accurate cloud advection model is thus required to forecast the future cloud distribution
and solar irradiance.
Contribution. Conventionally cloud advections are performed using wind fields forecasted by numerical weather prediction (NWP) models [10] or wind fields determined
using optical flow method [3]. In the former case, even when the NWP forecasted wind
field is accurate, the inaccuracy in the estimate of the altitude of the clouds may still
lead to significant error in the wind field and advected cloud distribution (see Fig. 1).
In the latter case, even though optical flow may accurately determine the current “wind
field” at the altitude of the clouds, this estimated “wind field” is assumed to persist for
the forecasting period. In case when the actual wind field changes significantly in the
hour ahead time scale, large error in the predicted cloud distribution may occur.
In this paper, we introduce a new algorithm for cloud motion prediction which assimilates
purely data driven “wind fields” F~k = (uk , vk )> (k – discrete time index) derived by an
optical flow estimation, into incompressible Navier-Stokes Equation (NSE) in two spatial
dimensions. It builds upon a simple observation: the dynamics of clouds represented by
COD images resembles the motion (transport) of a optical density C(x, t), associated
with a COD image, in a fluid flow ~u. This suggests that, to predict the dynamics of the
density C(x, t) given a possibly sparse (in time) sequence of COD images, one can use
the following procedure:
(i) given a sequence of COD images, construct a map F~k = (uk , vk )> , which transforms
C(x, tk−1 ) into C(x, tk ), by using some optical flow estimator;
(ii) assimilate F~k directly into the NSE, i.e. an initial velocity field for NSE is selected
so that the corresponding NSE’ solution, ~u passes through {F~k };
(iii) predict the dynamics of the velocity field ~u by integrating NSE forward in time until
the desired horizon, and then utilize a linear transport equation (see equation (1)
below) to predict the dynamics of C(x, t).
To obtain maps F~k in step (i) we used an optical flow estimator from [12] which relies
upon a combination of coarse-to-fine estimation by 3-level Gaussian pyramid, bicubic
interpolation of image derivatives, and robust penalty function augmented by median
filtering.
Clearly, the quality of the estimated optical flow has strong impact onto the prediction
accuracy. We stress that the COD images may be noisy due to the acquisition and
transmission processes, and, in fact, this noise may mislead the estimator. In addition,
it is well known that the optical flow estimation problem is generally ill-posed [6], i.e.
the COD images do not determine the unique map F~k . Moreover, the COD images are
usually sparse in time, and so they may not capture quite well fast underlying dynamic
processes. Hence, the optical flow estimator has to rely upon a regulariser which may
not account for the underlying physics of cloud motion. The latter may lead to some
non-physical patterns in the resulting estimate of the optical flow (see Fig. 2).
Fig. 2: Optical flow estimate displaying a number of non-physical patterns (see red
boxes).
To overcome this issue, at step (ii) above we further regularize the data driven sequence
of optical flows {F~k } obtained from (i): indeed, the solution of NSE captures some basic
physics of the wind dynamics, and it introduces a spatio-temporal coherence between
the subsequent optical flow estimates {F~k }. Indirect assimilation of COD images into
NSE has been previously considered in the literature (see [1], [6] and references therein).
The approach taken here is different in that we assimilate the estimated optical flows
directly into the NSE to simplify the computations and omit quite subtle numerical
difficulties arising in the process of indirect assimilation. The precise differences are
discussed section I-.1. Finally, the prediction is done by advecting COD images forward
in time by utilizing a discontinuous Galerkin (dG) approximation of the linear advection
equation. As a result, the proposed algorithm enables accurate cloud prediction at least 60
minutes ahead as it captures both the “current wind field” accurately determined by the
data driven optical flow estimate, and some basic physics of the wind dynamics.
n
n
Notation.
vector space, x ·
P C /R denotes the standard n-dimensional complex/real
y :=
xi y i denotes the canonical inner product in Cn /Rn , g denotes the complex
conjugate of g, A> denotes the transposed matrix, A? denotes the conjugate transposed
matrix, Ω ⊂ R2 denotes a computational domain, LR2 (Ω) denotes the space of Rmeasurable
squared-integrable functions over Ω, kf k2L2 (Ω) := Ω f 2 dxdy, (f, g)L2 (Ω) := Ω f gdxdy,
δt−tk ψ(x, y, t) := ψ(x, y, tk ), ΩT := Ω × (t0 , T ). ∂x and ∂y denote derivatives with
respect to x and y, curl ~u = vx − uy .
1) Formal problem statement.: Assume that COD images are observed over a domain
of interest, Ω (in two spatial dimensions). We say that the cloud optical depth at
position x := (x, y)> ∈ Ω and time t, C(x, t) is transported by a velocity field
~u = (u(x, y, t), v(x, y, t))> , which describes the flow of an incompressible fluid in Ω,
if
∂t C + uCx + vCy = 0 .
(1)
In computer vision literature, the latter equality is often reffered to as an optical flow
constraint, and it has the following interpretation: each pixel x(t) := (x(t), y(t))> of
a COD image moves with velocity ~u, i.e. ẋ(t) = u(x(t), y(t), t), ẏ = v(x(t), y(t), t),
and it carries the value of the optical density C(x, t) along the trajectory of x, i.e.
C(x(0), 0) = C(x(t), t).
We stress that the straightforward advection (warping), i.e. propagating each pixel x(t) of
a COD image subject to ~u at that pixel, may lead to developing artificial discontinuities
as displayed on Figure 3. The reason for this is that the warping is, in fact, equivalent
to a very basic numerical advection scheme, i.e. the position of the pixel x at time
t + 1 is obtained by using the forward Euler time integrator: x(t + 1) = x(t) +
dtu(x(t), y(t), t), y(t + 1) = y(t) + dtv(x(t), y(t), t). In this work we use a more
Fig. 3: Advecting a COD image by propagating each COD pixel subject to the velocity
field ~u at that pixel.
sophisticated discretization based on discontinuous Galerkin (dG) method which does
not introduce artificial discontinuities.
Dynamics of the fluid flow. As noted above, the dynamics of COD images resembles the
transport of a density/concentration of a mass in a fluid flow (in two spatial dimensions).
This suggests to model ~u as the unique solution of the following two-dimensional NavierStokes equation in the vorticity-streamfunction weak formulation:
d
(ω, φ)L2 (Ω) + b(ũ + ∇⊥ ψ, ω, φ) + a(ω, φ)L2 (Ω) = (f, φ)L2 (Ω) ,
dt
a(ψ, φ) = (ω, φ)L2 (Ω)
ω(0) = curl(~u0 ) , (x, t) ∈ ΩT
(2)
where φ is a test function, ũ ∈ R2 is the mean velocity field, ω is the vorticity, i.e.
ω = curl(~u), ~u0 is an initial velocity field, f is a control parameter, e.g. source/sink,
and
b(~u, w, v) := (~u · ∇, w)L2 (Ω) , a(ψ, φ) = (∇ψ, ∇φ)L2 (Ω)
= (ψx , φx )L2 (Ω) + (ψy , φy )L2 (Ω)
Note that the weak formulation above is obtained from the classical NSE by multiplying
it with a test function φ in order to relax the smoothness assumptions on the solution ω:
in the above formulation, ω needs to have just first order derivatives. We refer the reader
to [8] for the further mathematical details on various fomulations of the NSE. Note that,
by using integration by parts and selecting appropriate test/basis functions, it is not hard
to show that the weak formulation (2) incapsulates various boundary conditions for the
vorticity ω, namely periodic, homogeneous Dirichlet and homogeneous Neumann.
Note that ~u can be matched to a discrete (in time) sequence of COD images by recasting
the classical optical flow estimation problem as an optimal control problem: select the
control parameter f , and the initial velocity ~u0 such that the unique solution C of (1),
corresponding to the unique solution ~u of (2), is as close as possible (in some metric)
to the given sequence of COD images (see [1], [6] and references therein). We stress
that, for the case of COD images, the implementation of the latter control problem requires highly elaborate numerical schemes, and may be very expensive computationally.
Indeed, transport model (1) is described by a hyperbolic equation which has Dirichlet
boundary conditions on the inflow zone. The latter is a function of the coefficients ~u
(see Section II for the details) which makes it quite complicated to design a numerically
sound1 approximation of the gradient which is required to solve the aforementioned
control problem numerically.
For this reason, we propose a hybrid data assimilation strategy: (i) construct a map
F~k = (uk , vk )> which transforms COD image C(x, tk−1 ) into the consecutive image
C(x, tk ) by using an optical flow estimator, and (ii) to design an estimate of ~u which
is close (locally) to each of the F~k , and, at the same time, takes into account spatiotemporal patterns encoded in the COD images, and, importantly, is straightforward to
implement. Now we are ready to formalize the problem statement:
c −1
A) fitting problem: given a sequence of optical flow maps {F~k }N
k=1 estimated from a
c −1
sequence of COD images {C(x, tk )}N
k=1 , find an initial vorticity field q̂ and scalars
û, v̂ such that the cost function
Nc
X
u −(û+δ
ψ̂y )
k vk −(v̂−δ t−tk ψ̂ ) k2L2 (Ω)
J(q̂, û, v̂) :=
k
k=1
t−tk
x
(3)
is minimized along the solutions of (2), provided ũ = (û, v̂)> .
B) advection/prediction problem: given an estimate of ~u from A) construct a short-term
forecast of the future dynamics of C by using (1).
II. A LGORITHM
1) Optical flow estimation.: As noted in the Introduction, we constructed maps F~k
transforming C(x, tk−1 ) into C(x, tk ) by using the optical flow estimator from [12]
which aggregates state of the art optical flow estimaton techniques. This and a data
preprocessing step are described below, in section III.
1
i.e. the approximation which passes the so called gradient test
2) Fitting problem for NSE: Note that J defined in (3) is a non-linear and non-convex
functional. In what follows we provide a discrete representation of the gradient of J and
apply the L-BFGS version of the quasi-Newton method to solve the fitting problem (see
point A) above).
To compute a discrete gradient of J we proceed as follows. We set Ω = (0, Lx ) ×
(0, Ly ) and select the basis in L2 (Ω) generated by complex-valued functions2 φ(x, y) =
φc (x)φd (y), φn (x) = e2πinz/Lx . This system is orthonormal in L2 (Ω), i.e. (φi φj , φc φd )L2 (Ω) =
δi,c δj,d , provided the inner product is rescaled as follows: (f, g)L2 (Ω) := Lx1Ly (f, g)L2 (Ω) .
Now, denote by ω̂ and ψ̂ the solution of (2) which corresponds to the initial vorticity
ω̂(0) = q̂ and f = 0. We approximate the gradient of J by using Fourier-Galerkin (FG)
method. To this end, we define the projection operator PNx ,Ny :
X
PNx ,Ny v :=
vnm (t)φn (x)φm (y) .
|n|≤Nx /2
|m|≤Ny /2
It projects any v ∈ L2 (Ω) onto a (Nx + 1)(Ny + 1)-subspace of L2 (Ω) = L2 (0, Lx ) ×
L2 (0, Ly ), which is generated by functions Φ := {φc φd }|c|≤ Nx ,|d|≤ Ny . Now, Fourier2
2
Galerkin method suggests (i) to substitute PNx ,Ny q̂, PNx ,Ny ω and PNx ,Ny ψ into (2), and
(ii) to plug the test function φ = φc φd , for any |c| ≤ N2x and |d| ≤ N2y , into the resulting
system to generate a system of Ordinary Differential Equations (ODE) which defines
the dynamics of the vector of the projection coefficients ω̂ := {ω̂pq }|p|≤ Nx ,|q|≤ Ny . This
2
2
ODE takes the following form:
dω̂
= −B(ω̂)ω̂ − B(ũ)ω̂ − Aω̂ , ω̂(0) = q̂ ,
dt
(4)
P
x Ly
δp+n,c δq+m,d } with |c|, |n| ≤ N2x and |m|, |d| ≤ N2y
where B(ω̂) = { p,q ω̂pqp(pm−qn)L
2 L2 +q 2 L2
y
x
is a skew-symmetric matrix approximating the non-linear advection term in (2), and
A := diag(λ− Nx ,− Ny . . . λ Nx , Ny ) is a symmetric non-negative definite matrix composed
2
2
2
2
2 2
2 2
of eigenvalues of the Laplacian, λcd := 4πL2c + 4πL2d , c2 + d2 > 0, λ0,0 = 0. We refer
x
y
the reader to [8] for the further details on FG method.
Now, given q̂ we can approximate J as follows: J(q̂, û, v̂) ≈ J(PNx ,Ny q̂, û, v̂). As a
result, J depends on a finite number of parameters: the vector of projection coefficients
q̂ representing PNx ,Ny q̂, û and v̂. Effectively, to each trajectory of (4) which corresponds
to the initial condition q̂ and coefficients û and v̂, J assignes a number, the Euclidian
distance between the velocity field u(x, y, tk ) = ψ̂y (x, y, tk ), v(x, y, tk ) = ψ̂x (x, y, tk ),
computed from the trajectory, and the corresponding optical flow map F~k = (uk , vk )> .
2
In what follows, for a product of two basis functions, we always place the function of x to the left.
In other words, minimizing J is equivalent to projecting F~k onto the manifold generated
by trajectories of (4).
To fit q̂, û, v̂ to the “data” F~k = (uk , vk )> we need to compute the gradient of J
w.r.t. to q̂, û, v̂. The latter can be done by using a so-called adjoint method: one
computes Gateaux derivative of J w.r.t. q̂, obtaines a linear ODE for the gradient,
and, finally, uses its adjoint equation to find the gradient explicitly. Here, we denote the
solution
by λk and form the gradient as follows: ∇J(q̂, û, v̂) =
P c of? the adjoint equation
>
2 N
Ψ
λ
(0),
J
,
J
)
.
λ
,
û
v̂
k Ψ1,ı , Jû and Jv̂ are defined in the appendix . This
k=1 1,ı k
gradient is then used to approximate the solution of the fitting problem (3) by means of
the L-BFGS method [2] provided F~0 = (u0 , v0 )> is used as the starting point.
3) Advection/prediction problem.: In what follows we apply the dG method for solving
a linear advection equation with non-stationary coefficients:
∂t C(x, t) + ~u · ∇C(x, t) = 0,
x ∈ Ω ∈ R2
(5)
with Dirichlet boundary conditions C(x, t) = g on the inflow zone of the boundary of
Ω, ∂Ω. The inflow zone is composed of all x ∈ ∂Ω such that ~u(x, t) points inside Ω.
Since ~u varies in time, the inflow zone could “travel” along the boundary of Ω reflecting
changes of ~u. This makes it quite complicated to fit the initial vorticity q̂ to the COD
images. Indeed, C depends on ~u which depends on q̂. Hence, to find q̂ by minimizing
some distance between C and the COD images, one would need to differentiate the weak
formulation (8) of (5) w.r.t. the inflow zone. The latter does not seem to be practical
from the computational stand-point. One could have overcome this by resorting to the
vanishing viscosity method but this approach would require one to deal with numerical
boundary layers. To overcome these difficulties, we are assimilating F~k into (2) directly,
and then use (5) just for the advection/prediction.
Following [7] we approximate C by using dG approach as follows. Ω is divided into K
K
S
Dk . On each element
of non-overlapping rectangular elements, Dk , i.e., Ω ' Ωh =
k=1
Dk , C is approximated by Chk , which is expressed as the series,
Chk (x, t) =
N
+1
X
C k (xki , t)`ki (x),
x ∈ Dk ,
(6)
i=1
where `ki (x) are Lagrange interpolating polynomials in two dimensions defined by
Legendre-Gauss-Lobatto (LGL) points xi (see [7]). Substituting Chk into (5), we form
the residual Rhk on a single element:
Rhk = ∂t Chk (x, t) + ~u · ∇Chk (x, t),
x ∈ Dk .
(7)
The Galerkin methods involves taking `ki as test functions (i.e. same as the expansion/trial
functions) and forcing the residual to be orthogonal to each of these test functions. Doing
this, and then using integration by parts to move the spatial derivatives off the state and
onto the test functions gives the following weak statement on element Dk :
Z
Chk (x, t)t `kn (x) − fhk (x, t) · ∇`kn (x) dx =
Dk
Z
(8)
−
n̂ · f ∗ `kn (x)dx, n = 1 . . . N + 1,
∂Dk
where n̂ is the outward facing unit normal, fhk = (uChk , vChk )> and f ∗ is the numerical
flux function which we take to be the local Lax-Friedrichs flux:
C i u i + C e u e cs
+ n̂(ui − ue ),
(9)
f ∗ (Ci , Ce , ui , ue ) =
2
2
where subscripts i and e refer respectively to the interior and exterior values at a point
on the boundary, and cs is the maximum absolute value of the signal speed normal to
the boundary at that point, i.e.,
cs = max |n̂x uj + n̂y vj |.
j∈{i,e}
(10)
The surface integral in (8) allows the elements to ‘communicate’ with one another. Since
we are using rectangular elements, this surface integral is the sum of four line integrals,
each one over one face of the element. The exterior solution values in (9) and (10) refer
to the values of C on a neighbouring element sharing the node in question. In the case
that this node lies on the physical boundary, ∂Ω, the exterior value is determined by
a “physical” Dirichlet boundary condition g. If flow direction at that node is ‘into’ the
domain, Ω, then the exterior solution values is set to a prescribed value subject to the
boundary conditions (inflow zone). In this case, n̂ · ~u < 0 at a point on the boundary.
If, on the other hand, the flow direction at that boundary point is ‘out of’ the domain
(i.e. n̂ · u > 0), then a free exit boundary condition is imposed at that point by setting
the exterior values equal to the interior values.
The weak DG formulation on a single element Dk can be written as:
4
Mk
X
d k
Ih − Sx> Fx − Sy> Fy = −
(−1)i Mek,i Fi∗
dt
i=1
(11)
where Mk , Sx and Sy are the mass and stiffness matrices with the latter corresponding
to advection in the x- and y-directions. The vectors Ihk is a grid-function representing
Chk on the (N +1)2 quadrature points on element Dk , and the vectors Fx and Fy are grid
functions representing the first and second components of fhk respectively. The matrices,
Mek,i are edge-mass matrices for element Dk on face i, where the faces, i = 1 . . . 4, are
ordered: left, right, lower, upper. These matrices act on the vector Fi∗ which represents
the numerical flux over each node on face i. An easy way to define the mass and stiffness
matrices for 2D rectangular elements is to do so in terms of LGL quadrature points and
weights in the 1D interval I = [−1, 1] and also in terms of Lagrange polynomials defined
on this interval. We refer the reader to [13] for the further implementation details.
III. E XPERIMENTS
a) COD images: The aforementioned algorithm has been tested on data from two GOES
serving the west-hemisphere, GOES-15, the WEST satellite located over the Equator at
135◦ west longitude and GOES-13, the EAST satellite located over the Equator at 75◦
west longitude. The COD products of 15 minute interval and 4 km spatial resolution are
derived from GOES data using CLAVR-x [14] made available by the NOAA Advanced
Satellite Products Branch and Cooperative Institute for Meteorological Satellites Studies
at the University of Wisconsin Madisons Space Science and Engineering Center (SSEC).
All data are on a longitude-latitude grid with uniform 0.02◦ spacing. The area Ω covered
by the data extends from −140◦ East (longitude) and 39◦ North (latitude) to −124◦
degrees East (longitude) and 51◦ North (latitude). For the experiment we used COD
images recorded at 1 September 2013 over Ω. (see Figure 4).
Fig. 4: Example of a COD image over Ω, 01.09.2013, 21:15.
b) Optical flow estimates: We estimated the optical flow maps F~1 . . . F~4 from the COD
images displayed in Figure 5 by using the publicly available Matlab implementation
of the optical flow estimator from [12]. It is important to note that the COD images
required some preprocessing prior to invoking the estimator: without preprocessing the
resulting maps F~k were of little or no use. The preprocessing step involved a number of
transformations. First, the intensities of bright regions of a COD image are usually very
high compared to the average intensity outside these regions. Careful data inspection
revealed that the distortion preserves the local ordering and the local pattern. The
following log-transform was p
used to alleviate the distortion of intensity at each image
pixel (x, y): C̃(x, y) = σ · log (1 + (C(x, y)/σ)2 )/ log (2), where σ is a threshold
value. If C(x, y) = σ, then C̃(x, y) = C(x, y). For C(x, y) < σ the C̃(x, y) is nearly
a linear function of C(x, y), whereas for C(x, y) > σ the logarithm dominates in the
transformation. The threshold value σ was selected as the 90th percentile, which indicates
the magnitude such that 90% percent of the pixel intensities have values less than this
number. In other words, 10% of intensities were considered as potential outliers and
those were suppressed by the logarithm. The optical flow was computed on transformed
images with noticeable improvement comparing to the untransformed case.
Finally, the mean brightness of consecutive images C(x, tk ) and C(x, tk+1 ) tends to variate significantly possibly due to some problems occurring in the acquisition/transmission
processes. This variation was compensated by equating the median intensity values of
C(x, tk ) and C(x, tk+1 ): the medians are computed over the non-zero (“day”) pixels
and the first image C(x, tk ) is scaled to make both medians equal.
After the data preprocessing step, F~1 was estimated from two consecutive (15 minutes
apart) images (Fig. 5, top left image, 21:30) and so on. The resulting maps are presented
in Figure 5,6.
To fit NSE solution to the COD images we used the algorithm of Section II-.2. We took
Nx = Ny = 28, Lx = 1779 and Ly = 941.8 and timestep was taken to be the same
as in the dG advection (see below). The NSE velocity field is compared versus F~4 in
Figure 7. As expected, the NSE velocity looks more regular.
c) Forecasting results: In order to propagate COD using ~u, the following change of
variable is made:
x = rλ,
y = r sin ϕ,
(12)
where r is the radius of the earth, λ denotes longitude and ϕ denotes latitude. Figure 8
shows the standard long-lat coordinate system, and Figure 9 illustrates the coordinate
transformation given by (12). The new variables, x and y, are thus “space” variables
corresponding to the longitude, λ, and latitude, ϕ, respectively. Note that at the poles,
the transformation, (12), is invalid since the longitude coordinates coalesce. Thus, it is
not suitable for regions close to either pole.
Fig. 5: COD images/optical flows F~k (yellow arrows) at four consecutive times, 21:30
- 22:15, 01.09.2013.
The velocity field obtained through the optical flow procedure is in units pixel per time
unit. In order to use this field for COD propagation, it must first undergo two conversions/transformations. First, we must convert the field to km/h. The time conversion
is trivial as we know that the images are 15 minutes apart. For the space conversion,
we must take into account how much space each pixel represents depending on its
position. Assuming the earth to be spherical, the dimensions of a pixel do not depend
on its longitudinal coordinate. However, its dimensions do depend on the latitudinal
coordinate in the following way. The length of a degree of latitude does not vary much
with the latitudinal coordinate, so we assume that each degree of latitude represents the
same distance regardless of its coordinates. So for each pixel, we have
∆xlat = r∆ϕ
(13)
where ∆xlat is the distance a pixel represents along a line of longitude (i.e., distance
along latitude axis) and ∆ϕ is the latitude pixel spacing in radian measure which is
uniform over the image.
The length of a degree of longitude, however, varies significantly with the latitude
Fig. 6: COD image/optical flow map F~4 at (yellow arrows) 22:15
coordinate. We have
∆xlong = ∆λ · r cos ϕ
(14)
where ∆xlong is the distance a pixel represents along a line of latitude and ∆λ is the
longitude pixel spacing in radian measure which is also uniform over the image.
Having used the above the obtain the field, ~u in km/h, we must transform it from the
(λ, ϕ) coordinate system to the (x, y) coordinate system. To do so, the following is
applied:
u = U/ cos ϕ, v = V cos ϕ.
(15)
The new velocity field, (u, v) is valid on the x-y grid. Note that both velocity components
are transformed based on the latitudes of the velocity data on the original long-lat grid,
and not on the longitudes. Looking at Figure 9, it is clear that moving away from the
equator (|ϕ| increasing), the transformation, (12), compresses the grid longitudinally,
and expands it latitudinally. The velocity transformation, (15), accounts for this by
respectively increasing and decreasing the magnitudes of the latitudinal and longitudinal
components of the velocity field with increasing |ϕ| in transforming from the (λ, ϕ)
coordinate system to the (x, y) system.
To advect COD images we employed the algorithm of Section II-.3. We set Ω :=
[−15567, −13788]×[4009, 4951] and discretize Ω with 200×200 elements of order N =
3. The timestep was selected depending on the advection coefficients speed to ensure that
the CFL condition holds. The advection starts at 22:15 and uses the predicted velocity
Fig. 7: Optical flow F~4 (red) at 22:15 versus NSE velocity field (blue) at 22:15
r
Fig. 8: Longitude-latitude (λ-ϕ) coordinate system.
rsin
r
Fig. 9: Schematic of transformation from (λ, ϕ) to (x, y) coordinate system for a long-lat
region.
Date: 2013-09-01
22:30
22:45
23:00
23:15
NSE
34
43
42
48
Persistence
57
71
77
100
Optical flow
38
49
48
56
TABLE
I:
Relative
mean P absolute
percentage
P
100% s∈I |Ctrue (xs , t) − Cpredicted (xs , t)| / s∈I |Ctrue (xs , t)|.
errors:
field ~u from NSE to predict dynamics of COD images up to 23:30. We compared this
prediction against two most popular predictions, namely the so called persistency forecast
(the latest available COD image is used for the forecast) and optical flow forecast (map
F~4 displayed at Figure 7). The prediction results are summarized in the Table 1 which
shows that the NSE prediction outperforms the aforementioned popular predictors.
Here I is a set of indices of the grid points, which exclude the outliers. As noted above, in
Section II-.1, some COD image values are abnormally high due to acquisition problems.
These points (about 10%) are therefore excluded from the assessment.
IV. C ONCLUSION
This work presents a new algorithm for dynamic cloud optical density forecast from
satellite images. The algorithm relies upon ideas from data assimilation, numerical
analysis and classical image processing. It shows good performance on real satellite
images and outperforms the popular forecasting algorithms in the considered test cases.
A challenging topic for the future research is to develop a more accurate model for the
velocity field, e.g. take into account the fact that the 2D projection of a 3D velocity
field could be compressible, and to improve upon the linear advection model, e.g. to
introduce uncertain sources/sinks allowing the COD images changing intensity due to
drops/increases in temperature.
a) Acknowledgement: This research is partially supported3 by US Department of Energy
Contract NDE-EE0006017.
3
This report was prepared as an account of work sponsored by an agency of the United States government. Neither the United States government nor any agency thereof,
nor any of their employees, makes any warranty, expressed or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any
information, apparatus, product, or process disclosed, or represented that its use would not infringe privately owned rights. Reference herein to any specific commercial product,
process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United
States government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States government or any
agency thereof.
A PPENDIX
Let ω̂ := {ω̂pq }|p|≤ Nx ,|q|≤ Ny be the vector of the projection coefficients of PNx ,Ny ω̂,
2
2
and define q̂, vk , uk analogously. Note that ω̂ = Ψ √1 ,1 Ψ?√1 ,1 ω̂, provided Ψa,b is an
2
2 I 0 bJ
(Nx +1)(Ny +1) projection matrix defined by: Ψa,b = a 0 0 0 , i.e. ω̂− Nx +k,− Ny +s =
J 0 −bI
Nx Ny
, 2
2
2
2
and ω̂0,0 = 0 (ω̂ has zero mean!). We derive
ω̂ Nx −k, Ny −s for any 0 ≤ k, s ≤
2
2
the expression for the gradient of J by using a so-called adjoint method: one computes
Gateaux derivative of J w.r.t. q̂, obtaines a linear ODE for the gradient, and, finally,
uses its adjoint equation to find the gradient explicitly (m := (Nx + 1)(Ny + 1)):
Jû (q̂, û, v̂) = 2
Nc
X
(dk , ω û (tk ))C m − (uk − û, 1)L2 (Ω)
k=1
û
dω
= −B([ 10 ])ω̂ − B1 (ω̂)ω û − B(u)ω û − B(ω̂)ω û − Aω û ,
dt
Nc
X
(dk , ω v̂ (tk ))C m − (vk − v̂, 1)L2 (Ω) ,
Jv̂ (q̂, û, v̂) = 2
k=1
v̂
dω
= −B([ 01 ])ω̂ − B1 (ω̂)ω v̂ − B(u)ω v̂ − B(ω̂)ω v̂ − Aω v̂ ,
dt
ω û (0) = ω v̂ (0) = 0 , dk := A?x (vk + Ax ω̂(tk )) − A?y (uk − Ay ω̂(tk )) ,
∇q̂ J(q̂, û, v̂) = 2
Nc
X
Ψ?1,ı λk (0) ,
k=1
dω̂
= −B(ω̂)ω̂ − B(ũ)ω̂ − Aω̂ , ω̂(0) = q̂ ,
dt
dλk
= B1? (ω̂)λk − B(ω̂)λk − B(ũ)λk + Aλk ,
dt
λk (tk ) = A?x (vk + Ax ω̂(tk )) − A?y (uk − Ay ω̂(tk ) ,
2πin
Ny ) ,
} Nx
Ax,y := diag({
λnm Lx,y |n|≤ 2 ,|m|≤ 2
B1 (ω̂) = [B(Ψ √1 ,1 e1 )ω̂ . . . B(Ψ √1 ,1 em )ω̂]Ψ?√1 ,1 .
2
2
2
Note that all the differential equations above are linear, i.e. they can be represented in
the form ẋ = Q(t)x+s, but the equation for ω̂. This latter equation can be approximated
by the following implicit time integrator: ω̂0 = q̂ and
ω̂t+1 − ω̂t
ω̂t+1 + ω̂t
= (−B(ω̂t ) − B(ũ) − A)
.
dt
2
(16)
B(ω̂n ) and B(ũ) are skew-symmetric matrices, and −A is a symmetric non-negative
definite matrix, so that the above time integrator is unconditionally stable for dt > 0,
and it converges to the exact solution provided dt → 0 [15]. Once the discrete in time
sequence {ω̂t } is computed, one can approximate solutions of the aforementioned linear
equations by using the standard simplectic midpoint method:
Qt+1 + Qt xt+1 + xt ft+1 + ft
xt+1 − xt
=
+
.
dt
2
2
2
R EFERENCES
[1] D. Béréziat and I. Herlin. Solving ill-posed image processing problems using data assimilation. Numerical
Algorithms, 56(2):219–252, February 2011.
[2] R. Byrd et al. A limited memory algorithm for bound constrained optimization. SIAM J. Sci. Comp., 5(16),
1995.
[3] C. Chow. Cloud motion and stability estimation for intra-hour solar forecasting. Solar Energy, (115), 2015.
[4] P. Denholm et al. Production cost modeling for high levels of photovoltaics penetration. Energy Policy, 35,
2007.
[5] H. Hamann et al. Recent trends in variable generation forecasting and its value to the power system. IEEE
Trans. Sust. Energy, 2014.
[6] Isabelle Herlin and et al. Divergence-free motion estimation. In ECCV, pages 15–27, 2012.
[7] J. Hesthaven and T. Burton. Nodal Discontinuous Galerkin Methods. Springer, 2008.
[8] A. Majda and A. Bertozzi. Vorticity and incompressible flow. Cambridge Univ. Press, 2002.
[9] R. Margolis et al. Integration of solar into U.S. electric power system. Technical report, U.S. DoE, 2012.
[10] S. Miller et al. Cloud advection schemes for short-term satellite-based insolation forecasts. In Proc. of World
Renewable Energy Forum, 2012.
[11] R. Perez et al. A new operational satellite-to-irradiance model. Solar Energy, 73, 2002.
[12] D. Sun et al. A quantitative analysis of current practices in optical flow estimation and the principles behind
them. Int. J. of Computer Vision, 2013.
[13] T. Tchrakian et al.
A fast distributed data-assimilation algorithm for divergence-free advection.
arXiv:1707.07316v1, 2017.
[14] S. Thomas. Comparison of NOAA’s operational AVHRR-derived cloud amount to other satellite-derived cloud
climatologies. J. Appl. Meteorol., 2004.
[15] S. Zhuk and T. Tchrakian. Parameter estimation for euler equations with uncertain inputs. In Proc. of IEEE
Conference on Decision and Control, 2015.
| 1cs.CV
|
arXiv:1703.10101v1 [math.GR] 29 Mar 2017
Topological finite generation of compact open
subgroups of universal groups
Marc Burger1 and Shahar Mozes
∗ 2
1
2
ETH Zürich
The Hebrew University of Jerusalem
This paper is part of our ongoing study of the structure of co-compact
irreducible lattices in the product of two regular trees, see [BM00], [BM00b],
[BMZ09], [Rat04], [BM14]. The specific result obtained here, see Corollary 0.2,
is motivated by the following question: Let Td = (V, E) be the d-regular tree,
F < Symd a permutation group on d letters and U(F ) < Aut Td the universal
group attached to it; see [BM00], for definitions and properties. When is U(F )
the closure of the projection of a co-compact lattice Γ < U(F ) × G, where G is
locally compact, compactly generated? Work of D. Rattaggi (using [BM00b])
implies that this is so when F = A2n is the alternating group with 2n ≥ 6,
or F = M12 is the 12-Matthieu group, or F = ASL3 (2) is the special affine
group on F32 ; in addition there are co-compact lattices in U(A2n ) × U(A2m )
with dense projections for n ≥ 13, m ≥ 9.
Our starting observation is that if Γ < U(F ) × G is a co-compact lattice
with G as above, then any compact open subgroup of U(F ) is topologically
finitely generated. Our main result gives a characterization of the permutation
groups F for which this property holds. In fact, this result will follow from
a similar result concerning the finite generation of iterated wreath products.
More precisely, let D be a finite set and L < Sym(D) a permutation group.
Let Ln := L ≀ · · · ≀ L be the n’th iterated wreath product of L acting on D, and
L∞ := lim Ln the projective limit. Note that L∞ has a natural action on the
←
rooted tree whose vertices are labeled by words over D (the root is the empty
word).
M.B. thanks the hospitality of the IMS of the National University of Singapore, S.M.
acknowledges the support of ISF grants 1003/11 and 2095/15. We both thank the hospitality
of the University of Newcastle and the ANSI program ”Winter of disconnectedness”. We
thank the Isaac Newton Institute for Mathematical Sciences for support and hospitality
during the program on ”Non-positive curvature group action and cohomology”. This work
was supported by EPSRC Grant Number EP/K032208/1.
AMS Subject classification 2010: 20E18, 20E08, 22D05
∗
1
Theorem 0.1. The profinite group L∞ is topologically finitely generated if and
only if the two following conditions are satisfied:
a) L is perfect.
b) Every L-orbit in D has at least two elements.
If these conditions are satisfied, then L∞ is positively finitely generated.
This result was announced in [Moz98]. A version of the proof was written
up in the thesis of O. Amann [Ama03] following a manuscript by the authors.
The proof presented here substantially simplifies the original one.
Applied to our situation, let F1 = StabF (1), and D = {2, . . . , d}. Then
for every m ≥ 1, the pointwise stabilizer in U(F ) of the ball or radius m ≥ 1
centered at any vertex is a finite direct product of (F1 )∞ = limn (F1 ≀ · · · ≀ F1 )
to which our theorem applies and gives
Corollary 0.2. The profinite group obtained by considering the stabilizer in
U(F ) of a vertex is topologically finitely generated if and only if the following
two conditions are satisfied:
a) F1 is perfect.
b) Any F1 orbit in {2, . . . , d} has at least two elements.
Moreover, when those conditions are satisfied, any compact open subgroup of
U(F ) is positively finitely generated.
Recall that a compact group K is positively finitely generated if for some
ℓ ≥ 1, the Haar measure of the subset of ℓ-tuples in K ℓ generating a dense subgroup of K is positive. For background information concerning this property,
we refer to [Man96]. The question of finite generation of various infinite iterated wreath products was considered by several authors. See [Bha94], [Qui06],
[Qui04], [Bon10], [Van15]. All these works consider wreath products of transitive group actions.
Now we shortly discuss the necessity of the conditions in Theorem 0.1 and
to that end we start by fixing some notations used throughout this article.
The permutation groups Ln acting on Sn := D n are defined inductively by
L1 = L < Sym(D) and Ln+1 = Ln ⋉ LSn where an element in Ln+1 is a pair
(x, f ), f : Sn → L a map, and the product structure reads (x1 , f1 )(x2 , f2 ) =
(x1 x2 , f1x2 f2 ), and f x2 (w) = f (x2 w). The action of Ln+1 on Sn × D is given by
(x, f )(v, j) = (x(v), f (v)j).
Now we show the necessity of the two conditions.
2
a) If π: L → A is an abelian quotient of L, then
Ln ⋉ LSn −→ A
Q
π f (v)
(x, f ) 7−→
v∈Sn
is a surjective homomorphism; by induction this implies that An is a
quotient of Ln and hence L∞ admits AN as quotient. Thus, if L is not
perfect, L∞ is not topologically finitely generated.
b) If j ∈ D is L-fixed, then (j, . . . , j) ∈ Sn is Ln -fixed and
Sn
Q
v∈Sn
v6=(j,...,j)
L
is a normal subgroup of Ln+1 = Ln ⋉ L with quotient Ln × L. By
recurrence this implies that Ln is a quotient of Ln and hence LN of L∞ .
But if L 6= (e), the LN is not topologically finitely generated and hence
L∞ not as well.
The plan of the paper is the following: the proof of the converse of Theorem
0.1 follows a strategy devised by Bhattacharjee in [Bha94]. Namely, if pk (G)
denotes the probability for a k-tuple to generate the finite group G, we show
that under the hypothesis of Theorem 0.1, limn pk (Ln ) > 0 for some k. For this
we use a result of Bhattacharjee (see [Bha94]), which we recall below, relating
pk (Ln+1 ) to pk (Ln ) modulo a multiplication factor (1 − ζLn+1 /Ln (k − 1)) which
is defined in terms of the conjugacy classes of maximal subgroups of Ln+1
surjecting onto Ln . The main work consists then in classifying these conjugacy
classes.
1
A result of M. Bhattacharjee
We recall for the convenience of the reader the following:
Proposition 1.1. ([Bha94]) Let π: Y → X be a surjective homomorphism of
finite groups and k ≥ 1. Then
P
1
pk (X)
pk (Y ) ≥ 1 −
k−1
[M ]∈G(Y |X) [Y : M]
where the sum is over the set G(Y |X) of Y -conjugacy classes of proper maximal
subgroups of Y surjecting onto X.
Proof. We have pk (Y ) = pk (Y |X) pk (X), where pk (Y |X) is the probability for
a k-tuple to generate Y given that its image in X generates X. Then
pk (Y |X) = 1 − qk (Y |X)
3
where qk (Y |X) is the probability of the opposite event. Observing that a nongenerating k-tuple is always contained in a proper maximal subgroup M < Y ,
we obtain:
P
qk (Y |X) ≤
M Y maximal
=
|M|k
|Y |k
P
[Y : NY (M)] ·
P
1
.
[Y : M]k−1
[M ]∈G(Y |X)
≤
[M ]∈G(Y |X)
|M|k
|Y |k
P
1
[M ]∈G(Y |X) [Y ·M ]s ,
we conclude:
P
Corollary 1.2. Assume that for some k1 ≥ 2, ∞
k=1 ζLn+1 /Ln (k1 − 1) < +∞.
Let n1 be such that ζLn1 +1 /Ln1 (k1 − 1) < 1 and let k2 ≥ 2 be such that Ln1
is generated by k2 elements. Then L∞ = lim Ln is positively k-generated for
Denoting by ζY |X (s) =
←
k ≥ max(k1 , k2 ).
Proof. For every k ≥ k1 we have (Proposition 1.1)
Q
lim pk (Ln ) ≥
1 − ζLn+1 /Ln (k − 1) · pk (Ln1 ).
n→∞
n≥n1
Now observe that
Q
Q
1 − ζLn+1 /Ln (k − 1) ≥
n≥n1
n≥n1
1 − ζLn+1 /Ln (k1 − 1) > 0
and conclude by choosing k ≥ max(k1 , k2 ).
2
Maximal subgroups in wreath products
Let X be a finite group, Ω1 , . . . , Ωt transitive (non-empty) X-sets, t ≥ 1 and
B1 , . . . , Bt non-trivial perfect groups; these will be our standing assumptions
throughout this section. Let Y = X ⋉ (B1Ω1 × · · · × BtΩt ) be the semi-direct
product, where the action of X is by permuting factors, and π: Y → X the
projection map.
Definition 2.1. A standard normal subgroup of Y is a subgroup of the form
t
Q
i=1
NiΩi
where Ni ⊳ Bi
4
Observe that every subgroup M < Y contains a unique maximal standard
normal subgroup of Y .
Definition 2.2. A subgroup M < Y is clean if it contains no non-trivial
standard normal subgroup.
The following proposition summarizes the ingredients needed in the proof of
Theorem 0.1; this proposition is a corollary of more precise statements proven
in this section.
Proposition 2.3. Let M < Y be a clean, proper, maximal subgroup such that
π(M) = X.
Then one of the following holds:
1) t = 2, B1 , B2 are non-abelian simple and
(a) M ∩ (B1Ω1 × B2Ω2 ) is the graph of an isomorphism B1Ω1 → B2Ω2 .
(b) M = NY (M ∩ (B1Ω1 × B2Ω2 )).
In particular, there are at most |Ω1 | · |Out(B1 )| conjugacy classes of such
subgroups, and [Y : M] ≥ |B1 ||Ω1| .
2) t = 1, M ∩ B Ω 6= (e) and prw (M ∩ B Ω ) = B for all w ∈ B. There
is a normal subgroup U ⊳ B which is a product U = T r where T is
non-abelian simple. Moreover,
(a) M ∩ U Ω is a product of subdiagonals of T r·Ω corresponding to an
X-invariant block decomposition of r · Ω.
(b) M = NY (M ∩ U Ω ).
(c) For any given X-invariant block decomposition of r · Ω there are at
most |Out(T )|r|Ω| conjugacy classes of such subgroups.
(d) There are most (2|Ω|)r−1 arΩ such block decompositions of r · Ω where
aΩ is the number of X-invariant block decompositions of Ω.
(e) [Y : M] ≥ |T |
r·|Ω|
2
= |U|
|Ω|
2
.
3) t = 1, M ∩ B Ω 6= (e) and prω (M ∩ B Ω ) is a proper subgroup of B. Up to
conjugation prω (M ∩ B Ω ) = T for all ω ∈ Ω,
M = NY (T Ω )
and [Y : M] ≥ [B : T ]|Ω| .
5
4) t = 1, M ∩ B Ω = e.
Now we state and prove several lemmas which together will imply the above
proposition. We recall that we will work under the above standing assumptions
on the objects X, B1 , . . . , Bt , Ω1 , . . . , Ωt .
Lemma 2.4. Assume M < Y is proper, maximal, clean. Then Z(Bi ) = (e)
for all 1 ≤ i ≤ t.
Proof. Assume that for some j, Z(Bj ) 6= (e). Since M is clean and maximal,
Ω
Ω
Ω
M·Z(Bj )Ωj = Y and since Bj j ⊃ Z(Bj )Ωj this implies Bj j = (M ∩ Bj j )Z(Bj )Ωj .
Since Bj is perfect we obtain:
Ω
Ω
Ω
Ω
Ω
Bj j = [Bj j , Bj j ] = [M ∩ Bj j , M ∩ Bj j ] ⊂ M
which contradicts the assumption that M is clean.
Given a subgroup M < Y we let M 0 :=
T
y My −1 denote the kernel of
Q
Ω
the Y -action on Y /M. Also, given Ni < Bi , w ∈ Ωi , let Ni (ω) ⊂ tj=1 Bj j
denote the subgroup whose only non-identity component is at ω ∈ Ωi and
equals Ni .
y∈Y
Remark 2.5. If M < Y is clean, π(M) = X and Ni ⊳ Bi , then M 0 ∩ Ni (w) =
(e) for all ω ∈ Ωi .QIndeed, otherwise since π(M) = X, M would contain the
normal subgroup ω∈Ωi M 0 ∩ Ni (ω) which is standard.
Lemma 2.6. Assume M < Y is proper, clean, maximal and π(M) = X.
Then
t
Q
M0 ∩
BiΩi = (e).
i=1
Proof. If M 0 ∩
Qt
i=1
BiΩi 6= (e), then let j and ω ∈ Ωj be such that
t
Q
Ωi
0
Nj := prω M ∩
Bi
6= e.
i=1
Then Nj ⊳ Bj and
h
i
t
Q
[Nj (ω), Bj (ω)] = M 0 ∩
BiΩi , Bj (ω) ⊂ M 0 ∩ Bj (ω).
i=1
Since by Lemma 2.4, Z(Bj ) = (e) we must have [Bj (ω), Nj (ω)] 6= (e) and
hence M 0 ∩ Bj (ω) 6= (e). But this contradicts the cleanness assumption by
Remark 2.5.
6
Lemma 2.7. Let M < Y be clean, proper, maximal Q
with π(M) = X. If Y1 , Y2
are non-trivial normal subgroups of Y contained in ti=1 BiΩi such that
Y1 ∩ Y2 = (e).
Then both Y1 , Y2 act transitively and regularly on Y /M. In particular t = 2
or 1.
Proof.
The Y -action on Y /M is primitive and (Lemma 2.6) the subgroup
Qt
Ωi
i=1 Bi is faithful. The action of each of the normal subgroups Yi is transitive and faithful. Since they commute each of them must act freely and hence
regularly. Q
This implies that there cannot be three normal subgroups of Y contained in ti=1 BiΩi such that every two of them intersect only at the identity.
Thus t ∈ {1, 2}.
Lemma 2.8. Assume that M < Y is proper, clean, maximal with π(M) = X
and assume t = 2. Then B1 , B2 are non-abelian simple, M ∩ (B1Ω1 × B2Ω2 ) is
the graph of an isomorphism B1Ω1 → B2Ω2 and
M = NY M ∩ (B1Ω1 × B2Ω2 ) .
In addition, [Y : M] ≥ |B1 ||Ω1 | = |B2 ||Ω2 | .
Proof. It follows from Lemma 2.7 that both B1Ω1 and B2Ω2 act transitively and
regularly on Y /M. Hence, if (e) 6= Ni ⊳ Bi , then NiΩi being non-trivial, normal
in Y , acts transitively on Y /M, hence |NiΩi | = |BiΩi | and Ni = Bi which shows
that Bi is simple.
Next, B1Ω1 · M = Y , hence B1Ω1 × B2Ω2 = B1Ω1 · [M ∩ (B1Ω1 × B2Ω2 )] and thus
pr2 (M ∩ (B1Ω1 × B2Ω2 )) = B2Ω2 . Similarly, we have pr1 (M ∩ (B1Ω1 × B2Ω2 )) = B1Ω1 .
Since we also have M ∩B1Ω1 = M ∩B2Ω2 = (e), we deduce that M ∩(B1Ω1 ×B2Ω2 )
is the graph of an isomorphism B1Ω1 → B2Ω2 . Finally, we have
M ⊂ NY M ∩ (B1Ω1 × B2Ω2 )
and if equality does not hold then M ∩ (B1Ω1 × B2Ω2 ) is normal in Y and hence
equal to M 0 ∩ (B1Ω1 × B2Ω2 ), contradicting Lemma 2.6. The lower bound on
[Y : M] is clear.
Next, we need to estimate the number of Y -conjugacy classes of maximal
subgroups as in Lemma 2.8.
Lemma 2.9. The number of conjugacy classes of maximal subgroups as in
Lemma 2.8 is bounded by
|Out(B1 )| · |{v2 ∈ Ω2 : StabX (v2 ) = StabX (v1 )}|
where v1 ∈ Ω1 is some chosen element.
7
Proof. According to Lemma 2.8 the subgroup M is determined by an isomorphism F : B1Ω1 → B2Ω2 . Since B1 , B2 are non-abelian simple such an isomorphism is given by the following data:
(1) a bijection σ: Ω1 → Ω2
(2) for every w ∈ Ω1 an isomorphism ϕw : B1 → B2 .
Where
F (f )( σ(w) = ϕw f (w) ,
∀w ∈ Ω1 ,
∀f ∈ B1Ω1 .
Then M = NY (graph(F )) ⊃ graph(F ) and the condition that M surjects onto
X together with M ∩ B2Ω2 = (e) gives that for every x ∈ X there is a unique
hx ∈ B2Ω2 with (x, e, hx ) ∈ M. From this we deduce
(2.1)
hx1 x2 = hxx21 · hx2
∀x1 , x2 ∈ X.
Next, using that (x, e, hx ) normalizes the graph of F , we get for all f ∈ B1Ω1 ,
x ∈ X,
−1
−1
−1
−1
F (f x ) = hxx F (f )x (hxx )−1 .
Evaluating this at σ(ω) and for maps f ∈ B1Ω1 whose support is a singleton,
we get
(2.2)
xσ(w) = σ(xw) ∀w ∈ Ω1 ,
∀x ∈ X.
(2.3)
−1
ϕxw (b) = hx σ(w) ϕw (b) hx σ(w)
∀b ∈ B1 , ∀w ∈ Ω1 , ∀x ∈ X.
Now fix a point w0 ∈ Ω1 and observe that hx (σ(w0 )) ∈ Z(B2 ) = (e) for all
x ∈ StabX (w0 ). Together with (2.1) this implies that f2 : Ω2 → B2 given by
f2 (v) = hx2 (σ(w0 )) where x2 ∈ X is any element with x2 σ(w0 ) = v, is a well
defined map.
Rewriting the equation (2.1) in terms of f2 , we get
hx = f2x · f2−1 .
Using this one computes that the group (e, e, f2 ) M(e, e, f2 )−1 = M ′ corresponds to an isomorphism F ′ : B1Ω1 → B2Ω2 given by
1) the same equivariant bijection σ: Ω1 → Ω2 ;
2) a family of isomorphism ϕ′w : B1 → B2 which all coincide.
From this, one readily deduces the upper bound stated in Lemma 2.9.
8
Now we turn to the situation where t = 1, that is Y = X ⋉ B Ω , where we
have set B = B1 , Ω = Ω1 . This splits into four different cases which we will
analyze now. Our standing assumption on M < Y is that it is a proper, clean,
maximal subgroup which surjects onto X.
Lemma 2.10. Assume that M ∩ B Ω 6= (e), prw (M ∩ B Ω ) = B ∀ω ∈ Ω, and
that for every minimal normal subgroup N ⊳ B, M ∩ N Ω = (e). Then B is
the product of two isomorphic non-abelian simple groups. Furthermore:
M = NY (M ∩ B Ω ).
Proof. We claim that every non-trivial minimal normal subgroup N ⊳ B is
a direct factor. For ω ∈ Ω let Bω := {f ∈ B Ω : f (ν) ∈ N ∀ν 6= w}. From
Y = M · N Ω we deduce Bω = (M ∩ Bω ) N Ω and hence B = prω (M ∩ Bω ) N.
Observe that M ∩ Bω ⊳ M ∩ B Ω and thus prω (M ∩ Bω ) ⊳ B. Finally,
prω (M ∩ Bω ) ∩ N = prω (M ∩ N Ω ) = (e), which shows that N is a direct factor
of B. This implies using Lemma 2.7 that B is either non-abelian simple or
the product of two such groups. The assumption that for any minimal normal
subgroup N ⊳ B M ∩ N Ω = (e) implies that B is not simple.
If the last assertion of the lemma were not to hold, then M ∩ B Ω ⊳ Y and
hence M 0 ∩ B Ω = M ∩ B Ω 6= (e), contradicting Lemma 2.6.
Lemma 2.11. Assume that M ∩ B Ω 6= (e), prω (M ∩ B Ω ) = B ∀ω ∈ Ω and
there is N ⊳ B minimal normal with M ∩ N Ω 6= (e). Then N is a product of
(isomorphic) non-abelian simple groups.
(a) prw (M ∩ N Ω ) = N for all w ∈ Ω.
(b) M = NY (M ∩ N Ω ).
Proof. Since N Ω acts transitively, faithfully and not regularly on Y /M, it (and
hence also N) cannot be abelian. Since N is a minimal normal subgroup of B
it is therefore a product of isomorphic non-abelian simple groups.
Next, we observe that prω (M ∩ N Ω ) is normal in B = prω (M ∩ B Ω ), and
being non-trivial, we deduce prw (M ∩ N Ω ) = N for all w ∈ Ω. Again, if
M ∩ N Ω ⊳ Y , then M 0 ∩ N Ω = M ∩ N Ω 6= (e) contradicting Lemma 2.6.
It follows from Lemma 2.10 and Lemma 2.11 that when t = 1, M ∩B Ω 6= (e)
and prw (M ∩ B Ω ) = B for all w ∈ Ω, there is a normal subgroup U ⊳ B which
is a product U = T r where T is non-abelian simple and prw (M ∩ U Ω ) = U for
all w ∈ Ω. Since T is non-abelian simple, this implies already the assertion
2) (a) in Proposition 2.3. Concerning 2) (b), we observe that, as usual, if
M ∩ U Ω ⊳ Y , then M 0 ∩ U Ω = M ∩ U Ω 6= (e) which contradicts Lemma 2.6.
Next, we turn to assertions (c), (d) and (e) which will follow from the
following discussion. The subgroup M ∩ U Ω = M ∩ T r·Ω projects onto T for
9
each w ∈ r · Ω, where we recall that r · Ω denotes the disjoint union of r
copies of Ω with corresponding X-action. Since π(M) = X, the subgroup
H = M ∩ T r·Ω is a product of subdiagonals of T r·Ω corresponding to a Xinvariant block decomposition of r · Ω. Given a partition P of r · Ω such a
subgroup H is obtained in the following way: for each x ∈ r · Ω there is an
automorphism ϕx of T and H is the direct product over A ∈ P of the diagonal
subgroup of T A given by
ϕx (t) : t ∈ T, x ∈ A .
P
The index of H in T r·Ω is then |T |e where e =
A∈P (|A| − 1). Observe
that |A| ≥ 2 for every A ∈ P. Indeed otherwise for some ω ∈ Ω, M would
contain a T -factor of U(ω); but prω (M ∩ B Ω ) = B which implies in the case
of Lemma 2.11 where U = N is a minimal normal subgroup of B, that M ⊃
N(ω) and hence since the projection of M is X it follows that M ⊃ N Ω
contradicting the assumption that M is clean. In the case of Lemma 2.10
where U = B = T 2 , it follows that M would contain a T factor of B(ω).
Hence since the projection of M is X it follows that M ⊃ T Ω which since
T ⊳ B is a standard normal subgroup, contradicting the assumption that M
r|Ω|
is clean. This implies [Y : M] ≥ [T r·Ω : H] ≥ |T | 2 and establishes assertion
(e) of Proposition 2.3. Concerning assertion (c), just observe that we can
conjugate H by inner automorphisms of T r·Ω so that for a given partition we
have at most |Out(T )|r|Ω| conjugacy classes of such subgroups.
What remains is to estimate the number of X-invariant partitions of r · Ω.
e = Ω1 ∪ Ω2 .
To this end, let Ω1 , Ω2 be transitive X-sets and set Ω
If P is an X-invariant partition with |P| = k such that some A ∈ P
intersects Ω1 and Ω2 , then P induces partitions into k pieces of Ω1 and Ω2 .
Denoting by a(k, Ωi ) the number of X-invariant partitions with k pieces of Ωi
and by aΩi the total number of X-invariant partitions of Ωi , we have that the
number of X-invariant partitions of Ω1 ∪ Ω2 is estimated by:
aΩ1 · aΩ2 +
min(|Ω
1 |,|Ω2 |)
P
k a(k, Ω1 ) a(k, Ω2 ) ≤
k=1
(1 + m)aΩ1 aΩ2 ≤ 2m aΩ1 aΩ2 ,
where m = min(|Ω1 |, |Ω2|) and the estimate in the first line uses the observation
that to obtain a partition into k pieces of Ω1 ∪ Ω2 whose pieces meet both Ωi ’s
one needs to partition each Ωi into k pieces and pair them. The pairing of these
pieces is determined (by the transitivity of the action on each Ωi ) by choosing
for one piece of Ω1 a piece of Ω2 . Applying this inequality with Ω1 = (r − 1) · Ω
and Ω2 = Ω we get ar·Ω ≤ 2|Ω| a(r−1)Ω · aΩ and by recurrence,
(r−1)
ar·Ω ≤ 2|Ω|
(aΩ )r .
10
Lemma 2.12. Assume that M ∩ B Ω 6= (e) and prω (M ∩ B Ω ) =
6 B for all
ω ∈ Ω. Then up to conjugating M, there is a subgroup e 6= T
B such that
Ω
|Ω|
M = NY (T ). We have [Y : M] ≥ [B : T ] .
Proof. Let Tω := prω (M ∩ B Ω ), ω ∈ Ω. Then, since M surjects onto X,
all
M normalizes
Q the subgroups Tω are conjugated within B. In addition,
Ω
T
:
indeed,
it
is
the
unique
smallest
subgroup
of
B
containing
M ∩ BΩ
ω ω
and having direct product structure. Thus, modulo conjugating M, we may
assume Tω = T ∀ω ∈ Ω, hence M ⊂ NT (T Ω ). If they were not equal, we would
have that T ⊳ B, hence from M ·T Ω = Y we would deduce B Ω = (M ∩B Ω ) T Ω ,
but since M ∩ B Ω ⊂ T Ω this would imply T = B which is a contradiction.
The last case left is when M ∩ B Ω = (e), that is, M is a section of X in Y .
Concerning this, we cannot say anything in this degree of generality, and we
will have to estimate by recurrence the number of sections of Ln in groups of
the form Ln ⋉ B O , where O ⊂ Sn is an Ln -orbit.
3
Estimating ζLn+1 /Ln (k)
In order to prove Theorem 0.1, we will use Corollary 1.2 and show that for
some k ≥ 1,
∞
P
ζLn+1 /Ln (k) < +∞.
n=1
Thus we have to estimate
ζLn+1 /Ln (k) =
P
1
,
[Ln+1 : M]k
where the sum is over the conjugacy classes of proper, maximal subgroups M <
Ln+1 surjecting onto Ln . To this end, let Ms denote the maximal standard
normal subgroup of Ln+1 contained in M (see Definition 2.1). Let In be the
set of Ln orbits in Sn . Then
Q
(Bv )v ,
Ln+1 /Ms = Ln ⋉
v∈In
where Bv are quotients of L and M/Ms is a clean proper maximal subgroup
which surjects onto Ln . Let us observe first that we cannot have Bv = (e) for
all v ∈ In since otherwise Ms = LSn which together with the hypothesis that
M surjects onto Ln would imply that M = Ln+1 contradicting properness.
Proposition 2.3 implies then that there is either exactly one pair v1 6= v2 of
orbits for which Bv1 6= (e) and Bv2 6= (e), or there is exactly one orbit v ∈ In
for which Bv 6= (e). We are going to estimate the contribution to ζLn+1 /Ln
according to the four different cases of Proposition 2.3.
11
To this end we make some preliminary observations. Let
D = D1 ∪ · · · ∪ Dℓ
be a partition into L-orbits. Then by hypothesis b) of Theorem 0.1 we have
(3.4)
2 ≤ |Di |,
∀1 ≤ i ≤ ℓ .
The set of Ln -orbits in Sn is given by
In = {Di1 × · · · × Din : (i1 , . . . , in ) ∈ [1, ℓ]n }.
Then there are ℓn orbits of Ln in Sn and for each such orbit O we have
(3.5)
2n ≤ |O| ≤ |D|n .
Case 1: Every pair of orbits O1 6= O2 in Sn leads possibly to a contribution
to ζLn+1 /Ln (k) which is bounded by
C12 · C2 |D|n
,
(22n )k
where C1 is the number of simple quotients of L and C2 an upper bound on
the cardinality of their outer automorphism groups; this takes into account
the estimate (3.5).
Summing over all distinct pairs of orbits, the total contribution of Case 1
is bounded by
n
(ℓ2 |D|)n
|D|n
ℓ
2
2
≤
C
C
.
C1 C2
1 2
(22n )k
2 (22n )k
Case 2: According to Proposition 2.3 2 d) we need to estimate the number
of Ln -block decompositions of an orbit O ∈ In . Observe that a Ln -block
decomposition of Di1 × · · · × Din is obtained by taking the products of L-block
decompositions in each Dij . Therefore, if C3 is an upper bound on the number
of L-block decompositions of Di , 1 ≤ i ≤ ℓ, then the number of Ln -block
decompositions in any orbit O ∈ In is bounded by C3n .
Now fix O ∈ In such that BO 6= (e). Then every normal subgroup N ⊳ B
which is of the form T r for some non-abelian simple T contributes at most
|Out(T )|r|O|(2 |O|2 )r−1 C3r·O · |N|−
to ζLn+1 /Ln (k) and this can be rewritten as:
(2 |O|)r−1 C3r·n
|Out(T )|r |O|
12
|N|k/2
.
|O|
2
k
r
)|
< 1, then one sees easily that the
Now choose k large enough so that |Out(T
|N |k/2
sum over all O ∈ In and n ≥ 1 converges, taking into account that |O| ≥ 2n .
Case 3: The total contribution of this case is easily seen to be bounded by
C 6 · ℓn
,
(22n )k
where C6 bounds the number of subgroups of quotients of L.
Thus the sum over n ≥ 1 of the contributions to ζLn+1 /Ln (k) in the Cases
1, 2 and 3 converge already for k large enough depending on L.
Case 4: We will treat this case by estimating the number of sections of Ln
in Ln ⋉B O . To this end let O = Di1 ×· · ·×Din , we define Oj := Di1 ×· · ·×Dij ,
1 ≤ j ≤ n so that On = O. Let a(j) be the number of sections of Lj in Lj ⋉B Oj .
Let K be an upper bound on the number of non-abelian subgroups of
L ⋉ B Dj and C7 an upper bound on |Hom(L, L ⋉ B Dj )| for any quotient B of
L and any 1 ≤ j ≤ ℓ.
|
K! C7K .
Lemma 3.1. a(n) ≤ a(n − 1)|Din | |Sn−1
K
Proof. Let Ln = Ln−1 ⋉LSn−1 and represent a given section φ: Ln → Ln ⋉B On
by a map of the form
φ : Ln−1 ⋉ LSn−1 −→ (Ln−1 ⋉ LSn−1 ) ⋉ B On
(x, y) −→ (x, y), f (x, y) .
We analyze separately φ|Ln−1 and φ|LSn−1 .
We have φ(x, e) = ((x, e), f (x, e)) where f (x, e) is a map from On = On−1 × Din
to B. Expressing that φ is a homomorphism, we see at once that for every
s ∈ Din , the map
Ln−1 −→ Ln−1 ⋉ B On−1
x −→ x, f (x, e)(·, s)
is a section of Ln−1 in Ln−1 ⋉ B On−1 . The number of possible restrictions to
Ln−1 of a section φ as above is therefore bounded by a(n − 1)|Din | .
Next, we analyze φ|LSn−1
LSn−1 −→ LSn−1 ⋉ B On−1 ×Din
y −→ y, f (e, y) .
13
Observe that LSn−1 ⋉B On−1 ⋉Din is isomorphic to LSn−1 \On−1 ⋉(LOn−1 ⋉B On−1 ×Din )
and thus φ|LSn−1 is determined by
LSn−1 −→ LOn−1 ⋉ B On−1 ⋉Din
y −→ y On−1 , f (e, y) .
Now we observe that we have an isomorphism:
LOn−1 ⋉ B On−1 ×Di −→ (L ⋉ B Din )On−1
(z, g) −→ (z(α), gα )α∈On−1 ,
where gα (s) := g(α, s), s ∈ Din . Thus φ|LSn−1 is determined by a collection of
homomorphisms
φα : LSn−1 −→ L ⋉ B Din ,
parametrized by α ∈ On−1 . But φ|LSn−1 is the restriction of a homomorphism
defined on Ln−1 ⋉LSn−1 and the Ln−1 -action on On−1 is transitive, we conclude
that φα is determined once φα0 is fixed for some α0 . Thus we are left with
estimating the number of homomorphisms
LSn−1 −→ L ⋉ B Din .
Given such a homomorphism ϕ, for every ν ∈ Sn−1 the image of the corresponding L-factor is either trivial or non-abelian (since L is perfect). Since the
images via ϕ of different factors commute, they cannot be sent to the same
non-abelian subgroup of L ⋉ B Din .
Thus we conclude that the number of such homomorphisms is bounded by
|Sn−1 |
· K! C7K .
K
This concludes the proof of the lemma.
|On |
Lemma 3.2. a(n) ≤ C8
K+1
2
, where C8 = C7
|D|K .
Proof. Set |Dij | = dj , the n-tuple (i1 , . . . , in ) being given. In Lemma 3.1
estimate
|Sn−1 |
K! ≤ |D|K(n−1) .
K
Iterating the inequality thus obtained:
a(n) ≤ a(n − 1)dn |D|K(n−1) C7K
we get
a(n) ≤ a(1)d2 ...dn |D|K·αn (C7K )βn ,
14
where
αn = d3 . . . dn + 2d4 . . . dn + · · · + (n − 2) dn + (n − 1)
βn = d3 . . . dn + d4 . . . dn + · · · + dn + 1.
Then
h
3
n−1 i
2
+
+···+
αn = d 3 . . . d n 1 +
d3 d3 d4
d3 . . . dn
h
2
3
n − 1i
≤ d3 . . . dn 1 + + 2 + · · · + n−2
2 2
2
≤ 4d3 . . . dn
and
h
1
1 i
1
+
+···+
βn = d3 . . . dn 1 +
d3 d3 d4
d3 . . . d4
≤ 2d3 . . . dn .
As a result we obtain
a(n) ≤ a(1)1/d1
|On |
|D|K/d1 d2
K/2 |On |
≤ a(1)1/2 |D|K C7
|On |
(1+K)/2
≤ C7
|D|K
.
4|On |
K/d1 d2 2|On |
C7
Every orbit O in In contributes to ζLn+1 /Ln (k) at most
C |O|
C8
8
=
C
,
9
(2|O| )k
2k
|O|
C9 ·
where C9 is an upper bound on the number of quotient groups of L. Now
choose k ≥ 1 such that
C8
< 1.
2k
Then the sum over all orbits in In of the contribution in Case 4 is bounded by
C 2n
8
n
ℓ · k
2
and with this choice of k, the sum over n ≥ 1 converges.
This concludes the proof the theorem.
15
References
[Ama03] Olivier Amann, Groups of tree-automorphisms and their unitary representations,
Thesis, Eidgenoessische Technische Hochschule Zürich (Switzerland) (2003).
[Bha94] Meenaxi Bhattacharjee, The probability of generating certain profinite groups by
two elements, Israel J. Math. 86 (1994), no. 1-3, 311–329. MR1276141
[BM00a] Marc Burger and Shahar Mozes, Groups acting on trees: from local to global structure, Inst. Hautes Études Sci. Publ. Math. 92 (2000), 113–150 (2001). MR1839488
[BM00b]
, Lattices in product of trees, Inst. Hautes Études Sci. Publ. Math. 92
(2000), 151–194 (2001). MR1839489
[BM14]
, Lattices in products of trees and a theorem of H. C. Wang, Bull. Lond.
Math. Soc. 46 (2014), no. 6, 1126–1132. MR3291248
[BMZ09] Marc Burger, Shahar Mozes, and Robert J. Zimmer, Linear representations and
arithmeticity of lattices in products of trees, Essays in geometric group theory,
2009, pp. 1–25. MR2605353
[Bon10] Ievgen V. Bondarenko, Finite generation of iterated wreath products, Arch. Math.
(Basel) 95 (2010), no. 4, 301–308. MR2727305
[DM96] John D. Dixon and Brian Mortimer, Permutation groups, Graduate Texts in Mathematics, vol. 163, Springer-Verlag, New York, 1996. MR1409812
[Man96] Avinoam Mann, Positively finitely generated groups, Forum Math. 8 (1996), no. 4,
429–459. MR1393323
[Moz98] Shahar Mozes, Products of trees, lattices and simple groups, Proceedings of the
International Congress of Mathematicians, Vol. II (Berlin, 1998), 1998, pp. 571–
582.
[MS96] Avinoam Mann and Aner Shalev, Simple groups, maximal subgroups, and probabilistic aspects of profinite groups, Israel J. Math. 96 (1996), no. part B, 449–468.
MR1433701
[Qui04] Martyn Quick, Probabilistic generation of wreath products of non-abelian finite
simple groups, Comm. Algebra 32 (2004), no. 12, 4753–4768. MR2111594
[Qui06]
, Probabilistic generation of wreath products of non-abelian finite simple
groups. II, Internat. J. Algebra Comput. 16 (2006), no. 3, 493–503. MR2241619
[Rat04] Diego A. Rattaggi, Computations in groups acting on a product of trees: Normal
subgroup structures and quaternion lattices., Thesis, Eidgenoessische Technische
Hochschule Zürich (Switzerland) (2004).
[Van15] Matteo Vannacci, Finite generation of iterated wreath products in product action,
Arch. Math. (Basel) 105 (2015), no. 3, 205–214. MR3391479
16
| 4math.GR
|
Structural Attention Neural Networks for improved sentiment analysis
Alexandros Potamianos
Filippos Kokkinos
School of E.C.E.,
School of E.C.E.,
National Technical University of Athens, National Technical University of Athens,
15773 Athens, Greece
15773 Athens, Greece
[email protected]
[email protected]
arXiv:1701.01811v1 [cs.CL] 7 Jan 2017
Abstract
We introduce a tree-structured attention
neural network for sentences and small
phrases and apply it to the problem of sentiment classification. Our model expands
the current recursive models by incorporating structural information around a
node of a syntactic tree using both bottomup and top-down information propagation.
Also, the model utilizes structural attention to identify the most salient representations during the construction of the syntactic tree. To our knowledge, the proposed models achieve state of the art performance on the Stanford Sentiment Treebank dataset.
1
Introduction
Sentiment analysis deals with the assessment
of opinions, speculations, and emotions in text
(Zhang et al., 2012; Pang and Lee, 2008). It is
a relatively recent research area that has attracted
great interest as demonstrated by a series of shared
evaluation tasks, e.g., analysis of tweets (Nakov
et al., 2016). In (Turney and Littman, 2002),
the affective ratings of unknown words were predicted utilizing the affective ratings of a small set
of words (seeds) and the semantic relatedness between the unknown and the seed words. An example of sentence-level analysis was proposed in
(Malandrakis et al., 2013). Other application areas
include the detection of public opinion and prediction of election results (Singhal et al., 2015), correlation of mood states and stock market indices
(Bollen et al., 2011).
Recently, Recurrent Neural Network (RNN)
with Long-Short Term Memory (LSTM) (Hochreiter and Schmidhuber, 1997) or Gated Recurrent Units (GRU) (Chung et al., 2014) have been
applied to various Natural Language Processing
tasks. Tree structured neural networks, which
are found in literature as Recursive Neural Networks, hold a linguistic interest due to their close
relation to syntactic structures of sentences being
able to capture distributed information of structure
such as logical terms(Socher et al., 2012). These
syntactic structures are N-ary trees which represent either the underlying structure of a sentence,
known as constituency trees or the relations between words known as dependency trees.
This paper focuses on sentence-level sentiment
classification of movie reviews using syntactic
parse trees as input for the proposed networks. In
order to solve the task of sentiment analysis of
sentences, we work upon a variant of Recursive
Neural Networks which recursively create representation following the syntactic structure. The
proposed computation model exploits information
from subnodes as well as parent nodes of the node
under examination. This neural network is referred to as Bidirectional Recursive Network (Irsoy and Cardie, 2013). The model is further enhanced with memory units and the proposed structural attention mechanism. It is observed that different nodes of a tree structure hold information
of variable saliency. Not all nodes of a tree are
equally informative, so the proposed model selectively weights the contribution of each node
regarding the sentence level representation using
structural attention model.
We evaluate our approach on the sentence-level
sentiment classification task using one standard
movie review dataset (Socher et al., 2013). Experimental results show that the proposed model
outperforms the state-of-the art methods.
2
Tree-Structured GRUs
2.1
Recursive GRUs (Tree-Gru) upon tree structures
are an extension of the sequential GRUs that allow
information to propagate through network topologies. Similar to Recursive LSTM network on tree
structures (Tai et al., 2015), for every node of a
tree, the Tree-GRU has gating mechanisms that
modulate the flow of information inside the unit
without the need of a separate memory cell. The
activation hj of Tree-GRU for node j is the interpolation of the previous calculated activation hjk
of its kth child out of N total children and the candidate activation e
hj .
hj = zj ∗
N
X
hjk + (1 − zj ) ∗ e
hj
(1)
k=1
where zj is the update function which decide the
degree of update that will occur on the activation
based on the input vector xj and previously calculated representation hjk :
zj = σ(Uz ∗ xj +
N
X
Wzi ∗ hjk )
(2)
k=1
The candidate activation e
hj for a node j is computed similarly to that of a Recursive Neural Network as in (Socher et al., 2011):
e
hj = f (Uh ∗ xj +
N
X
Whk ∗ (hjk ∗ rj ))
(3)
Bidirectional TreeGRU
A natural extension of Tree-Structure GRU is the
addition of a bidirectional approach. TreeGRUs
calculate an activation for node j with the use of
previously computed activations lying lower in
the tree structure. The bidirectional approach for
a tree structure uses information both from under
and lower nodes of the tree for a particular node
j. In this manner, a newly calculated activation
incorporates content from both the children and
the parent of a particular node.
The bidirectional neural network can be trained in
two seperate phases: i) the Upward phase and ii)
the Downward phase. During the Upward phase,
the network topology is similar to the topology
of a TreeGru, every activation is calculated based
on the previously calculated activations which are
found lower on the structure in a bottom up fashion. When every activation has been computed,
from leaves to root, then the root activation is used
as input of the Downward phase. The Downward
phase calculates the activations for every child
of a node using content from the parent in a top
down fashion. The process of computing the
internal representations between the two phases is
separated, so in a first pass the network compute
the upward activation and after this is completed,
then the downward representations are computed.
The upward activation h↑j similarly to TreeGRU
for node j is the interpolation of the previous calculated activation h↑jk of its kth child out
of N total children and the candidate activation e
h↑ .
j
k=1
where rj is the reset gate which allows the network to forget effectively previous computed representations when the value is close to 0 and it is
computed as follows:
rj = σ(Ur ∗ xj +
N
X
Wrk ∗ hjk )
(4)
k=1
Every part of a gated recurrent unit
xj , hj , rj , zj , e
hj ∈ Rd where d is the input
vector dimensionality. σ is the sigmoid function
and f is the non-linear tanh function.The set of
matrices W k , U ∈ Rdxd used in 2 - 4 are the
trainable weight parameters which connect the
kth children node representation with the jth
node representation and the input vector xj .
h↑j = zj↑ ∗
N
X
h↑jk + (1 − zj↑ ) ∗ e
h↑j
(5)
k=1
The update gate, rest gate and candidate activation are computed as follows:
zj↑
= σ(Uz ∗
x↑j
+
N
X
Wzk ∗ h↑jk )
(6)
Wrk ∗ h↑jk )
(7)
k=1
rj↑
= σ(Ur ∗
x↑j
+
N
X
k=1
e
h↑j = f (Uh ∗ x↑j +
N
X
k=1
Wrk ∗ (h↑jk ∗ rj↑ ))
(8)
are used to form the final sentence representation
s ∈ Rd which is a weighted summation of every
node representation hj .
u>
j ∗ uw
aj = PN
>
i=1 ui ∗ uw
s=
N
X
ai hi
(14)
(15)
i=1
Figure 1: A tree-structured bidirectional neural
network with Gated Recurrent Units. The input
vectors x are given to the model in order to generate the phrase representations y ↑ and y ↓ .
The downward activation h↑j for node j is the
interpolation of the previous calculated activation
h↑p(j) , where the function p calculates the index of
the parent node, and the candidate activation e
h↓ .
j
h↓j = zj↓ ∗ h↓p(j) + (1 − zj↓ ) ∗ e
h↓j
(9)
The update gate, reset gate and candidate activation for the downward phase are computed as follows:
zj↓ = σ(Uzd ∗ h↑j + Wzd ∗ h↓p(j) )
(10)
rj↓ = σ(Urd ∗ h↑j + Wrd ∗ h↓p(j) )
(11)
e
h↓j = f (Uhd ∗ h↑j + Whd ∗ (h↓p(j) ∗ rj↓ ))
(12)
During downward phase, matrix U d ∈ Rdxd
connects the upward representation of node j with
the respective jth downward node while W d ∈
Rdxd connect the parent representation p(j).
2.2
Structural Attention
We introduce Structural Attention, a generalization of sequential attention model (Luong et al.,
2015) which extracts informative nodes out of a
syntactic tree and aggregates the representation of
those nodes in order to form the sentence vector. We feed representation hj of node through a
one-layer Multilayer Perceptron with Ww ∈ Rdxd
weight matrix to get the hidden representation uj .
uj = tanh(Ww ∗ hj )
(13)
Using the softmax function, the weights aj for
each node are obtained based on the similarity
of the hidden representation uj and a global context vetor uw ∈ Rd . The normalized weights aj
The proposed attention model is applied on structural content since all node representations contain
syntactic structural information during training because of the recursive nature of the network topology.
3
Experiments
We evaluate the performance of the aforementioned models on the task of sentiment classification of sentences sampled from movie reviews.
We use the Stanford Sentiment Treebank (Socher
et al., 2013) dataset which contains sentiment labels for every syntactically plausible phrase out of
the 8544/1101/2210 train/dev/test sentences. Each
phrase is labeled with respect to a 5-class sentiment value, i.e. very negative, negative, neutral, positive, very positive. The dataset can also
be used for a binary classification subtask by excluding any neutral phrases for the original splits.
The binary classification subtask is evaluated on
6920/872/1821 train/dev/test splits.
3.1
Sentiment Classification
For all of the aforementioned architectures at each
node j we use a softmax classifier to predict the
sentiment label ŷj . For example, the predicted label ŷj corresponds to the sentiment class of the
spanned phrase produced from node j. The classifier for unidirectional TreeGRU architectures uses
the hidden state hj produced from recursive computations till node j using a set xj of input nodes
to predict the label as follows:
p̂θ (y|xj ) = sof tmax(Ws ∗ hj )
(16)
where Ws ∈ Rdxc and c is the number of sentiment classes.
The classifier for bidirectional TreeBiGRU architectures uses both the hidden state h↑j and h↓j
produced from recursive computations till node j
during Upward and Downward Phase using a set
xj of input nodes to predict the label as follows:
p̂θ (y|xj ) = sof tmax(Ws↑ ∗ h↑j + Ws↓ ∗ h↓j ) (17)
where Ws↑ , Ws↓ ∈ Rdxc and c is the number of
sentiment classes. The predicted label ŷj is the
argument with the maximum confidence:
ŷj = argmax(p̂θ (y|xj ))
System
RNN
MV-RNN
RNTN
PVec
TreeLSTM
DRNN
DCNN
CNN-multichannel
DMN
TreeGRU
- without attention
- with attention
TreeBiGRU
- without attention
- with attention
(18)
y
For the Structural Attention models, we use for the
final sentence representation s to predict the sentiment label ŷj where j is the corresponding root
node of a sentence. The cost function used is the
negative log-likelihood of the ground-truth label
y k at each node:
E(θ) =
m
X
p̂θ (y k |xk ) +
k=1
λ
||θ||2
2
(19)
where m is the number of labels in a training sample and λ is the L2 regularization hyperparameter.
Network Variant
TreeGRU
-without attention
-with attention
TreeBiGRU
-without attention
-with attention
d
|θ|
300
300
7323005
7413605
300
300
8135405
8317810
Table 1: Memory dimensions d and total network
parameters |θ| for every network variant evaluated
3.2
Results
The evaluation results are presented in Table 2
in terms of accuracy, for several state-of-the-art
models proposed in the literature as well as for
the TreeGRU and TreeBiGRU models proposed
in this work. Among the approaches reported in
the literature, the highest accuracy is yielded by
DRNN and DMN for the binary scheme (88.6),
and by DMN for the fine-grained scheme (52.1).
We observe that the best performance is achieved
by TreeBiGRU with attention, for both binary
(89.5) and fine-grained (52.4) evaluation metrics,
exceeding any previously reported results. In addition, the attentional mechanism employed in the
proposed TreeGRU and TreeBiGRU models improve the performance for both evaluation metrics.
4
Hyperparameters and Training Details
The evaluated models are trained using the AdaGrad (Duchi et al., 2010) algorithm using 0.01
learning rate and a minibatch of size 25 sentences.
L2-regularization is performed on the model parameters with a λ value 10−4 . We use dropout
Binary
82.4
82.9
85.4
87.8
88.0
86.6
86.8
88.1
88.6
Fine-grained
43.2
44.4
45.7
48.7
51.0
49.8
48.5
47.4
52.1
88.6
89.0
50.5
51.0
88.5
89.5
51.3
52.4
Table 2: Test Accuracies achieved on the Stanford
Sentiment Treebank dataset. RNN, MV-RNN and
RNTN (Socher et al., 2013). PVec: (Mikolov et
al., 2013). TreeLSTM (Tai et al., 2015). DRNN
(Irsoy and Cardie, 2013). DCNN (Kalchbrenner et
al., 2014).CNN-multichannel (Kim, 2014). DMN
(Kumar et al., 2015)
with probability 0.5 on both the input layer and
the softmax layer.
The word embeddings are initialized using the
public available Glove vectors with a 300 dimensionality. The Glove vectors provide 95.5% coverage for the SST dataset. All initialized word vectors are finetuned during the training process along
with every other parameter. Every matrix is initialized with the identity matrix multiplied by 0.5
except for the matrices of the softmax layer and
the attention layer which are randomly initialized
from the normal Gaussian distribution. Every bias
vectors is initialized with zeros.
The training process lasts for 40 epochs. During
training, we evaluate the network 4 times every
epoch and keep the parameters which give the best
root accuracy on the development dataset.
5
Conclusion
In this short paper, we propose an extension of
Recursive Neural Networks that incorporates a
bidirectional approach with gated memory units
as well as an attention model on structure level.
The proposed models were evaluated on both finegrained and binary sentiment classification tasks
on a sentence level. Our results indicate that both
the direction of the computation and the attention
on a structural level can enhance the performance
of neural networks on a sentiment analysis task.
References
Johan Bollen, Huina Mao, and Xiaojun Zeng. 2011.
Twitter mood predicts the stock market. Journal of
Computational Science, 2(1):1–8.
Junyoung Chung, Çaglar Gülçehre, KyungHyun Cho,
and Yoshua Bengio. 2014. Empirical evaluation of
gated recurrent neural networks on sequence modeling. CoRR, abs/1412.3555.
John Duchi, Elad Hazan, and Yoram Singer. 2010.
Adaptive subgradient methods for online learning
and stochastic optimization. Technical report, EECS
Department, University of California, Berkeley.
Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long
short-term memory. Neural Comput., 9(8):1735–
1780.
Ozan Irsoy and Claire Cardie. 2013. Bidirectional recursive neural networks for token-level labeling with
structure. CoRR, abs/1312.0493.
Nal Kalchbrenner, Edward Grefenstette, and Phil Blunsom. 2014. A convolutional neural network for
modelling sentences. CoRR, abs/1404.2188.
Yoon Kim. 2014. Convolutional neural networks for
sentence classification. CoRR, abs/1408.5882.
Ankit Kumar, Ozan Irsoy, Jonathan Su, James Bradbury, Robert English, Brian Pierce, Peter Ondruska,
Ishaan Gulrajani, and Richard Socher. 2015. Ask
me anything: Dynamic memory networks for natural language processing. CoRR, abs/1506.07285.
Minh-Thang Luong, Hieu Pham, and Christopher D.
Manning.
2015.
Effective approaches to
attention-based neural machine translation. CoRR,
abs/1508.04025.
Nikolaos Malandrakis, Abe Kazemzadeh, Alexandros
Potamianos, and Shrikanth Narayanan. 2013. Sail:
A hybrid approach to sentiment analysis. In Proceedings SemEval, pages 438–442.
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing
systems, pages 3111–3119.
Preslav Nakov, Alan Ritter, Sara Rosenthal, Fabrizio
Sebastiani, and Veselin Stoyanov. 2016. Semeval2016 task 4: Sentiment analysis in twitter. In Proceedings of the 10th international workshop on semantic evaluation (SemEval 2016).
Bo Pang and Lillian Lee. 2008. Opinion mining and
sentiment analysis. Found. Trends Inf. Retr., 2(12):1–135.
Kartik Singhal, Basant Agrawal, and Namita Mittal.
2015. Modeling indian general elections: sentiment
analysis of political twitter data. In Information
Systems Design and Intelligent Applications, pages
469–477.
Richard Socher, Jeffrey Pennington, Eric H. Huang,
Andrew Y. Ng, and Christopher D. Manning. 2011.
Semi-supervised recursive autoencoders for predicting sentiment distributions. In Proceedings of the
Conference on Empirical Methods in Natural Language Processing, EMNLP ’11, pages 151–161,
Stroudsburg, PA, USA. Association for Computational Linguistics.
Richard Socher, Brody Huval, Christopher D. Manning, and Andrew Y. Ng. 2012. Semantic compositionality through recursive matrix-vector spaces.
In Proceedings of the 2012 Joint Conference on
Empirical Methods in Natural Language Processing and Computational Natural Language Learning, EMNLP-CoNLL ’12, pages 1201–1211. Association for Computational Linguistics.
Richard Socher, Alex Perelygin, Jean Y Wu, Jason
Chuang, Christopher D Manning, Andrew Y Ng,
and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment
treebank. In Proceedings of the conference on
empirical methods in natural language processing
(EMNLP).
Kai Sheng Tai, Richard Socher, and Christopher D.
Manning. 2015. Improved semantic representations from tree-structured long short-term memory
networks. CoRR, abs/1503.00075.
Peter Turney and Michael L Littman. 2002. Unsupervised learning of semantic orientation from a
hundred-billion-word corpus.
Zhu Zhang, Xin Li, and Yubo Chen. 2012. Deciphering word-of-mouth in social media: Text-based metrics of consumer reviews. ACM Trans. Manage. Inf.
Syst., 3(1):5:1–5:23.
| 9cs.NE
|
Noname manuscript No.
(will be inserted by the editor)
Identifying Hazardousness of Sewer-Pipeline
Gas-Mixture using Classification Methods
A Comparative Study
arXiv:1707.00561v1 [cs.NE] 16 May 2017
Varun Kumar Ojha · Parmartha Dutta ·
Atal Chaudhuri
Received: date / Accepted: date
Abstract In this work, we formulated a real-world problem related to sewerpipeline gas detection using the classification-based approaches. The primary
goal of this work was to identify the hazardousness of sewer-pipeline to offer
safe and non-hazardous access to sewer-pipeline workers so that the human
fatalities, which occurs due to the toxic exposure of sewer gas components,
can be avoided. The dataset acquired through laboratory tests, experiments,
and various literature-sources were organized to design a predictive model
that was able to identify/classify hazardous and non-hazardous situation of
sewer-pipeline. To design such prediction model, several classification algorithms were used and their performances were evaluated and compared, both
empirically and statistically, over the collected dataset. In addition, the performances of several ensemble methods were analyzed to understand the extent
of improvement offered by these methods. The result of this comprehensive
study showed that the instance-based-learning algorithm performed better
than many other algorithms such as multi-layer perceptron, radial basis function network, support vector machine, reduced pruning tree, etc. Similarly, it
was observed that multi-scheme ensemble approach enhanced the performance
of base predictors.
V. K. Ojha
IT4Innovations, VŠB Technical University of Ostrava, Ostrava, Czech Republic and Dept.
of Computer Science & Engineering, Jadavpur University, Kolkata, India
E-mail: [email protected]
P. Dutta
Dept. of Computer & System Sciences, Visva-Bharati University, India
E-mail: [email protected]
A Chaudhuri
Dept. of Computer Science & Engineering, Jadavpur University, Kolkata, India E-mail:
[email protected]
Neural Computing and Applications
DOI: 10.1007/s00521-016-2443-0
2
Varun Kumar Ojha et al.
Keywords Sewer gas detection · Neural network · Classification · KS test
1 Introduction
This is in the view of providing a solution to a real-world problem using technology, where the human fatalities need to be avoided. Hence, the technology
should be as simple as possible. In this work, we addressed a complex realworld problem related to sewer-pipeline gas detection, where sewer-pipeline
safety detection (in terms of non-toxic environment) was required to allow
maintenance and cleaning of the pipeline. The sewer gas detection is a highly
complex problem because of the presence of several toxic gases in a mixture
form, and a single gas detector may not offer reliable solution. Therefore, we
studied the complexity of this problem in terms of gas mixture. The primary
goal was to offer a simple solution with a high accuracy so that it was easy to
categorize the hazardous situation in straightforward way such as “hazardous”
or “non-hazardous.” To meet this simplicity, we formulated sewer-pipeline gas
detection problem as a classification problem.
Sewer-pipeline contains a mixture of several toxic gases such as hydrogen sulphide (H2 S), ammonia (NH3 ), methane (CH4 ), carbon dioxide (CO2 ),
nitrogen oxides (NOx ), etc., [1, 2, 3]. Usually, this mixture is generated due
to the biodegradation of the waste and the sewage into the sewer-pipeline.
Such toxic gas-mixture is fatal for those who come to the proximity/exposure
of these gases. Following this, an alarming number of human fatalities are
reported each year by the newspapers and the other agencies [4, 5,6]. The
authorities those are responsible for maintaining and cleaning of the sewer
pipeline provides various electronic portable gas detectors available in the market to the employed persons so that they can determine the safeness of the
sewer-environment before physically get involve into the maintenance work.
However, the available electronic portable gas detectors are not providing satisfactory results. It is evident from the recent comments from the judiciary to
these authorities. In a judgment to a civil appeal number 5322 of 2011, the
Supreme Court of India stated, “the State and its agencies/instrumentalities
cannot absolve themselves of the responsibility to put in place effective mechanism for ensuring safety of the workers employed for maintaining and cleaning
the sewage system [7].” Similarly, in another judgment, the Supreme Court of
India stated, “...entering sewer lines without safety gears should be made a
crime even in emergency situations... [8, 9].” This motivated us to carry out
our research in this domain and to come out with a simple solution so that
without having the minimum knowledge of the technicalities of gas composition and safety limits, a person is able to understand the environment of a
sewer system before entering.
To ensure the simplicity in model, we collected and preprocessed data to
realize sewer gas-detection as a binary class classification problem. However, in
this work, apart from the objective of constructing a prediction model, we set
a secondary objective, which was to analyze the performances of the classifiers,
Identifying Hazardousness of Sewer-Pipeline
3
both empirically and statistically. To meet these objectives, we used 12 base
predictors from four different categories such as neural network based classifiers, tree based classifiers, instance based classifiers, and rule based classifiers.
The algorithms were applied over the collected dataset and the performance
of the algorithms were collected in terms of the accuracy. The collected results
were then used for analyzing the performance superiority of the one algorithm
over another or the one category of algorithms over another.
We observed that the performance of the algorithms were independent of
the category they belong to. For example, the performance of instance based
k-nearest neighbor, logistic model tree, and support vector machine came from
three different categories, but they had a very competitive performance. However, we must consider the “No-free-lunch theorem” that suggests that some
algorithms perform better on some problem and some on another [10]. Therefore, to find out which predictor performs best in this case, we used 12 base
predictors and nine ensemble methods.
Rest of the article is organized as follows. A background study is provided
in Section 2.1, which leads to setting ground for describing our contribution to
the sewer-pipeline gas detection. In Section 2.2, we provide a detailed description of the data collection and preprocessing mechanisms, which constitute
the core and significant part for formulating gas detection problem as a binary classification problem. Section 2.3 deals with the brief descriptions of
the classifiers/algorithms and methods used for constructing the prediction
model. The design of comprehensive experiment set for the evaluation of the
classifiers is reported in Section 3. Whereas, Section 3 describes empirical and
statistical evaluation of the classifiers, discussions and conclusion are reported
in Sections 4 and 5, respectively.
2 Methodology
In this Section, we put together the background study, the data collection
mechanisms, and the classification methods definitions. The background study
describes the significance of the sewer-pipeline gas detection problem and the
data collection mechanism describes the formulation of gas-detection as a classification problem.
2.1 Background Study
Literature review was conducted in the perspective of electronic-nose (ENOSE) and gas-detection-system to cover a broad area of research in the field
of gas detection and modeling using intelligent computing techniques/algorithms.
Although not much work specifically on sewer gas-mixture-detection was reported in the past, few notable contributions were observed. Li et al. [11]
reported a noticeable research work on the development and design of an
electronic nose (E-NOSE) and gas detection system, where a neural network
4
Varun Kumar Ojha et al.
(NN)-based mixed gas (NOx, and CO) measurement system was developed.
On the other hand, Sirvastava et al. [12, 13] proposed a design of intelligent ENOSE system using backpropagation (BP) and neuro-genetic approach. Llobet
et al. [14] presented a pattern recognition approach, based on the wallet transformation for gas mixture analysis using single tin-oxide sensor. Liu et al. [13]
addressed a genetic-NN algorithm to recognize patterns of mixed gases (a
mixture of three component gases) using infrared gas sensor. Lee et al. [15]
illustrated uses of micro gas sensor array (GSA) combined with NN for recognizing combustible leakage gases. Ambard et al. [16] have demonstrated use
of NN for gas discrimination using a tin-oxide GSA for the gases H2 , CO and
CH4 . In [17], authors have illustrated a NN-based technique for developing a
gas sensory system for sensing gases in a dynamic environment. Pan et al. [18]
have shown several applications of E-NOSE. Wongchoosuka et al. [19] have
proposed an E-NOSE detection system based on carbon nanotube-SnO2 gas
sensors for detecting methanol. Zhang et al. [20] developed a knowledge-based
genetic algorithm for detecting mixed gas in mines. Won et al. [21] proposed
a system for estimation of hazardous gas release rate using optical sensor and
NN-based technique. The following salient points came out of the above mentioned articles:
– Mainly, BP and NN-based approaches were studied so far for detecting
gas-mixtures.
– Mostly, the E-Nose systems reported in the past were developed for the
gas-mixtures of only two or three gases and the sensors of the gases used
were less cross-sensitive to the other gases in mixtures.
– Cross-sensitivity during sensing is an important factor in gas detection system, which was least reported in literature as yet. However, Ojha et al. [22,
23, 24, 25,26, 27] offered a few methods such as neuro-genetic, neuro-swarm,
ant-colony-based, neuro-simulated annealing, etc., where cross-sensitivity
factor has been addressed to some extent. However, these works were primarily related to regression modeling.
– The impact of humidity and temperature on sensors remained ignored so
far.
– The gas detection system or E-Nose was viewed only in the framework of
regression problems and not classification problem.
Classification based approach led us to determine the hazardous and nonhazardous situation of a sewer-pipeline. In addition, the collection, organization, and the preprocessing of the collected data enabled us to address
the cross-sensitivity issue firmly. The cross-sensitivity issue occurs because
of the sensitivity of one gas-sensor towards multiple gases. So was our case,
where a semiconductor-based GSA was designed using five gas-sensors. Each
gas-sensor was typically meant for detecting its respective target gas. Hence,
when the GSA was used for collecting data for a mixture of gases, the crosssensitivity in the sensed values (collected data) became inevitable. Therefore,
rather than considering pure results of the respective gases, we registered the
cross-sensitive results as a part of our-collected data. Since a computation-
Identifying Hazardousness of Sewer-Pipeline
5
ally intelligent model learned from the data and also maintained the crosssensitivity patterns registered in terms of data values itself, a learned model
accurately predicts an unknown gas mixture.
2.2 Equipment and Data Collection Mechanism
(A/trained/Classifier/embedded/into/
electronic/(EEPROM)./
(Buzzer/light)/
Intelligent/unit./
Output/Unit
(Figure/2)
Gas/Sensor/
Array/
Before explaining the details of data collection and equipment, we need to
explain the basic design and the purpose of our work, which is to offer an
intelligent gas detection system (an electronic portable gas detector) that will
be a result of embedding learned-predictor (trained-classifier) into an electronic system. The data flow into our developed intelligent system is shown in
Fig. 1, which describes the entire process of the intelligent system design, which
is divided into three phases: 1) The data acquisition unit, which consists of
gas suction-motor chamber, GSA, and data acquisition-cum data-preprocessor
block; 2) An intelligent unit (classifier unit), which receives data from dataacquisition unit and classifying the acquired data patterns; 3) The output unit,
which prompts the result in terms of colored light and buzzer. Hence, our objective here was limited to only train a classifier using the collected data. We
describe the data collection process as follows.
Fig. 1 Block diagram of intelligent system design (real time data flow process)
At first, we collected the data samples from the data-sheets, literature, and
laboratories test of the collected gas mixture samples from sewer-pipelines.
Second, we designed our own metal oxide semiconductor (MOS) gas sensors
array (GSA) that was used for verifying the literature and laboratory data and
for generating the data samples for the purpose experiments. Our designed
GSA consists of five gas-sensors for sensing five different gases. They include
hydrogen sulphide (H2 S), ammonia (NH3 ), methane (CH4 ), carbon dioxide
(CO2 ), and nitrogen oxides (NOx ). Typically, MOS sensors are resistance-type
electrical sensors, where responses are change in circuit resistance proportional
to gas concentration, A resistance type sensor responds to change in resistance
due to change in the concentration of gases. The change in resistance is given
as δRs /R0 , where δRs is change in MOS sensor resistance and R0 is base
resistance or the sensing resistance at a specifics gas concentration in clean
air [19]. The R0 of the sensors MiCS - 4514, MQ - 7, MQ - 136, MQ - 135, and
MQ - 4 is 0.25 ppm, 100 ppm, 10 ppm, 100 ppm and 1000 ppm, respectively.
Here, ppm is the unit for measuring concentration of gas into air which is
defined as follows: 1 ppm is equal to 1 volume of a gas into 106 volume of air.
6
Varun Kumar Ojha et al.
A typical arrangement of a gas sensor array is shown in Fig. 2. The circuitry
shown in Fig. 2 (left) was developed in our laboratory. Here, the fabricated
and installed sensors were MiCS - 4514, MQ - 7, MQ - 136, MQ - 135, and
MQ - 4 for gases NO2 , CO, H2 S, NH3 , and CH4 , respectively [28, 29].
The gas sensors used were sensitive to not only their target gases, but they
were sensitive also to other gases in the gas-mixture [30, 31]. Hence, crosssensitivity effect over MOS sensors was confirmed [32]. It was moreover confirmed that the sensor responses were noisy and accordingly the pattern of such
noise were considered and recorded as an instance into our dataset. Hence, a
non-intelligent use of raw values of sensor response for hazardousness prediction may be misleading in operating (real-world) environment. Therefore, a
training electronic portable gas detector may be used to predict sewer hazardousness, accurately. So was the effort in this work to provide a classifier.
Data collection had vital role in training of a classifier. Data samples were
collected as per the following steps. At first, several manhole samples collected
from the Kolkata, India municipal area were tested in laboratory to identify
the presence of several toxic gases such as nitrogen dioxide (NO2 ), carbon
monoxide (CO), hydrogen sulphide (H2 S), ammonia (NH3 ), methane (CH4 ),
and carbon dioxide (CO2 ). Secondly, gas sensors were identified for each of the
respective gases. As a result we came out with the procurement of gas sensor
MiCS - 4514, MQ - 7, MQ - 136, MQ - 135, and MQ - 4 for NO2 , CO, H2 S,
NH3 , and CH4 , respectively. We collected data sheets form the companies
for the respective sensors. In the third step, a laboratory was setup for the
verification and collection of the sensor response of the respective gas sensors
in certain range of their concentration. Specifically, the concentration range in
ppm laid down in sensor manuals of sensors MiCS - 4514, MQ - 7, MQ - 136,
MQ - 135, and MQ - 4 are [0.25 - 5], [20 - 1000], [1 - 100], [10 - 300] and [300
- 10000] of the gases NO2 , CO, H2 S, NH3 , and CH4 , respectively. In addition,
the lab was setup (see Fig. 2 [right]), where gas cylinders were connected to
a gas concentration measuring unit called mass flow controller (MFC), which
was further connected to a gas chamber, where each gas was allowed to pass
in a specific concentration over an array of gas sensor. More specifically, the
behavior of each of the gas sensors was recorded.
Fig. 2 Laboratory-scale gas sensor array (GSA) [28, 29]
Identifying Hazardousness of Sewer-Pipeline
7
The following steps were used for preparing data sample for the classifiers’
training. First, hazardous (safety) limits of the component gases of manhole
gas mixture were collected. Secondly, three different levels, (i) above safetylimit, (ii) at safety-limit, and (iii) below safety-limit for each manhole gas were
recognized. Thirdly, gases were mixed in different combination to prepare several mixture sample that were used to pass over GSA. Table 1 indicates few
examples of such mixture of gases in different combinations. For example, when
we mix five gases each of which has three different recognized concentration
levels, we get 243 different combinations (35 ). In addition, we considered the
role of humidity and temperature to influence the sensor’s behavior. Accordingly, the data values were recorded. Hence, our collected dataset contained
seven input features and an output class. Each sample was labeled with “0” for
safe sample (if the responses of all five sensors were under the maximum safety
limit) or “1” for unsafe sample (if the responses of any among the five sensors were above the maximum safety limit). The safety limits of the manhole
gases are as follows: safety limit of NH3 is between 25 ppm and 40 ppm [33],
CO is in between 35 ppm and 100 ppm [34], H2 S is in between 50 ppm and
100 ppm [35], CO2 is in between 5000 ppm and 8000 ppm [36] and CH4 is in
between 5000 ppm and 10000 ppm [37]. Table 2 illustrates a fraction of the
collected data samples.
2.3 Classification Based Approach
We categorized the classifiers in the four different groups of classifiers. Each
category of classifiers contains three classifiers.
Table 1 Samples of gas-mixture in different concentration.
Concentration of gases in ppm
#
Humidity
Temperature
NO2
CO
H2S
NH
CH
Class
Status
1
2
3
:
7535
7536
7537
:
16036
16037
16038
65
65
65
20
20
20
0
0
0
10
10
10
10
10
10
20
20
20
2000
5000
10000
0
0
1
safe
safe
unsafe
65
65
65
30
30
30
0
0
0
10
10
10
10
10
10
20
20
20
2000
5000
10000
0
1
0
safe
unsafe
safe
75
75
75
50
50
50
20
20
20
50
50
50
50
50
50
50
100
100
10000
2000
5000
1
1
1
unsafe
unsafe
unsafe
8
Varun Kumar Ojha et al.
2.3.1 Network Based Classifiers
Multi-layer perceptron (MLP) is a computational model that imitates human
brain, and learn from environment, i.e., data. In our work, we used threelayered MLP, where layers are input layer, hidden layer, and output layer [38].
Radial Basis Function Network (RBF) is a special class of MLP, where inputs
are mapped onto a hidden layer that consists of radial basis function, which
does the non-linear mapping of input to a hidden layer [39].
Support vector machine (SVM) is a supervised learning computational model
that maps input to a high dimension feature space using kernel trick. Hence,
non-linear separable patterns in input space are linearly classified on a high
dimensional feature space [40].
2.3.2 Tree Based Classifiers
Reduced pruning tree (REP) is a tree based classifier method, where a treelike structure is designed for predicting target class based on the input variables [41, 42]. More specifically, the leaves of tree offers decision of the class
based on the conjunction of the input feature represented by the branches of
the tree. REP tree is a decision tree, where the tree size is reduced by pruning
inefficient branches [43].
Naive Bayes tree (NBT) is a special class of decision tree, where the leaf nodes
of decision tree that offer decision on the class is replaced by a Naive Bayes
classifier, which decides the class label, based on the features and learned
threshold [44].
Table 2 Samples of calibrated sensor responses based on the knowledge gathered from
literature, data-sheets, lab tests, and scaling process.
Sensors response (δRs /R0 )
#
Humidity
Temperature
InNO2
InCO
InH2S
InNH
InCH
Class
1
2
3
:
7535
7536
7537
:
16036
16037
16038
65
65
65
20
20
20
0.813
1.301
1.035
6.929
7.521
6.658
5.938
5.525
5.841
3.433
3.521
3.633
3.985
2.178
1.620
0
0
1
65
65
65
30
30
30
1.038
1.054
0.642
7.565
6.694
7.210
5.658
5.745
5.819
3.228
3.692
1.326
2.275
1.268
3.530
0
1
0
75
75
75
50
50
50
4.645
4.712
4.911
2.764
2.985
2.433
0.608
0.641
0.381
2.709
1.228
0.937
0.499
0.450
0.481
1
1
1
Identifying Hazardousness of Sewer-Pipeline
9
Logistic Model Trees (LMT) is similar to NBT that does the transformation
of leaves of a decision tree into a logistic regression node. A logistic regression maps independent variables to categorical dependent variables using a
logistic function [45, 46]. Hence, LMT is a simple idea, where nodes of a decision/classification tree are replaced by logistic regression model [47].
2.3.3 Rule Based Classifiers
Decision Table (DT) is a simple representation of data into a table based system, where the decision is made based on the features matching or searched
into a decision table. On a successful search, the majority class label is returned, otherwise the majority class label of the entire dataset is returned as
a decision for an unlabeled data [48].
PART is a rule based classification method based on partial decision tree that
generates a list of rules, used subsequently for making prediction of unknown
data instance. The rules are generated based on the partial decision tree, which
splits dataset into subsets until the entire dataset gets exhausted to form nodes
and leaf nodes of the tree [49].
Majority Predictor (Zero R) is the simplest possible form of classification
method. It is based on the majority of class label into a dataset. In simple
words, it always predicts the majority class.
2.3.4 Instance Based Classifiers
Instance-Based Learning (IBK) provides the concept description which is the
primary output of an IBK algorithm. It is a function that maps an instance to
a category (class label). The concept description function is updated based on
training procedure that involves two functions similarity and classification. The
similarity function computes the similarity between the training instances and
the pre-stored instances, and returns a numeric-value. Then, the classification
function provides class label to the instances based on the results of similarity
function. Accordingly, the concept description is updated [50].
K∗ (K Star) is an instance-based learner that uses an entropy-based similarity matching function for searching/matching test instances to the learned
instances [51].
Locally Weighted Learning (LWL). In a locally weighted learning, the prediction models are allowed to create at local points in a dataset or the specific
point of interest rather than creating model for entire dataset. Hence, a linear
regression or naive Bayes classifier or any other classifier may be used to create local models. In this case, we use Decision Stamp, which is a single level
decision tree model for prediction [52, 53].
10
Varun Kumar Ojha et al.
2.3.5 Ensemble Methods
In this work, we tried to exploit different method of making ensemble. For an
ensemble to perform well, we need to take into account two things which are
accuracy of predictors and diversity among the predictors [54]. For example,
Bagging maintains diversity by bootstrapping dataset, AddBoost combines
several weak predictors, Random Subspace maintains diversity by splitting
feature space, Random committee maintains diversity by creating predictors
using different random seeds, and Rotation forest maintains diversity by splitting and extracting feature subspace using principal component analysis. Similarly, in multi-scheme and voting scheme, we combine several predictors to
maintain diversity. Here, we describe the ensemble methods as follows.
Bagging. In Bagging, several copies of same predictor is created. Each copy
of the predictor learns a different replicate of learning set created from the
complete training set using bootstrapping. Finally, the predictor’s decision is
combined using plurality voting method [55].
Adaptive Boosting (AdaBoost) is an ensemble technique that combines several
weak predictors and inaccurate rules to create an accurate predictor [56].
Random Subspace (Random SUB). In random subspace ensemble method feature space is divided into several feature subset. Hence, predictors are constructed for each feature subset. Finally, the decision of each constructed predictors are combined using voting method [57]. Random Committee (Random
COM): In a random committee ensemble, several predictors are constructed
over similar dataset, but they use different random seeds to maintain diversity
in the ensemble.
Rotation Forest (Rotation FRST). In this approach, training set for the predictors are created by splitting feature set into K subsets, and Principal Component Analysis is applied to extract all the principle components [58]. Hence,
diversity among the predictors are maintained by K axis rotation to form new
feature set for training [58].
Ensemble Selection (Ensemble SEL). In the ensemble selection approach, the
ensemble starts with an empty bag, and the predictors (chosen from a library of trained predictors) maximizing the performance of ensemble are added
to the bag one by one to compute the decision of ensemble by using voting
method [59].
Voting Scheme (Vote). The voting scheme combines probability distribution
of several chosen predictors/classifiers (or predictors available in a bag for
making ensemble) using majority voting combination method [60].
Identifying Hazardousness of Sewer-Pipeline
11
Multi-Scheme (Multi). The multi-scheme ensemble approach uses a bag of
predictors and selects the output class by selecting a predictor from the bag
of predictors based on cross-validation performance of the predictors [60].
Weighted Predictor Ensemble (WPE). In this scheme of ensemble, the weight
of predictors were determined. Subsequently, the ensemble output of k many
predictors were computed as follows:
c
y = arg max
j=1
k
X
wj I (Pj = ωj ) ,
j=1
where c is the number of classes (here it is two), I (Pj = ωj ) is a function that
returns value one for the predicted class ωj .
3 Experimental Framework and Results
Our aim in the experiment design was to obtain a highly accurate model for
predicting hazardousness of the environment in a sewer pipeline. The sewerpipeline environment was represented by the collected dataset. The second
objective of the experiment design was to obtain results for analyzing the
classifiers (predictors). Accordingly, the results of the classifiers were collected.
Table 3 represents the parameter setting of the chosen classifiers. For the
evaluation of the classifiers, we repeated our experiments 10 times. Finally,
the results were compared based on empirical and statistical (Kolmogorov–
Smirnov test) evaluation. We used WEKA [61] and MATLAB tools [62] for
the purpose of our experiments.
We organized the experimental results into three parts as reflected in Table 4. The first part in the table describes the category wise performance of
classifier. Hence, the performance of the category of classifiers was evaluated.
We represented the performance of the classifiers as per their training and
test accuracy. An accuracy close to 1.0 indicates 100% classification accuracy.
Accordingly, the standard deviation (std) of training and test accuracies were
reported for understanding the consistency of the classifiers’ performance. In
Table 4, the performance of the classifiers were arranged as follows. The category is arranged in the ascending order of their average accuracy over 10-fold
CV test set, i.e., better performing classifier to the less performing classifier.
The dataset was portioned into 10 equal sets and each time 9 sets were used
for training and one set for testing. This process was repeated 10 times and
each time a unique test set was used.
In the second part, we organized the results according to rank of the classifiers’ performance over 10-fold test set. It may please be noted that for each
classifier, we collected 10 instances of 10-fold CV training and test results.
Hence, the results in Table 5 reflect averaged training and test accuracy of
the classifiers. However, ranking the classifiers based only on the average results does not say much about the quality of the classifier. Hence, in the
12
Varun Kumar Ojha et al.
Table 3 Parameter setting of different classifiers
category
Network-Based
Classifiers (F1)
Tree-Based
Classifiers (F2)
Instance-Based
Classifiers (F3)
Rule-Based
Classifiers (F4)
Ensemble
Classifiers (E1)
Ensemble
Classifiers (E2)
Classifiers
Parameters
MLP
Learning rate: 0.3, momentum factor: 0.2, iteration: 500, nodes in hidden layer: 100
Kernel: Gaussian basis function.
Kernel: Radial basis function
Minimum no. of instance per leaf: 2, split proportion: 0.001
Leaf node: nave Bayes classifier.
Node: logistic function, Number of instance per
node for splitting: 15
Similarity function: linear nearest neighbor search,
neighbor size: 1
Similarity function: entropy distance measure.
Similarity function: linear nearest neighbor search,
Weight function: Linear, Classifier: Decision
Stamp.
Evaluation metric: accuracy, Search method: best
first
Confidence threshold for pruning: 0.25
Ensemble size: 10. Classifier: REP Tree
Ensemble size: 10. Classifier: Decision Stamp.
Ensemble size: 10. Classifier: REP Tree
Ensemble size: 10. Classifier: Random Tree
Ensemble size: 10. Classifier: Random Tree
Ensemble size: 10. Classifier: REP Tree
Ensemble size: 12. Classifiers: F1, F2, F3 and F4
Ensemble size: 12. Classifiers: F1, F2, F3 and F4
Ensemble size: 12. Classifiers: F1, F2, F3 and F4
RBF
SVM
REP
NBT
LMT
IBK
K Star
LWL
DT
PART
Zero R
Bagging
AdaBoost
Random SEL
Random COM
Rotation FRST
Ensemble SEL
Vote
Multi Scheme
WPE
third part of the results, we used pairwise comparison of the classifiers using
Kolmogorov–Smirnov (KS) test, which ascertains whether the supremacy of
one classifier over the other is statistically significant or not. A comprehensive
matrix of the pairwise KS test results are presented in Table 6. The KS Test
is a non-parametric statistical test that determines the difference between the
cumulative frequency distribution (cfd) of two samples. In other words, it indicates whether the empirical cfd of one sample is equal “=”, larger “ ”, or
smaller “≺” than the other. It tells whether two dataset A and B are statistically similar “A=B”, dissimilar “A≺B”, where A being statistically dominated
by B, or dissimilar “A B”, where A being statistically dominant over B. In
our experiments, the KS test was evaluated with 5% significance level, i.e.,
with 95% confidence.
4 Discussions
Since the developed electronic portable gas detector shall be used by naive
persons who are engaged in maintaining sewer-pipeline, we are looking for
binary answer. Hence, our objective is to search for classification accuracy and
Identifying Hazardousness of Sewer-Pipeline
13
Table 4 Experimental Results of Classifiers over 10 fold cross validation error
category
NN-Based
Classifiers
Tree-Based
Classifiers
Instance-Based
Classifiers
Rule-Based
Classifiers
Ensemble
Classifiers
Classifiers
Training
avg. accuracy
std
Test
avg. accuracy
std
SVM
MLP
RBF
LMT
REP Tree
NB Tree
IBK
K Star
LWL
PART
Decision Table
Zero R
Multi
Rotation FRST
Random COM
Bagging
WPE
Ensemble SEL
Vote
Random SUB
AdaBoostM1
0.9407
0.8681
0.8064
0.9697
0.9528
0.9064
1.0000
0.9997
0.7613
0.9275
0.8672
0.7613
1.0000
1.0000
1.0000
0.9728
0.9635
0.9577
0.9423
0.9160
0.7613
0.9340
0.8664
0.8051
0.9360
0.9265
0.8898
0.9671
0.9638
0.7613
0.9062
0.8553
0.7613
0.9672
0.9622
0.9549
0.9395
0.9356
0.9330
0.9214
0.8720
0.7613
0.0041
0.0081
0.0187
0.0023
0.0067
0.0418
0.0023
0.0036
0.0128
0.0103
0.0154
0.0091
0.0035
0.0036
0.0073
0.0077
0.0056
0.0077
0.0143
0.0165
0.0091
0.0008
0.0029
0.0090
0.0039
0.0025
0.0469
0.0000
0.0001
0.0014
0.0097
0.0049
0.0010
0.0000
0.0000
0.0000
0.0008
0.0006
0.0009
0.0128
0.0103
0.0010
Table 5 Ranking algorithms according to their performance on test set (10 Fold CV).
Rank
category
Classifiers
Training
Test
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
E2
F3
F3
E1
E1
E1
E2
F2
E1
F1
F2
E2
F4
F2
F1
E1
F4
F1
F3
F4
E1
Multi
IBK
KStar
Rotation FRST
Random COM
Bagging
WPE
LMT
Ensemble SEL
SVM
REPTree
Vote
PART
NBTree
MLP
Random SUB
DT
RBF
LWL
ZeroR
AdaBoost
100.0000
100.0000
99.9653
100.0000
100.0000
97.2874
96.3564
96.7454
95.7173
94.0837
95.2469
93.9593
92.7331
89.4770
86.8566
90.9650
86.6874
80.7795
76.1285
76.1285
76.1302
96.8060
96.7945
96.4677
96.2725
95.7737
94.1025
93.9865
93.4674
93.3778
93.3647
92.4733
92.1314
90.8675
88.0795
86.6336
86.4397
85.4790
80.7571
76.1451
76.1451
76.1302
14
Varun Kumar Ojha et al.
RBF
REPTree
SVM
ZeroR
Multi-Scheme
Vote
AdaBoost
Bagging
Ensemble SEL
Random COM
≺
≺
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
≺
≺
≺
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
≺
≺
≺
≺
.
.
.
.
.
.
.
.
.
.
.
.
.
.
≺
.
.
.
.
.
.
.
.
.
.
.
.
.
≺
≺
≺
≺
≺
≺
.
.
.
.
.
.
.
.
.
.
.
.
≺
≺
≺
≺
≺
≺
≺
.
.
.
.
.
.
.
.
.
.
.
=
.
.
.
.
.
.
.
.
.
.
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
.
.
.
.
.
.
.
.
.
≺
≺
≺
≺
≺
≺
≺
.
.
.
.
.
.
.
.
=
=
.
.
.
.
.
.
.
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
.
.
.
.
.
.
≺
≺
≺
≺
≺
≺
≺
=
≺
≺
≺
.
.
.
.
.
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
.
.
.
.
WPE
PART
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Random SUB
NBTree
≺
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Rotation FRST
MLP
≺
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
LMT
≺
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
LWL
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
IBK
DT
IBK
Kstar
LMT
LWL
MLP
NBTree
PART
RBF
REPTree
SVM
ZeroR
Multi-Scheme
Vote
AdaBoost
Bagging
Ensemble SEL
Random COM
Random SUB
Rotation FRST
WPE
Kstar
Classifiers
DT
Table 6 Ranking algorithms according to their performance on test set (10 Fold CV).
≺
≺
=
≺
≺
≺
.
.
.
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
.
.
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
≺
.
the model (weights) with the highest accuracy so that such a combination may
be implemented into electronic portable gas detector form.
Moreover, it is also a difficult task to be certain with the accuracy of an
implemented electronic portable gas detector because the toxic exposure of a
gas is also proportional to the time and not only its safety limit. However,
with a real-time monitoring and requisite maintenance involved, the accuracy
of detector may be relaxed and hence, we resorted to choose 90% accuracy as
the accuracy for our developed detector. So, the classifier’s performance was
compared with a threshold setting of 90% accuracy.
First, let us discuss on the obtained results. For the classifiers belonging
to network-based category F1, the classifier SVM performs better than its
counterparts MLP and RBF both in terms of high accuracy (test accuracy
0.93403) and high consistency (std on test accuracy 0.0041). On the other
hand, the performance of MLP was reported next to SVM with high consistency. The performance of the RBF was found to be inconsistent and poorer
in comparison to its counterparts.
In the tree based category F2, the performance of LMT and REPTree was
comparable to whereas, NBTree has shown poor performance compared to its
counterparts.
Identifying Hazardousness of Sewer-Pipeline
15
In instance-base category F3, the performance of IBK and K Star was
comparative with a high accuracy and high consistency. LWL performed poor
with a very low accuracy.
When it came to the category of rule based classifier F4, PART has outperformed others in its category, but the consistency was not as high as the
consistency of the other well performing classifiers IBK, SVM, MLP, etc. The
classifier ZeroR consistently performed poor in comparison to all other classifiers.
In the ensemble category E1 and E2, the multi-scheme, Random COM, Rotation FRST, Bagging, WPE, and Ensemble SEL performed with high accuracies (over 90%) and consistency. However, the performance of the ensembles
Random Forest, Vote, and AdaBoost were not as satisfactory as compared to
the other ensembles. One of the reason behind poor performance of Random
SUB was the usage of subset of the features. Therefore, the feature selection
may not help in case of this dataset because of the high correlation maintained by each of the features with the output feature. Similarly, Voting used
probability measures to combine the predictors and AddBoost combined weak
predictors, whereas, the entirely better performing ensemble exploited the best
predictors. Hence, they performed better in this scenario.
Considering the assumption of 90% accuracy being a good predictor for implementation as gas detector, we can figure out from Table 5 that the classifiers
belong to category F3 (exception of the classifier LWL) had performed better
than the classifiers of other categories. However, the instance based classifier
IBK is not suitable for the implementation as electronic gas detector since it
required a large memory for its computation for saving all the instances of the
training set. IBK prediction is computed based on all training samples. Hence,
it takes long time to compute the output, which is unacceptable in real time.
The next category whose performance was found close to IBK were the
classifiers of category F2 (tree based classifier). Two classifies, LMT and REP
Tree qualified the 90% accuracy threshold. On the contrary, two classifiers
from each F1 and F4 had performed lower than 90% accuracy. However, SVM
performed significantly well with a very high accuracy 93.36%. Similarly, classifier PART from category F4 had an accuracy of 90.86%. However, since the
SVM produced less number of parameters than the tree based predictor and it
robustly accommodates the noisy attributes, it was recommended from these
experiments that SVM is a proper choice for the implementation of the proposed gas detector.
5 Conclusion
In this work, we explored a real world problem in the context of classification,
where we simplified the approach by offering binary decision to the problem. We explored the problem related to the detection of hazardousness of a
sewer pipeline environment. This is very crucial problem since it is related to
the safety of the persons who have to work under the toxic environment of
16
Varun Kumar Ojha et al.
the sewer-pipeline. Usually, a sewer-pipeline environment contains mixture of
toxic gases. Hence, we collected samples from sewer pipelines from different
locations. Then we examined those samples to identify data samples for our
experiments. We prepared a large dataset by collecting gas sensor responses
from laboratory tests, literature and scaled the collected gas sensor responses
to form a dataset where non-hazardous samples were labeled 0 and hazardous
samples were labeled 1. Finally, we applied 21 different classifiers over the identified dataset and their empirical and statistical performance were evaluated.
We discovered that for this problem, the instance based classifier performed
best followed by the performance of tree based classifiers. However, we found
that the performance of the classifiers were dependent on the ability and mechanism of the classifiers themselves and not on the information regarding which
category they belong to.
Acknowledgements This work was supported by the IPROCOM Marie Curie Initial
Training Network, funded through the People Programme (Marie Curie Actions) of the
European Unions Seventh Framework Programme.
References
1. J. Whorton, ““the insidious foe”–sewer gas,” Western Journal of Medicine, vol. 175,
no. 6, pp. 427–428, Dec. 2001.
2. R. J. Lewis, Sax’s Dangerous Properties of Industrial Materials, 12th ed. Wiley, 2010.
3. N. Gromicko, “Sewer gases in the home,” 2006, http://www.nachi.org/sewer-gaseshome.html.
4. T. Hindu, “Deaths in the drains,” 2014, http://www.thehindu.com/opinion/oped/deaths-in-the-drains/article5868090.ece?homepage=true., Accessed on 15 Dec 2015.
5. NDTV,
“He
died
on
diwali
inside
a
sewage
pipe,”
2014,
http://www.ndtv.com/opinion/he-died-on-diwali-inside-a-sewage-pipe-1245559,
Accessed on 15 Dec 2015.
6. S. Anand, “Dying in the gutters,” Tehelka Magazine, vol. 4, no. 47, Dec 2007,
Achttp://archive.tehelka.com/story main36.asp?filename=Ne081207DYING.asp,
cessed on: 15 Dec 2015.
7. T. Hindu, “Provide safety gear to sewer workers who enter manholes, says court,”
2011,
http://www.thehindu.com/todays-paper/tp-national/provide-safety-gear-tosewer-workers-who-enter-manholes-says-court/article2228688.ece, Accessed on 15 Dec
2015.
8. ——, “Sewer deaths,” 2014, http://www.thehindu.com/opinion/letters/sewer-deaths/
article5873493.ece, Accessed on 15 Dec 2015.
9. ——, “Supreme court orders states to abolish manual scavenging,” 2014,
http://www.thehindu.com/news/national/supreme-court-orders-states-to-abolishmanual-scavenging/article5840086.ece, Accessed on 15 Dec 2015.
10. D. H. Wolpert and W. G. Macready, “No free lunch theorems for optimization,” IEEE
Transactions on Evolutionary Computation, vol. 1, no. 1, pp. 67–82, 1997.
11. J. Li, “A mixed gas sensor system based on thin film saw sensor array and neural
network,” in Proceedings of the Twelfth Southern Biomedical Engineering Conference,
1993, pp. 179–181.
12. A. Srivastava, S. Srivastava, and K. Shukla, “On the design issue of intelligent electronic nose system,” in Proceedings of IEEE International Conference on Industrial
Technology 2000., vol. 2. IEEE, 2000, pp. 243–248.
13. ——, “In search of a good neuro-genetic computational paradigm,” in Proceedings of
IEEE International Conference on Industrial Technology 2000., vol. 1. IEEE, 2000,
pp. 497–502.
Identifying Hazardousness of Sewer-Pipeline
17
14. E. Llobet, R. Ionescu, S. Al-Khalifa, J. Brezmes, X. Vilanova, X. Correig, N. Barsan, and
J. W. Gardner, “Multicomponent gas mixture analysis using a single tin oxide sensor
and dynamic pattern recognition,” IEEE Sensors Journal, vol. 1, no. 3, pp. 207–213,
2001.
15. D.-S. Lee, S.-W. Ban, M. Lee, and D.-D. Lee, “Micro gas sensor array with neural
network for recognizing combustible leakage gases,” IEEE Sensors Journal, vol. 5, no. 3,
pp. 530–536, 2005.
16. M. Ambard, B. Guo, D. Martinez, and A. Bermak, “A spiking neural network for gas
discrimination using a tin oxide sensor array,” in 4th IEEE International Symposium
on Electronic Design, Test and Applications. IEEE, 2008, pp. 394–397.
17. H. Baha and Z. Dibi, “A novel neural network-based technique for smart gas sensors
operating in a dynamic environment,” Sensors, vol. 9, no. 11, pp. 8944–8960, 2009.
18. W. Pan, N. Li, and P. Liu, “Application of electronic nose in gas mixture quantitative
detection,” in IEEE International Conference on Network Infrastructure and Digital
Content. IEEE, 2009, pp. 976–980.
19. C. Wongchoosuk, A. Wisitsoraat, A. Tuantranont, and T. Kerdcharoen, “Portable electronic nose based on carbon nanotube- SnO2 gas sensors and its application for detection
of methanol contamination in whiskeys,” Sensors and Actuators B: Chemical, vol. 147,
no. 2, pp. 392–399, 2010.
20. Q. Zhang, H. Li, and Z. Tang, “Knowledge-based genetic algorithms data fusion and its
application in mine mixed-gas detection,” in Chinese Control and Decision Conference
(CCDC). IEEE, 2010, pp. 1334–1338.
21. W. So, J. Koo, D. Shin, and E. S. Yoon, “The estimation of hazardous gas release
rate using optical sensor and neural network,” Computer Aided Chemical Engineering,
vol. 28, pp. 199–204, 2010.
22. V. K. Ojha, P. Dutta, and H. Saha, “Performance analysis of neuro genetic algorithm
applied on detecting proportion of components in manhole gas mixture,” International
Journal of Artificial Intelligence \& Applications, vol. 3, no. 4, pp. 83–98, 2012.
23. V. K. Ojha and P. Dutta, “Performance analysis of neuro swarm optimization algorithm
applied on detecting proportion of components in manhole gas mixture,” Artificial Intelligence Research, vol. 1, no. 1, pp. 31–45, 2012.
24. V. K. Ojha, P. Dutta, A. Chaudhuri, and H. Saha, “Convergence analysis of backpropagation algorithm for designing an intelligent system for sensing manhole gases,” in
Hybrid Soft Computing Approaches. Springer India, 2016, pp. 215–236.
25. P. Dutta and V. K. Ojha, “Conjugate gradient trained neural network for intelligent
sensing of manhole gases to avoid human fatality,” in Advances in Secure Computing,
Internet Services, and Applications. IGI Global, 2013, pp. 257–280.
26. V. K. Ojha, P. Dutta, A. Chaudhuri, and H. Saha, “Understating continuous ant colony
optimization for neural network training: A case study on intelligent sensing of manhole
gas components,” International Journal of Hybrid Intelligent Systems, vol. 12, no. 4,
pp. 185–202, 2016.
27. ——, “A multi-agent concurrent neurosimulated annealing algorithm: A case study
on intelligent sensing of manhole gases,” International Journal of Hybrid Intelligent
Systems, vol. 12, no. 4, pp. 203–217, 2016.
28. S. Ghosh, A. Roy, S. Singh, H. Saha, V. K. Ojha, and P. Dutta, “Sensor array for
manhole gas analysis,” in 1st International Symposium on Physics and Technology of
Sensors (ISPTS). IEEE, 2012, pp. 9–12.
29. S. Ghosh, H. Saha, C. RoyChaudhuri, V. K. Ojha, and P. Dutta, “Portable sensor array
system for intelligent recognizer of manhole gas,” in Sixth International Conference on
Sensing Technology (ICST). IEEE, 2012, pp. 589–594.
30. C. Cantalini, L. Valentini, I. Armentano, L. Lozzi, J. Kenny, and S. Santucci, “Sensitivity
to NO2 and cross-sensitivity analysis to NH3 , ethanol and humidity of carbon nanotubes
thin film prepared by PECVD,” Sensors and Actuators B: Chemical, vol. 95, no. 1, pp.
195–202, 2003.
31. K. D. Mitzner, J. Sternhagen, and D. W. Galipeau, “Development of a micromachined
hazardous gas sensor array,” Sensors and Actuators B: Chemical, vol. 93, no. 1, pp.
92–99, 2003.
18
Varun Kumar Ojha et al.
32. J. Liu, Y. Zhang, Y. Zhang, and M. Cheng, “Cross sensitivity reduction of gas sensors
using genetic algorithm neural network,” in Optical Methods for Industrial Processes,
S. Farquharson, Ed., vol. 4201. Proceedings of SPIE, 2001.
33. K. J. Donham, “Exposure limits related to air quality and risk assessment,” Iowa Concentrated Animal Feeding Operations Air Quality Study, p. 164, 2002.
34. L. K. Weaver, “Carbon monoxide poisoning,” New England Journal of Medicine, vol.
360, no. 12, pp. 1217–1225, 2009.
35. S. Simonton, “Human health effects from exposure to low-level concentrations of hydrogen sulfide,” Occupational Health & Safety, Nov. 2007.
36. G. Shilpa, “New insight into panic attacks: Carbon dioxide is the culprit,” Journal
of Young Investigators, Nov. 2007, http://www.jyi.org/issue/new-insight-into-panicattacks-carbon-dioxide-is-the-culprit/.
37. D. W. Fahey and M. I. Hegglin, “Twenty questions and answers about the ozone layer:
2010 update,” Scientific assessment of ozone depletion, pp. 4–1, 2010.
38. A. S. Weigend, B. A. Huberman, and D. E. Rumelhart, “Predicting the future: A
connectionist approach,” International journal of neural systems, vol. 1, no. 03, pp.
193–209, 1990.
39. D. Lowe and D. Broomhead, “Multivariable functional interpolation and adaptive networks,” Complex System, vol. 2, pp. 321–355, 1988.
40. C. Cortes and V. Vapnik, “Support-vector networks,” Machine learning, vol. 20, no. 3,
pp. 273–297, 1995.
41. L. Olshen, C. J. Stone et al., “Classification and regression trees,” Wadsworth International Group, vol. 93, no. 99, p. 101, 1984.
42. J. R. Quinlan, C4. 5: programs for machine learning. Elsevier, 2014.
43. F. Esposito, D. Malerba, G. Semeraro, and V. Tamma, “The effects of pruning methods
on the predictive accuracy of induced decision trees,” Applied Stochastic Models in
Business and Industry, vol. 15, no. 4, pp. 277–299, 1999.
44. W. N. H. W. Mohamed, M. N. M. Salleh, and A. H. Omar, “A comparative study
of reduced error pruning method in decision tree algorithms,” in IEEE International
Conference on Control System, Computing and Engineering (ICCSCE), 2012. IEEE,
2012, pp. 392–397.
45. S. H. Walker and D. B. Duncan, “Estimation of the probability of an event as a function
of several independent variables,” Biometrika, vol. 54, no. 1-2, pp. 167–179, 1967.
46. D. R. Cox, “The regression analysis of binary sequences,” Journal of the Royal Statistical Society. Series B (Methodological), pp. 215–242, 1958.
47. N. Landwehr, M. Hall, and E. Frank, “Logistic model trees,” Machine Learning, vol. 59,
no. 1-2, pp. 161–205, 2005.
48. R. Kohavi, “The power of decision tables,” in Machine Learning: ECML-95. Springer,
1995, pp. 174–189.
49. E. Frank and I. H. Witten, “Generating accurate rule sets without global optimization,”
1998.
50. D. W. Aha, D. Kibler, and M. K. Albert, “Instance-based learning algorithms,” Machine
learning, vol. 6, no. 1, pp. 37–66, 1991.
51. J. G. Cleary, L. E. Trigg et al., “K*: An instance-based learner using an entropic distance
measure,” in Proceedings of the 12th International Conference on Machine learning,
vol. 5, 1995, pp. 108–114.
52. E. Frank, M. Hall, and B. Pfahringer, “Locally weighted naive bayes,” in Proceedings of
the Nineteenth conference on Uncertainty in Artificial Intelligence. Morgan Kaufmann
Publishers Inc., 2002, pp. 249–256.
53. C. G. Atkeson, A. W. Moore, and S. Schaal, “Locally weighted learning,” Artificial
Intelligence Review, vol. 11, no. 5, pp. 11–73, 1997.
54. R. Polikar, “Ensemble based systems in decision making,” IEEE Circuits and Systems
Magazine, vol. 6, no. 3, pp. 21–45, 2006.
55. L. Breiman, “Bagging predictors,” Machine learning, vol. 24, no. 2, pp. 123–140, 1996.
56. Y. Freund and R. E. Schapire, “A decision-theoretic generalization of on-line learning
and an application to boosting,” Journal of computer and system sciences, vol. 55,
no. 1, pp. 119–139, 1997.
Identifying Hazardousness of Sewer-Pipeline
19
57. T. K. Ho, “The random subspace method for constructing decision forests,” IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 20, no. 8, pp. 832–
844, 1998.
58. J. J. Rodriguez, L. I. Kuncheva, and C. J. Alonso, “Rotation forest: A new classifier
ensemble method,” IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 28, no. 10, pp. 1619–1630, 2006.
59. R. Caruana, A. Niculescu-Mizil, G. Crew, and A. Ksikes, “Ensemble selection from
libraries of models,” in Proceedings of the twenty-first international conference on Machine learning. ACM, 2004, p. 18.
60. L. I. Kuncheva, Combining pattern classifiers: methods and algorithms. John Wiley
& Sons, 2004.
61. “Weka 3: Data mining software in java,” accessed: 2016-05-01. [Online]. Available:
http://www.cs.waikato.ac.nz/ml/index.html
62. “Matlab: Statistics and machine learning toolbox,” accessed: 2016-05-01. [Online].
Available: http://www.mathworks.com/products/matlab/
| 2cs.AI
|
Fully Context-Aware Video Prediction
arXiv:1710.08518v1 [cs.CV] 23 Oct 2017
Wonmin Byeon12 , Qin Wang1 , Rupesh Kumar Srivastava3 , and Petros Koumoutsakos1
[email protected], [email protected], [email protected], [email protected]
1
ETH Zurich
2
Istituto Dalle Molle di Studi sull’Intelligenza Artificiale (The Swiss AI Lab IDSIA)
3
NNAISENSE
Abstract
This paper proposes a new neural network design for unsupervised learning through
video prediction. Current video prediction models based on convolutional networks,
recurrent networks, and their combinations often result in blurry predictions. Recent work has attempted to address this issue with techniques like separation of
background and foreground modeling, motion flow learning, or adversarial training.
We highlight that a contributing factor for this problem is the failure of current
architectures to fully capture relevant past information for accurately predicting the
future. To address this shortcoming we introduce a fully context-aware architecture,
which captures the entire available past context for each pixel using Parallel MultiDimensional LSTM units and aggregates it using context blending blocks. Our
model is simple, efficient and directly applicable to high resolution video frames.
It yields state-of-the-art performance for next step prediction on three challenging
real-world video datasets: Human 3.6M, Caltech Pedestrian, and UCF-101 and
produces sharp predictions of high visual quality.
1
Introduction
Learning from raw videos to predict future video frames without labels has recently become an
important direction of research. In this problem setting, a model is trained to predict future frames
conditioned on the past and learns a representation that captures the information (the appearance and
the motion of objects) in a video without external supervision. Such a representation can then be
transferred to other video analysis tasks such as action recognition [1] or utilized for model based
reinforcement learning [2].
Learning such predictive models for natural videos is a challenging problem due to various resolutions,
object occlusion, camera movement, dynamic scene and light changes between frames. In particular,
videos often include large and rapid motion which is hard to capture. To make modeling easier,
explicit motion priors [3, 4] are often incorporated into deep neural networks.
Srivastava et al. [1] introduced Long Short-Term Memory (LSTM) [5] based encoder-decoder models
for the task of video prediction. This architecture was able to model temporal dependencies but
did not take spatial regularities of the data into account, making it very difficult to generate precise
predictions. A more suitable architectural unit for video prediction networks extended the fullyconnected LSTM structure by incorporating convolutions in LSTM operations to better capture
spatio-temporal correlations (developed simultaneously by Xingjian et al. [6] and Stollenga et al.
[7]). Current models rely on these units or traditional convolutional units [8, 9] to capture past frame
information, but the resulting predicted frames are often blurry. Several possible methods to address
this problem have been suggested [10, 11, 12, 13, 14].
As usual for any modeling task, a possible source of blurry predictions or artifacts is lack of modeling
capacity or appropriate training data. However, even sufficiently powerful models training on
extensive training data can suffer from fundamental limitations. Imprecise predictions (resulting
in blurry outputs) from these models can be a manifestation of model uncertainty, which in turn
arises from two main sources: 1) the future may be ambiguous given the past, and 2) the model fails
to sufficiently capture relevant past information which might reduce uncertainty. It is well known
that to address the first issue, it is necessary to use appropriate loss functions [10, 14] so that large
uncertainties and multiple modes can be captured. This source of uncertainty has therefore received
more attention in recent work, and there is a lack of exploration of strategies to mitigate the second
source, which this paper attempts to address.
We first highlight a "blind spot problem" in recent network architectures for video prediction, showing
that they do not take the entire spatio-temporal context from past frames into account. This leads to
increased uncertainty about the future which can not be remedied using recent techniques explored
in the literature mentioned above. As a solution, we propose a new model for video prediction,
which integrates the full available context of each pixel in all possible directions (left, right, top,
bottom, and depth directions). This context-aware predictive model is rather simple compared to other
recent approaches and produces sharper predictions without using novel loss functions, motion or
background priors. It outperforms state-of-the-art approaches significantly in a variety of challenging
video prediction scenarios: car driving, human motion, and diverse human actions. We also evaluate
the effects of various modifications to the model such as directional weight-sharing and different
context blending schemes. Quantitative improvements on metrics are accompanied by results of high
visual quality showing sharper future predictions with reduced blurriness or other motion artifacts.
These results suggest that many current techniques for video prediction and analysis may benefit
from fully context-aware architectures, and draw attention to the importance of uncertainty arising
from missing context.
2
Related Work
Network Architectures: Published approaches for video analysis exploit different amounts of spatiotemporal information in different ways depending on model architecture. One common strategy is
to use models based on 3D Convolutional Neural Networks (CNNs) that use convolutions across
temporal and spatial dimensions to model all local correlations [3, 15] for supervised learning. Similar
architectures have been used for video prediction to directly generate the RGB values of pixels in
videos [8, 10, 16]. An inherent limitation of these models is that convolutions take only short-range
dependencies into account due to the limited size of the kernels.
This short-range context is not enough to capture large and variable motion between frames, and
therefore such networks need a large stack of convolutional layers to include a wider range of contexts.
Since the context is always limited by the architecture, loss of relevant context can lead to uncertainty
and often produces blurry results. One idea, applicable to situations with mostly static background,
is to explicitly model moving foreground objects separately from the static background using a
two-stream architecture to improve the precision of the predictions Simonyan et al. [3], Vondrick
et al. [16], and Finn et al. [13].
Incorporating recurrent neural networks can address the issue of limited context, and this is idea
behind the use of ConvLSTM based models which replace the internal transformations of an LSTM
cell with convolutions. Xingjian et al. [6] proposed this design for precipitation nowcasting; the
motivation being that the convolutional operation would model spatial dependencies, while LSTM
connectivity would offer increased temporal context. Since then it has been utilized for incorporating
long term temporal context in video prediction models [12, 13]. Specifically, Finn et al. [13] use it in
a model that learns to predict pixel motions instead of values, and Lotter et al. [12] develop a model
inspired by the predictive coding that learns to better predict future frames by incorporating previous
prediction errors. While these models improve prediction quality over past work, they still retain
blurriness in generated frames.
Kalchbrenner et al. [14] discussed that a general probabilistic model of videos should take into
account the entire history (all context in past frames and generated pixels of present frame) for
generating each new pixel. However, their proposed Video Pixel Networks (VPNs) use encoders
based on stacks of convolution layers for which the context available from past frames can be limited
in the same sense as the models mentioned above. In all these cases, covering more context requires
larger depth leading to more parameters and computation. In contrast, our proposed models avoid this
strong dependence by design since even a single layer model is fully-context aware for any image
size. We note that we do not use the decoders used by VPNs which additionally condition outputs
2
on the generated pixels of the current frame. This loses available context, but allows us to fairly
compare to other recent models which also generate all pixels in a frame in parallel. This also makes
application to larger resolution videos efficient – VPNs were only tested on a maximum resolution of
64×64, while we test on up to 256×256 frames.
Learning strategies: Mathieu et al. [10] attributed the issue of blurry predictions to the Mean
Squared Error (MSE) loss function used by several other methods. Since MSE assumes that the data
comes from a Gaussian distribution, it trains the model to produce the mean of possible outcomes
when the posterior distribution is multimodal. To reduce the ambiguity in prediction, they proposed
the use of L1 & Gradient Difference Loss (GDL) loss functions, together with adversarial training
based on Generative Adversarial Networks [17]. Vondrick et al. [16] also used adversarial training,
but developed a two-stream architecture to generate short videos conditioned on single input frames.
Context aggregation using MD-LSTMs: A general and principled approach to incorporate long
term context in multiple dimensions is Multidimensional LSTM (MD-LSTM; [18]), a specialization
of DAG-RNNs [19], which has been applied for various supervised tasks such as handwriting
recognition [20], language translation [21], image classification [22], and segmentation [23]. This
architecture uses MD-LSTM cells to sequentially connect the input elements between multiple
dimensions, thus combining temporal and non-temporal dependencies. It contains two MD-LSTM
modules per dimension to combine context from all possible directions.
In principle, the MD-LSTM can be applied to any high-dimensional domain problem (including
video prediction) to model all relevant dependencies in the data compactly. However, the fully
sequential nature of the model makes it unsuitable for parallelization and thus impractical for higher
dimensional data such as videos. The PyraMiD-LSTM [7] addressed this issue by re-arranging the
recurrent connection topology of each MD-LSTM module from cuboid to pyramidal and enabled an
application to volumetric segmentation. The change in topology reduces the total number of modules
needed and allows parallelization of computation inside each module. In section 4, we will utilize
similar ideas to design an improved model for video prediction. We note here that mathematically,
one of the parallelized MD-LSTM modules in PyraMiD-LSTM is equivalent to a typical ConvLSTM
which has LSTM connections along the temporal dimension.
3
Missing Contexts in Other Network Architectures
As mentioned earlier, blurry predictions can result from a video prediction model if it does not
adequately capture all relevant information in the past video frames which can be used to reduce
uncertainty. Figure 1 shows the recurrent connections of a pixel at time t with a 3 × 3 convolution
between two frames (left) and the information flow of a ConvLSTM predicting the pixel at time T + 1
(right). The covering context grows progressively over time (depth), but there are also blind areas
which cannot be used for prediction. In fact, as can be seen in Figure 1 (right, marked in gray color),
frames in the recent past have larger blind areas. Due to this structural issue, the network is unable to
capture the entire available context, and is likely to miss important spatio-temporal dependencies
leading to increased ambiguity in the predictions. The prediction will eventually fail when the object
appearance or motion in videos changes dramatically within a few frames.
One possible way to address limited context, widely used in CNNs for image analysis, is to expand
context by stacking multiple layers (sometimes with dilated convolutions [24]). However, stacking
layers still limits the available context to a maximum as dictated by the network architecture, and
the number of additional parameters required to gain sufficient context can be very large for high
resolution videos. Another technique that can help is using a multi-scale architecture, but fixed scale
factors may not generalize to all possible objects, their positions and motions. In the next section, we
introduce a solution for removing these blind areas by combining the full contexts from all directions.
Our approach does not require more parameters for higher resolution videos or a priori specification
of scale factors, and results in precise predictions without any artifacts.
4
Approach
Let xT1 = {x1 , ..., xT } be a given input sequence of length T . xt ∈ RH×W ×C is the t-th frame,
where t ∈ {1, ..., T }, H is the height, W the width, and C the number of channels. For simplicity,
assume C = 1. xT1 is then a cuboid of pixels bounded by six planes. The task is to predict p future
frame(s) in the sequence, xt+p
t+1 = {xt+1 , ..., xt+p } (next-frame prediction if p = 1). Therefore, our
3
prediction
frame t-1
frame T-3
frame t
frame T-2
frame T-1
frame T
frame T+1
Figure 1: (left) The Convolutional LSTM (ConvLSTM) structure between two frames. (right) The information
flow in ConvLSTM over time. This architecture introduces the blind areas (marked as a gray color) which cannot
be used to predict the pixel value at time T + 1. The closer time frame has larger blind areas.
t- direction
frame t-1
frame t
w- direction
frame t-1
frame t
w+ direction
frame t-1
frame t
h- direction
frame t-1
frame t
h+ direction
frame t-1
frame t
prediction
frame T-3
frame T-2
frame T-1
frame T
frame T+1
Figure 2: (top) The Parallel MD-LSTM (PMD) units between two frames for five directions: (t − 1), (w − 1),
(w + 1), (h − 1), and (h + 1), where h, w, and t indicates the current position for height, width, and time
dimensions. (down) The information flow using all directional PMD units. It covers all available contexts of the
history frames in a single architecture.
goal is to integrate information from the entire cuboid xT1 into a representation at the plane where
t = T , which can be used for predicting xt+p
t+1 . The proposed model includes two main components
that learn dependencies along spatio-temporal dimensions in a single architecture. The first is the
Parallel MD-LSTM unit (PMD unit) that connects dependencies to each pixel. The second is the
Context Blending Block that combines the output of PMD units from multiple directions.
4.1
Parallel MD-LSTM Unit
A parallel MD-LSTM (PMD) unit is a variant of the MD-LSTM unit, proposed by Stollenga et al.
[7], which is more amenable to parallelization. As mentioned in section 2, it implements a method of
modeling dependencies in multiple dimensions. Although it is mathematically similar to ConvLSTM,
we adopt this terminology to highlight that it is not necessary to limit convolutional operations to
spatial dimensions and LSTM connectivity to the temporal dimension.
For any sequence of K two dimensional planes xK
1 = {x1 , ..., xK }, the PMD unit computes the
current cell and hidden state ck , sk using input, forget, output gates ik , fk , ok , and the transformed
cell c̃k given the cell and hidden state from the previous plane, ck−1 , sk−1 .
ik
fk
ok
c̃k
ck
sk
= σ(Wi ∗ xk + Hi ∗ sk-1 + bi ),
= σ(Wf ∗ xk + Hf ∗ sk-1 + bf ),
= σ(Wo ∗ xk + Ho ∗ sk-1 + bo )
= tanh(Wc̃ ∗ xk + Hc̃ ∗ sk-1 + bc̃ ),
= fk ck-1 + ik c̃k ,
= ok tanh(ck ).
(1)
Here (∗) is the convolution operation, and ( ) the element-wise multiplication. W and H are the
weights for input-state and state-state. The size of weight matrices are dependent only on the kernel
size and number of units. If the kernel size is larger, more local context is taken into account. The
4
convolution operation provides an efficient computation on GPUs 1 , as well as takes the surrounding
local correlation into account.
4.2
Context Blending Block
This block captures the entire available context by combining the output of PMD units from all
relevant directions. This is the critical difference from the traditional ConvLSTM: the conditioning
directions are aligned not only with the time dimension but also with the spatial dimensions. As
shown in section 3, using a ConvLSTM would be equivalent to running a PMD unit along the time
dimension from k = 1 to k = T , which would only integrate information from a pyramid shaped
region of the cuboid and ignore several blind areas. For this reason, it is necessary to use four
additional PMD units, for which the conditioning directions are aligned with the spatial dimensions,
as shown in Figure 2 (top). When the current processing pixel is at the location l := (h, w, t), the
connecting recurrent directions are defined as h-, h+, w-, w+, and t-. The five PMD unit outputs from
these directions are hd where d ∈ D = {h-, h+, w-, w+, t-}. We suggest two ways to combine the
information from different directions.
Uniform blending (U-blending): this strategy is used in the traditional MD-LSTM [25, 23] and
PyraMiD LSTM [7]. The computation is simply the summation of all directions before the regular
fully-connected connections:
X
ml = f ((
sdl ) · Wl + bl ),
(2)
d∈D
N 1×N 2
N2
where Wl ∈ R
and bl ∈ R are a weight matrix and a bias. N 1 is the number of PMD units,
and N 2 is the number of (blending) blocks. f is an activation function.
d5
Weighted blending (W-blending): PMD units from all five directions output sd1
l , ...sl , where
d1, ..d5 are the indices of the directions. This block then concatenates s from all directions:
d1
sl
sd2
l
Sl = .
(3)
..
sd5
l
The vector Sl is then weighted as follows:
ml = f (Sl · Wl + bl ),
(4)
where Wl ∈ R(5×N 1)×N 2 (5 is the number of directions). The main difference between the two
blending approaches is whether the information for each pixel is equally important for all directions
or needs to be rated and prioritized. In the latter case, this priority is learned while training.
The PMD computation for each direction (top) and the context mapping to past frames (down) are
visualized in Figure 2. Compared to Figure 1, the architecture in Figure 2 covers all available
context at every frame (no blind areas) in order to predict the pixel values at time T + 1.
4.3
Directional Weight-Sharing (DWS)
Visual data tend to have structurally similar local patterns along opposite directions. This is the
reason why horizontal flipping is a commonly used data augmentation technique in computer vision.
We explore the use of a similarly inspired weight-sharing technique for regularizing the proposed
networks. The weights and biases of the PMD units in opposite directions are shared i.e. weights for
h- and h+ are shared, as are w- and w+. This strategy has several benefits in practice: 1) it lowers the
number of parameters to be learned, 2) it incorporates knowledge about structural similarity into the
model, and 3) it improves generalization.
4.4
Training
x̂l = g(ml ) is an output of the top-most (output) layer, where g is an output activation function. The
model minimizes the loss between the predicted pixel (model output) x̂l and the target pixel xl+1 .
Mathieu et al. [10] introduced several loss functions for better prediction. Here, Lp loss and the
1
Unlike the traditional MD-LSTM structure, each pixel in the same plane is independent from the conditioning
pixels (see Figure 2-top). Thus, all pixels in the plane can be processed concurrently.
5
Image Gradient Difference Loss (GDL) are combined. Let y and x̂ are the target and the predicted
frame. The objective function is defined as follows:
L(y, x̂) = λp Lp (y, x̂) + λgdl Lgdl (y, x̂)
Lp (y, x̂) = ||y − x̂||p
X
Lgdl (y, x̂) =
||yi,j − yi−1,j | − |x̂i,j − x̂i−1,j || + ||yi,j−1 − yi,j | − |x̂i,j−1 − x̂i,j ||,
(5)
i,j
where |.| is the absolute value function, x̂i,j and yi,j are the pixel elements from the frame x̂ and y,
respectively. λp and λgdl are the weights for each loss. In our experiments, λgdl is set to 1 when
p = 1 and 0 when p = 2. λp is always set to 1.
We use ADAM [26] as the optimizer with an initial learning rate of 1e − 3. The learning rate is
decayed every 5 epochs with the decay rate 0.99. Weights are initialized using the Xavier’s normalized
initializer [27] and the states of the LSTMs are initialized to zero (no prior).
5
Experiments
We evaluated the proposed approach on three real-world scenarios with distinct characteristics: human
motion prediction, car-mounted camera video prediction, and human activity prediction.
The first scenario is to predict human motions using the Human 3.6M dataset [28]. These videos
have static backgrounds and are recorded with a fixed camera. Therefore, the model can focus on
the appearance and the motion of humans. The second scenario is to model a real-world driving
environment, with potential applications in autonomous driving. Videos are taken from cameras
mounted on moving vehicles, and include inconsistent camera movement but with (relatively) similar
background, for instance, sky, road, or others cars appear in many of the scenes; but sometimes there
are unique backgrounds like a car parking space. To model such videos, not only does a model need
to predict the motion of foreground objects, but it also needs to model camera movements. Moreover,
training and testing videos are taken from the different cities: KITTI dataset [29] (train) from Germany
and CalTech Pedestrian dataset [30] (test) from the US. Such a setup ensures that the model does not
memorize particular backgrounds and has to use past information for predictions. The last scenario
is to capture people performing a diverse range of activities (applying makeup, playing sports etc.)
using the UCF-101 dataset [31]. These videos, collected from YouTube (https://www.youtube.com/),
feature varied backgrounds as well as rapid camera movement.
All input pixel values are normalized to the range [0, 1]. For the human motion and car-mounted
videos, the models are trained to use ten frames as input to predict the next frame. For the human
activity videos, the input consists four frames (for fair comparison to previous work). Quantitative
evaluation on the test sets is performed based on mean Peak Signal-to-Noise Ratio (PSNR) and the
Structural Similarity Index Measure (SSIM) [32]2 . These commonly used numerical measures are
known to be not fully representative of human vision. Therefore, we highly recommend looking at
the visual results in Figure 3 and the website https://sites.google.com/view/contextvp.
Network architecture: Several video prediction models based on the proposed architecture were
tested, with variable number of layers (one, three, or four) and types of context blending (uniform
or weighted). When using a single hidden layer, 64 PMD units from each of the five directions are
directly connected to the output layer. The three and four layer models include PMD units followed
by context blending blocks. For the three layer model, the number of hidden units are 16 (PMD1)
- 20 (Blending1) - 32 (PMD2) - 45 (Blending2) - 64 (PMD3) - 80 (Blending3). For the four layer
model, the number of hidden units are 32 (PMD1) - 32 (Blending1) - 64 (PMD2) - 64 (Blending2) 32 (PMD3) - 32 (Blending3) - 64 (skip connection with Blending1) - 64 (PMD4) - 64 (Blending4) 128 (skip connection with Blending2). The skip connection simply concatenates outputs of different
layers. The output layer always uses the sigmoid activation function which outputs values in the range
(0, 1). Each PMD unit has convolutional kernels of size 3 × 3 (for human motion and car-mounted
camera videos) or 5 × 5 (for human action prediction). The activation function f in both Equation 2
and Equation 4 are set to linear. Use of nonlinear activations (e.g., ReLU [33] or tanh) or layer
normalization [34] in the blending blocks does not affect the performance in our experiments. We
report the results from L1 (p = 1 in Equation 5) with the GDL loss for our experiments. Finn et al.
2
Mean Squared Error (MSE) is also reported for the car-mounted camera video prediction to compare with
PredNet [12].
6
Table 1: Evaluation of Next-Frame Predictions on the Human3.6M dataset. The model is trained on ten frames
and predicts the next frame. The results are averaged over test videos. Higher values of PSNR and SSIM, lower
values of MSE indicate better results. L and DWS indicate the number of layers and Directional Weight-Sharing,
respectively.
Method
PSNR
SSIM
Copy-Last-Frame
32
-
BeyondMSE [10]
PredNet [12]
DNA [13]
26.7
38.9
42.1
0.992
Ours (1-L)
Ours (3-L)
Ours (4-L, U-blending)
Ours (4-L, W-blending)
Ours (4-L, W-blending, DWS)
38.1
41.2
42.3
44.8
45.2
0.990
0.992
0.994
0.996
0.996
[13] found that L1 with the GDL loss function performs better than L2 but the performance in our
case was very similar.
Human Motion Prediction: We first evaluate our model on Human3.6M dataset [35]. The dataset
includes seven human subjects (three females and four males). Five subjects are used for training and
the other two for validation and testing. The videos are subsampled to 10 fps and downsampled to
64 × 64 resolution. These and others details of the experimental setup are same as Finn et al. [13].
Note that the model of Finn et al. [13] learns a mask to reconstruct the fixed background easily (copying pixels directly from the previous frame) since videos in this dataset contain a static background
and camera with small human movements. This masking scheme is suitable for this scenario but is
restricted to such conditions. In contrast, our model learns the entire structure of appearance and
motion (as well as their correlations) without such domain specific properties.
Table 1 shows the comparison of the prediction results with the state-of-the-art approach (DNA) Finn
et al. [13] as well as PredNet [12], BeyondMSE [10], and Copy-Last-Frame. Copy-Last-Frame is
included to ensure that our model does not simply copy the last frame. Our approach consistently
outperforms all other approaches, and notable improvements are obtained by increasing the number
of layers. Single layer networks already outperform BeyondMSE which is built on 3D-CNN, and the
three layer networks outperform PredNet which uses ConvLSTM. Finally, the four layer networks
with weighted context blending and directional weight-sharing outperform the best performing
approach on this dataset (DNA) which also uses ConvLSTM and explicitly models a background
mask and motion separately. All compared approaches have between five to seven layers. These
results show that our approach can efficiently capture the internal representation using simpler
networks. Figure 3 (row 1) shows sample visual prediction results.
Car-mounted Camera Video Prediction: As mentioned earlier, the model is trained on the KITTI
dataset [29] (Germany) and tested on the CalTech Pedestrian dataset [30] (USA) Every ten frames
from “city”, “Residential”, and “Road” videos are sampled for training resulting in ≈41K frames.
Frames from both datasets are center-cropped and down-sampled to 128 × 160 pixels. We used the
exact data preparation as PredNet [12] for direct comparison.
The car-mounted camera videos are taken from moving vehicles and consist of a wide range of
motions. Compared to Human3.6M dataset, which has static background and small motion flow, this
dataset has diverse and large motion of cars (in different scales) and also has a camera movement. To
understand such videos, a model is required to learn not only small movement of pedestrians, but
also relatively large motion of surrounding vehicles and backgrounds.
We compare our approach with the Copy-Last-Frame baseline and the PredNet which is the current
state-of-the-art model for this dataset. Note that the scores provided in Lotter et al. [12] are averaged
over nine frames (time step 2-10), but ours are computed only on the last predicted frame. We
therefore re-calculated the scores of PredNet using their trained network. As shown in Table 2, our
four layer model with weighted context blending and directional weight-sharing outperforms the
7
true
predicted
true
predicted
true
predicted
time →
Figure 3: Next frame prediction for human motion (row 1) and car-cam video (row 2-3) sequences. The first
rows are the true frames and the second rows are predictions. For car-cam video sequences, the model was
trained on the KITTI dataset and the prediction results are from the CalTech Pedestrian dataset. More examples
of the prediction results are online: https://sites.google.com/view/contextvp
Table 2: Evaluation of Next frame prediction on the CalTech Pedestrian dataset (trained on the KITTI dataset).
The model is trained on ten frames and predicts the next frame. The results are averaged over test videos. Higher
values of PSNR and SSIM, lower values of MSE indicate better results. L and DWS indicate the number of
layers and Directional Weight-Sharing, respectively. (+) This score is provided by [36]. (*) The scores provided
in Lotter et al. [12] are averaged over nine frames (time step 2-10), but ours are computed only on the last
predicted frame. We therefore re-calculated the scores of PredNet using their trained network.
MSE
(×10−3 )
PSNR
SSIM
Copy-Last-Frame
7.95
23.3
0.779
+BeyondMSE [10]
*PredNet [12]
Dual Motion GAN [36]
3.26
2.42
2.41
27.6
-
0.881
0.905
0.899
Ours (3-L, W-blending)
Ours (4-L, W-blending)
Ours (4-L, W-blending, DWS)
2.18
2.04
1.96
27.9
28.4
28.6
0.914
0.919
0.926
Method
state-of-the-art on all metrics.Some samples of the prediction results on CalTech Pedestrian dataset
are provided in Figure 3 (row 2-3). Samples are selected to show the results from diverse driving
scenarios: moving into the indoor parking space, (row 2) and the front car changing the lane (row 3).
Our model is able to adapt predictions to the current scene and make sharp predictions without any
artifacts.
Human Action Prediction: The last evaluation is on the UCF-101 Dataset [31]. Although many
videos in this dataset contain small movements between frames, they contain much more diversity in
objects, backgrounds and camera motions compared to previous datasets. Our experimental setup
follows that of Mathieu et al. [10]. About 500,000 training videos are selected from the UCF-101
training set, and 10% of UCF-101 test set is used for testing (378 videos). All frames are resized to
256 × 256. Note that Mathieu et al. [10] used randomly selected sequences of 32 × 32 patches from
the Sports-1M dataset [38] for training because the motion between frames in the UCF-101 dataset
8
Table 3: Evaluation of Next-Frame Predictions on the UCF-101 dataset. The model is trained on four frames
and predicts the next frame. The results are averaged over test videos. Higher values of PSNR and SSIM, lower
values of MSE indicate better results. L and DWS indicate the number of layers and Directional Weight-Sharing,
respectively.
Method
time
T −1
PSNR
SSIM
EpicFlow [37]
BeyondMSE [10]
DVF [11]
31.6
32
33.4
0.93
0.92
0.94
Ours (3-L, W-blending)
Ours (4-L, W-blending)
Ours (4-L, W-blending, DWS)
32.2
34.3
34.8
0.90
0.92
0.92
T
T +1
T −1
T
T +1
true
predicted
Figure 4: Examples of failed prediction for car-cam video sequences. The first rows are the true frames and the
second rows are predictions. T − 1 and T − 2 are the last two input frames, and T is the current output frame.
As can be seen in the predicted frame at time T , some details of objects which were not appeared in the last
frames, are missed.
are too small to capture dynamics. Our model however, is directly trained on UCF-101 subsequences
of length four with the original resolution. Motion masks generated using Epicflow [39] provided by
Mathieu et al. [10] are used for validation and testing, so the evaluation is focused on regions with
significant motion when computing PSNR and SSIM.x
Table 3 presents the quantitative evaluation on the UCF-101 dataset. Our best model, a four layered network with weighted context blending and directional weight-sharing, again outperforms
BeyondMSE and predictions generated using the outputs of EpicFlow. These results indicate that this
model can capture all relevant spatial-temporal information and use it to make sharp predictions in
very diverse settings without requiring adversarial training.
Failure Cases: Figure 4 shows some examples of failing cases in prediction. Although most of
objects and background in the predicted frame at time T + 1 look sharp, some object (marked as
a red circle) appears to be blurry. This particular object hardly appears in the last (input) frames
before prediction (see true frames at time T − 1, T , and T + 1 in Figure 4), so the input does not
have sufficient information to predict details of the object (see the predicted frames at time T + 1 in
Figure 4). Thus, this failure case is caused by future ambiguity mentioned in section 1.
6
Conclusion and Future Directions
This paper identified the issue of missing context in current video prediction models, which is a
contributor to uncertain predictions about the future and leads to generation of blurry frames. To
address this issue, we adapted ideas from recent work in volumetric segmentation of biomedical data
to develop a prediction architecture that can capture all the relevant context efficiently. The resulting
model outperformed existing approaches for video prediction in a variety of scenarios, demonstrating
the importance of fully context-aware models.
Our model did not incorporate the use of other orthogonal ideas to improve video prediction, such
as use of multiple scales, explicit background/motion flow modeling, or novel training strategies.
Since these ideas have been previously explored for models with incomplete context, a promising
future direction is to evaluate their influence on fully context-aware models. Our work suggests that
availability of full context should be a required feature of any video prediction baseline to rule out
9
multiple sources of uncertainty. Finally, an important line of investigation is transfer learning for
various supervised or reinforcement learning tasks.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
N. Srivastava, E. Mansimov, and R. Salakhudinov. “Unsupervised learning of video representations using lstms”. In: International Conference on Machine Learning. 2015, pp. 843–
852.
R. S. Sutton and A. G. Barto. Reinforcement Learning: An Introduction. Cambridge, MA: MIT
Press, 1998.
K. Simonyan and A. Zisserman. “Two-stream convolutional networks for action recognition in
videos”. In: Advances in neural information processing systems. 2014, pp. 568–576.
J. Yue-Hei Ng, M. Hausknecht, S. Vijayanarasimhan, O. Vinyals, R. Monga, and G. Toderici.
“Beyond short snippets: Deep networks for video classification”. In: Proceedings of the IEEE
conference on computer vision and pattern recognition. 2015, pp. 4694–4702.
S. Hochreiter and J. Schmidhuber. Long Short-Term Memory. Tech. rep. FKI-207-95. Revised
1996 (see www.idsia.ch/˜juergen, www7.informatik.tu-muenchen.de/˜hochreit). Fakultät für
Informatik, Technische Universität München, 1995.
S. Xingjian, Z. Chen, H. Wang, D.-Y. Yeung, W.-K. Wong, and W.-c. Woo. “Convolutional
LSTM network: A machine learning approach for precipitation nowcasting”. In: Advances in
Neural Information Processing Systems. 2015, pp. 802–810.
M. F. Stollenga, W. Byeon, M. Liwicki, and J. Schmidhuber. “Parallel multi-dimensional
LSTM, with application to fast biomedical volumetric image segmentation”. In: Advances in
Neural Information Processing Systems. 2015, pp. 2998–3006.
M. Ranzato, A. Szlam, J. Bruna, M. Mathieu, R. Collobert, and S. Chopra. “Video (language) modeling: a baseline for generative models of natural videos”. In: arXiv preprint
arXiv:1412.6604 (2014).
J. Oh, X. Guo, H. Lee, R. L. Lewis, and S. Singh. “Action-conditional video prediction using
deep networks in atari games”. In: Advances in Neural Information Processing Systems. 2015,
pp. 2863–2871.
M. Mathieu, C. Couprie, and Y. LeCun. “Deep multi-scale video prediction beyond mean
square error”. In: arXiv preprint arXiv:1511.05440 (2015).
Z. Liu, R. Yeh, X. Tang, Y. Liu, and A. Agarwala. “Video Frame Synthesis using Deep Voxel
Flow”. In: arXiv preprint arXiv:1702.02463 (2017).
W. Lotter, G. Kreiman, and D. Cox. “Deep predictive coding networks for video prediction
and unsupervised learning”. In: arXiv preprint arXiv:1605.08104 (2016).
C. Finn, I. Goodfellow, and S. Levine. “Unsupervised learning for physical interaction through
video prediction”. In: Advances In Neural Information Processing Systems. 2016, pp. 64–72.
N. Kalchbrenner, A. v. d. Oord, K. Simonyan, I. Danihelka, O. Vinyals, A. Graves, and K.
Kavukcuoglu. “Video pixel networks”. In: arXiv preprint arXiv:1610.00527 (2016).
D. Tran, L. D. Bourdev, R. Fergus, L. Torresani, and M. Paluri. “C3D: generic features for
video analysis”. In: CoRR, abs/1412.0767 2 (2014), p. 7.
C. Vondrick, H. Pirsiavash, and A. Torralba. “Generating videos with scene dynamics”. In:
Advances In Neural Information Processing Systems. 2016, pp. 613–621.
I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville,
and Y. Bengio. “Generative adversarial nets”. In: Advances in neural information processing
systems. 2014, pp. 2672–2680.
A. Graves, S. Fernández, and J. Schmidhuber. “Multi-dimensional Recurrent Neural Networks”.
In: Artificial Neural Networks – ICANN 2007: 17th International Conference, Porto, Portugal,
September 9-13, 2007, Proceedings, Part I. Ed. by J. M. de Sá, L. A. Alexandre, W. Duch,
and D. Mandic. Berlin, Heidelberg: Springer Berlin Heidelberg, 2007, pp. 549–558. ISBN:
978-3-540-74690-4.
P. Baldi and G. Pollastri. “The principled design of large-scale recursive neural network
architectures–dag-rnns and the protein structure prediction problem”. In: Journal of Machine
Learning Research 4.Sep (2003), pp. 575–602.
A. Graves and J. Schmidhuber. “Offline Handwriting Recognition with Multidimensional
Recurrent Neural Networks”. In: NIPS. 2009.
N. Kalchbrenner, I. Danihelka, and A. Graves. “Grid long short-term memory”. In: arXiv
preprint arXiv:1507.01526 (2015).
10
[22] W. Byeon, M. Liwicki, and T. M. Breuel. “Texture classification using 2d lstm networks”. In:
Pattern Recognition (ICPR), 2014 22nd International Conference on. IEEE. 2014, pp. 1144–
1149.
[23] W. Byeon, T. M. Breuel, F. Raue, and M. Liwicki. “Scene Labeling With LSTM Recurrent
Neural Networks”. In: CVPR. 2015.
[24] F. Yu and V. Koltun. “Multi-scale context aggregation by dilated convolutions”. In: arXiv
preprint arXiv:1511.07122 (2015).
[25] A. Graves, S. Fernández, and J. Schmidhuber. “Multi-dimensional recurrent neural networks”.
In: Proceedings of the 17th International Conference on Artificial Neural Networks. 2007.
[26] D. Kingma and J. Ba. “Adam: A method for stochastic optimization”. In: arXiv preprint
arXiv:1412.6980 (2014).
[27] X. Glorot and Y. Bengio. “Understanding the difficulty of training deep feedforward neural
networks.” In: Aistats. Vol. 9. 2010, pp. 249–256.
[28] C. Ionescu, D. Papava, V. Olaru, and C. Sminchisescu. “Human3. 6m: Large scale datasets
and predictive methods for 3d human sensing in natural environments”. In: IEEE transactions
on pattern analysis and machine intelligence 36.7 (2014), pp. 1325–1339.
[29] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun. “Vision meets robotics: The KITTI dataset”. In:
The International Journal of Robotics Research 32.11 (2013), pp. 1231–1237.
[30] P. Dollár, C. Wojek, B. Schiele, and P. Perona. “Pedestrian detection: A benchmark”. In:
Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on. IEEE.
2009, pp. 304–311.
[31] K. Soomro, A. R. Zamir, and M. Shah. “UCF101: A dataset of 101 human actions classes
from videos in the wild”. In: arXiv preprint arXiv:1212.0402 (2012).
[32] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. “Image quality assessment: from
error visibility to structural similarity”. In: IEEE transactions on image processing 13.4 (2004),
pp. 600–612.
[33] V. Nair and G. E. Hinton. “Rectified linear units improve restricted boltzmann machines”.
In: Proceedings of the 27th international conference on machine learning (ICML-10). 2010,
pp. 807–814.
[34] J. L. Ba, J. R. Kiros, and G. E. Hinton. “Layer normalization”. In: arXiv preprint
arXiv:1607.06450 (2016).
[35] C. Ionescu, D. Papava, V. Olaru, and C. Sminchisescu. “Human3.6M: Large Scale Datasets and
Predictive Methods for 3D Human Sensing in Natural Environments”. In: IEEE Transactions
on Pattern Analysis and Machine Intelligence 36.7 (2014), pp. 1325–1339.
[36] X. Liang, L. Lee, W. Dai, and E. P. Xing. “Dual Motion GAN for Future-Flow Embedded
Video Prediction”. In: arXiv preprint arXiv:1708.00284 (2017).
[37] J. Revaud, P. Weinzaepfel, Z. Harchaoui, and C. Schmid. “Epicflow: Edge-preserving interpolation of correspondences for optical flow”. In: Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition. 2015, pp. 1164–1172.
[38] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei. “Large-scale
Video Classification with Convolutional Neural Networks”. In: CVPR. 2014.
[39] J. Revaud, P. Weinzaepfel, Z. Harchaoui, and C. Schmid. “EpicFlow: Edge-Preserving Interpolation of Correspondences for Optical Flow”. In: Computer Vision and Pattern Recognition.
2015.
11
| 1cs.CV
|
A Comparison of Stealthy Sensor Attacks on Control Systems
arXiv:1710.02597v1 [cs.SY] 6 Oct 2017
Navid Hashemi1 , Carlos Murguia2 , and Justin Ruths1
Abstract— As more attention is paid to security in the context
of control systems and as attacks occur to real control systems
throughout the world, it has become clear that some of the
most nefarious attacks are those that evade detection. The
term stealthy has come to encompass a variety of techniques
that attackers can employ to avoid detection. Here we show
how the states of the system (in particular, the reachable set
corresponding to the attack) can be manipulated under two
important types of stealthy attacks. We employ the chi-squared
fault detection method and demonstrate how this imposes
a constraint on the attack sequence either to generate no
alarms (zero-alarm attack) or to generate alarms at a rate
indistinguishable from normal operation (hidden attack).
I. INTRODUCTION
For many decades, Control Theory operated in a challenging but happy place in which problems pitted designers
against the world, a haphazard place of disturbances and
uncertainty. The past decade has seen the rise in concern
over attacks on control systems, which necessarily requires
us to shift our focus to a problem of designer against attacker,
a strategic and knowledgeable entity that seeks to exploit the
weaknesses of our systems and control frameworks.
Control systems have become an attractive target to attackers due to accessibility, impact, and obfuscation. Largescale control systems such as process control plants are
increasingly moving toward Ethernet-like technology to communicate data throughout the system. This new architecture
provides new capabilities but also opens systems up to the
same types of cyber attacks that banking and database companies endure. These systems also represent major industry
or municipal infrastructure, which means damaging them
makes large impact. Finally, these systems are large and
complex enough - and often not monitored well enough - for
attackers to manipulate the system without being detected.
The literature of attack detection has concerned itself with
designing methods to effectively monitor systems and detect
anomalies [1]-[6]. The origin of many of these methods
arise from fault detection, but have been retooled to consider
antagonistic and strategic “faults”. A key component of this
body of work is to understand the limits of these detectors,
and identifying attacks that are stealthy to these methods
is a critical way to benchmark detector performance. The
*This work was partially supported by the National Research Foundation
(NRF), Prime Minister’s Office, Singapore, under its National Cybersecurity
R&D Programme (Award No. NRF2014NCR-NCR001-40) and administered by the National Cybersecurity R&D Directorate.
1 These authors are with the Departments of Mechanical and Systems
Engineering at the University of Texas at Dallas, Richardson, Texas, USA
Navid.Hashemi, jruths @utdallas.edu
2 C. Murguia is with the iTrust Centre at the Singapore University of Technology and Design, Singapore murguia [email protected]
term stealthy has taken on several meanings in the literature.
It has been used to address attacks that do not induce the
detectors to raise alarms; we rename these zero-alarm attacks
to be more precise [1], [7], [8]. It has also referred to an
attack that changes the alarm rate of the detector by only a
small amount; we call these perturbation attacks [9], [10].
We define a hidden attack which exactly mimics the alarm
rate of the detector. Stealthy is also used to describe attacks
that effect the uncontrollable and unobservable modes of the
system and, therefore, do not propagate to any measurement
or estimated state of the system [2]. Replay attacks also
fall into the category of stealthy attacks as they replay past
(recorded) data back to the monitoring equipment [11].
The attacks on unobservable/uncontrollable modes and
replay attacks completely circumvent the detectors, which
is interesting and relevant to the broader context of security,
but requires a countering strategy that goes beyond detectors.
The perturbation attack has a relatively small effect on the
system compared with zero-alarm and hidden attacks, thus
we also omit it from this study.
We use this manuscript to present a distribution-based
perspective on attack detection. While this in and of itself is
not novel, this way of looking at and describing attacks has
yet to be captured clearly in the literature. As part of this we
present a equitable comparison between the impact of zeroalarm and hidden attacks. We use the set of states reachable
by the system when driven by the attacker input as a metric
for this comparison. To achieve this, we present several novel
results on techniques to formulate and algorithms to find
ellipsoidal outer bounds on the reachable sets of the system
corresponding to attacks.
II. BACKGROUND
In this work, we study stochastic discrete-time linear timeinvariant (LTI) systems
(
xk+1 = F xk + Guk + vk ,
(1)
yk = Cxk + ηk ,
in which the state xk ∈ Rn , k ∈ N, evolves due to the
state update provided by the state matrix F ∈ Rn×n , the
control input uk ∈ Rm filtered by the input matrix G ∈
Rn×m , and the i.i.d. zero-mean Gaussian system noise vk
with covariance matrix R1 . The output yk ∈ Rp aggregates
a linear combination, given by the observation matrix C ∈
Rp×n , of the states and zero-mean Gaussian measurement
noise with covariance matrix R2 . We assume that the pair
(F, C) is detectable and (F, G) is stabilizable.
In this work, we consider the scenario that the actual
measurement yk can be corrupted by an additive attack,
δk ∈ Rp . At some point in the process of measuring and
transmitting the output to the controller the attacked output
becomes
ȳk = yk + δk = Cxk + ηk + δk .
(2)
If the attacker has access to the measurements, then it is
possible for the attack δk to cancel some or all of the original
measurement yk - so an additive attack can achieve arbitrary
control over the “effective” output of the system.
As our approach leverages a fault-detection approach, we
require an estimator of some type to produce a prediction
of the system behavior. In this work we use the steady state
Kalman filter
x̂k+1 = F x̂k + Guk + L(ȳk − C x̂k ),
(3)
where x̂k ∈ Rn is the estimated state. The observer gain L
is designed to minimize the steady state covariance matrix
P := limk→∞ Pk := E[ek eTk ] in the absence of attacks,
where ek := xk − x̂k denotes the estimation error. Existence
of P is guaranteed since the pair (F, C) is assumed to be
detectable [12]. Next, we define the residual sequence rk
rk := ȳk − C x̂k ,
alarm time(s) k ∗ are produced. The Σ−1 factor in the definition of zk rescales the distribution (E[zk ] = p, E[zk zkT ] =
2p) so that the threshold α can be designed independent of
the specific statistics of the noises vk and ηk ; instead, it can
be selected simply based on the number of sensors (i.e., the
dimension of the output, p).
It is important to note that because of the infinite support
of noises vk and ηk , the distance measure zk , distributed
according to a chi-squared distribution, also has infinite
support. Therefore, even in the absence of attacks, we expect
that the detector will generate alarms because some values
drawn from the distance measure distribution will exceed
the threshold α. Such alarms in the absence of an attack are
called false alarms. Because we can characterize the chisquared distribution analytically, we have an exact relation
between the choice of the threshold α and the expected rate
of false alarms A generated by the chi-squared detector.
Lemma 1: [8]. Assume that there are no attacks to the
system and consider chi-squared detector, with threshold α ∈
R>0 , rk ∼ N (0, Σ). Let α = α∗ := 2P −1 (1−A∗ , p2 ), where
P −1 (·, ·) denotes the inverse regularized lower incomplete
gamma function, then A = A∗ .
(4)
A. Undetected Attacks
the difference between what we actually receive (ȳk ) and
expect to receive (C x̂k ), which evolves according to
(
ek+1 = F − LC ek − Lηk + vk − Lδk ,
(5)
rk = Cek + ηk + δk .
In the absence of attacks (i.e., δk = 0), it is straightforward
to show that the rk random variable falls according to a zero
mean Gaussian distribution with covariance [8]
Σ = E[rk rkT ] = CP C T + R2 .
(6)
In this work, we consider only one detector, the popular
chi-squared detector. Although other alternatives exist, the
chi-squared is easily the dominant choice for most research
and it also provides a transparent choice to highlight the
key messages we wish to communicate in this work. Similar
analysis can be done with these other detector choices using
attacks derived in our other work [7], [8], [13]. In the case
of the chi-squared detector, a quadratic distance measure zk
is created to be sensitive to changes in the variance of the
distribution as well as the expected value,
zk = rkT Σ−1 rk .
(7)
Since rk ∼ N (0, Σ), the zk random variable, as the sum
of the squares of normally distributed random variables,
falls according to the chi-square distribution. Since rk ∈
Rp , this chi-squared distribution has p degrees of freedom.
The chi-squared detector is summarized as follows: for
given a threshold α ∈ R>0 and the distance measure
zk = rkT Σ−1 rk
(
zk ≤ α −→ no alarm,
(8)
zk > α −→ alarm: k ∗ = k,
Some of the most insidious attacks on industrial control
systems feature attack strategies that manipulate the system
while all the time staying undetected. The effect of the attack
can aggregate during this “stealthy” execution of the attack
and the damage caused by the attack can spread. If obvious
attacks were used, single components might be damaged,
but it would give operators the opportunity to react in time
to prevent further damage. When attacks are undetected,
single damaged components might lead to other components
being damaged without operators realizing the changes to
the system. Past attacks on industrial control systems seem
to favor these undetected attacks, such as the famous Stuxnet
worm incident [14].
In many industrial settings fault detection is accomplished
simply by assigning a collection of static rules (e.g., if a
pressure in a vessel exceeds a given value). These offer littleto-no protection against stealthy adversarial attacks as the
attack can deviate the actual system state while reporting
a state that is within normal operating conditions. When
detectors are implemented in control systems, these detectors
limit what the attacker is able to accomplish if he/she
seeks to remain undetected. We advance two notions of
undetected attacks (we phrase these with respect to the
chi-squared detector, however, the concept of these attack
classes generalize to other detectors). These attack models
require strong attacker knowledge and access, namely we
assume that the attacker has perfect knowledge of the system
dynamics, the Kalman filter, control inputs, measurements,
and chi-squared procedure. In addition, the attacker has read
and write access to all the sensors at each time step. The
goal of these stealthy attacks is to construct a worst case
scenario. In the same spirit of designing buildings for a 1000year earthquake, we aim to design the control infrastructure
against a strong opponent. If designers and operators are
comfortable with the security performance given this kind
of strong attacker, they will also accept the performance for
less powerful attackers.
chi-squared with
2 degrees of freedom
1) Zero-alarm attacks generate attack sequences that
maintain the distance measure at or below the threshold, i.e., zk ≤ α. These attacks generate no alarms
during the attack. To satisfy this condition we define
the attack as
zero alarm attack
1
δk = −Cek − ηk + Σ 2 δ̄k ,
hidden attack
(9)
where δ̄k ∈ Rp is any vector such that δ̄kT δ̄k ≤ α
1
and Σ 2 is the symmetric square root of Σ (recall
the attacker has read access to the sensor, yk , and
knowledge of the estimator, x̂k ). This attack sequence
leads the distance measure to become
zk = rkT Σ−1 rk
= (Cek + ηk + δk )T Σ−1 (Cek + ηk + δk )
1
1
= (Σ 2 δ̄k )T Σ−1 (Σ 2 δ̄k ) ≤ α.
(10)
Since zk ≤ α, no alarms are raised. A schematic
of a zero-alarm attack is shown in Fig. 1. Although
generating no alarms seems like a successful strategy
to avoid detection, it is important to remember that
in the attack-free case alarms are raised due to the
infinite support of the distance measure distribution.
Thus, before the attack, alarms are raised at a rate
A > 0 and after the attack the alarm rate becomes
zero, A = 0. While the detector does not monitor
changes in the false alarm rate, it is possible that an
operator might notice this discrepancy. This leads us
to develop a second class of undetectable attacks. We
are also motivated to develop the following attacks
because they exploit the stochasticity to inject larger,
more potent attacks.
2) Hidden attacks generate attack sequences that raise
alarms at the same rate as the false alarm rate of the
detector (i.e., alarms are raised at the same rate during
the attack as are false alarms in the attack-free case).
In hidden attacks, the attack sequence δ̄k in (9) is a
random variable designed such that
Pr(zk > α) = Pr(δ̄kT δ̄k > α) = A.
(11)
In other words, on average out of N time steps:
(1 − A)N time steps δ̄kT δ̄k ≤ α and the remaining
AN time steps δ̄kT δ̄k > α. The chi-squared detector
tuned to a false alarm rate A effectively splits the zk
distribution into a part zk ≤ α and a part zk > α,
where α is selected using Lemma 1. Hidden attacks
ensure that the proportion of zk values larger than
α observed by the detector during the attack match
the proportion expected in the attack-free case [15]. A
schematic of a hidden attack is shown in Fig. 1.
Fig. 1. The original (attack-free) zk distribution (top) is chi-squared with
p degrees of freedom (this paper uses examples in which p = 2). The
threshold α is selected to satisfy a false alarm rate of A, implying that in
the attack-free distribution, the area under the distribution curve that falls
beyond α is A. In zero-alarm attacks (middle), δ̄k is selected such that zk
is no larger than α, implying that no alarms are raised under zero-alarm
attacks. In hidden attacks (bottom), δ̄k is designed so that the fraction of the
distribution that falls beyond α matches that of the attack-free distribution,
which means that the alarm rate under the hidden attack is equal to the
false alarm rate. The definition of the zero alarm and hidden attacks do not
stipulate the shape of the density functions above and below α, although
the allocation of mass in the density function greatly influences the effect
of the attack on the reachable states.
B. Feedback
In order for the attack to propagate from the estimation
error to the state, we need to incorporate a model of
feedback in the control system. In this paper we assume
static estimator feedback uk = K x̂. With this feedback the
closed-loop system becomes
(
xk+1 = (F + GK)xk + GKek + vk ,
(12)
ek+1 = (F − LC)ek − Lδk − Lηk + vk .
The estimation error updates according to, without attacks,
ek+1 = (F − LC)ek − Lηk + vk ,
(13)
and with attacks of the form in (9),
1
ek+1 = F ek − LΣ 2 δ̄k + vk .
(14)
Remark 1: Note that if the spectral radius ρ[F ] > 1,
then kE[ek ]k (and also kE[xk ]k due to the interconnection)
diverges to infinity as k grows for any non-stabilizing k.
That is, attacks of the form (9) may destabilize the system
if ρ[F ] > 1. If ρ[F ] ≤ 1, then kE[ek ]k may or may not
diverge to infinity depending on algebraic and geometric
multiplicities of the eigenvalues on the unit circle. Thus we
consider open-loop stable system matrices, ρ[F ] < 1.
III. REACHABLE SET BOUNDS
In order to compare the effects of these different stealthy
attacks, we require a metric to quantify the impact of each
attack. A popular choice to quantify system impact due to a
disturbance is the set of states reachable by the action of the
disturbance. Here, we show two techniques to derive outer
Fig. 2. The p̄-probable ellipsoid captures a level set of the distribution of
reachable states corresponding to when the system is driven by a truncated
distance measure, such that zk ≤ z̄, where Pr(zk ≤ z̄) = p̄.
ellipsoidal bounds on the reachable states. The first, based
on Linear Matrix Inequalities (LMIs), constructs a convex
optimization problem, the solution of which is the ellipsoid
that bounds the states driven by attacks. This approach
provides more conservative estimates of the reachable set, but
allows for the opportunity to simultaneously design system
components, such as estimator and controller gain matrices,
to reduce the size of the reachable set (see, e.g., [10], [15]).
The second approach provides extremely tight bounds for
the reachable set through the use of geometric ellipsoidal
methods.
The different definitions of the zero alarm attack and the
hidden attack naturally give rise to different reachable sets.
The challenge of the hidden attack definition is that there are
no constraints on the location of the A mass that falls beyond
α. This means that A of the probability density function of
the distance function can be made arbitrarily large, which
in turn makes the reachable sets driven by hidden attacks
arbitrarily large. Therefore, it is not meaningful to define the
outer bound of the reachable set corresponding to hidden
attacks - it would simply be the entire state space. Instead,
we introduce the notion of a p̄-probable ellipsoid which
encompasses the reachable set when the zk distribution is
truncated at z̄, where Pr(zk ≤ z̄) = p̄. This ellipsoid can
be interpreted graphically as a level set of the distribution
function of the reachable states (see Fig. 2). It is worth noting
that the probability p̄ corresponds to the truncation of the zk
distribution and does not specify the probability that a point
in the true reachable state set is in the p̄-probable ellipsoid.
While the later probability is closer to what we want to
know, this would require knowing the complete distribution
of reachable states which is what we aim to find in the first
place. Notwithstanding, there is a one-to-one mapping from
p̄ to level sets of the reachable set distribution, so increasing
(resp., decreasing) p̄ necessarily expands (resp., shrinks) the
ellipsoid level set, so this is not much of a restriction.
While we consider more general choices of p̄ in other work
(see [15]), here we focus on the most immediate choice p̄ =
1 − A and so z̄ = α. Recall that the hidden attack (see Fig.
1 and (11)) only requires the attacker to satisfy one statistic
of the attack δ̄k , namely Pr(zk = δ̄kT δ̄k ≤ α) = 1 − A. For a
general hidden attack, we have no further information about
the distribution of zk (recall the shape of the distribution,
beyond this one constraint, is completely free for the attacker
to choose, see Fig. 1). Thus the only choice of p̄ that can be
evaluated for a general hidden attack is p̄ = 1 − A. This also
simplifies the comparison of hidden attacks with zero alarm
attacks. Selecting z̄ = α as the truncation point of the zk
distribution implies then that we truncate the attacks such that
zk = δ̄kT δ̄k ≤ α. This truncation to quantify the p̄-probable
ellipsoidal bound for the reachable set due to hidden attacks
now imposes the same constraint that exists in the case of
zero alarm attacks, see (10).
When we look at the complete reachable state of the
system, we can decompose the contributions due to system
noise and due to attack separately. Using the superposition
principle of linear systems, the estimation error ek can be
written as ek = evk + eδk , where evk denotes the part of ek
driven by noise and eδk is the part driven by attacks. We can
now write the estimation error dynamics in (14) as follows,
where we assume the attack starts at k = k ∗ ,
evk+1 = F evk + vk ,
1
2
eδk+1 = F eδk − LΣ δ̄k ,
evk∗ = ek∗
(15)
eδk∗ = 0.
(16)
Similarly, the state of the system xk can be written as xk =
xvk + xδk , where xvk denotes the part of xk driven by noise
and xδk is the part driven by attacks. Using this new notation,
we can write the system dynamics in (12) as follows,
xvk+1 = (F + GK)xvk − GKevk + vk ,
xvk∗ = xk∗
(17)
xδk+1
xδk∗
(18)
= (F +
GK)xδk
−
GKeδk ,
= 0.
With these definitions, there are two reachable sets we
aim to identify: the reachable states due to noise and due to
attack. Because the state equation depends on the estimation
error, in general, we must first identify the reachable estimation error due to noise and due to attack. Interestingly, the
noise equations (15) and (17) have a special symmetry due
to the zero initial conditions, i.e., xv1 = ev1 = 0. By writing
out evk and xvk for each k = 1, 2, . . . , it quickly becomes
clear that xvk = evk for all k ∈ N. Thus, for the contribution
driven by noise, we need only to solve the evk equation. The
reachable set of the estimation error driven by noise equals
the reachable set of the states driven by noise.
Notice that the noise, a multivariate Gaussian distribution,
also has unbounded support, thus it also has an infinite
reachable set. We use the notion of a p̄-probable reachable set
to define a finite reachable set; for an equitable contribution
by noise and attack, we again select p̄ = 1 − A and truncate
the distribution with v̄ such that
Pr(vkT R1−1 vk ≤ v̄) = p̄ = 1 − A,
(19)
where R1 is the covariance matrix of the system noise vk .
Since vkT R1−1 vk is a chi-squared random variable with n
degrees of freedom, the value of v̄ can be determined by
Lemma 1. Thus for noises, k ∈ N,
Rvx = Rve = evk ∈ Rn (15), vkT R1−1 vk ≤ v̄ .
(20)
For attacks, ∀ k ≥ k ∗ ,
Rδe = eδk ∈ Rn (16), δ̄kT δ̄k ≤ α ,
Rδx = xδk ∈ Rn (18), eδk ∈ Rδe .
(21)
(22)
A. LMI Approach
In general, it is analytically intractable to compute a reachable set R exactly. Instead, using Linear Matrix Inequalities
(LMIs), for some positive definite matrix P, we derive outer
ellipsoidal bounds of the form E = {ξk | ξkT Pξk ≤ 1} containing R. Our LMI results leverage the following lemma;
this approach parallels work in [10], [15] however, the attack
definitions are different, so they should be reformulated here.
Lemma 2: [16]. Let Vk be a positive definite function,
V1 = 0, and ζkT ζk ≤ κ ∈ R>0 . If there exists a constant
a ∈ (0, 1) such that the condition below holds, then Vk ≤ 1:
1−a T
ζ ζk ≤ 0.
(23)
κ k
We present a generic solution to identify the outer bounding ellipsoids we need. We consider a linear system driven by
an input that is elliptically bounded, which, as we will show,
represent the dynamics in (15)-(18) and the corresponding
constraints in (20)-(22).
Proposition 1: Given a LTI system ξk+1 = Aξk + Bµk ,
A ∈ Rn×n and B ∈ Rn×p , with the constraint µTk Rµk ≤
1, R > 0, for all k ∈ N, if there exists a ∈ (0, 1) and
positive definite matrix P ∈ Rn×n that solves the convex
optimization,
min − log det P,
P
s.t. P > 0, and
(24)
T
T
aP
−
A
PA
−A
PB
≥ 0,
−B T PA
(1 − a)R − B T PB
Vk+1 − aVk −
then the reachable states R ⊆ E = {ξk ∈ Rn | ξkT Pξk ≤ 1}
and the ellipsoid E has minimum volume.
1
Proof: Let Vk = ξkT Pξk and ζk = R 2 µk in (23) in
1
Lemma 2, where R 2 is the symmetric square root of the
positive definite matrix R. It is easy to confirm that ζkT ζk =
1
1
(R 2 µ)T (R 2 µ) = µTk Rµk ≤ 1 with κ = 1. Substituting
the dynamic equation for ξk+1 in Vk+1 yields an expression
that when factored into quadratic form νkT Qνk ≥ 0, with
νk = [ξkT , µTk ]T , the matrix Q is the LMI in the optimization
problem above. Thus the bounding ellipsoid is given by E =
{ξk | Vk = ξkT Pξk ≤ 1}.
To ensure that the ellipsoid bound is as tight as possible,
1
we minimize (det P)− 2 since this quantity is proportional
to the volume of E. We instead minimize log det P −1 as
it shares the same minimizer and because for P > 0 this
objective is convex [17].
We now use this generic result to outer bound the four
reachable sets we need (In is the n × n identity matrix).
Theorem 1: The reachable sets
Rve = Rvx , Rδe , and Rδx ,
are contained in the minimum volume ellipsoids
Eev = Exv , Eeδ , and Exδ ,
respectively, characterized by the positive definite matrices
Pev = Pxv , Peδ , and Pxδ ,
respectively, which are the solutions to the convex optimization in Proposition 1 with to the following choices of A, B,
and R, respectively:
•
Pev = Pxv : A = F , B = In , R = v̄1 R1−1 ,
•
Peδ : A = F , B = −LΣ 2 , R =
•
Pxδ : A = F + GK, B = −GK, R = Peδ .
1
1
α Ip ,
Proof: The proofs of each case are quite similar. We
prove the case for Pxδ and the rest follow a same pattern. In
(22), we identified that eδk ∈ Rδe . Instead we impose eδk ∈ Eeδ .
These sets are not equal, but since the ellipsoid contains the
reachable set, this still satisfies the requirement of (22) - it
does so with extra conservatism by also including estimation
errors eδk ∈ Eeδ \ Rδe .
Setting A = F + GK is straightforward comparing (18)
to Proposition 1. Define the input vector µk = eδk such that
µTk Rµk = µTk Peδ µk ≤ 1,
(25)
since (eδk )T Peδ eδk ≤ 1 by the definition of the ellipsoid Eeδ .
With this definition of µk , the corresponding input matrix
that satisfies (18) is B = −GK.
Having derived the set of reachable states due to noise
and due to attack, both bounded by ellipsoids, we now
compose these together to yield the total reachable set of
states. The superposition of two ellipsoidal sets has been
studied extensively and labeled the geometric (Minkowski)
sum such that E1 ⊕ E2 = {x + y | x ∈ E1 , y ∈ E2 }.
The complete reachable set is then Ex = Exv ⊕ Exδ . It is
possible to compose another convex optimization and LMI
to combine the ellipsoids [15], the geometric sum provides a
tighter resulting ellipsoid. When these techniques are used to
design controller and estimator gains, optimization methods
are preferred, but in this work we do not follow this line of
inquiry.
B. Geometric Approach
A geometric approach to finding the ellipsoidal bounds for
the reachable set of states comes from the observation that
the equations in (15)-(18) contain inputs that are ellipsoidally
bounded, i.e., v̄1 vkT R1−1 vk ≤ 1 and α1 δ̄kT δ̄k ≤ 1 (in fact these
are spherically bounded). The geometric sum introduced
above provides an operation that simultaneously computes
all possible combinations between two geometric sets. For
example, the dynamics for k ∈ N,
ξk+1 = Aξk + Bµk ,
ξ1 = 0,
with µTk Rµk ≤ 1, (26)
can be interpreted as an ellipsoidal update. For k = 1,
ξ2 = Aξ1 + Bµ1 .
(27)
The µT1 Rµ1 ≤ 1 bound identifies that any possible value
of µ1 belongs to an ellipse µ1 ∈ {µ | µT Rµ ≤ 1}.
Many ellipsoid calculations are more concise when the
ellipsoid is characterized by its shape matrix, Q, E(Q) =
{µ | µT Q−1 µ ≤ 1}. With this definition it is easy to express
the linear transformation of an ellipse: if ξ = M µ and
µ ∈ E(Q), then ξ ∈ E(M QM T ) [18]. Thus in this example
µ1 ∈ E(R−1 ) and ξ2 ∈ E(BR−1 B T ). Continuing,
ξ3 = Aξ2 + Bµ2 ,
−1
T
(28)
µT2 Rµ2
−1 T
where ξ2 ∈ E(BR B ) and
≤ 1. In this case
ξ3 ∈ E(ABR−1 B T AT ) ⊕ E(BR B ), where ⊕ represents
the geometric sum. Although the geometric sum of two ellipsoids is not necessarily an ellipsoid, there are straightforward
techniques to tightly fit an ellipsoid around the resulting
shape (see details in [18]) so we will consider for the rest of
this paper that the geometric sum of two ellipsoids produces
an ellipsoid. It is worth noting that this fitting does embed
an element of conservatism in the result due to the fitting;
therefore, we will minimize the number of times the fitting
needs to occur in our proposed algorithm. It is also important
to note that the geometric sum of two ellipsoids is only valid
if the two ellipsoids are independent. Here each ellipsoid
corresponds to a different realization µk .
We see now that all possible values that ξk can take
on will belong to an ellipsoid, and one that is iteratively
updated along the lines of the discussion above. Now we
will specialize these observations to the context here to find
Ex . We take the same approach as in the LMI method by
splitting the dynamics into a contribution driven by noise
and a contribution driven by the attack, such that again
Ex = Exv ⊕ Exδ .
Theorem 2: Given the estimation and state equations (15)
and (17) for the system driven by noise, the ellipse Exv
contains all possible values of xvk , where,
Exv =
∞
M
E v̄F k R1 (F k )T .
(29)
k=0
In other words Rvx ⊆ Exv .
N
−1
X
F N −1−k vk .
(30)
k=1
Note that vi and vj are independent and equivalently
bounded. Therefore, the terms of (30) have identical ellipsoids, E(v̄R1 ) transformed by different powers of F ,
ExvN =
=
N
−1
M
k=1
N
−2
M
Since ρ[F ] < 1, det F < 1 and the volume goes to zero as
k becomes large. The practical application of this is that
one can simply take N terms of the limiting geometric
sum in (29) to achieve an accurate approximation of the
bounding ellipsoid. The convergence of this sum (and hence
the number of terms that should be chosen) depends on the
spectrum of F .
Theorem 3: Given the estimation and state equations (16)
and (18) for the system driven by attack, the ellipse Exδ
contains all possible values of xδk , where,
Exδ =
∞
M
E αHLΣLT H T ,
(33)
k=1
where H = (F + GK)k − F k . In other words Rδx ⊆ Exδ .
Proof: Expanding the recursive definitions in (18) and
substituting in (16), we find
xδN =
N
−2
X
1
(F + GK)N −1−k − F N −1−k LΣ 2 δk . (34)
k=1
The rest of the proof follows the same line as the proof of
Theorem 2 and so we omit the details.
Remark 3: Similarly, ρ[F + GK] < 1 because the controller matrix is selected to make the closed-loop system
stable. Thus Theorem 3 benefits from the same practical
advantage of constructing a good approximation of the
ellipsoid Exδ with finitely many terms.
IV. E MPIRICAL R EACHABLE S ETS
Proof: Recall for the contribution driven by noise we
need only to solve the evk equation (15). Expanding the
recursive definition we can express evN (xvN ) in terms of all
the past terms,
xvN = evN =
(proportional to the determinant of their shape matrix) with
higher powers of F become vanishingly small,
det v̄F k R1 (F k )T = v̄ det(R1 )(det F )2k .
(32)
E F N −1−k (v̄R1 )(F N −1−k )T ,
E v̄F k R1 (F k )T .
(31)
k=0
The ellipsoid that bounds all possible trajectories is the
limiting ellipsoid as N goes to infinity.
Remark 2: We assume the matrix F is stable, otherwise
the attacker can easily achieve arbitrarily large reachable
sets simply by decoupling the controller from the openloop system. Because of this, the volume of the ellipsoids
We now demonstrate these tools and provide a comparison between zero-alarm and hidden attacks. Although we
generate a common bounding ellipsoid for both attacks, we
also run extensive Monte-Carlo simulations to derive an
approximation of the empirical reachable set the ellipsoids
are meant to bound. We consider the following system for
this study with the chi-squared detector tuned to a false alarm
rate A = 0.05 (5%):
0.84
0.23
0.07 −0.32
1 0
, G=
, C=
,
−0.47 0.12
0.23
0.58
2 1
0.045
−0.011
1.404 −1.042
R1 =
, K=
,
−0.011
0.02
1.842
1.008
0.0276
0.0448
2 0
2.086 0.134
L=
, R2 =
, Σ=
.
−0.01998 −0.0290
0 2
0.134 2.230
F =
Fig. 1 clearly shows the ambiguity in designing zero
alarm attacks and hidden attacks. In a zero alarm attack,
the density function can be arbitrarily shaped on zk ≤ α.
For simplicity, consider that we use a uniform distribution
of width w1 and centered at zk = c1 , such that the support
of the distribution is over [c1 − w21 , c1 + w21 ]. In other work,
we have shown that in terms of steady-state deviation of
3
3
Fig. 3. The empirical reachable state sets in black with ellipsoidal bounds derived by the LMI approach (blue) and geometric approach (red) for zero
alarm attacks (a) ZA.A, (b) ZA.B, and (c) ZA.C, as well as hidden attacks (d) H.A, (e) H.B, (f) H.C, and (g) H.D.
TABLE I
PARAMETERS FOR ZERO ALARM (ZA) AND HIDDEN (H) ATTACKS .
c1
w1
c2
w2
Fig. 4. The volume (blue is larger volume) of the reachable sets for different
zero alarm zk distributions of the form shown in Table I.
the state, there exists a magnitude and “direction” of δ̄k that
yields the strongest attack [13]. It is intuitive that maximizing
the norm of the attack δ̄kT δ̄k = α leads to stronger attacks
than δ̄kT δ̄k < α. While we show an analytic result for the
steady state deviation of the state [13], showing the same
for the reachable sets is more nuanced. To demonstrate that
this intuition holds empirically, we calculate the volume of
the empirical reachable set attained through simulation. We
approximate the volume by fitting an ellipsoid to the point
cloud, where the lengths of the principle axes
√ in
√ are given
terms of the eigenvalues of the data, as 1/ λ1 and 1/ λ2 .
The volumes are plotted in Fig. 4 for different choices of w1
and c1 and clearly shows the largest volume ellipsoids (dark
blue) are generated by attacks for which c1 = α and w1 ≈ 0.
We select three sets of values for the pair (w1 , c1 ) labeled
ZA.A, ZA.B, and ZA.C for our comparison (see Table I).
ZA.A
α/8
α/10
-
ZA.B
α/2
α
-
ZA.C
α
0
-
H.A
α
0
1.5α
α
H.B
α
0
2α
0
H.C
α
0
10α
0
H.D
α
0
100α
0
For hidden attacks, there are two regions to define (below
and beyond zk = α). Based on our observations (and
intuition), we set the 1 − A portion of the distribution that
falls at or below α as a point mass at α. From Fig. 4,
an attacker who wishes to maximize their influence on the
reachable set would naturally make this choice. As discussed
before, the second mass lies beyond α and could theoretically
cause arbitrarily large reachable sets. Here we select the
A mass in four different configurations (parameterized by
a second uniform distribution section centered at c2 and
with width w2 , see Table I): spread uniformly from (α, 2α]
(labeled H.A), a point mass at 2α (H.B), a point mass at 10α
(H.C), and a point mass at 100α (H.D).
In Fig. 3, we display the empirical reachable sets for all
seven of these attacks as well as the LMI (blue) and geometric (red) outer ellipsoidal bounds derived with our methods.
We first observe that both techniques are able to rather tightly
bound the reachable set of states due to zero alarm attacks,
although the geometric approach provides slightly tighter
ellipsoid bounds. For hidden attacks, while it takes high
magnitude attacked zk values (e.g., c2 = 10α, 100α) to
see a distinct growth in the volume of the reachable set, it
is possible to grow the reachable set arbitrarily large.
The more substantial takeaway from this empirical study is
that when we use conventional detectors that use a single cut
in the distribution to determine if the current zk is more likely
to come from the original attack-free distribution or some
other (attacked) distribution, we lack the ability to constrain
the attacker due to the A fraction of the distribution that
falls beyond the detector threshold α. We require either a
combination of detectors or modified definitions of current
detectors to synthesize the information necessary to limit attackers further. When attackers hide in the infinite support of
the noise, as in a hidden attack, we require some mechanism
to effectively truncate or bound the impact of an attacker.
Some obvious solutions are available, such as enforcing
finite support of all noises, however, these approaches have
not been integrated into conventional detector methods. In
addition saying a disturbance as finite support is different
from practically using this assumption; this gap must be
addressed before this type of approach could be used.
V. CONCLUSION
We have presented a thorough exposition of the current
ideology on using fault detection type detectors for identification of (sensor) attacks on control systems. In particular, we
compared two attacks in which the opponent aims to remain
stealthy - one in which the attack sequence is generated
so as to not raise any alarms and one in which the attack
sequence raises alarms at the same rate they occur randomly
in the absence of attacks. We developed two approaches
to determine ellipsoidal and p̄-probable ellipsoidal bounds
(when the reachable set is infinite) on the reachable states
of the system in response to the attack and to the inherent
system noise. We demonstrated these concepts and methods
with a numerical example that emphasizes the need for work
that goes beyond traditional detectors.
R EFERENCES
[1] A. Cárdenas, S. Amin, Z. Lin, Y. Huang, C. Huang, and S. Sastry, “Attacks against process control systems: Risk assessment, detection, and
response,” in Proceedings of the 6th ACM Symposium on Information,
Computer and Communications Security, 2011, pp. 355–366.
[2] F. Pasqualetti, F. Dorfler, and F. Bullo, “Attack detection and identification in cyber-physical systems,” IEEE Transactions on Automatic
Control, vol. 58, pp. 2715–2729, 2013.
[3] Y. Mo, E. Garone, A. Casavola, and B. Sinopoli, “False data injection
attacks against state estimation in wireless sensor networks,” in Decision and Control (CDC), 2010 49th IEEE Conference on, 2010, pp.
5967–5972.
[4] C. Kwon, W. Liu, and I. Hwang, “Security analysis for cyber-physical
systems against stealthy deception attacks,” in American Control
Conference (ACC), 2013, 2013, pp. 3344–3349.
[5] F. Miao, Q. Zhu, M. Pajic, and G. J. Pappas, “Coding sensor outputs
for injection attacks detection,” in Decision and Control (CDC), 2014
IEEE 53rd Annual Conference on, 2014, pp. 5776–5781.
[6] C. Z. Bai, F. Pasqualetti, and V. Gupta, “Security in stochastic
control systems: Fundamental limitations and performance bounds,”
in American Control Conference (ACC), 2015, 2015, pp. 195–200.
[7] C. Murguia and J. Ruths, “Characterization of a cusum model-based
sensor attack detector,” in proceedings of the 55th IEEE Conference
on Decision and Control (CDC), 2016.
[8] ——, “Cusum and chi-squared attack detection of compromised
sensors,” in proceedings of the IEEE Multi-Conference on Systems
and Control (MSC), 2016.
[9] Y. Mo and B. Sinopoli, “On the performance degradation of cyberphysical systems under stealthy integrity attacks,” IEEE Transactions
on Automatic Control, vol. 61, pp. 2618–2624, 2016.
[10] C. Murguia, N. van de Wouw, and J. Ruths, “Reachable sets of hidden
cps sensor attacks: Analysis and synthesis tools,” in proceedings of the
IFAC World Congress, 2016.
[11] Y. Mo and B. Sinopoli, “Secure control against replay attacks,” in
Communication, Control, and Computing, 2009. Allerton 2009. 47th
Annual Allerton Conference on, 2009, pp. 911–918.
[12] K. J. Aström and B. Wittenmark, Computer-controlled Systems (3rd
Ed.). Upper Saddle River, NJ, USA: Prentice-Hall, Inc., 1997.
[13] T. R, C. Murguia, and J. Ruths, “Tuning windowed chi-squared
detectors for sensor attacks,” in eprint arXiv (submitted to ACC2018),
2017.
[14] R. Langner, “Stuxnet: Dissecting a cyberwarfare weapon,” IEEE
Security & Privacy, vol. 9, no. 3, pp. 49–51, 2011.
[15] C. Murguia and J. Ruths, “On reachable sets of hidden cps sensor
attacks,” in eprint arXiv (submitted to ACC2018), 2017.
[16] N. D. That, P. T. Nam, and Q. P. Ha, “Reachable set bounding for
linear discrete-time systems with delays and bounded disturbances,”
Journal of Optimization Theory and Applications, vol. 157, pp. 96–
107, 2013.
[17] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix
Inequalities in System and Control Theory, ser. Studies in Applied
Mathematics. Philadelphia, PA: SIAM, 1994, vol. 15.
[18] A. A. Kurzhanskiy and P. Varaiya, “Ellipsoidal toolbox (et),” in
Decision and Control, 2006 45th IEEE Conference on. IEEE, 2006,
pp. 1498–1503.
| 3cs.SY
|
arXiv:1504.01431v2 [cs.CC] 5 Nov 2015
If the Current Clique Algorithms are Optimal,
so is Valiant’s Parser
Amir Abboud
Stanford University
[email protected]
Arturs Backurs
MIT
[email protected]
Virginia Vassilevska Williams
Stanford University
[email protected]
Abstract
The CFG recognition problem is: given a context-free grammar G and a string w of length
n, decide if w can be obtained from G. This is the most basic parsing question and is a core
computer science problem. Valiant’s parser from 1975 solves the problem in O(nω ) time, where
ω < 2.373 is the matrix multiplication exponent. Dozens of parsing algorithms have been
proposed over the years, yet Valiant’s upper bound remains unbeaten. The best combinatorial
algorithms have mildly subcubic O(n3 / log3 n) complexity.
Lee (JACM’01) provided evidence that fast matrix multiplication is needed for CFG parsing,
and that very efficient and practical algorithms might be hard or even impossible to obtain.
Lee showed that any algorithm for a more general parsing problem with running time O(|G| ·
n3−ε ) can be converted into a surprising subcubic algorithm for Boolean Matrix Multiplication.
Unfortunately, Lee’s hardness result required that the grammar size be |G| = Ω(n6 ). Nothing
was known for the more relevant case of constant size grammars.
In this work, we prove that any improvement on Valiant’s algorithm, even for constant size
grammars, either in terms of runtime or by avoiding the inefficiencies of fast matrix multiplication, would imply a breakthrough algorithm for the k-Clique problem: given a graph on n
nodes, decide if there are k that form a clique.
Besides classifying the complexity of a fundamental problem, our reduction has led us to
similar lower bounds for more modern and well-studied cubic time problems for which faster
algorithms are highly desirable in practice: RNA Folding, a central problem in computational
biology, and Dyck Language Edit Distance, answering an open question of Saha (FOCS’14).
1
Introduction
Context-free grammars (CFG) and languages (CFL), introduced by Chomsky in 1956 [Cho59], play
a fundamental role in computability theory [Sip96], formal language theory [HMU06], programming
languages [ASU86], natural language processing [JM00], and computer science in general with
applications in diverse areas such as computational biology [DEKM98] and databases [KSSY13].
They are essentially a sweet spot between very expressive languages (like natural languages) that
computers cannot parse well, and the more restrictive languages (like regular languages) that even
a DFA can parse.
In this paper, we will be concerned with the following very basic definitions. A CFG G in
Chomsky Normal Form over a set of terminals (i.e. alphabet) Σ consists of a set of nonterminals
T , including a specified starting symbol S ∈ T , and a set of productions (or derivation rules) of
the form A → B C or A → σ for some A, B, C ∈ T and σ ∈ Σ. Each CFG G defines a CFL
L(G) of strings in Σ∗ that can be obtained by starting with S and recursively applying arbitrary
derivation rules from the grammar. The CFG recognition problem is: given a CFG G and a string
w ∈ Σ∗ determine if w can be obtained from G (i.e. whether w ∈ L(G)). The problem is of most
fundamental and practical interest when we restrict G to be of fixed size and let the length of the
string n = |w| to be arbitrary.
The main question we will address in this work is: what is the time complexity of the CFG
recognition problem?
Besides the clear theoretical importance of this question, the practical motivation is overwhelming. CFG recognition is closely related to the parsing problem in which we also want to output a
possible derivation sequence of the string from the grammar (if w ∈ L(G)). Parsing is essential:
this is how computers understand our programs, scripts, and algorithms. Any algorithm for parsing
solves the recognition problem as well, and Ruzzo [Ruz79] showed that CFG recognition is at least
as hard as parsing, at least up to logarithmic factors, making the two problems roughly equivalent.
Not surprisingly, the critical nature of CFG recognition has led to the development of a long
list of clever algorithms for it, including classical works [Val75, Ear70, CS70, You67, Kas65, Knu65,
DeR69, Bak79, Lan74], and the search for practical parsing algorithms, that work well for varied
applications, is far from over [PK09, RSCJ10, SBMN13, CSC13]. For example, the canonical CYK
algorithm from the 1960’s [CS70, Kas65, You67] constructs a dynamic programming table D of size
n × n such that cell D(i, j) contains the list of all nonterminals that can produce the substring of w
from position i to position j. The table can be computed with linear time per entry, by enumerating
all derivation rules A → B C and checking whether for some i ≤ k ≤ j, D(i, k) contains B and
D(k + 1, j) contains C (and if so, add A to D(i, j)). This gives an upper bound of O(n3 ) for the
problem. Another famous algorithm is Earley’s from 1970 [Ear70] which proceeds by a top-down
dynamic programming approach and could perform much faster when the grammar has certain
properties. Variants of Earley’s algorithm were shown to run in mildly subcubic O(n3 / log2 n) time
[GRH80, Ryt85]. 1
In 1975 a big theoretical breakthrough was achieved by Valiant [Val75] who designed a sophisticated recursive algorithm that is able to utilize many fast boolean matrix multiplications to speed
up the computation of the dynamic programming table from the CYK algorithm. The time complexity of the CFG problem decreased to O(g 2 nω ), where ω < 2.373 is the matrix multiplication
1
As typical, we distinguish between “truly subcubic” runtimes, O(n3−ε ) for constant ε > 0, and “mildly subcubic”
for all other subcubic runtimes.
1
exponent [Wil12, Gal14] and g is the size of the grammar; because in most applications g = O(1),
Valiant’s runtime is often cited as O(nω ). In 1995 Rytter [Ryt95] described this algorithm as
“probably the most interesting algorithm related to formal languages” and it is hard to argue with
this quote even today, 40 years after Valiant’s result. Follow-up works proposed simplifications
of the algorithm [Ryt95], generalized it to stochastic CFG parsing [BS07], and applied it to other
problems [Aku98, ZTZU10].
Despite its vast academic impact, Valiant’s algorithm has enjoyed little success in practice.
The theoretically fastest matrix multiplication algorithms are not currently practical, and Valiant’s
algorithm can often be outperformed by “combinatorial” methods in practice, even if the most
practical truly subcubic fast matrix multiplication algorithm (Strassen’s [Str69]) is used. Theoretically, the fastest combinatorial algorithms for Boolean Matrix Multiplication (BMM) run in time
O(n3 / log4 n) [Cha15, Yu15]. To date, no combinatorial algorithm for BMM or CFG recognition
with truly subcubic running time is known.
In the absence of efficient algorithms and the lack of techniques for proving superlinear unconditional lower bounds for any natural problem, researchers have turned to conditional lower bounds
for CFG recognition and parsing. Since the late 1970’s, Harrison and Havel [HH74] observed that
any algorithm for the problem would imply an algorithm that can verify a Boolean matrix multipli√
√
cation of two n × n matrices. This reduction shows that a combinatorial O(n1.5−ε ) recognition
algorithm would imply a breakthrough subcubic algorithm for BMM. Ruzzo [Ruz79] showed that
a parsing algorithm that says whether each prefix of the input string is in the language, could
√
√
even compute the BMM of two n × n matrices. Even when only considering combinatorial
algorithms, these Ω(n1.5 ) lower bounds left a large gap compared to the cubic upper bound. A
big step towards tight lower bounds was in the work of Satta [Sat94] on parsing Tree Adjoining
Grammars, which was later adapted by Lee [Lee02] to prove her famous conditional lower bound
for CFG parsing. Lee proved that BMM of two n × n matrices can be reduced to “parsing” a
string of length O(n1/3 ) with respect to a CFG of size Θ(n2 ), where the parser is required to say
for each nonterminal T and substring w[i : j] whether T can derive w[i : j] in a valid derivation of
w from the grammar. This reduction proves that such parsers cannot be combinatorial and run in
O(gn3−ε ) time without implying a breakthrough in BMM algorithms.
Lee’s result is important, however suffers from significant limitations which have been pointed
out by many researchers (e.g. [Ruz79, Lee02, Sah14a, Sah14b]). We describe a few of these below.
Despite the limitations, however, the only progress after Lee’s result is a recent observation by Saha
[Sah15] that one can replace BMM in Lee’s proof with APSP by augmenting the production rules
with probabilities, thus showing an APSP-based lower bound for Stochastic CFG parsing. Because
Saha uses Lee’s construction, her lower bound suffers from exactly the same limitations.
The first (and most major) limitation of Lee’s lower bound is that it is irrelevant unless the size
of the grammar is much larger than the string, in particular it is cubic only when g = Ω(n6 ). A
CFG whose description needs to grow with the input string does not really define a CFL, and as
Lee points out, this case can be unrealistic in many applications. In programming languages, for
instance, the grammar size is much smaller than the programs one is interested in, and in fact the
grammar can be hardcoded into the parser. A parsing algorithm that runs in time O(g3 n), which is
not ruled out by Lee’s result, could be much more appealing than one that runs in O(gn2.5 ) time.
The second limitation of both Lee’s and Ruzzo’s lower bounds is the quite demanding requirement from the parser to provide extra information besides returning some parse tree. These lower
bounds do not hold for recognizers nor any parser with minimal but meaningful output.
2
Theoretically, it is arguably more fundamental to ask: what is the time complexity of CFG
recognition and parsing that can be obtained by any algorithm, not necessarily combinatorial?
Lee’s result cannot give a meaningful answer to this question. To get a new upper bound for BMM
via Lee’s reduction, one needs a parser that runs in near-linear time. Lee’s result does not rule
out, say, an O(n1.11 ) time parser that uses fast matrix multiplication; such a parser would be an
amazing result.
Our first observation is that to answer these questions and understand the complexity of CFG
recognition, we may need to find a problem other than BMM to reduce from. Despite the apparent
similarity in complexities of both problems - we expect both to be cubic for combinatorial algorithms
and O(nω ) for unrestricted ones - there is a big gap in complexities because of the input size. When
the grammar is fixed, a reduction cannot encode any information in the grammar and can only use
the n letters of the string, i.e. O(n) bits of information, while an instance of BMM requires Θ(n2 )
bits to specify. Thus, at least with respect to reductions that produce a single instance of CFG
recognition, we do not expect BMM to imply a higher than Ω(n1.5 ) lower bound for parsing a fixed
size grammar.
Main Result In this paper we present a tight reduction from the k-Clique problem to the recognition of a fixed CFG and prove a new lower bound for CFG recognition that overcomes all the
above limitations of the previously known lower bounds. Unless a breakthrough k-Clique algorithm
exists, our lower bound completely matches Valiant’s 40-year-old upper bound for unrestricted algorithms and completely matches CYK and Earley’s for combinatorial algorithms, thus resolving
the complexity of CFG recognition even on fixed size grammars.
Before formally stating our results, let us give some background on k-Clique. This fundamental
graph problem asks whether a given undirected unweighted graph on n nodes and O(n2 ) edges contains a clique on k nodes. This is the parameterized version of the famously NP-hard Max-Clique
(or equivalently, Max-Independent-Set) [Kar72]. k-Clique is amongst the most well-studied problems in theoretical computer science, and it is the canonical intractable (W[1]-complete) problem
in parameterized complexity.
A naive algorithm solves k-Clique in O(nk ) time. By a reduction from 1985 to BMM on matrices
of size nk/3 ×nk/3 it can be solved with fast matrix multiplication in O(nωk/3 ) time [NP85] whenever
k is divisible by 3 (otherwise, more ideas are needed [EG04]). No better algorithms are known, and
researchers have wondered if improvements are possible [Woe04, Aut84]. As is the case for BMM,
obtaining faster than trivial combinatorial algorithms, by more than polylogarithmic factors, for kClique is a longstanding open question. The fastest combinatorial algorithm runs in O(nk / logk n)
time [Vas09].
Let 0 ≤ F ≤ ω and 0 ≤ C ≤ 3 be the smallest numbers such that 3k-Clique can be solved
combinatorially in O(nCk ) time and in O(nF k ) time by any algorithm, for any (large enough)
constant k ≥ 1. A conjecture in graph algorithms and parameterized complexity is that C = 3
and F = ω. It is known that an algorithm refuting this conjecture immediately implies a faster
exact algorithm for MAX-CUT [Wil05, Woe08]. Note that even a linear time algorithm for BMM
(ω = 2) would not prove that F < 2. A well known result by Chen et al. [CCF+ 05, CHKX06]
shows that F > 0 under the Exponential Time Hypothesis. A plausible conjecture about the
parameterized complexity of Subset-Sum implies that F ≥ 1.5 [ALW14]. There are many other
negative results that intuitively support this conjecture: Vassilevska W. and Williams proved that
a truly subcubic combinatorial algorithm for 3-Clique implies such algorithm for BMM as well
3
[WW10]. Unconditional lower bounds for k-Clique are known for various computational models,
such as Ω(nk ) for monotone circuits [AB87]. The planted Clique problem has also proven to be
very challenging (e.g. [AAK+ 07, AKS98, HK11, Jer92]). Max-Clique is also known to be hard to
efficiently approximate within nontrivial factors [Has99].
Formally, our reduction from k-Clique to CFG recognition proves the following theorem.
Theorem 1. There is context-free grammar GC of constant size such that if we can determine if
a string of length n can
be obtained from GC in T (n) time, then k-Clique on n node graphs can be
k/3+1
solved in O T n
time, for any k ≥ 3. Moreover, the reduction is combinatorial.
To see the tightness of our reduction, let 1 ≤ f ≤ ω and 1 ≤ c ≤ 3 denote the smallest numbers
such that CFG recognition can be solved in O(nf ) time and combinatorially in O(nc ) time. An
immediate corollary of Theorem 1 is that f ≥ F and c ≥ C. Under the plausible assumption
that current k-Clique algorithms are optimal, up to no(1) improvements, our theorem implies that
f ≥ ω and c ≥ 3. Combined with Valiant’s algorithm we get that f = ω and with standard CFG
parsers we get that c = 3. Because our grammar size g is fixed, we also rule out O(h(g) · n3−ε ) time
combinatorial CFG parsers for any computable function h(g).
In other words, we construct a single fixed context-free grammar GC for which the recognition
problem (and therefore any parsing problem) cannot be solved any faster than by Valiant’s algorithm and any combinatorial recognizer will not run in truly subcubic time, without implying a
breakthrough algorithm for the Clique problem. This (conditionally) proves that these algorithms
are optimal general purpose CFG parsers, and more efficient parsers will only work for CFL with
special restricting properties. On the positive side, our reduction might hint at what a CFG should
look like to allow for efficient parsing.
The online version of CFG recognition is as follows: preprocess a CFG such that given a string
w that is revealed one letter at a time, so that at stage i we get w[1 · · · i], we can say as quickly as
possible whether w[1 · · · i] can be derived from the grammar (before seeing the next letters). One
usually tries to minimize the total time it takes to provide all the |w| = n answers. This problem
has a long history of algorithms [Wei76, GRH80, Ryt85] and lower bounds [HS65, Gal69, Sei86].
The current best upper bound is O(n3 / log2 n) total running time, and the best lower bound is
Ω(n2 / log n). Since this is a harder problem, our lower bound for CFG recognition also holds for it.
1.1
More Results
The main ingredient in the proof of Theorem 1 is a lossless encoding of a graph into a string that
belongs to a simple CFL iff the graph contains a k-Clique. Besides classifying the complexity of a
fundamental problem, this construction has led us to new lower bounds for two more modern and
well-studied cubic time problems for which faster algorithms are highly desirable in practice.
RNA Folding The RNA folding problem is a version of maximum matching and is one of the
most relevant problems in computational biology. Its most basic version can be neatly defined as
follows. Let Σ be a set of letters and let Σ′ = {σ ′ | σ ∈ Σ} be the set of “matching” letters, such
that for every letter σ ∈ Σ the pair σ, σ ′ match. Given a sequence of n letters over Σ ∪ Σ′ the RNA
folding problem asks for the maximum number of non-crossing pairs {i, j} such that the ith and
j th letter in the sequence match.
The problem can be viewed as a Stochastic CFG parsing problem in which the CFG is very restricted. This intuition has led to many mildly subcubic algorithms for the problem [Son15, Aku98,
4
BTZZU11, PTZZU11, VGF13, FG09, ZTZU10]. The main idea is to adapt Valiant’s algorithm by
replacing his BMM with a (min, +)-matrix product computation (i.e. distance or tropical product). Using the fastest √known algorithm for (min, +)-matrix multiplication it is possible to solve
RNA folding in O(n3 /2 log n ) time [Wil14]. The fastest combinatorial algorithms, however, run in
roughly O(n3 / log2 n) time [Son15, BTZZU11], and we show that a truly subcubic such algorithm
would imply a breakthrough Clique algorithm. Our result gives a negative partial answer to an
open question raised by Andoni [And14].
Theorem 2. If RNA Folding on a sequence of length
n can be solved in T (n) time, then k-Clique
k/3+O(1)
on n node graphs can be solved in O T n
time, for any k ≥ 3. Moreover, the reduction
is combinatorial.
Besides a tight lower bound for combinatorial algorithms, our result also shows that a faster than
O(nω ) algorithm for RNA Folding is unlikely. Such an upper bound is not known for the problem,
leaving a small gap in the complexity of the problem after this work. However, we observe that
the known reductions from RNA Folding to (min, +)-matrix multiplication produce matrices with
very special “bounded monotone” structure: the entries are bounded by n and every row and
column are monotonically increasing. This exact structure has allowed Chan and Lewenstein to
solve the “bounded monotone” (min, +)-convolution problem in truly subquadratic time [CL15].
Their algorithm uses interesting tools from additive combinatorics and it seems very plausible that
the approach will lead to a truly subcubic (non-combinatorial) algorithm for “bounded monotone”
(min, +) product and therefore for RNA Folding.
Dyck Edit Distance The lower bound in Theorem 1 immediately implies a similar lower bound
for the Language Edit Distance problem on CFLs, in which we want to be able to return the
minimal edit distance between a given string w and a string in the language. That is, zero if w is
in the language and the length of the shortest sequence of insertions, deletions, substitutions that
is needed to convert w to a string that is in the language otherwise. This is a classical problem
introduced by Aho and Peterson in the early 70’s with cubic time algorithms [AP72, Mye95] and
many diverse applications [KSSY13, GCS+ 00, FM80]. Very recently, Rajasekaran and Nicolae
[RN14] and Saha [Sah14b] obtained truly subcubic time approximation algorithms for the problem
for arbitrary CFGs.
However, in many applications the CFG we are working with is very restricted and therefore
easy to parse in linear time. One of the simplest CFGs with big practical importance is the Dyck
grammar which produces all strings of well-balanced parenthesis. The Dyck recognition problem
can be easily solved in linear time with a single pass on the input. Despite the grammar’s very
special structure, the Dyck Edit Distance problem is not known to have a subcubic algorithm.
In a recent breakthrough, Saha [Sah14a] presented a near-linear time algorithm that achieves a
logarithmic approximation for the problem. Dyck edit distance can be viewed as a generalization of
the classical string edit distance problem whose complexity is essentially quadratic [CLRS09, BI15],
and Saha’s approximation algorithm nearly matches the best known approximation algorithms for
string edit distance of Andoni, Krauthgamer, and Onak [AKO10], both in terms of running time
and approximation factor. This naturally leads one to wonder whether the complexity of (exact)
Dyck edit distance might be also quadratic. We prove that this is unlikely unless ω = 2 or there
are faster clique finding algorithms.
5
Theorem 3. If Dyck edit distance on a sequence of length
n can be solved in T (n) time, then
k+O(1)
3k-Clique on n node graphs can be solved in O T n
time, for any k ≥ 1. Moreover, the
reduction is combinatorial.
Our result gives an answer to an open question of Saha [Sah14a], who asked if Lee’s lower
bound holds for the Dyck Edit Distance problem, and shows that the search for good approximation
algorithms for the problem is justified since efficient exact algorithms are unlikely.
Remark A simple observation shows that the longest common subsequence (LCS) problem on
two sequences x, y of length n over an alphabet Σ can be reduced to RNA folding on a sequence of
length 2n: if y = y1 . . . yn then let ŷ := yn′ . . . y1′ and then RN A(x ◦ ŷ) = LCS(x, y). A quadratic
lower bound for LCS was recently shown under the Strong Exponential Time Hypothesis (SETH)
[ABV15, BK15], which implies such a lower bound for RNA folding as well (and, with other ideas
from this work, for CFG recognition and Dyck Edit Distance). However, we are interested in higher
lower bounds, ones that match Valiant’s algorithm and basing such lower bounds on SETH would
imply that faster matrix multiplication algorithms refute SETH - a highly unexpected breakthrough.
Instead, we base our hardness on k-Clique and devise more delicate constructions that use the
cubic-time nature of our problems.
Proof Outlines In our three proofs, the main approach is the following. We will first preprocess
a graph G in O(nk+O(1) ) time in order to construct an encoding of it into a string of length
O(nk+O(1) ). This will be done by enumerating all k-cliques and representing them with carefully
designed gadgets such that a triple of clique gadgets will “match well” if and only if the triple make
a 3k-clique together, that is, if all the edges between them exists. We will use a fast (subcubic)
CFG recognizer, RNA folder, or an Dyck Edit Distance algorithm to speed up the search for such
a “good” triple and solve 3k-Clique faster than O(n3k ). These clique gadgets will be constructed
in similar ways in all of our proofs. The main differences will be in the combination of these cliques
into one sequence. The challenge will be to find a way to combine O(nk ) gadgets into a string in a
way that a “good” triple will affect the overall score or parse-ability of the string.
Notation and Preliminaries All graphs in this paper will be on n nodes and O(n2 ) undirected
and unweighted edges. We associate each node with an integer in [n] and let v̄ denote the encoding
of v in binary and we will assume that it has length exactly 2 log n for all nodes in V (G). When
a graph G is clear from context, we will denote the set of all k-cliques of G by Ck . We will denote
concatenation of sequences with x ◦ y, and the reverse of a sequence x by xR . Problem definitions
and additional problem specific preliminaries will be given in the corresponding section.
2
Clique to CFG Recognition
This section we show our reduction from Clique to CFG recognition and prove Theorem 1.
Given a graph G = (V, E), we will construct a string w of length O(k2 · nk+1 ) that encodes
G. The string will be constructed in O(k2 · nk+1 ) time which is linear in its length. Then, we will
define our context free grammar GC which will be independent of G or k and it will be of constant
size, such that our string w will be in the language defined by GC if and only if G contains a 3k
clique. This will prove Theorem 1.
6
Let Σ = {0, 1, $, #, astart , amid , aend , bstart , bmid , bend , cstart , cmid , cend } be our set of 13 terminals
(alphabet). As usual, ε will denote the empty string. We will denote the derivation rules of a
context free grammar with → and the derivation (by applying one or multiple rules) with =⇒ .
The string First, we will define node and list gadgets:
N G(v) = # v̄ #
and
LG(v) = #
u∈N (v)
($ ūR $) #
Consider some t = {v1 , . . . , vk } ∈ Ck . We now define “clique node” and “clique list” gadgets.
CN G(t) =
k
v∈t (N G(v))
and CLG(t) = (
k
v∈t LG(v))
and our main clique gadgets will be:
CGα (t) = astart CN G(t) amid CN G(t) aend
CGβ (t) = bstart CLG(t) bmid CN G(t) bend
CGγ (t) = cstart CLG(t) cmid CLG(t) cend
Finally, our encoding of a graph into a sequence is the following:
w=(
t∈Ck CGα (t)) (
t∈Ck CGβ (t)) (
The Clique Detecting Context Free Grammar.
GC is:
t∈Ck CGγ (t))
The set of non-terminals in our grammar
T = {S, W, W′ , V, Sαγ , Sαβ , Sβγ , S⋆αγ , S⋆αβ , S⋆βγ , Nαγ , Nαβ , Nβγ }.
The “main” rules are:
S → W astart Sαγ cend W
S⋆αγ → amid Sαβ bmid Sβγ cmid
S⋆αβ → aend W bstart
S⋆βγ → bend W cstart
And for every xy ∈ {αβ, αγ, βγ} we will have the following rules in our grammar. These rules
will be referred to as “listing” rules.
Sxy → S⋆ xy
Sxy → # Nxy $ V #
Nxy → # Sxy # V $
Nxy → σ Nxy σ
∀σ ∈ {0, 1}
Then we also add “assisting” rules:
W→ ε |σW
′
W → ε|σW
∀σ ∈ Σ
′
∀σ ∈ {0, 1}
′
V→ ε |$W $V
Our Clique Detecting grammar GC has 13 non-terminals T , 13 terminals Σ, and 38 derivation
rules. The size of GC , i.e. the sum of the lengths of the derivation rules, is 132.
7
The proof. This proof is essentially by following the derivations of the CFG, starting from the
starting symbol S and ending at some string of terminals, and showing that the resulting string
must have certain properties. Any encoding of a graph into a string as we describe will have these
properties iff the graph contains a 3k-clique. The correctness of the reduction will follow from the
following two claims.
Claim 1. If GC =⇒ w then G contains a 3k-clique.
Proof. The derivation of w must look as follows. First we must apply the only starting rule,
S =⇒ w1 astart Sαγ cend w2
where astart appears in CGα (tα ) for some tα ∈ Ck and cstart appears in CGγ (tγ ) for some tγ ∈ Ck ,
and w1 is the prefix of w before CG(tα ) and w2 is the suffix of w after CG(tγ ). Then we can get,
Sαγ =⇒ CN G(tα ) S⋆αγ CLG(tγ )
by repeatedly applying the xy-“listing” rules where xy = αγ and finally terminating with the rule
Sαγ → S⋆αγ . By Lemma 1 below, this derivation is only possible if the nodes of tα ∪ tγ make a
2k-clique (call this observation (*)). Then we have to apply the derivation:
S⋆αγ =⇒ amid Sαβ bmid Sβγ cmid
and for some tβ ∈ Ck we will have,
Sαβ =⇒ CN G(tα ) S⋆αβ CLG(tβ ), and Sβγ =⇒ CN G(tβ ) S⋆βγ CLG(tγ ),
where in both derivations we repeatedly use “listing” rules before exiting with the Sxy → S⋆ xy rule.
Again, by Lemma 1 we get that the nodes of tα ∪ tβ are a 2k clique in G, and that the nodes of
tβ ∪ tγ are a 2k clique as well (call this observation (**)). Finally, we will get the rest of w using
the derivations:
S⋆ αβ =⇒ aend w3 bstart , and S⋆ βγ =⇒ bend w4 cstart ,
where w3 is the substring of w between CG(tα ) and CG(tβ ), and similarly w4 is the substring of w
between CG(tβ ) and CG(tγ ).
Combining observations (*) and (**), that we got from the above derivation scheme and
Lemma 1, we conclude that the nodes of tα ∪ tβ ∪ tγ form a 3k-clique in G, and we are done.
To complete the proof we will now prove Lemma 1.
Lemma 1. If for some t, t′ ∈ Ck and xy ∈ {αβ, αγ, βγ} we can get the derivation Sxy
CN G(t) S⋆ xy CLG(t′ ), only using the “listing” rules, then t ∪ t′ forms a 2k clique in G.
=⇒
Proof. Any sequence of derivations starting at Sxy and ending at S⋆ xy will have the following form.
From Sxy we can only proceed to non-terminals V and Nxy . The non-terminal V does not produce
any other non-terminals. A single instantiation of Sxy can only produce (via the second “listing”
rule) a single instantiation of Nxy . In turn, from Nxy we can only proceed to non-terminals V and
Sxy , and again, a single instantiation of Nxy can produce a single instantiation of Sxy . Thus, we
produce some terminals (on the left and on the right) from the set {#, $, 0, 1} and then we arrive
to Sxy again. This can repeat an arbitrary number of times, until we apply the rule Sxy → S⋆ xy .
8
Thus, the derivations must look like this:
Sxy =⇒ ℓ S⋆xy r
for some strings ℓ, r of the form {#, $, 0, 1}∗ , and our goal is to prove that ℓ, r satisfy certain
properties.
It is easy to check that V can derive strings of the following form p = ($ {0, 1}∗ $)∗ , that is,
it produces a list (possibly of length 0) of binary sequences (possibly of length 0) surrounded by
$ symbols (between every two neighboring binary sequences there are two $). A key observation
is that repeated application of the fourth “listing” rule gives derivations Nxy =⇒ s Nxy sR , for
any s ∈ {0, 1}∗ . Combining these last two observations, we see that when starting with Sxy we can
only derive strings of the following form, or terminate via the rule Sxy → S⋆ xy .
Sxy =⇒ # Nxy $ p1 # =⇒ # s Nxy sR $ p1 # =⇒ # s # Sxy # p2 $ sR $ p1 #
(1)
for some p1 , p2 of the form ($ {0, 1}∗ $)∗ .
Now consider the assumption in the statement of the lemma, and recall our constructions of
“clique node gadget” and “clique list gadget” . By construction, CN G(t) is composed of k2 node
gadgets (N G) separated by # symbols, and CLG is composed of k2 list gadgets (LG) separated by
# symbols. Note also that the list gadgets contain O(n) node gadgets within them and those are
separated by $ symbols, and there are no # symbols within the list gadgets.
For every i ∈ [k2 ], let ℓi be the ith N G in CN G(t) and let ri be the ith LG in CLG(t′ ). Then,
for every 2 ≤ i ≤ k, in the derivation Sxy =⇒ CN G(t) S⋆xy CLG(t′ ), we must have had the
derivation
Sxy =⇒ ℓ1 · · · ℓi−1 ( Sxy ) rk2 −i+2 · · · rk2 =⇒ ℓ1 · · · ℓi−1 ( ℓi Sxy rk2 −i+1 ) rk2 −i+2 · · · rk2
and by (1) this implies that the binary encoding of the node v ∈ t that appears in the ith N G
in CN G(t) must appear in one of the N Gs that appear in the (k2 − i + 1) LG in CLG(t′ ) which
corresponds to a node u ∈ t′ . Since LG(u) contains a list of neighbors of the node u, this implies
that v ∈ N (u) and {u, v} ∈ E. Also note that u ∈
/ N (u) and therefore u does not appear in LG(u)
and therefore v cannot be equal to u if this derivation occurs.
Now, consider any pair of nodes v ∈ t, u ∈ t′ . By the construction of CNG and CLG, we must
have an index i ∈ [k2 ] such that the ith NG in CN G(t) is N G(v) and the (k 2 − i + 1) LG in
CLG(t′ ) is CLG(u). By the previous argument, we must have that u 6= v and {u, v} ∈ E is an
edge. Given that t, t′ are k-cliques themselves, and any pair of nodes v ∈ t, u ∈ t′ must be neighbors
(and therefore different), we conclude that t ∪ t′ is a 2k-clique.
Claim 2. If G contains a 3k-clique, then GC =⇒ w.
Proof. This claim follows by following the derivations in the proof of Claim 1 with any triple
tα , tβ , tγ ∈ Ck of k-cliques that together form a 3k-clique.
We are now ready to prove Theorem 1.
Reminder of Theorem 1 There is context-free grammar GC of constant size such that if we
can determine if a string of length n can beobtained from GC in T (n) time, then k-Clique on
n node graphs can be solved in O T nk/3+1 time, for any k ≥ 3. Moreover, the reduction is
combinatorial.
9
Proof. Given an instance of 3k-Clique, a graph G = (V, E) we construct the string w as described
above, which will have length O(k2 · nk+1 ), in O(k2 · nk+1 ) time. Given a recognizer for GC as in
the statement, we can check whether GC =⇒ w in O T nk/3+1 time (treating k as a constant).
By Claims 1 and 2, GC =⇒ w iff the graph G contains a 3k-clique.
3
Clique to RNA folding
In this section we prove Theorem 2 by reducing k-Clique to RNA folding, defined below.
Let Σ be an alphabet of letters of constant size. For any letter σ ∈ Σ there will be exactly one
“matching” letter which will be denoted by σ ′ . Let Σ′ = {σ ′ | σ ∈ Σ} be the set of matching letters
to the letters in Σ. Throughout this section we will say that a pair of letters {x, y} match iff y = x′
or x = y ′ .
Two pairs of indices (i1 , j1 ), (i2 , j2 ) such that i1 < j1 and i2 < j2 are said to “cross” iff at
least one of the following three conditions hold: (i) i1 = i2 or i1 = j2 , or j1 = i2 , or j1 = j2 ; (ii)
i1 < i2 < j1 < j2 ; (iii) i2 < i1 < j2 < j1 . Note that by our definition, non-crossing pairs cannot
share any indices.
Definition 1 (RNA Folding). Given a sequence S of n letters from Σ ∪ Σ′ , what is the maximum
number of pairs A = {(i, j) | i < j and i, j ∈ [n]} such that for every pair (i, j) ∈ A the letters
S[i] and S[j] match and there are no crossing pairs in A. We will denote this maximum value by
RN A(S).
It is interesting to note that RNA can be seen as a language distance problem with respect
to some easy to parse grammar. Because of the specific structure of this grammar, our reduction
from Section 2 does not apply. However, the ideas we introduced allow us to replace our clique
detecting grammar with an easier grammar if we ask the parser to return more information, like
the distance to a string in the grammar. At a high level, this is how we get the reduction to RNA
folding presented in this section.
To significantly simplify our proofs, we will reduce k-Clique to a more general weighted version
of RNA folding. Below we show that this version can be reduced to the standard RNA folding
problem with a certain overhead.
Definition 2 (Weighted RNA Folding). Given a sequence S of n letters from Σ ∪ Σ′ and a weight
function w : Σ → [M ], what is the maximum weight of a set of pairs A = {(i, j) | i < j and i, j ∈
[n]} such that for every pair (i, j) ∈ A the
Pletters S[i] and S[j] match and there are no crossing
pairs in A. The weight of A is defined as (i,j)∈A w(S[i]). We will denote this maximum value by
W RN A(S).
Lemma 2. An instance S of Weighted RNA Folding on a sequence of length n, alphabet Σ ∪ Σ′ ,
and weight function w : Σ → [M ] can be reduced to an instance Ŝ of RNA Folding on a sequence
of length O(M n) over the same alphabet.
w(S )
w(S )
Proof. Let S = S1 · · · Sn , we set Ŝ := S1 1 · · · Sn n , that is, each symbol Si is repeated w(Si )
times. First, we can check that W RN A(S) ≤ RN A(Ŝ). This holds because we can replace each
matching pair {a, a′ } in the folding achieving weighted RNA score of W RN A(S) with w(a) such
pairs in the (unweighted) RNA folding instance Ŝ giving the same contribution to RN A(Ŝ).
10
Now we will show that RN A(Ŝ) ≤ W RN A(S). Suppose thet there are symbols a and a′ in Ŝ
that are paired. The symbol a comes from a sequence s1 of a symbols of length w(a). The sequence
s1 was produced from a single symbol a when transforming S into Ŝ. Similarly, the symbol a′
comes from a sequence s2 of a′ symbols of length w(a). Also, assume that there exists a symbol in
s1 that is paired to a symbol that is outside of s2 or there exists a symbol in s2 that is paired to a
symbol outside of s1 . While we can find such symbols a and a′ , we repeat the following procedure.
Choose a and a′ that satisfy the above properties. And choose them so that the number of other
symbols between a and a′ is as small as possibly. Break ties arbitrarily. We match all symbols in
s1 to their counterparts in s2 . Also, we rematch all symbols that were previously matched to s1
or s2 among themselves. We can check that we can rematch these symbols so that the number of
matched pairs do not decrease.
Therefore, we can assume that in some optimal folding of Ŝ, for any pair a ∈ Σ, a′ ∈ Σ′ that is
matched the corresponding substrings s1 and s2 are completely paired up. Thus, to get a folding
of S that achieves W RN A(S) at least RN A(Ŝ) we can now simply fold the corresponding symbols
to s1 and s2 , for any such pair a, a′ .
3.1
The Reduction
Given a graph G = (V, E) on n nodes and O(n2 ) unweighted undirected edges, we will describe
how to efficiently construct a sequence SG over an alphabet Σ of constant size, such that the RNA
score of SG will depend on whether G contains a 3k-clique. The length of SG will be O(kd nk+c )
for some small fixed constants c, d > 0 independent of n and k, and the time to construct it from
G will be linear in its length. This will prove that a fast (e.g. subcubic) RNA folder can be used
as a fast 3k-clique detector (one that runs much faster than in O(n3k ) time).
Our main strategy will be to enumerate all k-cliques in the graph and then search for a triple of
k-cliques that have all the edges between them. We will be able to find such a triple iff the graph
contains a 3k-clique. An RNA folder will be utilized to speed up the search for such a “good”
triple. Our reduction will encode every k-clique of G using a “short” sequence of length O(nc ) such
that the RNA folding score of a sequence composed of the encodings of a triple of sequences will
be large iff the triple is “good”. Then, we will show how to combine the short encodings into our
long sequence SG such that the existence of a “good” triple affects the overall score of an optimal
folding.
The RNA Sequence Our sequence SG will be composed of many smaller gadgets which will be
combined in certain ways by other padding gadgets. We construct these gadgets now and explain
their useful properties. The proofs of these properties are postponed until after we present the
whole construction of SG .
For a sequence s ∈ Σ∗ let p(s) ∈ (Σ′ )∗ be the sequence obtained from s by replacing every letter
σ ∈ Σ with the matching letter σ ′ ∈ Σ′ . That is, if s = s1 · · · sn then p(s) = s′1 · · · s′n .
Our alphabet Σ will contain the letters 0, 1 and some additional symbols which we will add as
needed in our gadgets. We will set the weights so that w(0) = w(1) = 1, and the extra symbols
we add will be more “expensive”. For example, we will add the $ symbol to the alphabet and set
w($) = 10 · log n. We define node gadgets as,
N G(v) = $2n v̄ $2n
11
and list or neighborhood gadgets as,
LG(v) =
u∈N (v) ($
ū $) ◦
u∈N
/ (v) ($
$).
These gadget are constructed so that for any two nodes u, v ∈ V (G), the RNA folding score of
the sequence N G(v) ◦ p (LG(u))R is large (equal to some fixed value E1 ) if v is in the neighborhood
of u, that is (u, v) ∈ E(G), and smaller otherwise (at most E1 − 1). This is because the more
expensive $ symbols force an optimal folding to match v̄ with exactly one p(w̄)R , since otherwise a
$′ will remain unmatched while many $ are free. The construction also allows the folding to pick
any w ∈ N (u) to fold together with v̄, without affecting the score from the $ symbols. Then, we
use the fact that v̄ ◦ p(w̄)R achieves maximal score iff v = w. This is formally proved in Claim 3.
Let ℓ1 = 10k2 · n log n and note that ℓ1 is an upper bound on the total weight of all the symbols
in the gadgets N G(v) and LG(v), for any node v ∈ V (G). Let Ck be the set of k-cliques in G and
consider some t = {v1 , . . . , vk } ∈ Ck . We will now combine the node and list gadgets into larger
gadgets that will be encoding k-cliques.
We will add the # symbol to the alphabet and set w(#) = ℓ1 , i.e. a single # letter is more
expensive than entire k2 node or list gadget. We will encode a clique in two ways. The first one is,
CN G(t) =
k
v∈t (#N G(v)#)
CLG(t) = (
k
v∈t (#LG(v)#)) .
and the second one is,
These clique gadgets are very useful because of the following property. For any two k-cliques
t1 , t2 ∈ Ck , the RNA folding score of the sequence CN G(t1 ) ◦ p (CLG(t2 ))R is large (equal to some
fixed value E2 ) if t1 and t2 , together, form a 2k-clique, and is smaller otherwise (at most E2 − 1).
That is, the RNA folding score of the sequence tells us whether any pair of nodes u ∈ t1 , v ∈ t2
are connected (u, v) ∈ E(G). There are two ideas in the construction of these gadgets. First, we
copy the gadgets corresponding to the k nodes of the cliques k times, resulting in k2 gadgets, and
we order them in a way so that for any pair of nodes u ∈ t1 , v ∈ t2 there will be a position i such
that the gadget of u in CN G(t1 ) and the gadget of v in p(CLG(t2 ))R are both at position i. Then,
we use the expensive # separators to make sure that in an optimal RNA folding of CN G(t1 ) and
p(CLG(t2 ))R , the gadgets at positions i are folded together, and not to other gadgets - otherwise
some # symbol will not be paired. This is formally proved in Claim 4.
Let ℓ2 = 10 · k2 · ℓ1 = O(n log n) and note that it is an upper bound on the total weight of
all the symbols in the CN G(t) and CLG(t) gadgets. Finally, we introduce a new letter to the
alphabet g and set its weight to w(g) = ℓ2 , which is much more expensive than the entire gadgets
we constructed before, and then define our final clique gadgets. Moreover, we will now duplicate our
alphabet three times to force only “meaningful” foldings between our gadgets. It will be convenient
to think of α, β, γ as three types such that we will be looking for three k-clique, one from type α one
from β and one from γ. For any pair of types xy ∈ {αβ, αγ, βγ} we will construct a new alphabet
Σxy = {σxy | σ ∈ Σ} in which we mark each letter with the pair of types it should be participating
in. For a sequence s ∈ (Σ ∪ Σ)∗ we use the notation [s]xy to represent the sequence in (Σxy ∪ Σ′xy )∗
in which we replace every letter σ with the letter σxy .
We will need three types of these clique gadgets in order to force the desired interaction between
them.
CGα (t) = [g CN G(t) g]αγ ◦ [g′ p(CLG(t))R g′ ]αβ
12
CGβ (t) = [g CN G(t) g]αβ ◦ [g′ p(CLG(t))R g′ ]βγ
CGγ (t) = [g CN G(t) g ]βγ ◦ [g′ p(CLG(t))R g′ ]αγ
These clique gadgets achieve exactly what we want: for any three k-cliques tα , tβ , tγ ∈ Ck the
RNA folding score of the sequence CGα (tα ) ◦ CGβ (tβ ) ◦ CGγ (tγ ) is large (equal to some value E3 ) if
tα ∪ tβ ∪ tγ is a 3k-clique and smaller otherwise (at most E3 − 1). In other words, an RNA folder can
use these gadgets to determine if three separate k-cliques can form a 3k-clique. This is achieved by
noticing that the highest priority for an optimal folding would be to match the gxy letters with their
counterparts g′xy , which leaves us with three of sequences to fold: Sαγ = [CN G(tα )◦p(CLG(tγ ))R ]αγ
and Sαβ = [p(CLG(tα ))R ◦ CN G(tβ )]αβ and Sβγ = [p(CLG(tβ ))R ◦ CN G(tγ )]βγ . The maximal
score (E2 ) in each one of these three sequences can be achieved iff every pair of our k-cliques form
a 2k-clique which happens iff they form a 3k-clique. This is formally proved in Claim 5.
The only remaining challenge is to combine all the sequences corresponding to all the O(nk )
k-cliques in the graph into one sequence in a way that the existence of a “good” triple, one that
makes a 3k-clique, affects the RNA folding score of the entire sequence. Note that if we naively
concatenate all the clique gadgets into one sequence, the optimal sequence will choose to fold clique
gadgets in pairs instead of triples since folding a triple makes other gadgets unable to fold without
crossings. Instead, we will use the structure of the RNA folding problem again to implement a
“selection” gadget that forces exactly three clique gadgets to fold together in any optimal folding.
We remark that the implementation of such “selection” gadgets is very different in the three proofs
in this paper: In Section 2 we use the derivation rules, in this section we use the fact that even
when folding all expensive separators in a sequence to the left or right we are left with an interval
that is free to fold with other parts of the sequence, and in Section 4 we rely on the restriction of
Dyck that an opening bracket can match only to closing brackets to its right.
Towards this end, we introduce some extremely expensive symbols α, β, γ. Let ℓ3 = 10ℓ2 be
an upper bound on the total weight the CGx (t) gadgets, and set w(α) = w(β) = w(γ) = ℓ3 . Our
“clique detecting” RNA sequence is defined as follows.
SG =
α2n
◦ β 2n
◦ γ
k
t∈Ck
k
2nk
t∈Ck
t∈Ck
k
α′ CGα (t) α′ α2n
k
β ′ CGβ (t) β ′ β 2n
k
γ ′ CGγ (t) γ ′ γ 2n
The added padding makes sure that all but one CGα gadget are impossible to fold without
giving up an extremely valuable α, α′ pair, and similarly all but one CGβ and one CGγ cannot be
folded. To see this, assume all the α′ are paired (left or right) and note that if both α′ symbols
surrounding a clique gadget CGα (t) are paired to one side (say, left) then the only non crossing
pairs that the gadget could participate in are either with α symbols (but those cannot be matches)
or within itself. Our marking of symbols with pairs of types xy make it so that a clique gadget
cannot have any matches with itself. Therefore, if all α′ symbols are matched, then all but one
CGα (t) gadgets do not participate in any foldings. The argument for β, γ is symmetric. We are left
with a folding of a sequence of three clique gadgets CGα (tα ), CGβ (tβ ), CGγ (tγ ) which can achieve
maximal score iff tα ∪ tβ ∪ tγ is a 3k-clique.
This proves our main claim that the (weighted) RNA folding score of our clique detecting
sequence SG is large (equal to some fixed value EC ) if the graph contains a 3k-clique and smaller
(at most EC − 1) otherwise. See Claim 6 for the formal proof.
13
Our final alphabet Σ has size 18 (together with Σ′ this makes 36 symbols).
[
Σ = {α, β, γ} ∪
{0, 1, $, #, g}xy
xy∈{αβ,αγ,βγ}
Observe that SG can be constructed from G in O(nk+1 ) time, by enumerating all subsets of
k nodes and that it has length O(nk+1 ). The construction of SG should be seen as a heavy
preprocessing and encoding of the graph, after which we only have to work with k-cliques. The
largest weight we use in our construction is ℓ3 = O(kO(1) n log n) and therefore using Lemma 2
we can reduce the computation of the weighted RNA of SG to an instance of (unweighted) RNA
folding on a sequence of length O(|SG |kO(1) n log n) = Õ(kO(1) nk+2 ) which proves Theorem 2.
Formal Proofs We will start with the proof that the list and node gadgets have the desired
functionality. Let E1 = (10n + 1) · log n.
Claim 3. For any xy ∈ {αβ, αγ, βγ} and two nodes u, v ∈ V (G) we have that the weighted RNA
folding score W RN A([N G(u) ◦ p(LG(v))R ]xy ) is E1 if u ∈ N (v) and at most E1 − 1 otherwise.
Proof. Since all letters in the sequence we are concerned with have the same mark xy, we will omit
the subscripts. If u ∈ N (v) then p(ū)R appears in LG(v) and we can completely match it with
the sequence ū in N G(u), giving a score of 2 log n, then we match all the n $′ symbols in LG(v)
to some $ symbols in N G(u) and gain an extra score of n · 10 log n. Therefore, in this case, the
weighted RNA score is E1 = log n + 10n log n.
Now we assume that u ∈
/ N (v) and show that in the optimal folding the score is at most E1 − 1.
First, note that the sequence p(LG(v))R has fewer $′ symbols (it has 2n such symbols) than the
sequence N G(u) (which has 4n such symbols). By not pairing a symbol $′ in p(LG(v))R , we lose a
score of w($) which is much more than the entire weight of the non-$ symbols in N G(u). Therefore,
any matching which leaves some $′ unmatched is clearly sub-optimal, and we can assume that all
the $′ symbols are matched. Given this, the substring ū can only be folding to at most one of
the substrings p(v¯i )R in LG(v) for some vi ∈ N (v). This folding can only achieve score 2 log n − 1
because u ∈
/ N (v). Thus, the total score of the optimal matching is no more than E1 − 1.
Next, we prove that the “clique node gadgets” and “clique list gadgets” check that two k-cliques
form one bigger 2k-clique. Let E2 = 2k2 · ℓ1 + k2 · E1 .
Claim 4. For any xy ∈ {αβ, αγ, βγ} and two k-cliques t1 , t2 ∈ Ck we have that the weighted RNA
folding score W RN A([CN G(t1 ) ◦ p(CLG(t2 ))R ]xy ) is E2 if t1 ∪ t2 is a 2k-clique and at most E2 − 1
otherwise.
Proof. We will omit the irrelevant xy subscripts. First, note that the sequences CN G(t1 ) and
p(CLG(t2 ))R have the same number of # and #′ symbols, respectively. By not pairing a single
one of them with its counterpart, we lose a contribution of w(#) to the WRNA score, which is
much more than we could gain by pairing all the symbols in all the node and list gadgets (that is,
the rest of the sequence). Therefore, we assume that all the # and #′ symbols are paired. Let
t1 = {u1 , . . . , uk } and t2 = {v1 , . . . , vk }. We can now say that
X X
W RN A(CN G(t1 ) ◦ p(CLG(t2 ))R ) = (2k2 )w(#) +
W RN A(N G(ui ) ◦ p(LG(vj ))R ).
i∈[k] j∈[k]
14
and by Claim 3 we know that W RN A(N G(ui ) ◦ p(LG(vj ))R ) = E1 if ui and vj are connected and
less otherwise. Therefore, we can only get the maximal E2 = 2k2 · ℓ1 + k2 · E1 iff every pair of
nodes, one from t1 and one from t2 are connected. Since u ∈
/ N (u) for all u ∈ V (G) and since t1 , t2
are k-cliques, we conclude that t1 ∪ t2 is a 2k-clique.
We are now ready to prove the main property of our clique gadgets: a sequence of three
clique gadgets (one from each type) achieves maximal score iff they form a 3k-clique together. Let
E3 = 6ℓ2 + 3E2 .
Claim 5. For any xy ∈ {αβ, αγ, βγ} and three k-cliques tα , tβ , tγ ∈ Ck we have that the weighted
RNA folding score W RN A(CGα (tα ) ◦ CGβ (tβ ) ◦ CGγ (tγ )) is E3 if t1 ∪ t2 ∪ t3 is a 3k-clique and
at most E3 − 1 otherwise.
Proof. If for some xy ∈ {αβ, αγ, βγ} there is a symbol gxy which is not paired up with its counterpart, we lose a contribution to the W RN A score that is more than we could get by pairing up
all symbols that are not gxy . Therefore, we have the equality
W RN A(CGα (tα ) ◦ CGβ (tβ ) ◦ CGγ (tγ )) = 3 · 2ℓ2
+ W RN A([CN G(tα ) ◦ p(CLG(tγ ))R ]αγ )
+ W RN A([CN G(tβ ) ◦ p(CLG(tγ ))R ]βγ )
+ W RN A([p(CLG(tα ))R ◦ CN G(tβ )]αβ )).
By Claim 4, the last three summands are equal to E2 if all our three k-cliques are pairwise 2kcliques and otherwise at least one of the summands is less than E2 . The claim follows by noticing
that tα ∪ tβ ∪ tγ is a 3k-clique iff the three k-cliques are pairwise 2k-cliques.
We are now ready to prove our main claim about SG . This proof shows that our “selection”
gadgets achieve the desired property of having exactly one clique from each type fold in an optimal
matching. Let N = O(nk ) be the size of Ck which is the number of k-cliques in our graph and
therefore the number of clique gadgets we will have from each type. We will set EC = 6N + E3 .
Claim 6. The weighted RNA score of SG is EC if G contains a 3k-clique and at most EC − 1
otherwise.
Proof. Let x ∈ {α, β, γ} and define tx ≥ 0 denote the number of x′ symbols in SG that are not
paired. Because any clique gadget CGx can only have matches with letters from clique gadgets
CGy for some y ∈ {α, β, γ} such that y 6= x, we can say that at most tx /2 + 1 clique gadget
sequences CGx can have letters that participate in the folding.
Recall that by definition of our weights, the total weight of any clique gadget is much less than
ℓ3 /10 where ℓ3 is the weight of a letter α, β, γ and recall the definition of N = |Ck |. We will use
the inequalities:
W RN A(SG ) ≤ ((tα /2 + 1) + (tβ /2 + 1) + (tβ /2 + 1)) · ℓ3 /10 + ((2N − ta ) + (2N − tb ) + (2N − tc ))ℓ3 ,
and:
W RN A(SG ) ≥ ((2N − ta ) + (2N − tb ) + (2N − tc ))ℓ3 .
15
Since ℓ3 ≫ ℓ3 /10, we must have that tα = tβ = tγ = 0 in any optimal folding of SG . Now we get
that:
W RN A(SG ) = 6N ℓ3 + W RN A(CGα (tα ) ◦ CGβ (tβ ) ◦ CGγ (tγ ))
for some k-cliques tα , tβ , tγ ∈ Ck . By Claim 5, the last summand can be equal to E3 iff the graph
has a 3k-clique, and must be at most E3 − 1 otherwise. This, and the fact that EC = 6N ℓ3 + E3
completes the proof.
We are now ready to show that the construction of SG from graph G proves Theorem 2.
Reminder of Theorem 2 If RNA Folding on a sequence of length
n can be solved in T (n) time,
then k-Clique on n node graphs can be solved in O T nk/3+O(1) time, for any k ≥ 3. Moreover,
the reduction is combinatorial.
Proof. Given a graph G on n nodes we construct the sequence SG as described above. The sequence
can be constructed in O(kO(1) · nk+1 ) time, by enumerating all subsets of k nodes and that it has
length O(kO(1) · nk+1 ). The largest weight we use in our construction is ℓ3 = O(kO(1) n log n) and
therefore using Lemma 2 we can reduce the computation of the weighted RNA of SG to an instance
of (unweighted) RNA folding on a sequence of length O(|SG |kO(1) n log n) = Õ(kO(1) nk+2 ). Thus,
an RNA folder as in the statement returns the weighted RNA folding score of SG in T (nk/3+2 ) time
(treating k as a constant) and by Claim 6 this score determines whether G contains a 3k-clique.
All the steps in our reduction are combinatorial.
4
Clique to Dyck Edit Distance
In this section we prove Theorem 3 by reducing k-Clique to the Dyck Edit Distance problem,
defined below.
The Dyck grammar is defined over a fixed size alphabet of opening brackets Σ and of closing
brackets Σ′ = {σ ′ | σ ∈ Σ}, such that σ can only be closed by σ ′ . A string S belongs to the Dyck
grammar if the brackets in it are well-formed. More formally, the Dyck grammar is defined by the
rules S → SS and S → σ S σ ′ for all σ ∈ Σ and S → ε. This grammar defines the Dyck context
free language (which can be parsed in linear time).
The Dyck Edit Distance problem is: given a string S over Σ ∪ Σ′ find the minimum edit distance
from S to a string in the Dyck CFL. In other words, find the shortest sequence of substitutions
and deletions that is needed to convert S into a string that belongs to Dyck. We will refer to this
distance as the Dyck score or cost of S.
Let us introduce alternative ways to look at the Dyck Edit Distance problem that will be useful
for our proofs. Two pairs of indices (i1 , j1 ), (i2 , j2 ) such that i1 < j1 and i2 < j2 are said to “cross”
iff at least one of the following three conditions hold
• i1 = i2 or i1 = j2 , or j1 = i2 , or j1 = j2 ;
• i1 < i2 < j1 < j2 ;
• i2 < i1 < j2 < j1 .
Note that by our definition, non-crossing pairs cannot share any indices. We define an alignment
A of a sequence S of length n to be a set of non-crossing pairs (i, j), i < j, i, j ∈ [n]. If (i, j) is in
our alignment we say that letter i and letter j are aligned. We say that an aligned pair is a match
if S[i] = σ for some σ ∈ Σ and S[j] = σ ′ , i.e. an opening bracket and the corresponding closing
16
bracket. Otherwise, we say that the aligned pair is a mismatch. Mismatches will correspond to
substitutions in an edit distance transcript. A letter at an index i that does not appear in any
of the pairs in the alignment is said to be deleted. We define the cost of an alignment to be the
number of mismatches plus the number of deleted letters. One can verify that any alignment of
cost E corresponds to an edit distance transcript from S to a string in Dyck of cost E, and vice
versa.
4.1
The Reduction
Given a graph G = (V, E) on n nodes and O(n2 ) unweighted undirected edges, we will describe
how to efficiently construct a sequence SG over an alphabet Σ of constant size, such that the Dyck
score of SG will depend on whether G contains a 3k-clique. The length of SG will be O(kd nk+c )
for some small fixed constants c, d > 0 independent of n and k, and the time to construct it from
G will be linear in its length. This will prove that a fast (e.g. subcubic) algorithm for Dyck Edit
Distance can be used as a fast 3k-clique detector (one that runs much faster than in O(n3k ) time).
As in the other sections, our main strategy will be to enumerate all k-cliques in the graph and
then search for a triple of k-cliques that have all the edges between them. We will be able to find
such a triple iff the graph contains a 3k-clique. A Dyck Edit Distance algorithm will be utilized
to speed up the search for such a “good” triple. Our reduction will encode every k-clique of G
using a “short” sequence of length O(nc ) such that the Dyck score of a sequence composed of the
encodings of a triple of sequences will be large iff the triple is “good”. Then, we will show how to
combine the short encodings into our long sequence SG such that the existence of a “good” triple
affects the overall score of an optimal alignment.
The Sequence Our sequence SG will be composed of many smaller gadgets which will be combined in certain ways by other padding gadgets. We construct these gadgets now and explain their
useful properties. The proofs of these properties are postponed until after we present the whole
construction of SG .
Recall that we associate every node in V (G) with an integer in [n] and let v̄ denote the encoding
of v in binary and we will assume that it has length exactly 2 log n for all nodes. We will use the
fact that there is no node with encoding 0̄. For a sequence s ∈ Σ∗ let p(s) ∈ (Σ′ )∗ be the sequence
obtained from s by replacing every letter σ ∈ Σ with the closing bracket σ ′ ∈ Σ′ . That is, if
s = s1 · · · sn then p(s) = s′1 · · · s′n .
Our alphabet Σ will contain the letters 0, 1 and some additional symbols which we will add as
needed in our gadgets like $, #. We will use the numbers ℓ0 , . . . , ℓ5 such that ℓi = (1000 · n2 )i+1 ,
which can be bounded by nO(1) . We define node gadgets as,
N G(v) = $ℓ1 v̄ $ℓ1
and list or neighborhood gadgets as,
LG(v) =
ℓ0
u∈N (v) ($
ū $ℓ0 ) ◦
ℓ0
u∈N
/ (v) ($
0̄ $ℓ0 ).
These gadget are constructed so that for any two nodes u, v ∈ V (G), the Dyck score of the
sequence N G(v) ◦ p (LG(u))R is small (equal to some fixed value E1 ) if v is in the neighborhood of
u, that is (u, v) ∈ E(G), and larger otherwise (at least E1 + 1). This is proved formally in Claim 7,
by similar arguments as in Section 3.
17
Note that ℓ2 is an upper bound on the total length of all the symbols in the gadgets N G(v)
and LG(v), for any node v ∈ V (G). Let Ck be the set of k-cliques in G and consider some
t = {v1 , . . . , vk } ∈ Ck . We will now combine the node and list gadgets into larger gadgets that will
be encoding k-cliques. We will encode a clique in two ways. The first one is,
CN G(t) =
v∈t
(#ℓ2 N G(v) #ℓ2 )k
CLG(t) = (
v∈t
(#ℓ2 LG(v) #ℓ2 ))k .
and the second one is,
Note that ℓ3 is an upper bound on the total length of all the symbols in the CN G(t) and CLG(t)
gadgets. We will add the symbol g to the alphabet. Moreover, we will now duplicate our alphabet
three times to force only “meaningful” alignments between our gadgets. It will be convenient to
think of α, β, γ as three types such that we will be looking for three k-cliques, one from type α one
from β and one from γ. For any pair of types xy ∈ {αβ, αγ, βγ} we will construct a new alphabet
Σxy = {σxy | σ ∈ Σ} in which we mark each letter with the pair of types it should be participating
in. For a sequence s ∈ (Σ ∪ Σ′ )∗ we use the notation [s]xy to represent the sequence in (Σxy ∪ Σ′xy )∗
in which we replace every letter σ with the letter σxy .
We will need three types of these clique gadgets in order to force the desired interaction between
them.
CGα (t) = aℓ4 (x′α )ℓ5 [gℓ3 CN G(t) gℓ3 ]αγ
CGβ (t) = bℓ4 (x′β )ℓ5 [(g′ )ℓ3 CLG(t) (g′ )ℓ3 ]αβ
CGγ (t) = cℓ4 (x′γ )ℓ5 [(g′ )ℓ3 CLG(t) (g′ )ℓ3 ]βγ
◦
◦
◦
[gℓ3 p(CN G(t))R gℓ3 ]αβ yℓα5 (a′ )ℓ4
[gℓ3 p(CN G(t))R gℓ3 ]βγ yℓβ5 (b′ )ℓ4
[(g′ )ℓ3 p(CLG(t))R (g′ )ℓ3 ]αγ yℓγ5 (c′ )ℓ4
These clique gadgets achieve exactly what we want: for any three k-cliques tα , tβ , tγ ∈ Ck
the Dyck score of the sequence CGα (tα ) ◦ CGβ (tβ ) ◦ CGγ (tγ ) is small (equal to some value E3 )
if tα ∪ tβ ∪ tγ is a 3k-clique and larger otherwise (at least E3 + 1). This is formally proved in
Claim 9, again by similar arguments as in Section 3 (but more complicated because of the possible
mismatches).
The main difference over the proof of Section 3 is the way we implement the “selection” gadgets.
We want to combine all the clique gadgets into one sequence in a way that the existence of a “good”
triple, one that makes a 3k-clique, affects the Dyck score of the entire sequence. The ideas we used in
the RNA proof do not immediately work here because of “beneficial mismatches” of the separators
we add with themselves and because in Dyck (σ, σ ′ ) match but (σ ′ , σ) do not (while in RNA we do
not care about the order). We will use some new ideas.
Our “clique detecting” sequence is defined as follows.
SG =
x α ℓ5 (
t∈Ck CGα (t))
y′α
◦ x β ℓ5 (
t∈Ck CGβ (t))
y′β
t∈Ck CGγ (t))
y′γ
◦ x γ ℓ5 (
ℓ5
ℓ5
ℓ5
As the xα , y′α symbols are very rare and “expensive” an optimal alignment will match them to
some of their counterparts within the α part of the sequence. However, when the x′α , yα letters are
matched, we cannot match the adjacent a, a′ symbols - which are also quite expensive. Therefore,
18
the optimal behavior is to match the xα , y′α from exactly one “interval” from the α part. A similar
argument holds for the β, γ parts. This behavior leaves exactly one clique gadget from each type
to be aligned freely with each other as a triple. By the construction of these gadgets, an optimal
score can be achieved iff there is a 3k-clique.
This proves our main claim that the Dyck score of our clique detecting sequence SG is small
(equal to some fixed value EC ) if the graph contains a 3k-clique and larger (at least EC + 1)
otherwise. See Claim 10 for the formal proof.
When k is fixed, SG can be constructed from G in O(nk+O(1) ) time, by enumerating all subsets
of k nodes and that it has length O(nk+O(1) ). This proves Theorem 3.
Our final alphabet Σ has size 24 (together with Σ′ this makes 48 symbols).
[
Σ = {a, b, c} ∪
{0, 1, $, #, g, x, y}xy
xy∈{αβ,αγ,βγ}
Formal Proofs
Let E1 = log n · (n − 1) + (2ℓ1 − n · 2ℓ0 )/2.
Claim 7. For any xy ∈ {αβ, αγ, βγ}, if v ∈ N (u), then
Dyck([N G(v) ◦ p(LG(u))]xy ) = E1
and > E1 otherwise.
Proof. We will omit the subscripts xy since they do not matter for the proof. We want to claim that
the binary sequence v̄ is aligned to at most one binary sequence p(z̄ R ). If this is not so, then there
are 2ℓ0 symbols $′ from p(LG(u)) that will be mismatched or deleted thus contributing at least
S1 = ℓ0 to the Dyck score. There will be at least 2ℓ1 − (n − 1)2ℓ0 symbols $ that are not matched to
their counterparts $′ . Those will contribute at least S2 = ℓ1 − (n − 1)ℓ0 to the Dyck score. We get
that the Dyck score is at least S1 + S2 > E1 . Now we assume that the binary sequence v̄ is aligned
to exactly one other binary sequence which we denote by p(z¯1 R ). There are S2 = 2 log(n) · (n − 1)
symbols from binary sequences from p(LG(u)) that are not matched to their counterparts. Also,
there are S3 = 2ℓ1 − n · 2ℓ0 symbols $ that are not matched to their counterparts. The contribution
of the unmatched symbols is ≥ (S2 + S3 )/2 = E1 to the Dyck cost. We want to show that the
equality can be achieved iff v ∈ N (u). From the proof it is clear that, if v ∈ N (u), then we can
achieve equality by choosing z1 to be the element from N (u) that is equal to v. Also, if we achieve
equality, the symbols corresponding to S2 and S3 contribute ≥ (S2 + S3 )/2 to the Dyck score.
These symbols contribute (S2 + S3 )/2 to the Dyck cost iff the mismatches happen only between
themselves and all symbols $′ are matched to their counterparts. The only remaining symbols that
could potentially contribute to the score correspond to v and z1 . They contribute 0 to the Dyck
score iff v̄ = z¯1 R , that is, v ∈ N (u).
For the next proofs we will use the following definition.
Definition 3. Given two sequences P and T , we define
pattern(P, T ) :=
min
Q is a contiguous
subsequence of T
Let E2 = k2 · E1 .
19
Dyck(P ◦ Q).
Claim 8. For any xy ∈ {αβ, αγ, βγ} and two k-cliques t1 , t2 ∈ Ck , if t1 ∪ t2 is a 2k clique then
Dyck([CN G(t1 ) ◦ CLG(t2 )]xy ) = E2
and > E2 otherwise.
Proof. We will omit the subscripts xy since they do not matter for the proof. We have that
X
k · pattern(N G(v), CLG(t2 )).
Dyck(CN G(t1 ) ◦ CLG(t2 )) ≥
v∈t1
Suppose that for some v ∈ t1 , N G(v) is aligned to more than one gadget p(LG(u)). Then symbols
# between these gadgets p(LG(u)) will be substituted or deleted. The cost of these operations is
≥ ℓ2 > E2 . Therefore, we have that any one of k2 gadgets N G(v) is aligned to at most one gadget
p(LG(u)) for some u ∈ t2 . By the construction of CN G and CLG, we have that
Dyck(CN G(t1 ) ◦ CLG(t2 ))
XX
≥
pattern(N G(v), (#′ )2ℓ2 p(LG(u))(#′ )2ℓ2 )
v∈t1 u∈t2
=
XX
v∈t1 u∈t2
Dyck(N G(v) ◦ p(LG(u))),
where the last equality follows because # does not appear among symbols of N G(v). Now we have
that
XX
Dyck(CN G(t1 ) ◦ CLG(t2 )) ≥
E1 = k2 · E1 = E2 ,
v∈t1 u∈t2
where we use Claim 7. If we have equality, it means that we have equality in all k2 invocations of
Claim 7, which implies that v ∈ N (u) for all v ∈ t1 , u ∈ t2 . And this gives that there is a biclique
between vertices of t1 and t2 . Also, it is possible to verify that we can achieve the equality if there
is a biclique.
Let E3 = 3(ℓ4 + E2 ).
Claim 9. For any triple of k-cliques tα , tβ , tγ ∈ Ck , the union tα ∪ tβ ∪ tγ is a 3k-clique, then
Dyck(
xα ℓ5 CGα (tα )(yα ′ )ℓ5
◦ xβ ℓ5 CGβ (tβ )(yβ ′ )ℓ5
◦ xγ ℓ5 CGγ (tγ )(yγ ′ )ℓ5
) = E3
and > E3 otherwise.
Proof. We need to lower bound
3
3
CN Gαγ (tα )gℓαγ
gℓαβ
3
CN Gαβ (tα )gℓαβ
yα ℓ5 (a′ )ℓ4 (yα ′ )ℓ5
xβ ℓ5 bℓ4 (xβ ′ )ℓ5 (g′αβ )ℓ3
3
CLGαβ (tβ )(g′αβ )ℓ3 gℓβγ
3
CN Gβγ (tβ )gℓβγ
yβ ℓ5 (b′ )ℓ4 (yβ ′ )ℓ5
xγ ℓ5 cℓ4 (xγ ′ )ℓ5 (g′βγ )ℓ3
CLGβγ (tγ )(g′βγ )ℓ3 (g′αγ )ℓ3
CLGαγ (tγ )(g′αγ )ℓ3 yγ ℓ5 (c′ )ℓ4 (yγ ′ )ℓ5 ).
3
Dyck( xα ℓ5 aℓ4 (xα ′ )ℓ5 gℓαγ
20
Assume that some symbol a is aligned to a symbol to the right of xα ′ . Then sequence (xα ′ )ℓ5
will contribute ≥ ℓ5 /2 > E3 to the Dyck score and we are done. (We prove later that we can
achieve Dyck score E3 if there is a clique.) Now let r denote the number of symbols from sequence
j = xα ℓ5 aℓ4 (xα ′ )ℓ5 that are aligned to a symbol that does not belong to sequence j. Let s denote
the set of these r symbols. Let l denote symbols xα that are aligned to a symbol xα ′ . There are
2ℓ5 + ℓ4 − r − 2l symbols from j that are not considered yet. These symbols are not matched
to their counterparts and, therefore, contribute at least ⌈(2ℓ5 + ℓ4 − r − 2l)/2⌉ to the Dyck score
(we divide by 2 because the symbols can be mismatched among themselves in pairs). We have
that ⌈(2ℓ5 + ℓ4 − r − 2l)/2⌉ + r ≥ ℓ4 /2 + ⌈r/2⌉ by the definition of l (it implies that l ≤ ℓ5 ).
We note that ℓ4 /2 = Dyck(j). Dyck(j) ≤ ℓ4 /2 can be obtained by aligning symbols xα with
symbols xα ′ and mismatching symbols a in pairs. The reverse inequality follows by observing that
all symbols a will be mismatched. Also we note that, if we mismatch symbols from s among
themselves, this costs ⌈r/2⌉. All this gives that we can assume that symbols in j do not interact
with symbols that are not in j when we want to bound the Dyck score. Similarly, we can argue
when j = yα ℓ5 (a′ )ℓ4 (yα ′ )ℓ5 , xγ ℓ5 bℓ4 (xγ ′ )ℓ5 , yβ T (b′ )ℓ4 (yβ ′ )ℓ5 , xγ ℓ5 cℓ4 (xγ ′ )ℓ5 , yγ ℓ5 (c′ )ℓ4 (yγ ′ )ℓ5 . Thus, we
need to show that
CN Gαγ (tα )
3
gℓαγ
3
gℓαβ
CN Gαβ (tα )
3
gℓαβ
(g′αβ )ℓ3
CLGαβ (tβ )
(g′αβ )ℓ3
3
gℓβγ
CN Gβγ (tβ )
3
gℓβγ
(g′βγ )ℓ3
CLGβγ (tγ )
(g′βγ )ℓ3
(g′αγ )ℓ3
CLGαγ (tγ )
(g′αγ )ℓ3
3
Dyck( gℓαγ
) ≥ 3E2
with equality iff there is a biclique between vertices of tα and tβ , between vertices of tα and tγ and
between vertices of tβ and tγ . Let h be the argument to Dyck function, that is, we want to show
that Dyck(h) ≥ 3E2 with the stated condition for the equality.
Consider three gadgets CN G and three gadgets CLG as above. We can assume that no symbol
of any of these six gadgets is aligned to any symbol gxy or g′xy . Assume that it is not the case. Then
we can delete all symbols from the gadgets that are aligned to symbols gxy or g′xy . After this, we
rematch gxy or g′xy among themselves. We can check that we can always make this rematching of
symbols gxy or g′xy so that the cost do not increase. Furthermore, if some CN Gxy gadget is aligned
with a CN Gx′ y′ (or CLGx′ y′ ) for (x, y) 6= (x′ , y ′ ), then there are two substrings of the type gab or
g′ab that don’t have their counterpart between CN Gxy and CN Gx′ y′ (or CLGx′ y′ ). Hence, their
contribution to the Dyck score is at least 2ℓ3 /2 = ℓ3 ≫ 3E2 . Thus for all (x, y) the only gadget
that CN Gxy can be aligned with is CLGxy and vice versa. This means that we can assume that
all the g and g ′ symbols are completely aligned.
We have shown that the Dyck cost of the string is exactly
Dyck(CN Gαγ (tα ) ◦ CLGαγ (tα ))
+Dyck(CN Gαβ (tβ ) ◦ CLGαβ (tβ ))
+Dyck(CN Gβγ (tγ ) ◦ CLGβγ (tγ )).
We want to show that this is ≥ 3E2 with equality iff tα ∪ tβ ∪ tγ is a 3k-clique. This was shown
in Claim 8.
We now turn to the proof of the claim about the behavior of our “selection gadgets”. Let EC
be a fixed integer to be defined later that depends on n, N , and k.
21
Claim 10. If the G contains a 3k-clique, then Dyck(SG ) = EC and Dyck(SG ) > EC otherwise.
The proof of this claim will require several claims and lemmas. We start with some lemmas
about general properties of Dyck Edit Distance.
Lemma 3. Let Z1 be a substring of sequence Z. Assume that Z1 is of even length. If all symbols
symbols from Z1 participate in mismatches and deletions only, then we can modify the alignment
so that there is no symbol in Z1 that is aligned to a symbol that is not in Z1 .
Proof. Let l denote number of symbols from Z1 that are aligned to symbols that are outside of Z1 .
Let s denote the set of all these symbols outside of Z1 that are aligned to Z1 . There are two cases
to consider.
• l is even. Then the Dyck cost induced by symbols in s and in Z1 is at least S1 := l+(|Z1 |−l)/2
by the properties from the statement of the lemma. We do the following modification to the
alignment. We align all symbols in s among themselves in pairs. This induces cost S2 := l/2.
We align all symbols in Z1 among themselves. This induces cost S3 = |Z1 |/2. The total
induced cost after the modification is S2 + S3 ≤ S1 and we satisfy the requirement in the
lemma.
• l is odd. Then the Dyck cost induced by symbols in s and in Z1 is at least S1 := l + (|Z1 | −
l + 1)/2. We do the following modification to the alignment. We align all symbols in s
among themselves in pairs except one symbol, which we delete (remember that l is odd). This
induces cost S2 := (l + 1)/2. We align all symbols in Z1 among themselves. This induces cost
S3 = |Z1 |/2. The total induced cost after the modification is S2 + S3 ≤ S1 and we satisfy the
requirement in the lemma.
Consider optimal alignment of some string w ∈ (Σ ∪ Σ′ )∗ .
Lemma 4. Let Z be a maximal substring of w consisting entirely of symbols z (for some symbol
z appearing in w). Let Z0 be a maximal substring of w consisting entirely of symbols z0 . Let
|Z| = |Z0 | and let there are matches between Z and Z0 . Also, there are no other maximal substrings
of w containing z other than Z. Then we can increase the Dyck score by at most 2 by modifying
the alignment and get that the symbols of Z0 , that are matched to z, form a substring of Z0 and
the substring is suffix or prefix of Z0 (we can choose whether it is a suffix of a prefix). We can also
assume that the rest of symbols in Z0 are deleted or mismatched among themselves.
Proof. Wlog, we will show that we can make the substring to be the suffix. We write Z0 = Z1 Z2 Z3
so that the first symbol of Z2 is the first symbol of Z0 that is aligned to some symbol z and the
last symbol of Z2 is the last symbol of Z0 that is aligned to some symbol z. First, we modify the
alignment as follows. If the length of Z1 is even, we don’t do anything. Otherwise, consider the first
symbol of Z1 . If it is aligned to some symbol, delete the first symbol of Z1 and the symbol aligned
to it by increasing the Dyck cost by 1. Similarly delete the last symbol of Z3 and the symbol
aligned to it if Z3 is of odd length. Now, by Lemma 3, we can assume that all symbols in Z1 and Z3
are mismatched among themselves (except, possibly, the first symbol of Z1 and the last symbol of
Z3 ). Now we are at the state when all symbols from Z0 are mismatched among themselves except
few that are matched with symbols z from Z. Now we can rematch these symbols from Z with
symbols z0 from Z0 so that z0 come from suffix of Z0 . We see that this do not increase the Dyck
score besides two possible deletions.
22
Now we prove some claims about the properties of the optimal alignment of SG . These claims
essentially show that any “bad behaviour”, in which we do not align exactly one clique gadget of
each type, is suboptimal.
Let El be some value that depends only on n and k.
Claim 11. For any gadget CGβ (t) from the sequence, if none of the x′β , yβ symbols are matched
to their counterparts, then all the b symbols from CGβ (t) must be matched to their counterparts b′
from CGβ (t), and in this case, the gadget contributes El to the Dyck score. Analogous claims hold
for α, γ.
Proof. By Lemma 3 we can assume that all symbols x′β are mismatched among themselves. The
same we can say about symbols yβ . Let Z be the substring of the gadget between symbols x′β
and yβ . If Z has matches with some symbols, then, by the construction of w, no symbol b or
b′ is matched to its counterpart and by Lemma 3, we get that b and b′ are mismatched within
themselves. But now we can decrease the Dyck score by deleting all symbols from Z and all symbols
that Z is aligned to outside the gadget. This increases the Dyck score but then we can decrease
it substantially by matching all symbols b to their counterparts b′ from the gadget. In the end we
get smaller Dyck score because ℓ4 ≥ 100|Z|. Now it remains to consider the case when symbols
in Z do not participate in matches. But then by Lemma 3, we again conclude that symbols in Z
participate only in mismatches and only among themselves. Let s be the union of all symbols b
and b′ of the gadget and all symbols that these symbols are aligned to outside the gadget. Let l
denote the number of symbols from s that are not coming from the gadget. Consider two cases.
• l = 0. We satisfy requirements of the claim by matching all symbols b to their counterparts
b′ .
• There is no symbol among the l ≥ 1 symbols that participate in a match. We have that all
symbols in s contribute at least l to the Dyck score. We modify the alignment as follows.
We match all symbols b to their counterparts b′ . We match the rest l symbols from s among
themselves. If there is odd number of them, we delete one. This contributes at most S1 :=
(l + 1)/2 to the Dyck score after the modification. We have that S1 ≤ l for l ≥ 1.
• Complement of the previous two cases: there is a symbol among the l ≥ 1 symbols that
participates in a match. Wlog, symbols b from the gadget participate in at least one matching.
Then all symbols b′ from the gadget do not participate in any matching and by Lemma 3 we
have that all symbols b′ from the gadget are mismatched among themselves. Therefore, we
can assume that l ≤ ℓ4 . We need to consider two subcases.
– l is even. The symbols in s contribute at least S1 = (2ℓ4 − l)/2 to the Dyck score. We do
the following modification to the algorithm. We match all symbols b to their counterparts
b′ . We mismatch l symbols in pairs among themselves. After the modification, the Dyck
contribution of symbols from s is S2 := l/2. We see that S2 ≤ S1 .
– l is odd. The symbols in s contribute at least S1 = (2ℓ4 − l + 1)/2 to the Dyck score (at
least one symbol is deleted because 2ℓ4 − l is odd). We do the following modification to
the algorithm. We match all symbols b to their counterparts b′ . We mismatch l symbols
in pairs among themselves except that we delete one symbol. After the modification, the
Dyck contribution of symbols from s is S2 := (l + 1)/2. We see that S2 ≤ S1 .
Because symbols in between b and b′ don’t have their counterparts among themselves, the gadget
contributes El := (|CGβ (t)| − 2ℓ4 )/2 to the Dyck cost. This quantity only depends on n and k and
this can be verified from the construction of SG .
23
Claim 12. In some optimal alignment, we can assume that there is some symbol xβ that is aligned
to its counterpart x′β . Analogous statements can be proved about symbols yβ , y′β , xα , x′α , yα , y′α ,
xγ , x′γ , yγ , y′γ .
Proof. Suppose that xβ is not aligned to any symbol x′β . We will modify the alignment so that
all symbols xβ are aligned to x′β coming from the first substring f of SG consisting entirely of
x′β . From the statement we have that all symbol xβ and all symbols from f are mismatched or
deleted. Therefore, by Lemma 3 we get that all symbols xβ are mismatched among themselves and
all symbols in f are mismatched among themselves. Now we modify the alignment as follows to
achieve our goal. We delete all symbols b between xβ and f . Also, we delete all the symbols that
the deleted bs were aligned to before the deletion. This increases Dyck cost by at most 2ℓ4 . Then
we match all symbols xβ to x′β in pairs. This decreases Dyck cost by ℓ5 . As a result we decreases
the Dyck cost because ℓ5 ≫ 2ℓ4 .
Claim 13. In some optimal alignment, there is a symbol x′β that is matched to a symbol xβ and
there is a symbol yβ that is matched to a symbol y′β so that the symbols x′β and yβ come from the
same gadget CGβ (t). Analogous statements can be proved about symbols xα , x′α , yα , y′α , xγ , x′γ ,
yγ , y′γ .
Proof. By Claim 12, xβ is aligned to some sequence consisting of x′β . Suppose that the sequence
comes from gadget CGβ (t1 ) for some t1 . Also, by Claim 12, y′β is aligned to some sequence consisting
of yβ . Suppose that the sequence comes from gadget CGβ (t2 ) for some t2 . We want to prove that
t1 = t2 . Suppose that this is not the case and CGβ (t1 ) comes to the left of CGβ (t2 ) (the order
can’t be reverse by the construction of SG and because the alignments can’t cross). Suppose that
there is some other gadget CGβ (t3 ) in between CGβ (t1 ) and CGβ (t2 ). Then we can verify that
CGβ (t1 ) satisfy the conditions of Claim 11 and we can assume that all symbols in CGβ (t3 ) are
aligned with symbols in CGβ (t3 ). Therefore, we can remove gadget CGβ (t3 ) from SG (because it
does not interact with symbols outside it) and this decreases the Dyck cost by El . We do that until
CGβ (t1 ) is to the left from CGβ (t2 ) and they are neighboring. Now we will change the alignment
so that xβ is aligned to a symbol x′β from CGβ (t2 ) and as a result we will decrease Dyck cost. Now
we can verify from the construction of SG that symbols b, yβ , b′ from CGβ (t1 ) do not participate
in matches. Also, symbols b and x′β from CGβ (t2 ) do not participate in matches. By Lemma 3, we
conclude that all these symbols have mismatches among themselves. Let g be the sequence between
symbols x′β and yβ in CGβ (t1 ). Now we modify the alignment so that the symbols in sequence g
have mismatches only among themselves and the symbols that were aligned to symbols in g are
deleted. This increases Dyck cost by at most 10|g| ≤ 100ℓ3 =: S1 . Some symbols x′β are aligned
to symbols outside CGβ (t1 ). We transfer these alignmets of symbols x′β from CGβ (t1 ) to symbols
x′β in CGβ (t2 ) so that we only have mismatches among x′β in CGβ (t1 ) and we don’t change the
Dyck cost. Now we can align all symbols b in CGβ (t1 ) to b′ in CGβ (t1 ) in pairs. This decreases
the Dyck cost by ℓ4 . In the end, we decreased the Dyck cost by ℓ4 − S1 > 0 and we proved what
we wanted.
Claim 14. In some optimal alignment, the only symbols x′β that symbols xβ are matched to, come
from the same gadget CGβ (t), and the only symbols yβ that symbols y′β are matched to, come from
the same gadget CGβ (t). In both cases it is the same gadget CGβ (t). Analogous statements can be
proved about symbols xα , x′α , yα , y′α , xγ , x′γ , yγ , y′γ .
24
Proof. Suppose that xβ is matched to symbols x′β coming from two different gadgets CGβ (t1 ) and
CGβ (t2 ). CGβ (t1 ) comes earlier in SG than CGβ (t2 ). Assume that there is no gadget CGβ (t3 ) in
between CGβ (t1 ) and CGβ (t2 ) in sequence SG . We can make this assumption because otherwise we
can remove CGβ (t3 ) as in Claim 11. We can check that all symbols x′β and yβ from CGβ (t3 ) do not
participate in matches and thus we satisfy the requirements of Claim 11. Now we can check that
symbols b, yβ , b′ in CGβ (t1 ) do not participate in matches. Also, symbols between x′β and yβ in
CGβ (t1 ) do not participate in matches. Also, symbols b′ do not participate in matches. Therefore,
by Lemma 3 we conclude that all these symbols participate in mismatches only among themselves.
Let X1 denote sequence of x′β from CGβ (t1 ) and X2 denote sequence of x′β from CGβ (t2 ). By
Lemma 4, we can assume that symbols x′β from X1 and X2 that are matched to xβ form suffix in
both sequences and the rest of symbols in both sequences are mismatched among themselves or
deleted. The corresponding modifications increases the Dyck cost by at most ≤ 4 =: S1 . Let Z1
be the suffix of X1 and Z2 be the suffix of X2 . |Z1 | + |Z2 | is less or equal to the total number of
symbols xβ in SG by the construction of SG . Suppose that |Z1 | is even. We mismatch all symbols
in Z1 among themselves and match resulting unmatched |Z1 | symbols xβ to x′β from X2 . This does
not change Dyck cost. Suppose that |Z1 | is odd, then there is a deletion among symbols in X1 that
are not in Z1 . We do mismatches among symbols Z1 and the one deleted. We match the resulting
unmatched |Z1 | symbols xβ to x′β from X2 . We can check that we can do this matching so that the
Dyck cost do not increase. Now we can match all symbols b from CGβ (t1 ) to their counterparts
b′ in CGβ (t1 ). This decreases the Dyck cost by S2 := ℓ4 . In total, we decrease the Dyck cost by
≥ S2 − S1 > 0.
We note that in the proof of Claim 14 we remove 3N − 3 cliques from the graph (N is the
number of k-cliques in the graph), each removal costing El . After all the removals, we arrive to a
sequence of form as required in Claim 9. Thus, we set EC := (3N − 3)El + Ec and our proof for
Claim 10 is finished.
We are now ready to show that the construction of SG from the graph G proves Theorem 3.
Reminder of Theorem 3 If Dyck edit distance on a sequence of lengthn can be solved in T (n)
time, then 3k-Clique on n node graphs can be solved in O T nk+O(1) time, for any k ≥ 1.
Moreover, the reduction is combinatorial.
Proof. Given a graph G on n nodes we construct the sequence SG as described above. The sequence
can be constructed in O(kO(1) · nk+O(1) ) time, by enumerating all subsets of k nodes and that it has
length O(kO(1) · nk+O(1) ). Thus, an algorithm for Dyck Edit Distance as in the statement returns
Dyck score of SG in T (nk/3+O(1) ) time (treating k as a constant) and by Claim 10 this score
determines whether G contains a 3k-clique. All the steps in our reduction are combinatorial.
Acknowledgements. We would like to thank Piotr Indyk for a discussion that led to this work,
and Roy Frostig for introducing us to many modern works on CFG parsing. We also thank Alex Andoni, Ryan Williams, and the anonymous reviewers for comments. A.B. was supported by NSF and
Simons Foundation. A.A. and V.V.W. were supported by a Stanford School of Engineering Hoover
Fellowship, NSF Grant CCF-1417238, NSF Grant CCF-1514339, and BSF Grant BSF:2012338.
25
References
[AAK+ 07] Noga Alon, Alexandr Andoni, Tali Kaufman, Kevin Matulef, Ronitt Rubinfeld, and
Ning Xie. Testing k-wise and almost k-wise independence. In Proceedings of the 39th
Annual ACM Symposium on Theory of Computing, San Diego, California, USA, June
11-13, 2007, pages 496–505, 2007.
[AB87]
Noga Alon and Ravi B Boppana. The monotone circuit complexity of boolean functions. Combinatorica, 7(1):1–22, 1987.
[ABV15]
Amir Abboud, Arturs Backurs, and Virginia Vassilevska Williams. Tight hardness
results for lcs and other sequence similarity measures. In 56th IEEE Annual Symposium
on Foundations of Computer Science, FOCS, 2015. to appear.
[AKO10]
Alexandr Andoni, Robert Krauthgamer, and Krzysztof Onak. Polylogarithmic approximation for edit distance and the asymmetric query complexity. In Foundations
of Computer Science (FOCS), 2010 51st Annual IEEE Symposium on, pages 377–386.
IEEE, 2010.
[AKS98]
Noga Alon, Michael Krivelevich, and Benny Sudakov. Finding a large hidden clique
in a random graph. Random Struct. Algorithms, 13(3-4):457–466, 1998.
[Aku98]
Tatsuya Akutsu. Approximation and exact algorithms for RNA secondary structure
prediction and recognition of stochastic context-free languages. In Algorithms and
Computation, 9th International Symposium, ISAAC ’98, Taejon, Korea, December
14-16, 1998, Proceedings, pages 337–346, 1998.
[ALW14]
Amir Abboud, Kevin Lewi, and Ryan Williams. Losing weight by gaining edges.
In Algorithms - ESA 2014 - 22th Annual European Symposium, Wroclaw, Poland,
September 8-10, 2014. Proceedings, pages 1–12, 2014.
[And14]
Alexandr
Andoni.
Question
on
RNA
http://sublinear.info/index.php?title=Open_Problems:61, 2014.
accessed 31-March-2015].
[AP72]
Alfred V Aho and Thomas G Peterson. A minimum distance error-correcting parser
for context-free languages. SIAM Journal on Computing, 1(4):305–312, 1972.
[ASU86]
Alfred V. Aho, Ravi Sethi, and Jeffrey D. Ullman. Compilers: Principles, Techniques,
and Tools. Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1986.
[Aut84]
Miscellaneous Authors. Queries and problems. SIGACT News, 16(3):38–47, 1984.
[Bak79]
James K Baker. Trainable grammars for speech recognition. The Journal of the
Acoustical Society of America, 65(S1):S132–S132, 1979.
[BI15]
Arturs Backurs and Piotr Indyk. Edit Distance Cannot Be Computed in Strongly
Subquadratic Time (unless SETH is false). In STOC’15, 2015.
[BK15]
Karl Bringmann and Marvin Künnemann. Quadratic conditional lower bounds for
string problems and dynamic time warping. arXiv preprint arXiv:1502.01063, 2015.
[BS07]
José-Miguel Benedı́ and Joan-Andreu Sánchez. Fast stochastic context-free parsing: A
stochastic version of the valiant algorithm. In Pattern Recognition and Image Analysis,
pages 80–88. Springer, 2007.
26
folding.
[Online;
[BTZZU11] Rolf Backofen, Dekel Tsur, Shay Zakov, and Michal Ziv-Ukelson. Sparse RNA folding:
Time and space efficient algorithms. Journal of Discrete Algorithms, 9(1):12–31, 2011.
[CCF+ 05]
Jianer Chen, Benny Chor, Mike Fellows, Xiuzhen Huang, David W. Juedes, Iyad A.
Kanj, and Ge Xia. Tight lower bounds for certain parameterized np-hard problems.
Inf. Comput., 201(2):216–231, 2005.
[Cha15]
Timothy M. Chan. Speeding up the four russians algorithm by about one more logarithmic factor. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium
on Discrete Algorithms, SODA 2015, San Diego, CA, USA, January 4-6, 2015, pages
212–217, 2015.
[CHKX06] Jianer Chen, Xiuzhen Huang, Iyad A. Kanj, and Ge Xia. Strong computational lower
bounds via parameterized complexity. J. Comput. Syst. Sci., 72(8):1346–1367, 2006.
[Cho59]
Noam Chomsky. On certain formal properties of grammars. Information and Control,
2(2):137 – 167, 1959.
[CL15]
Timothy M. Chan and Moshe Lewenstein. Clustered Integer 3SUM via Additive Combinatorics. In STOC’15, 2015.
[CLRS09]
Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein. Introduction to Algorithms, Third Edition. The MIT Press, 3rd edition, 2009.
[CS70]
John Cocke and Jacob T. Schwartz. Programming languages and their compilers:
Preliminary notes (technical report) (2nd revised ed.). Technical report, CIMS, NYU,
1970.
[CSC13]
Shay B Cohen, Giorgio Satta, and Michael Collins. Approximate pcfg parsing using
tensor decomposition. In HLT-NAACL, pages 487–496, 2013.
[DEKM98] R. Durbin, S. Eddy, A. Krogh, and G. Mitchison. Biological Sequence Analysis. Cambridge University Press, 1998.
[DeR69]
Franklin L DeRemer. Practical translators for LR (k) languages. PhD thesis, Massachusetts Institute of Technology, 1969.
[Ear70]
Jay Earley. An efficient context-free parsing algorithm. Commun. ACM, 13(2):94–102,
1970.
[EG04]
Friedrich Eisenbrand and Fabrizio Grandoni. On the complexity of fixed parameter
clique and dominating set. Theoretical Computer Science, 326(1):57–67, 2004.
[FG09]
Yelena Frid and Dan Gusfield. A simple, practical and complete o (\ frac {nˆ 3}{\ log
n})-time algorithm for RNA folding using the Four-Russians speedup. In Algorithms
in Bioinformatics, pages 97–107. Springer, 2009.
[FM80]
Charles N. Fischer and Jon Mauney. On the role of error productions in syntactic error
correction. Computer Languages, 5(3):131–139, 1980.
[Gal69]
Hervé Gallaire. Recognition time of context-free languages by on-line turing machines.
Information and Control, 15(3):288–295, 1969.
[Gal14]
François Le Gall. Powers of tensors and fast matrix multiplication. In International
Symposium on Symbolic and Algebraic Computation, ISSAC ’14, Kobe, Japan, July
23-25, 2014, pages 296–303, 2014.
27
[GCS+ 00]
Robin R Gutell, JJ Cannone, Z Shang, Y Du, and MJ Serra. A story: unpaired
adenosine bases in ribosomal RNAs. Journal of molecular biology, 304(3):335–354,
2000.
[GRH80]
S. Graham, W. L. Ruzzo, and M. Harrison. An improved context-free recognizer. ACM
Trans. Program. Lang. Syst., 2(3):415–462, 1980.
[Has99]
Johan Hastad. Clique is hard to approximate within n1−ε .
182(1):105–142, 1999.
[HH74]
Ivan M. Havel and Michael A. Harrison. On the parsing of deterministic languages. J.
ACM, 21:525–548, 1974.
[HK11]
Elad Hazan and Robert Krauthgamer. How hard is it to approximate the best nash
equilibrium? SIAM J. Comput., 40(1):79–91, 2011.
[HMU06]
John E. Hopcroft, Rajeev Motwani, and Jeffrey D. Ullman. Introduction to Automata
Theory, Languages, and Computation (3rd Edition). Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 2006.
[HS65]
Juris Hartmanis and Richard E Stearns. On the computational complexity of algorithms. Transactions of the American Mathematical Society, pages 285–306, 1965.
[Jer92]
Mark Jerrum. Large cliques elude the metropolis process. Random Struct. Algorithms,
3(4):347–360, 1992.
[JM00]
Daniel Jurafsky and James H. Martin. Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition. Prentice Hall PTR, Upper Saddle River, NJ, USA, 1st edition, 2000.
[Kar72]
Richard M. Karp. Reducibility among combinatorial problems. In Proceedings of a
symposium on the Complexity of Computer Computations, held March 20-22, 1972,
at the IBM Thomas J. Watson Research Center, Yorktown Heights, New York., pages
85–103, 1972.
[Kas65]
Tadao Kasami. An efficient recognition and syntax algorithm for context-free languages. Technical Report AFCRL-65-758, Air Force Cambridge Research Lab, Bedford, MA., 1965.
[Knu65]
Donald E Knuth. On the translation of languages from left to right. Information and
control, 8(6):607–639, 1965.
[KSSY13]
Flip Korn, Barna Saha, Divesh Srivastava, and Shanshan Ying. On repairing structural
problems in semi-structured data. Proc. VLDB Endow., 6(9):601–612, July 2013.
[Lan74]
Bernard Lang.
Springer, 1974.
[Lee02]
Lillian Lee. Fast context-free grammar parsing requires fast boolean matrix multiplication. J. ACM, 49(1):1–15, 2002.
[Mye95]
Gene Myers. Approximately matching context-free languages. Information Processing
Letters, 54(2):85–92, 1995.
[NP85]
J. Nešetřil and S. Poljak. On the complexity of the subgraph problem. Commentationes
Math. Universitatis Carolinae, 26(2):415–419, 1985.
Acta Mathematica,
Deterministic techniques for efficient non-deterministic parsers.
28
[PK09]
Adam Pauls and Dan Klein. K-best a* parsing. In Proceedings of the Joint Conference
of the 47th Annual Meeting of the ACL and the 4th International Joint Conference
on Natural Language Processing of the AFNLP: Volume 2-Volume 2, pages 958–966.
Association for Computational Linguistics, 2009.
[PTZZU11] Tamar Pinhas, Dekel Tsur, Shay Zakov, and Michal Ziv-Ukelson. Edit distance with
duplications and contractions revisited. In Combinatorial Pattern Matching, pages
441–454. Springer, 2011.
[RN14]
Sanguthevar Rajasekaran and Marius Nicolae. An error correcting parser for context
free grammars that takes less than cubic time. arXiv preprint arXiv:1406.3405, 2014.
[RSCJ10]
Alexander M Rush, David Sontag, Michael Collins, and Tommi Jaakkola. On dual
decomposition and linear programming relaxations for natural language processing.
In Proceedings of the 2010 Conference on Empirical Methods in Natural Language
Processing, pages 1–11. Association for Computational Linguistics, 2010.
[Ruz79]
Walter Ruzzo. On the complexity of general context-free language parsing and recognition (extended abstract). In Proceedings of the 6th Colloquium, on Automata, Languages and Programming, pages 489–497, 1979.
[Ryt85]
Wojciech Rytter. Fast recognition of pushdown automaton and context-free languages.
Information and Control, 67(1–3):12 –22, 1985.
[Ryt95]
Wojciech Rytter. Context-free recognition via shortest paths computation: A version
of valiant’s algorithm. Theor. Comput. Sci., 143(2):343–352, 1995.
[Sah14a]
Barna Saha. The dyck language edit distance problem in near-linear time. In 55th IEEE
Annual Symposium on Foundations of Computer Science, FOCS 2014, Philadelphia,
PA, USA, October 18-21, 2014, pages 611–620, 2014.
[Sah14b]
Barna Saha. Faster language edit distance, connection to all-pairs shortest paths and
related problems. CoRR, abs/1411.7315, 2014.
[Sah15]
Barna Saha. Language edit distance and maximum likelihood parsing of stochastic
grammars: Faster algorithms and connection to fundamental graph problems. In 56th
IEEE Annual Symposium on Foundations of Computer Science, FOCS, 2015. to appear.
[Sat94]
Giorgio Satta. Tree-adjoining grammar parsing and boolean matrix multiplication.
Comput. Linguist., 20:173–191, 1994.
[SBMN13] Richard Socher, John Bauer, Christopher D Manning, and Andrew Y Ng. Parsing with
compositional vector grammars. In In Proceedings of the ACL conference. Citeseer,
2013.
[Sei86]
Joel I Seiferas. A simplified lower bound for context-free-language recognition. Information and Control, 69(1):255–260, 1986.
[Sip96]
Michael Sipser. Introduction to the Theory of Computation. International Thomson
Publishing, 1st edition, 1996.
[Son15]
Yinglei Song. Time and space efficient algorithms for RNA folding with the FourRussians technique. arXiv preprint arXiv:1503.05670, 2015.
[Str69]
Volker Strassen. Gaussian elimination is not optimal.
13(4):354–356, 1969.
29
Numerische Mathematik,
[Val75]
Leslie G. Valiant. General context-free recognition in less than cubic time. Journal of
Computer and System Sciences, 10(2):308 – 315, 1975.
[Vas09]
Virginia Vassilevska. Efficient algorithms for clique problems. Inf. Process. Lett.,
109(4):254–257, 2009.
[VGF13]
Balaji Venkatachalam, Dan Gusfield, and Yelena Frid. Faster algorithms for RNAfolding using the Four-Russians method. In WABI, pages 126–140, 2013.
[Wei76]
Reinhold Weicker. General context-free language recognition by a ram with uniform
cost criterion in time n2 logn. Technical Report 182, Pennsylvania State University,
1976.
[Wil05]
Ryan Williams. A new algorithm for optimal 2-constraint satisfaction and its implications. Theoretical Computer Science, 348(2):357–365, 2005.
[Wil12]
Virginia Vassilevska Williams. Multiplying matrices faster than coppersmith-winograd.
In Proceedings of the 44th symposium on Theory of Computing, pages 887–898. ACM,
2012.
[Wil14]
Ryan Williams. Faster all-pairs shortest paths via circuit complexity. STOC, 2014.
[Woe04]
Gerhard J. Woeginger. Space and time complexity of exact algorithms: Some open
problems. In Parameterized and Exact Computation, volume 3162 of Lecture Notes in
Computer Science, pages 281–290, 2004.
[Woe08]
Gerhard J Woeginger. Open problems around exact algorithms. Discrete Applied
Mathematics, 156(3):397–405, 2008.
[WW10]
V. Vassilevska Williams and R. Williams. Subcubic equivalences between path, matrix
and triangle problems. In Proc. FOCS, pages 645–654, 2010.
[You67]
Daniel H. Younger. Recognition and parsing of context-free languages in time n3 .
Information and Control, 10(2):189–208, 1967.
[Yu15]
Huacheng Yu. An improved combinatorial algorithm for boolean matrix multiplication.
In Automata, Languages, and Programming - 42nd International Colloquium, ICALP
2015, Kyoto, Japan, July 6-10, 2015, Proceedings, Part I, pages 1094–1105, 2015.
[ZTZU10]
Shay Zakov, Dekel Tsur, and Michal Ziv-Ukelson. Reducing the worst case running
times of a family of RNA and CFG problems, using Valiant’s approach. In WABI,
pages 65–77, 2010.
30
| 8cs.DS
|
MARFCAT: Transitioning to Binary and Larger
Data Sets of SATE IV
Serguei A. Mokhov1,2 , Joey Paquet1 , Mourad Debbabi1 , Yankui Sun2
arXiv:1207.3718v2 [cs.CR] 10 May 2013
1
Concordia University
Montreal, QC, Canada
{mokhov,paquet,debbabi}@encs.concordia.ca
2
Tsinghua University
Beijing, China
[email protected]
Abstract
We present a second iteration of a machine learning approach to static code analysis and
fingerprinting for weaknesses related to security, software engineering, and others using the
open-source MARF framework and the MARFCAT application based on it for the NIST’s
SATE IV static analysis tool exposition workshop’s data sets that include additional test
cases, including new large synthetic cases. To aid detection of weak or vulnerable code,
including source or binary on different platforms the machine learning approach proved to
be fast and accurate to for such tasks where other tools are either much slower or have much
smaller recall of known vulnerabilities. We use signal and NLP processing techniques in our
approach to accomplish the identification and classification tasks. MARFCAT’s design from
the beginning in 2010 made is independent of the language being analyzed, source code,
bytecode, or binary. In this follow up work with explore some preliminary results in this
area. We evaluated also additional algorithms that were used to process the data.
Contents
1 Introduction
2
2 Related Work
3
3 Data Sets
4
4 Methodology
4.1 Methodology Overview . . . . . . . . . . . . . . . . .
4.2 CVEs and CWEs – the Knowledge Base . . . . . . .
4.3 Categories for Machine Learning . . . . . . . . . . .
4.4 Algorithms . . . . . . . . . . . . . . . . . . . . . . .
4.4.1 Signal Pipeline . . . . . . . . . . . . . . . . .
4.4.2 NLP Pipeline . . . . . . . . . . . . . . . . . .
4.5 Binary and Bytecode Analysis . . . . . . . . . . . . .
4.6 Wavelets . . . . . . . . . . . . . . . . . . . . . . . . .
4.7 Demand-Driven Distributed Evaluation with GIPSY
4.8 Export . . . . . . . . . . . . . . . . . . . . . . . . . .
4.8.1 SATE . . . . . . . . . . . . . . . . . . . . . .
4.8.2 Forensic Lucid . . . . . . . . . . . . . . . . .
4.8.3 SAFES . . . . . . . . . . . . . . . . . . . . .
4.9 Experiments . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5
5
6
6
7
7
7
7
8
9
10
10
10
11
11
1
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
5 Results
5.1 Preliminary Results Summary . . . . . . . .
5.2 Version SATE-IV.1 . . . . . . . . . . . . . .
5.2.1 Half-Training Data For Training and
5.3 Version SATE-IV.2 . . . . . . . . . . . . . .
5.4 Version SATE-IV.5 . . . . . . . . . . . . . .
5.4.1 Wavelet Experiments . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
Full For Testing
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
11
11
14
14
19
20
20
6 Conclusion
6.1 Shortcomings . . . . .
6.2 Advantages . . . . . .
6.3 Practical Implications
6.4 Future Work . . . . .
6.5 Acknowledgments . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
21
21
22
22
23
23
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
A Classification Result Tables
24
B Forensic Lucid Report Example
24
List of Tables
1
2
3
4
CVE Stats for Wireshark 1.2.0, Low-Pass FFT Filter Preprocessing . . . . . .
CVE Stats for Wireshark 1.2.0, Separating DWT Wavelet Filter Preprocessing
CWE Stats for Wireshark 1.2.0, Separating DWT Wavelet Filter Preprocessing
CWE Stats for Wireshark 1.2.0, Low-Pass FFT Filter Preprocessing . . . . . .
.
.
.
.
25
26
27
28
List of Figures
1
2
3
4
5
1
Machine-learning-based static code analysis testing algorithm using the signal
pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Machine-learning-based static code analysis testing algorithm using the NLP
pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A wave graph of a fraction of the CVE-2009-2562-vulnerable packet-afs.c in
Wireshark 1.2.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Spectrograms of CVE-2009-2562-vulnerable packet-afs.c in Wireshark 1.2.0,
fixed Wireshark 1.2.9 and Wireshark 1.2.18 . . . . . . . . . . . . . . . . . . . . .
A spectrogram of CVE-2009-2562-vulnerable packet-afs.c in Wireshark 1.2.0,
after SDWT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
9
13
13
21
Introduction
This is a follow up work on the first incarnation of MARFCAT detailed in [Mok10d, Mok11].
Thus, the majority of the results content here addresses the newer iteration duplicating only
the necessary background and methodology information (reduced). The reader is deferred to
consult the expanded background information and results in that previous work freely accessible
online (and the arXiv version of that is still occasionally updated).
2
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
We elaborate on the details of the expanded methodology and the corresponding results of
application of the machine learning techniques along with signal and NLP processing to static
source and binary code analysis in search for weaknesses and vulnerabilities. We use the tool,
named MARFCAT, a MARF-based Code Analysis Tool [Mok13], first exhibited at the Static
Analysis Tool Exposition (SATE) workshop in 2010 [ODBN10] to machine-learn from the (Common Vulnerabilities and Exposures) CVE-based vulnerable as well as synthetic CWE-based cases
to verify the fixed versions as well as non-CVE based cases from the projects written in same programming languages. The 2nd iteration of this work was prepared based on SATE IV [ODBN12]
and uses its updated data set and application. On the NLP side, we employ simple classical NLP techniques (n-grams and various smoothing algorithms), also combined with machine
learning for novel non-NLP applications of detection, classification, and reporting of weaknesses
related to vulnerabilities or bad coding practices found in artificial constrained languages, such
as programming languages and their compiled counterparts. We compare and contrast the NLP
approach to the signal processing approach in our results summary and illustrate concrete results
and for the same test cases.
We claim that the presented machine learning approach is novel and highly beneficial in
static analysis and routine testing of any kind of code, including source code and binary deployments for its efficiency in terms of speed, relatively high precision, robustness, and being a
complimentary tool to other approaches that do in-depth semantic analysis, etc. by prioritizing those tools’ targets. All that can be used in automatic manner in distributed and scalable
diverse environments to ensure the code safety, especially the mission critical software code in
all kinds of systems. It uses spectral, acoustic and language models to learn and classify such a
code.
This document, like its predecessor, is a “rolling draft” with several updates expected to be
made as the project progresses beyond SATE IV. It is accompanied with the updates to the
open-source MARFCAT tool itself [Mok13].
Organization
The related work, some of the present methodology is based on, is referenced in Section 2. The
data sets are described in Section 3. The methodology summary is in Section 4. We present
some of the results in Section 5 from the SAMATE reference test data set. Then we present
a brief summary, description of the limitations of the current realization of the approach and
concluding remarks in Section 6. In the Appendix there are classification result tables for specific
test cases illustrating top results by precision.
2
Related Work
To our knowledge this was the first time a machine learning approach was attempted to static
code analysis with the first results demonstrated during the SATE2010 workshop [Mok10d,
Mok13, ODBN10]. In the same year, a somewhat similar approach independently was presented
[BSSV10] for vulnerability classification and prediction using machine learning and SVMs, but
working with a different set of data.
Additional related work (to various degree of relevance or use) is further listed (this list is not
exhaustive). A taxonomy of Linux kernel vulnerability solutions in terms of patches and source
code as well as categories for both are found in [MLB07]. The core ideas and principles behind the
MARF’s pipeline and testing methodology for various algorithms in the pipeline adapted to this
case are found in [Mok08b, Mok10b] as it was the easiest implementation available to accomplish
3
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
the task. There also one can find the majority of the core options used to set the configuration
for the pipeline in terms of algorithms used. A binary analysis using machine learning approach
for quick scans for files of known types in a large collection of files is described in [MD08].
This includes the NLP and machine learning for NLP tasks in DEFT2010 [Mok10c, Mok10a]
with the corresponding DEFT2010App and its predecessor for hand-written image processing
WriterIdentApp [MSS09]. Tlili’s 2009 PhD thesis covers topics on automatic detection of
safety and security vulnerabilities in open source software [Tli09]. Statistical analysis, ranking,
approximation, dealing with uncertainty, and specification inference in static code analysis are
found in the works of Engler’s team [KTB+ 06, KAYE04, KE03]. Kong et al. further advance
static analysis (using parsing, etc.) and specifications to eliminate human specification from the
static code analysis in [KZL10]. Spectral techniques are used for pattern scanning in malware
detection by Eto et al. in [ESI+ 09]. Some researchers propose a general data mining system for
incident analysis with data mining engines in [IYE+ 09]. Hanna et al. describe a synergy between
static and dynamic analysis for the detection of software security vulnerabilities in [HLYD09]
paving the way to unify the two analysis methods. Other researchers propose a MEDUSA
system for metamorphic malware dynamic analysis using API signatures in [NJG+ 10]. Some
of the statistical NLP techniques we used, are described at length in [MS02]. BitBlaze (and its
web counterpart, WebBlaze) are other recent types of tools that to static and dynamic binary
code analysis for vulnerabilities fast, developed at Berkeley [Son10a, Son10b]. For wavelets, for
example, Li et al. [LjXP+ 09] have shown wavelet transforms and k-means classification can be
used to identify communicating applications on a network fast and is relevant to our study of
the code in any form, text or binary.
3
Data Sets
We use the SAMATE data set to practically validate our approach. The SAMATE reference
data set contains C/C++, Java, and PHP language tracks comprising CVE-selected cases as
well as stand-alone cases and the large generated synthetic C and Java test cases (CWE-based,
with a lot of variants of different known weaknesses). SATE IV expanded some cases from
SATE2010 by increasing the version number, and dropped some other cases (e.g., Chrome).
The C/C++ and Java test cases of various client and server OSS software are compilable
into the binary and object code, while the synthetic C and Java cases generated for various
CWE entries provided for greater scalability testing (also compilable). The CVE-selected cases
had a vulnerable version of a software in question with a list of CVEs attached to it, as well
as the most known fixed version within the minor revision number. One of the goals for the
CVE-based cases is to detect the known weaknesses outlined in CVEs using static code analysis
and also to verify if they were really fixed in the “fixed version” [ODBN12]. The cases with
known CVEs and CWEs were used as the training models described in the methodology. The
summary below is a union of the data sets from SATE2010 and SATE IV. The preliminary list
of the CVEs that the organizers expect to locate in the test cases were collected from the NVD
[NIS13a, ODBN12] for Wireshark 1.2.0, Dovecot, Tomcat 5.5.13, Jetty 6.1.16, and Wordpress
2.0. The specific test cases with versions and language at the time included CVE-selected:
• C: Wireshark 1.2.0 (vulnerable) and Wireshark 1.2.18 (fixed, up from Wireshark 1.2.9 in
SATE2010)
• C: Dovecot (vulnerable) and Dovecot (fixed)
• C++: Chrome 5.0.375.54 (vulnerable) and Chrome 5.0.375.70 (fixed)
4
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
• Java: Tomcat 5.5.13 (vulnerable) and Tomcat 5.5.33 (fixed, up from Tomcat 5.5.29 in
SATE2010)
• Java: Jetty 6.1.16 (vulnerable) and Jetty 6.1.26 (fixed)
• PHP: Wordpress 2.0 (vulnerable) and Wordpress 2.2.3 (fixed)
originally non-CVE selected in SATE2010:
• C: Dovecot
• Java: Pebble 2.5-M2
Synthetic CWE cases produced by the SAMATE team:
• C: Synthetic C covering 118 CWEs and ≈ 60K files
• Java: Synthetic Java covering ≈ 50 CWEs and ≈ 20K files
4
Methodology
In this section we outline the methodology of our approach to static source code analysis. Most
of this methodology is an updated description from [Mok10d]. The line number determination
methodology is also detailed in [Mok10d, ODBN10], but is not replicated here. Thus, the
methodology’s principles overview is described in Section 4.1, the knowledge base construction
is in Section 4.2, machine learning categories in Section 4.3, and the high-level algorithmic
description is in Section 4.4.
4.1
Methodology Overview
The core methodology principles include:
• Machine learning and dynamic programming
• Spectral and signal processing techniques
• NLP n-gram and smoothing techniques (add-δ, Witten-Bell, MLE, etc.)
We use signal processing techniques, i.e. presently we do not parse or otherwise work at the
syntax and semantics levels. We treat the source code as a “signal”, equivalent to binary, where
each n-gram (n = 2 presently, i.e. two consecutive characters or, more generally, bytes) are used
to construct a sample amplitude value in the signal. In the NLP pipeline, we similarly treat the
source code as a “characters”, where each n-gram (n = 1..3) is used to construct the language
model.
We show the system the examples of files with weaknesses and MARFCAT learns them
by computing spectral signatures using signal processing techniques or various language models (based on options) from CVE-selected test cases. When some of the mentioned techniques
are applied (e.g., filters, silence/noise removal, other preprocessing and feature extraction techniques), the line number information is lost as a part of this process.
When we test, we compute either how similar or distant each file is from the known trained-on
weakness-laden files or compare trained language models with the unseen language fragments in
5
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
the NLP pipeline. In part, the methodology can approximately be seen as some fuzzy signaturebased “antivirus” or IDS software systems detect bad signature, except that with a large number
of machine learning and signal processing algorithms and fuzzy matching, we test to find out
which combination gives the highest precision and best run-time.
At the present, however, we are looking at the whole files instead of parsing the finer-grain
details of patches and weak code fragments. This aspect lowers the precision, but is relatively
fast to scan all the code files.
4.2
CVEs and CWEs – the Knowledge Base
The CVE-selected test cases serve as a source of the knowledge base to gather information of
how known weak code “looks like” in the signal form [Mok10d], which we store as spectral
signatures clustered per CVE or CWE (Common Weakness Enumeration). The introduction by
the SAMATE team of a large synthetic code base with CWEs, serves as a part of knowledge
base learning as well. Thus, we:
• Teach the system from the CVE-based cases
• Test on the CVE-based cases
• Test on the non-CVE-based cases
For synthetic cases we do similarly:
• Teach the system from the CWE-based synthetic cases
• Test on the CWE-based synthetic cases
• Test on the CVE and non-CVE-based cases for CWEs from synthetic cases
We create index files in XML in the format similar to that of SATE to index all the file of
the test case under study. The CVE-based cases after the initial index generation are manually
annotated from the NVD database before being fed to the system. The script that does the
initial index gathering in the OSS distribution of MARFCAT is called collect-files-meta.
pl written in Perl. The synthetic cases required a special modification to that resulting in
collect-files-meta-synthetic.pl where there are no CVEs to fill in but CWEs alone, with
the auto-prefilled explanations since the information in the synthetic cases is not arbitrary and
controlled for identification.
4.3
Categories for Machine Learning
The tow primary groups of classes we train and test on include are naturally the CVEs [NIS13a,
NIS13b] and CWEs [VM13]. The advantages of CVEs is the precision and the associated meta
knowledge from [NIS13a, NIS13b] can be all aggregated and used to scan successive versions of
the the same software or derived products (e.g., WebKit in multiple browsers). CVEs are also
generally uniquely mapped to CWEs. The CWEs as a primary class, however, offer broader
categories, of kinds of weaknesses there may be, but are not yet well assigned and associated
with CVEs, so we observe the loss of precision. Since we do not parse, we generally cannot
deduce weakness types or even simple-looking aspects like line numbers where the weak code
may be. So we resort to the secondary categories, that are usually tied into the first two, which
we also machine-learn along, such as issue types (sink, path, fix) and line numbers.
6
MARFCAT: A MARF Approach to SATE IV
4.4
Mokhov, Paquet, Debbabi, Sun
Algorithms
In our methodology we systematically test and select the best (a tradeoff between speed and
accuracy) combination(s) of the algorithm implementations available to us and then use only
those for subsequent testing. This methodology is augmented with the cases when the knowledge
base for the same code type is learned from multiple sources (e.g., several independent C test
cases).
4.4.1
Signal Pipeline
Algorithmically-speaking, the steps that are performed in the machine-learning signal based
analysis are in Figure 1. The specific algorithms come from the classical literature and other
sources and are detailed in [Mok08b] and the related works. To be more specific for this work,
the loading typically refers to the interpretation of the files being scanned in terms of bytes
forming amplitude values in a signal (as an example, 8kHz or 16kHz frequency) using either unigram, bi-gram, or tri-gram approach. Then, the preprocessing allows to be none at all (“raw”,
or the fastest), normalization, traditional frequency domain filters, wavelet-based filters, etc.
Feature extraction involves reducing an arbitrary length signal to a fixed length feature vector
of what thought to be the most relevant features are in the signal (e.g., spectral features in FFT,
LPC), min-max amplitudes, etc. Classification stage is then separated either to train by learning
the incoming feature vectors (usually k-means clusters, median clusters, or plain feature vector
collection, combined with e.g., neural network training) or testing them against the previously
learned models.
4.4.2
NLP Pipeline
The steps that are performed in NLP and the machine-learning based analysis are presented
in Figure 2. The specific algorithms again come from the classical literature (e.g., [MS02]) and
are detailed in [Mok10b] and the related works. To be more specific for this work, the loading
typically refers to the interpretation of the files being scanned in terms of n-grams: uni-gram,
bi-gram, or tri-gram approach and the associated statistical smoothing algorithms, the results
of which (a vector, 2D or 3D matrix) are stored.
4.5
Binary and Bytecode Analysis
In this iteration we also perform preliminary Java bytecode and compiled C code static analysis
and produce results using the same signal processing, NLP, combined with machine learning
and data mining techniques. At this writing, the NIST SAMATE synthetic reference data
set for Java and C was used. The algorithms presented in Section 4.4 are used as-is in this
scenario with the modifications to the index files. The modifications include removal of the
line numbers, source code fragments, and lines-of-text counts (which are largely meaningless
and ignored. The byte counts may be recomputed and capturing a byte offset instead of a line
number was projected. The filenames of the index files were updated to include -bin in them
to differentiate from the original index files describing the source code. Another point is at the
moment the simplifying assumption is that each compilable source file e.g., .java or .c produce
the corresponding .class and .o files that we examine. We do not examine inner classes or
linked executables or libraries at this point.
7
MARFCAT: A MARF Approach to SATE IV
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Mokhov, Paquet, Debbabi, Sun
// Construct an index mapping CVEs to files and locations within files
Compile meta-XML index files from the CVE reports (line numbers, CVE, CWE,
fragment size, etc.). Partly done by a Perl script and partly annotated manually;
foreach source code base, binary code base do
// Presently in these experiments we use simple mean clusters of
feature vectors or unigram language models per default MARF
specification ([Mok08b])
Train the system based on the meta index files to build the knowledge base (learn);
begin
Load (interpret as a wave signal or n − gram);
Preprocess (none, FFT-filters, wavelets, normalization, etc.);
Extract features (FFT, LPC, min-max, etc.);
Train (Similarity, Distance, Neural Network, etc.);
end
Test on the training data for the same case (e.g., Tomcat 5.5.13 on Tomcat 5.5.13)
with the same annotations to make sure the results make sense by being high and
deduce the best algorithm combinations for the task;
begin
Load (same);
Preprocess (same);
Extract features (same);
Classify (compare to the trained k-means, or medians, or language models);
Report;
end
Similarly test on the testing data for the same case (e.g., Tomcat 5.5.13 on Tomcat
5.5.13) without the annotations as a sanity check;
Test on the testing data for the fixed case of the same software (e.g., Tomcat 5.5.13
on Tomcat 5.5.33);
Test on the testing data for the general non-CVE case (e.g., Tomcat 5.5.13 on Pebble
or synthetic);
end
Figure 1: Machine-learning-based static code analysis testing algorithm using the signal
pipeline
4.6
Wavelets
As a part of a collaboration project with Dr. Yankui Sun from Tsinghua University, waveletbased signal processing for the purposes of noise filtering is being introduced with this work to
compare it to no-filtering, or FFT-based classical filtering. It’s been also shown in [LjXP+ 09] that
wavelet-aided filtering could be used as a fast preprocessing method for a network application
identification and traffic analysis [LKW08].
We rely in part on the the algorithm and methodology found in [AS01, SCL+ 03, KBC05,
KBC06], and at this point only a separating 1D discrete wavelet transform (SDWT) has been
tested (see Section 5.4.1).
Since the original wavelet implementation [SCL+ 03] is in MATLAB [Mat12a], [Sch07], we
used in part the codegen tool from the MATLAB Coder toolbox [Mat12b], [Mat12c] to generate
8
MARFCAT: A MARF Approach to SATE IV
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
Mokhov, Paquet, Debbabi, Sun
Compile meta-XML index files from the CVE reports (line numbers, CVE, CWE,
fragment size, etc.). Partly done by a Perl script and partly annotated manually;
foreach source code base, binary code base do
// Presently in these experiments we use simple unigram language models
per default MARF specification ([Mok10b])
Train the system based on the meta index files to build the knowledge base (learn);
begin
Load (n-gram);
Train (statistical smoothing estimators);
end
Test on the training data for the same case (e.g., Tomcat 5.5.13 on Tomcat 5.5.13)
with the same annotations to make sure the results make sense by being high and
deduce the best algorithm combinations for the task;
begin
Load (same);
Classify (compare to the trained language models);
Report;
end
Similarly test on the testing data for the same case (e.g., Tomcat 5.5.13 on Tomcat
5.5.13) without the annotations as a sanity check;
Test on the testing data for the fixed case of the same software (e.g., Tomcat 5.5.13
on Tomcat 5.5.33);
Test on the testing data for the general non-CVE case (e.g., Tomcat 5.5.13 on Pebble
or synthetic);
end
Figure 2: Machine-learning-based static code analysis testing algorithm using the NLP
pipeline
a rough C/C++ equivalent in order to (manually) translate some fragments into Java (the
language of MARF and MARFCAT). The specific function for up/down sampling used by the
wavelets function in [Mot09] written also C/C++ was translated to Java in MARF as well with
unit tests added.
4.7
Demand-Driven Distributed Evaluation with GIPSY
To enhance the scalability of the approach, we convert the MARFCAT stand-alone application to
a distributed one using an eductive model of computation (demand-driven) implemented in the
General Intensional Programming System (GIPSY)’s multi-tier run-time system [Han10, Ji11,
Vas05, Paq09], which can be executed distributively using Jini (Apache River), or JMS [JMP13].
To adapt the application to the GIPSY’s multi-tier architecture, we create a problem-specific
generator and worker tiers (PS-DGT and PS-DWT respectively) for the MARFCAT application.
The generator(s) produce demands of what needs to be computed in the form of a file (source
code file or a compiled binary) to be evaluated and deposit such demands into a store managed
by the demand store tier (DST) as pending. Workers pickup pending demands from the store,
and them process then (all tiers run on multiple nodes) using a traditional MARFCAT instance.
Once the result (a Warning instance) is computed, the PS-DWT deposit it back into the store
9
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
with the status set to computed. The generator “harvests” all computed results (warnings) and
produces the final report for a test cases. Multiple test cases can be evaluated simultaneously
or a single case can be evaluated distributively. This approach helps to cope with large amounts
of data and avoid recomputing warnings that have already been computed and cached in the
DST.
The initial basic experiment assumes the PS-DWTs have the training sets data and the test
cases available to them from the start (either by a copy or via an NFS/CIFS-mounted volumes);
thus, the distributed evaluation only concerns with the classification task only as of this version.
The follow up work will remove this limitation.
In this setup a demand represents a file (a path) to scan (actually an instance of the FileItem
object), which is deposited into the DST. The PS-DWT picks up that and checks the file per
training set that’s already there and returns a ResultSet object back into the DST under the
same demand signature that was used to deposit the path to scan. The result set is sorted from
the most likely to the list likely with a value corresponding to the distance or similarity. The
PS-DGT picks up the result sets and does the final output aggregation and saves report in one
of the desired report formats (see Section 4.8 picking up the top two results from the result set
and testing against a threshold to accept or reject the file (path) as vulnerable or not. This
effectively splits the monolithic MARFCAT application in two halves in distributing the work
to do where the classification half is arbitrary parallel.
Simplifying assumptions:
• Test case data and training sets are present on each node (physical or virtual) in advance (via a copy or a CIFS or NFS volume), so no demand driven training occurs, only
classification
• The demand assumes to contain only file information to be examined (FileItem)
• PS-DWT assumes a single pre-defined configuration, i.e. configuration for MARFCAT’s
option is not a part of the demand
• PS-DWT assume CVE or CWE testing based on its local settings and not via the configuration in a demand
4.8
4.8.1
Export
SATE
By default MARFCAT produces the report data in the SATE XML format, according to the
SATE IV requirements. In this iteration other formats are being considered and realized. To
enable multiple format output, the MARFCAT report generation data structures were adapted
case-based output.
4.8.2
Forensic Lucid
The first one, is Forensic Lucid, the author Mokhov’s PhD topic, a language to specify and
evaluate digital forensic cases by uniformly encoding the evidence and witness accounts (evidential statement or knowledge base) of any case from multiple sources (system specs, logs, human
accounts, etc.) as a description of an incident to further perform investigation and event reconstruction. Following the data export in Forensic Lucid in the preceding work [MPD08, MPD10,
Mok08a] we use it asa format for evidential processing of the results produced by MARFCAT.
10
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
The work [MPD08] provides details of the language; it will suffice to mention here that the report
generated by MARFCAT in Forensic Lucid is a collection of warnings as observations with the
hierarchical notion of nested context of warning and location information. These will form an
evidential statement in Forensic Lucid. The example scenario where such evidence compiled via
a MARFCAT Forensic Lucid report would be in web-based applications and web browser-based
incident investigations of fraud, XSS, buffer overflows, etc. linking CVE/CWE-based evidence
analysis of the code (binary or source) security bugs with the associated web-based malware
propagation or attacks to provide possible events where specific attacks can be traced back to
the specific security vulnerabilities.
4.8.3
SAFES
The third format, for which the export functionality is not done as of this writing, SAFES, is
the 3rd format for output of the MARFCAT. SAFES is becoming a standard to reporting such
information and the SATE organizers began endorsing it as an alternative during SATE IV.
4.9
Experiments
The below is the current summary of the conducted experiments:
• Re-testing of the newer fixed versions such as Wireshark 1.2.18 and Tomcat 5.5.33.
• Half-based testing of the previous versions by reducing the training set by half and but
testing for all known CVEs or CWEs for Wireshark 1.2.18, Tomcat 5.5.33, and Chrome
5.0.375.54.
• Testing the new test cases of Dovecot, Jetty 6.1.x, and Wordpress 2.x as well as Synthetic
C and Synthetic Java.
• Binary test on the Synthetic C and Synthetic Java test cases.
• Performing tests using wavelets for preprocessing.
5
Results
The preliminary results of application of our methodology are outlined in this section. We
summarize the top precisions per test case using either signal-processing or NLP-processing
of the CVE-based and synthetic cases and their application to the general cases. Subsequent
sections detail some of the findings and issues of MARFCAT’s result releases with different
versions. Some experiments we compare the results with the previously obtained ones [Mok10d]
where compatible and appropriate.
The results currently are being gradually released in the iterative manner that were obtained
through the corresponding versions of MARFCAT as it was being designed and developed.
5.1
Preliminary Results Summary
The results summarize the half-training-full-testing data vs. that of regular ones reported in
[Mok10d].
• Wireshark:
11
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
– CVEs (signal): 92.68%, CWEs (signal): 86.11%,
– CVEs (NLP): 83.33%, CWEs (NLP): 58.33%
• Tomcat:
– CVEs (signal): 83.72%, CWEs (signal): 81.82%,
– CVEs (NLP): 87.88%, CWEs (NLP): 39.39%
• Chrome:
– CVEs (signal): 90.91%, CWEs (signal): 100.00%,
– CVEs (NLP): 100.00%, CWEs (NLP): 88.89%
• Dovecot (new, 2.x):
– 14 warnings; but it appears all quality or false positive
– (very hard to follow the code, severely undocumented)
• Pebble:
– none found during quick testing
• Dovecot 1.2.x: (ongoing of this writing)
• Jetty: (ongoing of this writing)
• Wordpress: (ongoing of this writing)
What follows are some select statistical measurements of the precision in recognizing CVEs
and CWEs under different configurations using the signal processing and NLP processing techniques.
“Second guess” statistics provided to see if the hypothesis that if our first estimate of a
CVE/CWE is incorrect, the next one in line is probably the correct one. Both are counted if
the first guess is correct.
A sample signal visiusalization in the middle of a vulnerable file packet-afs.c in Wireshark
1.2.0 to CVE-2009-2562 is in Figure 3 in the wave form. The low “dips” represent the text line
endings (coupled with a preceding character (bytes) in bigrams (two PCM-signed bytes assumed
encoded in 8kHz representing the amplitude; normalized), which are often either semicolons,
closing or opening braces, brackets or parentheses). Only a small fragment is shown of roughly
300 bytes in length to be visually comprehensive of a nature of a signal we are dealing with.
In Figure 4, there are 3 spectrograms generated for the same file packet-afs.c. The first two
columns represent the CVE-2009-2562-vulnerable file, both versions are the same with ehanced
contrast to see the detail. The subsequent pairs are of the same file in Wireshark 1.2.9 and
Wireshark 1.2.18, where CVE-2009-2562 is no longer present. Small changes are noticeable
primarily in the bottom left and top right corners of the images, and even smaller elsewhere in
the images.
12
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
Figure 3: A wave graph of a fraction of the CVE-2009-2562-vulnerable packet-afs.c in Wireshark 1.2.0
Figure 4: Spectrograms of CVE-2009-2562-vulnerable packet-afs.c in Wireshark 1.2.0, fixed
Wireshark 1.2.9 and Wireshark 1.2.18
13
MARFCAT: A MARF Approach to SATE IV
5.2
Mokhov, Paquet, Debbabi, Sun
Version SATE-IV.1
5.2.1
Half-Training Data For Training and Full For Testing
This is one of the experiment per discussion with Aurelien Delaitre and SATE organizers. The
main idea is to test robustness and precision of the MARFCAT approach by artificially reducing
known weaknesses (their locations) to learn from by 50%, but test on the whole 100% to see how
much does precision degrade with such a reduction. Specifically, we supply only CWE classes
testing for this experiment (CVE classes make little sense). Only the first 50% of the entries
entries were used for training for Wireshark 1.2.0, Tomcat 5.5.13, and Chrome 5.0.375.54, while
the full 100% were used to test the precision changes. The below are the results.
It should be noted that CWE classification is generally less accurate due to a lot of dissimilar
items “stuffed” (by NVD) into very broad categories such as NVD-CWE-Other and NVD-CWEnoinfo when the data were collected. Additionally, since we arbitrarily picked the first 50% of
the training data, some of the CWEs simply were left out completely and not trained on if they
were entirely in the omitted half, so their individual precision is obviously 0% when tested for.
The archive contains the .log and the .xml files (the latter for now are in SATE format only
with the scientific notation +E3 removed). The best reports are:
report-cweidnoprepreprawfftcheb-wireshark-1.2.0-half-train-cwe.xml
report-cweidnoprepreprawfftdiff-wireshark-1.2.0-half-train-cwe.xml
The experiments are subdivided into regular (signal) and NLP based testing.
Signal
• Wireshark 1.2.0:
Reduction of the training data by half resulted in ≈ 14% precision drop compared to the
previous result (best 86.11% see the NIST report [Mok11], vs. 72.22% overall).
New results (by algorithms, then by CWEs):
14
MARFCAT: A MARF Approach to SATE IV
guess
1st
1st
1st
1st
1st
1st
2nd
2nd
2nd
2nd
2nd
2nd
guess
1st
1st
1st
1st
1st
2nd
2nd
2nd
2nd
2nd
run
1
2
3
4
5
6
1
2
3
4
5
6
run
1
2
3
4
5
1
2
3
4
5
algorithms
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
class
NVD-CWE-noinfo
CWE-20
CWE-119
NVD-CWE-Other
CWE-189
NVD-CWE-noinfo
CWE-20
CWE-119
NVD-CWE-Other
CWE-189
-raw
-raw
-raw
-raw
-raw
-raw
-raw
-raw
-raw
-raw
-raw
-raw
Mokhov, Paquet, Debbabi, Sun
-fft
-fft
-fft
-fft
-fft
-fft
-fft
-fft
-fft
-fft
-fft
-fft
-cheb
-diff
-eucl
-cos
-mink
-hamming
-cheb
-diff
-eucl
-cos
-mink
-hamming
good
26
26
22
25
17
17
30
30
24
32
23
24
good
68
38
18
9
0
84
39
29
11
0
bad
10
10
14
23
19
19
6
6
12
16
13
12
bad
39
22
14
8
12
23
21
3
6
12
%
72.22
72.22
61.11
52.08
47.22
47.22
83.33
83.33
66.67
66.67
63.89
66.67
%
63.55
63.33
56.25
52.94
0.00
78.50
65.00
90.62
64.71
0.00
• Tomcat 5.5.13:
Drop from 81.82% (see NIST report’s Table 7, p. 70) to 75% top result as a result (about
7 points) of training data reduction by 50%.
New precision estimates:
15
MARFCAT: A MARF Approach to SATE IV
guess
1st
1st
2nd
2nd
guess
1st
1st
1st
1st
1st
1st
1st
2nd
2nd
2nd
2nd
2nd
2nd
2nd
run
1
2
1
2
run
1
2
3
4
5
6
7
1
2
3
4
5
6
7
algorithms
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
class
CWE-264
CWE-255
CWE-200
CWE-22
CWE-79
CWE-119
CWE-20
CWE-264
CWE-255
CWE-200
CWE-22
CWE-79
CWE-119
CWE-20
-raw
-raw
-raw
-raw
Mokhov, Paquet, Debbabi, Sun
-fft
-fft
-fft
-fft
-diff
-hamming
-diff
-hamming
good
6
5
6
8
good
1
2
1
6
1
0
0
1
2
1
7
3
0
0
bad
2
9
2
6
bad
0
0
0
3
4
2
2
0
0
0
2
2
2
2
%
75.00
35.71
75.00
57.14
%
100.00
100.00
100.00
66.67
20.00
0.00
0.00
100.00
100.00
100.00
77.78
60.00
0.00
0.00
• Chrome 5.0.375.54:
Chrome result is for completeness even though it is not a test case for SATE IV. Chrome is
poor for some reason—drop from 100% (Table 5, p. 68) to 44.44%, but it’s only 9 entries.
The first result in the table is erroenous, i.e., with a poor recall (the sum of 2 + 0 < 9,
should be total 9).
16
MARFCAT: A MARF Approach to SATE IV
guess
1st
1st
1st
1st
1st
2nd
2nd
2nd
2nd
2nd
guess
1st
1st
1st
1st
1st
1st
1st
2nd
2nd
2nd
2nd
2nd
2nd
2nd
run
1
2
3
4
5
1
2
3
4
5
run
1
2
3
4
5
6
7
1
2
3
4
5
6
7
algorithms
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
-cweid -nopreprep
class
CWE-94
CWE-20
CWE-79
NVD-CWE-noinfo
NVD-CWE-Other
CWE-399
CWE-119
CWE-94
CWE-20
CWE-79
NVD-CWE-noinfo
NVD-CWE-Other
CWE-399
CWE-119
-raw
-raw
-raw
-raw
-raw
-raw
-raw
-raw
-raw
-raw
Mokhov, Paquet, Debbabi, Sun
-fft
-fft
-fft
-fft
-fft
-fft
-fft
-fft
-fft
-fft
-cos
-eucl
-cheb
-hamming
-mink
-cos
-eucl
-cheb
-hamming
-mink
good
2
4
3
3
2
2
4
4
4
3
good
6
3
2
2
1
0
0
6
3
3
3
2
0
0
bad
0
5
6
6
7
0
5
5
5
6
bad
3
2
2
2
7
4
4
3
2
1
1
6
4
4
%
100.00
44.44
33.33
33.33
22.22
100.00
44.44
44.44
44.44
33.33
%
66.67
60.00
50.00
50.00
12.50
0.00
0.00
66.67
60.00
75.00
75.00
25.00
0.00
0.00
NLP Generally this genre of classification was poor as before in this experiment, all around
40-45% percent precision.
• Wireshark 1.2.0:
New results (by algorithms, then by CWEs):
guess
1st
2nd
guess
1st
1st
1st
1st
1st
2nd
2nd
2nd
2nd
2nd
run
1
1
run
1
2
3
4
5
1
2
3
4
5
algorithms
-cweid -nopreprep -char -unigram -add-delta
-cweid -nopreprep -char -unigram -add-delta
class
NVD-CWE-noinfo
NVD-CWE-Other
CWE-119
CWE-20
CWE-189
NVD-CWE-noinfo
NVD-CWE-Other
CWE-119
CWE-20
CWE-189
good
15
23
good
11
1
2
1
0
17
1
4
1
0
bad
21
13
bad
7
1
3
9
1
1
1
1
9
1
%
41.67
63.89
%
61.11
50.00
40.00
10.00
0.00
94.44
50.00
80.00
10.00
0.00
17
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
• Tomcat 5.5.13:
Intriguingly, the best result is higher than with all of the date in the past report (42.42%
below vs. previous 39.39%).
guess
1st
2nd
guess
1st
1st
1st
1st
1st
1st
1st
1st
2nd
2nd
2nd
2nd
2nd
2nd
2nd
2nd
run
1
1
run
1
2
3
4
5
6
7
8
1
2
3
4
5
6
7
8
algorithms
-cweid -nopreprep -char -unigram -add-delta
-cweid -nopreprep -char -unigram -add-delta
class
CWE-255
CWE-264
CWE-119
CWE-20
CWE-22
CWE-200
CWE-79
CWE-16
CWE-255
CWE-264
CWE-119
CWE-20
CWE-22
CWE-200
CWE-79
CWE-16
good
14
18
good
1
2
1
1
7
1
1
0
1
2
1
1
11
1
1
0
bad
19
15
bad
0
0
0
0
9
3
6
1
0
0
0
0
5
3
6
1
%
42.42
54.55
%
100.00
100.00
100.00
100.00
43.75
25.00
14.29
0.00
100.00
100.00
100.00
100.00
68.75
25.00
14.29
0.00
good
4
5
good
1
1
1
1
0
0
0
1
1
1
1
0
0
1
bad
5
4
bad
0
0
0
1
1
2
1
0
0
0
1
1
2
0
%
44.44
55.56
%
100.00
100.00
100.00
50.00
0.00
0.00
0.00
100.00
100.00
100.00
50.00
0.00
0.00
100.00
• Chrome 5.0.375.54:
Here drop is twice as much (≈ 44% vs. 88%).
guess
1st
2nd
guess
1st
1st
1st
1st
1st
1st
1st
2nd
2nd
2nd
2nd
2nd
2nd
2nd
18
run
1
1
run
1
2
3
4
5
6
7
1
2
3
4
5
6
7
algorithms
-cweid -nopreprep -char -unigram -add-delta
-cweid -nopreprep -char -unigram -add-delta
class
NVD-CWE-noinfo
CWE-79
CWE-20
CWE-94
CWE-399
NVD-CWE-Other
CWE-119
NVD-CWE-noinfo
CWE-79
CWE-20
CWE-94
CWE-399
NVD-CWE-Other
CWE-119
MARFCAT: A MARF Approach to SATE IV
5.3
Mokhov, Paquet, Debbabi, Sun
Version SATE-IV.2
These runs represent using the same SATE2010 training data for Tomcat 5.5.13 and Wireshark
1.2.0 to test the updated fixed versions (as from SATE2010) to Tomcat 5.5.33 and Wireshark
1.2.18 using the same settings. At this run, no new CVEs that may have happened from the
previous fixed versions of Tomcat 5.5.29 and Wireshark 1.2.9 respectively in 2010 were added
to the training data for the versions being tested in this experiment as to see if any old issues
reoccur or not. In this short summary, both signal and NLP testing reveal no same known issues
found.
• SATE-IV.2-train-test-test-run-quick-tomcat-5-5-33-cve
This is CVE-based classical signal classification.
A typical MARFCAT run. Tomcat 5.5.13 used for training. For most reports, no warnings
were spotted based on what was learned from 5.5.13, so the reports convey earlier CVEs
were fixed.
Empty reports like:
report-noprepreprawfftcheb-train-test-test-run-quick-tomcat-5-5-33-cve.xml
However, the -cos report is noisy and non-empty:
report-noprepreprawfftcos-train-test-test-run-quick-tomcat-5-5-33-cve.xml
Overly detailed log files are also provided.
• SATE-IV.2-train-test-test-run-quick-tomcat-5-5-33-cwe
This is classical CWE-based testing.
A typical MARFCAT CWE run. Tomcat 5.5.13 used for training.
No warnings found based on the CVE data learned.
Most of the reports are empty, e.g.:
report-nopreprepcharunigramadddelta-train-test-test-run-quick-tomcat-5-5-33-cve-nlp.
xml
The -cos report is not as noisy as for CVEs, but still contains a couple of false positives.
report-cweidnoprepreprawfftcos-train-test-test-run-quick-tomcat-5-5-33-cwe.
xml
Overly detailed log files also provided.
Training and testing indexes are provided (*_test.xml and *_train.xml).
• SATE-IV.2-train-test-test-run-quick-tomcat-5-5-33-cve-nlp
This is CVE-based NLP testing.
A typical MARFCAT NLP run. Tomcat 5.5.13 used for training. Usually a slow run, so
only one configuration is tried. No warnings found based on the CVE data learned.
The only empty report is:
report-nopreprepcharunigramadddelta-train-test-test-run-quick-tomcat-5-5-33-cve-nlp.
xml
However, the -cos report is noisy and non-empty:
19
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
report-noprepreprawfftcos-train-test-test-run-quick-tomcat-5-5-33-cve.xml
Overly detailed log files also provided.
Training and testing indexes are provided (*_test.xml and *_train.xml).
• SATE-IV.2-train-test-test-run-quick-tomcat-5-5-33-cwe-nlp
This is CWE-based NLP testing.
A typical MARFCAT CWE NLP run. Tomcat 5.5.13 used for training. Usually a slow
run, so only one configuration is tried. No warnings found based on the CVE data learned.
The only empty report is:
report-cweidnopreprepcharunigramadddelta-train-test-test-run-quick-tomcat-5-5-33-cwe-nlp
xml
Overly detailed log files also provided.
• SATE-IV.2-train-test-test-run-quick-wireshark-1-2-18-cve
Test Wireshark 1.2.18 using the training data from Wireshark 1.2.0 and classical CVEbased processing.
Majority of algorithms returned empty reports. -cos was as noisy as usual, but -mink
was non-empty but quite short (though also presumed with false positives).
Empty reports:
report-noprepreprawfftcheb-train-test-test-run-quick-wireshark-1-2-18-cve.xml
report-noprepreprawfftdiff-train-test-test-run-quick-wireshark-1-2-18-cve.xml
report-noprepreprawffteucl-train-test-test-run-quick-wireshark-1-2-18-cve.xml
report-noprepreprawffthamming-train-test-test-run-quick-wireshark-1-2-18-cve.xml
Non empty reports:
report-noprepreprawfftcos-train-test-test-run-quick-wireshark-1-2-18-cve.xml
report-noprepreprawfftmink-train-test-test-run-quick-wireshark-1-2-18-cve.xml
Verbose log files and input index files are also supplied for the most cases.
[TODO]
5.4
5.4.1
Version SATE-IV.5
Wavelet Experiments
The preliminary experiments using the separating discreet wavelet transform (DWT) filter are
summarized in Table 2 and Table 3 for CVEs and CWEs respectively. For comparison, the
low-pass FFT filter is used for the same as shown in Table 1 and Table 4 respectively. For the
CVE experiments, the wavelet transforms overall produces better precision across configurations
(larger number of configurations produce higher precision result) than those with the low-pass
FFT filter. While the top precision result remains the same, it is shown than when filtering
is wanted, the wavelet transform is perhaps a better choice for some configurations, e.g., from
4 and below as well as for the Second Guess statistics. The very top result for the CWE
based processing so far exceeds the overall precision of separating DWT vs. low-pass FFT,
which then drops below for the subsequent configurations. -cos was dropped from Table 3 for
20
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
technical reasons. In Figure 5 is a spectrogram with the SDWT preprocessing in the pipeline.
More exploration in this area is under way for more advanced wavelet filters than the simple
separating DWT filter as to see whether they would outperform -raw or not and at the same
time minimizing the run-time performance decrease with the extra filtering.
Figure 5: A spectrogram of CVE-2009-2562-vulnerable packet-afs.c in Wireshark 1.2.0, after
SDWT
6
Conclusion
We review the current results of this experimental work, its current shortcomings, advantages,
and practical implications.
6.1
Shortcomings
Following is a list of the most prominent issues with the presented approach. Some of them are
more “permanent”, while others are solvable and intended to be addressed in the future work.
Specifically:
• Looking at a signal is less intuitive visually for code analysis by humans. (However, can
produce a problematic “spectrogram” in some cases).
• Line numbers are a problem (easily “filtered out” as high-frequency “noise”, etc.). A
whole “relativistic” and machine learning methodology developed for the line numbers
in [Mok10d] to compensate for that. Generally, when CVEs are the primary class, by
accurately identifying the CVE number one can get all the other pertinent details from
the CVE database, including patches and line numbers making this a lesser issue.
• Accuracy depends on the quality of the knowledge base (see Section 4.2) collected. Some of
this collection and annotation is manual to get the indexes right, and, hence, error prone.
“Garbage in – garbage out.”
• To detect more of the useful CVE or CWE signatures in non-CVE and non-CWE cases
requires large knowledge bases (human-intensive to collect), which can perhaps be shared
by different vendors via a common format, such as SATE, SAFES or Forensic Lucid.
• No path tracing (since no parsing is present); no slicing, semantic annotations, context,
locality of reference, etc. The “sink”, “path”, and “fix” results in the reports also have to
be machine-learned.
• A lot of algorithms and their combinations to try (currently ≈ 1800 permutations) to get
the best top N. This is, however, also an advantage of the approach as the underlying
framework can quickly allow for such testing.
21
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
• File-level training vs. fragment-level training – presently the classes are trained based
on the entire file where weaknesses are found instead of the known file fragments from
CVE-reported patches. The latter would be more fine-grained and precise than whole-file
classification, but slower. However, overall the file-level processing is a man-hour limitation
than a technological one.
• Separating wavelet filter performance is rather adversely affects the precision to low levels.
• No nice GUI. Presently the application is script/command-line based.
6.2
Advantages
There are some key advantages of the approach presented. Some of them follow:
• Relatively fast (e.g., Wireshark’s ≈ 2400 files train and test in about 3 minutes) on a
now-commodity desktop or a laptop.
• Language-independent (no parsing) – given enough examples can apply to any language,
i.e. methodology is the same no matter C, C++, Java or any other source or binary
languages (PHP, C#, VB, Perl, bytecode, assembly, etc.) are used.
• Can automatically learn a large knowledge base to test on known and unknown cases.
• Can be used to quickly pre-scan projects for further analysis by humans or other tools
that do in-depth semantic analysis as a means to prioritize.
• Can learn from SATE’08, SATE’09, SATE’10, and SATE IV reports.
• Generally, high precision (and recall) in CVE and CWE detection, even at the file level.
• A lot of algorithms and their combinations to select the best for a particular task or class
(see Section 4.3).
• Can cope with altered code or code used in other projects (e.g., a lot of problems in Chrome
were found it WebKit, used by several browsers).
6.3
Practical Implications
Most practical implications of all static code analyzers are obvious—to detect source code weaknesses and report them appropriately to the developers. We outline additional implications this
approach brings to the arsenal below:
• The approach can be used on any target language without modifications to the methodology or knowing the syntax of the language. Thus, it scales to any popular and new
language analysis with a very small amount of effort.
• The approach can nearly identically be transposed onto the compiled binaries and bytecode, detecting vulnerable deployments and installations—akin to virus scanning of binaries, but instead scanning for infected binaries, one would scan for security-weak binaries
on site deployments to alert system administrators to upgrade their packages.
• Can learn from binary signatures from other tools like Snort [Sou13].
• The approach is easily extendable to the embedded code and mission-critical code found
in aircraft, spacecraft, and various autonomous systems.
22
MARFCAT: A MARF Approach to SATE IV
6.4
Mokhov, Paquet, Debbabi, Sun
Future Work
There is a great number of possibilities in the future work. This includes improvements to
the code base of MARFCAT as well as resolving unfinished scenarios and results, addressing
shortcomings in Section 6.1, testing more algorithms and combinations from the related work,
and moving onto other programming languages (e.g., ASP, C#). Furthermore, foster collaboration with the academic, industry (such as VeraCode, Coverity) and government vendors and
others who have vast data sets to test the full potential of the approach with the others and a
community as a whole. Then, move on to dynamic code analysis as well applying similar techniques there. Other near-future work items include realization of the SVM-based classification,
data export in SAFES and Forensic Lucid formats, a lot of wavelet filtering improvements, and
distributed GIPSY cluster-based evaluation.
To improve detection and classification of the malware in the network traffic or otherwise
we employ machine learning approach to static pcap payload malicious code analysis and fingerprinting using the open-source MARF framework and its MARFCAT application, originally
designed for the SATE static analysis tool exposition workshop. We first train on the known
malware pcap data and measure the precision and then test it on the unseen, but known data
and select the best available machine learning combination to do so. This work elaborates
on the details of the methodology and the corresponding results of application of the machine
learning techniques along with signal processing and NLP alike to static network packet analysis in search for malicious code in the packet capture (pcap) data. malicious code analysis
[BOB+ 10, SEZS01, SXCM04, HJ07, HRSS07, Sue07, RM08, BOA+ 07] We show the system the
examples of pcap files with malware and MARFCAT learns them by computing spectral signatures using signal processing techniques. When we test, we compute how similar or distant
each file is from the known trained-on malware-laden files. In part, the methodology can approximately be seen as some signature-based “antivirus” or IDS software systems detect bad
signature, except that with a large number of machine learning and signal processing algorithms,
we test to find out which combination gives the highest precision and best run-time. At the
present, however, we are looking at the whole pcap files. This aspect lowers the precision, but
is fast to scan all the files. The malware database with known malware, the reports, etc. serves
as a knowledge base to machine-learn from. Thus, we primarily:
• Teach the system from the known cases of malware from their pcap data
• Test on the known cases
• Test on the unseen cases
6.5
Acknowledgments
The authors would like to express thanks and gratitude to the following for their help, resources,
advice, and otherwise support and assistance:
• NIST SAMATE group
• Dr. Brigitte Jaumard
• Sleiman Rabah
• Open-Source Community
23
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
This work is partially supported by the Faculty of ENCS, Concordia University, NSERC, and the
2011-2012 CCSEP scholarship. The wavelet-related work of Yankui Sun is partially supported
by the National Natural Science Foundation of China (No. 60971006).
A
Classification Result Tables
What follows are result tables with top classification results ranked from most precise at the
top. This include the configuration settings for MARF by the means of options (the algorithm
implementations are at their defaults [Mok08b]).
B
Forensic Lucid Report Example
An example report encoding the reported data in Forensic Lucid for Wireshark 1.2.0 after using
simple FFT-based feature extraction and Chebyshev distance as a classifier. The report provides
the same data, compressed, as the SATE XML, but in the Forensic Lucid syntax for automated
reasoning and event reconstruction during a digital investigation. The example is a an evidential
statement context encoded for the use in the investigator’s knowledge base of a particular case.
#FORENSICLUCID
evidential statement report_marfcat_0_0_2_SATE_IV_4
{
weakness_1 @ [id:1, tool_specific_id:1, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_1 = (locations_wk_1, 1, 0, 1.0);
locations_wk_1 = locations @ [tool_specific_id:1, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_340( [line => 828, path => "wireshark-1.2.0/epan/dissectors/packet-afs.c")
observation location_id_340( [line => 1718, path => "wireshark-1.2.0/epan/dissectors/packet-afs.c")
observation location_id_340( [line => 1729, path => "wireshark-1.2.0/epan/dissectors/packet-afs.c")
observation location_id_340( [line => 1740, path => "wireshark-1.2.0/epan/dissectors/packet-afs.c")
observation location_id_340( [line => 1747, path => "wireshark-1.2.0/epan/dissectors/packet-afs.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_2 @ [id:2, tool_specific_id:2, cweid:119, cwename:"Buffer Errors (CWE119)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_2 = (locations_wk_2, 1, 0, 1.0);
locations_wk_2 = locations @ [tool_specific_id:2, cweid:119, cwename:"Buffer Errors (CWE119)"];
observation location_id_411( [line => 830, path => "wireshark-1.2.0/epan/dissectors/packet-ber.c")
observation location_id_411( [line => 861, path => "wireshark-1.2.0/epan/dissectors/packet-ber.c")
observation location_id_411( [line => 885, path => "wireshark-1.2.0/epan/dissectors/packet-ber.c")
textoutput="";
observation grade = ([ severity => 1, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_3 @ [id:3, tool_specific_id:3, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_3 = (locations_wk_3, 1, 0, 0.004878625561362933);
locations_wk_3 = locations @ [tool_specific_id:3, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_433( [line => 669, path => "wireshark-1.2.0/epan/dissectors/packet-btl2cap.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 204.97576364943077], 1, 0, 0.004878625561362933);
end;
weakness_4 @ [id:4, tool_specific_id:4, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_4 = (locations_wk_4, 1, 0, 1.0);
locations_wk_4 = locations @ [tool_specific_id:4, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_550( [line => 248, path => "wireshark-1.2.0/epan/dissectors/packet-dcerpc-nt.c")
observation location_id_550( [line => 252, path => "wireshark-1.2.0/epan/dissectors/packet-dcerpc-nt.c")
observation location_id_550( [line => 256, path => "wireshark-1.2.0/epan/dissectors/packet-dcerpc-nt.c")
observation location_id_550( [line => 1138, path => "wireshark-1.2.0/epan/dissectors/packet-dcerpc-nt.c")
observation location_id_550( [line => 1142, path => "wireshark-1.2.0/epan/dissectors/packet-dcerpc-nt.c")
observation location_id_550( [line => 1146, path => "wireshark-1.2.0/epan/dissectors/packet-dcerpc-nt.c")
observation location_id_550( [line => 1201, path => "wireshark-1.2.0/epan/dissectors/packet-dcerpc-nt.c")
observation location_id_550( [line => 1205, path => "wireshark-1.2.0/epan/dissectors/packet-dcerpc-nt.c")
observation location_id_550( [line => 1209, path => "wireshark-1.2.0/epan/dissectors/packet-dcerpc-nt.c")
observation location_id_550( [line => 1314, path => "wireshark-1.2.0/epan/dissectors/packet-dcerpc-nt.c")
observation location_id_550( [line => 1318, path => "wireshark-1.2.0/epan/dissectors/packet-dcerpc-nt.c")
observation location_id_550( [line => 1322, path => "wireshark-1.2.0/epan/dissectors/packet-dcerpc-nt.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_5 @ [id:5, tool_specific_id:5, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
24
MARFCAT: A MARF Approach to SATE IV
Table
guess
1st
1st
1st
1st
1st
1st
2nd
2nd
2nd
2nd
2nd
2nd
guess
1st
1st
1st
1st
1st
1st
1st
1st
1st
1st
1st
1st
1st
1st
1st
1st
1st
1st
1st
1st
1st
1st
2nd
2nd
2nd
2nd
2nd
2nd
2nd
2nd
2nd
2nd
2nd
2nd
2nd
2nd
Mokhov, Paquet, Debbabi, Sun
1: CVE Stats for Wireshark 1.2.0, Low-Pass FFT Filter Preprocessing
run algorithms
good bad
%
1 -nopreprep -low -fft -cheb -flucid
37
4
90.24
2 -nopreprep -low -fft -diff -flucid
37
4
90.24
3 -nopreprep -low -fft -eucl -flucid
27
14
65.85
4 -nopreprep -low -fft -hamming -flucid
23
18
56.10
5 -nopreprep -low -fft -mink -flucid
22
19
53.66
6 -nopreprep -low -fft -cos -flucid
36
114
24.00
1 -nopreprep -low -fft -cheb -flucid
38
3
92.68
2 -nopreprep -low -fft -diff -flucid
38
3
92.68
3 -nopreprep -low -fft -eucl -flucid
34
7
82.93
4 -nopreprep -low -fft -hamming -flucid
26
15
63.41
5 -nopreprep -low -fft -mink -flucid
31
10
75.61
6 -nopreprep -low -fft -cos -flucid
39
111
26.00
run class
good bad
%
1 CVE-2009-3829
6
0
100.00
2 CVE-2009-4376
6
0
100.00
3 CVE-2010-0304
6
0
100.00
4 CVE-2010-2286
6
0
100.00
5 CVE-2010-2283
6
0
100.00
6 CVE-2009-3551
6
0
100.00
7 CVE-2009-3549
6
0
100.00
8 CVE-2009-3241
15
9
62.50
9 CVE-2009-2560
9
6
60.00
10 CVE-2010-1455
30
24
55.56
11 CVE-2009-2563
6
5
54.55
12 CVE-2009-2562
6
5
54.55
13 CVE-2009-2561
6
7
46.15
14 CVE-2009-4378
6
7
46.15
15 CVE-2010-2287
6
7
46.15
16 CVE-2009-3550
6
8
42.86
17 CVE-2009-3243
13
23
36.11
18 CVE-2009-4377
12
22
35.29
19 CVE-2010-2285
6
11
35.29
20 CVE-2009-2559
6
11
35.29
21 CVE-2010-2284
6
12
33.33
22 CVE-2009-3242
7
16
30.43
1 CVE-2009-3829
6
0
100.00
2 CVE-2009-4376
6
0
100.00
3 CVE-2010-0304
6
0
100.00
4 CVE-2010-2286
6
0
100.00
5 CVE-2010-2283
6
0
100.00
6 CVE-2009-3551
6
0
100.00
7 CVE-2009-3549
6
0
100.00
8 CVE-2009-3241
16
8
66.67
9 CVE-2009-2560
10
5
66.67
10 CVE-2010-1455
44
10
81.48
11 CVE-2009-2563
6
5
54.55
12 CVE-2009-2562
6
5
54.55
13 CVE-2009-2561
6
7
46.15
14 CVE-2009-4378
6
7
46.15
25
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
Table 2: CVE Stats for Wireshark 1.2.0, Separating DWT Wavelet Filter Preprocessing
guess run algorithms
good bad
1st
1 -nopreprep -sdwt -fft -diff -spectrogram -graph -flucid
37
4
1st
2 -nopreprep -sdwt -fft -cheb -spectrogram -graph -flucid
37
4
1st
3 -nopreprep -sdwt -fft -eucl -spectrogram -graph -flucid
27
14
1st
4 -nopreprep -sdwt -fft -hamming -spectrogram -graph -flucid
26
15
1st
5 -nopreprep -sdwt -fft -mink -spectrogram -graph -flucid
22
19
1st
6 -nopreprep -sdwt -fft -cos -spectrogram -graph -flucid
38
65
2nd
1 -nopreprep -sdwt -fft -diff -spectrogram -graph -flucid
39
2
2nd
2 -nopreprep -sdwt -fft -cheb -spectrogram -graph -flucid
39
2
2nd
3 -nopreprep -sdwt -fft -eucl -spectrogram -graph -flucid
35
6
2nd
4 -nopreprep -sdwt -fft -hamming -spectrogram -graph -flucid
29
12
2nd
5 -nopreprep -sdwt -fft -mink -spectrogram -graph -flucid
31
10
2nd
6 -nopreprep -sdwt -fft -cos -spectrogram -graph -flucid
39
64
guess run class
good bad
1st
1 CVE-2009-3829
6
0
1st
2 CVE-2009-2562
6
0
1st
3 CVE-2009-4378
6
0
1st
4 CVE-2010-2286
6
0
1st
5 CVE-2010-0304
6
0
1st
6 CVE-2009-4376
6
0
1st
7 CVE-2010-2283
6
0
1st
8 CVE-2009-3551
6
0
1st
9 CVE-2009-3550
6
0
1st
10 CVE-2009-3549
6
0
1st
11 CVE-2009-2563
6
2
1st
12 CVE-2009-2560
11
4
1st
13 CVE-2009-3241
15
9
1st
14 CVE-2010-1455
31
23
1st
15 CVE-2009-2561
6
6
1st
16 CVE-2010-2287
6
6
1st
17 CVE-2009-2559
6
6
1st
18 CVE-2009-3243
16
16
1st
19 CVE-2010-2285
6
7
1st
20 CVE-2009-4377
12
16
1st
21 CVE-2010-2284
6
9
1st
22 CVE-2009-3242
6
17
2nd
1 CVE-2009-3829
6
0
2nd
2 CVE-2009-2562
6
0
2nd
3 CVE-2009-4378
6
0
2nd
4 CVE-2010-2286
6
0
2nd
5 CVE-2010-0304
6
0
2nd
6 CVE-2009-4376
6
0
2nd
7 CVE-2010-2283
6
0
2nd
8 CVE-2009-3551
6
0
2nd
9 CVE-2009-3550
6
0
2nd
10 CVE-2009-3549
6
0
2nd
11 CVE-2009-2563
6
2
26
2nd
12 CVE-2009-2560
12
3
2nd
13 CVE-2009-3241
16
8
2nd
14 CVE-2010-1455
43
11
%
90.24
90.24
65.85
63.41
53.66
36.89
95.12
95.12
85.37
70.73
75.61
37.86
%
100.00
100.00
100.00
100.00
100.00
100.00
100.00
100.00
100.00
100.00
75.00
73.33
62.50
57.41
50.00
50.00
50.00
50.00
46.15
42.86
40.00
26.09
100.00
100.00
100.00
100.00
100.00
100.00
100.00
100.00
100.00
100.00
75.00
80.00
66.67
79.63
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
Table 3: CWE Stats for Wireshark 1.2.0, Separating DWT Wavelet Filter
guess run algorithms
good
1st
1 -cweid -nopreprep -sdwt -fft -diff -flucid
31
1st
2 -cweid -nopreprep -sdwt -fft -eucl -flucid
29
1st
3 -cweid -nopreprep -sdwt -fft -mink -flucid
17
1st
4 -cweid -nopreprep -sdwt -fft -hamming -flucid
14
2nd
1 -cweid -nopreprep -sdwt -fft -diff -flucid
33
2nd
2 -cweid -nopreprep -sdwt -fft -eucl -flucid
34
2nd
3 -cweid -nopreprep -sdwt -fft -mink -flucid
27
2nd
4 -cweid -nopreprep -sdwt -fft -hamming -flucid
23
guess run class
good
1st
1 CWE399
4
1st
2 CWE189
4
1st
3 NVD-CWE-Other
11
1st
4 CWE20
30
1st
5 NVD-CWE-noinfo
34
1st
6 CWE119
8
2nd
1 CWE399
4
2nd
2 CWE189
4
2nd
3 NVD-CWE-Other
11
2nd
4 CWE20
34
2nd
5 NVD-CWE-noinfo
53
2nd
6 CWE119
11
Preprocessing
bad
%
5
86.11
7
80.56
19
47.22
22
38.89
3
91.67
2
94.44
9
75.00
13
63.89
bad
%
0
100.00
0
100.00
1
91.67
10
75.00
34
50.00
8
50.00
0
100.00
0
100.00
1
91.67
6
85.00
15
77.94
5
68.75
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_5 = (locations_wk_5, 1, 0, 0.003778693428627627);
locations_wk_5 = locations @ [tool_specific_id:5, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_652( [line => 77, path => "wireshark-1.2.0/epan/dissectors/packet-dtls.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 264.64173897356557], 1, 0, 0.003778693428627627);
end;
weakness_6 @ [id:6, tool_specific_id:6, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_6 = (locations_wk_6, 1, 0, 0.004125022212036806);
locations_wk_6 = locations @ [tool_specific_id:6, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_763( [line => 8447, path => "wireshark-1.2.0/epan/dissectors/packet-gsm_a_rr.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 242.42293704067873], 1, 0, 0.004125022212036806);
end;
weakness_7 @ [id:7, tool_specific_id:7, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_7 = (locations_wk_7, 1, 0, 1.0);
locations_wk_7 = locations @ [tool_specific_id:7, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_863( [line => 945, path => "wireshark-1.2.0/epan/dissectors/packet-infiniband.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_8 @ [id:8, tool_specific_id:8, cweid:119, cwename:"Buffer Errors (CWE119)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_8 = (locations_wk_8, 1, 0, 1.0);
locations_wk_8 = locations @ [tool_specific_id:8, cweid:119, cwename:"Buffer Errors (CWE119)"];
observation location_id_877( [line => 2746, path => "wireshark-1.2.0/epan/dissectors/packet-ipmi-se.c")
observation location_id_877( [line => 2748, path => "wireshark-1.2.0/epan/dissectors/packet-ipmi-se.c")
observation location_id_877( [line => 2752, path => "wireshark-1.2.0/epan/dissectors/packet-ipmi-se.c")
textoutput="";
observation grade = ([ severity => 1, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_9 @ [id:9, tool_specific_id:9, cweid:998, cwename:"Other (NVD-CWE-Other)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_9 = (locations_wk_9, 1, 0, 1.0);
locations_wk_9 = locations @ [tool_specific_id:9, cweid:998, cwename:"Other (NVD-CWE-Other)"];
27
MARFCAT: A MARF Approach to SATE IV
Table 4:
guess run
1st
1
1st
2
1st
3
1st
4
1st
5
1st
6
2nd
1
2nd
2
2nd
3
2nd
4
2nd
5
2nd
6
guess run
1st
1
1st
2
1st
3
1st
4
1st
5
1st
6
2nd
1
2nd
2
2nd
3
2nd
4
2nd
5
2nd
6
Mokhov, Paquet, Debbabi, Sun
CWE Stats for Wireshark 1.2.0, Low-Pass FFT Filter Preprocessing
algorithms
good bad
%
-cweid -nopreprep -low -fft -diff -flucid
30
6
83.33
-cweid -nopreprep -low -fft -cheb -flucid
30
6
83.33
-cweid -nopreprep -low -fft -eucl -flucid
25
11 69.44
-cweid -nopreprep -low -fft -mink -flucid
20
16 55.56
-cweid -nopreprep -low -fft -cos -flucid
36
40 47.37
-cweid -nopreprep -low -fft -hamming -flucid
12
24 33.33
-cweid -nopreprep -low -fft -diff -flucid
31
5
86.11
-cweid -nopreprep -low -fft -cheb -flucid
31
5
86.11
-cweid -nopreprep -low -fft -eucl -flucid
30
6
83.33
-cweid -nopreprep -low -fft -mink -flucid
22
14 61.11
-cweid -nopreprep -low -fft -cos -flucid
48
28 63.16
-cweid -nopreprep -low -fft -hamming -flucid
16
20 44.44
class
good bad
%
CWE399
6
1
85.71
CWE20
48
12 80.00
NVD-CWE-Other
18
7
72.00
CWE189
6
3
66.67
NVD-CWE-noinfo
61
61 50.00
CWE119
14
19 42.42
CWE399
6
1
85.71
CWE20
48
12 80.00
NVD-CWE-Other
18
7
72.00
CWE189
6
3
66.67
NVD-CWE-noinfo
78
44 63.93
CWE119
22
11 66.67
observation location_id_882( [line => 792, path => "wireshark-1.2.0/epan/dissectors/packet-ipmi.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_10 @ [id:10, tool_specific_id:10, cweid:119, cwename:"Buffer Errors (CWE119)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_10 = (locations_wk_10, 1, 0, 1.0);
locations_wk_10 = locations @ [tool_specific_id:10, cweid:119, cwename:"Buffer Errors (CWE119)"];
observation location_id_969( [line => 523, path => "wireshark-1.2.0/epan/dissectors/packet-lwres.c")
textoutput="";
observation grade = ([ severity => 1, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_11 @ [id:11, tool_specific_id:11, cweid:20, cwename:"Input Validation (CWE20)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_11 = (locations_wk_11, 1, 0, 1.0);
locations_wk_11 = locations @ [tool_specific_id:11, cweid:20, cwename:"Input Validation (CWE20)"];
observation location_id_1099( [line => 62, path => "wireshark-1.2.0/epan/dissectors/packet-paltalk.c")
textoutput="";
observation grade = ([ severity => 1, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_12 @ [id:12, tool_specific_id:12, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_12 = (locations_wk_12, 1, 0, 0.004878625561362927);
locations_wk_12 = locations @ [tool_specific_id:12, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_1174( [line => 897, path => "wireshark-1.2.0/epan/dissectors/packet-radius.c")
observation location_id_1174( [line => 906, path => "wireshark-1.2.0/epan/dissectors/packet-radius.c")
observation location_id_1174( [line => 913, path => "wireshark-1.2.0/epan/dissectors/packet-radius.c")
observation location_id_1174( [line => 1005, path => "wireshark-1.2.0/epan/dissectors/packet-radius.c")
observation location_id_1174( [line => 1227, path => "wireshark-1.2.0/epan/dissectors/packet-radius.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 204.975763649431], 1, 0, 0.004878625561362927);
28
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
end;
weakness_13 @ [id:13, tool_specific_id:13, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_13 = (locations_wk_13, 1, 0, 1.0);
locations_wk_13 = locations @ [tool_specific_id:13, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_1282( [line => 1131, path => "wireshark-1.2.0/epan/dissectors/packet-sflow.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_14 @ [id:14, tool_specific_id:14, cweid:998, cwename:"Other (NVD-CWE-Other)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_14 = (locations_wk_14, 1, 0, 1.0);
locations_wk_14 = locations @ [tool_specific_id:14, cweid:998, cwename:"Other (NVD-CWE-Other)"];
observation location_id_1303( [line => 2141, path => "wireshark-1.2.0/epan/dissectors/packet-smb-pipe.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_15 @ [id:15, tool_specific_id:15, cweid:998, cwename:"Other (NVD-CWE-Other)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_15 = (locations_wk_15, 1, 0, 1.0);
locations_wk_15 = locations @ [tool_specific_id:15, cweid:998, cwename:"Other (NVD-CWE-Other)"];
observation location_id_1307( [line => 8757, path => "wireshark-1.2.0/epan/dissectors/packet-smb.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_16 @ [id:16, tool_specific_id:16, cweid:189, cwename:"Numeric Errors (CWE189)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_16 = (locations_wk_16, 1, 0, 1.0);
locations_wk_16 = locations @ [tool_specific_id:16, cweid:189, cwename:"Numeric Errors (CWE189)"];
observation location_id_1307( [line => 2195, path => "wireshark-1.2.0/epan/dissectors/packet-smb.c")
textoutput="";
observation grade = ([ severity => 2, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_17 @ [id:17, tool_specific_id:17, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_17 = (locations_wk_17, 1, 0, 0.008328136212759968);
locations_wk_17 = locations @ [tool_specific_id:17, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_1307( [line => 8457, path => "wireshark-1.2.0/epan/dissectors/packet-smb.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 120.07488523877028], 1, 0, 0.008328136212759968);
end;
weakness_18 @ [id:18, tool_specific_id:18, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_18 = (locations_wk_18, 1, 0, 0.008328136212759964);
locations_wk_18 = locations @ [tool_specific_id:18, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_1309( [line => 955, path => "wireshark-1.2.0/epan/dissectors/packet-smb2.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 120.07488523877032], 1, 0, 0.008328136212759964);
end;
weakness_19 @ [id:19, tool_specific_id:19, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_19 = (locations_wk_19, 1, 0, 0.004321352067642762);
locations_wk_19 = locations @ [tool_specific_id:19, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_1333( [line => 813, path => "wireshark-1.2.0/epan/dissectors/packet-ssl-utils.c")
observation location_id_1333( [line => 843, path => "wireshark-1.2.0/epan/dissectors/packet-ssl-utils.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 231.40905539443497], 1, 0, 0.004321352067642762);
end;
weakness_20 @ [id:20, tool_specific_id:20, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_20 = (locations_wk_20, 1, 0, 0.0021114804997331383);
locations_wk_20 = locations @ [tool_specific_id:20, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_1334( [line => 153, path => "wireshark-1.2.0/epan/dissectors/packet-ssl-utils.h")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 473.60134281438354], 1, 0, 0.0021114804997331383);
end;
weakness_21 @ [id:21, tool_specific_id:21, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_21 = (locations_wk_21, 1, 0, 0.003463630817441021);
locations_wk_21 = locations @ [tool_specific_id:21, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_1335( [line => 275, path => "wireshark-1.2.0/epan/dissectors/packet-ssl.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 288.71437306901373], 1, 0, 0.003463630817441021);
end;
weakness_22 @ [id:22, tool_specific_id:22, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_22 = (locations_wk_22, 1, 0, 0.004125022212036806);
locations_wk_22 = locations @ [tool_specific_id:22, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_1583( [line => 1799, path => "wireshark-1.2.0/epan/packet.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 242.42293704067873], 1, 0, 0.004125022212036806);
end;
29
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
weakness_23 @ [id:23, tool_specific_id:23, cweid:399, cwename:"Resource Management Errors (CWE399)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_23 = (locations_wk_23, 1, 0, 1.0);
locations_wk_23 = locations @ [tool_specific_id:23, cweid:399, cwename:"Resource Management Errors (CWE399)"];
observation location_id_1611( [line => 345, path => "wireshark-1.2.0/epan/sigcomp-udvm.c")
textoutput="";
observation grade = ([ severity => 3, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_24 @ [id:24, tool_specific_id:24, cweid:119, cwename:"Buffer Errors (CWE119)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_24 = (locations_wk_24, 1, 0, 1.0);
locations_wk_24 = locations @ [tool_specific_id:24, cweid:119, cwename:"Buffer Errors (CWE119)"];
observation location_id_1611( [line => 321, path => "wireshark-1.2.0/epan/sigcomp-udvm.c")
textoutput="";
observation grade = ([ severity => 1, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_25 @ [id:25, tool_specific_id:25, cweid:20, cwename:"Input Validation (CWE20)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_25 = (locations_wk_25, 1, 0, 0.001495003320843141);
locations_wk_25 = locations @ [tool_specific_id:25, cweid:20, cwename:"Input Validation (CWE20)"];
observation location_id_2012( [line => 89, path => "wireshark-1.2.0/plugins/docsis/packet-bpkmreq.c")
textoutput="";
observation grade = ([ severity => 1, tool_specific_rank => 668.8948352542972], 1, 0, 0.001495003320843141);
end;
weakness_26 @ [id:26, tool_specific_id:26, cweid:20, cwename:"Input Validation (CWE20)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_26 = (locations_wk_26, 1, 0, 0.0014959114047375394);
locations_wk_26 = locations @ [tool_specific_id:26, cweid:20, cwename:"Input Validation (CWE20)"];
observation location_id_2013( [line => 90, path => "wireshark-1.2.0/plugins/docsis/packet-bpkmrsp.c")
textoutput="";
observation grade = ([ severity => 1, tool_specific_rank => 668.4887867242726], 1, 0, 0.0014959114047375394);
end;
weakness_27 @ [id:27, tool_specific_id:27, cweid:20, cwename:"Input Validation (CWE20)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_27 = (locations_wk_27, 1, 0, 0.002153585613826869);
locations_wk_27 = locations @ [tool_specific_id:27, cweid:20, cwename:"Input Validation (CWE20)"];
observation location_id_2020( [line => 72, path => "wireshark-1.2.0/plugins/docsis/packet-dsaack.c")
textoutput="";
observation grade = ([ severity => 1, tool_specific_rank => 464.341883405798], 1, 0, 0.002153585613826869);
end;
weakness_28 @ [id:28, tool_specific_id:28, cweid:20, cwename:"Input Validation (CWE20)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_28 = (locations_wk_28, 1, 0, 0.00229165238295895);
locations_wk_28 = locations @ [tool_specific_id:28, cweid:20, cwename:"Input Validation (CWE20)"];
observation location_id_2022( [line => 72, path => "wireshark-1.2.0/plugins/docsis/packet-dsarsp.c")
textoutput="";
observation grade = ([ severity => 1, tool_specific_rank => 436.36635618741343], 1, 0, 0.00229165238295895);
end;
weakness_29 @ [id:29, tool_specific_id:29, cweid:20, cwename:"Input Validation (CWE20)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_29 = (locations_wk_29, 1, 0, 0.002184230355798278);
locations_wk_29 = locations @ [tool_specific_id:29, cweid:20, cwename:"Input Validation (CWE20)"];
observation location_id_2023( [line => 72, path => "wireshark-1.2.0/plugins/docsis/packet-dscack.c")
textoutput="";
observation grade = ([ severity => 1, tool_specific_rank => 457.82716889058463], 1, 0, 0.002184230355798278);
end;
weakness_30 @ [id:30, tool_specific_id:30, cweid:20, cwename:"Input Validation (CWE20)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_30 = (locations_wk_30, 1, 0, 0.0023006295251237975);
locations_wk_30 = locations @ [tool_specific_id:30, cweid:20, cwename:"Input Validation (CWE20)"];
observation location_id_2025( [line => 73, path => "wireshark-1.2.0/plugins/docsis/packet-dscrsp.c")
textoutput="";
observation grade = ([ severity => 1, tool_specific_rank => 434.6636383996635], 1, 0, 0.0023006295251237975);
end;
weakness_31 @ [id:31, tool_specific_id:31, cweid:20, cwename:"Input Validation (CWE20)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_31 = (locations_wk_31, 1, 0, 0.001897888480826156);
locations_wk_31 = locations @ [tool_specific_id:31, cweid:20, cwename:"Input Validation (CWE20)"];
observation location_id_2032( [line => 72, path => "wireshark-1.2.0/plugins/docsis/packet-regack.c")
textoutput="";
observation grade = ([ severity => 1, tool_specific_rank => 526.9013485790784], 1, 0, 0.001897888480826156);
end;
weakness_32 @ [id:32, tool_specific_id:32, cweid:20, cwename:"Input Validation (CWE20)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_32 = (locations_wk_32, 1, 0, 0.002216818195096963);
locations_wk_32 = locations @ [tool_specific_id:32, cweid:20, cwename:"Input Validation (CWE20)"];
observation location_id_2035( [line => 73, path => "wireshark-1.2.0/plugins/docsis/packet-regrsp.c")
textoutput="";
observation grade = ([ severity => 1, tool_specific_rank => 451.096983149879], 1, 0, 0.002216818195096963);
end;
weakness_33 @ [id:33, tool_specific_id:33, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
30
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_33 = (locations_wk_33, 1, 0, 0.0028814675905206645);
locations_wk_33 = locations @ [tool_specific_id:33, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_2097( [line => 433, path => "wireshark-1.2.0/plugins/opcua/opcua_complextypeparser.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 347.04537482557834], 1, 0, 0.0028814675905206645);
end;
weakness_34 @ [id:34, tool_specific_id:34, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_34 = (locations_wk_34, 1, 0, 0.0028288900371324934);
locations_wk_34 = locations @ [tool_specific_id:34, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_2107( [line => 616, path => "wireshark-1.2.0/plugins/opcua/opcua_serviceparser.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 353.49553601371184], 1, 0, 0.0028288900371324934);
end;
weakness_35 @ [id:35, tool_specific_id:35, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_35 = (locations_wk_35, 1, 0, 0.003058220966230374);
locations_wk_35 = locations @ [tool_specific_id:35, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_2110( [line => 340, path => "wireshark-1.2.0/plugins/opcua/opcua_simpletypes.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 326.9874907805045], 1, 0, 0.003058220966230374);
end;
weakness_36 @ [id:36, tool_specific_id:36, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_36 = (locations_wk_36, 1, 0, 0.0018096494904338023);
locations_wk_36 = locations @ [tool_specific_id:36, cweid:999, cwename:"Insufficient Information (NVD-CWE-noinfo)"];
observation location_id_2112( [line => 132, path => "wireshark-1.2.0/plugins/opcua/opcua_transport_layer.c")
observation location_id_2112( [line => 169, path => "wireshark-1.2.0/plugins/opcua/opcua_transport_layer.c")
observation location_id_2112( [line => 181, path => "wireshark-1.2.0/plugins/opcua/opcua_transport_layer.c")
observation location_id_2112( [line => 195, path => "wireshark-1.2.0/plugins/opcua/opcua_transport_layer.c")
observation location_id_2112( [line => 226, path => "wireshark-1.2.0/plugins/opcua/opcua_transport_layer.c")
observation location_id_2112( [line => 250, path => "wireshark-1.2.0/plugins/opcua/opcua_transport_layer.c")
textoutput="";
observation grade = ([ severity => 5, tool_specific_rank => 552.593198454295], 1, 0, 0.0018096494904338023);
end;
weakness_37 @ [id:37, tool_specific_id:37, cweid:119, cwename:"Buffer Errors (CWE119)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_37 = (locations_wk_37, 1, 0, 1.0);
locations_wk_37 = locations @ [tool_specific_id:37, cweid:119, cwename:"Buffer Errors (CWE119)"];
observation location_id_2321( [line => 149, path => "wireshark-1.2.0/wiretap/daintree-sna.c")
observation location_id_2321( [line => 205, path => "wireshark-1.2.0/wiretap/daintree-sna.c")
textoutput="";
observation grade = ([ severity => 1, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
weakness_38 @ [id:38, tool_specific_id:38, cweid:189, cwename:"Numeric Errors (CWE189)"]
where
dimension id, tool_specific_id, cweid, cwename;
observation sequence weakness_38 = (locations_wk_38, 1, 0, 1.0);
locations_wk_38 = locations @ [tool_specific_id:38, cweid:189, cwename:"Numeric Errors (CWE189)"];
observation location_id_2327( [line => 228, path => "wireshark-1.2.0/wiretap/erf.c")
textoutput="";
observation grade = ([ severity => 2, tool_specific_rank => 0.0], 1, 0, 1.0);
end;
}
References
[AS01]
A. F. Abdelnour and I. W. Selesnick. Nearly symmetric orthogonal wavelet bases. In Proc.
IEEE Int. Conf. Acoust., Speech, Signal Processing (ICASSP), May 2001.
[BOA+ 07] M. Bailey, J. Oberheide, J. Andersen, Z. M. Mao, F. Jahanian, and J. Nazario. Automated
classification and analysis of Internet malware. Technical report, University of Michigan, April
2007. http://www.eecs.umich.edu/techreports/cse/2007/CSE-TR-530-07.pdf.
+
[BOB 10] Hamad Binsalleeh, Thomas Ormerod, Amine Boukhtouta, Prosenjit Sinha, Amr M. Youssef,
Mourad Debbabi, and Lingyu Wang. On the analysis of the zeus botnet crimeware toolkit.
In Eighth Annual Conference on Privacy, Security and Trust, PST 2010, August 17-19, 2010,
Ottawa, Ontario, Canada, pages 31–38. IEEE, 2010.
[BSSV10] Mehran Bozorgi, Lawrence K. Saul, Stefan Savage, and Geoffrey M. Voelker. Beyond heuristics: Learning to classify vulnerabilities and predict exploits. In Proceedings of the 16th ACM
SIGKDD international conference on Knowledge Discovery and Data Mining, KDD’10, pages
105–114, New York, NY, USA, 2010. ACM.
+
[ESI 09] Masashi Eto, Kotaro Sonoda, Daisuke Inoue, Katsunari Yoshioka, and Koji Nakao. A proposal
of malware distinction method based on scan patterns using spectrum analysis. In Proceedings
31
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
of the 16th International Conference on Neural Information Processing: Part II, ICONIP’09,
pages 565–572, Berlin, Heidelberg, 2009. Springer-Verlag.
[Han10] Bin Han. Towards a multi-tier runtime system for GIPSY. Master’s thesis, Department of
Computer Science and Software Engineering, Concordia University, Montreal, Canada, 2010.
[HJ07]
K. Hwang and D. Jung. Anti-malware expert system. In H. Martin, editor, Proceedings of
the 17th Virus Bulletin International Conference, pages 9–17, Vienna, Austria: The Pentagon,
Abingdon, OX143YP, England, September 2007.
[HLYD09] Aiman Hanna, Hai Zhou Ling, Xiaochun Yang, and Mourad Debbabi. A synergy between
static and dynamic analysis for the detection of software security vulnerabilities. In Robert
Meersman, Tharam S. Dillon, and Pilar Herrero, editors, OTM Conferences (2), volume 5871
of Lecture Notes in Computer Science, pages 815–832. Springer, 2009.
[HRSS07] N. Hnatiw, T. Robinson, C. Sheehan, and N. Suan. Pimp my PE: Parsing malicious and
malformed executables. In H. Martin, editor, Proceedings of the 17th Virus Bulletin International Conference, pages 9–17, Vienna, Austria: The Pentagon, Abingdon, OX143YP, England,
September 2007.
+
[IYE 09] Daisuke Inoue, Katsunari Yoshioka, Masashi Eto, Masaya Yamagata, Eisuke Nishino, Jun’ichi
Takeuchi, Kazuya Ohkouchi, and Koji Nakao. An incident analysis system NICTER and its
analysis engines based on data mining techniques. In Proceedings of the 15th International
Conference on Advances in Neuro-Information Processing – Volume Part I, ICONIP’08, pages
579–586, Berlin, Heidelberg, 2009. Springer-Verlag.
[Ji11]
Yi Ji. Scalability evaluation of the GIPSY runtime system. Master’s thesis, Department of
Computer Science and Software Engineering, Concordia University, Montreal, Canada, March
2011.
[JMP13] Yi Ji, Serguei A. Mokhov, and Joey Paquet. Unifying and refactoring DMF to support concurrent Jini and JMS DMS in GIPSY. In Bipin C. Desai, Sudhir P. Mudur, and Emil I.
Vassev, editors, Proceedings of the Fifth International C* Conference on Computer Science
and Software Engineering (C3S2E’12), pages 36–44, New York, NY, USA, June 2010–2013.
ACM. Online e-print http://arxiv.org/abs/1012.2860.
[KAYE04] Ted Kremenek, Ken Ashcraft, Junfeng Yang, and Dawson Engler. Correlation exploitation in
error ranking. In Foundations of Software Engineering (FSE), 2004.
[KBC05] Manesh Kokare, P. K. Biswas, and B. N. Chatterji. Texture image retrieval using new rotated complex wavelet filters. IEEE Transaction on Systems, Man, and Cybernetics-Part B:
Cybernetics, 6(35):1168–1178, 2005.
[KBC06] Manesh Kokare, P. K. Biswas, and B. N. Chatterji. Rotation-invariant texture image retrieval
using rotated complex wavelet filters. IEEE Transaction on Systems, Man, and CyberneticsPart B: Cybernetics, 6(36):1273–1282, 2006.
[KE03] Ted Kremenek and Dawson Engler. Z-ranking: Using statistical analysis to counter the impact
of static analysis approximations. In SAS 2003, 2003.
+
[KTB 06] Ted Kremenek, Paul Twohey, Godmar Back, Andrew Ng, and Dawson Engler. From uncertainty to belief: Inferring the specification within. In Proceedings of the 7th Symposium on
Operating System Design and Implementation, 2006.
[KZL10] Ying Kong, Yuqing Zhang, and Qixu Liu. Eliminating human specification in static analysis.
In Proceedings of the 13th international conference on Recent advances in intrusion detection,
RAID’10, pages 494–495, Berlin, Heidelberg, 2010. Springer-Verlag.
+
[LjXP 09] Ru Li, Ou jie Xi, Bin Pang, Jiao Shen, and Chun-Lei Ren. Network application identification
based on wavelet transform and k-means algorithm. In Proceedings of the IEEE International
Conference on Intelligent Computing and Intelligent Systems (ICIS2009), volume 1, pages 38–
41, November 2009.
[LKW08] Kriangkrai Limthong, Fukuda Kensuke, and Pirawat Watanapongse. Wavelet-based unwanted
traffic time series analysis. In 2008 International Conference on Computer and Electrical
Engineering, pages 445–449. IEEE Computer Society, 2008.
32
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
[Mat12a] MathWorks. MATLAB. [online], 2000–2012. http://www.mathworks.com/products/
matlab/.
[Mat12b] MathWorks. MATLAB Coder. [online], 2012. http://www.mathworks.com/help/toolbox/
coder/coder_product_page.html, last viewed June 2012.
[Mat12c] MathWorks. MATLAB Coder: codegen – generate C/C++ code from MATLAB code.
[online], 2012. http://www.mathworks.com/help/toolbox/coder/ref/codegen.html, last
viewed June 2012.
[MD08] Serguei A. Mokhov and Mourad Debbabi. File type analysis using signal processing techniques
and machine learning vs. file unix utility for forensic analysis. In Oliver Goebel, Sandra
Frings, Detlef Guenther, Jens Nedon, and Dirk Schadt, editors, Proceedings of the IT Incident
Management and IT Forensics (IMF’08), LNI140, pages 73–85. GI, September 2008.
[MLB07] Serguei A. Mokhov, Marc-André Laverdière, and Djamel Benredjem. Taxonomy of linux kernel
vulnerability solutions. In Innovative Techniques in Instruction Technology, E-learning, Eassessment, and Education, pages 485–493, University of Bridgeport, U.S.A., 2007. Proceedings
of CISSE/SCSS’07.
[Mok08a] Serguei A. Mokhov. Encoding forensic multimedia evidence from MARF applications as Forensic Lucid expressions. In Tarek Sobh, Khaled Elleithy, and Ausif Mahmood, editors, Novel
Algorithms and Techniques in Telecommunications and Networking, proceedings of CISSE’08,
pages 413–416, University of Bridgeport, CT, USA, December 2008. Springer. Printed in January 2010.
[Mok08b] Serguei A. Mokhov. Study of best algorithm combinations for speech processing tasks in machine learning using median vs. mean clusters in MARF. In Bipin C. Desai, editor, Proceedings
of C3S2E’08, pages 29–43, Montreal, Quebec, Canada, May 2008. ACM.
[Mok10a] Serguei A. Mokhov. Complete complimentary results report of the MARF’s NLP approach to
the DEFT 2010 competition. [online], June 2010. http://arxiv.org/abs/1006.3787.
[Mok10b] Serguei A. Mokhov. Evolution of MARF and its NLP framework. In Proceedings of C3S2E’10,
pages 118–122. ACM, May 2010.
[Mok10c] Serguei A. Mokhov. L’approche MARF à DEFT 2010: A MARF approach to DEFT 2010. In
Proceedings of the 6th DEFT Workshop (DEFT’10), pages 35–49. LIMSI / ATALA, July 2010.
DEFT 2010 Workshop at TALN 2010; online at http://deft.limsi.fr/actes/2010/pdf/2_
clac.pdf.
[Mok10d] Serguei A. Mokhov. The use of machine learning with signal- and NLP processing of source code
to fingerprint, detect, and classify vulnerabilities and weaknesses with MARFCAT. [online],
October 2010. Online at http://arxiv.org/abs/1010.2511.
[Mok11] Serguei A. Mokhov. The use of machine learning with signal- and NLP processing of
source code to fingerprint, detect, and classify vulnerabilities and weaknesses with MARFCAT. Technical Report NIST SP 500-283, NIST, October 2011. Report: http://www.
nist.gov/manuscript-publication-search.cfm?pub_id=909407, online e-print at http:
//arxiv.org/abs/1010.2511.
[Mok13] Serguei A. Mokhov. MARFCAT – MARF-based Code Analysis Tool. Published electronically
within the MARF project, http://sourceforge.net/projects/marf/files/Applications/
MARFCAT/, 2010–2013. Last viewed April 2012.
[Mot09] Motorola. Efficient polyphase FIR resampler for numpy: Native C/C++ implementation of
the function upfirdn(). [online], 2009. http://code.google.com/p/upfirdn/source/browse/
upfirdn.
[MPD08] Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. Formally specifying operational semantics and language constructs of Forensic Lucid. In Oliver Göbel, Sandra Frings, Detlef
Günther, Jens Nedon, and Dirk Schadt, editors, Proceedings of the IT Incident Management and IT Forensics (IMF’08), LNI140, pages 197–216. GI, September 2008. Online at
http://subs.emis.de/LNI/Proceedings/Proceedings140/gi-proc-140-014.pdf.
[MPD10] Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. Towards automatic deduction and
33
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
event reconstruction using Forensic Lucid and probabilities to encode the IDS evidence. In
S. Jha, R. Sommer, and C. Kreibich, editors, Proceedings of RAID’10, LNCS 6307, pages
508–509. Springer, September 2010.
[MS02] Christopher D. Manning and Hinrich Schutze. Foundations of Statistical Natural Language
Processing. MIT Press, 2002.
[MSS09] Serguei A. Mokhov, Miao Song, and Ching Y. Suen. Writer identification using inexpensive
signal processing techniques. In Tarek Sobh and Khaled Elleithy, editors, Innovations in Computing Sciences and Software Engineering; Proceedings of CISSE’09, pages 437–441. Springer,
December 2009. ISBN: 978-90-481-9111-6, online at: http://arxiv.org/abs/0912.5502.
[NIS13a] NIST. National Vulnerability Database. [online], 2005–2013. http://nvd.nist.gov/.
[NIS13b] NIST. National Vulnerability Database statistics. [online], 2005–2013. http://web.nvd.nist.
gov/view/vuln/statistics.
+
[NJG 10] Vinod P. Nair, Harshit Jain, Yashwant K. Golecha, Manoj Singh Gaur, and Vijay Laxmi.
MEDUSA: MEtamorphic malware dynamic analysis using signature from API. In Proceedings
of the 3rd International Conference on Security of Information and Networks, SIN’10, pages
263–269, New York, NY, USA, 2010. ACM.
[ODBN10] Vadim Okun, Aurelien Delaitre, Paul E. Black, and NIST SAMATE. Static Analysis Tool
Exposition (SATE) 2010. [online], 2010. See http://samate.nist.gov/SATE2010Workshop.
html.
[ODBN12] Vadim Okun, Aurelien Delaitre, Paul E. Black, and NIST SAMATE. Static Analysis Tool
Exposition (SATE) IV. [online], March 2012. See http://samate.nist.gov/SATE.html.
[Paq09] Joey Paquet. Distributed eductive execution of hybrid intensional programs. In Proceedings of
the 33rd Annual IEEE International Computer Software and Applications Conference (COMPSAC’09), pages 218–224, Seattle, Washington, USA, July 2009. IEEE Computer Society.
[RM08] A. Newaz M. E. Rafiq and Yida Mao. A novel approach for automatic adjudification of new
malware. In Nagib Callaos, William Lesso, C. Dale Zinn, Jorge Baralt, Jaouad Boukachour,
Christopher White, Thilidzi Marwala, and Fulufhelo V. Nelwamondo, editors, Proceedings of
the 12th World Multi-Conference on Systemics, Cybernetics and Informatics (WM-SCI’08),
volume V, pages 137–142, Orlando, Florida, USA, June 2008. IIIS.
[Sch07] Rob Schreiber. MATLAB. Scholarpedia, 2(6):2929, 2007. http://www.scholarpedia.org/
article/MATLAB.
[SCL+ 03] Ivan Selesnick, Shihua Cai, Keyong Li, Levent Sendur, and A. Farras Abdelnour. MATLAB
implementation of wavelet transforms. Technical report, Electrical Engineering, Polytechnic
University, Brooklyn, NY, 2003. Online at http://taco.poly.edu/WaveletSoftware/.
[SEZS01] M. G. Schultz, E. Eskin, E. Zadok, and S. J. Stolfo. Data mining methods for detection of
new malicious executables. In Proceedings of IEEE Symposium on Security and Privacy, pages
38–49, Oakland, 2001.
[Son10a] Dawn Song. BitBlaze: Security via binary analysis. [online], 2010. Online at http://bitblaze.
cs.berkeley.edu.
[Son10b] Dawn Song. WebBlaze: New techniques and tools for web security. [online], 2010. Online at
http://webblaze.cs.berkeley.edu.
[Sou13] Sourcefire. Snort: Open-source network intrusion prevention and detection system (IDS/IPS).
[online], 1999–2013. http://www.snort.org/.
[Sue07] M. Suenaga. Virus linguistics – searching for ethnic words. In H. Martin, editor, Proceedings of
the 17th Virus Bulletin International Conference, pages 9–17, Vienna, Austria: ThePentagon,
Abingdon, OX143YP, England, September 2007.
[SXCM04] A. H. Sung, J. Xu, P. Chavez, and S. Mukkamala. Static analyzer of vicious executables
(SAVE). In Proceedings of 20th Annual of Computer Security Applications Conference, pages
326–334, December 2004.
[The13] The MARF Research and Development Group. The Modular Audio Recognition Framework
and its Applications. [online], 2002–2013. http://marf.sf.net and http://arxiv.org/abs/
34
MARFCAT: A MARF Approach to SATE IV
[Tli09]
[Vas05]
[VM13]
Mokhov, Paquet, Debbabi, Sun
0905.1235, last viewed April 2012.
Syrine Tlili. Automatic detection of safety and security vulnerabilities in open source software.
PhD thesis, Concordia Institute for Information Systems Engineering, Concordia University,
Montreal, Canada, 2009. ISBN: 9780494634165.
Emil Iordanov Vassev. General architecture for demand migration in the GIPSY demand-driven
execution engine. Master’s thesis, Department of Computer Science and Software Engineering,
Concordia University, Montreal, Canada, June 2005. ISBN 0494102969.
Various contributors and MITRE. Common Weakness Enumeration (CWE) – a communitydeveloped dictionary of software weakness types. [online], 2006–2013. See http://cwe.mitre.
org.
35
Index
API
DEFT2010App, 4
FileItem, 10
ResultSet, 10
Warning, 9
WriterIdentApp, 4
C, 4, 5, 7, 9, 11, 22
C++, 4, 9, 22
Chrome
5.0.375.54, 4, 11, 14, 16, 18
5.0.375.70, 4
CVE
CVE-2009-2559, 25, 26
CVE-2009-2560, 25, 26
CVE-2009-2561, 25, 26
CVE-2009-2562, 12, 13, 21, 25, 26
CVE-2009-2563, 25, 26
CVE-2009-3241, 25, 26
CVE-2009-3242, 25, 26
CVE-2009-3243, 25, 26
CVE-2009-3549, 25, 26
CVE-2009-3550, 25, 26
CVE-2009-3551, 25, 26
CVE-2009-3829, 25, 26
CVE-2009-4376, 25, 26
CVE-2009-4377, 25, 26
CVE-2009-4378, 25, 26
CVE-2010-0304, 25, 26
CVE-2010-1455, 25, 26
CVE-2010-2283, 25, 26
CVE-2010-2284, 25, 26
CVE-2010-2285, 25, 26
CVE-2010-2286, 25, 26
CVE-2010-2287, 25, 26
CWE
CWE-119, 15–18
CWE-16, 18
CWE-189, 15, 17
CWE-20, 15–18
CWE-200, 16, 18
CWE-22, 16, 18
CWE-255, 16, 18
CWE-264, 16, 18
CWE-399, 17, 18
CWE-79, 16–18
36
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
CWE-94, 17, 18
NVD-CWE-noinfo, 15, 17, 18, 27, 28
NVD-CWE-Other, 15, 17, 18, 27, 28
Dovecot
Dovecot
Dovecot
Dovecot
1.2.0, 4
1.2.17, 4
1.2.x, 11
2.0.beta6.20100626, 5
Files
*_test.xml, 19, 20
*_train.xml, 19, 20
collect-files-meta-synthetic.pl, 6
collect-files-meta.pl, 6
packet-afs.c, 12, 13, 21
report-cweidnopreprepcharunigramadddelta-train-test-test-run-quick-tomcat-5-5-33-cwe-nlp.
xml, 20
report-cweidnoprepreprawfftcheb-wireshark-1.2.0-half-train-cwe.xml, 14
report-cweidnoprepreprawfftcos-train-test-test-run-quick-tomcat-5-5-33-cwe.
xml, 19
report-cweidnoprepreprawfftdiff-wireshark-1.2.0-half-train-cwe.xml, 14
report-nopreprepcharunigramadddelta-train-test-test-run-quick-tomcat-5-5-33-cve-nlp.
xml, 19
report-noprepreprawfftcheb-train-test-test-run-quick-tomcat-5-5-33-cve.xml, 19
report-noprepreprawfftcheb-train-test-test-run-quick-wireshark-1-2-18-cve.xml,
20
report-noprepreprawfftcos-train-test-test-run-quick-tomcat-5-5-33-cve.xml, 19,
20
report-noprepreprawfftcos-train-test-test-run-quick-wireshark-1-2-18-cve.xml,
20
report-noprepreprawfftdiff-train-test-test-run-quick-wireshark-1-2-18-cve.xml,
20
report-noprepreprawffteucl-train-test-test-run-quick-wireshark-1-2-18-cve.xml,
20
report-noprepreprawffthamming-train-test-test-run-quick-wireshark-1-2-18-cve.
xml, 20
report-noprepreprawfftmink-train-test-test-run-quick-wireshark-1-2-18-cve.xml,
20
Forensic Lucid, 10, 11, 21, 23, 24
Frameworks
MARF, 1, 3, 8, 9, 23, 24
GIPSY, 9, 23
Java, 4, 5, 7, 9, 11, 22
Jetty
6.1.16, 4, 5
6.1.26, 5
6.1.x, 11
37
MARFCAT: A MARF Approach to SATE IV
Jini, 9
JMS, 9
Libraries
MARF, 1, 3, 8, 9, 23, 24
MARF, 1, 3, 8, 9, 23, 24
Applications
MARFCAT, 1–3, 5, 6, 9–11, 23
MARFCAT, 1–3, 5, 6, 9–11, 23
Options
-char, 17, 18
-cheb, 15, 17, 25, 26, 28
-cos, 15, 17, 19, 20, 25, 26, 28
-cweid, 15–18, 27, 28
-diff, 15, 16, 25–28
-eucl, 15, 17, 25–28
-fft, 15–17, 25–28
-flucid, 25–28
-graph, 26
-hamming, 15–17, 25–28
-low, 25, 28
-mink, 15, 17, 20, 25–28
-nopreprep, 15–18, 25–28
-raw, 15–17, 21
-sdwt, 26, 27
-spectrogram, 26
-unigram, 17, 18
Pebble, 5, 8, 9
Perl, 6
PHP, 4, 5
Test cases
Chrome 5.0.375.54, 4, 11, 14, 16, 18
Chrome 5.0.375.70, 4
Dovecot 1.2.0, 4
Dovecot 1.2.17, 4
Dovecot 1.2.x, 11
Dovecot 2.0.beta6.20100626, 5
Jetty 6.1.16, 4, 5
Jetty 6.1.26, 5
Jetty 6.1.x, 11
Pebble 2.5-M2, 5, 8, 9
Tomcat 5.5.13, 4, 5, 8, 9, 14, 15, 18–20
Tomcat 5.5.29, 5, 19
Tomcat 5.5.33, 5, 8, 9, 11, 19
Wireshark 1.2.0, 4, 12–14, 17, 19–21, 24–28
38
Mokhov, Paquet, Debbabi, Sun
MARFCAT: A MARF Approach to SATE IV
Mokhov, Paquet, Debbabi, Sun
Wireshark 1.2.18, 4, 11–13, 19, 20
Wireshark 1.2.9, 4, 12, 13, 19
Wordpress 2.0, 4, 5
Wordpress 2.2.3, 5
Wordpress 2.x, 11
TODO, 20
Tomcat
5.5.13, 4, 5, 8, 9, 14, 15, 18–20
5.5.29, 5, 19
5.5.33, 5, 8, 9, 11, 19
Tools
codegen, 8
Wireshark
1.2.0, 4, 12–14, 17, 19–21, 24–28
1.2.18, 4, 11–13, 19, 20
1.2.9, 4, 12, 13, 19
Wordpress
2.0, 4, 5
2.2.3, 5
2.x, 11
39
| 6cs.PL
|
Notes on rate equations in nonlinear continuum mechanics∗
Daniel Aubram†
arXiv:1709.10048v2 [physics.class-ph] 29 Sep 2017
September 29, 2017
Abstract: The paper gives an introduction to rate equations in nonlinear continuum mechanics which
should obey specific transformation rules. Emphasis is placed on the geometrical nature of the operations
involved in order to clarify the different concepts. The paper is particularly concerned with common classes
of constitutive equations based on corotational stress rates and their proper implementation in time for
solving initial boundary value problems. Hypoelastic simple shear is considered as an example application
for the derived theory and algorithms.
Contents
1 Introduction
2
2 Continuum Mechanics
2.1 Motion of a Body . . . . . . . . . . . . . . .
2.2 Deformation Gradient and Strain . . . . . .
2.3 Stress and Balance of Momentum . . . . . .
2.4 Constitutive Theory and Frame Invariance .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2
2
4
6
8
3 Rates of Tensor Fields
10
3.1 Fundamentals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 Corotational Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4 Rate Constitutive Equations
15
4.1 Hypoelasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.2 Hypoelasto-Plasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3 Hypoplasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
5 Objective Time Integration
20
5.1 Fundamentals and Geometrical Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.2 Algorithm of Hughes and Winget . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.3 Algorithms Using a Corotated Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
6 Hypoelastic Simple Shear
28
6.1 Analytical Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
6.2 Numerical Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
7 Conclusions
A Differential Geometry
A.1 Manifolds . . . . . . . . .
A.2 Tensors and Tensor Fields
A.3 Pushforward and Pullback
A.4 Tensor Analysis . . . . . .
32
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
33
33
34
37
38
∗ This work is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0). To view a copy
of this license, visit http://creativecommons.org/licenses/by/4.0/ or send a letter to Creative Commons, PO Box 1866,
Mountain View, CA 94042, USA.
† Chair of Soil Mechanics and Geotechnical Engineering, Technische Universität Berlin (TU Berlin), Secr. TIB1-B7, GustavMeyer-Allee 25, D-13355 Berlin, Germany; E-mail: [email protected]; URL: http://goo.gl/PlqPqn
1
1
Introduction
Many problems in physics and engineering science can be formalized as a set of balance equations for the
quantity of interest subject to a number of initial and/or boundary conditions. Additional closure relations
are often required which connect the primary unknowns with the dependent variables and render the set
of equations mathematically well-posed. The most important closure relations in continuum mechanics
[23, 48, 49, 70, 96, 95] are employed to determine the state of stress from the state of strain and are referred
to as the constitutive equations. Rate constitutive equations describe the rate of change of stress as a
function of the strain rate and a set of state variables.
The choice of a reference system to formulate the problem under consideration is a matter of convenience
and, from a formal viewpoint, all reference systems are equivalent. There are in fact preferred systems in
nonlinear continuum mechanics, particularly the one being fixed in space (Eulerian or spatial description),
and the other using fixed coordinates assigned to the particles of the material body in a certain configuration
in space (Lagrangian or material description) [48, 94]. Lagrangian coordinate lines are convected during
the motion of the body, and referring to them leads to the convected description [49, 79]. The arbitrary
Lagrangian-Eulerian (ALE) formulation is an attempt to generalize the material and spatial viewpoints
and to combine their advantages [4, 5, 6, 7, 9, 10, 33, 97]. The equivalence of reference systems for all
these descriptions requires that each term of the governing equations represents an honest tensor field
which transforms according to the transformation between the reference systems —a property referred to as
objectivity or, more generally, covariance [23, 49, 72].
As an example, consider a bar in simple tension which undergoes a rigid rotation. Then in a fixed spatial
(i.e. Eulerian) reference system the stress field transforms objectively if its components transform with the
matrix of that rigid rotation. In a Lagrangian reference system, on the other hand, the stress components
remain unaffected by such rigid motion because it does not stretch material lines. For reasons of consistency
it is required that, if the stress transforms objectively under rigid motions, the constitutive equation should
transform accordingly. This claim is commonly referred to as material frame indifference [57, 61, 95] and
has been the focus of much controversy during the last decades [12, 69, 87].
Further complexity is introduced if time derivatives are involved, as in rate constitutive equations, because
both the regarded quantity and the reference system are generally time-dependent. This has led to the
definition of countless rates of second-order tensors; see [30, 50, 51, 52, 64] for early discussions. Today
the most prominent examples include the Zaremba-Jaumann rate [39, 104] and the Green-Naghdi rate [27].
However, all objective rates are particular manifestations of the Lie derivative [49, 69, 72].
This paper gives an introduction to basic notions of nonlinear continuum mechanics and rate constitutive
equations. It is particularly concerned with constitutive equations based on corotational stress rates and their
proper implementation in time for solving mechanical initial boundary value problems. Section 2 addresses
kinematics, stress and balance of momentum as well as fundamentals of constitutive theory. Various rates of
second-order tensor fields are reviewed in Section 3, and classes of constitutive equations that employ such
rates are summarized in Section 4. In Section 5 we discuss procedures to integrate rate equations over a finite
time interval. We also provide detailed derivations of two widely-used numerical integration algorithms that
retain the property of objectivity on a discrete level. Applications of theory and algorithms are presented in
Section 6 using the popular example of hypoelastic simple shear. The paper closes in Section 7 with some
concluding remarks. Since we make extensive use of geometrical concepts and notions which have not yet
become standard practice in continuum mechanics, they are briefly introduced in Appendix A.
2
2.1
Continuum Mechanics
Motion of a Body
The starting point of any study about objectivity and rate equations in continuum mechanics is the motion
of a material body in the ambient space. As a general convention, we use upper case Latin for coordinates,
vectors, and tensors of the reference configuration, and objects related to the Lagrangian formulation. Lower
case Latin relates to the current configuration, the ambient space, or to the Eulerian formulation.
2
Definition 2.1. The ambient space, S, is an m-dimensional Riemannian manifold with metric g, and the
reference configuration of the material body is the embedded submanifold B ⊂ S with metric G induced by
the spatial metric. We assume that both B and S have the same dimension. Point resp. locations in space
are denoted by x ∈ S, and X ∈ B are the places of the particles of the body in the reference configuration. For
reasons of notational brevity, we refer to B as the body and to X as a particle. Particles carry the properties
of the material under consideration.
♦
Definition 2.2. The configuration of B in S at time t ∈ [0, T ] ⊂ R is an embedding
ϕt : B → S
X 7→ x = ϕt (X) ,
def
and the set C = {ϕt | ϕt : B → S} is called the configuration space. The deformation of the body is the
diffeomorphism B → ϕt (B). The motion of B in S is a family of configurations dependent on time t ∈ I ⊂ R,
def
i.e. a curve c : I → C, t 7→ c(t) = ϕt , and with ϕt (·) = ϕ(·, t) at fixed t. We assume that this curve is
sufficiently smooth. ϕt (B) is referred to as the current configuration of the body at time t, and x = ϕt (X)
is the current location of the particle X.
♦
Definition 2.3. The differentiable atlas of S consists of charts (V, σ), where V(x) ⊂ S is a neighborhood
def
1
m
i
m
of
∂x ∈ S and σ(x)
i = {x ,∗. . . , x }x = {x }x ∈ R . The holonomic basis of the tangent space at x is
∂xi x ∈ Tx S, dx x ∈ Tx S is its dual in the cotangent space, and the metric coefficients on S are
∂
∂
∂
∂
def
,
=
g
(x),
(x)
gij (x) =
∂xi ∂xj x
∂xi
∂xj
at every x ∈ V, taken with respect to the local coordinates {xi }x . The torsion-free connection ∇ has
coefficients denoted by γi jk .
♦
Definition 2.4. The charts of neighborhoods U(X) ⊂ B are denoted by (U, β), with local coordinate
∂
functions β(X) = {X I }X ∈ Rm . Therefore, ∂X
∈ TX B is the holonomic basis X, and the dual
I
X
I
∗
basis is {dX }X ∈ TX B. The metric of the ambient space induces a metric on B, with metric coefficients
def
GIJ (X) =
∂
∂
∂X I , ∂X J X
at every X ∈ U ⊂ B.
♦
Definition 2.5. In accordance with Definition A.6, the localization of the motion his the map
σ ◦ ϕt ◦ β −1 β (ϕ−1 (V)∩U ) ,
t
def
i
I
i
−1
with ϕ−1
)(X I ) are the spatial coordinates associated
t (V)∩U assumed non-empty, and ϕt (X ) = (x ◦ϕt ◦β
with that localization.
♦
Definition 2.6. It is assumed that both B and S are oriented with the same orientation, and their volume
densities be dV and dv, respectively. The relative volume change is given by Proposition A.13, that is,
dv ◦ ϕ = J dV ,
where J(X, t) is the Jacobian of the motion ϕ.
♦
Definition 2.7. Let ϕt be a continuously differentiable, i.e. C 1 -motion of B in S, then
def
V t (X) =
i
∂ϕt
def ∂ϕt
(X) =
∂t
∂t
β(X)
∂
∂ def i
= Vt (X) i (x)
∂xi
∂x
def
is called the Lagrangian or material velocity field over ϕt at X, where x = ϕt (X), V t (X) = V (X, t) for t
being fixed, and, V t : B → T S. Provided that ϕt is also regular, the spatial or Eulerian velocity field of ϕt
is defined through
def
v t = V t ◦ ϕ−1
: ϕt (B) → T S ,
t
so that v t is the “instantaneous” velocity at x ∈ ϕt (B) ⊂ S, and V (X, t) = v(ϕ(X, t), t). By abuse of
language, both V and v are occasionally called the material velocity in order to distinguish it from other,
non-material velocity fields.
♦
3
Definition 2.8. Depending on whether x = ϕ(X, t) ∈ S or X ∈ B serve as the independent variables
describing a physical field, one refers to qt : ϕt (B) → Tsr (S) as the Eulerian or spatial formulation and to
def
Qt = (qt ◦ ϕt ) : B → Tsr (S) as the Lagrangian or material formulation of that field, respectively.
♦
1
Proposition 2.1. For a regular C -motion, the Lie derivative of an arbitrary, possibly time-dependent,
spatial tensor field tt ∈ Tpq (S) along the spatial velocity v can be expressed by
Lv tt = ϕt ⇑
d
(ϕt ⇓ tt ) .
dt
d
(ψt,s ⇓ tt ), where ψt,s , with s, t ∈ [t0 , T ] ⊂ R, is the timeProof. By Definition A.34, Lv (tt ) = ψt,s ⇑ dt
dependent flow generated by the spatial velocity on S (Definition A.33). By Definition 2.7, the latter is
obtained from
ψt,s = ϕt ◦ ϕ−1
s :
S ⊃ ϕs (B) → ϕt (B) ⊂ S .
The assertion follows by applying the chain rule for pushforward and pullback (Proposition A.5), and noting
⇓= ϕs ⇑.
that (ϕs ⇓)−1 = ϕ−1
s
Proposition 2.2.
∂J
= J (tr d) ◦ ϕ .
∂t
Proof. ϕ ⇓ dv = J dV by Definition 2.6 in conjunction with Proposition A.13, so J dV is a time-dependent
volume form on B. Hence, from Propositions A.14 and 2.1,
dV
that is,
∂
∂t J
∂
∂J
= (J dV ) = ϕ ⇓ £v dv = ϕ ⇓ ((div v) dv) = ((div v) ◦ ϕ) J dV ,
∂t
∂t
= J (div v) ◦ ϕ. Since skew-symmetric tensors have zero trace, div v = tr l = tr d.
Definition 2.9. The material time derivative of an arbitrary time-dependent tensor field qt ∈
defined through
def ∂q
(x, t) + (v · ∇q) (x, t) ,
q̇(x, t) =
∂t x
where q̇t ∈ Trs (ϕ(B)), x = ϕ(X, t), and the term
2.2
∂
∂t q
is called the local time derivative of q.
Trs (ϕ(B))
is
♦
Deformation Gradient and Strain
Definition 2.10. The deformation gradient at X ∈ B is the tangent map over ϕ at X ∈ B, that is,
def
F (X) = Tϕ(X) : TX B → Tϕ(X) S (cf. Definition A.25); the time-dependency has been dropped for notational
brevity.
♦
Remark 2.1. The deformation gradient is a two-point tensor (cf. Definition A.10) and can locally be represented by
∂
F (X) = F iI (X) i ⊗ dX I ,
∂x
i
∂ϕ
∂
⋆
∗
in which F iI = ∂X
I , and ∂xi attached to ϕ(X) is being understood. Note that globally, F : B → ϕ T S ⊗T B
⋆
is a two-point tensor field, where ϕ T S denotes the induced bundle of T S over ϕ.
△
Proposition 2.3. Let t ∈ T02 (S), s ∈ T20 (S), T ∈ T02 (B), and S ∈ T20 (B), then (compositions with point
mappings are suppressed)
ϕ ⇓ s = F −1 · s · F −T ∈ T20 (B) ,
ϕ ⇓ t = F T · t · F ∈ T02 (B) ,
ϕ ⇑ T = F −T · T · F −1 ∈ T02 (S) ,
ϕ ⇑ S = F · S · F T ∈ T20 (S) .
Proof. By Definitions 2.10, A.27, and A.28.
Remark 2.2. The pullback and pushforward operators involve the tangent map Tϕ = F , and not ϕ itself.
This circumstance would justify the replacement of ϕ ⇓ by the symbol F ⇓, referred to as the F -pullback,
and ϕ ⇑ by F ⇑, called the F -pushforward.
△
4
Definition 2.11. The right Cauchy-Green tensor or deformation tensor is the tensor field defined through
def
C = (F T ◦ ϕ) · F ∈ T11 (B).
♦
def 1
2 (C
Definition 2.12. The Green-Lagrange strain or material strain is defined by E =
the second-order identity tensor on B, with components δ IJ .
− I), in which I is
♦
Remark 2.3. Note that both C and E are proper strain measures on the material body B, and that
E ♭ = 21 (C ♭ − G), where G = GIJ dX I ⊗ dX J is the metric on B.
△
Definition 2.13. The left Cauchy-Green tensor is a spatial or Eulerian strain measure defined through
def
b = F ◦ ϕ−1 · F T ∈ T11 (S). In a local chart of S,
b = GIJ gjk F iI F kJ
∂
⊗ dxj .
∂xi
The base points have been suppressed. The components of F T are given by Proposition A.4
def
eEA = 12 (i−c),
♦
def
Definition 2.14. The Euler-Almansi strain or spatial strain is defined by
in which c = b−1
is called the Finger tensor, and i is the second-order identity tensor on S with components δ ij .
♦
Proposition 2.4. Let g ∈ T02 (S) be the spatial metric, and ϕ a regular configuration, then (i) C ♭ = ϕ ⇓ g,
(ii) ϕ ⇑ G = c♭ , and (iii) ϕ ⇑ E ♭ = e♭EA .
Proof. The proofs of (i) and (ii) can be done in local coordinates with the aid of the formulas presented
in this section; cf. [3] for details. From this, (iii) becomes
ϕ ⇑ E ♭ = 21 (ϕ ⇑ C ♭ − ϕ ⇑ G) = 21 (g − c♭ ) = e♭EA .
Remark 2.4. It should be emphasized that associated tensors are different objects. For brevity, however,
the same name is used for all of them; e.g. all C, C ♭ , and C ♯ denote the right Cauchy-Green tensor. Note
that C ♭ plays a role of a material metric induced by the spatial metric g through the configuration ϕ. △
Definition 2.15. If ϕ : B → S is a regular configuration, then the deformation gradient has a unique
right polar decomposition F = R · U , and a unique left polar decomposition F = V · R. The two-point
tensor R(X) : TX B → Tx S, where x = ϕ(X), includes the rotatory part of the deformation and is proper
orthogonal, that is, R−1 = RT resp. det R = +1. The right stretch tensor U (X) : TX B → TX B and the
left stretch tensor V (x) : Tx S → Tx S are symmetric and positive definite for every X ∈ B and x ∈ S,
respectively.
♦
Remark 2.5. It will be usually clear from the context whether V denotes the left stretch tensor or the
material velocity, respectively, whether U denotes the right stretch tensor or the material displacement. △
Proposition 2.5. Both R-pushforward and R-pullback commute with index raising and index lowering,
e.g. R ⇑ (T ♭ ) = (R ⇑ T )♭ .
Proof. By Proposition 2.3 and noting that R is orthogonal, i.e. R−1 = RT .
Definition 2.16. The Lagrangian or material logarithmic strain is defined through the spectral decomposition
m
X
def
ε = ln U =
(ln λα ) Ψ (α) ⊗ Ψ (α) ∈ T11 (B) ,
α=1
where λα and Ψ α , with α ∈ {1, . . . , m}, are the eigenvalues and eigenvectors of the right stretch tensor,
respectively. The eigenvalues play the role of principal stretches. The Eulerian or spatial logarithmic strain
reads
m
X
def
e = ln V =
(ln λα ) ψ (α) ⊗ ψ (α) ∈ T11 (ϕ(B)) ,
α=1
where ψ α = R · Ψα are the eigenvectors of V . In the literature, the Eulerian logarithmic strain is often
referred to as the Hencky strain.
♦
P3
P3
Proposition 2.6. (i) F = α=1 λα ψ (α)⊗Ψ (α) , (ii) R = α=1 ψ (α)⊗Ψ (α) , and (iii) ε = RT ·e·R = R ⇓ e,
where R ⇓ is the R-pullback (cf. Remark 2.2).
5
Proof. (i) and (ii) follow directly from the previous definitions, and (iii) is a consequence of V = R·U ·RT .
Definition 2.17. Let ϕt : B → S be a regular C 1 -motion, then the Lagrangian or material rate of dedef ∂
∂
formation tensor D is defined by 2D(X, t) = ∂t
C(X, t) = 2 ∂t
E(X, t). The Eulerian or spatial rate of
def
deformation tensor field d is defined by d♭t = ϕt ⇑ (D♭t ), where dt : S → T S ⊗ T ∗S is a spatial tensor field
for fixed time t.
♦
Proposition 2.7. d♭ = Lv (e♭EA ) =
1
2
Lv g.
Proof. By Definition 2.17 together with Propositions 2.4 and 2.1.
Proposition 2.8. dij = 21 (∇i vj + ∇j vi ) resp. d♭ =
1
2
(∇v ♭ )T + ∇v ♭ .
Proof. By Proposition 2.7, Definition A.30, Propositions A.12 and A.7, and noting that the spatial metric
is time-independent.
Remark 2.6. It should be emphasized that associated tensors are different objects. For brevity, however,
the same name is used for all of them; e.g. all C, C ♭ , and C ♯ denote the right Cauchy-Green tensor.
△
Definition 2.18. The spatial velocity gradient is defined by
∂
def
def
−1
T
l = (∇v) =
◦ ϕ−1 = Ḟ · F −1 .
F ·F
∂t
def
Morevover, l = d + ω, where d = 21 (l + lT ) is the spatial rate of deformation tensor (Definition 2.17) and
def 1
2 (l
ω =
− lT ) = 21 ((∇v)T − ∇v)
is called the vorticity, with ω t : S → T S ⊗ T ∗S for fixed t.
♦
Definition 2.19. The infinitesimal strain is the linear approximation (linearization) to the Green-Lagrange
strain about a stress-free and undeformed state in the direction of an infinitesimal displacement u:
def
εlin = LINu E = 21 ((∇u)T + ∇u)
2.3
resp.
(εlin )ij = 12 (∇i uj + ∇j ui ) .
♦
Stress and Balance of Momentum
We are particularly concerned with isothermal mechanical problems that are governed by conservation of
mass and balance of linear and angular momentum. This section summarizes some basic relations for which
detailed derivations are available in the standard textbooks; e.g. [36, 48, 49, 96]. Notations and definitions of
the previous section are used throughout. In addition, let the material body be in its reference configuration
at time t = 0 such that
ϕ0 (B) = B
and
J(X, 0) = 1 .
Moreover, we assume that subsets U ⊂ B of the material body and subsets ϕt (U) ⊂ ϕt (B) ⊂ S embedded in
the ambient space have at least piecewise C 1 -continuous boundaries ∂U and ∂(ϕt (U)) = ϕt (∂U), respectively.
The outward normals to these boundaries are denoted by N ∗ ∈ Γ (T ∗B) and n∗ ∈ Γ (T ∗S), respectively.
Definition 2.20. A Cauchy traction vector field is a generally time-dependent vector field t on the boundary
∂(ϕt (B)) representing the force per unit area acting on an oriented surface
R element with outward normal
n∗ . If the ambient space is the linear Euclidian space, i.e. S = Rm , then ∂(ϕt (U )) t da represents the total
surface force acting on the body. The Cauchy traction vector at time t and point x ∈ ∂(ϕt (B)) is written
t(x, t, n∗ (x)) = tt (x, n∗ (x))
∈ Tx S .
♦
Theorem 2.1 (Cauchy’s Stress Theorem). Let the Cauchy traction vector
field t be a continuous function of its arguments, then there exists a unique time-dependent spatial 20 -tensor field σ t ∈ T20 (S) such
that
t = σ · n∗ ,
resp.
ti (x, t, n∗ (x)) = σ ij (x, t) nj (x) in spatial coordinates xi ,
that is, t depends linearly on n∗ .
6
j
∂
∂
i
Definition 2.21. The tensor σ(x, t) = σ ij (x, t) ∂x
i ⊗ ∂xj , as well as its associates with components σ j , σi
and σij , respectively, are referred to as the Cauchy stress.
♦
Definition 2.22. Let ρt : ϕt (B) → R be the spatial mass density of the body, bt ∈ Γ (T S) the external force
per unit mass, and σ t ∈ T20 (S) the Cauchy stress, with ρt (x) = ρ(x,Rt), bt (x) = b(x, t), and σ t (x) = σ(x, t)
d
at fixed t. Moreover, let ρ satisfy conservation of mass such that dt
ϕt (U ) ρ dv = 0, then ρ, b, and σ satisfy
spatial balance of linear momentum if
ρv̇ = ρb + div σ ,
where v̇ t ∈ Γ (T S) is the spatial acceleration field and a superposed dot denotes the material time derivative
(Definition 2.9).
♦
Definition 2.23. The first Piola-Kirchhoff stress is the tensor field P t : B → T S⊗T ∗ B obtained by applying
the Piola transformation (Definition A.38) to the second leg of the Cauchy stress, that is,
def
P t (X) = Jt (X) (σ t · F −T
t ) ◦ ϕt (X)
for every X ∈ B, and σ = σ ♯ being understood.
♦
Remark 2.7. In Definition 2.23, the placement of parentheses and the composition with the point map
are important: as σ(x, t) · F −T (x, t) has its values at (x, t), one has to switch the point arguments. In
material coordinates {X I }, spatial coordinates {xi }, and by omitting the point maps and arguments, one
has P iI = Jσ ij (F −1 )j I . Similar to the deformation gradient, P t (X) is a two-point tensor at every X ∈ B,
having the one “material” leg at X, and a “spatial” leg at x = ϕ(X, t) ∈ S.
△
Proposition 2.9. Since t = σ · n∗ is the force per unit of deformed area in the current configuration of the
body, T = P · N∗ resp. T i = P iI N I is the same force measured per unit reference area (or undeformed area),
and
(t da) ◦ ϕ = T dA
Proof. By Definition 2.23 and Proposition A.16.
def
Proposition 2.10. Let ρref (X) = ρ(ϕ(X, 0), 0) be the reference mass density at time t = 0, V (X, t) be the
def
material velocity, and B(X, t) = b(ϕ(X, t), t), then spatial balance of linear momentum (Definition 2.22)
has the equivalent Lagrangian resp. material form
ρref
∂V
= ρref B + DIV P .
∂t
Proof. Conservation of mass requires ρ(ϕ(X, t), t) J(X, t) = ρref (X) for all X ∈ B by Theorem A.1 and
Proposition A.13. Moreover, DIV P = J(div σ ◦ ϕ) by the Piola identity (Theorem A.3).
Definition 2.24. The second Piola-Kirchhoff stress S t ∈ T20 (B), with S t (X) = S(X, t) holding t fixed, is
the tensor field obtained by pullback of the first leg of P , that is,
def
−1
−T
S t = F −1
t · P t = Jt F t · (σ t · F t ) ◦ ϕt .
In components, S IJ = J(F −1 )I i (F −1 )j J σ ij .
♦
def
Definition 2.25. The Kirchhoff stress is defined through τ = (J ◦ ϕ−1 ) σ.
♦
Proposition 2.11. S = ϕ ⇓ τ .
Proof. By the previous definitions and Proposition 2.3.
Definition 2.26. Let R be the rotation two-point tensor obtained from polar decomposition of the deformation gradient (cf. Definition 2.15), then the corotated Cauchy stress is defined through R-pullback
(cf. Remark 2.2) of the Cauchy stress:
def
S = R⇓σ .
7
♦
Proposition 2.12. R-pullback commutes with index raising and index lowering, yielding
S♯ = R−1 · (σ ♯ · R−T ) ◦ ϕ = (RT · σ ♯ ) ◦ ϕ · R
and
S♭ = (RT · σ ♭ ) ◦ ϕ · R ,
where σ ♯ is the associated Cauchy stress with all indices raised and σ ♭ = σij dxi ⊗ dxj is the associated
Cauchy stress with all indices lowered (cf. Definition A.16).
Proof. By Proposition 2.3 again, and noting that R is proper orthogonal, i.e. R−1 = RT .
Theorem 2.1 (Symmetrie of Cauchy Stress). Let conservation of mass and balance of linear momentum be satisfied, then balance of angular momentum is satisfied if and only if
σ = σT
resp.
σ ij = σ ji ,
that is, if the Cauchy stress is symmetric. Symmetrie of Cauchy stress is equivalent to symmetry of the
second Piola-Kirchhoff stress, i.e. S = S T .
2.4
Constitutive Theory and Frame Invariance
For isothermal mechanical problems governed by balance of linear momentum (Definition 2.22) alone, the
motion ϕ : B × [0, T ] → S is generally treated as the primary unknown. The reference mass density, ρref ,
and the external force per unit mass, b, are usually given. The Jacobian J is known by the knowledge of ϕ,
hence the current density ρ can be determined from ρ = J −1 ρref . The acceleration v̇ can likewise be derived
from ϕ; equivalently, the m components of v̇ can be determined from the set of m equations of balance
of linear momentum. Therefore, in three dimensions one is left with six unknowns: the independent stress
components of σ = σ T . To close the set of model equations, these stress components are usually determined
from suitable constitutive equations.
Sets of axioms based on rational thermomechanical principles are routinely postulated to constrain and
simplify the constitutive equations. These will not be repeated here. Instead we refer to [95] and the key
papers and lecture notes [19, 18, 17, 27, 57, 58, 59, 60] particularly concerned with constitutive theory.
Additional citations are given in the text.
Definition 2.27. A relative motion or change of observer 1 is a time-dependent family of orientationpreserving diffeomorphisms θt : S → S ′ . A relative rigid motion or change of Euclidian observer requires
that θt = θtiso is a spatial isometry preserving the distance of every two points:
g ′ = θtiso ⇑ g ,
that is,
g(u, w) = g ′ (θtiso ⇑ u, θtiso ⇑ w) ,
where u, w ∈ Γ (T S), θtiso ⇑ u = (T θtiso · u) ◦ θtiso , and the tangent map
def
Tx S → Tx′ =θt (x) S ′
Tx θtiso = Qt (x) :
is proper orthogonal at every x ∈ S by Proposition A.6 such that Q−1
= QT
t
t and det Qt = +1. In this case
Qt is called a rotation, with Qt (x) = Q(x, t) at fixed t.
♦
For notational brevity the index “t” will be dropped in what follows. We also refrain from explicitly indicating
the dependence of a function on a mapping; e.g. for a scalar field f : B → R, a map ϕ : B → S, and x ∈ S,
we simply write f (x) instead of the correct (f ◦ ϕ−1 )(x).
Definition 2.28. A tensor field s ∈ Tpq (S) on the ambient space is called spatially covariant under the
action of a relative motion θ : S → S ′ if it transforms according to pushforward
s′ = θ ⇑ s
∈ Tpq (S ′ ) .
The field s is called objective if the transformation according to pushforward is restricted to relative rigid
motions θ = θiso .
♦
1 Both are equivalent provided that the different observers use charts having the same orientation relative to the orientation
of the spatial volume density dv [3].
8
Definition 2.29. A constitutive operator H is understood as a map between dual material tensor fields.
However, it can be equivalently formulated in terms of spatial fields by using the transformation rules outlined
in the previous sections. Conceptually, but without loss of generality, the constitutive response is denoted
by
S = H(C, A)
and
σ = h(F , g, α) ,
in the material description and spatial description, respectively. Besides S, C, σ, F , and g, which have
def
def
been defined in the sections above, the probably non-empty sets A = {A1 , . . . , Ak } resp. α = {α1 , . . . , αk }
consist of generally tensor-valued internal state variables (or history variables).
♦
Remark 2.8. From a formal viewpoint, a constitutive operator is a tensor bundle morphism between dual
tensor bundles over the same base space [69]. Bundle morphisms formalize mappings between tensor fields
and guarantee that the domain and co-domain of the constitutive operator are evaluated at the same base
point and the same time instant; see [67, 74] for more details on bundles and morphisms.
△
A minimal requirement in the formulation of constitutive equations is the following principle [69].
Principle 2.1 (Constitutive Frame Invariance). Any constitutive equation must conform to the principle of constitutive frame invariance (CFI) which requires that material fields, fulfilling the equation formulated
by an observer, will also fulfill the equation formulated by another Euclidian observer and vice versa. The
principle is expressed by the equivalence
S = H(C, A)
⇔
S ′ = H ′ (C ′ , A′ )
for any relative rigid motion resp. change of Euclidian observer θiso : S → S ′ .
Remark 2.9. CFI has been introduced by Romano and co-workers [67, 69, 72] in the context of a rigorous
geometric constitutive theory. It is intended as a substitute to the classical, but improperly stated principle
of material frame-indifference (MFI) [61, 95], which has been introduced by Noll [57] as the “principle of
objectivity of material properties”. MFI and the related concepts of indifference with respect to superposed
rigid body motions (IRBM), Euclidian frame indifference (EFI), and form-invariance (FI), cf. [87, 12], have
been the focus of much controversy over the years, until recently. In contrast to that, CFI employs basic
and properly settled geometric notions to account for the fact that distinct observers will formulate distinct
constitutive relations involving distinct material tensors.
△
Definition 2.30. The pushforward of a constitutive operator by a relative motion θ : S → S ′ is defined
consistent with Definition A.28 by the identity
(θ ⇑ H)(θ ⇑ C, θ ⇑ A) = θ ⇑ (H(C, A)) .
♦
Proposition 2.13. (See [69, prop. 9.1]) A constitutive equation conforms to the principle of CFI if and
only if the constitutive operator is frame invariant, that is,
H ′ = θiso ⇑ H ,
or equivalently,
h′ = θiso ⇑ h ,
for any change of Euclidian observer θiso : S → S ′ .
Proof. The assertion follows from Definitions 2.28 and 2.30 by a direct verification of the equivalence with
the statement of Principle 2.1.
The CFI Principle 2.1 requires that any constitutive equation must conform to it. In this work we are
particularly concerned with spatial rate constitutive equations, according to the following definition.
Definition 2.31. A spatial rate constitutive equation is understood as a map between a rate of strain and
a rate of stress in the spatial description. The spatial rate constitutive equations considered here take the
general form
def
◦ def
s = h(s, g, α, d) = m(s, g, α) : d ,
where s represents any objective rate of any spatial stress measure s satisfying (s)′ = θiso ⇑ s for any change
of Euclidian observer θiso : S → S ′ , and d is the spatial rate of deformation tensor.
♦
◦
◦
◦
Remark 2.10. The particular classes of constitutive equations that fall into the category formalized by
Definition 2.31 include hypoelasticity, hypoelasto-plasticity, and hypoplasticity. These will be discussed in
more detail in Sect. 4.
△
9
3
3.1
Rates of Tensor Fields
Fundamentals
In the following sections we inspect the transformation properties of the common tensor fields in spatial
rate constitutive equations under the action of any relative motion (resp. change of observer) and under the
action of a relative rigid motion (resp. change of Euclidian observer). In particular, a distinction is drawn
between spatially covariant rates, objective rates, and corotational rates of second-order tensors.
In accordance with Definition 2.27, and by dropping the index t in what follows, let θ : S → S ′ denote a
relative motion, and θ = θiso if the relative motion is an isometry, i.e. rigid. The tangent map of a relative
def
motion is generally time-dependent and denoted by F θ = T θ, and the proper orthogonal tangent map of a
def
relative rigid motion is the rotation two-point tensor field denoted by Q = T θiso . Here and in the following
we assume that both F θ and Q are continuously differentiable in time.
Proposition 3.1. The deformation gradient of a motion ϕ : B → S is spatially covariant.
Proof. By Definition 2.28, Proposition 2.3 and the chain rule,
F ′ = F θ · F = θ ⇑F .
Composition with the point mappings ϕ and θ have been suppressed.
def
Definition 3.1. The spin of a relative rigid motion is defined through Λ = Q̇T · Q.
♦
T
Proposition 3.2. Orthogonality of Q implies skew-symmetry of the spin, that is, Λ = −Λ .
Proof. By direct calculation,
˙
İ = QT · Q = Q̇T · Q + QT · Q̇ = Λ + ΛT = 0 .
Proposition 3.3. Let l = d + ω be the spatial velocity gradient, being composed of the spatial rate of
deformation d and the vorticity ω (cf. Definition 2.18). Then,
(i) both l and ω are neither spatially covariant, nor objective, and
(ii) d is objective, but not spatially covariant.
Proof. (i) Spatial covariance is a stronger version of objectivity under relative rigid motions, thus it suffices
to proof that the velocity gradient is not objective. First, note that F ′ = θiso ⇑ F = Q · F by Proposition 3.1.
Moreover, by Propositions 3.2, 2.3, and 2.5, and Definition 2.18, one has
l′ = (F˙ ′ ) · (F ′ )−1 = (Q̇ · F + Q · Ḟ ) · F −1 · QT = Q̇ · QT + Q · l · QT = lθiso + θiso ⇑ l ,
which is clearly non-objective, that is, it does not conform to Definition 2.28. Substitution of l = d + ω,
with d defined as the symmetric part of l, shows that lθiso = ω θiso and
ω ′ = Q̇ · QT + Q · ω · QT = ω θiso + θiso ⇑ ω ,
which proofs the first assertion.
(ii) A direct consequence of the proof of (i) is that d is indeed objective under relative rigid motions:
d′ = Q · d · QT = θiso ⇑ d .
In the case where θ : S → S ′ is an arbitrary relative motion, the tangent F θ is generally not orthogonal.
Definition 2.18 and Proposition 3.1 yield
−1
l′ = (F˙ ′ ) · (F ′ )−1 = F˙ θ · F −1
θ + F θ · l · F θ = lθ + θ ⇑ l .
Therefore, d′ = θ ⇑ d if and only if lθ ≡ ω θ , that is, if dθ ≡ 0.
Proposition 3.4. Cauchy stress σ is spatially covariant while its material time derivative σ̇ is not even
objective.
10
Proof. Transformation of the Cauchy traction vector field t using Cauchy’s stress theorem 2.1 and Proposition 2.3 yields
−T
t′ = θ ⇑ t = F θ · (σ · n∗ ) = (F θ · σ · F T
· n∗ ) = (θ ⇑ σ) · (θ ⇑ n∗ ) = σ ′ · (n∗ )′ .
θ ) · (F θ
For a relative rigid motion θ = θiso this becomes
t′ = θiso ⇑ t = Q · t = Q · (σ · n∗ ) = (Q · σ · QT ) · (Q · n∗ ) = (θiso ⇑ σ) · (θiso ⇑ n∗ ) = σ ′ · (n∗ )′ ,
by using the property Q · QT = QT · Q = I. Therefore, the Cauchy stress is spatially covariant. However,
its material time derivative is not even objective because
˙
(σ˙ ′ ) = Q · σ · QT = Q̇ · σ · QT + Q · σ̇ · QT + Q · σ · Q̇T 6= Q · σ̇ · QT .
A large amount of literature is concerned with the discussion and/or development of objective rates [30, 50,
51, 52, 62, 64, 90, 93, 92, 101, 103]. This would lead one to assume that there is a rate that is preferable
to others. However, the decisive conclusion for a rate, e.g. in a constitutive equation, could not be drawn
from its objectivity property alone, but has to consider the intended application of that rate [103]. In fact,
any possible objective rate of spatial second-order tensors is a particular manifestation of the Lie derivative
[49, 78, 77]. The Lie derivative (Definition A.34) is a geometric object that has an important property: if a
tensor is spatially covariant, then its Lie Derivative also is.
Theorem 3.1 (Spatial Covariance of Lie Derivative). Let ϕ : B → S be the motion of a material body
B with spatial velocity v, and let θ : S → S ′ be a relative motion such that ϕ′ = θ ◦ ϕ is the superposed
motion of B with spatial velocity v ′ . Moreover, let s ∈ Tpq (S) be a spatially covariant tensor field such that
s′ = θ ⇑ s, then
Lv′ s′ = θ ⇑ (Lv s) .
Proof. Let ϕi (X I ) = (xi ◦ ϕ)(X I ), θi (ϕj ) = (xi ◦ θ)(ϕj ), and (ϕ′ )i (X I ) = (xi ◦ ϕ′ )(X I ) be the spatial
coordinates xi on S arising from the localizations of ϕ, θ, and ϕ′ = θ ◦ ϕ, respectively. Then, by the chain
rule and Definition 2.7,
∂(xi ◦ θ ◦ ϕ)
∂(ϕ′ )i
◦ (ϕ′ )−1 =
◦ (ϕ′ )−1
∂t
∂t
i
∂θi ∂ϕj
∂θ i
∂θi
−1
i
−1
◦
θ
=
w
+
v ◦ θ−1 ,
◦ θ−1 +
◦
ϕ
=
j
∂t
∂ϕ
∂t
∂ϕj
(v ′ )i =
that is, v ′ = w + θ ⇑ v, where w, with components wi , represents the spatial velocity of θ. The rest of the
proof can be done as in [49, pp. 101–102], which is repeated here for completeness. By Proposition A.11,
Lv ′ s ′ =
∂s′
∂s′
+ £w+θ⇑v s′ =
+ £w s′ + θ ⇑ (£v (θ ⇓ s′ )) .
∂t
∂t
In accordance with Proposition 2.1, the flow associated with w is given by θt ◦θs−1 , for s, t ∈ R. Definition A.34
and Proposition A.5 then yield
d
∂s′
+ £w s′ + θ ⇑ (£v s) =
(θt ◦ θs−1 ) ⇓ s′t
+ θ ⇑ (£v s)
∂t
dt
t=s
d
d
=
θt ⇓ ◦(θs−1 ) ⇓ ◦θt ⇑ (θt ⇓ s′t )
+ θ ⇑ (£v s) =
θs ⇑ (θt ⇓ s′t )
+ θ ⇑ (£v s)
dt
dt
t=s
t=s
d
d
+ θ ⇑ (£v s) = θ ⇑
+ £v s = θ ⇑ (Lv s) .
θs ⇑ st
st
=
dt
dt t=s
t=s
Lv′ s′ =
There are two spatially covariant, and thus objective, stress rates that can be directly obtained from the Lie
derivative. Note that, with respect to spatial coordinates xi , the components of the Lie derivative of the
contravariant Kirchhoff stress τ ♯ ∈ T20 (S) are given by
(Lv τ )ij = τ̇ ij − τ kj (∇k v i ) − τ ik (∇k v j ) ,
where the general coordinate formula of Proposition A.12 has been applied. From this one obtains the
coordinate-invariant expression Lv (τ ♯ ) = τ̇ ♯ − l · τ ♯ − τ ♯ · lT . Moreover, by recalling that the Kirchhoff stress
is defined through τ = J σ, and that the spatial form of Proposition 2.2 is J˙ = Lv J = J div v, one has
J −1 Lv (τ ♯ ) = Lv (σ ♯ ) + σ ♯ div v = σ̇ ♯ − l · σ ♯ − σ ♯ · lT + σ ♯ tr d .
11
Definition 3.2. The rates defined through
def
τ Ol = Lv (τ ♯ )
◦
def
σ Tr = J −1 Lv (τ ♯ ) .
◦
and
are called the (upper) Oldroyd rate of Kirchhoff stress [62] and Truesdell rate of Cauchy stress [93, 92],
respectively.
♦
3.2
Corotational Rates
Both the Oldroyd rate and the Truesdell rate —as they are Lie derivatives— do not commute with index
raising and index lowering, because the Oldroyd and Truesdell rates of the metric tensor does not vanish.
For example, the components of the Oldroyd rate of the inverse metric g ♯ are
g Ol
◦
ij
= ġ ij − g kj (∇k v i ) − g ik (∇k v j ) = −(∇j v i + ∇i v j ) = −2 dij .
Hence, g Ol = Lv (g ♯ ) = −2d♯ , and Lv (g ♭ ) = 2d♭ by Proposition 2.7. Consequently the stress invariants
arising in constitutive equations of isotropic solids, and which are formed by the metric tensor, are not
stationary if the stress rate vanishes. This fact, however, conflicts with Prager’s requirement [64, 30, 52] and
constitutes a drawback of the Oldroyd and Truesdell rates in plasticity theory [14, 103] as well as in some
advanced applications [65, 89]. The so-called corotational rates circumvent this drawback.
◦
Definition 3.3. Let s be a second-order spatial tensor field continuously differentiable in time and let
Λ = −ΛT be a spin tensor, then
◦
def
s = ṡ − Λ · s + s · Λ
is called the corotational rate of s defined by the spin Λ.
♦
Definition 3.4. Let Λ(x, t) = −ΛT (x, t) be a given spin tensor for all x ∈ ϕ(B, t) and t ∈ [0, T ], with
ϕ(B, 0) = B. Consider the following evolution equation
∂R
= (Λ ◦ ϕ) · R ,
∂t
with
R|t=0 = I ,
where R(X, t) : TX B → Tϕ(X,t) S is a proper orthogonal two-point tensor for fixed X ∈ B and each t ∈ [0, T ],
such that RT · R = I B , R · RT = I S , and det R = +1. Solutions to the problem generate a one-parameter
group of rotations to which R belongs, thus Λ is called the generator of that group [37, 77].
♦
Remark 3.1. From the previous definition the term corotational can be justified as follows. In a rotating
Euclidian frame with spin Λ = Ṙ · RT the Cauchy stress is given by σ ′ = R ⇓ σ = RT · σ · R. Then,
◦
the corotational rate σ represents the rate of change of σ ′ observed in the fixed frame where σ is measured.
Clearly,
∂(R ⇓ σ)
◦
◦
.
σ = R · (σ˙ ′ ) · RT ,
or equivalently,
R⇓σ =
△
∂t
There are infinitely many objective rates and corotational rates. Not every corotational rate is objective,
and vice versa. Whether or not a corotational rate is objective depends on its defining spin tensor.
Definition 3.5. The Zaremba-Jaumann rate of Cauchy stress [39, 104] is obtained from Definition 3.3 by
def
setting Λ = ω, where ω = 21 (l−lT ) ∈ T11 (S) is the vorticity tensor according to Definition 2.18:
def
σ ZJ = σ̇ − ω · σ + σ · ω .
◦
♦
In contrast to the Oldroyd and Truesdell rates, which are non-corotational in terms of Definition 3.3, all
corotational rates do commute with index raising and index lowering, thus satisfy Prager’s requirement.
Proposition 3.5. Definition 3.3 identically applies for all associated tensor fields s ∈ T11 (S), s♯ = g ♯ · s ∈
T20 (S), and s♭ = g ♭ · s ∈ T02 (S) irrespective of index placement. That is, any corotational rate of the metric
tensor vanishes.
12
Proof. We proof this, without loss of generality, for the Zaremba-Jaumann rate. Keeping the property
∇(v ♭ ) = (∇v)♭ in mind, then the components of the Zaremba-Jaumann rate of the inverse metric g ♯ are
ij
◦
g ZJ = ġ ij − ω ik g kj − ω ik g jk = −ω ik g kj + g ik ωk j = −ω ij + ω ij = 0 .
Definition 3.6. Let F = R · U denote the right polar decomposition of the deformation gradient, with R
∂
being proper orthogonal. Similar to the velocity gradient given by the relation ∂t
F = (l ◦ ϕ) · F , let the
spatial rate of rotation Ω be defined through
∂R def
= (Ω ◦ ϕ) · R .
∂t
def
Choosing the spin Λ = Ω in Definition 3.3 then yields the Green-Naghdi rate of Cauchy stress [27]:
def
σ GN = σ̇ − Ω · σ + σ · Ω .
◦
♦
Proposition 3.6. Vorticity ω = 21 (l − lT ) associated with the Zaremba-Jaumann rate and spatial rate of
rotation Ω = Ṙ · RT associated with the Green-Naghdi rate are related by
ω = Ω + 12 R · (U̇ · U −1 − U −1 · U̇ ) · RT .
Proof. By time differentiation of F = R · U .
Remark 3.2. The tensor Ω is a kind of angular velocity field describing the rate of rotation of the material,
whereas ω describes the rate of rotation of the principal axes of the rate of deformation tensor d = l − ω [20].
In contrast to Ω, vorticity contains terms due to stretching. Therefore, the Green-Naghdi rate (Definition 3.6)
is identical to the material time derivative of the Cauchy stress in the absence of rigid body rotation, while
the Zaremba-Jaumann rate (Definition 3.5) is generally not. The Green-Naghdi rate requires knowledge of
total material motion resp. material deformation through R = Tϕ · U −1 , while the Zaremba-Jaumann rate,
by virtue of vorticity, is derivable from the instantaneous motion at current time; in fact l is the generator
∂
F = l · F . This makes the Zaremba-Jaumann rate more attractive to problems where a past
of F through ∂t
◦
◦
material motion is unavailable. By using Propostion 3.6, it can be shown that ω = Ω resp. σ ZJ = σ GN if
and only if the motion of the material body is a rigid rotation, a pure stretch, or if the current configuration
has been chosen as the reference configuration such that F = R = I, U = I, and Ḟ = Ṙ + I · U̇ ; see also
[95, pp. 54–55] and [20, 21, 16]. The last condition is used in Sect. 5 to compare different time integration
algorithms for large deformations based on the Zaremba-Jaumann and Green-Naghdi rates.
△
Proposition 3.7. Let θ : S → S ′ be a relative motion superposed to the motion ϕ : B → S of a material
body, then both the Zaremba-Jaumann rate and the Green-Naghdi rate of Cauchy stress
(i) transform objectively if θ = θiso is rigid, but they
(ii) are not spatially covariant.
Proof. (i) Recall that σ ′ = θiso ⇑ σ = Q · σ · QT ,
Q · σ̇ · QT = (σ˙ ′ ) − Q̇ · σ · QT − Q · σ · Q̇T = (σ˙ ′ ) − Q̇ · QT · σ ′ + σ ′ · Q̇ · QT ,
and Q · ω · QT = ω ′ − Q̇ · QT by Proposition 3.3(i). The pushforward of the Zaremba-Jaumann rate
(Definition 3.5) along the relative rigid motion θiso then becomes
Q · σ ZJ · QT = Q · (σ̇ − ω · σ + σ · ω) · QT = Q · σ̇ · QT − Q · ω · QT · σ ′ + σ ′ · Q · ω · QT
= (σ˙ ′ ) − Q̇ · QT · σ ′ + σ ′ · Q̇ · QT − (ω ′ − Q̇ · QT ) · σ ′ + σ ′ · (ω ′ − Q̇ · QT )
◦
◦
= (σ˙ ′ ) − ω ′ · σ ′ + σ ′ · ω ′ = (σ ZJ )′ ,
that is, σ ZJ is objective. Similarly, one has Ω ′ = Q̇ · QT + Q · Ω · QT , showing that the Green-Naghdi rate
◦
σ GN is objective, too [20, 40].
◦
(ii) Recall that if θ : S → S ′ is an arbitrary relative motion with generally non-orthogonal tangent F θ , then
l′ = lθ + θ ⇑ l by the proof of Proposition 3.3(ii), where lθ = Ḟ θ · F −1
θ . Moreover,
T
= ωθ + θ ⇑ ω .
ω′ = 12 lθ − lT
θ + θ ⇑ l − (θ ⇑ l)
13
Since σ ′ = θ ⇑ σ by Proposition 3.4, it is easy to show that
(σ˙ ′ ) = θ⋆ (σ̇) + lθ · σ ′ + σ ′ · lT
θ ,
for σ ≡ σ ♯ being understood. Now proceed as in the proof of (i), clearly,
θ ⇑ (σ ZJ ) = θ ⇑ (σ̇ − ω · σ + σ · ω) = θ ⇑ (σ̇) − (θ ⇑ ω) · σ ′ + σ ′ · (θ ⇑ ω)
= (σ˙ ′ ) − l · σ ′ − σ ′ · lT − (ω ′ − ω ) · σ ′ + σ ′ · (ω ′ − ω )
◦
θ
◦
θ
θ
ZJ ′
′
′
′
= (σ ) − lθ · σ + ω θ · σ − σ ·
lT
θ
θ
′
− σ · ωθ .
Then it follows immediately that θ ⇑ (σ ZJ ) = (σ ZJ )′ if and only if lθ ≡ ω θ , i.e. if θ is a rigid motion with
dθ ≡ 0. To proof (ii) for the Green-Naghdi rate, note that R′ = F θ · R, which leads to
◦
◦
Ω ′ = (R˙ ′ ) · (R′ )−1 = (Ḟ θ · R + F θ · Ṙ) · R−1 · F −1
θ = lθ + θ ⇑ Ω .
Ω ′ is not skew unless lθ is skew, that is, unless F θ is a pure rotation. The rest of the proof is similar to that
◦
for σ ZJ .
Remark 3.3. Proposition 3.7(ii) is remarkable because one would never have seen it in classical continuum
mechanics in linear Euclidian space. Marsden and Hughes [49, box 6.1] draw a proof from the transformation
property of the Lie derivative (Theorem 3.1).
△
Proposition 3.8. (See also [78, p. 222].) Let S = R ⇓ σ be the corotated Cauchy stress, and let V ⇓ denote
the pullback by the left stretch tensor V (cf. Remark 2.2), then the Green-Naghdi rate of Cauchy stress can
be obtained from
∂S
◦
= L(V ⇓v) σ .
σ GN = R ⇑
∂t
Proof. The first identity can be shown by a direct calculation
∂S
∂
∂
◦
T
R⇑
= R ⇑ (R ⇓ σ) = R ·
(R · σ · R) · RT = σ̇ − Ω · σ + σ · Ω = σ GN .
∂t
∂t
∂t
If R is obtained from the left polar decomposition F = V · R, then ϕ ⇓= F ⇓= R ⇓ ◦V ⇓ by the chain
rule for pullbacks (Proposition A.5). Moreover, R ⇓= ϕ ⇓ ◦V ⇑ and R ⇑= V ⇓ ◦ϕ ⇑, so finally, using
Proposition A.11,
∂
∂
R ⇑ (R ⇓ σ) = V ⇓ ϕ ⇑ (ϕ ⇓ (V ⇑ σ)) = V ⇓ (Lv (V ⇑ σ)) = L(V ⇓v) σ .
∂t
∂t
It can be summarized that the Zaremba-Jaumann rate and the Green-Naghdi rate are corotational, objective, and satisfy Prager’s requirement, but they are not spatially covariant. The Oldroyd rate and the
Truesdell rate meet the stronger condition of spatial covariance, but they include stretching parts, thus are
not corotational, and they do not commute with index raising and lowering applied to their argument.
Remark 3.4. Although the Green-Naghdi rate is related to a Lie derivative through Proposition 3.8, that
Lie derivative is not spatially covariant. The restriction arises from the flow generated by the “stretched”
spatial velocity field V ⇓ v employed.
△
Remark 3.5. The spatial rate of deformation or stretching d is a fundamental kinematic quantity. In
quoting [101], however, it should be noticed that “[...] by now the stretching has been known simply as a
symmetric part of the velocity gradient, [...] and it has not been known whether or not it is really a rate of
the change of a strain measure.” Assume that ϕ : B → S is a kind of motion of a three-dimensional body
B in S = R3 for which the principal axes of stretch are fixed, and let Λ, in a Cartesian coordinate system,
be the diagonal matrix containing the principal stretches λ1 , λ2 , λ3 . Then, according to [20], the rate of
deformation is the diagonal matrix
d = Λ̇Λ−1 ,
with components d11 = λ̇1 /λ1 = (ln˙V ) = ė11 , d22 = ė22 , and d33 = ė33 , and e = ln V being the spatial
11
logarithmic strain or Hencky strain (Definition 2.16). Certain corotational and objective rates of e can equal
the rate of deformation for certain particular left stretch tensors V [31, 34]. In their seminal paper [101]
14
finally prove that for all V there is a unique corotational rate, called the logarithmic rate, of the Hencky
strain e which is identical to the rate of deformation:
◦
def
d = (ln V )log = elog = ė − Ω log · e + e · Ω log .
◦
The calculation of the so-called logarithmic spin Ω log , however, is complicated for general cases [101]. Andef 1
2 (b
other important identity is that the upper Oldroyd rate of the so-called Finger strain a =
b = F · F T is the left Cauchy-Green tensor, equals spatial rate of deformation [14, 32]:
− i), where
aOl = Lv (a♯ ) = ȧ♯ − l · a♯ − a♯ · lT = d♯ .
◦
Therefore, d is indeed an honest strain rate.
4
△
Rate Constitutive Equations
There are basically two main groups of rate-independent constitutive equations (or material models) that are
used in computational solid mechanical applications at large deformation. The elements of the first group are
typically based on thermodynamical principles postulated at the outset, and they are commonly addressed
with the prefix “hyper”: hyperelasticity , hyperelasto-plasticity, and hyperplasticity. The constitutive equations belonging to the second group usually ignore balance of energy and the axiom of entropy production.
Many of them are are based on an ad hoc extension of existing small-strain constitutive equations to the finite
deformation range. Elements of the second group are called Eulerian or spatial rate constitutive equations
and are commonly addressed with the prefix “hypo”: hypoelasticity, hypoelasto-plasticity, and hypoplasticity.
The following section gives a general introduction to spatial rate constitutive equations belonging to the
second group. In spite of their shortcomings discussed, for example, in [80, 77], we point out that these
material models remain widely used in computational continuum mechanics. This is because the same
integration algorithms can be employed at both infinitesimal and finite deformations, as will be shown
in Sect. 5. Many, if not the majority of finite element codes in solid mechanics employ rate constitutive
equations for problems involving small or large inelastic deformations.
In this section we address only rate constitutive equations accounting for finite deformations. Readers who
are not familiar with elasticity and classical elasto-plasticity at small strains should consult introductory
texts on plasticity theory [15, 77]. We remark, however, that the general formulas presented here carry over
to the case of infinitesimal deformations if the objective stress rate and rate of deformation are replaced with
the common material time derivatives of stress and infinitesimal strain, respectively:
◦
σ ⇒ σ̇
and
d ⇒ ε̇lin .
The term material model or just model will be used as a synonym for constitutive equation. Without
indicating it further, stress measures are taken with all indices raised, and strain measures with all indices
def
def
lowered, e.g. σ = σ ♯ and d = d♭ . The dependence of a function on a point map, for example, on the
motion ϕ, will be usually clear from the context. Moreover, we do not indicate time-dependence of a
function explicitly, hence the argument or index t will be suppressed.
Definition 4.1. It proves convenient to define the following measures of stress and rate of deformation.
(i) The negative mean Cauchy stress and the Cauchy stress deviator
def
p = − 31 tr σ
and
def
σ dev = σ + p g ♯ ,
respectively.
(ii) The von Mises stress or equivalent shear stress
q
def p
q = 3J2 = 32 kσ dev k ,
in which
def 1
2
2
1
2 tr(σ dev ) = 3 (I1 (σ)) − I2 (σ)
= 12 sij sji = 21 (s211 + s222 + s233 + 2s212 + 2s223 +
= 16 ((σ1 − σ2 )2 + (σ2 − σ3 )2 + (σ3 − σ1 )2 ) .
J2 =
15
2s213 ) = 21 (s21 + s22 + s23 )
is the negative second principal invariant of the Cauchy stress deviator. Here I1 (σ) and I2 (σ) denote the first
and second the principal invariants of the Cauchy stress, respectively. Moreover, sij , with i, j ∈ {1, . . . , 3}
are the components of σ dev and σ1 , σ2 , σ3 the principal stresses in three-dimensional Euclidian space.
(iii) The equivalent shear strain rate and the volumetric strain rate
r
2
def
def
diso =
tr(d2dev )
and
dvol = tr d = dkk ,
3
def
respectively, with ddev = d − 31 (tr d)g.
4.1
♦
Hypoelasticity
The use of spatial rate constitutive equations to characterize the mechanical behavior of materials is very
attractive, especially from a numerical viewpoint. In addition, there is only a limited number of materials,
e.g. rubber, whose elastic response resp. stress state can be derived as a whole, either from a finite strain
measure (say C), or a free energy function. Truesdell [93] points out:
While the last few years have brought physical confirmation to the [hyperelastic; note from the author]
finite strain theory for rubber, there remain many physical materials which are linearly elastic under
small enough strain but which in large strain behave in a fashion the finite strain theory is not intended
to represent.
This observation led to the development of hypoelastic rate constitutive equations [92, 93, 95].
Definition 4.2. The general hypoelastic constitutive equation is defined through
def
σ ⋆ = h(σ, g, d) = a(σ, g) : d
◦
(linearity in d) ,
where σ ⋆ can be any objective rate of Cauchy stress, and a(σ, g) is a spatial fourth-order tensor-valued
function. To achieve the equivalence h(σ, g, d) = a(σ, g) : d, the function h is required to be continuously
differentiable in a neighborhood of d = 0, so that h is linear in d; note that a(σ, g) = Dh(σ, g, 0), and that
h(σ, g, 0) = 0, i.e. zero rate of deformation produces zero objective stress rate. If only rate-independent
response should be modeled, then h must be positively homogeneous of first degree in d, i.e. h(σ, g, ad) =
ah(σ, g, d) for all a > 0.
♦
◦
Definition 4.3. A material is hypoelastic of grade n, if a(σ, g) is a polynomial of degree n in the components
of σ [92, 95]. For n = 0, representing hypoelasticity of grade zero, the tensor a(g) is independent of σ. The
simplest ad hoc choice compatible with this idea is the constant isotropic elasticity tensor
aijkl = K g ij g kl + 2G(g ik g jl + g il g jk −
1
3
g ij g kl ) .
Here g ij are the components of the inverse metric, K = λ+ 23 µ is the bulk modulus or modulus of compression,
G = µ is the shear modulus, and λ, µ are the Lamé constants. The considered grade-zero hypoelastic rate
constitutive equation takes the equivalent forms
def
σ ⋆ = K(tr d) g ♯ + 2G d♯dev
◦
and
resp.
◦
def
σ ⋆ = λ(tr d) g ♯ + 2µ d♯ .
◦
def
(σ ⋆ )ij = Kdkk g ij + 2G dij
dev
♦
Remark 4.1. Within the hypoelasticity framework the stress is not necessarily path-independent such that
hypoelastic constitutive equations generally produce non-zero dissipation in a closed cycle [77]. Bernstein [11]
proposed conditions to proof if a certain hypoelastic model represents an elastic or even hyperelastic material,
i.e. elastic in the sense of Cauchy and Green, respectively. If a certain hypoelastic model is elastic, additional
conditions must hold so that the model represents a hyperelastic material. Simo and Pister [80] show that
any grade-zero hypoelastic constitutive equation with constant isotropic tensor according to Definition 4.3
cannot represent an elastic material. Instead, the components of a must be nontrivial functions of the
Jacobian J of the motion, and must also reduce to the linear elastic case for J = 1 [80].
△
16
Remark 4.2. Two decades ago, Xiao et al. [101, 100] proved that the grade-zero hypoelastic constitutive
def
equation σ ⋆ = K(tr d) g ♯ + 2G d♯dev is exactly integrable to define an isotropic elastic constitutive equation
◦
in the sense of Cauchy if and only if the stress rate σ ⋆ on the left hand side is the so-called logarithmic stress
rate
def
◦
σ log = σ̇ − Ω log · σ + σ · Ω log .
◦
The resulting finite strain constitutive equation is (see also Remark 3.5)
σ = K(tr e) g ♯ + 2G e♯dev ,
def
where e = ln V is the spatial logarithmic strain. Furthermore, Xiao et al. [102] show that if σ is replaced
def
with the Kirchhoff stress τ = J σ, then the integrable-exactly hypoelastic constitutive rate equation τ log =
λ(tr d) g ♯ + 2µ d♯ defines the isotropic hyperelastic (i.e. Green-elastic) relation
◦
τ = λ(ln(det V )) g ♯ + 2µ e♯ .
△
Remark 4.3. According to the classical treatment in [95, p. 404], the stress rate σ ⋆ in the general hypoelastic constitutive equation (Definition 4.2) can be any spatial objective rate of the Cauchy stress expressible
◦
◦
in the form σ ⋆ = σ ZJ +f (σ, g, d), with a 20 -tensor-valued function f linear in all arguments. Such objective
rates include, for example, the Oldroyd rate and Truesdell rate (Definition 3.2), but not the Green-Naghdi
rate (Definition 3.6). However, quoting Truesdell and Noll [95, p. 404], “[...] any advantage claimed for
◦
one such rate over another is pure illusion.” Indeed, any objective rate of the Cauchy stress can replace σ ⋆
provided that the right hand side is properly adjusted.
△
◦
Remark 4.4. Although Truesdell and Noll [95, p. 405] point out that hypoelasticity of grade zero “[...] is
not invariant under change of invariant stress rate” several article are concerned with the following question
[40, 105]: which objective stress rate should be applied to hypoelasticity of grade zero with constant isotropic
elasticity tensor according to Definition 4.3? That question arises after Dienes [20] and others show that
for hypoelasticity of grade zero the choice of the Zaremba-Jaumann rate of the Cauchy stress would lead to
oscillating stress response in simple shear, which is indeed unacceptable (cf. Sect. 6). Nowadays, researchers
agree that the question as posed is meaningless because the claim for a constant isotropic elasticity tensor
under large deformations is yet unacceptable [80]. However, for arbitrary rate constitutive equations the
question remains. According to Atluri [2] and Nemat-Nasser [55], it is not the Zaremba-Jaumann rate that
generates the spurious stresses, but the constitutive rate equation relating the Zaremba-Jaumann rate of the
response functions to their dependent variables. In particular, Atluri [2, p. 145] points out that
[...] all stress-rates are essentially equivalent when the constitutive equation is properly posed [i.e. if the
terms by which the rates differ are incorporated into the constitutive equation; note from the author].
The author of the present paper strongly believes that this and other issues associated with rate constitutive
equations could be considered obsolete if stated properly within the context of geometric constitutive theory
[67, 69, 72].
△
4.2
Hypoelasto-Plasticity
Elasto-plastic constitutive equations in finite element codes for large deformation solid mechanical applications are mostly based on an ad hoc extension of classical small-strain elasto-plasticity to the finite deformation range [77]. The presumed “elastic” part is described by a hypoelastic model, hence the term
hypoelasto-plasticity has been coined for that class of constitutive equations. In classical plasticity theory,
plastic flow is understood as an irreversible process characterized in terms of the past material history.
Definition 4.4. The past material history up to current time t ∈ R is defined as a map
]− ∞, t] ∋ τ
7→
{σ(x, τ ), α(x, τ )} ,
def
where α(x, τ ) = {α1 (x, τ ), . . . , αk (x, τ )} is a set of (possibly tensor-valued) internal state variables, often
referred to as the hardening parameters or plastic variables.
♦
17
-tensor fields, and let the internal plastic variables
2
0
Definition 4.5. Let T2sym be the set of symmetric
def
def
α = {α1 , . . . , αk } belong to the set identified through H = {α | α ∈ H}. A state of an elasto-plastic
material is the pair (σ, α) ∈ T2sym × H. The ad hoc extension of classical small-strain elasto-plasticity to the
finite deformation range then consists of the following elements [77]:
(i) Additive decomposition. The spatial rate of deformation tensor is additively decomposed into elastic and
plastic parts:
def
d = de + dp ,
def
dij = deij + dpij .
in components,
(ii) Stress response. A hypoelastic rate constitutive equation of the form
σ ⋆ = a(σ, g) : (d − dp )
◦
characterizes the “elastic” response, where σ ⋆ represents any objective stress rate.
◦
(iii) Elastic domain and yield condition. A differentiable function f : T2sym × T2sym × H → R is called the
yield condition, and
def
Aσ = {(σ, α) ∈ T2sym × H | f (σ, g, α) ≤ 0}
is the set of admissible states in stress space; the explicit dependency on the metric g is necessary in order
to define invariants of σ and α. An admissible state (σ, α) ∈ Aσ satisfying f (σ, g, α) < 0 is said to belong
to the elastic domain or to be an elastic state, and for f (σ, g, α) = 0 the state is an elasto-plastic state lying
on the yield surface. States with f > 0 are not admissible.
(iv) Flow rule and hardening law. The evolution equations for dp and α are called the flow rule and hardening
law, respectively:
def
dp = λ m(σ, g, α)
and
def
α⋆ = −λ h(σ, g, α) .
◦
Here m and h are prescribed functions, and λ ≥ 0 is called the consistency parameter or plastic multiplier.
The flow rule is called associated if m = Dσ f , and non-associated if m is obtained from a plastic potential
g 6= f as m = Dσ g. Within the isotropic hardening laws α usually represents the current radius of the yield
surface, whereas α represents the center of the yield surface (back stress) in kinematic hardening laws.
(v) Loading/unloading and consistency conditions. It is assumed that λ ≥ 0 satisfies the loading/unloading
conditions
λ ≥ 0,
f (σ, g, α) ≤ 0 ,
and
λ f (σ, g, α) = 0 ,
as well as the consistency condition
λ f˙(σ, g, α) = 0 .
♦
Example 4.1 (Von Mises Plasticity). Consider a well-known hypoelasto-plastic rate constitutive equation which is commonly referred to as J2 -plasticity with isotropic hardening or von Mises plasticity in computational solid mechanics [53, 37, 77]. This model is applicable to metals and other materials because it
includes the von Mises yield condition
def
f (σ, g, σ y ) = q(σ, g) − σ y ,
where q is the von Mises stress (Definition 4.1), and σ y is the current yield stress given by the linear hardening
rule
def
σ y (εp ) = σ y0 + E p εp .
The initial yield stress σ y0 and the plastic modulus E p are material constants in addition to the elastic
constants E and ν (or K and G), and the equivalent plastic strain εp is understood as a function of the
plastic rate of deformation tensor dp . Including the linear hardening rule produces bilinear elasto-plastic
response with isotropic hardening mechanism. Bilinear in this context means that a one-dimensional bar in
simple tension behaves elastic with Young’s modulus E until reaching the initial yield stress. Then plastic
flow occurs and the material hardens according to the linear hardening rule. The elasto-plastic tangent
modulus is given by the constant
p
def E E
.
Et =
E + Ep
18
Let the hypoelastic response be characterized by
σ ZJ = a(g) : (d − dp ) ,
◦
where a(g) is the constant isotropic elasticity tensor (Definition 4.3). Plastic flow is assumed to be associated,
that is,
r
∂f
3 ♭
3 ♭
def
p def p
d = ddev = λ
=λ
σ
=λ
n ,
∂σ
2q dev
2
def
where n = σ dev /kσdev k, with tr n = 0. Therefore, plastic straining is purely deviatoric, and the hardening
law, representing the evolution of the radius of the von Mises yield surface, is given by
r
2
y
p p
p
tr((dp )2 ) = λ E p
resp.
ε̇p = λ .
σ̇ = E ε̇ = E
3
After substitution into the consistency condition during plastic loading, the plastic multiplier is obtained as
r
2G
3
λ=
n : ddev ,
3G + E p 2
which completes the model. Some algebraic manipulation finally results in the hypoelasto-plastic spatial
rate constitutive equation
◦
σ ZJ = a ep (σ, g, εp ) : d ,
in which the elasto-plastic material tangent tensor is given by
def
aep (σ, g, εp ) = a(g) −
6G2
n⊗n
3G + E p
at plastic loading, and by a ep (σ, g, εp ) = a(g) at elastic loading and unloading, and neutral loading, respectively. The distinction of these types of loading is done with the aid of the yield condition and hardening
rule, that is, the dependency of the function aep on εp is implicit.
∃
4.3
Hypoplasticity
The notion of hypoplasticity, which is entirely different from that of hypoelasto-plasticity, has been introduced
by Kolymbas [43], but the ideas behind are much older. Starting in the 1970’s [29, 41], the development
of hypoplastic rate constitutive equations has a clear focus on granular materials and applications in soil
mechanics [8, 24, 28, 42, 56, 98].
Definition 4.6. The general hypoplastic constitutive equation for isotropic materials takes the form
def
σ ⋆ = h(σ, g, α, d) ,
◦
def
where σ ⋆ represents any objective stress rate and α(x, t) = {α1 (x, t), . . . , αk (x, t)} is a set of (possibly
tensor-valued) internal state variables.
♦
◦
Hypoplasticity can be understood as a generalization of hypoelasticity. In contrast to hypoelasticity, the
hypoplastic response function h is generally nonlinear in d in order to describe dissipative behavior. Hypoplastic constitutive modeling basically means to fit the almost arbitrary tensor-valued response function
h to experimental data. That makes it to an deductive design approach, whereas elasto-plastic constitutive
modeling is inductive. A basic requirement is that the desired function be as simple as possible. In the
◦
simplest hypoplastic model the objective stress rate σ ⋆ is regarded a nonlinear function of g and d only.
def
Example 4.2. Consider the case where σ ⋆ = h(σ, g, d). If rate-independent material should be described,
then h is required to be positively homogeneous of first degree in d, so that for every a > 0, h(σ, g, ad) =
ah(σ, g, d). In this case, however,
◦
h(σ, g, d) =
∂h(σ, g, d)
: d = m(σ, g, d) : d
∂d
19
by Euler’s theorem on homogeneous functions, where m is a spatial fourth-order tensor-valued function that
explicitly depends on d ; a proof can be done by differentiating both sides of f (ad) = af (d) with respect
to a and then applying the chain rule. The values of m are referred to as material tangent tensors, as for
other classes of rate constitutive equations. Due to constitutive frame invariance (Principle 2.1) the response
def
function h must be isotropic in all variables. It then follows that σ ⋆ = h(σ, g, d) has the representation
(see [95, eq. (13.7)] and [43, eq. (15)])
◦
σ ⋆ = ψ0 g ♯ + ψ1 σ + ψ2 d + ψ3 σ 2 + ψ4 d2 + ψ5 (σ · d + d · σ)
◦
+ ψ6 (σ 2 · d + d · σ 2 ) + ψ7 (σ · d2 + d2 · σ) + ψ8 (σ 2 · d2 + d2 · σ 2 ) ,
where ψ0 , . . . , ψ8 are polynomials of the ten basic invariants tr σ, tr(σ 2 ), tr(σ 3 ), tr d, tr(d2 ), tr(d3 ), tr(σ·d),
tr(σ·d2 ), tr(σ 2 ·d), tr(σ 2 ·d2 ). Note that in the above equation the symbol ♯ denoting index raising has been
omitted for the d-terms.
∃
5
5.1
Objective Time Integration
Fundamentals and Geometrical Setup
Determination of the motion ϕ : B × [0, T ] → S from balance of linear momentum (Definition 2.22) requires
the total Cauchy stress σ at every time t ∈ [0, T ]. If a given constitutive equation calculates only a rate
of stress but not total stress, then the latter represents the solution of an initial value problem. A formal
description of this situation is given below. For simplicity, we consider only rate constitutive equations
that determine an objective corotational rate of Cauchy stress according to Definition 3.3. Recall that
examples of such rates are the widely-used Zaremba-Jaumann rate (Definition 3.5) and Green-Naghdi rate
(Definition 3.6).
Definition 5.1. The considered class of corotational rate constitutive equations takes the general form
def
def
σ ⋆ = h(σ, α, d) = m(σ, α) : d ,
◦
where σ ⋆ is any objective corotational rate of Cauchy stress, α is a set of state variables in addition to stress,
d is the rate of deformation, and m is the material tangent tensor. By defining
◦
def
h̄(σ, α, d, Λ) = h(σ, α, d) + Λ · σ − σ · Λ ,
where Λ = −ΛT is the spin tensor associated with σ ⋆ , the constitutive equation takes the equivalent form
σ̇ = h̄(σ, α, d, Λ). Moreover, each element in the set α is assumed to have evolution equations similar to
those of stress, that is, α̇ = k(σ, α, d, Λ).
♦
◦
Definition 5.2. An incremental decomposition of time is a disjoint union
def
[0, T ] =
N[
−1
[tn , tn+1 ] ,
n=0
motivating a sequence (t0 = 0, t1 = t0 + ∆t1 , . . . , tn+1 = tn + ∆tn+1 , . . . , tN = T ) of discrete time steps
tn+1 = tn +∆tn+1 with time increment ∆tn+1 . For simplicity, we assume that the time increment is constant,
that is, ∆tn+1 ≡ ∆t such that tn+1 = (n + 1)∆t for t0 = 0. Let ϕn (B) and ϕn+1 (B) be configurations of the
material body B at time tn and tn+1 , respectively, then the incremental decomposition of stress is accordingly
defined by
def
σ n+1 = σ n + ∆σ ,
in which
σ n : ϕn (B) →
T02 (S)
,
def
σ n = σ(tn ) ,
and
def
∆σ =
Z
tn+1
σ̇(t) dt .
tn
Similar holds for the state variables.
♦
20
Definition 5.3. A time integration of the rate constitutive equation in Definition 5.1 determines the stress
and material state, {σ, α}, at time t = tn+1 by considering the differential equations
σ̇(t) = h̄(t, σ(t), α(t), d(t), Λ(t))
and
α̇(t) = k(t, σ(t), α(t), d(t), Λ(t))
subject to the initial condition {σ, α}|t=tn = {σn , αn }. The time integration is called incrementally objective [38] if the stress is exactly updated (i.e. without the generation of spurious stresses) for rigid motions
ϕt : B → S over the incremental time interval [tn , tn+1 ], that is, if
σ n+1 = ϕ ⇑ σ n = Q · σ n · QT ,
def
where Q = T ϕ is proper orthogonal. The same is required for tensor-valued state variables, if any.
♦
Since the rate constitutive equations are generally non-linear functions of their arguments, the time integration must be carried out numerically by employing suitable time integration methods; also called stress
integration methods in the present context. The choice of the stress integration method plays a crucial
role in numerical simulation of solid mechanical problems because it affects the stability of the solution
process and the accuracy of the results. Most stress integration methods are customized for small-strain
elasto-plastic resp. hypoelasto-plastic constitutive rate equations that include yield conditions. Early works
include [45, 46, 54, 75, 81, 82, 99], and a comprehensive treatise is that of Simo and Hughes [77].
The time integration is usually split into two different phases: the objective update, which is only present at
finite deformations, and the actual integration of the stress rate. The initial value problem associated with
stress integration can be solved either by explicit schemes or implicit schemes. Explicit stress integration
methods are formulations using known quantities at the beginning of the time step, like the forward Euler
scheme. The procedure is straightforward, and the resulting equations are almost identical to the analytical
set up. However, the simplicity of the implementation fronts the stability constraint and error accumulation
during calculation, since generally no yield condition is enforced. Accuracy can be increased by partitioning
the time increment into a number of substeps, and to perform automatic error control [82, 83].
Implicit stress integration methods are based on quantities taken with respect to the end of the time step,
like the backward Euler scheme. Operator-split procedures are preferred to solve the coupled system of
nonlinear equations. From a geometric standpoint, the implicit stress update with operator-split projects
an elastically estimated trial state onto the yield surface. The plastic multiplier serves as the projection
magnitude; the plastic multiplier is zero in case of elastic loading, unloading, and neutral loading. The
yield condition is naturally enforced at the end of the time increment. Therefore, at the same increment
size, implicit algorithms can be more accurate than explicit algorithms. The numerical implementation
is, however, more complicated because generally the plastic multiplier has to be obtained from the yield
condition by an iterative procedure. This also generates computational overburden. In spite of this, implicit
stress integration methods became standard in small-strain elasto-plasticity and hypoelasto-plasticity.
Remark 5.1. Consider the initial value problem defined through
ẏ(t) = f (t, y(t))
subject to the initial condition (tn , y n ). A solution to that problem is a function y that solves the differential
equation for all t ∈ [tn , tn+1 ] and satisfies y(tn ) = y n . Different approaches are available to obtain an
approximate solution. The explicit forward Euler method, for example, uses a first-order approximation to
the time derivative:
y(tn + ∆t) − y(tn )
.
ẏ(t) ≈
∆t
def
Setting y(tn ) = y n and noting that ẏ(tn ) = f (tn , y(tn )) by definition, then
y n+1 = y n + ∆t f (tn , y n ) .
In contrast to explicit methods, the implicit integration methods use quantities defined at the end of the
time increment. Since these are generally unknown, they have to be estimated and subsequently corrected
by an iteration. For example, the backward Euler method uses the approximation
ẏ(t) ≈
y(tn+1 ) − y(tn+1 − ∆t)
,
∆t
21
by which
y n+1 = y n + ∆t f (tn+1 , y n+1 ) .
Combination of both methods yields the generalized midpoint rule
y n+1 = y n + ∆t fn+θ ,
where
def
fn+θ = θf (tn+1 , y n+1 ) + (1 − θ)f (tn , y n ) ,
The Crank-Nicolson method is obtained by setting θ =
1
2.
mit θ ∈ [0, 1] .
△
The rotational terms of the stress rate (Definition 5.1) present at finite deformation render the integration of
rate constitutive equations expensive compared to the infinitesimal case. Subsequent to the work of Hughes
and Winget [38], who have introduced the notion of incremental objectivity formalized in Definition 5.3,
several authors have developed or improved incrementally objective algorithms, e.g. [25, 37, 63, 66, 73,
77]. One basic methodology in formulating objective integration methods utilizes a corotated or rotationneutralized representation. Within this approach, the basic quantities and evolution equations are locally
transformed to a rotating coordinate system that remains unaffected by relative rigid motions; the local
coordinate system “corotates” with the relative rotation of the body. Then, the constitutive equation is
integrated in the corotated representation by using the algorithms outlined above, and is finally rotated
back to the current spatial configuration at time tn . The main advantage of this class of algorithms is that
the integration of the rate constitutive equation can be carried out by the same methods at both infinitesimal
and finite deformations.
The remainder of this section is largely based on [77, ch. 8] and [37]. It introduces the integration of rate
constitutive equations for finite deformation problems. The main concern is the numerical method designed
in such a way that the requirement of incremental objectivity is identically satisfied. Concerning details
on stress integration methods at infinitesimal deformations the reader is referred to the cited literature,
particularly [77].
A geometrical setup for objective integration of rate equations is introduced as follows. Consider a regular
and sufficiently smooth motion ϕ : B × [0, T ] → S of a material body B in the m-dimensional Euclidian
space S = Rm . Let v(x, t) be the spatial velocity field of that motion, defined for every x = ϕ(X, t) ∈ S
and time t ∈ [0, T ], where X ∈ B, and v t : ϕt (B) → T S at fixed t. The spatial rate of deformation d, with
d♭t : ϕt (B) → T20 (S), is understood as a measure of strain rate; indeed d has been proven to be an honest
strain rate just a few years ago (cf. Remark 3.5).
Definition 5.4. Let time be incrementally decomposed according to Definition 5.2 such that tn+1 = tn +
∆t ∈ [0, T ], with t0 = 0 and ∆t > 0. Moreover, let
def
ϕ(B, tn ) = ϕn (B) = {xn = ϕn (X) | X ∈ B}
⊂ S = Rm
be a given configuration of B at time tn ∈ [0, T ], and
def
u = v∆t :
ϕn (B) → T S
a given incremental spatial displacement field imposed on ϕn (B) which is constant over the time increment
[tn , tn+1 ]. The incremental material displacement field is then defined through U = u ◦ ϕn .
♦
Definition 5.5. The spatial finite strain tensor e at time t ∈ [0, T ] is defined conceptually through
Z t
def
e(x, t) = e(x, 0) +
d(x, τ ) dτ ,
0
where e(x, 0) is given. The overall accuracy of the stress integration method is then affected by the approxdef R t
imate evaluation of the finite strain increment ∆e = tnn+1 d(t) dt constant in [tn , tn+1 ].
♦
According to Simo and Hughes [77, p. 279], the aim in developing objective integration algorithms
[...] is to find algorithmic approximations for spatial rate-like objects, in terms of the incremental displacements u(xn ) and the time increment ∆t, which exactly preserve proper invariance under superposed
rigid body motions [...].
22
To achieve the objectives, we use the following.
Definition 5.6. A one-parameter family of configurations is the linear interpolation between ϕn and ϕn+1
defined through
def
ϕn+θ = θϕn+1 + (1 − θ)ϕn ,
with θ ∈ [0, 1] .
Conceptually, the intermediate configuration ϕn+θ is related to an intermediate time tn+θ = θtn+1 + (1 −
θ)tn = tn + θ∆t. The configuration ϕn+1 in the Euclidian ambient space Rm can be determined by adding
the incremental displacements to the configuration at time tn , that is, ϕn+1 (X) = ϕn (X) + U (X) ∈ T Rm .
Accordingly, the deformation gradient of ϕn+θ is given by the relationship
Fn+θ = T ϕn+θ = θFn+1 + (1 − θ)F n ,
with θ ∈ [0, 1] .
The relative incremental deformation gradient of the configuration ϕn+θ (B) with respect to the configuration
ϕn (B) is then defined through
def
fn+θ = Fn+θ · Fn−1 ,
with θ ∈ [0, 1] .
♦
Definition 5.7. The relative incremental displacement gradient is the tensor field ∇n+θ u ∈ T11 (S) which
has the local representative
!
∂ui (xn+θ )
∂
def
i
j
(∇n+θ u)(xn+θ ) =
+ u (xn+θ )γj k dxkn+θ ⊗
.
∂xin+θ
∂xkn+θ
def
def
at xn+θ = ϕn+θ (X), where ui (xn+θ ) = U I (X)|X=ϕ−1
n+θ
(xn+θ )
are the components of the incremental dis-
placements referred to the configuration ϕn+θ (B). The spatial connection coefficients γj ik are understood to
be taken with respect to xn+θ . In a spatial Cartesian coordinate system {z b }, recall that ∇n+θ u can likewise
be expressed by
∂ubz (xn+θ )
∂u(xn+θ )
=
(∇n+θ u)(xn+θ ) ≡
ed ⊗ eb ,
d
∂xn+θ
∂zn+θ
def
d
where zn+θ
= z d (xn+θ ).
♦
Definition 5.8. The relative right Cauchy-Green tensor Cn+1 and the relative Green-Lagrange strain En+1
of the configuration ϕn+1 (B) with respect to the configuration ϕn (B) are defined by
def
T
Cn+1 = fn+1
· fn+1
def
2E n+1 = Cn+1 − I
and
∈ T11 (ϕn (B)) ,
respectively; composition with the map ϕ has been dropped. I is the unit tensor on ϕn (B). On the other
hand, define the relative left Cauchy-Green tensor bn+1 and the relative Euler-Almansi strain eEA
n+1 of the
configuration ϕn+1 (B) relative to ϕn (B) through
def
T
bn+1 = fn+1 · fn+1
and
def
−1
2eEA
n+1 = i − bn+1
∈ T11 (S) ,
respectively. Here i is the unit tensor on S.
♦
By these definitions and the basic relationships of Section 2, the next results are obtained; see [77, secs. 8.1
and 8.3] for full proofs with respect to Cartesian frames.
Proposition 5.1. Let [tn , tn+1 ] be an incremental time interval, tn+1 = tn + ∆t, and let θ ∈ [0, 1]. Then,
(i) the relative incremental deformation and displacement gradients are equivalent to
fn+θ = I + θ(∇n u)T
−1
(∇n+θ u)T = (∇n u)T · fn+θ
,
and
respectively, where ∇n u(xn ) ≡ ∂u(xn )/∂xn in Cartesian coordinates,
(ii) objective approximations to the spatial rate of deformation in [tn , tn+1 ] are
1 −T
−1
f
· E ♭n+1 · fn+θ
,
∆t n+θ
1 −T
−1
♭
=
f
· f T · (eEA
and
n+1 ) · fn+1 · fn+θ ,
∆t n+θ n+1
1
=
(∇n+θ u)T + ∇n+θ u + (1 − 2θ)(∇n+θ u)T · ∇n+θ u ,
2∆t
d♭n+θ =
d♭n+θ
dn+θ
23
(iii) an algorithmic approximations to the vorticity is
ω n+θ =
1
(∇n+θ u)T − ∇n+θ u .
2∆t
Proposition 5.2. Let s ∈ T02 (S) be a covariant second-order spatial tensor field in [tn , tn+1 ], then (objective)
algorithmic approximations to its Lie derivative are provided through
Lv sn+θ =
1 −T
−1
f
· f T · sn+1 · fn+1 · fn+θ
.
∆t n+θ n+1
Proof. This follows from the second equation in Proposition 5.1(ii) by similarity to Proposition 2.7 and
using Propositions 2.1 and 2.3.
Remark 5.2. Propositions 5.1 and 5.2 are remarkable in the sense that, by Definition 5.7, the approximations to the discrete rate of deformation, the discrete vorticity, and the discrete Lie derivative in [tn , tn+1 ]
can be obtained with the aid of either total or incremental deformation gradients. These discrete spatial
variables thus do not depend on the choice of a reference configuration, which is consistent to the continuous
theory.
△
Definition 5.9. The algorithmic finite strain increment or incremental finite strain tensor is defined by
def
∆ẽn+θ = dn+θ ∆t ,
where dn+θ is according to Proposition 5.1(ii). Hence, ∆ẽn+θ is likewise incrementally objective.
♦
Remark 5.3. For the choice θ = 0, ∆ẽn+θ coincides with the relative Green-Lagrange strain En+1 . In
case of θ = 1, ∆ẽn+θ is identical to the relative Euler-Almansi strain eEA
n+1 . This follows directly from
Definition 5.8 and Proposition 5.1(ii). Hughes [37] has shown that Definition 5.9 is a first-order approximation
to the finite strain increment (Definition 5.5) for all θ. If θ = 21 , then the approximation is second-order
accurate. Moreover, for θ = 12 , referred to as the midpoint strain increment, the approximation is linear in
u (cf. Proposition 5.1(ii)). Therefore, the midpoint strain increment is the most attractive expression from
the viewpoint of implementation.
△
5.2
Algorithm of Hughes and Winget
The algorithm of Hughes and Winget [38] is probably the most widely used objective stress integration
method in nonlinear finite element programs. It considers a class of constitutive rate equations of the form
◦
σ ZJ = m(σ, α) : d or, equivalently,
def
σ̇ = m(σ, α) : d + ω · σ − σ · ω ,
where σ ZJ is the Zaremba-Jaumann rate of Cauchy stress (Definition 3.5). Recall from Sect. 3.2 that σ ZJ is
a corotational rate of σ defined by the spin ω = −ωT . The spin generates a one-parameter group of proper
orthogonal transformations by solving
◦
◦
Ṙ = ω · R ,
subject to
R|t=0 = I ,
where R is a proper orthogonal two-point tensor (cf. Definition 3.4).
Based on the observations summarized in Remark 5.3, Hughes and Winget [38] employ time-centered (θ = 21 )
approximations of d and ω. This leads to a midpoint strain increment and midpoint rotation increment as
follows.
def
def
Proposition 5.3. Let ∆ẽn+1/2 = dn+1/2 ∆t and ∆r̃n+1/2 = ω n+1/2 ∆t, then
∆ẽn+1/2 = 12 (∇n+1/2 u)T + ∇n+1/2 u
and
∆r̃ n+1/2 = 12 (∇n+1/2 u)T − ∇n+1/2 u ,
respectively, where
(∇n+1/2 u)T = 2(fn+1 − I)(fn+1 + I)−1 .
Proof. By straightforward application of Proposition 5.1. Similar expressions have been derived in [63,
eq. (41)] and [66, eq. (36)].
24
The proof of the next statement is one of the main concerns of [38].
Proposition 5.4. The generalized midpoint rule (Remark 5.1), with θ = 21 , is used to approximately integrate Ṙ = ω · R subject to R|t=0 = I, resulting in the proper orthogonal relative rotation
−1
1
∆R = Rn+1 · RT
(I + 12 ∆r̃ n+1/2 ) .
n = (I − 2 ∆r̃ n+1/2 )
Definition 5.10. The objective stress integration method of Hughes and Winget [38] can be summarized as
def
σ n+1 = σ ′n+1 + ∆σ ,
where
def
def
σ ′n+1 = ∆R ⇑ σ n = ∆R · σ n · ∆RT ,
∆σ = y(σ ′n+1 , α′n+1 , ∆ẽn+1/2 ) ,
def
α′n+1 = ∆R ⇑ αn ,
∆R is calculated according to Proposition 5.4, and ∆R ⇑ denotes the associated pushforward. The update,
properly adjusted, has to be applied to any tensor-valued material state variable. The complete procedure
is in Alg. 1, and incremental objectivity has been proven in [38, 37].
♦
Algorithm 1: Objective integration of rate equations according to Hughes and Winget [38].
Input: geometry xn , incremental displacements u, stress σ n , and state variables αn
Output: σ n+1 , αn+1 , and material tangent tensor m
1
2
3
4
5
6
compute fn+1 = I + (∇n u)T and (∇n+1/2 u)T = 2(fn+1 − I)(fn+1 + I)−1 ;
obtain midpoint strain increment ∆ẽn+1/2 and rotation increment ∆r̃ n+1/2 (Prop. 5.3);
compute ∆R = (I − 21 ∆r̃ n+1/2 )−1 (I + 12 ∆r̃ n+1/2 );
rotate stress and state variables by ∆R, resulting in σ ′n+1 and α′n+1 , respectivly;
integrate constitutive equation using ∆ẽn+1/2 , σ ′n+1 , α′n+1 as for infinitesimal deformations;
compute material tangent tensor m if necessary;
Remark 5.4. The stress update can be interpreted as follows. The full amount of relative rotation ∆R
over the time increment [tn , tn+1 ] is applied instantaneously to the stress at time tn , σ n , in order to account
for rigid body motion. The rotated stress σ ′n+1 , more precisely, the ∆R-pushforward of σ n , the rotated
state variables α′n+1 etc., are then passed to the procedure that integrates the rate constitutive equation
without any rotational terms by the methods outlined in Section 5.1. It is emphasized that no choice of such
a integration procedure, e.g. explicit or implicit, is defined by Hughes and Winget’s algorithm. However, in
case where the material tangent tensor, m(σ, α), is an isotropic function of its arguments and explicit stress
integration is employed, the stress increment can be obtained in closed-form from
∆σ = m(σ ′n+1 , α′n+1 ) : ∆ẽn+1/2 .
△
Remark 5.5. Alg. 1 places a restriction to the magnitude of k∆r̃ n+1/2 k = kωn+1/2 k∆t. From the approximation 21 k∆r̃n+1/2 k ≈ tan k 12 ∆r̃n+1/2 k, [77, eq. 8.3.24], it follows that ∆R according to Proposition 5.4 is
defined only for k 21 ∆r̃ n+1/2 k < 180◦.
△
5.3
Algorithms Using a Corotated Configuration
The class of algorithms discussed in the following are ideally suited for corotational rate constitutive equations
(Definition 5.1). These algorithms go back at least to Nagtegaal and Veldpaus [53] and Hughes [37]. Recall
from Section 3.2 that any corotational rate of a spatial second-order tensor involves a spin Λ = −ΛT . The
spin generates a one-parameter group of rotations associated with the initial value problem Ṙ = Λ·R subject
to R|t=0 = I, see Definition 3.4, where R is proper orthogonal, i.e. a rotation. The crucial observation that
leads to the considered class of algorithms can then be stated as follows.
Proposition 5.5. Let σ ⋆ = σ̇ − Λ · σ + σ · Λ be any corotational rate of Cauchy stress defined by the spin
tensor Λ = Ṙ · RT , then
◦
σ ⋆ = h(σ, α, d)
◦
and
R⇑
∂
(R ⇓ σ) = R ⇑ (R ⇓ h(R ⇓ σ, R ⇓ α, R ⇓ d))
∂t
are equivalent rate constitutive equations.
25
Proof. Equivalence of the left hand sides of both equations has been shown in Proposition 3.8 for the
◦
◦
particular choice of the Green-Naghdi rate, σ ⋆ = σ GN . On the other hand, by Definition 2.30 in conjunction
with Definition A.28,
R ⇓ (h(σ, α, d)) = (R ⇓ h)(R ⇓ σ, R ⇓ α, R ⇓ d) .
Pushforward by R on both sides then yields
(R ⇑ ◦R ⇓)(h(σ, α, d)) = h(σ, α, d) = R ⇑ (R ⇓ h(R ⇓ σ, R ⇓ α, R ⇓ d))
as desired.
The proposition formalizes how to replace a corotational rate by the usual time derivative. Consequently,
a corotational rate constitutive equation can be integrated by transforming all variables to the corotating
R-system, performing the update of the stress and state variables, and then rotating the updated stress
tensor back to the current configuration.
Definition 5.11. Let σ ⋆ = m(σ, α) : d be a corotational rate constitutive equation for the Cauchy stress
defined by the spin Λ = Ṙ · RT (see also Sect. 3.2). A general class of objective algorithms based on a
corotated configuration is then defined by
◦
def
σ n+1 = Rn+1 · (Sn + ∆Sn+θ ) · RT
n+1 ,
where θ ∈ [0, 1],
def
Sn = Rn ⇓ σ n = RT
n · σ n · Rn ,
def
∆Sn+θ = fn+θ (Sn+θ , An+θ , ∆En+θ ) ,
def
∆En+θ = Rn+θ ⇓ ẽn+θ = RT
n+θ · ∆ẽn+θ · Rn+θ ,
def
and A = R ⇓ α. The stress increment ∆Sn+θ is evaluated at some rotation-neutralized intermediate configuration specified by the actual integration algorithm using θ ∈ [0, 1] and an associated rotation Rn+θ .
This evaluation is denoted conceptually, but without loss of generality, by the response function fn+θ representing an explicit or implicit stress integration method (cf. Sect. 5.1). The tensor ∆En+θ is called the
algorithmic corotated finite strain increment, and ∆ẽn+θ (Definition 5.9) is regarded as given. The algorithm
is incrementally objective provided that Rn , Rn+θ , and Rn+1 are properly determined:
Case (i): Λ = Ω, R = R. The rate constitutive equation is formulated in terms of the Green-Naghdi stress
◦
◦
rate, that is, σ ⋆ = σ GN . Recall that R is the rotation tensor resulting from the polar decomposition of the
deformation gradient, and Ω = Ṙ · RT .
Case (ii): Λ = ω, R 6= R. The rate constitutive equation is formulated in terms of the Zaremba-Jaumann
◦
◦
stress rate, that is, σ ⋆ = σ ZJ , requiring proper integration of Ṙ = ω · R (cf. Sect. 5.2).
♦
Remark 5.6. The above algorithm employs the generalized midpoint rule (Remark 5.1), that is, Sn+1 =
Sn + ∆Sn+θ , to emphasize that the general procedure is not affected by the choice of θ ∈ [0, 1]. If, for
example, an explicit stress integration procedure (θ = 0) is applied to the rate constitutive equation in the
R-system, then the stress increment can be calculated in closed-form:
∆Sn = (R ⇓ m)(Sn , An ) : ∆En .
△
Remark 5.7. The most obvious procedure to determine Rn+1 and Rn+θ in case (i) of Definition 5.11
is the polar decomposition of the total deformation gradients Fn+1 and Fn+θ , respectively. Alternative
procedures that circumvent polar decomposition have been proposed by Flanagan and Taylor [25] and Simo
and Hughes [77, sec. 8.3.2]. An algorithmic approximations to the vorticity tensor ω in case of the ZarembaJaumann rate (case (ii) in Definition 5.11) is provided by Proposition 5.1(iii). The rotation group can be
approximately integrated, for example, by using the general algorithm of Simo and Hughes [77, sec. 8.3.2],
or by the particular procedure of Hughes and Winget [38]; see also algorithm of Hughes [37] below.
△
The widely-used algorithm of Hughes [37] can be obtained from the general objective integration algorithm
in Definition 5.11 by making particular approximations to the orthogonal group of rotations and by using
time-centering, i.e. θ = 12 , in the calculation of the algorithmic finite strain increment. Time-centering is
employed in accordance with the incrementally objective algorithm developed by Hughes and Winget [38];
see Sect. 5.2. Hughes [37] originally uses the example of von Mises plasticity (cf. Example 4.1) and carried
out implicit time integration in the corotated configuration. However, Definition 5.11 generally places no
restrictions on the actual integration procedure (explicit, implicit, or generalized midpoint rule).
26
Definition 5.12. The corotated midpoint strain increment is defined by
def
∆En+1/2 = RT
n+1/2 · ∆ẽn+1/2 · Rn+1/2 ,
where ∆ẽn+1/2 = 21 ((∇n+1/2 u)T +∇n+1/2 u) is the second-order accurate midpoint strain increment (cf. Proposition 5.3 and Remark 5.3).
♦
Remark 5.8. Since ∆ẽn+1/2 = dn+1/2 ∆t by Definition 5.9, an algorithmic approximation to the corotated
rate of deformation tensor is thus given by
def
Dn+1/2 = RT
n+1/2 · dn+1/2 · Rn+1/2 .
Summation of ∆En+1/2 = Dn+1/2 ∆t over a time interval [t0 , tn+1 ] gives an excellent approximation to the
Lagrangian logarithmic strain [37]. Therefore, an algorithmic approximation to the Eulerian logarithmic
strain (Definition 2.16) can be obtained by applying Proposition 2.6(iii), leading to
!
X
ln V n+1 ≈ Rn+1 · E0 +
∆En+1/2 · RT
n+1 ,
n
where E0 is given.
△
The rotations Rn+1 and Rn+1/2 need to be determined in order to complete the algorithm of Definition 5.11.
◦
◦
In case of R = R, or equivalently, σ ⋆ = σ GN , Hughes [37] suggests polar decomposition of the total
deformation gradients Fn+1 = fn+1 · Fn and Fn+1/2 = 21 (Fn+1 + Fn ) in order to determine Rn+1 and
◦
◦
Rn+1/2 , respectively. In case of σ ⋆ = σ ZJ , where R 6= R, the rotation and half-step rotation are defined
through
Rn+1 = ∆R · Rn
and
Rn+1/2 = ∆R1/2 · Rn ,
where ∆R is the time-centered approximation to the incremental rotation according to Hughes and Winget [38];
see Proposition 5.4. For computation of the proper orthogonal square root ∆R1/2 the reader is referred to
the literature, e.g. [26, 35, 37, 91]. The complete integration procedure is summarized in Alg. 2.
Algorithm 2: Objective integration of rate equations according to Hughes [37].
Input: geometry xn , incremental displacements u, stress σ n , state variables αn , and rotation Rn
Output: σ n+1 , αn+1 , and material tangent tensor m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
compute fn+1 = I + (∇n u)T and (∇n+1/2 u)T = 2(fn+1 − I)(fn+1 + I)−1 ;
obtain midpoint strain increment ∆ẽn+1/2 and rotation increment ∆r̃ n+1/2 (Prop. 5.3);
◦
switch corotational rate σ ⋆ do
◦
case Green-Naghdi rate σ GN (R = R) do
compute Fn+1 = fn+1 · Fn and F n+1/2 = 21 (Fn+1 + Fn );
perform polar decomposition to obtain Rn+1 and Rn+1/2 ;
◦
case Zaremba-Jaumann rate σ ZJ (R 6= R) do
compute ∆R = (I − 21 ∆r̃n+1/2 )−1 (I + 21 ∆r̃ n+1/2 ) and ∆R1/2 ;
update Rn+1 = ∆R · Rn and Rn+1/2 = ∆R1/2 · Rn ;
corotate midpoint strain increment: ∆En+1/2 = RT
n+1/2 · ∆ẽn+1/2 · Rn+1/2 ;
integrate constitutive equation using ∆En+1/2 , Sn+1/2 , An+1/2 as for infinitesimal deformations;
compute material tangent tensor if necessary;
back-rotate updated stress to the current configuration: σ n+1 = Rn+1 · Sn+1 · RT
n+1 ;
back-rotate updated state variables and material tangent tensor to the current configuration;
Remark 5.9. According to the basic Definition 5.6, the deformation gradient of the motion is updated by
F n+1 = fn+1 · Fn , where Fn = Rn · Un , and Rn is proper orthogonal. Now, suppose that the current
configuration at time tn is taken as the reference configuration, i.e. B = ϕn (B), and no data is available
of configurations prior to tn such that Fn = Rn = Rn = I and Un = I. Then, by Remark 3.2, one has
ω ≡ Ω and the Zaremba-Jaumann and Green-Naghdi stress rates are identical. Moreover, the approximation
∆R to the incremental rotation according to [38], viz. Proposition 5.4, identically approximates the proper
orthogonal part Rn+1 of the polar decomposition of the deformation gradient Fn+1 ≡ fn+1 at time tn+1 .
This follows immediately from the definitions.
△
27
Remark 5.10. If ∆R, in case of Λ = ω, is determined from Proposition 5.4, then the same restriction is
placed to the magnitude of kωn+1/2 k∆t as in Alg. 1; see Remark 5.5.
△
Remark 5.11. The algorithms based on a corotated description go beyond the one of Hughes and Winget [38]
outlined in Sect. 5.2. The difference is that the algorithm of Hughes and Winget [38] rotates the stress and
state variables to the current configuration before passing it to the constitutive equation, whereas the algorithms using a corotated configuration use rotation-neutralized variables for calculation of the stress rate.
However, all these algorithms satisfy the requirement of incremental objectivity provided that the rotation
tensors are properly determined.
△
Remark 5.12. The main advantage of the investigated algorithms of [38, 37] is that the integration of
the constitutive rate equation can be carried out by the same methods at both infinitesimal and finite
deformations. That is, the objective algorithms comply with the usual small-strain algorithms if deformations
are infinitesimal. From a computational viewpoint this is very attractive, because the same material model
subroutine can be employed for both cases without changes. The “rotational” part of the stress update,
then, is done outside the subroutine. The algorithms, however, rely heavily on the use of corotational rate
constitutive equations. If the desired constitutive rate equation is based on a non-corotational rate, like the
Truesdell and Oldroyd rates (Definition 3.2), then additional terms need to be handled.
△
6
Hypoelastic Simple Shear
Hypoelastic simple shear is an excellent problem to analyze fundamental relations in nonlinear continuum
mechanics and to test implementations of objective time integration algorithms for rate equations. This
is because material deformations due to simple shear include both finite strains and finite rotations; it is
in fact a compound action of pure shear and pure rotation. Several papers are concerned with analytical
solutions of simple shear, mostly in connection with a discussion of objective stress rates for constitutive
rate equations [2, 16, 20, 21, 25, 40, 47]. They also serve as references for the numerical solution. Notations
and definitions of the previous sections are used throughout.
6.1
Analytical Solution
Definition 6.1. Let B ⊂ S = R3 be the initial configuration of a material body in the Euclidian space,
X ∈ B the initial location of a material particle, and ϕ : B × [0, T ] → S the motion of the body. Let
def
def
{Z 1 , Z 2 , Z 3 }X = {Z A }X and {z 1 , z 2 , z 3 } = {z a }x respectively denote the coordinate tuples of X and
x = ϕ(X, t) ∈ S with respect to an ortho-normalized frame in S = R3 . Simple shear then prescribes a planar
parallel motion of the form
def
z a = ϕa (Z A , t)
where the ϕa , a ∈ {1, 2, 3}, are respectively defined through
def
ϕ1 = Z 1 + k(t)Z 2 ,
def
ϕ2 = Z 2 ,
and
def
ϕ3 = Z 3 ,
and k(t) ∈ R with initial condition k(0) = 0. The problem statement is depicted in Fig. 1.
♦
Through the definition of an ortho-normalized frame of reference, every second-order tensor can be represented by a (3 × 3)-matrix of its components with respect to that frame. In particular, the deformation
gradient takes the form
1 k 0
1 0 0
0 k 0
∂ϕa
= 0 1 0 = 0 1 0 + 0 0 0 = I + ∇u ,
F =
∂Z A
0 0 1
0 0 1
0 0 0
where J = det F = 1 (zero volume change) and ∇u is the displacement gradient.
Cauchy-Green tensors are given by
1
k
0
1 + k2 k
T
T
2
k
1
C =U U = k 1+k 0
and
b=VV =
0
0
1
0
0
28
The right and left
0
0 ,
1
Z 2, z2
k/2
U
k/2
Ψ1
Ψ2
ψ1
ψ2
1
β
β
deformed
F = RU
ψα = R Ψα
Z 1, z1
undeformed
1
Figure 1: Simple shear and schematic diagram of associated right polar decomposition F = RU .
respectively. Here U denotes the right stretch tensor and V is the left stretch tensor, which can be obtained
from the right and left polar decompositions F = RU and F = VR, respectively. R is the proper orthogonal
rotation, which is often referred to in the literature as the material rotation.
To solve for R and U , let Ψ 1 , Ψ 2 , Ψ 3 be the ortho-nomalized eigenvectors, and λ21 , λ22 , λ23 the eigenvalues of
C. The eigenvalues are all real-valued and positive, because C is symmetric and positive definite. Define
the matrices
2
λ1 0
0
def
def
and
Ψ = (Ψ 1 , Ψ 2 , Ψ 3 ) ,
Λ2 = 0 λ22 0
0
0 λ23
so that Λ2 = Ψ T C Ψ is the principal axis transformation of C in {Z A }. From this, one obtains the
component matrix U = Ψ Λ Ψ T of the right stretch, and finally, R = F U −1 .
Remark 6.1. Be aware of the notation involved here. Ψ is a matrix whose columns arrange the components
∂Ψ α
of the Ψα , with α ∈ {1, 2, 3}, in the Cartesian coordinate system {Z A }. Consequently, Ψ = ∂Z
is the
A
Jacobian matrix of the transformation Z A 7→ Ψ α , in which {Ψ α } denotes the coordinate system spanned
by the eigenvectors Ψα . The basis vectors E A belonging to the Z A are thus related to the tangent map by
∂Ψ α
E A = ∂Z
△
A Ψα .
The right polar decomposition F = RU describes a stretch U of the material body in the direction of the
principal axes Ψα , followed by a rotation R (Fig. 1). Within the left polar decomposition F = VR, the body
is first rotated, and then stretched by V in the direction of the rotated principal axes ψ α = R Ψα . Since
T
V = R U RT = (R Ψ ) Λ (Ψ R) = ψΛ ψ T ,
def
where ψ = (ψ 1 , ψ 2 , ψ 3 ), the stretches U and V have the same eigenvalues λ1 , λ2 , λ3 , called the principal
stretches; hence b has the same eigenvalues as C. Like before, the three principal stretches are real-valued
and positive. Having the principal stretches, one is able to determine the Lagrangian logarithmic strain
ε = ln U and Eulerian logarithmic strain e = ln V (Definition 2.16), which play an important role in
nonlinear continuum mechanics. Recall from Proposition 2.6(iii) that both are related by
ε = RT eR
using matrix notation.
29
The particular eigenvalue problem C Ψ (α) = λ2i Ψ (α) associated with simple shear results in the characteristic
polynomial
1 − λ2
k
0
k
1 + k 2 − λ2
0
0 = det
2
0
0
1−λ
= (1 − λ2 )(1 + k 2 − λ2 )(1 − λ2 ) − (1 − λ2 )k 2 .
p
It immediately follows λ23 = λ3 = 1, that is, the eigenvector Ψ 3 is equal to the basis vector in Z 3 -direction.
In the remaining two dimensions, the characteristic polynomial reduces to (1 − λ2 )(1 + k 2 − λ2 ) − k 2 = 0
resp. λ2 + λ−2 = 2 + k 2 . Hence, the other two eigenvalues can be obtained from
s
2
2 + k2
2 + k2
2
−1,
±
(λ )1/2 =
2
2
and they are related by λ2 = λ−1
1 . This yields
λ
0
Λ = 0 λ−1
0
0
and finally
cos β
U = sin β
0
0
0
1
sin β
2
1+sin β
cos β
0
in which β(t) has been defined through
√
1−sin β
√
2
and
0
0
1
√
1+sin β
Ψ =
√
2
0
and
√
1+sin β
− √2
√
1−sin β
√
2
cos β
R = − sin β
0
0
sin β
cos β
0
0
0
,
1
0
0 ,
1
k(t) = 2 tan β(t) .
The spatial velocity gradient l = d + ω is readily available from
0 k̇ 0
l = Ḟ F −1 = 0 0 0 ,
0 0 0
so that the spatial rate of deformation and the vorticity take the form
0
k̇/2 0
0
k̇/2 0
d = k̇/2
and
ω = −k̇/2
0
0
0
0 ,
0
0
0
0
0
0
respectively. Moreover, spatial rate of rotation is given by
0
Ω = ṘR−1 = −β̇
0
β̇
0
0
0
0 .
0
Now, consider the grade-zero hypoelastic constitutive rate equation
def
σ ⋆ = λ(tr d) g ♯ + 2µ d♯ = λ(tr d) I + 2G d ,
◦
where λ, µ are the Lamé constants, σ ⋆ is a generic objective rate of Cauchy stress σ, and I is the second◦
order unit tensor. G = µ represents the shear modulus of the material. Particular choices for σ ⋆ are the
Zaremba-Jaumann stress rate and the Green-Naghdi stress rate which, in matrix notation, are calculated
from
◦
◦
σ ZJ = σ̇ − ωσ + σω
and
σ GN = σ̇ − Ωσ + σΩ ,
◦
respectively. For simple shear, analytical solutions of the Cauchy stress have been derived by Dienes [20],
Johnson and Bammann [40], and Flanagan and Taylor [25], among other, based on the grade-zero hypoelastic
30
Non-dimensional stress σ 12 /G [-]
2.5
Zaremba-Jaumann rate
Green-Naghdi rate
2
1.5
1
0.5
0
0
0.5
1
1.5
2
Strain k/2 [-]
Figure 2: Comparison of the shear stress in hypoelastic simple shear using the Zaremba-Jaumann and
Green-Naghdi stress rates.
def
def
constitutive rate equation above, and using the particular choices σ ⋆ = σ ZJ and σ ⋆ = σ GN . In case of the
Zaremba-Jaumann stress rate,
◦
σ 11 = −σ 22 = G(1 − cos k)
and
◦
◦
◦
σ 12 = G sin k ,
whereas for the Green-Naghdi stress rate2 ,
and
σ 11 = −σ 22 = 4G(cos 2β ln(cos β) + β sin 2β − sin2 β)
σ 12 = 2G cos 2β(2β − 2 tan 2β ln(cos β) − tan β) .
In both cases, σ 33 = 0 holds. The functions for the shear stress component σ 12 = σ 21 are plotted in Fig. 2.
It can be seen that shear stress increases monotonically when using the Green-Naghdi stress rate. However,
the Zaremba-Jaumann stress rate results in unphysical harmonic oscillation of the stress when applied to
hypoelasticity of grade zero.
6.2
Numerical Solution
Let k = 0.4 (β = 0.1974) and G = 5000 kN m−2 , then the formulas above yield
1.2198
0
0
0.63399 −0.77334
0
0.8198 0 ,
Λ=
Ψ = 0.77334 0.63399
0
0
1
0
0
0.98058 0.19612 0
0.98058 0.19612
U = 0.19612 1.05903 0 ,
R = −0.19612 0.98058
0
0
1
0
0
−0.03897 0.19483 0
0.03897 0.19483
ε = 0.19483 0.03897 0 ,
e = 0.19483 −0.03897
0
0
0
0
0
0
0 ,
1
0
0 ,
1
0
0 .
0
Note that tr ε = tr e = 0, that is, logarithmic strain is consistent with isochoric response in simple shear.
Moreover, using the Zaremba-Jaumann stress rate results in
σ 11 = −σ 22 = 394.7 kN m−2
and
σ 12 = 1947.1 kN m−2 ,
and for the Green-Naghdi stress rate,
2 The
σ 11 = −σ 22 = 387.2 kN m−2
and
σ 12 = 1948.9 kN m−2 .
term tan 2β in the expression for σ12 was incorrectly printed as tan2 β in [20, eq. (5.25)]; see also [40, 22].
31
Relative error [%]
5
Zaremba-Jaumann rate
Green-Naghdi rate
4
3
2
1
0
1
10
100
Number of substeps [-]
12
12
12
Figure 3: Relative shear stress error (σnum
− σexact
)/σexact
in simple shear for k = 1.0 (β = π/4) using the
Zaremba-Jaumann stress rate and Green-Naghdi stress rate.
While both stress rates are approximately equal at k = 0.4, they significantly differ at k = 1.0 (cf. Fig. 2).
For the Zaremba-Jaumann stress rate,
σ 11 = −σ 22 = 2298.5 kN m−2
and
σ 12 = 4207.4 kN m−2 ,
and for the Green-Naghdi stress rate,
σ 11 = −σ 22 = 2079.5 kN m−2
and
σ 12 = 4348.9 kN m−2 .
Two series of numerical simulations have been carried out using implementations of the objective integration
algorithm of Hughes [37] (Alg. 2) outlined in Sect. 5.3. One series employed the Zaremba-Jaumann stress
rate and the other the Green-Naghdi stress rate. Here we used the fact that, by Remark 5.9 in conjunction
with Remark 3.2, the Green-Naghdi rate reduces to the Zaremba-Jaumann rate if the current configuration is
taken as the reference configuration. Clearly, for the calculations employing the Zaremba-Jaumann rate the
total deformation gradient in each calculational cycle was set equal to the incremental deformation gradient
(Definition 5.6).
In each calculation the maximum shear strain applied was k = 1.0 (β = π/4), but the number of substeps
to reach the maximum was continuously increased respectively the size of the applied strain increments was
continuously decreased. Fig. 3 shows that the relative error between the numerically calculated stress and
the exact solutions presented above is reduced with increasing number of substeps.
7
Conclusions
We have presented basic notions of rate equations in nonlinear continuum mechanics by placing emphasis
on the geometrical background. The application of these notions to second-order tensors has led to a
clear distinction between the properties their rates may possess under different transformations: objective,
covariant, and corotational. Objectivity in constitutive theory has been formalized by the basic principle of
constitutive frame invariance, which is intended as a substitute to the classical principle of material frameindifference. We have then discussed classes of objective and corotational rate constitutive equations for large
deformation problems and their numerical integration in time. The focus has been on formulations using
the Green-Naghdi and Zaremba-Jaumann corotational stress rates as well as on two incrementally objective
integration algorithms employed by several finite element codes. Finally, simple shear of hypoelastic material
at finite deformations has been considered as an example application of both the fundamental relations and
the numerical algorithms. The analytical and numerical results presented can also be used for the verification
of future developments. Further research will be concerned with the implications of geometric continuum
mechanics and constitutive theory [69, 71, 72], particularly with respect to the integration of rate constitutive
equations in numerical simulations.
32
Acknowledgements
This work was partially supported by the German Research Foundation (DFG) Grant SA 310/21-1 and
SA 310/21-2, which is gratefully acknowledged.
A
Differential Geometry
This appendix summarizes some basic notions of differential geometry essential for the main text. Differential
geometry [1, 3, 13, 49, 84, 85, 86, 88] has been found to be the most natural way in formulating continuum
mechanics [1, 3, 44, 49, 68, 69, 70, 71, 72, 76]. The arguments are similar to those that lead to an accelerated
progress in theoretical physics in the first half of the 20th century; see references for details. In fact, some
recent derivations even seem to have wiped away long lasting debates in the field. We assume that the reader
is familiar with linear algebra and calculus in linear spaces.
The Einstein summation convention is forced in the present paper. By this convention, the sum is taken
over all possible values of a coordinate index variable whenever it appears twice, and as both a subscript and
a superscript, in a single term. For example, the local representative of aP
vector v with respect to a basis
n
{e1 , . . . , en } in an n-dimensional space is written v = v i ei instead of v = i=1 v i ei , with i ∈ {1, . . . , n}.
A.1
Manifolds
Definition A.1. A topological space is a set of elements, called points, together with a collection of subsets
that carries the information of relations or interconnections between the points. A topological space M is
referred to as a Hausdorff space, if every two points X, Y ∈ M, X 6= Y , can be separated by neighborhoods
U(X) ⊂ M and V(Y ) ⊂ M such that U ∩ V = ∅. A topological space M with metric d : S × S → R such
that 1. d(X, Y ) = 0 if and only if X = Y , 2. d(X, Y ) = d(Y, X), and 3. d(O, Y ) ≤ d(O, X) + d(X, Y ), for
X, Y, O ∈ M, is called a metric space.
♦
Definition A.2. A homeomorphism is a continuous bijective mapping ϕ : M → N between topological
spaces which has a continuous inverse and preserves the topology of M.
♦
Definition A.3. Let U(X) ⊂ M be an open neighborhood of the point X ∈ M, then the pair (U, β)
including the homeomorphism
β : M ⊃ U → X ⊂ Rn
def
X 7→ β(X) = {x1 , x2 , . . . , xn }X = {xi }X , ∃ β −1 ,
is called a chart or local coordinate system on M, where n = dim(M). The tuple {xi }X is called the
def
def S
coordinates of X in the chart (U, β). An atlas of M = i∈I⊂N Ui is a collection A(M) = {(Ui , βi )}i∈I⊂N
of a finite number of charts that covers M.
♦
Definition A.4. A chart transition or change of coordinates is a composition
β ′ ◦ β −1
β(U ∩U ′ )
: β (U ∩ U ′ ) → β ′ (U ∩ U ′ ) ,
in which (U, β), (U ′ , β ′ ) are charts on M, and U ∩ U ′ 6= ∅. An atlas is called differentiable, if for every two
charts the chart transition is differentiable.
♦
Definition A.5. A differentiable manifold is a Hausdorff space with differentiable atlas.
♦
Definition A.6. Let ϕ : M → N be continuous, U(X) ⊂ M and V(x) ⊂ N neighborhoods of X ∈ M and
x ∈ N , respectively, and let (U, β), (V, σ) be charts. Then for non-empty ϕ−1 (V) ∩ U, then the localization
of ϕ,
σ ◦ ϕ ◦ β −1 β(ϕ−1 (V)∩U ) : β ϕ−1 (V) ∩ U → σ (V ∩ ϕ(U)) ,
describes the chart transition concerning ϕ with respect to β and σ. The map ϕ is called differentiable at
X ∈ ϕ−1 (V) ∩ U, if its localization is differentiable at β(X). A bijective differentiable map ϕ is referred to
as a diffeomorphism, if both ϕ and ϕ−1 are continuous differentiable.
♦
33
Remark A.1. In this paper we simply assume that every chart transition is a diffeomorphism. If xi are
the coordinate functions of (V, σ) and X I are those of (U, β), then it would be convenient to define
def
ϕi = xi ◦ ϕ ◦ β −1
def
ϕi (X I ) = (xi ◦ ϕ ◦ β −1 )(X I ) .
resp.
△
Definition A.7. A map ϕ : M → N is called an embedding, if ϕ(M) ⊂ N is a submanifold in N and
M → ϕ(M) is a diffeomorphism.
♦
Definition A.8. Let M be an n-dimensional differentiable manifold, U ⊂ M a subset, and (U, β) a chart
with coordinate functions β(X) = {xi }X for every X ∈ U. The tangent space TX M at X is a local vector
∂
def ∂
∂
= ∂xi X . Conceptually,
space spanned by the vectors of the holonomic basis ∂x
1 , . . . , ∂xn
def
TX M = {X} × Vn .
def S
The disjoint union T M = X∈M TX M of all tangent spaces at all points of the manifold is called the
tangent bundle of M. An element (X, w) ∈ T M, called a tangent vector, will often be denoted by w X , or
just w if the base point X is clear from the context.
♦
Proposition A.1. For the previous situation, the coordinate differentials dx1 , . . . , dxn
a dual basis at X.
Proof. By dxi (X) ·
∂
∂xj (X)
=
∂xi
∂xj (X)
def
X
=
= δ ij (X).
dxi
X
form
def
∗
Definition A.9. The co-vector space dual to the tangent space TX M is called the cotangent space TX
M =
def
∗
{X} × Vn∗ , and elements of TX
M are called differential 1-forms, or just 1-forms. The union T ∗M =
S
∗
♦
X∈M TX M is referred to as the cotangent bundle of M.
A.2
Tensors and Tensor Fields
Definition A.10. A
-tensor T (X) at point X of a differentiable manifold M is a multilinear mapping
p
q
∗
∗
T (X) : TX
M × . . . × TX
M × TX M × . . . × TX M → R .
{z
}
{z
} |
|
q−fold
p−fold
The space of all pq -tensors at all points X ∈ M is denoted by Tqp (M). If N is another differentiable
manifold, a pq rs -two-point tensor over a map ϕ : M → N is a multilinear mapping
∗
∗
T (X) : Tϕ(X)
N ×. . .×Tϕ(X)
N × Tϕ(X) N ×. . .×Tϕ(X)N
{z
}
|
{z
} |
q−fold
p−fold
∗
∗
× TX
M×. . .×TX
M × TX M×. . .×TX M
{z
}
|
|
{z
r−fold
}
♦
→ R.
s−fold
∂
Definition A.11. Let
(U,
β),
where
U
⊂
M,
be
a
local
chart
on
M
such
that
∈ TX M is a local
∂xi
i
p
∗
basis at X ∈ U, and dx ∈ TX M is its dual. The components of a q -tensor in the chart (U, β) are then
defined through
∂
∂
def
i1 ...ip
ip
i1
,...,
.
T
dx , . . . , dx ,
j1 ...jq = T
∂xj1
∂xjq
Based on index placements, T is said to be contravariant of order p and covariant of order q.
Proposition A.2. Under a chart transition with Jacobian matrix
of a pq -tensor transform according to the rule
T
i′1 ...i′p
j1′ ...jq′
′
=
′
i′
∂x
∂xi
and its inverse
∂xip ∂xj1
∂xjq i1 ...ip
∂xi1
. . . ip
′ T
′ ...
j1 ...jq .
i
j
1
∂x
∂x ∂x 1
∂xjq
34
∂xi
∂xi′
♦
the components
Proof. The assertion follows by multilinearity of a tensor and the transformation properties of dxi and
∂
∂xi .
Remark A.2. From the definition of a tensor it should be clear that every 1-form a∗ = ai dxi is a 01 -tensor,
def
1
∂
∗
∗
i
and every vector v = v i ∂x
i is a
△
0 -tensor since a (v) = v(a ) = ai v ∈ R.
def
Definition A.12. The operation a∗ (v) = a∗ · v is called the (single) contraction of the tensors a∗ and v.
In a local chart X with coordinate functions {xi }X , one has
∂
∂
= ai v j δ ij = ai v i .
a∗ · v = ai dxi · v j j = ai v j dxi ·
∂x
∂xj
Dependence on the point X being understood. In general, the contraction two tensors T and S in the i-th
covariant slot of T and the j-th contravariant slot of S is defined as if the covariant slot is a 1-form and the
contravariant slot is a vector. If the slots are not specified, and T abcd and Sijkl are the components of T
and S, respectively, then the (single) contraction T · S simply means T abcd Sijkd in components. The double
contraction condenses the last two slots of T and S:
T :S,
in components,
T abcd Sijcd .
Moreover, the contraction of a 11 -tensor T is called its trace, written tr T = T ii .
♦
Definition A.13. Let T ∈ Tqp (M) and S ∈ Tsr (M) at a point X ∈ M, then the tensor product T ⊗ S is
the p+r
q+s -tensor defined by
(T ⊗ S) a∗1 , . . . , a∗p , v 1 , . . . , v q , b∗1 , . . . , b∗r , w1 , . . . , ws
def
= T a∗1 , . . . , a∗p , v 1 , . . . , v q S (b∗1 , . . . , b∗r , w 1 , . . . , w s ) ,
∗
where v 1 , . . . , v q , w1 , . . . , w s ∈ TX M and a∗1 , . . . , a∗p , b∗1 , . . . , b∗r ∈ TX
M.
Proposition A.3. A pq -tensor T has the local representative
T (X) = T
i1 ...ip
j1 ...jq (X)
♦
∂
∂
⊗ . . . ⊗ ip ⊗ dxj1 ⊗ . . . ⊗ dxjq .
∂xi1
∂x
def
and Tqp (M) = T M ⊗ . . . ⊗ T M ⊗ T ∗M ⊗ . . . ⊗ T ∗M .
|
{z
} |
{z
}
p−fold
q−fold
Proof. By Definitions A.11 and A.13.
Definition A.14. A Riemannian manifold is the pair (M, g), where M is a differentiable manifold and g
is a metric. If U ⊂ M is a subset and (U, β) a chart with coordinate functions β(X) = {xi }X for every
X ∈ U, then the metric can be locally represented by
∂
∂
def
def
i
j
g(X) = gij (X) dx ⊗ dx ,
where gij (X) =
,
≥0
∂xi ∂xj X
are the metric coefficients at every point X ∈ M.
♦
Definition A.15. Contracting the metric tensor with the inverse metric g −1 = g ij
δi j , gives the second-order identity tensor on M,
def
I M = g · g −1 = δi j dxi ⊗
∂
∂
∂xi ⊗∂xj ,
where gik g kj =
∂
∂
= dxi ⊗
.
j
∂x
∂xi
The fourth-order symmetric identity tensor or symmetrizer 1M , with components
def 1
δi k δj l + δi l δj k ,
1ij kl =
2
def
yields the symmetric part Sym (T ) = 1M : T of a second-order tensor T ∈ T11 (M).
35
♦
0
Definition A.16. Let S ∈ Tqp (M), then S ♭ ∈ Tp+q
(M) is the associated tensor with all indices lowered, and
p+q
♯
S ∈ T0 (M) is the associated tensor with all indices raised. Here ♭ is called the index lowering operator,
and ♯ is the index raising operator.
♦
q
def
def
Definition A.17. The Frobenius norm of T ∈ T11 (M) is defined through kT k =
tr(T 2 ), where T 2 =
T ♯ · T ♭ ∈ T11 (M) is the squared tensor T .
♦
Remark A.3. Tensor indices can be raised by the inverse metric coefficients, and lowered by the metric
∂
j
1
coefficients. For example, let T = T ij ∂x
i ⊗ dx ∈ T1 (M) in a given chart, then the associated tensors are
T ♭ = g · T = gik T kj dxi ⊗ dxj
T ♯ = T · g −1 = T ik g kj
and
def
∂
∂
⊗
.
∂xi ∂xj
△
def
Remark A.4. Note that g = g ♭ = (I M )♭ and g −1 = g ♯ = (I M )♯ . Moreover, the trace of T ∈ T11 (M) can
be written tr T = T ♯ : g.
△
Definition A.18. Let M, N be Riemannian manifolds and T (X) : TX M → Tϕ(X) N a general twopoint tensor over a diffeomorphism ϕ : M → N . Moreover, let U ∈ TX M and v ∈ Tx N be vectors on
M and N , respectively, with X ∈ M and x = ϕ(X) ∈ N , then the transpose of T is the linear map
T T (x) : Tx (ϕ(M)) → Tϕ−1 (x) M defined through
D
E
def
hv, T (U )ix = T T(v), U
.
X
For ϕ = Id resp. N = M the transpose of an ordinary (one-point) tensor is obtained.
♦
Proposition A.4. The components of T T are given by
(T T )i I (x) = gij (x) T jJ (ϕ−1 (x)) GIJ (ϕ−1 (x))
∂
∂
∈ Tx N , where gij (x) are the metric coefficients on N
with respect to local bases ∂X
∈ TX M and ∂x
i
I
and GIJ (X) are the inverse metric coefficients on M.
Proof. By the definitions of the transpose, metric and inverse metric; see [49, 3] for details.
Definition A.19. With T , U , and v be as before, the operations
T −1 · T (U ) = U
T −T · T T (v) = v
and
involve the inverse T −1 (X) and the inverse transpose T −T (x). Moreover, a two-point tensor T (X) : TX M →
Tϕ(X) N is called orthogonal provided that T T · T = I M and T · T T = I N . If T is orthogonal and the
determinant det T = +1, then T is called proper orthogonal.
♦
In what follows, M and N are differentiable manifolds, ϕ : M → N is a diffeomorphism, (U, β) and (V, σ)
def
are charts of U ⊂ M and V ⊂ N , respectively, and ϕi (X I ) = (xi ◦ ϕ ◦ β −1 )(X I ) are the coordinates xi on
N arising from the coordinates X I on U via localization of ϕ.
Definition A.20. The tangent bundle of M has been denoted T M. In in a more rigorous definition, it is
the triplet (T M, τM , M) including the projection τM : T M → M. At X ∈ M, with dim(M) = n, the
−1
tangent space τM
(X) = TX M = {X} × Vn is called fibre over X, and Vn is the fibre space. If (T N , τN , N )
′
is another tangent bundle, then a continuous map ϕ : M → N induces the bundle (ϕ⋆ T N , τN
, M) with
′
⋆
⋆
τN : ϕ T N → M. The restriction of ϕ T N to x = ϕ(X) ∈ N is the tangent space Tϕ(X) N .
♦
Definition A.21. A vector field v on M is identified with the tangent bundle section
v : M → TM,
with τM (v(X)) = X, ∀X ∈ M. A 1-form field is a section of the cotangent bundle: a∗ : M → T ∗M. The
sets of all sections of T M and T ∗M are denoted by Γ (T M) and Γ (T ∗M), respectively. As well, if some
manifold N has the tangent bundle T N , u ∈ Γ (T N ) is a vector field, and ϕ : M → N is continuous, then the
def
related vector field over ϕ is the induced section ϕ⋆ u : M → T N defined through (ϕ⋆ u)(X) = u(ϕ(X)).
36
♦
Definition A.22. The local basis sections of T M restricted to U,
∂
∂
: U → T M|U
,
.
.
.
,
∂x1
∂xn
define a local basis for all X ∈ M, and dx1 , . . . , dxn : U → T ∗M|U are their duals. Hence, for every
−1
∂
∗
i
fibre τM
(X) at X ∈ U, v(X) = v i (X) ∂x
i (X) and a (X) = ai (X) dx (X), respectively. For the fields to
i
j i
i
be continuously differentiable, the mappings x → v (x ) and x → aj (xi ) on β(U) ⊂ Rn are required to be
continuously differentiable.
♦
Remark A.5. One may construct tensor fields of any order by fibrewise
tensor-multiplication of vector and
1-form fields. For example, if w ∈ Γ (T M) and b∗ ∈ Γ (T ∗M), a 11 -tensor field T would be
def
T = (w ⊗ b∗ ) ∈ Γ (T M ⊗ T ∗M) ,
def
and (w⊗b∗ )(X) = w(X)⊗b∗ (X). Thus T is a section of the 11 -tensor bundle T11 (M) = T M⊗T ∗M → M.
Two-point tensor fields over maps ϕ : M → N are defined analogously by taking into account the sections
induced by ϕ.
△
p
p
p
Definition A.23. The q -tensor bundle of M is denoted by (Tq (M), τM , M), or just Tq (M), and the
def
set of all sections of it is denoted by Tpq (M) = Γ (Tqp (M)).
♦
Definition A.24. The index lowering and index raising operators for tensors carry over to tensor fields by
defining the so-called musical isomorphisms ♭ : T M → T ∗M and ♯ : T ∗M → T M, respectively.
♦
A.3
Pushforward and Pullback
Definition A.25. The tangent map and cotangent map over ϕ are defined through
T ∗ϕ : T ∗N → T ∗M
Tϕ : TM → TN
∂
∂ϕi ∂
→
7
∂X I
∂X I ∂xi
dxi 7→
∂ϕi
dX I ,
∂X I
respectively.
♦
∂
Remark A.6. If V = V I ∂X
I ∈ TX M is a vector at X ∈ M, then
∂
∂ϕi I
def
V
(X) i (ϕ(X))
T ϕ (V ) (X) =
∂X I
∂x
∈ Tϕ(X) N ,
i
∂ϕ
∂
that is, T ϕ(V ) = V I ∂X
I ∂xi without indicating the base point. As the tangent map is linear, a two-point
∗
tensor F (X) ∈ Tϕ(X) N ⊗ TX
M can be defined yielding the same result:
i
∂
∂ϕi ∂
∂ def
I
J
I ∂ϕ
=V
⊗ dX
= F ·V .
T ϕ (V ) = V
·
∂X I ∂xi
∂X J ∂xi
∂X I
|
{z
}
def
=F
def
Hence, we may write T ϕ ◦ V = T ϕ · V . Moreover, if a∗ ∈ T ∗N , then T ∗ϕ(a∗ ) = a∗ · F .
△
Remark A.7. If V : M → T M is a vector field and not just a vector emanating from a specific base point,
then T ϕ(V ) is a vector field over ϕ. T ϕ(V ) becomes an honest vector field on N when the base points are
switched.
△
Definition A.26. The pushforward by ϕ of a vector field V ∈ Γ (T M) is the vector field ϕ ⇑ V ∈ Γ (T N )
defined by
def
ϕ ⇑ V = (T ϕ · V ) ◦ ϕ−1 .
def
The pullback by ϕ of a vector field w ∈ Γ (T N ) is a vector field on M defined through ϕ ⇓ w = (T (ϕ−1 ) ◦
w) ◦ ϕ ∈ Γ (T M).
♦
37
Remark A.8. The definition carries over to real functions resp. scalar fields. For example, let f : M → R,
def
then the pushforward ϕ ⇑ f = f ◦ϕ−1 is a scalar field on N , and ϕ ⇑= ϕ−1 ⇓ defines the pullback of a scalar
field as the inverse operation. Note that f has the same values at X ∈ M as ϕ ⇑ f has at x = ϕ(X) ∈ N . △
Proposition A.5. For a composition of maps ϕ and ψ,
(ψ ◦ ϕ) ⇓ = ϕ ⇓ ◦ ψ ⇓
(ψ ◦ ϕ) ⇑ = ψ ⇑ ◦ ϕ ⇑ .
and
Proof. By the chain rule; see, for example, [1, 49].
∗
∗
∗
∗
Definition A.27. The pullback and pushforward of 1-form fields a ∈ Γ (T N ) and B ∈ Γ (T M) on N
and M, respectively, are being defined according to their action on vector fields. Clearly,
def
ϕ ⇓ a∗ = (a∗ ◦ ϕ) · T ϕ
def
ϕ ⇑ B ∗ = (B ∗ ◦ ϕ−1 ) · T (ϕ−1 ) ,
and
respectively.
♦
Definition A.28. Let T ∈ Tpq (M) and t ∈ Tpq (N ), then
and
def
(ϕ ⇑ T )(x) a∗1 , . . . , a∗p , w 1 , . . . , w q = T (X) (ϕ ⇓ a∗1 ), . . . , (ϕ ⇓ a∗p ), (ϕ ⇓ w 1 ), . . . , (ϕ ⇓ w q ) ,
def
(ϕ ⇓ t)(X) B ∗1 , . . . , B ∗p , V 1 , . . . , V q = t(x) (ϕ ⇑ B ∗1 ), . . . , (ϕ ⇑ B ∗p ), (ϕ ⇑ V 1 ), . . . , (ϕ ⇑ V q ) ,
where X ∈ M and x = ϕ(X).
♦
Remark A.9. In general, pushforward and pullback do not commute with index raising and lowering,
e.g. ϕ ⇑ (T ♭ ) 6= (ϕ ⇑ T )♭ .
△
Definition A.29. An diffeomorphism ϕ : M → N between Riemannian manifolds (M, G) and (N , g) is
called an isometry if
g = ϕ⇑G ,
G(U , V ) = g(ϕ ⇑ U , ϕ ⇑ V ) ,
or equivalently,
for U , V ∈ Γ (T M).
♦
Proposition A.6. The tangent map of an isometry ϕ : M → N is orthogonal, and proper orthogonal, with
det(T ϕ) = +1, if ϕ is also orientation-preserving.
Proof. Let G be the metric on M, and g the metric on N . Then, by Definition A.26 and the definition of
an isometry,
hU , V iX = hT ϕ · U , T ϕ · V ix=ϕ(X)
for every X ∈ M. On the other hand, Definition A.18 of the transpose yields
hT ϕ · U , T ϕ · V ix = h(T ϕ)T · (T ϕ · U ), V iX .
Comparison of both equations shows that
(T ϕ)T · T ϕ = I M ,
(T ϕ)−1 = (T ϕ)T
that is,
at every X ∈ M .
Proofing the second assertion requires the notion of orientation, which is briefly introduced below.
A.4
Tensor Analysis
Definition A.30. Let v, w ∈ Γ (T N ) be vector fields on a Riemannian manifold N , and v continuously
differentiable.
In a chart (V, σ) on N with coordinates xi , the covariant derivative or gradient of v is the
proper 11 -tensor field defined through
∂ def
=
∇v(x) = ∇j v (x) dx ⊗
∂xi
i
j
38
∂
∂v i
k
i
+ v γk j (x) dxj ⊗
,
∂xj
∂xi
and the covariant derivative of v along w is the proper vector field defined through
i
∂v j
∂
def
def
k j
i
∇w v(x) = w(x) · ∇v(x) =
w
+
v
w
γ
.
k j (x)
∂xj
∂xi
If a∗ ∈ Γ (T ∗N ) is continuously differentiable, then ∇a∗ ∈ T02 (N ), defined by
∂ai
def
k
i
j
−
a
γ
∇a∗ (x) =
k
i j (x) dx ⊗ dx ,
∂xj
is a proper 11 -tensor field. The term “proper” is meant in the sense that the components of the covariant
derivative transform under chart transitions according to the tensorial transformation rule (Proposition A.2).
′
In particular, under a chart transition such that xi 7→ xi , the connection coefficients transform according to
′
′
γk ji =
′
′
∂xk ∂xj ∂xi
∂xj ∂ 2 xm
γk′ j i′ + m′ k i ,
′
k
j
i
∂x ∂x ∂x
∂x ∂x ∂x
with i, i′ , j, j ′ , k, k ′ , m, m′ ∈ {1, . . . , ndim }.
♦
Definition A.31. The operator ∇ : Γ (T N )×Γ (T N ) → Γ (T N ) introduced in Definition A.30 is referred to
as the connection on N , and γk ji are the connection coefficients. A connection ∇ is called torsion-free if
γk ji = γi jk . In case of a Riemannian manifold the connection coefficients are called Christoffel symbols of
the second kind.
♦
Theorem A.1. Let N be a Riemannian manifold, and gij and g ij be the coefficients of the metric and
inverse metric, respectively, then there is a unique and torsion-free connection whose coefficients are given
by
∂gil
∂gij
∂gjl
def 1
.
+
−
γi kj = g kl
2
∂xi
∂xj
∂xl
Proof. Detailed derivations can be found, for example, in [13, 49, 85].
Proposition A.7. Let N be a Riemannian manifold with connection ∇ and metric g, then
∇g = 0 .
def
Proof. Using the previous definitions, Proposition A.1, and g = gij dxi ⊗ dxj ,
∂gjk
∂gjk
∂gkj
∂gki
∂gij
l
∇i gjk =
= 0.
−
2g
γ
=
−
+
−
jl k i
∂xi
∂xi
∂xk
∂xi
∂xj
Proposition A.8. On a Riemannian manifold N , ∇(u♭ ) = (∇u)♭ for any u ∈ Γ (T N ).
Proof. This follows from a straightforward calculation by using Proposition A.7.
Definition A.32. The divergence of a tensor field t ∈ Tpq (N ) is defined as the contraction of its covariant
derivative ∇t on the last contravariant leg. For examples, if t is a 30 -tensor field, then
ij def
(div t)
= ∇k tijk .
♦
Proposition A.9. Let N be a Riemannian manifold with metric coefficients gij in a positively oriented
chart and v ∈ Γ (T N ) a vector field, then
1
∂ p
i
.
|
det
g
|
v
div v = ∇i v i = p
kl
| det gkl | ∂xi
Proof. By Definition A.30 and Proposition A.1; see [3] for details.
Definition A.33. The evolution in time of a differentiable manifold N is described by a mapping ψt,s :
N → N , where t, s are points in a time interval I ⊂ R. The mapping ψt,s is called a time-dependent flow
on N provided that
ψt,s ◦ ψs,r = ψt,r
and
ψt,t = idN .
♦
39
u
v(ψt,s (Ps ))
Luv ∆t
P (t) = ψt,s (Ps )
⋆
(ψt,s
v)(Ps )
⋆
v)(Ps )
v(Ps ) = (ψs,s
Ps = ψs,s (Ps )
Figure 4: Lie derivative of a time-independent vector field v along a time-dependent vector field u; reprint
from [3, fig. 3.5]
Remark A.10. If Xs = ψs,s (Xs ) ∈ N is the starting point at starting time t = s, then X = c(t) =
ψt,s (Xs ) = ψ(Xs , s, t) is the point at t = t, for s, t fixed. Hence, c : I → N is a curve on N with the initial
condition c(s) = Xs . The flow ψt,s is closely connected to a time-dependent vector field u : N × I → T N
through u(ψt,s (Xs ), t) = ċ(t), with c(s) = Xs . Conversely, c(t) is the unique integral curve of u starting at
Xs at time t = s; thus u generates the flow, so ψt,s needs not to be given explicitly.
△
Definition A.34. The Lie derivative of a time-dependent tensor field Tt ∈ Tpq (N ) along a time-dependent
vector field ut ∈ Γ (T N ) is defined by
def
Lu T = lim
∆t→0
ψt,s ⇓ T t − T s
d
ψt,s ⇓ T t − ψs,s ⇓ T s
= lim
=
ψt,s ⇓ T t
∆t→0
∆t
∆t
dt
,
t=s
def
where ∆t = t − s, and ψt,s ⇓ denotes the pullback concerning the flow ψt,s associated with ut . Therefore,
the Lie derivative approximately answers the question how a tensor field T changes under some flow. The
so-called autonomous Lie derivative is obtained by holding t fixed in Tt , that is,
def
⋆
⋆
ψt,s
T s − ψs,s
Ts
d ⋆
=
ψ Ts
∆t→0
∆t
dt t,s
£u T = lim
t=s
= Lu T −
∂T
.
∂t
If T is time-independent, Lu T ≡ £u T . Fig. 4 illustrates the concept.
♦
Proposition A.10. The Lie derivative of the tensor field Tt along the vector field ut is obtained by pulling
back the tensor field according to the flow associated with ut at some starting time, performing the common
time derivative, and then pushing forward the result using the inverse of the pullback. Clearly,
def
Lu T = ψt,s ⇑
d
(ψt,s ⇓ Tt ) .
dt
Proof. We refer to [3] and [1, sect. 5.4] for a detailed discussion.
Proposition A.11. Let ϕ : M → N be a diffeomorphism, ut , v t ∈ Γ (T N ), and T ∈ Tpq (N ), then
£u+v = £u + £v
and
ϕ ⇓ (£u T ) = £(ϕ⇓u) (ϕ ⇓ T ) .
Proof. See, for example, [49].
Proposition A.12. Let N be a Riemannian manifold, and ut ∈ Γ (T N ) be a time-dependent vector field.
In a chart (V, σ) on N with coordinates xi , the components of the Lie derivative of a time-dependent tensor
field Tt ∈ T11 (M) along ut are computed from
(Lu T )ij =
∂T ij
∂T ij
∂ui
∂uk
+ uk k − T kj k + T ik j .
∂t
∂x
∂x
∂x
40
If N has a torsion-free connection ∇, then
(Lu T )ij =
∂T ij
+ uk ∇k T ij − ∇k T kj ui + ∇j T ik uk .
∂t
Proof. We refer again to [49] for detailed proof.
Remark A.11. As pushforward and pullback do not commute with index raising and lowering, the Lie
derivative also does not commute with these operations in general, that is, for example, Lu (T ♭ ) 6= (Lu T )♭ .△
In this section, we let M and N be Riemannian manifolds with some orientation, ϕ : M → N be an
orientation-preserving diffeomorphism, (U, β) be a positively oriented chart of U ⊂ M with respect to the
orientation of M, and (V, σ) be a positively oriented chart of V ⊂ N , with non-empty ϕ−1 (V) ∩ U ⊂ M.
def
Furthermore, let ϕi (X I ) = (xi ◦ ϕ ◦ β −1 )(X I ) be the coordinates xi on N arising from the coordinates X I
on U via localization of ϕ.
Remark A.12. A precise definition of orientation requires exterior calculus, which is probably one of the
most exotic fields of modern differential geometry, at least from an engineer’s point of view. The interested
reader is referred to the literature suggested at the beginning of this appendix.
△
Definition A.35. Let the tuple (w1 , . . . , w n ) of vector fields w1 , . . . , wn ∈ Γ (T N ) be positively oriented
with respect to the orientation of N , then the volume density dv on N is defined through
q
def
dv(w1 , . . . , wn ) = dethwi , w j i ,
where dethw i , wj i is the determinant of the matrix (Wij ) whose elements are given by the inner products
def
Wij = hw i , wj i.
♦
Remark A.13. Note that dv(e1 , . . . , en ) = 1 for a positively oriented ortho-normalized basis {e1 , . . . , en }
in T N . In ordinary R3 , the volume of the parallelepiped spanned by the three vectors w1 , w2 , w3 ∈ R3 is
def p
given by V (w 1 , w 2 , w3 ) = dethw i , wj i provided that w1 , w 2 , and w3 are positively oriented.
△
Proposition A.13. Let dV and dv be the volume densities on M and N , respectively, then
ϕ ⇓ dv = dv ◦ ϕ = Jϕ dV ,
where
Jϕ (X) = det
using local coordinates.
∂ϕi
∂X I
p
det gij (ϕ(X))
p
det GIJ (X)
Proof. The proof is most easily obtained using local representatives of dV and dv; cf. [3, 49].
Definition A.36. The proper scalar field Jϕ : M → R introduced by Proposition A.13 is called the Jacobian
of ϕ or relative volume change with respect to dV and dv. Since ϕ was assumed orientation-preserving,
Jϕ > 0.
♦
Proposition A.14. (See [1, 49] for a proof.) £u dv = (div u) dv.
Theorem A.1 (Change of Variables). Let f : ϕ(M) → R, then
Z
Z
f dv .
ϕ ⇓ (f dv) =
M
ϕ(M)
Proof. This is well-known from the analysis of real functions.
The following relations, including the divergence theorem, play a fundamental role in both differential geometry and continuum mechanics. A full derivation is beyond the scope of this paper and can be found
elsewhere, e.g. [1, 3, 49].
41
Definition A.37. Let N be an oriented n-dimensional manifold with compatible oriented boundary ∂N
def
such that the normals to ∂N , n∗ ∈ Γ (T ∗N ), point outwards. The area density da = dv ∂N is the volume
density on (n − 1)-dimensional ∂N induced by the volume density dv on N . Conceptually, we write
dv = n∗ ∧ da to emphasize that dv and da are linked by the outward normals.
♦
Theorem A.2 (Divergence Theorem). Let w ∈ Γ (T N ) be a vector field, then for the situation defined
above,
Z
Z
w · n∗ da .
(div w) dv =
N
∂N
Proposition A.15. Let M be oriented and ϕ : M → N an orientation preserving diffeomorphism with
tangent F = T ϕ. Let dA and da be the volume forms on ∂M and ∂(ϕ(M)), respectively. Then da =
dA ◦ ϕ−1 if and only if the outward normals on ∂(ϕ(M)) and ∂M are related by
n∗ = ϕ⋆ (Jϕ N ∗ ) = (Jϕ ◦ ϕ−1 )F −T · (N ∗ ◦ ϕ−1 ) ,
where n∗ ∈ Γ (T ∗N ), N ∗ ∈ Γ (T ∗M), and Jϕ is the Jacobian of ϕ.
Definition A.38. Let M, ϕ, etc., be as before, then the Piola transform of a spatial vector field y ∈ Γ (T N )
is the vector field on M given by
def
Y = Jϕ ϕ ⇓ y = Jϕ F −1 · (y ◦ ϕ)
∈ Γ (T M) .
♦
Theorem A.3 (Piola Identity). If Y is the Piola transform of y, then the divergence operators DIV on
M and div on N are related by
DIV Y = (div y ◦ ϕ) Jϕ .
Proposition A.16.
Z
∂ϕ(M)
y · n∗ da =
Z
∂M
Y · N∗ dA
resp.
(y · n∗ da) ◦ ϕ = Y · N∗ dA .
References
[1] R. Abraham, J. E. Marsden, and T. Ratiu. Manifolds, Tensor Analysis, and Applications. AddisonWesley, Reading, Massachusetts, 1983.
[2] S. N. Atluri. On constitutive relations at finite strain: Hypo-elasticity and elasto-plasticity with isotropic
or kinematic hardening. Computer Methods in Applied Mechanics and Engineering, 43:137–171, 1984.
[3] D. Aubram. Differential Geometry Applied to Continuum Mechanics.
Number 44 in Veröffentlichungen des Grundbauinstitutes der Technischen Universität Berlin. Shaker Verlag, Aachen, 2009.
http://dx.doi.org/10.14279/depositonce-2185.
[4] D. Aubram. An Arbitrary Lagrangian-Eulerian Method for Penetration into Sand at Finite Deformation.
Number 62 in Veröffentlichungen des Grundbauinstitutes der Technischen Universität Berlin. Shaker
Verlag, Aachen, 2013. http://dx.doi.org/10.14279/depositonce-3958.
[5] D. Aubram. Development and experimental validation of an arbitrary Lagrangian-Eulerian (ALE)
method for soil mechanics. Geotechnik, 38(3):193–204, 2015.
[6] D. Aubram, F. Rackwitz, and S. A. Savidis.
Contribution to the non-Lagrangian formulation of geotechnical and geomechanical processes. In Th. Triantafyllidis, editor, Holistic Simulation of Geotechnical Installation Processes: Theoretical Results and Applications, volume 82 of
Lecture Notes in Applied and Computational Mechanics. Springer International Publishing, 2017.
http://dx.doi.org/10.1007/978-3-319-52590-7_3.
[7] D. Aubram, F. Rackwitz, P. Wriggers, and S. A. Savidis. An ALE method for penetration into
sand utilizing optimization-based mesh motion. Computers and Geotechnics, 65:241–249, 2015.
http://dx.doi.org/10.1016/j.compgeo.2014.12.012.
[8] E. Bauer. Calibration of a comprehensive constitutive equation for granular materials. Soils and
Foundations, 36(1):13–26, 1996.
42
[9] D. J. Benson. An efficient, accurate, simple ALE method for nonlinear finite element programs. Computer Methods in Applied Mechanics and Engineering, 72(3):305–350, 1989.
[10] D. J. Benson. Computational methods in Lagrangian and Eulerian hydrocodes. Computer Methods in
Applied Mechanics and Engineering, 99(2-3):235–394, 1992.
[11] B. Bernstein. Hypo-elasticity and elasticity. Archive for Rational Mechanics and Analysis, 6:89–104,
1960.
[12] A. Bertram and B. Svendsen. On material objectivity and reduced constitutive equations. Archives of
Mechanics, 53(6):653–675, 2001.
[13] R. L. Bishop and S. I. Goldberg. Tensor Analysis on Manifolds. The Macmillan Company, New York,
1968.
[14] O. T. Bruhns, A. Meyers, and H. Xiao. On non-corotational rates of Oldroyd’s type and relevant issues
in rate constitutive formulations. Proceedings of the Royal Society of London. Series A, 460:909–928,
2004.
[15] W. F. Chen and D. J. Han. Plasticity for Structural Engineers. Springer-Verlag New York, Inc., 1988.
[16] Y. M. Cheng and Y. Tsui. Limitations to the large strain theory. International Journal for Numerical
Methods in Engineering, 33:101–114, 1992.
[17] B. D. Coleman and M. E. Gurtin. Thermodynamics with internal state variables. The Journal of
Chemical Physics, 47(2):597–613, 1967.
[18] B. D. Coleman and V. J. Mizel. Existence of caloric equations of state in thermodynamics. The Journal
of Chemical Physics, 40(4):1116–1125, 1964.
[19] B. D. Coleman and W. Noll. The thermodynamics of elastic materials with heat conduction and
viscosity. Archive for Rational Mechanics and Analysis, 13:167–178, 1963.
[20] J. K. Dienes. On the analysis of rotation and stress rate in deforming bodies. Acta Mechanica, 32:217–
232, 1979.
[21] J. K. Dienes. A discussion of material rotation and stress rate. Acta Mechanica, 65:1–11, 1986.
[22] J. K. Dienes. Finite deformation of materials with an ensemble of defects. Report LA-13994-MS, Los
Alamos National Laboratory, Los Alamos, USA, 2003.
[23] A. C. Eringen. Mechanics of Continua. Robert E. Krieger Publishing Company, New York, 2nd edition,
1980.
[24] C. Fang, Y. Wang, and K. Hutter. Shearing flows of a dry granular material – hypoplastic constitutive
theory and numerical simulations. International Journal for Numerical and Analytical Methods in
Geomechanics, 30:1409–1437, 2006.
[25] D. P. Flanagan and L. M. Taylor. An accurate numerical algorithm for stress integration with finite
rotations. Computer Methods in Applied Mechanics and Engineering, 62:305–320, 1987.
[26] L. P. Franca. An algorithm to compute the square root of a 3 × 3 positive definite matrix. Computers
and Mathematics with Applications, 18(5):459–466, 1989.
[27] A. E. Green and P. M. Naghdi. A general theory of an elastic-plastic continuum. Archive for Rational
Mechanics and Analysis, 18:251–281, 1965.
[28] G. Gudehus. A comprehensive constitutive equation for granular materials. Soils and Foundations,
36(1):1–12, 1996.
[29] G. Gudehus and D. Kolymbas. A constitutive law of the rate type for soils. In W. Wittke, editor,
Proceedings Third International Conference on Numerical Methods in Geomechanics, Aachen, Germany,
pages 319–329. A. A. Balkema, 1979.
[30] Z.-H. Guo. Time derivatives of tensor fields in non-linear continuum mechanics. Archiwum Mechaniki
Stosowanej, 15(1):131–161, 1963.
[31] M. E. Gurtin and K. Spear. On the relationship between the logarithmic strain rate and the stretching
tensor. International Journal of Solids and Structures, 19(5):437–444, 1983.
43
[32] P. Haupt. Continuum Mechanics and Theory of Materials. Springer-Verlag Berlin Heidelberg, 2nd
edition, 2002.
[33] C. W. Hirt, A. A. Amsden, and J. L. Cook. An arbitrary Lagrangian-Eulerian computing method for
all flow speeds. Journal of Computational Physics, 14:227–253, 1974.
[34] A. Hoger. The material time derivative of logarithmic strain. International Journal of Solids and
Structures, 22(9):1019–1032, 1986.
[35] A. Hoger and D. E. Carlson. Determination of the stretch and rotation in the polar decomposition of
the deformation gradient. Quarterly of Applied Mathematics, 42:113–117, 1984.
[36] G. A. Holzapfel. Nonlinear Solid Mechanics - A Continuum Approach for Engineering. John Wiley &
Sons, Ltd., 2000.
[37] T. J. R. Hughes. Numerical implementation of constitutive models: Rate-independent deviatoric plasticity. In S. Nemat-Nasser, R. J. Asaro, and G. A. Hegemier, editors, Theoretical Foundation for
Large-Scale Computations for Nonlinear Material Behavior, pages 29–63. Martinus Nijhoff Publishers,
Dordrecht, Niederlande, 1984.
[38] T. J. R. Hughes and J. Winget. Finite rotation effects in numerical integration of rate constitutive equations arising in large-deformation analysis. International Journal for Numerical Methods in Engineering,
15(12):1862–1867, 1980.
[39] G. Jaumann. Geschlossenes System physikalischer und chemischer Differentialgesetze. Sitzungsberichte
der Mathematisch-Naturwissenschaftlichen Classe der Kaiserlichen Akademie der Wissenschaften 2a,
120:385–530, 1911.
[40] G. C. Johnson and D. J. Bammann. A discussion of stress rates in finite deformation problems. International Journal of Solids and Structures, 20(8):725–737, 1984.
[41] D. Kolymbas. Ein nichtlineares viskoplastisches Stoffgesetz für Böden, volume 77 of Veröffentlichungen
des Institutes für Bodenmechanik und Felsmechanik der Universität Fridericiana in Karlsruhe. 1978.
(Institute Series; in German).
[42] D. Kolymbas. Eine konstitutive Theorie für Böden und andere körnige Stoffe, volume 109 of Veröffentlichungen des Institutes für Bodenmechanik und Felsmechanik der Universität Fridericiana in Karlsruhe. 1988. (Institute Series; in German).
[43] D. Kolymbas. An outline of hypoplasticity. Archive of Applied Mechanics, 61(3):143–151, 1991.
[44] K. Kondo, editor. RAAG Memoirs of the Unifying Study of Basic Problems in Engineering and Physical
Sciences by Means of Geometry. Gakujutsu Bunken Fukyu-Kai, Tokyo, 1955–1968. (4 vols.).
[45] R. D. Krieg and S. W. Key. Implementation of a time independent plasticity theory into structural computer programs, volume 20 of Constitutive Equations in Viscoplasticity: Computational and Engineering
Aspects, pages 125–137. Applied Mechanics Division, ASME, New York, 1976.
[46] R. D. Krieg and D. N. Krieg. Accuracy of numerical solution methods for the elastic, perfectly plastic
model. Journal of Pressure Vessel Technology, 99:510–515, 1977.
[47] E. H. Lee, R. L. Mallet, and T. B. Wertheimer. Stress analysis for anisotropic hardening in finitedeformation plasticity. Journal of Applied Mechanics, 50:554–560, 1983.
[48] L. E. Malvern. Introduction to the Mechanics of a Continuous Medium. Prentice Hall, Inc., New Jersey,
1969.
[49] J. E. Marsden and T. J. R. Hughes. Mathematical Foundations of Elasticity. Dover Publications, New
York, 1994. (originally published: Prentice-Hall, 1983).
[50] E. F. Masur. On the definition of stress rate. Quarterly of Applied Mathematics, 19:160–163, 1961.
[51] E. F. Masur. On tensor rates in continuum mechanics. Zeitschrift für Angewandte Mathematik und
Physik (ZAMP), 16(2):191–201, 1965.
[52] P. M. Naghdi and W. L. Wainwright. On the time derivatives of tensors in mechanics of continua.
Quarterly of Applied Mathematics, 19(2):95–109, 1961.
44
[53] J. C. Nagtegaal and F. E. Veldpaus. On the implementation of finite strain plasticity equations in a
numerical model. In J. F. T. Pittman, O. C. Zienkiewicz, R. D. Wood, and J. M. Alexander, editors,
Numerical Analysis of Forming Processes. John Wiley & Sons, Ltd., 1984. chapter 12.
[54] G. C. Nayak and O. C. Zienkiewicz. Elasto-plastic stress analysis. a generalization for various constitutive relations including strain softening. International Journal for Numerical Methods in Engineering,
5:113–135, 1972.
[55] S. Nemat-Nasser. On finite plastic flow of crystalline solids and geomaterials. Journal of Applied
Mechanics, 50:1114–1126, 1983.
[56] A. Niemunis and I. Herle. Hypoplastic model for cohesionless soils with elastic strain range. Mechanics
of Cohesive-Frictional Materials, 2:279–299, 1997.
[57] W. Noll. A mathematical theory of the mechanical behavior of continuous media. Archive for Rational
Mechanics and Analysis, 2:197–226, 1958.
[58] W. Noll. Materially uniform simple bodies with inhomogeneities. Archive for Rational Mechanics and
Analysis, 27(1):1–32, 1967.
[59] W. Noll. A new mathematical theory of simple materials. Archive for Rational Mechanics and Analysis,
48(1):1–50, 1972.
[60] W. Noll. Lectures on the foundations of continuum mechanics and thermodynamics. Archive for Rational
Mechanics and Analysis, 52(1):62–92, 1973.
[61] W. Noll.
Five Contributions to Natural
http://www.math.cmu.edu/~wn0g/noll/, 2004.
Philosophy,
On
Material
Frame-Indifference.
[62] J. G. Oldroyd. On the formulation of rheological equations of state. Proceedings of the Royal Society
of London. Series A, 200:523–541, 1950.
[63] P. M. Pinsky, M. Ortiz, and K. S. Pister. Numerical integration of rate constitutive equations in finite
deformation analysis. Computer Methods in Applied Mechanics and Engineering, 40:137–158, 1983.
[64] W. Prager. An elementary discussion of definitions of stress rate. Quarterly of Applied Mathematics,
18(4):403–407, 1961.
[65] L. Radhakrishna, L. N. Katkar, and T. H. Date. Jaumann transport in relativistic continuum mechanics.
General Relativity and Gravitation, 13(10):939–946, 1981.
[66] M. M. Rashid. Incremental kinematics for finite element applications. International Journal for Numerical Methods in Engineering, 36:3937–3956, 1993.
[67] G. Romano and R. Baretta. Continuum mechanics on manifolds. Lecture notes, Department of Structural Engineering, University of Naples Federico II, Italy, 2009.
[68] G. Romano and R. Baretta. Covariant hypo-elasticity. European Journal of Mechanics A/Solids,
30:1012–1023, 2011.
[69] G. Romano and R. Baretta. Geometric constitutive theory and frame invariance. International Journal
of Non-Linear Mechanics, 51:75–86, 2013.
[70] G. Romano and R. Baretta. Continuum mechanics on manifolds. Lecture notes, Department of Structural Engineering, University of Naples Federico II, Italy, 2014. http://wpage.unina.it/romano/.
[71] G. Romano, R. Baretta, and M. Diaco. Geometric continuum mechanics. Meccanica, 49:111–133, 2014.
[72] G. Romano, R. Baretta, and M. Diaco.
A geometric rationale for invariance, covariance and constitutive relations.
Continuum Mechanics and Thermodynamics, 2017.
https://doi.org/10.1007/s00161-017-0595-5.
[73] R. Rubinstein and S. N. Atluri. Objectivity of incremental constitutive relations over finite time steps in
computational finite deformation analysis. Computer Methods in Applied Mechanics and Engineering,
36:277–290, 1983.
[74] R. Saunders. The Geometry of Jet Bundles, volume 142 of London Mathematical Society Lecture Note
Series. Cambridge University Press, Cambridge, UK, 1989.
45
[75] H. L. Schreyer, R. F. Kulak, and J. M. Kramer. Accurate numerical solutions for elastic-plastic models.
Journal of Pressure Vessel Technology, 101:226–234, 1979.
[76] R. Segev. Notes on metric independent analysis of classical fields. Mathematical Methods in the Applied
Sciences, 36(5):497–566, 2013.
[77] J. C. Simo and T. J. R. Hughes. Computational Inelasticity. Springer-Verlag Berlin, 1998.
[78] J. C. Simo and J. E. Marsden. On the rotated stress tensor and the material version of the DoyleEricksen formula. Archive for Rational Mechanics and Analysis, 86(3):213–231, 1984.
[79] J. C. Simo, J. E. Marsden, and P. S. Krishnaprasad. The Hamiltonian structure of nonlinear elasticity:
The convective representation of solids, rods, and plates. Archive for Rational Mechanics and Analysis,
104(2):125–183, 1988.
[80] J. C. Simo and K. S. Pister. Remarks on rate constitutive equations for finite deformation problems:
Computational implications. Computer Methods in Applied Mechanics and Engineering, 46:201–215,
1984.
[81] J. C. Simo and R. L. Taylor. Consistent tangent operators for rate-independent elastoplasticity. Computer Methods in Applied Mechanics and Engineering, 48:101–118, 1985.
[82] S. W. Sloan. Substepping schemes for the numerical integration of elastoplastic stress-strain relations.
International Journal for Numerical Methods in Engineering, 24:839–911, 1987.
[83] S. W. Sloan, A. J. Abbo, and D. Sheng. Refined explicit integration of elastoplastic models with
automatic error control. Engineering Computations, 18(1/2):121–154, 2001.
[84] M. Spivak. Calculus on Manifolds. Benjamin, New York, 1965.
[85] M. Spivak. A Comprehensive Introduction to Differential Geometry, volume 2. Publish or Perish, Inc.,
Houston, Texas, 2nd edition, 1979.
[86] M. Spivak. A Comprehensive Introduction to Differential Geometry, volume 1. Publish or Perish, Inc.,
Houston, Texas, 3rd edition, 1999.
[87] B. Svendsen and A. Bertram. On frame-indifference and form-invariance in constitutive theory. Acta
Mechanica, 132:195–207, 1999.
[88] J. L. Synge and A. Schild. Tensor Calculus. Dover Publications, New York, 1978. (originally published:
University of Toronto Press, 1949).
[89] J.-L. Thiffault. Covariant time derivatives for dynamical systems. Journal of Physics A: Mathematical
and General, 34:5875–5885, 2001.
[90] T. Y. Thomas. Kinematically preferred co-ordinate systems. Proceedings of the National Academy of
Sciences, 41:762–770, 1955.
[91] T. C. T. Ting. Determination of C 1/2 , C −1/2 and more general isotropic tensor functions of C. Journal
of Elasticity, 15:319–323, 1985.
[92] C. Truesdell. Hypo-elasticity. Journal of Rational Mechanics and Analysis, 4(1):83–133, 1955.
[93] C. Truesdell. The simplest rate theory of pure eleasticity. Communications on Pure and Applied
Mathematics, 8:123–132, 1955.
[94] C. Truesdell. The Elements of Continuum Mechanics. Springer-Verlag Berlin Heidelberg New York,
1966.
[95] C. Truesdell and W. Noll. The Non-Linear Field Theories of Mechanics. Springer-Verlag Berlin Heidelberg New York, 3rd edition, 2004.
[96] C. Truesdell and R. A. Toupin. The Classical Field Theories, volume III/1 of Encyclopedia of Physics,
pages 226–793. Springer-Verlag Berlin Göttingen Heidelberg, 1960.
[97] J. G. Trulio. Theory and structure of the AFTON codes. Report AFWL-TR-66-19, Air Force Weapons
Laboratory, Kirtland Air Force Base, New Mexico, USA, 1966.
46
[98] P.-A. von Wolffersdorff. A hypoplastic relation for granular materials with a predefined limit state
surface. Mechanics of Cohesive-Frictional Materials, 1:251–271, 1996.
[99] M. L. Wilkins. Calculation of elastic-plastic flow. Report UCRL-7322, Lawrence Radiation Laboratory,
University of California, Livermore, USA, 1963.
[100] H. Xiao, O. T. Bruhns, and A. Meyers. Hypo-elasticity model based upon the logarithmic stress rate.
Journal of Elasticity, 47:51–68, 1997.
[101] H. Xiao, O. T. Bruhns, and A. Meyers. Logarithmic strain, logarithmic spin and logarithmic rate.
Acta Mechanica, 124:89–105, 1997.
[102] H. Xiao, O. T. Bruhns, and A. Meyers. Existence and uniqueness of the integrable-exactly hypoelastic
◦
equation τ ∗ = λ(trD)I + 2µD and its significance to finite inelasticity. Acta Mechanica, 138:31–50,
1999.
[103] H. Xiao, O. T. Bruhns, and A. Meyers. The choice of objective rates in finite elastoplasticity: General
results on the uniqueness of the logarithmic rate. Proceedings of the Royal Society of London. Series A,
456:1865–1882, 2000.
[104] S. Zaremba. Sur une forme perfectionée de la théorie de la relaxation. Bulletin international de
l’Académie des Sciences de Cracovie, Classe des Sciences Mathematiques et Naturelles, 124:594–614,
1903.
[105] X. Zhou and K. K. Tamma. On the applicability and stress update formulations for corotational stress
rate hypoelasticity constitutive models. Finite Elements in Analysis and Design, 39:783–816, 2003.
47
| 5cs.CE
|
DeepHeart: Semi-Supervised Sequence Learning for Cardiovascular Risk
Prediction
Brandon Ballinger, Johnson Hsieh, Avesh Singh,
Nimit Sohoni, Jack Wang
Cardiogram
San Francisco, CA
Abstract
We train and validate a semi-supervised, multi-task LSTM
on 57,675 person-weeks of data from off-the-shelf wearable heart rate sensors, showing high accuracy at detecting multiple medical conditions, including diabetes (0.8451),
high cholesterol (0.7441), high blood pressure (0.8086), and
sleep apnea (0.8298). We compare two semi-supervised training methods, semi-supervised sequence learning and heuristic pretraining, and show they outperform hand-engineered
biomarkers from the medical literature. We believe our work
suggests a new approach to patient risk stratification based
on cardiovascular risk scores derived from popular wearables
such as Fitbit, Apple Watch, or Android Wear.
Introduction
In medicine, each label represents a human life at risk —
for example, a person who recently suffered a heart attack
or experienced an abnormal heart rhythm. As a result, even
widely-used predictive models like CHA2DS2-VASc may
be derived from as few as 25 positive labels (Y H Lip et al.
2009). However, popular wearables, such as Fitbit and Apple
Watch, generate trillions of unlabeled sensor data points per
year, including rich signals like resting heart rate and heart
rate variability, which have been shown to correlate with
health conditions as diverse as diabetes, sleep apnea, atrial
fibrillation, heart failure, sudden cardiac death, and irritable
bowel syndrome (Kamath, Watanabe, and Upton 2012).
Using consumer-grade heart rate sensors in a medical
context presents several challenges. First, the sensors themselves have significant error (Gillinov 2016). Second, they
vary the rate of measurement to preserve battery life. Third,
since wearables are used in an ambulatory setting, daily activities like walking, exercise, stress, consuming alcohol, or
drinking coffee (Figure 1) may confuse simple heuristics.
Deep neural networks (LeCun, Bengio, and Hinton 2015)
have shown high accuracy at pattern recognition from noisy,
complex inputs, including automated detection of diabetic
retinopathy from images (Gulshan et al. 2016), skin cancer
from mobile phone cameras (Esteva et al. 2017), and the
onset of health conditions from electronic medical records
(Razavian, Marcus, and Sontag 2016; Lipton et al. 2015;
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
Geoffrey H. Tison, Gregory M. Marcus,
Jose M. Sanchez, Carol Maguire
Jeffrey E. Olgin, Mark J. Pletcher
Department of Medicine
University of California
San Francisco, CA
Choi et al. 2016; Che et al. 2015). However, purely supervised deep learning requires many labeled examples. No
such database exists for wearable data, and in most hospital systems, recruiting even a hundred patients into a new
medical study is labor-intensive and expensive.
To solve this, we take a two-pronged approach. First, we
employ a mobile phone application to recruit 14,011 participants from across the world, collecting 200 million unlabeled sensor measurements. Data from these participants is
then used in one of two semi-supervised training procedures.
In semi-supervised sequence learning (Dai and Le 2015), an
LSTM (Hochreiter and Schmidhuber 1997) is pretrained as a
sequence autoencoder; weights from pretraining are used as
initialization for a second supervised phase, making use of
a limited pool of labeled data from participants with known
diagnoses. In heuristic pretraining, the neural network is instead pretrained to compute heart-rate-derived biomarkers
from the medical literature, partially bridging the gap between feature engineering and deep learning.
The rest of this paper is organized as follows. We first
describe a novel data set derived from 14,011 participants
with wearable heart rate monitors, recruited in partnership
with the Health eHeart Study of the University of California, San Francisco. We then describe DeepHeart, our model
architecture, and the semi-supervised training methods used,
semi-supervised sequence learning and heuristic pretraining.
Then we present experimental results showing that semisupervised deep neural networks achieve higher accuracy
in detecting diabetes, sleep apnea, high blood pressure, and
high cholesterol than several strong baselines derived from
biomarkers in the medical literature. We conclude by describing some of the historical barriers to wide-scale deployment of artificial intelligence in medicine and how we plan
to address them in future work.
Data and Study Cohort
Electronic medical records contain important discrete
events, such as lab tests and diagnoses, while continuous
wearable data from mobile health platforms, such as Apple
HealthKit or Google Fit, cover the majority of the patient’s
time outside of a medical setting. In order to train a deep
neural network to predict diagnoses given wearable data, we
sought to combine the two data sources.
Figure 1: Examples of wearable heart rate data showing different physiological states: exercise, sleeping, stress, and medical
conditions.
Multi-Channel, Multi-Timescale Sensor Data
Study Cohort
We recruited 14,011 users of a popular Apple Watch app
into an online, IRB-approved study run in partnership with
the cardiology department of the University of California,
San Francisco (Tison et al. 2017a). Participants completed
a medical history, including previous diagnoses, blood test
results, and medications. A mobile app—Cardiogram—
integrated with HealthKit continuously stored, processed,
and displayed participant heart rate, step count, and other
activity data.
We chose to focus on four highly prevalent but often undiagnosed conditions associated with cardiovascular risk: high
cholesterol, hypertension (high blood pressure), sleep apnea,
and diabetes. 45.8% of diabetes cases are undiagnosed (Beagley et al. 2014). Sleep apnea is estimated to affect 22 million people in the US alone, over 80% of whom are undiagnosed (Young et al. 1997). Together, high blood pressure
and high cholesterol account for 17.3% of global deaths (Alwan and others 2011), as they are both major risk factors for
heart attacks.
Condition
Table 1: Cohort Statistics
Number of PersonPeople
Weeks of
Training
Data
PersonWeeks of
Testing
Data
Unlabeled
1,462
10,616
N/A
High Cholesterol
Hypertension
Sleep Apnea
Diabetes
2,331
2,230
1,016
462
9,220
7,651
3,662
1,678
3,243
2,762
1,141
644
Consumer wearables produce multiple channels of data—
most commonly, step count measured using a wrist-worn
accelerometer and optical heart rate measured using photoplethysmography. For example, the Apple Watch measures
heart rate every five seconds when in workout mode and
roughly every five minutes during other times of the day.
Likewise, step counts are recorded every few minutes while
the user is walking.
We encoded sensor measurements into a tensor Xu,t,c
with indices for the user-week u, timestep t, and input channel c. Heart rate and step count were encoded as separate
input channels. To accommodate variable timescales, we encoded intra-channel time-deltas as a special dt channel:
dttransf ormed = 0.1log(
dt
)
5000
(1)
Without transforming dt, the model does not train due
to the wide spread of values (ranging from 5000 to
28,800,000).
Sequences contained up to 4096 timesteps of sensor data,
and shorter sequences were zero-padded.
Pre-Processing
Each participant was randomly assigned to either the training, tuning, or testing set. Each participant’s data was then
split into week-long chunks, and any weeks with ≤ 672
heart rate measurements (≤ 8 hours per day of background
measurements) or ≤ 30 minutes of continuous heart rate
recordings were omitted. After filtering, there were 57,675
person-weeks of data in total, divided into 33,628 for training, 18,555 for tuning, and 12,790 for validation. The training, tuning, and testing sets contained completely disjoint
sets of participants.
length of 5. Each convolutional layer contains 128 convolutional channels. After each convolutional layer, we apply
dropout with probability 0.2 to prevent overfitting, and apply
max pooling with pool length 2 to reduce dimensionality.
Bidirectional LSTM
The convolutional layers are followed by four bidirectional
LSTM layers that model longer temporal patterns in wearable sensor data and their corresponding multi-timescale diagnoses. Each bidirectional LSTM layer contains 128 units
(64 in each direction).
A dropout of 0.2 is applied to this final LSTM layer, and
the result is run through a simple output unit: a convolution
of filter length 1. The purpose of this convolution is to map
the outputs of the LSTM into a single prediction per task,
per timestep. For this final layer, we use a tanh activation so
outputs range from -1 to 1.
Training Methods
Figure 2: DeepHeart architecture.
In addition to ordinary supervised training, we experimented
with two pretraining architectures. Semi-supervised training
is advantageous here because unlabeled data from wearables
is abundant, but medical condition labels are more scarce.
Unsupervised Sequence Pretraining
Multi-Task Labeled Output
Similar to X, our output is a tensor Yu,t,c with indices for
the user-week u, timestep t, and output channel c. There are
separate output channels for prevalent cases of each condition: high cholesterol, hypertension, sleep apnea, and diabetes. Diagnoses are derived from validated health surveys
collected in partnership with the Health eHeart Study of the
University of California, San Francisco. In addition, we included an output channel of ECG readings1 . ECG readings
were aligned to the closest sensor measurement in X, and diagnoses were aligned to the last timestep in X. Each Yt [task]
is a discrete label: either positive (user has condition, or abnormal ECG reading) or negative (user is not diagnosed with
condition, or ECG reading is normal). We used output masking to score the models only when Yt [task] has a value.
Model Architecture
Figure 2 shows our model architecture. The inputs are multichannel, multi-timescale sensor measurements, and the outputs are multi-task, multi-timescale diagnoses. Table 3 summarizes hyperparameter tuning experiments.
Temporal Convolutions
The first three layers are temporal convolutions, which have
been shown to efficiently extract features and model temporal translation invariance (Amodei et al. 2016). The first
layer has a wide filter length of 12, similar to AlexNet
(Krizhevsky, Sutskever, and Hinton 2012), and the next two
layers use the residual units (He et al. 2016) with a filter
1
Our research builds on previous work to detect atrial fibrillation using ECGs. We chose to keep this output task so our model
could benefit from ECG readings.
Semi-supervised sequence learning has been shown to improve performance when presented with large amounts of
unlabeled training data (Dai and Le 2015). We experimented
with this approach by first training a sequence-to-sequence
autoencoder, consisting of the 3 convolutional and 4 recurrent layers. We then use the weights from the autoencoder
as initializing parameters for the corresponding layers in the
supervised architecture. In order to force the autoencoder to
not merely memorize the input, we added Gaussian noise to
the input sequence.
Weakly-Supervised Heuristic Pretraining
Previous work has applied statistical methods on beat-tobeat variability to detect heart arrhythmias (McManus et al.
2013). Inspired by this, we synthesized a dataset for pretraining using a time-windowed heart rate variability metric. We
defined four output tasks, with window sizes of 5 seconds,
30 seconds, 5 minutes, and 30 minutes, computed the average absolute difference between successive heart rate measurements, and used these as output channels in a weaklysupervised pretraining phase. The resulting weights were
used for initialization in supervised training. In this way, we
use pretraining as a technique to bridge feature engineering
and deep learning.
Experiments
To evaluate our architecture, we ran a series of experiments
summarized in Table 2. For each condition on our data set,
we report the c-statistic—the area under the ROC curve—
on the test set. Each experiment used the Adam optimizer
(Kingma and Ba 2014) and a squared error loss, since we
found squared error to be more numerically stable than
cross-entropy.
Name
Table 2: Experiment Results (c-statistics)
Diabetes High Cholesterol Sleep Apnea
High Blood Pressure
Logistic Regression
Support Vector Machine
Decision Tree
Random Forest
Multi-layer Perceptron
0.7906
0.5722
0.4142
0.5515
0.7846
0.5941
0.5996
0.5152
0.5647
0.4327
0.6583
0.5490
0.6415
0.5907
0.6172
0.6389
0.6106
0.6625
0.6310
0.7195
LSTM, No Pretraining
0.8451
0.6736
0.8041
0.7991
LSTM, Heuristic Pretraining
LSTM, Unsupervised Pretraining
0.8366
0.7998
0.7148
0.7441
0.7951
0.8298
0.7427
0.8086
Baselines
We compared our architecture with multiple machine learning algorithms using hand-engineered biomarkers derived
from the medical literature, including resting heart rate (Fox
et al. 2007), time-windowed average heart rate (Kamath,
Watanabe, and Upton 2012), time-windowed standard deviation of heart rates (Kamath, Watanabe, and Upton 2012),
time-windowed spectrum entropy (Kamath, Watanabe, and
Upton 2012), time-windowed root mean squared of successive beats per minute differences (McManus et al. 2013),
and time-windowed entropy of BPM differences (McManus
et al. 2013). We used 5-minute and 30-minute time windows,
and included global versions of standard deviation and RMS
of successive BPM differences, for 13 features in total. We
trained several standard machine learning algorithms on all
of these features, including scikit-learn’s (Pedregosa et al.
2011) implementation of logistic regression, support vector machines, decision trees, random forests, and multi-layer
perceptrons.
LSTM without Pretraining
We first trained a purely-supervised LSTM without pretraining, which performed significantly better than the baselines
on sleep apnea (0.80 vs 0.55-0.66), hypertension (0.80 vs
0.61-0.72), high cholesterol (0.67 vs 0.43-0.60), and diabetes (0.85 vs 0.41-0.79). The gap is smallest for diabetes,
with the least amount of labeled data, suggesting potential benefit from combining hand-engineered features with
deep learning. Previous work from the medical community
has suggested that diabetes may impact heart rate variability
through effects on the autonomic nervous system (Kudat et
al. 2006). Our work builds on top of these previous findings,
and suggests that heart rate variability changes driven by diabetes can be detected via consumer, off-the-self wearable
heart rate sensors.
Heuristic Pretrained LSTM
When the supervised LSTM is initialized with weights from
heuristic pretraining, we saw noticeable improvements for
high cholesterol (0.71 vs 0.67), neutral results for sleep apnea and diabetes, and a significant loss for high blood pressure (0.74 vs 0.80). The gain on high cholesterol is particularly surprising, given that the baseline results, which are
based on similar hand-engineered heuristic features, did not
perform well. One explanation is that the neural network is
able to find non-linear relationships for the underlying patterns that generate these heuristic features, and the insights
discovered by the heuristic pretraining were useful for predicting high cholesterol.
On high blood pressure, the heuristic-pretrained LSTM
performs significantly worse than the LSTM without pretraining. One explanation is that the heuristic of average
BPM differences is a poor predictor of high blood pressure, and gradient descent cannot escape from this space of
poorly-performing predictors.
Semi-supervised Sequence Learning
The last experiment applied semi-supervised sequence
learning: we first pretrained DeepHeart as a sequence autoencoder, using the encoder weights as initialization for a
second, supervised phase, as described in (Dai and Le 2015).
This approach resulted in a significant improvement on high
cholesterol, high blood pressure, and sleep apnea.
The 0.83 c-statistic at detecting sleep apnea is surprising
given that few participants sleep with their watch. This suggests that wearable-derived biomarkers measured during the
day capture a distinct signature of sleep apnea.
Among all four health conditions, high cholesterol has the
least direct physiological relationship with heart rate variability in the medical literature. While we are able to detect
high cholesterol with c-statistic of 0.74, it is the lowest performing prediction out of the four disease states, suggesting that much of the correlation may be a reflection of confounding variables such as age, sex, or usage of medications
like beta blockers. Nevertheless, a c-statistic of 0.74 from
consumer-grade wearables alone is a surprising and novel
finding.
Effect of Varying Amounts of Labeled Data
To quantify the impact of each pretraining technique, we
evaluated DeepHeart’s performance when trained on varying fractions of the labeled data (5%, 10%, 20%, 50%, 70%,
and 100%), using no pretraining, heuristic pretraining, and
semi-supervised sequence learning (Figure 4).
For hypertension and sleep apnea, unsupervised pretraining yields a 10x improvement in data-efficiency: the DNN
achieves nearly the same accuracy with 10% of labeled data
using unsupervised pretraining as it does with 100% of labeled data using no pretraining. For diabetes and cholesterol,
the effect is muted. For diabetes, this is due to a scarcity of
labels (1,678, vs 3,662-9,220 for the other conditions). For
high cholesterol, the lack of correlation between the number
of labels and accuracy of the algorithm likely confirms that
high cholesterol is primarily detected through the effect of
confounding variables such as age, sex, and medications.
Figure 4: Accuracy as a function of percentage of training labels used for each condition. Sleep apnea and hypertension,
which show the greatest effect, have both a direct physiological connection to heart rate variability and a large amount of
labeled data in our training set. For diabetes and high cholesterol, the effect of adding more labeled data is muted due
to the low absolute number of labels and less direct physiological connection, respectively. Error bars show the 95th
percentile confidence interval.
Figure 3: ROC curves (sensitivity-specificity curves) comparing the accuracy of heuristic pretraining and semisupervised pretraining to the strongest overall baseline in
detecting diabetes, sleep apnea, high cholesterol, and high
blood pressure.
The shape of the ROC curves (Figure 3) are also interesting to note. It is clear that the semi-supervised sequence
LSTM performs very well for detecting diabetes, sleep apnea, and hypertension, three of the most highly prevalent
conditions with undiagnosis rates of 45.8%, 80%, and 20%
respectively. For high cholesterol, although the c-statistic is
lower, there are high-precision operating points which rule
out more than half of the study population. This suggests
wearables could be used to focus diagnostic testing on a
high-risk subset of the population, leading to cost-effective
screening for metabolic syndrome and getting at-risk patients care sooner.
Ablative Analysis of Input Channels
To determine the impact of each input channel, we tested
models using heart rate and step count in isolation.
A model trained only on heart rate timeseries performed
significantly worse on diabetes and high blood pressure (decrease of 0.0653 and 0.0721 in AUC, respectively), and the
same on sleep apnea and high blood pressure (within the
95% confidence interval), suggesting that heart rate is being interpreted as a response to activity. Within the medical literature, heart rate recovery after exercise is associated with lower risks of cardiovascular disease (MorshediMeibodi et al. 2002). A model using only aggregate step
counts achieved an AUC of 0.7011 (diabetes), 0.5811 (sleep
apnea), 0.5714 (hbp), and 0.5601 (high cholesterol), suggesting that although overall physical activity plays an important role in predicting the onset of disease, the heart’s
response to physical activity is a more salient biomarker that
can be captured using deep learning.
Discussion
We’ve shown that two methods, semi-supervised sequence
learning and heuristic pretraining, address a key technical
challenge in applying deep learning to medicine: achieving high accuracy with limited labeled data. Furthermore,
we’ve shown high accuracy at detecting four common health
conditions—diabetes, sleep apnea, hypertension, and high
cholesterol–using readily available, off-the-shelf heart rate
monitors. Since these four conditions are highly prevalent,
commonly undiagnosed, and reduced HRV is associated
with the earliest stages of disease progression (Schroeder
et al. 2005; Schroeder et al. 2003), this suggests a new approach to public health screening using deep learning.
While these results are promising, artificial intelligencebased systems have shown high accuracy at medical tasks—
even outperforming human physicians—since the days of
MYCIN (Yu et al. 1979), CADUCEUS (Banks 1985), and
INTERNIST-I (Wolfram 1995). For AI-based systems to
broadly improve human health, several key challenges beyond prediction accuracy must be addressed.
A first challenge is confounding factors. Statistical models in medicine are typically used for etiology, not just prediction, and cause (disease) and effect (treatment) may be intertwined in subtle ways within medical data sets (Shmueli
and others 2010). For example, a person with previouslydiagnosed hypertension may be prescribed beta blockers,
which affect heart rate variability, and therefore an algorithm
with high prediction accuracy at identifying hypertension
may be picking up signals of both the underlying disease
and the treatment. While there are well-known algorithms to
adjust for confounders in linear models, confounder-aware
neural network architectures are a critical future area of research.
A second challenge is deployment. While hospital departments such as the intensive care unit or radiology department often have ample training data available, deploying a
new algorithm in a hospital is typically slow or impossible
because errors are costly (a severe error may cost a human
life), the technical implementation is cumbersome (an integration with an electronic medical record system), the organizational complexity is high (high regulatory burden, necessity to train physicians on new systems), and the financial incentives are often misaligned (for example, preventing
readmissions may reduce the hospital’s revenue). An alternative is to take an outside-in approach: design care pathways for ambulatory, mobile-based screening that operate
outside of the hospital but guide participants in to the appropriate point of care when a neural network detects high
risk. Here, false negatives represent the status quo (no diagnosis), and with sufficiently high precision the cost of false
positives (e.g., blood tests, mobile ECGs) is exceeded by the
cost savings to the healthcare system of each true positive.
Since an algorithm can be packaged in a mobile app, it can
be updated and quickly deployed to participants. We are currently deploying our algorithm in ambulatory, mobile trials
in partnership with major medical centers to show real-world
efficacy and build a rigorous clinical evidence base (Tison et
al. 2017b).
A third challenge is interpretability. Deep learning systems perform well but are “black boxes,” whereas medical
systems prefer to provide an explanation of the chain of reasoning. We intend to incorporate mechanisms such as differentiable attention (Xu et al. 2015; Chan et al. 2016) to help
both physicians and regulators understand the internal workings of the neural network and its potential failure modes.
A last challenge is the time-scale and volume of data.
Many participants in our study have more than a year of
heart rate and step count data, corresponding to nearly
one million time steps. We intend to test neural architectures that can better model long-range dependencies, such
as Clockwork RNNs (Koutnik et al. 2014), Phased LSTMs
(Neil, Pfeiffer, and Liu 2016), and Gaussian Process RNNs
(Joseph Futoma and O’Brien 2017) in order to better model
changes to a participant’s physiological state over a long period of time.
While these challenges may seem daunting, the constraints of medicine—scarce labeled data, confounding
factors, difficulty of deployment, and the necessity of
interpretability—also represent opportunities to develop
new techniques. This work is a first step in showing how
health conditions can be detected using techniques first developed in natural language processing and computer vision.
Over time, we plan to extend this work to address the challenges above and deploy neural networks that can improve
health in the real world.
References
[Alwan and others 2011] Alwan, A., et al. 2011. Global
status report on noncommunicable diseases 2010. World
Health Organization.
[Amodei et al. 2016] Amodei, D.; Anubhai, R.; Battenberg, E.; Case, C.; Casper, J.; Catanzaro, B.; Chen, J.;
Chrzanowski, M.; Coates, A.; Diamos, G.; et al. 2016. End
to end speech recognition in english and mandarin.
[Banks 1985] Banks, G. 1985. Artificial intelligence in medical diagnosis: the internist/caduceus approach. Critical reviews in medical informatics 1(1):23–54.
[Beagley et al. 2014] Beagley, J.; Guariguata, L.; Weil, C.;
and Motala, A. A. 2014. Global estimates of undiagnosed
diabetes in adults. Diabetes research and clinical practice
103(2):150–160.
[Chan et al. 2016] Chan, W.; Jaitly, N.; Le, Q.; and Vinyals,
O. 2016. Listen, attend and spell: A neural network for large
vocabulary conversational speech recognition. In Acoustics,
Speech and Signal Processing (ICASSP), 2016 IEEE International Conference on, 4960–4964. IEEE.
[Che et al. 2015] Che, Z.; Kale, D.; Li, W.; Bahadori, M. T.;
and Liu, Y. 2015. Deep computational phenotyping. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 507–516.
ACM.
[Choi et al. 2016] Choi, E.; Bahadori, M. T.; Schuetz, A.;
Stewart, W. F.; and Sun, J. 2016. Doctor ai: Predicting
clinical events via recurrent neural networks. In Machine
Learning for Healthcare Conference, 301–318.
[Dai and Le 2015] Dai, A. M., and Le, Q. V. 2015. Semisupervised sequence learning. In Advances in Neural Information Processing Systems, 3079–3087.
[Esteva et al. 2017] Esteva, A.; Kuprel, B.; Novoa, R. A.;
Ko, J.; Swetter, S. M.; Blau, H. M.; and Thrun, S. 2017.
Dermatologist-level classification of skin cancer with deep
neural networks. Nature 542(7639):115–118.
[Fox et al. 2007] Fox, K.; Borer, J. S.; Camm, A. J.; Danchin,
N.; Ferrari, R.; Sendon, J. L. L.; Steg, P. G.; Tardif, J.-C.;
Tavazzi, L.; Tendera, M.; et al. 2007. Resting heart rate in
cardiovascular disease. Journal of the American College of
Cardiology 50(9):823–830.
[Gillinov 2016] Gillinov, M. 2016. Accuracy of commercially available heart rate monitors.
https://
clinicaltrials.gov/ct2/show/NCT02818244.
[Gulshan et al. 2016] Gulshan, V.; Peng, L.; Coram, M.;
Stumpe, M. C.; Wu, D.; Narayanaswamy, A.; Venugopalan,
S.; Widner, K.; Madams, T.; Cuadros, J.; et al. 2016. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs.
Jama 316(22):2402–2410.
[He et al. 2016] He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016.
Identity mappings in deep residual networks. In European
Conference on Computer Vision, 630–645. Springer.
[Hochreiter and Schmidhuber 1997] Hochreiter, S., and
Schmidhuber, J. 1997. Long short-term memory. Neural
computation 9(8):1735–1780.
[Joseph Futoma and O’Brien 2017] Joseph Futoma, Sanjay Hariharan, K. H. M. S. N. B. M. C. A. B., and O’Brien,
C. 2017. An improved multi-output gaussian process rnn
with real-time validation for early sepsis detection.
[Kamath, Watanabe, and Upton 2012] Kamath,
M.
V.;
Watanabe, M.; and Upton, A. 2012. Heart rate variability
(HRV) signal analysis: clinical applications. CRC Press.
[Kingma and Ba 2014] Kingma, D., and Ba, J. 2014. Adam:
A method for stochastic optimization. arXiv preprint
arXiv:1412.6980.
[Koutnik et al. 2014] Koutnik, J.; Greff, K.; Gomez, F.; and
Schmidhuber, J. 2014. A clockwork rnn. In International
Conference on Machine Learning, 1863–1871.
[Krizhevsky, Sutskever, and Hinton 2012] Krizhevsky, A.;
Sutskever, I.; and Hinton, G. E. 2012. Imagenet classification with deep convolutional neural networks. In Advances
in neural information processing systems, 1097–1105.
[Kudat et al. 2006] Kudat, H.; Akkaya, V.; Sozen, A.;
Salman, S.; Demirel, S.; Ozcan, M.; Atilgan, D.; Yilmaz, M.;
and Guven, O. 2006. Heart rate variability in diabetes patients. Journal of international medical research 34(3):291–
296.
[LeCun, Bengio, and Hinton 2015] LeCun, Y.; Bengio, Y.;
and Hinton, G.
2015.
Deep learning.
Nature
521(7553):436–444.
[Lipton et al. 2015] Lipton, Z. C.; Kale, D. C.; Elkan, C.; and
Wetzell, R. 2015. Learning to diagnose with lstm recurrent
neural networks. arXiv preprint arXiv:1511.03677.
[McManus et al. 2013] McManus, D. D.; Lee, J.; Maitas, O.;
Esa, N.; Pidikiti, R.; Carlucci, A.; Harrington, J.; Mick, E.;
and Chon, K. H. 2013. A novel application for the detection
of an irregular pulse using an iphone 4s in patients with atrial
fibrillation. Heart Rhythm 10(3):315–319.
[Morshedi-Meibodi et al. 2002] Morshedi-Meibodi, A.; Larson, M. G.; Levy, D.; ODonnell, C. J.; and Vasan, R. S.
2002. Heart rate recovery after treadmill exercise testing and
risk of cardiovascular disease events (the framingham heart
study). The American journal of cardiology 90(8):848–852.
[Neil, Pfeiffer, and Liu 2016] Neil, D.; Pfeiffer, M.; and Liu,
S.-C. 2016. Phased lstm: Accelerating recurrent network
training for long or event-based sequences. In Advances in
Neural Information Processing Systems, 3882–3890.
[Pedregosa et al. 2011] Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.;
Prettenhofer, P.; Weiss, R.; Dubourg, V.; et al. 2011. Scikitlearn: Machine learning in python. Journal of Machine
Learning Research 12(Oct):2825–2830.
[Razavian, Marcus, and Sontag 2016] Razavian, N.; Marcus,
J.; and Sontag, D. 2016. Multi-task prediction of disease onsets from longitudinal laboratory tests. In Machine Learning
for Healthcare Conference, 73–100.
[Schroeder et al. 2003] Schroeder, E. B.; Liao, D.; Chambless, L. E.; Prineas, R. J.; Evans, G. W.; and Heiss, G. 2003.
Hypertension, blood pressure, and heart rate variability. Hypertension 42(6):1106–1111.
[Schroeder et al. 2005] Schroeder, E. B.; Chambless, L. E.;
Liao, D.; Prineas, R. J.; Evans, G. W.; Rosamond, W. D.;
and Heiss, G. 2005. Diabetes, glucose, insulin, and heart
rate variability. Diabetes care 28(3):668–674.
[Shmueli and others 2010] Shmueli, G., et al. 2010. To explain or to predict? Statistical science 25(3):289–310.
[Tison et al. 2017a] Tison, G. H.; Hsu, K.; Hsieh, J. T.;
Ballinger, B. M.; Pletcher, M. J.; Marcus, G. M.; and Olgin, J. E. 2017a. Abstract 21029: Achieving high retention
in mobile health research using design principles adopted
from widely popular consumer mobile apps. Circulation
136(Suppl 1):A21029–A21029.
[Tison et al. 2017b] Tison, G. H.; Singh, A. C.; Ohashi,
D. A.; Hsieh, J. T.; Ballinger, B. M.; Olgin, J. E.; Marcus,
G. M.; and Pletcher, M. J. 2017b. Abstract 21042: Cardiovascular risk stratification using off-the-shelf wearables and
a multi-task deep learning algorithm. Circulation 136(Suppl
1):A21042–A21042.
[Wolfram 1995] Wolfram, D. A. 1995. An appraisal of
internist-i. Artificial intelligence in medicine 7(2):93–116.
[Xu et al. 2015] Xu, K.; Ba, J.; Kiros, R.; Cho, K.; Courville,
A.; Salakhudinov, R.; Zemel, R.; and Bengio, Y. 2015.
Show, attend and tell: Neural image caption generation with
visual attention. In International Conference on Machine
Learning, 2048–2057.
[Y H Lip et al. 2009] Y H Lip, G.; Nieuwlaat, R.; Pisters, R.;
Lane, D.; and Crijns, H. 2009. Refining clinical risk stratification for predicting stroke and thromboembolism in atrial
fibrillation using a novel risk factor-based approach the euro
heart survey on atrial fibrillation. 137:263–72.
[Young et al. 1997] Young, T.; Evans, L.; Finn, L.; and Palta,
M. 1997. Estimation of the clinically diagnosed proportion
of sleep apnea syndrome in middle-aged men and women.
Sleep 20(9):705–706.
[Yu et al. 1979] Yu, V. L.; Fagan, L. M.; Wraith, S. M.;
Clancey, W. J.; Scott, A. C.; Hannigan, J.; Blum, R. L.;
Buchanan, B. G.; and Cohen, S. N. 1979. Antimicrobial
selection by a computer. a blinded evaluation by infectious
diseases experts. Jama 242(12):1279–1282.
Width
32
32
32
32
32
32
32
32
64
64
64
64
64
64
64
128
128
128
128
128
128
128
Conv Depth
2
2
2
2
4
4
4
4
2
2
2
4
4
4
4
2
2
2
4
4
4
4
LSTM Depth
2
2
4
4
2
2
4
4
2
2
4
2
2
4
4
2
4
4
2
2
4
4
Initial Filter
12
5
12
5
12
5
12
5
12
5
12
12
5
12
5
5
12
5
12
5
12
5
AF
0.7548
0.7340
0.7621
0.7633
0.7480
0.7652
0.7440
0.7336
0.7317
0.7514
0.7891
0.7903
0.7855
0.8079
0.7958
0.7770
0.8212
0.7695
0.8076
0.7646
0.8143
0.7702
Diabetes
0.8709
0.8724
0.8748
0.8659
0.8540
0.8622
0.8711
0.8628
0.8696
0.8719
0.8647
0.8517
0.8655
0.8805
0.8756
0.8653
0.8718
0.8727
0.8668
0.8658
0.8723
0.8640
High BP
0.8494
0.8375
0.8455
0.7995
0.8502
0.8258
0.8233
0.8403
0.8494
0.8544
0.8444
0.8550
0.8483
0.8453
0.8434
0.8529
0.8262
0.8427
0.8616
0.8030
0.8414
0.8555
Apnea
0.8610
0.8594
0.8664
0.8567
0.8583
0.8592
0.8605
0.8626
0.8599
0.8585
0.8597
0.8622
0.8559
0.8584
0.8632
0.8607
0.8604
0.8630
0.8613
0.8516
0.8600
0.8575
Chol
0.8549
0.8541
0.8511
0.8257
0.8352
0.8211
0.8442
0.8444
0.8594
0.8525
0.8467
0.8513
0.8442
0.8647
0.8618
0.8549
0.8361
0.8273
0.8409
0.8286
0.8639
0.8407
Avg
0.8382
0.8315
0.8400
0.8222
0.8291
0.8267
0.8286
0.8287
0.8340
0.8377
0.8409
0.8421
0.8399
0.8514
0.8480
0.8422
0.8431
0.8350
0.8476
0.8227
0.8504
0.8376
Table 3: Hyperparameter tuning results, showing AUC ROC for each health condition, as well as the average across all health
conditions. Here, Width is the number of output channels in each convolutional and LSTM layer, Conv Depth is the number
of convolutional layers, LSTM Depth is the number of LSTM layers, and Initial Filter is the initial filter length of the first
convolutional layer.
| 2cs.AI
|
Non-linear Learning for Statistical Machine Translation
First Author
Affiliation / Address line 1
Affiliation / Address line 2
Affiliation / Address line 3
email@domain
Second Author
Affiliation / Address line 1
Affiliation / Address line 2
Affiliation / Address line 3
email@domain
arXiv:1503.00107v1 [cs.CL] 28 Feb 2015
Abstract
The modeling method is extended to log-linear
models by Och and Ney (2002), as shown in Equation 2, where hm (e|f ) is the mth feature function
and λm is the corresponding weight.
Modern statistical machine translation
(SMT) systems usually use a linear combination of features to model the quality
of each translation hypothesis. The linear
combination assumes that all the features
are in a linear relationship and constrains
that each feature interacts with the rest features in an linear manner, which might
limit the expressive power of the model
and lead to a under-fit model on the current data. In this paper, we propose a nonlinear modeling for the quality of translation hypotheses based on neural networks,
which allows more complex interaction
between features. A learning framework is
presented for training the non-linear models. We also discuss possible heuristics
in designing the network structure which
may improve the non-linear learning performance. Experimental results show that
with the basic features of a hierarchical
phrase-based machine translation system,
our method produce translations that are
better than a linear model.
P r(e|f ) = pλM (e|f )
1
PM
exp[ m=1 λm hm (e|f )]
=P
PM
′
e′ exp[ m=1 λm hm (e |f )]
Because the normalization term in Equation 2 is
the same for all translation hypotheses of the same
source sentence, the score of each hypothesis, denoted by sL , is actually a linear combination of all
features, as shown in Equation 3.
sL (e) =
M
X
λm hm (e|f )
(3)
m=1
1 Introduction
One of the core problems in the research of statistical machine translation is the modeling of translation hypotheses. Each modeling method defines a
score of a target sentence e = e1 , e2 , ..., ei , ..., eI ,
given a source sentence f = f1 , f2 , ..., fj ...fJ ,
where each ei is the ith target word and fj is
the jth source word. The well-known modeling method starts from the Source-Channel
model (Brown et al., 1993)(Equation 1). The scoring of e decomposes to the calculation of a translation model and a language model.
P r(e|f ) = P r(e)P r(f |e)/P r(f )
(2)
(1)
The log-linear models are flexible to incorporate new features and show significant
advantage over the traditional source-channel
models, thus become the state-of-the-art modeling method and are applied in various translation
settings
(Yamada and Knight, 2001;
Koehn et al., 2003;
Chiang, 2005;
Liu et al., 2006).
It is worth noticing that log-linear models try to
separate good and bad translation hypotheses using a linear hyper-plane. However, complex interactions between features make it difficult to linearly separate good translation hypotheses from
bad ones (Clark et al., 2014).
Taking features in a typical phrase-based machine translation system (Koehn et al., 2003) as
an example, the language model feature favors
shorter hypotheses; the word penalty feature encourages longer hypotheses. The phrase translation probability feature selects phrases that occurs more frequently in the training corpus, which
sometimes are long with lower translation probability, as in translating named entities or id-
ioms; sometimes are short but with high translation probability, as in translating verbs or pronouns. These three features jointly decide the
choice of translations. Simply use the weighted
sum of their values may not be the best choice for
modeling translations.
As a result, log-linear models may under-fit the
data. This under-fitting may prevents the further
improvement of translation quality.
In this paper, we propose a non-linear modeling of translation hypotheses based on neural networks. The traditional features of a machine translation system are used as the input to the network. By feeding input features to nodes in a hidden layer, complex interactions among features are
modeled, resulting in much stronger expressive
power than traditional log-linear models. (Section 3)
Employing a neural network as non-linear
models for SMT has two issues to be tackled.
The first issue is the parameter learning. Loglinear models rely on minimum error rate training (MERT) (Och, 2003) to achieve best performance. When the scoring function become nonlinear, the intersection points of these non-linear
functions could not be effectively calculated and
enumerated. Thus MERT is no longer suitable for
learning the parameters. To solve the problem ,
we present a framework for effective training including several criteria to transform the training
problem into a binary classification task, a unified
objective function and an iterative training algorithm. (Section 4)
The second issue is the structure of neural network. Single layer neural networks are equivalent
to linear models; two-layer networks with sufficient nodes are capable of learning any continuous
function (Bishop, 1995). Adding more layers into
the network could model complex functions with
less nodes, but also brings the problem of vanishing gradient (Erhan et al., 2009). We adapt a
two-layer feed-forward neural network to keep the
training process efficient. We notice that one major problem that prevent a neural network training
reaching a good solution is that there are too many
local minimums in the parameter space. Thus we
discuss how to constrain the learning of neural networks with our intuition and observations of the
features. (Section 5)
Experiments are conducted to compare various settings and verify the effectiveness of our
proposed learning framework. Experimental results show that our framework could achieve better
translation quality even with the same traditional
features as previous linear models. (Section 6)
2 Related work
Many research has been attempting to bring nonlinearity into the training of SMT. These efforts
could be roughly divided into the following three
categories.
The first line of research attempted to reinterpret original features via feature transformation or additional learning. For example, Maskey
and Zhou (2012) use a deep belief network to
learn representations of the phrase translation and
lexical translation probability features. Clark et
al. (2014) used discretization to transform realvalued dense features into a set of binary indicator
features. Lu et al. (2014) learned new features using a semi-supervised deep auto encoder. These
work focus on the explicit representation of the
features and usually employ extra learning procedure. Our proposed method only take the original
feature with no transformation as input. Feature
transformation or combination are performed implicitly during the training of the network and integrated with the optimization of translation quality.
The second line of research attempted to use
non-linear models instead of log-linear models,
which is most similar in spirit with our work. Duh
and Kirchhoff (2008) used the boosting method
to combine several results of MERT and achieved
improvement in a re-ranking setting. Liu et
al. (2013) proposed an additive neural network
which employed a two-layer neural network for
embedding-based features. To avoid local minimum, they still rely on a pre-training and posttraining from MERT or PRO. Comparing to these
efforts, our proposed method takes a further step
that it is integrated with iterative training, instead
of re-ranking, and works without the help of any
pre-trained linear models.
The third line of research attempted to add
non-linear features/components into the loglinear learning framework.
Neural network
based models are trained as language models (Vaswani et al., 2013; Auli and Gao, 2014),
translation models (Gao et al., 2014) or joint language and translation models (Auli et al., 2013;
Devlin et al., 2014). Liu et al. (2013) also introduced word embedding for source and target side
of translation rule as local features. In this paper we focus on enhancing the expressive power
of the modeling, which is independent of the research of enhancing translation system with new
designed features. We believe additional improvement could be achieved by incorporating more features into our framework.
3 Non-linear Translation
The non-linear modeling of translation hypotheses could be used in both phrase-based system and
syntax-based systems. In this paper, we take the
hierarchical phrase based machine translation system (Chiang, 2005) as an example and introduce
how we fit the non-linearity into the system.
3.1
Decoding
The
basic
decoding
algorithm
could
be kept almost the same as traditional
phrase-based
or
syntax-based
translation
systems
(Yamada and Knight, 2001;
Koehn et al., 2003;
Chiang, 2005;
Liu et al., 2006). For example, in the experiments of this paper, we use a CKY style decoding
algorithm following Chiang (2005).
Our non-linear translation system is different
from traditional systems in the way to calculate
the score for each hypothesis. Instead of calculating the score as a linear combination, we use neural networks (Section 3.2) to perform a non-linear
combination of feature values.
We also use the cube-pruning algorithm (Chiang, 2005) to keep the decoding
efficient. Although the non-linearity in model
scores may cause more search errors in finding
the highest scoring hypothesis, in practice it still
achieves reasonable results.
output
layer
Mo
hidden
layer
Mh
input
Figure 1: A two-layer feed-forward neural network.
defined as:
sN (e) = σo (Mo ·σh (Mh ·hm
1 (e|f )+bh )+bo ) (4)
where M , b is the weight matrix, bias vector of
the neural nodes, respectively; σ is the activation
function, which is often set to non-linear functions
such as the tanh function or sigmoid function; subscript h and o indicates the parameters of hidden
layer and output layer, respectively.
3.3 Features
We use the standard features of a typical hierarchical phrase based translation
system(Chiang, 2005).
Adding new features
into the framework is left as a future direction.
The features as listed as following:
• p(α|γ) and p(γ|α): conditional probability
of translating α as γ and translating α as γ,
where α and γ is the left and right hand side
of a initial phrase (or hierarchical translation
rule), respectively;
• pw (α|γ) and pw (γ|α): lexical probability of
translating words in α as words in γ and
translating words in γ as words in α;
• plm : language model probability;
3.2
Two-layer Neural Networks
We employ a two-layer neural network as the nonlinear model for scoring translation hypotheses.
The structure of a typical two-layer feed-forward
neural network includes an input layer, a hidden
layer, and a output layer (as shown in Figure 1).
We use the input layer to accept input features,
the hidden layer to combine different input features, the output layer with only one node to output the model score for each translation hypothesis
based on the value of hidden nodes. More specifically, the score of hypothesis e, denoted as sN , is
• wc: accumulated count of individual words
generated during translation;
• pc: accumulated count of initial phrases used;
• rc: accumulated count of hierarchical rule
phrases used;
• gc: accumulated count of glue rule used in
this hypothesis;
• uc: accumulated count of unknown source
word;
• nc: accumulated count of source phrases that
translate into null;
4 Non-linear Learning Framework
Traditional machine translation systems rely on
MERT to tune the weight of different features.
MERT performs efficient search by enumerating
the score function of all the hypotheses and using intersections of these linear functions to form
the ”upper-envelope” of the model score function (Och, 2003). When the scoring function is
non-linear, it is not feasible to find the intersections of these functions. In this section, we discuss
alternatives to train the parameter for non-linear
models.
4.1
Training Criteria
The task of machine translation is a complex problem with structural output space. Decoding algorithms search for the translation hypothesis with
the highest score, according to a given scoring
function, from an exponentially large set of candidate hypotheses. The purpose of training is to select the scoring function, so that the function score
the hypotheses ”correctly”. The correctness is often introduced by some extrinsic metrics, such as
BLEU (Papineni et al., 2002).
~ or
We denote the scoring function as s(f , e; θ),
~
simply s, which is parametrized by θ; denote the
set of all candidate hypotheses as C; denote the extrinsic metric as eval(·). Note that, in linear cases,
s is a linear function as in Equation 3, while in
the non-linear case described in this paper, s is the
scoring function in Equation 4.
Ideally, the training objective is to select a scoring function s, from all functions S, that scores
the correct translation (or references), denoted as
ê, higher than any other hypotheses (Equation 5).
s = {s ∈ S|s(ê) > s(e) ∀e ∈ C}
(5)
In practice, the candidate set C is exponentially
large and hard to enumerate; the correct translation
ê may not even exist in the current search space for
various reasons, e.g. unknown source word. As a
result, we seek the following three alternatives as
approximations to the ideal objective.
Best v.s. Rest (BR) To score the best hypothesis
in the n-best set ẽ higher than the rest hypotheses. This objective is very similar to
MERT in that it tries to optimize the score
of ẽ and doesn’t concern about the ranking
of rest hypothesis. In this case, the n-best set
Cnbest is used to approximate C, and ẽ to approximate ê.
Best v.s. Worst (BW) To score the best hypothesis higher than the worst hypothesis in the
n-best set. This objective is motivated by the
practice of separating the ”hope” and ”fear”
translation hypothesis (Chiang, 2012). We
take a simpler strategy which uses the best
and worst hypothesis in Cnbest as the ”hope”
and ”fear” hypothesis, respectively, in order
to avoid multi-pass decoding.
Pairwise (PW) To score the better hypotheses
in sampled hypothesis pairs higher than the
worse ones in the same pair. This objective
is adapted from the Pairwise Ranking Optimization (PRO) (Hopkins and May, 2011),
which tries to ranking all the hypotheses instead of selecting the best one. We use the
same sampling strategy as their original paper.
Note that each of the above criterions transforms the original problem of selecting best hypotheses from an exponential space to a certain
pair-wise comparison problem, which could be
easily trained as standard binary classifiers.
4.2 Training Objective
For the binary classification task, we use a hinge
loss following Watanabe (2012). Because the network has a lot of parameters compared with the
linear model, we use a L1 norm instead of L2
norm as the regularization term, to favor sparse solutions. We define our training objective function
in Equation 6.
1 X X
arg min
δ(f , e1 , e2 ; θ) + λ · ||θ||1
θ
N
f ∈D (e1 ,e2 )∈T
with
δ(·) = max{s(f , e1 ; θ) − s(f , e2 ; θ) + 1, 0}
(6)
D is the given training data; (e1 , e2 ) is a training hypothesis-pair, with the assumption that e1 is
the one with higher eval(·) score; N is the total
number of hypothesis-pairs in D; T is the set of
hypothesis-pairs for each source sentence.
The set T is decided by the criterion used for
training. For the BR setting, the best hypothesis is
paired with every other hypothesis in the n-best list
(Equation 7); while for the BW setting, it is only
paired with the worst hypothesis (Equation 8). The
generation of T in PW setting is the same with
PRO sampling, we refer the readers to the original
paper of Hopkins and May (2011).
TBR = {(e1 , e2 )|e1 = arg max eval(e),
e∈Cnbest
(7)
e2 ∈ Cnbest and e1 6= e2 }
TBW = {(e1 , e2 )|e1 = arg max eval(e),
e∈Cnbest
e2 = arg min eval(e)}
e∈Cnbest
(8)
4.3
Algorithm 1 Iterative Training Algorithm
Input: the set of training sentences D, max number of iteration I
1: θ 0 ← RandomInit(),
2: for i = 0 to I do
3:
Ti ← ∅;
4:
for each f ∈ D do
5:
Cnbest ← NbestDecode(f ; θ i )
6:
T ← GeneratePair(Cnbest )
7:
Ti ← Ti ∪ T
8:
end for
i−1
9:
Tall ← WeightedCombine(∪k=0
Tk , Ti )
i+1
i
10:
θ
← Optimize(Tall , θ )
11: end for
Training Procedure
In standard training algorithm for classification, the training instances stays the same in
each iteration. In machine translation, decoding algorithms usually return a very different
n-best set with different parameters. This is
due to the exponentially large size of search
space. MERT and PRO extend the current nbest set by merging the n-best set of all previous iterations into a pool (Papineni et al., 2002;
Hopkins and May, 2011). In this way, the enlarged n-best set may give a better approximation
of the true hypothesis set C and may lead to better
and more stable training results.
We argue that the training should still focus on
hypotheses obtained in current round, because in
each iteration the searching for the n-best set is independent of previous iterations. To compromise
the above two goals, in our practice, training hypothesis pairs are first generated from the current
n-best set, then merged with the pairs generated
from all previous iterations. In order to make the
model focus more on pairs from current iteration,
we assign pairs in previous iterations a small constant weight and assign pairs in current iteration a
relatively large constant weight. This is inspired
by the AdaBoost algorithm (Schapire, 1999) in
weighting instances.
Following the spirit of MERT, we propose a iterative training procedure (Algorithm 1).
As shown in Algorithm 1, the training procedure starts by randomly init model parameters θ 0
(line 1). In ith iteration, the decoding algorithm
decodes each sentence f to get the n-best set Cnbest
(line 5). Training hypothesis pairs T are extracted
from Cnbest according to the training criterion described in Section 4.2 (line 6). New collected pairs
Ti are combined with pairs from previous iterations before used for training (line 9). θ i+1 is obtained by solving Equation 6 using the Conjugate
Sub-Gradient method (Le et al., 2011) (line 10).
5 Structure of the Network
Although neural networks bring strong expressive
power to the modeling of translation hypothesis,
training a neural network is prone to resulting in
local minimum which may affect the training results. We speculate that one reason for these local minimums is the structure of a well-connected
network has too many parameters. Take a neural network with k nodes in the input layer and m
nodes in the hidden layer as an example. Every
node in the hidden layer is connected to each of
the k input nodes. This simple structure resulting
in at least k × m parameters.
In Section 4.2, we use L1 norm in the objective function in order to get sparser solutions. In
this section, we propose some constrained network
structures according to our prior knowledge of the
features. These structures have much less parameters or simpler structures comparing to original
neural networks, thus reduce the possibility of getting stuck in local minimums.
5.1 Network with two-degree Hidden Layer
We find the first pitfall of the standard two-layer
neural network is that each node in the hidden
layer receives input from every input layer node.
Features used in SMT are usually manually designed, which has their concrete meanings. For a
network of several hidden nodes, combining every
features into every hidden node may be redundant
and not necessary to represent the quality of a hypothesis.
As a result, we take a harsh step and constrain
the nodes in hidden layer to have a in-degree of
two, which means each hidden node only accepts
inputs from two input nodes. We do not use any
other prior knowledge about features in this setting. So for a network with k nodes in the input layer, the hidden layer should contain Ck2 =
k(k − 1)/2 nodes to accept all combinations from
the input layer. We name this network structure as
Two-Degree Hidden Layer Network (TDN).
It is easy to see that a TDN has Ck2 × 2 =
k(k − 1) parameters for the hidden layer because
of the constrained degree. This is one order of
magnitude less than a standard two-layer network
with the same number of hidden nodes, which has
Ck2 × k = k2 (k − 1)/2 parameters.
Note that we perform a 2-degree combination
that looks similar in spirit with those combination of atomic features in large scale discriminative learning for other NLP tasks, such as POS tagging and parsing. However, unlike the practice in
these tasks that directly combines values of different features to generate a new feature type, we first
linearly combine the value of these features and
perform non-linear transformation on these values
via an activation function.
5.2
constraints: given a disjoint partition of features:
G1 , G1 , Gk , every hidden node takes input from a
set of input nodes, where any two nodes in this set
come from two different feature groups. We name
this network structure as Grouped Network (GN).
In practice, we divide the basic features in Section 3.3 into five groups: language model features,
translation probability features, lexical probability
features, the word count feature, and the rest of
count features.
6 Experiments and Results
6.1 General Settings
We conduct experiments on a large scale machine
translation tasks. The parallel data comes from
LDC, including LDC2002E18, LDC2003E14,
LDC2004E12, LDC2004T08, LDC2005T10,
LDC2007T09, which consists of 8.6 million
of sentence pairs. Monolingual data includes
Xinhua portion of Gigaword corpus. We use
multi-references data MT03 as training data,
MT02 as development data, and MT04, MT05
as test data. These data are mainly in the same
genre, avoiding the extra consideration of domain
adaptation.
Data
LDC
Gigaword
MT03
MT02
MT04
MT05
Network with Grouped Features
It might be a too strong constraint to require the
hidden node have in-degree of 2. In order to relax this constraint, we need more prior knowledge from the features. Our first observation is
that there are different types of features. These
types are different from each other in terms of
value ranges, sources, importance, etc. For example, language model features usually take a very
small value of probability, and word count feature
takes a integer value and usually has a much higher
weight in linear case than other count features.
The second observation is that features in the
same group are basically of the same type and
may not have complex interaction with each other.
For example, it is reasonable to combine language
model features with word count features in a hidden node. But it may not be necessary to combine
the count of initial phrases and the count of unknown words into a hidden node.
Based on the above two intuitions, we design
a new structure of network that has the following
Usage
TM train
LM train
train
dev
test
test
Sents.
8,260,093
14,684,074
919
878
1,789
1,083
Table 1: Experimental data and statistics.
The Chinese side of the corpora is word segmented using ICTCLAS1. Our translation system
is an in-house implementation of the hierarchical
phrase-based translation system(Chiang, 2005).
We set the beam size to 20. We train a 5-gram
language model on the monolingual data with
MKN smoothing(Chen and Goodman, 1998). For
each parameter tuning experiments, we ran the
same training procedure 3 times and present
the average results.
The translation quality is evaluated use 4-gram case-insensitive
BLEU (Papineni et al., 2002). Significant test
is performed using bootstrap re-sampling implemented by Clark et al. (2011). We employ a twolayer neural network with 11 input layer nodes,
1
http://ictclas.nlpir.org/
Criteria
BRc
BR
BW
PW
MT03(train)
35.02
38.66
39.55
38.61
MT02(dev)
36.63
40.04
39.36
38.85
MT04
34.96
38.73
38.72
38.73
MT05
34.15
37.50
37.81
37.98
Table 2: BLEU4 in percentage on different training criteria (”BR”, ”BW” and ”PW” refer to experiments
with ”Best v.s. Rest”, ”Best v.s. Worst” and ”Pairwise” training criteria, respectively. ”BRc ” indicates
generate hypothesis pairs from n-best set of current iteration only presented in Section 4.3.
corresponding to features listed in Section 3.3 and
1 output layer node. The number of nodes in the
hidden layer varies in different settings. The sigmoid function is used as the activation function for
each node in the hidden layer. For the output layer
we use a linear activation function. We try different λ for the L1 norm from 0.01 to 0.00001 and
use the one with best performance on the development set. We solve the optimization problem with
ALGLIB package2 .
6.2
Experiments of Training Criteria
This set experiments evaluates different training
criteria discussed in Section 4.1. We generate
hypothesis-pair according to BW, BR and PW criteria, respectively, and perform training with these
pairs. In the PW criterion, we use the sampling
method of PRO (Hopkins and May, 2011) and get
the 50 hypothesis pairs for each sentence. We use
20 hidden nodes for all three settings to make a
fair comparison.
The results are presented in Table 2. The
first two rows compare training with and without the weighted combination of hypothesis pairs
we discussed in Section 4.3. As the result suggested, with the weighted combination of hypothesis pairs from previous iterations, the performance
improves significantly on both test sets.
Although the system performance on the dev
set varies, the performance on test sets are almost comparable. This suggest that although the
three training criteria are based on different assumptions, their are basically equivalent for training translation systems.
We also compares the three training criteria in
their number of new instances per iteration and
final training accuracy (Table 3). Compared to
BR which tries to separate the best hypothesis
from the rest hypotheses in the n-best set, and PW
which tries to obtain a correct ranking of all hy2
http://www.alglib.net/
Criteria
BR
BW
PW
Pairs/iteration
19
1
100
Accuracy(%)
70.7
79.5
67.3
Table 3: Comparison of different training criteria
in number of new instances per iteration and training accuracy.
potheses, BW only aims at separating the best and
worst hypothesis of each iteration, which is a easier task for learning a classifiers. It requires the
least training instances and achieves the best performance in training. Note that, the accuracy for
each system in Table 3 are the accuracy each system achieves after training stops. They are not calculated on the same set of instances, thus not directly comparable. We use the differences in accuracy as an indicator for the difficulties of the corresponding learning task.
For the rest of this paper, we use the BW criterion because it is much simpler compared to sampling method of PRO (Hopkins and May, 2011).
6.3 Experiments of Network Structures
We make several comparisons of the network
structures and compare them with a baseline hierarchical phrase-based translation system (HPB)
(Table 4).
We first compares the neural network with different number of hidden nodes. The systems
TLayer20 , TLayer30 and TLayer50 are standard
two-layer feed forward neural network with 20,
30 and 50 hidden layer nodes3 . We can see that
training a larger network do lead to an improvement in translation quality. However training a
larger network is often time-consuming. We experimented with neural networks with 100 and
more hidden nodes (TLayer100 ). But TLayer30
takes 10 times longer in training time for each iter3
TLayer20 is the same system as BW in Table 2
Systems
HPB
TLayer20
TLayer30
TLayer50
TDN
GN
MT03(train)
39.25+
39.55∗
39.70+
39.26
39.60+
39.73+
MT02(dev)
39.07
39.36∗
39.71∗
38.97
38.94
39.41+
MT04
38.81
38.72
38.89
38.72
38.99∗
39.45+
MT05
38.01
37.81
37.90
38.79+
38.13
38.51+
TestAverage
38.41(-)
38.27(-0.14)
38.40(-0.01)
38.76(+0.35)
38.56(+0.15)
38.98(+0.57)
Table 4: BLEU4 in percentage for comparing of systems using different network structures (HPB refers
to the baseline hierarchical phrase-based system. TLayer , TDN, GN refer to the standard 2-layer network, Two-Degree Hidden Layer Network, Grouped Network, respectively. Subscript of TLayer indicates the number of nodes in the hidden layer.) + , ∗ marks results that are significant better than the
baseline system with p < 0.01 and p < 0.05.
ation than TLayer20 and did not finish by the time
of submission deadline.
We then compared the two network structures
proposed in Section 5. The Two-Degree Hidden
Layer Network (TDN) already perform comparable to the baseline system. But it constrain all input to the hidden node to be of degree 2, which is
likely to be too restrictive. With the grouped feature, we could design networks such as GN, which
shows significant improvement over the baseline
systems and achieves the best performance among
all neural systems. Note that GN is in a much
larger scale, but is also sparse in parameters and
takes significant less training time than standard
neural networks.
7 Conclusion
In this paper, we discuss a non-linear framework
for modeling translation hypothesis for statistical machine translation system. We also present
a learning framework including training criterion
and algorithms to integrate our modeling into a
state of the art hierarchical phrase based machine
translation system. Compared to previous effort
in bringing in non-linearity into machine translation, our method uses a single two-layer neural
networks and performs training independent with
any previous linear training methods (e.g. MERT).
Our method also trains its parameters without any
pre-training or post-training procedure. Experiment shows that our method could improve the
baseline system even with the same feature as
input, in a large scale Chinese-English machine
translation task.
In training neural networks with hidden nodes,
we use heuristics to reduce the complexity of network structures and obtain extra advantages over
standard networks. It shows that heuristics and intuitions of the data and features are still important
to a machine translation system.
As future work, it is necessary to integrate more
features into our learning framework. It is also interesting to see how the non-linear modeling fit in
to more complex learning tasks which involves domain specific learning techniques.
References
[Auli and Gao2014] Michael Auli and Jianfeng Gao.
2014. Decoder integration and expected BLEU
training for recurrent neural network language models. In Proceedings of the 52nd Annual Meeting of
the Association for Computational Linguistics, ACL
2014, June 22-27, 2014, Baltimore, MD, USA, Volume 2: Short Papers, pages 136–142.
[Auli et al.2013] Michael Auli, Michel Galley, Chris
Quirk, and Geoffrey Zweig. 2013. Joint language
and translation modeling with recurrent neural networks. In Proceedings of the 2013 Conference on
Empirical Methods in Natural Language Processing, EMNLP 2013, 18-21 October 2013, Grand Hyatt Seattle, Seattle, Washington, USA, A meeting of
SIGDAT, a Special Interest Group of the ACL, pages
1044–1054.
[Bishop1995] Christopher M. Bishop. 1995. Neural
Networks for Pattern Recognition. Oxford University Press, Inc., New York, NY, USA.
[Brown et al.1993] Peter F. Brown, Stephen Della
Pietra, Vincent J. Della Pietra, and Robert L. Mercer. 1993. The mathematic of statistical machine
translation: Parameter estimation. Computational
Linguistics, 19(2):263–311.
[Chen and Goodman1998] S. F. Chen and J. T. Goodman. 1998. An empirical study of smoothing techniques for language modeling. Technical report,
Computer Science Group, Harvard University, Technical Report TR-10-98.
[Chiang2005] David Chiang. 2005. A hierarchical
phrase-based model for statistical machine translation. In annual meeting of the Association for Computational Linguistics.
of the Conference on Empirical Methods in Natural Language Processing, EMNLP ’11, pages 1352–
1362, Stroudsburg, PA, USA. Association for Computational Linguistics.
[Chiang2012] David Chiang. 2012. Hope and fear for
discriminative training of statistical translation models. J. Mach. Learn. Res., 13(1):1159–1187, April.
[Koehn et al.2003] Philipp Koehn, Franz Josef Och,
and Daniel Marcu. 2003. Statistical phrase-based
translation. In HLT-NAACL.
[Clark et al.2011] Jonathan H. Clark, Chris Dyer, Alon
Lavie, and Noah A. Smith. 2011. Better hypothesis testing for statistical machine translation: Controlling for optimizer instability. In Proceedings of
the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies: Short Papers - Volume 2, HLT ’11, pages 176–
181, Stroudsburg, PA, USA. Association for Computational Linguistics.
[Le et al.2011] Quoc V. Le, Jiquan Ngiam, Adam
Coates, Ahbik Lahiri, Bobby Prochnow, and Andrew Y. Ng. 2011. On optimization methods for
deep learning. In Proceedings of the 28th International Conference on Machine Learning, ICML
2011, Bellevue, Washington, USA, June 28 - July 2,
2011, pages 265–272.
[Clark et al.2014] Jonathan Clark, Chris Dyer, and Alon
Lavie. 2014. Locally non-linear learning for statistical machine translation via discretization and structured regularization. Transactions of the Association
for Computational Linguistics, 2:393–404.
[Devlin et al.2014] Jacob Devlin,
Rabih Zbib,
Zhongqiang Huang, Thomas Lamar, Richard M.
Schwartz, and John Makhoul. 2014. Fast and
robust neural network joint models for statistical
machine translation. In Proceedings of the 52nd
Annual Meeting of the Association for Computational Linguistics, ACL 2014, June 22-27, 2014,
Baltimore, MD, USA, Volume 1: Long Papers, pages
1370–1380.
[Duh and Kirchhoff2008] Kevin Duh and Katrin Kirchhoff. 2008. Beyond log-linear models: Boosted
minimum error rate training for n-best re-ranking. In
Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics on Human
Language Technologies: Short Papers, HLT-Short
’08, pages 37–40, Stroudsburg, PA, USA. Association for Computational Linguistics.
[Erhan et al.2009] Dumitru Erhan, Pierre antoine Manzagol, Yoshua Bengio, Samy Bengio, and Pascal
Vincent. 2009. The difficulty of training deep architectures and the effect of unsupervised pre-training.
In David V. Dyk and Max Welling, editors, Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics (AISTATS-09), volume 5, pages 153–160. Journal of Machine Learning
Research - Proceedings Track.
[Gao et al.2014] Jianfeng Gao, Xiaodong He, Wen-tau
Yih, and Li Deng. 2014. Learning continuous
phrase representations for translation modeling. In
Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, ACL 2014,
June 22-27, 2014, Baltimore, MD, USA, Volume 1:
Long Papers, pages 699–709.
[Hopkins and May2011] Mark Hopkins and Jonathan
May. 2011. Tuning as ranking. In Proceedings
[Liu et al.2006] Yang Liu, Qun Liu, and Shouxun Lin.
2006. Tree-to-string alignment template for statistical machine translation. In Proceedings of the 44th
Annual Meeting of the Association of Computational
Linguistics. The Association for Computer Linguistics.
[Liu et al.2013] Lemao Liu, Taro Watanabe, Eiichiro
Sumita, and Tiejun Zhao. 2013. Additive neural
networks for statistical machine translation. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, ACL 2013, 4-9
August 2013, Sofia, Bulgaria, Volume 1: Long Papers, pages 791–801.
[Lu et al.2014] Shixiang Lu, Zhenbiao Chen, and
Bo Xu. 2014. Learning new semi-supervised deep
auto-encoder features for statistical machine translation. In Proceedings of the 52nd Annual Meeting of
the Association for Computational Linguistics (Volume 1: Long Papers), pages 122–132, Baltimore,
Maryland, June. Association for Computational Linguistics.
[Maskey and Zhou2012] Sameer Maskey and Bowen
Zhou.
2012.
Unsupervised deep belief features for speech translation. In INTERSPEECH
2012, 13th Annual Conference of the International
Speech Communication Association, Portland, Oregon, USA, September 9-13, 2012.
[Och and Ney2002] Franz Josef Och and Hermann Ney.
2002. Discriminative training and maximum entropy models for statistical machine translation.
pages 295–302.
[Och2003] Franz Josef Och. 2003. Minimum error rate
training in statistical machine translation. In ACL
’03: Proceedings of the 41st Annual Meeting on Association for Computational Linguistics, pages 160–
167, Morristown, NJ, USA. Association for Computational Linguistics.
[Papineni et al.2002] Kishore Papineni, Salim Roukos,
Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a
method for automatic evaluation of machine translation. In ACL ’02: Proceedings of the 40th Annual
Meeting on Association for Computational Linguistics, pages 311–318, Morristown, NJ, USA. Association for Computational Linguistics.
[Schapire1999] Robert E. Schapire. 1999. A brief introduction to boosting. In Proceedings of the 16th
International Joint Conference on Artificial Intelligence - Volume 2, IJCAI’99, pages 1401–1406, San
Francisco, CA, USA. Morgan Kaufmann Publishers
Inc.
[Vaswani et al.2013] Ashish Vaswani, Yinggong Zhao,
Victoria Fossum, and David Chiang. 2013. Decoding with large-scale neural language models improves translation. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language
Processing, EMNLP 2013, 18-21 October 2013,
Grand Hyatt Seattle, Seattle, Washington, USA, A
meeting of SIGDAT, a Special Interest Group of the
ACL, pages 1387–1392.
[Watanabe2012] Taro Watanabe. 2012. Optimized
online rank learning for machine translation. In
Proceedings of the 2012 Conference of the North
American Chapter of the Association for Computational Linguistics: Human Language Technologies,
NAACL HLT ’12, pages 253–262, Stroudsburg, PA,
USA. Association for Computational Linguistics.
[Yamada and Knight2001] Kenji Yamada and Kevin
Knight. 2001. A syntax-based statistical translation model. In Proceedings of the 39th Annual Meeting of the Association of Computational Linguistics,
pages 523–530.
| 9cs.NE
|
arXiv:1612.08257v1 [math.AC] 25 Dec 2016
Tensor products and direct limits of almost
Cohen-Macaulay modules
Cristodor Ionescu and Samaneh Tabejamaat
December 28, 2016
Abstract
We investigate the almost Cohen-Macaulay property and the Serretype condition (Cn ), n ∈ N, for noetherian algebras and modules. More
precisely, we find permanence properties of these conditions with respect
to tensor products and direct limits.
1
Introduction
All rings considered will be commutative, with unit and noetherian. All modules
are supposed to be finitely generated.
Almost Cohen-Macaulay rings appeared from a flaw in Matsumura’s book
[13] and were first studied by Han [9] and afterwards by Kang [11], who introduced the notion of almost Cohen-Macaulay modules. The first author considered the condition (Cn ), where n is a natural number [10], inspired by the
well-known condition (Sn ) of Serre and characterized almost Cohen-Macaulay
rings using this notion. The notion of a module satisfying the condition (Cn )
was defined and studied by the second author and A. Mafi [17].
We study the behaviour of the condition (Cn ), n ∈ N and of almost CohenMacaulayness with respect to tensor product of A-modules and of A-algebras.
We show that a ring having a module of finite projective dimension satisfying
the condition (Cn ), must satisfy itself the condition (Cn ), a property that was
already proved for the condition (Sn ). We investigate also the behaviour of the
condition (Cn ) and of almost Cohen-Macaulayness to direct limits.
2
Tensor products of almost Cohen-Macaulay
algebras
We begin by recalling the notions and basic facts that will be needed in the
paper.
Definition 2.1 [11, Definition 1.2] Let A be a ring and M be an A-module. We
say that M is an almost Cohen-Macaulay A-module if P ∈ Supp(M ) we have
1
depth(P, M ) = depthAP (MP ). A is called an almost Cohen-Macaulay ring if it
is an almost Cohen-Macaulay A-module.
Remark 2.2 Let (A, m) be a noetherian local ring and M a finitely generated
A-module. Then it follows at once by [11, Lemma 1.5 and Lemma 2.4] that M
is almost Cohen-Macaulay iff dim(M ) ≤ depth(M ) + 1.
Definition 2.3 [17, Definition 2.1] Let A be a ring, n a natural number and
M an A−module. We say that M satisfies the condition (Cn ) if for every
P ∈ Supp(M ) we have depth(MP ) ≥ min(n, htM (P )) − 1. If A satifies the
condition (Cn ) as an A-module, we say that the ring A satisfies (Cn ).
Remark 2.4 Recall that an A−module M is said to satisfy Serre’s condition
(Sn ) if for every P ∈ Supp(M ) we have depth(MP ) ≥ min(n, htM (P )). Hence
obviously if M satisfies the condition (Sn ), it satisfies also the condition (Cn ).
Lemma 2.5 Let n ∈ N and M be an A-module. If M satisfies the condition
(Sn ), then it satisfies the condition (Cn+1 ).
Proof: Let P ∈ Supp(M ). Then by the condition (Sn ) we have depth(MP ) ≥
min(htM (P ), n). Suppose first that n < htM (P ). Then min(htM (P ), n) = n,
hence depth(MP ) ≥ n. But n + 1 ≤ htM (P ), hence min(n + 1, htM (P )) = n + 1.
Then min(n + 1, htM (P )) − 1 = n + 1 − 1 = n. It follows that depth(MP ) ≥ n =
min(n + 1, htM (P )) − 1.
Suppose now that n ≥ htM (P ). Then min(n, htM (P )) = htM (P ), hence
by (Sn ) we have depth(MP ) = htM (P ). But n + 1 > htM (P ), hence min(n +
1, htM (P )) − 1 = htM (P ) − 1 < htM (P ) = depth(MP ).
Lemma 2.6 Let u : A → B be a flat morphism of noetherian rings. If B is
almost Cohen-Macaulay, then all the fibers of u are almost Cohen-Macaulay.
Proof: Let P ∈ Spec(A) and let QBP /P BP ∈ Spec(BP /P BP ). Then the morphism AP → BQ is flat and local and by [10, Proposition 2.2,a)] it follows that
BQ /P BQ = (BP /P BP )QBP /P BP is almost Cohen-Macaulay. From [11, Lemma
2.6] we obtain that BP /P BP is almost Cohen-Macaulay.
Proposition 2.7 Let k be a field, A and B be k−algebras such that A ⊗k B is
noetherian and n ∈ N. Then:
i) If A ⊗k B satisfies the condition (Cn ), then A and B satisfy the condition
(Cn );
ii) If A and B satisfy the condition (Cn ) and one of them satisfies the condition (Sn ), then A ⊗k B satisfies the condition (Cn );
iii) If A and B satisfy the condition (Sn ), then A⊗k B satisfies the condition
(Cn+1 ).
Proof: i) Follows from [10, Proposition 3.12].
ii) Follows from [10, Proposition 3.13].
iii) Follows from [18, Theorem 6,b)] and Lemma 2.5.
2
Corollary 2.8 Let k be a field, A and B be k−algebras such that A ⊗k B is
noetherian. Then:
i) If A ⊗k B is almost Cohen-Macaulay, then A and B are almost CohenMacaulay;
ii) If A and B are almost Cohen-Macaulay and moreover one of them is
Cohen-Macaulay, then A ⊗k B is almost Cohen-Macaulay.
Proof: Follows from 2.7 and [10, Theorem 3.3].
Example 2.9 Let A = k[x4 , x5 , xy, y](x4 ,x5 ,xy,y) . Then A is a noetherian local
domain of dimension 2, hence by remark 2.2 it is almost Cohen-Macaulay. Since
x5 y = x4 xy ∈ x4 A and x5 (xy)3 = x8 y 3 ∈ x4 A, it follows that depth(A) = 1.
But A ⊗k A is not almost Cohen-Macaulay, because by [7, Lemma 2] we get
dim(A ⊗k A) = 4 and depth(A ⊗k A) = 2. The first example with this property
was given by Tabaâ [16, Exemple].
Proposition 2.10 Let u : A → B be a morphism of noetherian rings, M a
finitely generated A-module, N a finitely generated B-module and n a natural
number. Suppose that N is a flat A-module. We consider the structure of
M ⊗A N as a B-module. Then:
i) If M ⊗A N satisfies the condition (Cn ), then M satisfies the condition
(Cn );
ii) If M and NP /P NP satisfy the condition (Sn ), for any P ∈ Spec(A), then
M ⊗A N satisfies the condition (Sn );
iii) If M satisfies the condition (Sn ) and NP /P NP satisfies the condition
(Cn ) for any P ∈ Supp(M ), then M ⊗A N satisfies the condition (Cn );
iv) If M satisfies the condition (Cn ) and NP /P NP satisfies the condition
(Sn ) for any P ∈ Supp(M ), then M ⊗A N satisfies the condition (Cn ).;
v) If M and NP /P NP satisfy the condition (Sn ), for any P ∈ Supp(M ),
then M ⊗A N satisfies the condition (Cn+1 ).
Proof: i) Let P ∈ Supp(M ), Q ∈ Min(P B). By [4, Proposition 1.2.16 and
Theorem A.11], the flatness of u implies that
depthBQ (MP ⊗AP NQ ) = depthAP (MP ) + depthBQ (NQ /P NQ )
and
dimBQ (MP ⊗AP NQ ) = dimAP (MP ) + dimBQ (NQ /P NQ ).
But dimBQ (NQ /P NQ ) = 0 and MP ⊗AP NQ satisfies the condition (Cn ), hence
depthAP (MP ) ≥ min(n, dimAP (MP )) − 1,
that is M satisfies the condition (Cn ).
ii) Let Q ∈ SuppB (M ⊗A N ) and P = Q ∩ A. As above we have
depthBQ (MP ⊗AP NQ ) ≥ min(n, dimAP (MP )) + min(n, dimBQ (NQ /P NQ )) ≥
≥ min(n, dimBQ (MP ⊗AP NQ )).
iii) and iv) The proof is similar to the proof of assertion i).
v) Follow from ii) and lemma 2.5.
3
Proposition 2.11 Let u : A → B be a morphism of noetherian rings, M a
finitely generated A-module, N a finitely generated B-module and n a natural
number. Suppose that N is a flat A-module. We consider the structure of
M ⊗A N as a B-module. Then:
i) If M ⊗A N is almost Cohen-Macaulay, then M and NP /P NP are almost
Cohen-Macaulay for any P ∈ Supp(M );
ii) If M is almost Cohen-Macaulay and NP /P NP is Cohen-Macaulay for
any P ∈ Supp(M ), then M ⊗A N is almost Cohen-Macaulay;
iii) If M is Cohen-Macaulay and NP /P NP is almost Cohen-Macaulay for
any P ∈ Supp(M ), then M ⊗A N is almost Cohen-Macaulay.
Proof: i) From [10, Theorem 3.3] and Proposition 2.10 i), it follows at once that
M is almost Cohen-Macaulay. Let P ∈ SuppA (M ). We have
SuppBP /P BP (NP /P NP ) =
= {QBP /P BP | Q ∈ SuppB (N ) ∩ V (P B), Q ∩ (A \ P ) = ∅ }.
Hence, let QBP /P BP ∈ SuppBP /P BP (NP /P NP ). Since M ⊗A N is an almost
Cohen-Macaulay B−module, by [11, Lemma 2.6] we get that MP ⊗AP NQ is an
almost Cohen-Macaulay BQ −module, as Q ∈ Supp(M ⊗A N ). Since u is flat,
MP is finitely generated and NQ is finitely generated BQ -module and flat AP module, by [12, Proposition 2.7] it follows that NQ /P NQ is an almost CohenMacaulay BQ /P BQ -module. Now by [11, Lemma 2.6] we obtain that NP /P NP
is an almost Cohen-Macaulay BP /P BP -module.
ii) and iii) Follows from 2.10 and [10, Theorem 3.3].
Lemma 2.12 Let R be a commutative ring, A and B be R-algebras such that
A⊗R B is noetherian. Let P ∈ Spec(A⊗R B) and set p := P ∩A, q = P ∩B, r :=
P ∩ R. Assume that Ap is flat over Rr . Then:
i) If (A ⊗R B)P is almost Cohen-Macaulay, then Bq and Ap /rAp are almost
Cohen-Macaulay;
ii) If Ap /rAp and Bq are almost Cohen-Macaulay and one of them is CohenMacaulay, then (A ⊗R B)P is almost Cohen-Macaulay.
Proof: Follows at once from [3, Corollary 2.8].
Corollary 2.13 Let R be a commutative ring, A and B be R-algebras such that
A⊗R B is noetherian. Let P ∈ Spec(A⊗R B) and set p := P ∩A, q = P ∩B, r :=
P ∩ R. Assume that Ap and Bq are flat over Rr . Then:
i) If (A ⊗R B)P is almost Cohen-Macaulay, then Ap and Bq are almost
Cohen-Macaulay;
ii) If Ap and Bq are almost Cohen-Macaulay and one of them is CohenMacaulay, then (A ⊗R B)P is almost Cohen-Macaulay.
Proof: i) Follows from 2.12 and the flatness of Ap and Bq over Rr .
ii) Since Rr → Ap is flat, by [10, Proposition 2.2] it follows that Ap /rAp is
almost Cohen-Macaulay. Note that in case Ap is Cohen-Macaulay, Ap /rAp is
Cohen-Macaulay too. Now the assertion follows from 2.12, ii).
4
Example 2.14 In the previous corollary, the assumption in ii) that one of the
rings is Cohen-Macaulay is necessary. Indeed, consider again the ring in example 2.9, that is A = B = k[x4 , x5 , xy, y](x4 ,x5 ,xy,y) and let R = k[x4 ]. By
[3, Proposition 2.1, 2)], there is P ∈ Spec(A ⊗R B) such that P ∩ A is the
maximal ideal of A. Then by [3, Corollary 2.6] it follows that (A ⊗R B)P is
not almost Cohen-Macaulay. As we saw in 2.9 the ring Ap = Bq is almost
Cohen-Macaulay.
Corollary 2.15 Let R be a commutative ring, A and B be flat R-algebras such
that A ⊗R B is noetherian. Then:
i) If A ⊗R B is almost Cohen-Macaulay, then Ap and Bq are almost CohenMacaulay for any p ∈ Spec(A) and q ∈ Spec(B) such that p ∩ R = q ∩ R;
ii) If for any p ∈ Spec(A) and q ∈ Spec(B) such that p ∩ R = q ∩ R, Ap
and Bq are almost Cohen-Macaulay and one of them is Cohen-Macaulay, then
A ⊗R B is almost Cohen-Macaulay.
Proof: i) By [3, Proposition 2.1] there exists P ∈ Spec(A ⊗R B) such that
P ∩ A = p, P ∩ B = q. Now apply 2.13, i).
ii) Follows from 2.13, ii).
3
Tensor products of almost Cohen-Macaulay
modules
Recall the following definition(cf. [19, Definition 1.1]):
Definition 3.1 If M and N are finitely generated non-zero A-modules, the
grade of M with respect to N, is defined by
grade(M, N ) = inf{i | ExtiA (M, N ) 6= 0}.
Remark 3.2 By [6, Proposition 1.2, h)] we see that grade(M, N ) is the length
of a maximal N -regular sequence in Ann(M ).
Remark 3.3 By [4, Proposition 1.2.10, a)] we get
grade(M, N ) = inf{(depth(NP ) | P ∈ Supp(M )} =
= inf{(depth(NP ) | P ∈ Supp(M ) ∩ Supp(N )}.
Definition 3.4 If M and N are finitely generated A-modules and pd(M ) < ∞,
we say that M is N -perfect if grade(M, N ) = pd(M ).
We will use several times the following fact:
Remark 3.5 (Intersection Theorem cf. [15, Theorem 8.4.4]) If pdA (M ) < ∞,
then dim(N ) ≤ pd(M ) + dim(M ⊗A N ).
5
Proposition 3.6 Let A be a noetherian local ring, M and N be finitely generated non-zero A-modules, n ∈ N. Assume that:
a) M is N −perfect and pd(M ) := p ≤ n;
b) N satisfies the condition (Cn );
c) TorA
i (M, N ) = 0, ∀i > 0.
Then M ⊗A N satisfies the condition (Cn−p ).
Proof: Let P ∈ Supp(M ⊗A N ). We have
p = grade(M, N ) ≤ grade(MP , NP ) ≤ pd(MP ) ≤ pd(M ) = p,
hence pd(MP ) = p. By [2, Theorem 1.2] we obtain
depth(M ⊗A N )P = depth(MP ⊗AP NP ) = depth(NP ) − pd(MP ) =
= depth(NP ) − p ≥ min(htN P, n) − 1 − p = min(htN P − p, n − p) − 1
and it is enough to show that min(htN P − p, n − p) ≥ min(htM⊗A N P, n − p).
But since grade(MP , NP ) = grade((M ⊗A N )P , NP ), by [19, Theorem 2.1] we
have p ≤ htN (P ) − htM⊗A N (P ) and this concludes the proof.
Corollary 3.7 Let n ∈ N, A be a noetherian local ring satisfying the condition
(Cn ) and let M be a perfect A-module such that pd(M ) = p ≤ n. Then M
satisfies the condition (Cn−p ).
Proposition 3.8 Let A be a noetherian local ring and let M and N be finitely
generated non-zero A-modules such that TorA
i (M, N ) = 0, ∀i > 0. If pd(M ) <
∞ and M ⊗A N is almost Cohen-Macaulay, then N is almost Cohen-Macaulay.
Proof: Applying remarks 2.2 and 3.5 and [2, Theorem 1.2] we get
dim(N ) ≤ pd(M ) + dim(M ⊗A N ) ≤
≤ pd(M ) + depth(M ⊗A N ) + 1 = depth(N ) + 1,
hence applying again remark 2.2 it follows that N is almost Cohen-Macaulay.
Corollary 3.9 (see [15, Corollary 8.4.5]) Let A be a local ring having an almost
Cohen-Macaulay module of finite projective dimension. Then A is an almost
Cohen-Macaulay ring.
We can prove a more general form of 3.8:
Proposition 3.10 Let A be a noetherian local ring and n ∈ N. Let also M and
N be finitely generated non-zero A-modules such that TorA
i (M, N ) = 0, ∀i > 0.
If pd(M ) < ∞ and M ⊗A N satisfies the condition (Cn ), then N satisfies the
condition (Cn ).
6
Proof: Let P ∈ SuppA (N ). Assume first that P ∈ SuppA (M ). Then P ∈
Supp(M ⊗A N ) and we have two cases:
a) depth(M ⊗A N )P < n − 1. Then (M ⊗A N )P is almost Cohen-Macaulay and
by Proposition 3.8 it follows that NP is almost Cohen-Macaulay too. Hence
depth(NP ) ≥ min(n, dim(NP )) − 1.
b) depth(M ⊗A N )P ≥ n − 1. Then from[2, Theorem 1.2] we obtain:
depth(NP ) = depth(M ⊗A N )P + pd(MP ) ≥
≥ n − 1 ≥ min(n − 1, htN (P ) − 1) = min(n, htN (P )) − 1.
Assume now that P ∈
/ SuppA (M ). Let Q ∈ Min(AnnA (M ) + P ). By [19, Theorem 2.1] we have
depth(NP ) ≥ grade(AQ /P AQ , NQ ) ≥ depth(NQ ) − dim(AQ /P AQ ).
By remark 3.5 we have
dim(AQ /P AQ ) ≤ pd(MQ ) + dim(MQ /P MQ ) = pd(MQ ),
hence by [2, Theorem 1.2] we obtain
depth(NP ) ≥ depth(NQ ) − pd(MQ ) = depth(MQ ⊗AQ NQ ).
If depth(MQ ⊗AQ NQ ) ≥ n − 1 we have
depth(NP ) ≥ n − 1 ≥ min(n − 1, htN (P ) − 1) = min(n, htN (P )) − 1,
as required.
In the case depth(MQ ⊗AQ NQ ) < n − 1, by definition MQ ⊗AQ NQ is almost
Cohen-Macaulay, hence from Proposition 3.8 it follows that NQ is almost CohenMacaulay. But then NP is almost Cohen-Macaulay too and consequently
depth(NP ) ≥ min(n, htN (P )) − 1.
Corollary 3.11 Let A be a local ring having a module of finite projective dimension satisfying the condition (Cn ). Then A satisfies the condition (Cn ).
Remark 3.12 A similar result for the condition (Sn ) was proved in [20, Proposition 4.1].
4
Direct limits of almost Cohen-Macaulay rings
Proposition 4.1 Let (Ai , fij )i,j∈Λ be a direct system of noetherian rings, n ∈
N and let A := lim Ai . Assume that:
−→
i∈Λ
a) the ring A is noetherian;
b) the morphism fij is flat for any i ≤ j;
c) the ring Ai satisfies the condition (Cn ) for any i ∈ Λ.
Then A satisfies the condition (Cn ).
7
Proof: Let P ∈ Spec(A), B := AP , k := B/P B and for any i ∈ Λ, put Pi :=
P ∩ Ai , Bi := (Ai )Pi and ki := Bi /Pi Bi . There exists i0 ∈ Λ such that P = Pi B
for any i ≥ i0 , hence k = B ⊗Bi ki . But the morphism Bi → B is flat, hence
dim(Bi ) = dim(B) and depth(Bi ) = depth(B) for any i ≥ i0 . The assertion
follows from 2.2.
Corollary 4.2 Let (Ai , fij )i,j∈Λ be a direct system of noetherian rings and let
A := lim Ai . Assume that:
−→
i∈Λ
a) the ring A is noetherian;
b) the morphism fij is flat for any i ≤ j;
c) the ring Ai is almost Cohen-Macaulay for any i ∈ Λ.
Then A is almost Cohen-Macaulay.
Proof: Let us give a direct proof, slightly different to the one coming out at once
from 4.1. We will apply [5, Corollary 2.3]. Let P ⊆ Q be two prime ideals in A
and for any i ∈ Λ let Pi := P ∩ Ai , Qi := Q ∩ Ai , Bi := (Ai )Pi , Ci := (Ai )Qi . Let
also B := AP and C := AQ . Then, as in the proof of 4.1, there exists i0 ∈ Λ such
that P = Pi B and Q = Qi B for any i ≥ i0 . Since Ai is almost Cohen-Macaulay
it follows by [5, Corollary 2.3] that depth(Bi ) ≤ depth(Ci ) and consequently, by
flatness, depth(B) = depth(Bi ) ≤ depth(Ci ) = depth(C). Now we apply again
[5, Corollary 2.3].
Proposition 4.3 Let ((Ai , mi , ki ), fij )i,j∈Λ be a direct system of noetherian local rings and let A := lim Ai . Assume that:
−→
i∈Λ
a) mi Aj = mj , ∀i ≤ j;
b) for any i ∈ Λ and for any ideal I of Ai we have IA ∩ Ai = I;
c) Ai is an almost Cohen-Macaulay ring for any i ∈ Λ.
Then A is almost Cohen-Macaulay.
Proof: By [14] and [8, (0III , 10.3.1.3)], the ring A is a local noetherian ring,
with maximal ideal m = mi A, ∀i ∈ Λ. Let i0 ∈ Λ be such that dim(Aj ) =
dim(A), ∀j ≥ i0 (see [1, Lemma 3.10]) and let x1 , . . . , xs ∈ Ai0 be a system of
parameters. Set Qj = (x1 , . . . , xs )Aj , ∀j ≥ i0 and Q = (x1 , . . . , xs )A. Then
p
p
m = mi0 A = Qi0 A ⊆ Qi0 A ⊆ m,
whence x1 , . . . , xs is a system of parameters in A. The same argument shows that
x1 , . . . , xs is a system of parameters in Aj , ∀j ≥ i0 . If Ai is CM for all i ∈ Λ, it
follows by [1, Corollary 3.7] that A is Cohen-Macaulay and we are done. Assume
not. Since Aj is almost CM, by [11, Theorem 1.7], there exists l ∈ {1, . . . , s}
such that {x1 , . . . , xbl , . . . , xs } is a regular sequence. We assume l = 1, that is
{x2 , . . . , xs } is a regular sequence in Ai0 . As {x1 , . . . , xs−1 , xs } is not a regular
sequence in Ai0 , it cannot be a regular sequence in Aj , ∀j ≥ i0 . But x1 , . . . , xs
is a system of parameters in Aj , ∀j ≥ i0 . We will show that we can assume
that x2 , . . . , xs remains a regular sequence in Aj , ∀j ≥ i0 . If this is so, it follows
that x2 , . . . , xs is a regular sequence in A and again by [11, Theorem 1.7] we
8
get that A is almost Cohen-Macaulay. So, assume that there exists j1 ≥ i0 such
that {x2 , . . . , xs } is not a regular sequence in Aj1 . First we observe that then
{x2 , . . . , xs } is not regular in Ak , ∀k ≥ j1 . Since Aj1 is almost Cohen-Macaulay,
it follows that, for example, {x1 , x
c2 , x3 , . . . , xs } = {x1 , x3 , . . . , xs } is a regular
c2 , x3 , . . . , xs }
sequence in Aj1 . Assume that there exists j2 ≥ j1 such that {x1 , x
is not a regular sequence in Aj2 . Then there exists j2 ≥ j1 such that, for example,
{x1 , x2 , x4 , . . . , xs } is a regular sequence. Continuing we obtain an index l0 ∈ Λ,
such that {x1 , . . . , xi−1 , xbi , xi+1 , . . . , xs } is not a regular sequence in Al , ∀ l ≥ l0
and ∀ i = 1, . . . , s. Contradiction.
Example
4.4 Let K ⊂ L be an infinite algebraic field extension. Then L =
S
Ki , where Ki are the finite field subextensions of L. For any i ∈ Λ, let
i∈Λ
Ai := Ki [[X 4 , X 3 Y, XY 3 , Y 4 ]]. It is easy to see that for any i ∈ Λ we have
depth(Ai ) = 1 and dim(Ai ) = 2, that is Ai is an
ring
S almost Cohen-Macaulay
Ai ( L[[X 4 , X 3 Y, XY 3 , Y 4 ]]
which is not Cohen-Macaulay. Moreover A :=
i∈Λ
and the family (Ai )i∈Λ satisfies the conditions in 4.3. Hence A is an almost
Cohen-Macaulay ring.
Acknowledgements: This paper was partiallly prepared while the first author
was visiting Institut Teknologi Bandung. He thanks this institution for support
and hospitality.
References
[1] M. Asgharzadeh, M. Dorreh and M. Tousi. Direct limits of Cohen-Macaulay
rings. J. Pure Appl. Algebra, 218:1730–1744, 2014. .
[2] M. Auslander. Modules over unramified regular local rings. Illinois J.
Math., 5:631–647, 1961.
[3] S. Bouchiba, J. Conde-Lago and J. Majadas. Cohen-Macaulay, Gorenstein,
Complete Intersection and Regular Defect for the Tensor Product of Algebras, preprint arXiv 1512.0280.
[4] W. Bruns and J. Herzog. Cohen-Macaulay rings, 2nd Edition. Cambridge
Univ. Press, Cambridge, 1998.
[5] L. Chu, Z. Tang and H. Tang. A note on almost Cohen-Macaulay modules.
J. Algebra Appl. 14:1550136, 2015.
[6] H.-B. Foxby. On the µi in a minimal injective resolution. Math. Scand.,
29:175–186, 1971. 175–186.
9
[7] R. Fröberg. A note on the Stanley-Reisner ring of a join and a suspension.
Manuscripta Math., 60:89–91, 1988.
[8] A. Grothendieck. Eléments de géometrie algèbrique, Publ. Math. IHES, 24,
1967.
[9] Y. Han. D-rings. Acta Math. Sinica 4:1047–1052, 1998.
[10] C. Ionescu. More properties of almost Cohen-Macaulay rings. J. Commut.
Algebra 7:363–372, 2015.
[11] M.C. Kang. Almost Cohen-Macaulay modules. Comm. Algebra 29:781–787,
2001.
[12] A. Mafi and S. Tabejamaat. Results on almost Cohen-Macaulay modules.
J. Algebraic Systems 3:147–150, 2016.
[13] H. Matsumura. Commutative Algebra, 2nd Edition. Benjamin Cummings,
Reading Massachusets, 1980.
[14] T. Ogoma. Noetherian property of inductive limits of Noetherian local
rings. Proc. Japan. Acad., Ser. A, Math. Sci., 67:68–69, 1991.
[15] J. R. Strooker. Homological Questions in Local Algebra. London Mathematical Society Lecture Note Series 145, Cambridge Univ. Press, Cambridge,
1990.
[16] M. Tabaâ. Sur le produit tensoriel d’algèbres. Math. Scand. 119:5–13, 2016.
[17] S. Tabejamaat and A. Mafi. About a Serre-type condition for modules. J.
Algebra Appl., 16:1750206, 2017.
[18] M. Tousi and S. Yassemi. Tensor products of some special rings. J. Algebra
268:672–675, 2003.
[19] S. Yassemi, L. Khatami and T. Sharif. Grade and Gorenstein dimension.
Comm. Algebra, 29:5085–5094, 2001.
[20] K. Yoshida. Tensor products of perfect modules and maximal surjective
Buchsbaum Modules. J. Pure Appl. Algebra, 123:313–326, 1998.
Simion Stoilow Institute of
Mathematics of the Romanian
Academy
P. O. Box 1-764
Bucharest 014700
10
Romania
email: [email protected]
and
Department of Mathematics
Payame Noor University
P.O. Box 19395-3697
Tehran
Iran
email: samanetabejamaat [email protected]
11
| 0math.AC
|
SCHEDULING MASSIVELY PARALLEL MULTIGRID FOR
MULTILEVEL MONTE CARLO METHODS
arXiv:1607.03252v1 [cs.CE] 12 Jul 2016
BJÖRN GMEINER∗ , DANIEL DRZISGA† , ULRICH RÜDE∗ , ROBERT SCHEICHL‡ , AND
BARBARA WOHLMUTH†
Abstract. The computational complexity of naive, sampling-based uncertainty quantification
for 3D partial differential equations is extremely high. Multilevel approaches, such as multilevel
Monte Carlo (MLMC), can reduce the complexity significantly, but to exploit them fully in a parallel
environment, sophisticated scheduling strategies are needed. Often fast algorithms that are executed
in parallel are essential to compute fine level samples in 3D, whereas to compute individual coarse
level samples only moderate numbers of processors can be employed efficiently. We make use of
multiple instances of a parallel multigrid solver combined with advanced load balancing techniques.
In particular, we optimize the concurrent execution across the three layers of the MLMC method:
parallelization across levels, across samples, and across the spatial grid. The overall efficiency and
performance of these methods will be analyzed. Here the ”scalability window” of the multigrid solver
is revealed as being essential, i.e., the property that the solution can be computed with a range of
process numbers while maintaining good parallel efficiency. We evaluate the new scheduling strategies
in a series of numerical tests, and conclude the paper demonstrating large 3D scaling experiments.
1. Introduction. Data uncertainties are ubiquitous in many application fields,
such as subsurface flow or climate prediction. Inherent uncertainties in input data
propagate to uncertainties in quantities of interest, such as the time it takes pollutants leaking from a waste repository to reach a drink water well. This situation
has driven the development of novel uncertainty quantification (UQ) methods; most
commonly, using partial differential equations (PDEs) to model the physical processes
and stochastic models to incorporate data uncertainties. Simulation outputs are then
statistics (mean, moments, cumulative distribution function) of the quantities of interest. However, typical sampling-and-averaging techniques for computing statistics
quickly become infeasible, when each sample involves the numerical solution of a PDE.
1.1. Mathematical model and UQ methods. Let us consider an abstract,
possibly nonlinear system of PDEs with uncertain data
M(u; ω) = 0,
(1.1)
where the solution u is sought in some suitable space V of functions v : D ⊂ Rd → Rk
with k ∈ N and D open and bounded, subject to suitable boundary conditions. M is
a differential operator depending on a set of random parameters parametrised by an
element ω of the abstract sample space (Ω, F, P) that encapsulates the uncertainty in
the data, with Ω the set of all outcomes, F the σ-algebra (the “set” of all events),
and P the associated probability measure. As a consequence the solution u itself is a
random field, i.e. u = u(x, ω), with realizations in V .
We are typically only interested in functionals Q(u) ∈ R of u. To compute them
we need to approximate the solution u numerically, e.g. using finite element methods,
which introduces bias error. The cost C typically grows inverse proportionally to some
power of the bias error, i.e. C = O(ε−r ) where ε denotes the bias error tolerance.
This is a challenging computational task that requires novel methodology combined
∗ Institute of System Simulation, University Erlangen-Nuremberg, 91058 Erlangen, Germany
([email protected], [email protected])
† Institute for Numerical Mathematics, Technische Universität München, 85748 Garching, Germany ([email protected], [email protected]), Partly funded by WO671/11-1 (DFG)
‡ Dept. Mathematical Sciences, University of Bath, Bath BA2 7AY, UK ([email protected])
1
with cutting-edge parallel computing for two reasons: firstly, real life applications
lead to PDE systems in three dimensions that often can only be solved effectively
and accurately on a parallel computer (even without data uncertainties); secondly,
typical uncertainties in applications, such as a random diffusion coefficient k(x, ω), are
spatially varying on many scales and cannot be described by a handful of stochastic
parameters. This limits considerably the types of UQ methods that are applicable.
For low dimensional problems, stochastic Galerkin, stochastic collocation and
polynomial chaos methods have been shown to provide efficient and powerful UQ tools
(see, e.g., [13, 42, 26] and the references therein), but in general their complexity grows
exponentially with the stochastic dimension. The cost of sampling methods, such as,
e.g., Monte Carlo, does not grow with the stochastic dimension, but classical Monte
Carlo is notoriously slow to converge. Multilevel Monte Carlo (MLMC) simulation
[14, 6] can help to significantly accelerate the convergence. It has been applied and
extended to a range of applications, see [2, 28, 7, 11, 22].
The idea of MLMC is to reduce algorithmic complexity by performing as much
computational work as possible on coarse meshes. To this end, MLMC uses a hierarchy of discretisations of (1.1) of increasing accuracy to estimate statistics of Q(u) more
efficiently, i.e. using a large number of coarse samples to fully capture the variability,
but only a handful of fine samples to eliminate the bias due to the spatial discretisation. Here, we employ multilevel methods not only to accelerate the stochastic part,
but also to provide a scalable solver for individual realizations of (1.1).
1.2. Parallel methods and algorithms. Current leading-edge supercomputers provide a peak performance in the order of a hundred petaflop/s (i.e. 1017 floating
point operations per second) [33]. However, all these computers draw their computational power from parallelism, with current processor numbers already at Pmax ≈ 107
see [9]. The technological trend indicates that future exascale computers may use
Pmax ≈ 109 . Consequently, designing efficient fast parallel algorithms for high performance computers is a challenging task today and will be even more so in the future.
MLMC methods are characterized by three algorithmic levels that are potential
candidates for parallel execution. As in a standard Monte Carlo method, the algorithm
uses a sequence of classical deterministic problems (samples) that can be computed in
parallel. The size of these subproblems varies depending on which level of resolution
the samples are computed on. We therefore distinguish between parallelism within an
MLMC level and parallelism across MLMC levels. The third algorithmic level is the
solver for each deterministic PDE problem which can be parallelized itself. Indeed, the
total number of samples on finer MLMC levels is typically moderate, so that the first
two levels of parallelism will not suffice to exploit Pmax processors. Parallel solvers
for elliptic PDEs are now able to solve systems with 1.1 × 1013 degrees of freedom
on petascale machines [16] with compute times of a few minutes using highly parallel
multigrid methods [5, 17]. In this paper, we will illustrate for a simple model problem
in three spatial dimensions, how these different levels of parallelism can be combined
and how efficient parallel MLMC strategies can be designed.
To achieve this, we extend the massively parallel Hierarchical Hybrid Grids (HHG)
framework [3, 19] that exhibits excellent strong and weak scaling behavior [21, 1] to
the MLMC setting. We use the fast multigrid solver in HHG to generate spatially
correlated samples of the random diffusion coefficient, as well as to solve the resulting
subsurface flow problems efficiently. Furthermore, the hierarchy of discretisations in
HHG provides the ideal multilevel framework for the MLMC algorithm.
Parallel solvers may not yield linear speedup and the efficiency may deteriorate on
2
a large parallel computer system when the problems become too small. In this case,
too little work can be executed concurrently and the scalar overhead dominates. This
effect is well-known and can be understood prototypically in the form of Amdahl’s law
[21]. In the MLMC context, problems of drastically different size must be solved. In
general, a solver, when applied to a problem of given size, will be characterized by its
scalability window, i.e., the processor range for which the parallel efficiency remains
above an acceptable threshold. Because of memory constraints, the scalability window
will open at a certain minimal processor number. For larger processor numbers the
parallel efficiency will deteriorate until the scalability window closes. In practice,
additional restrictions imposed by the system and the software permit only specific
processor numbers within the scalability window to be used.
MLMC typically leads to a large number of small problems, a small number of
very large problems, and a fair number of intermediate size problems. On the coarser
levels, the problem size is in general too small to use the full machine. The problem is
outside the scalability window and solver-parallelism alone is insufficient. On the other
hand, the efficiency of parallelization across samples and across levels typically does
not deteriorate, since only little data must be extracted from each sample to compute
the final result of the UQ problem. However, on finer levels we may not have enough
samples to fill the entire machine. Especially for adaptive MLMC, where the number
of samples on each level is not known a priori but must be computed adaptively using
data from all levels, this creates a challenging load balancing problem.
A large scalability window of the solver is essential to devise highly efficient execution strategies, but finding the optimal schedule is restricted by a complex combination
of mathematical and technical constraints. Thus the scheduling problem becomes in
itself a high-dimensional, multi-constrained, discrete optimisation problem. Developing suitable approaches in this setting is one of the main objectives of this paper. See
[34, 35] for earlier static and dynamic load balancing approaches.
The paper is structured as follows: In Section 2, we briefly review the MLMC
method and its adaptive version. Section 3 introduces the model problem. Here,
we use an alternative PDE-based sampling technique for Matérn covariances [25, 29]
that allows us to reuse the parallel multigrid solver. In Sections 4 and 5, we define a
classification of different parallel execution strategies and develop them into different
parallel scheduling approaches. In Section 6, we study the parallel efficiency of the
proposed strategies and demonstrate their flexibility and robustness, before finishing
in Section 7 with large-scale experiments on advanced supercomputer systems.
2. The Multilevel Monte Carlo method. To describe the MLMC method,
we assume that we have a hierarchy of finite element (FE) discretisations of (1.1).
Let {V` }`≥0 be a nested sequence of FE spaces with V` ⊂ V , mesh size h` > 0 and
M` degrees of freedom. In the Hierarchical Hybrid Grids (HHG) framework [3, 19],
the underlying sequence of FE meshes is obtained via uniform mesh refinement from
a coarsest grid T0 , and thus h` ' 2−` h0 and M` ' 23` M0 in three space dimensions.
Denoting by u` = u` (x, ω) ∈ V` the FE approximation of u on Level `, we have
M` (u` ; ω) = 0,
` ≥ 0.
(2.1)
Here, the (non)linear operator M` and the functional of interest Q` (u` , ω) may also
involve numerical approximations.
3
2.1. Standard Monte-Carlo Simulation. The standard Monte Carlo (MC)
estimator for the expected value E[Q] of Q(u) on level L ≥ 0 is given by
N
X
b MC,N = 1
Q
Qi ,
L
N i=1 L
(2.2)
where QiL = QL (uiL , ω i ), i = 1, . . . , N , are N independent samples of QL (uL ).
There are two sources of error: (i) The bias error due to the FE approximation.
Assuming that |QiL − Q(ui , ω i )| = O(ML−α ), for almost all ω i and a constant α > 0,
it follows directly that there exists a constant Cb , independent of ML , such that
|E[QL − Q]| ≤ Cb ML−α ≤ εb
(2.3)
for ML ≥ (εb /Cb )1/α (cf. [36]).
(ii) There is a sampling error due to the finite number N of samples in (2.2).
The total error is typically quantified via the mean square error (MSE), given by
2
b MC,N := E[(Q
b MC,N − E[Q])2 ] = (E[QL − Q])2 + N −1 V[QL ],
(2.4)
e Q
L
L
where V[QL ] denotes the variance of the random variable QL (uL ). The first term in
(2.4) can be bounded in terms of (2.3), and the second term in is smaller than a sample
tolerance ε2s if N ≥ V[QL ]ε−2
s . We note that for L sufficiently large, V[QL ] ≈ V[Q].
To ensure that the total MSE is less than ε2 we choose
ε2s = θε2
and ε2b = (1 − θ)ε2 ,
for any fixed 0 < θ < 1.
(2.5)
Thus, to reduce (2.4) we need to choose a sufficiently fine FE mesh and a sufficiently large number of samples. This very quickly leads to an intractable problem
for complex PDE problems in 3D. The cost for one sample QiL of QL depends on the
complexity of the FE solver and of the random field generator. Typically it will grow
like Cc MLγ , for some γ ≥ 1 and some constant Cc , independent of i and of ML . Thus,
b MC,N )2 ≤ ε2 (the ε-cost) is
the total cost to achieve a MSE e(Q
L
b MC,N = O(M γ N ) = O(ε−2−γ/α ).
Cost Q
(2.6)
L
For the coefficient field and for the output functional studied below, we have only
α = 1/6. In that case, even if γ = 1, to reduce the error by a factor 2 the cost
grows by a factor of 28 = 256, which quickly leads to an intractable problem even in
a massively parallel environment.
2.2. Multilevel Monte-Carlo Simulation. Multilevel Monte Carlo (MLMC)
simulation [14, 6, 2] seeks to reduce the variance of the estimator and thus to reduce
computational time, by recursively using coarser FE models as control variates. By
exploiting the linearity of the expectation operator, we avoid estimating E[Q] directly
on the finest level L and do not compute all samples to the desired accuracy (bias
PL
error). Instead, using the simple identity E[QL ] = E[Q0 ] + `=1 E[Y` ], we estimate
the mean on the coarsest level (Level 0) and correct this mean successively by adding
estimates of the expected values of Y` (ω) := Q` (u` , ω) − Q`−1 (u`−1 , ω), for ` ≥ 1.
Setting Y0 := Q0 , the MLMC estimator is then defined as
b ML
Q
L :=
L
X
`=0
4
Yb`MC,N` ,
(2.7)
where the numbers of samples N` , ` = 0, . . . , L, are chosen to minimize the total cost
of this estimator for a given prescribed sampling error (see Eqn. (2.10) below). Note
that we require the FE solutions u` (x, ω i ) and u`−1 (x, ω i ) on two levels to compute
a sample Y`i of Y` , for ` ≥ 1, and thus two PDE solves, but crucially both with the
same ω i and thus with the same PDE coefficient (see Algorithm 1).
Algorithm 1 Multilevel Monte Carlo.
1. For all levels ` = 0, . . . , L do
a. For i = 1, . . . , N` do
i.
Set up (2.1) for ω i on Level ` and ` − 1 (if ` > 0).
ii. Compute u` (ω i ) and u`−1 (ω i ) (if ` > 0), as well as Y`i .
PN` i
Y .
b. Compute Yb MC,N` = 1
`
2.
N`
i=1
`
b ML using (2.7).
Compute Q
L
The cost of this estimator is
b ML ) =
Cost(Q
L
L
X
N` C` ,
(2.8)
`=0
where C` is the cost to compute one sample of Y` on level `. For simplicity, we use
independent samples across all levels, so that the L + 1 standard MC estimators in
b ML simply expands to
(2.7) are independent. Then, the MSE of Q
L
L
2
X
b ML = E[QL − Q] 2 +
e Q
N`−1 V[Y` ] .
L
(2.9)
l=0
This leads to a hugely reduced variance of the estimator since both FE approximations
Q` and Q`−1 converge to Q and thus V[Y` ] → 0, as M`−1 → ∞.
By choosing ML ≥ (εb /Cb )−1/α , we can ensure again that the bias error is less
than εb , but we still have some freedom to choose the numbers of samples N` on each
of the levels, and thus to ensure that the sampling error is less than ε2s . We will
b ML ) in (2.8) subject to the constraint
use this freedom to minimize the cost Cost(Q
L
PL
−1
2
`=0 N` V[Y` ] = εs , a simple discrete, constrained optimization problem with respect to N0 , . . . , NL (cf. [14, 6]). It leads to
!s
L p
X
V[Y` ]
−2
N` = εs
V[Y` ]C`
.
(2.10)
C`
`=0
Finally, under the assumptions that
C` ≤ Cc M`γ
and V[Y` ] ≤ Cv M`−β ,
(2.11)
for some 0 < β ≤ 2α and γ ≥ 1 and for two constants Cc and Cv , independent of i
b ML )2 ≤ ε2 can be bounded by
and of M` , the ε-cost to achieve e(Q
L
−2
b ML
Cost(Q
L ) = εs
L p
X
V[Y` ]C`
!2
≤ CML ε−2−max(0,
γ−β
α
).
(2.12)
`=0
Typically β ≈ 2α for smooth functionals Q(·). For CDFs we typically have β = α.
5
There are three regimes: γ < β, γ = β and γ > β. In the case of the exponential
b ML ) = O(ε−γ/α ), which is a
covariance, typically γ > β and β = 2α and thus Cost(Q
L
full two orders of magnitude faster than the standard MC method. Moreover, MLMC
is optimal for this problem, in the sense that its cost is asymptotically of the same
order as the cost of computing a single sample to the same tolerance ε.
2.3. Adaptive Multilevel Monte Carlo. In Algorithm 2 we present a simple
sequential, adaptive algorithm from [14, 6] that uses the computed samples to estimate
bias and sampling error and thus chooses the optimal values for L and N` . Alternative
adaptive algorithms are described in [15, 7, 11]. For the remainder of the paper we
will restrict to uniform mesh refinement, i.e. h` = 2−` h0 and M` = O(8` M0 ) in 3D.
Algorithm 2 Adaptive Multilevel Monte Carlo.
1. Set ε, θ, L = 1 and N0 = N1 = NInit .
2. For all levels ` = 0, . . . , L do
a. Compute new samples of Y` until there are N` .
b. Compute Yb`MC,N` and s2` , and estimate C` .
3. Update the estimates for N` using (2.14) and
if YbLMC,NL > (8α − 1)εb , increase L → L + 1 and set NL = NInit .
4. If all N` and L are unchanged,
Go to 5.
Else Return to 2.
b ML = PL Yb MC,N` .
5. Set Q
L
`=0 `
To estimate the bias error, let us assume that M` is sufficiently large, so that we
are in the asymptotic regime, i.e. |E[Q` − Q]| ≈ Cb M`−α in (2.3). Then (cf. [11])
1 b MC,N`
Y
.
(2.13)
8α − 1 `
2
PN`
Also, using the sample estimator s2` := N1` i=1
Y`i − Yb`MC,N` to estimate V[Y` ]
and the CPU times from the runs up-to-date to estimate C` , we can estimate
!s
L q
X
s2`
N` ≈ ε−2
s2` C`
.
(2.14)
s
C`
|E[Q` − Q]| ≤
`=0
3. Model problem and deterministic solver. As an example, we consider
an elliptic PDE in weak form: Find u(·, ω) ∈ V := H01 (D) such that
Z
Z
∇v(x) · k(x, ω)∇u(x, ω) dx =
f (x)v(x) dx, for all v ∈ V and ω ∈ Ω. (3.1)
D
D
This problem is motivated from subsurface flow. The solution u and the coefficient
k are random fields on D × Ω related to fluid pressure and rock permeability. For
simplicity, we only consider D = (0, 1)3 , homogeneous Dirichlet conditions and a
deterministic source term f . If k(·, ω) is continuous (as a function of x) and kmin (ω) :=
minx∈D k(x, ω) > 0 almost surely (a.s.) in ω ∈ Ω, then it follows from the LaxMilgram Lemma that this problem has a unique solution (cf. [4]). As quantities of
interest in Section
7, we consider Q(u) := u(x∗ ), for some x∗ ∈ D, or alternatively
R
∂u
1
Q(u) := |Γ| Γ −k ∂n ds, for some two-dimensional manifold Γ ⊂ D can be of interest.
6
3.1. Discretisation. To discretise (3.1), for each ω ∈ Ω, we use standard P1
finite elements on a sequence of uniformly refined simplicial meshes {T` }`≥0 . Let V`
be the FE space associated with T` , N` the set of interior vertices, h` the mesh size
and M` = |N` | the number of degrees of freedom. Now, problem (3.1) is discretised
by restricting it to functions u` , v` ∈ V` . Using the nodal basis {φj : xj ∈ N` } of V`
P
(`)
and expanding u` (·, ω) := j∈N` Uj (ω)φj , this can be written as a linear equation
system where the entries of the system matrix are assembled elementwise based on
on a four node quadrature formula
A(`) (ω)U(`) (ω) = F(`) , where
(`)
Ai,j (ω)
:=
X
τ ∈T`
∇φi · ∇φj
τ
|τ |
4
X
4
k(xτk , ω)
, and
(`)
Fi
Z
:=
f φi dx.
D
k=1
Here xτk , 1 ≤ k ≤ 4 denote the four vertices of the element τ .
The quantity of interest Q(u) is simply approximated by Q(u` ). For Q(u` ) to
converge to Q(u), as ` → ∞, we need stronger assumptions on the random field k.
Let k(·, ω) ∈ C 0,t (D), i.e. Hölder-continuous with coefficient t ∈ (0, 1), and suppose
kmin (ω) and kk(·, ω)kC 0,t have bounded second moments. It was shown in [36] that
−tq/3
E [(Q(u) − Q(u` ))q ] = O htq
=
O
M
, q = 1, 2,
(3.2)
`
`
Hence, the bound in (2.3) holds with α = 3t , and since
X
V [Q(u` ) − Q(u`−1 )] ≤ E (Q(u` ) − Q(u`−1 ))2 ≤ 2
E (Q(u) − Q(ur ))2
r=`,`−1
the bound in (2.11) holds with β = 2α =
2t
3.
3.2. PDE-based sampling for lognormal random fields. A coefficient function k of particular interest is the lognormal random field k(·, ω) := exp(Z(·, ω)), where
Z(·, ω) is a mean-free, stationary Gaussian random field with exponential covariance
|x − y|
E[Z(x, ω)Z(y, ω)] = σ 2 exp −
.
(3.3)
λ
The two parameters in this model are the variance σ 2 and the correlation length λ.
Individual samples k(·, ω) of this random field are in C 0,t (R3 ), for any t < 1/2. In
particular, this means that the convergence rates in (2.3) and (2.11) are α = 1/3 − δ
and β = 2/3−δ in this case, for any δ > 0. The field Z(·, ω) belongs to the larger class
of Matérn covariances [25, 26], which also includes smoother, stationary lognormal
fields, but we will only consider the exponential covariance in this paper.
Two of the most common approaches to realise the random field Z above are
Karhunen-Loeve (KL) expansion [13] and circulant embedding [8, 20]. While the
KL expansion is very convenient for analysis and essential for polynomial expansion
methods such as stochastic collocation, it can quickly dominate all the computational
cost for short correlation lengths λ in three dimensions. Circulant embedding, on the
other hand, relies on the Fast Fourier Transform, which may pose limits to scalability
in a massively parallel environment. An alternative way to sample Z(x, ω) is to
exploit the fact that in three dimensions, mean-free Gaussian fields with exponential
covariance are solutions to the stochastic partial differential equation (SPDE)
(κ2 − ∆)Z(x, ω) =d W (x, ω),
7
(3.4)
where the right hand side W is Gaussian white noise with unit variance and =d
denotes equality in distribution. As shown by Whittle [41], a solution of this SPDE
will be Gaussian with exponential covariance σ 2 = (8πκ)−1 and λ = 2/κ.
In [25], the authors show how this SPDE can be solved using a FE discretisation
and this will be the approach we use to bring our fast parallel multigrid methods to
bear again. Since we only require samples of k(·, ω) = exp(Z(·, ω)) at the vertices of
T` , we discretise (3.4) using again standard P1 finite elements. If now Z`0 (·, ω) ∈ V`0
denotes the FE approximation to Z 0 , then we approximate k(xj , ω) in our qudrature
formula by exp(Z`0 (xj , ω)), for all xj ∈ N` . It was shown in [25, 31] that Z`0 converges
1/3−δ
in a certain weak sense to Z 0 with O(M`
), for any δ > 0. Since (3.4) is in principle
e := (−1, 2)3 ⊃ D
posed on all of R3 we embed the problem into the larger domain D
e (see [31]).
with artifical, homogeneous Neumann boundary conditions on ∂ D
4. Performance parameters and execution strategies. Although MLMC
methods can achieve better computational complexity than standard MC methods,
efficient parallel execution strategies are challenging and depend strongly on the performance characteristics of the solver, in our case a multigrid method. The ultimate
goal is to distribute the Pmax processors to the different subtasks such that the total
run time of the MLMC is minimal. This can be formulated as a high dimensional,
multi-constraint discrete optimization problem. More precisely, this scheduling problem is in general NP-complete, see, e.g., [10, 12, 23, 37] and the references therein,
precluding exact solutions in practically relevant situations.
4.1. Characteristic performance parameters. To design an efficient scheduling strategy, we rely on predictions of the time-to solution and have to take into
account fluctuations. Static strategies thus possibly suffer from a significant load imbalance and may result in poor parallel efficiency. Dynamic strategies, such as the
greedy load balancing algorithms in [34], which take into account run-time data are
more robust, especially when run-times vary strongly within a level.
For the best performance, the number of processors P` per sample on level ` should
lie within the scalability window {P`min , P`min + 1, . . . , P`max } of the PDE solver, where
the parallel efficiency is above a prescribed threshold of, e.g., 80%. Due to machine
constraints, P` may be restricted to a subset, such as {P`min , 2P`min , . . . , 2S P`min },
where S ∈ N0 characterizes the size of the scalability window and P`max = 2S P`min .
Efficient implementations of 3D multigrid schemes such as, e.g., within HHG [3, 18,
19], have excellent strong scalability and a fairly large scalability window, with a
typical value of S = 4 for a parallel efficiency threshold of 80%. The HHG solver has
not only good strong scaling properties but also exhibits excellent weak scalability.
We can thus assume that P`min = 23` P0min and P`max = 23` P0max for PDEs in 3D. The
value of P0min is the number of processors for which the main memory capacity is
fully utilized. Multigrid PDE solvers typically achieve the best parallel efficiency for
P = P`min , when each subdomain is as large as possible and the ratio of computation
to communication is maximal (cf. [17]).
In the following, the time-to solution for the ith sample on level ` executing on
2θ P`min = 23`+θ P0min processors is denoted by t(i, `, θ). We assume that
t(i, `, θ) ≈ C`,θ (ω i ) t`,θ ,
1 ≤ i ≤ N` , 0 ≤ ` ≤ L, 0 ≤ θ ≤ S.
(4.1)
Here, t`,θ is a reference time-to solution per sample on level `. Several natural choices
exist, such as the mode, median, mean or minimum over a sample set. The term
C`,θ (ω i ) encapsulates fluctuations across samples. It depends on the robustness of
8
the PDE solver, as well as on the type of parallel computer system. It is scaled such
that it is equal to one if there are no run-time variations. Fig. 4.1 (right) shows a
typical run-time distribution for 2048 samples each of which was computed on 512
processors with ` = 0 and σ 2 = 0.5 in (3.3).
Assuming no efficiency loss due to load imbalances and an optimal parallel efficiency for θ = 0, the theoretical optimal mean run-time for the MLMC method is
topt
mlmc =
L
P0min X
N` 23` E(C`,0 )t`,0 .
Pmax
(4.2)
`=0
There are three main sources of inefficiency in parallel MLMC algorithms: (i)
a partly idle machine due to large run-time variations between samples scheduled
in parallel, (ii) non-optimal strong scalability properties of the solver, i.e., t`,θ >
2t`,θ−1 , or (iii) over-sampling, i.e., more samples than required are scheduled to fill
the machine. In the following we address (ii) and (iii) in more detail.
The strong parallel efficiency of a solver can be charaterized in terms of Eff` (θ) :=
t`,0 /(2θ t`,θ ). In order to predict t`,θ , 1 ≤ θ ≤ S, we define a surrogate cost function
depending on 0 ≤ θ ≤ S that is motivated by Amdahl’s law [21]:
t`,θ ≈ t`,0 (B + 2−θ (1 − B)),
Eff` (θ) ≈ (2θ B + (1 − B))−1 .
(4.3)
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
10000
B=0
B = 0.01
HHG
B = 0.1
B=1
512
1024
Number of samples
Parallel efficiency
The serial fraction parameter B in (4.3) quantifies the amount of non-parallelizable
work. It can be calibrated from time measurements. For a solver with good scalability
properties, B is almost constant over the levels so that we use a single value on all
levels. Fig. 4.1 (left) shows the typical range of the scalability window, i.e., S = 4, for
` = 0 and P0min = 512. We also see the influence of different serial fraction parameters
B ∈ {0, 0.01, 0.1, 1} on the parallel efficiency and the good agreement of the cost model
(4.3) with averaged measured run-times. The fitted serial fraction parameter B lies in
the range of [0.01, 0.03] for different types of PDE within the HHG framework. In an
adaptive strategy, we can also use performance measurements from past computations
to fit better values of B in the cost predictions for future scheduling steps.
1000
100
10
1
2048
4096
8192
40
41
42
43
44
45
46
47
48
49
50
Time [s]
Number of processors P
Fig. 4.1: Left: Parallel efficiency for different serial fraction parameters B, Right:
Example of a run-time histogram for a multigrid solver using full multigrid-cycles.
Let J` (θ) ∈ N denote the number of samples that can be at most computed
simultaneously on level ` if 23`+θ P0min processors are used per sample, and by k`seq (θ)
we denote the number of required sequential cycles to run in total a minimum of N`
samples. Then, J` (θ), k`seq (θ) and the associated relative load imbalance Imb` (θ) are
9
given by
J` (θ) =
Pmax
,
23`+θ P0min
k`seq (θ) =
N`
,
J` (θ)
Imb` (θ) := 1 −
23`+θ P0min N`
. (4.4)
k`seq (θ)Pmax
We note that 0 ≤ Imb` (θ) < 1, with Imb` (θ) = 0 when no load imbalance occurs. For
Imb` (θ) > 0, part of the machine will be idle either due to the Pmax /(23`+θ P0min ) 6∈ N
or due to N` /J` (θ) 6∈ N.
The remaining processors in the last sequential steps can be used to compute
additional samples that improve the accuracy, but are not necessary to achieve the
required tolerance, or we can schedule samples on other levels in parallel (see the next
section). The product
η` (θ) := (1 − Imb` (θ))Eff` (θ)
(4.5)
will be termed MLMC level efficiency and we note that it also depends on N` .
4.2. Classification of concurrent execution strategies. We classify execution strategies for MLMC methods in two ways, either referring to the layers of
parallelisms or to the resulting time-processor diagram.
4.2.1. Layers of parallel execution. Especially on the finer grid levels in
MLMC, the number of samples is too small to fully exploit modern parallel systems
by executing individual samples in parallel. Multiple layers of parallelism must be
identified. In the context of MLMC methods, three natural layers exist:
Level parallelism: The estimators on level ` = 0, . . . , L may be computed in parallel.
`
Sample parallelism: The samples {Y`i }N
i=1 on level ` may be evaluated in parallel.
Solver parallelism: The PDE solver to compute sample Y`i may be parallelized.
The loops over the levels and over the samples are inherently parallel, except for
some minimal postprocessing to compute the statistical quantities of interest. The
challenge is how to balance the load between different levels of parallelism and how
to schedule the solvers for each sample. Especially in the adaptive setting, without
a priori information, an exclusive use of level parallelism is not always possible, but
in most practical cases, a minimal number of required levels and samples is known a
priori. For the moment, we assume L and N` , 0 ≤ ` ≤ L, to be fixed and given. In
general, these quantities have to be determined dynamically (cf. Alg. 2).
The concurrent execution can now be classified according to the number
PL of layers
of parallelism that are exploited: one, two, or three. Typically, Pmax > l=0 N` and
Pmax NL on modern supercomputers, and thus solver parallelism is mandatory for
large-scale computing. Thus, the only possible one-layer approach on supercomputers
is the solver-only strategy. For a two-layer approach, one can either exploit the solver
and level layers or the solver and sample layers. Since the number of levels L is, in
general, quite small, the solver-level strategy has significantly lower parallelization
potential than a solver-sample strategy. Finally, the three-layer approach takes into
account all three possible layers of parallelism and is the most flexible one.
4.2.2. Concurrency in the processor-time diagram. An alternative way to
classify different parallel execution models is to consider the time-processor diagram,
where the scheduling of each sample Y`i , 1 ≤ i ≤ N` , 0 ≤ ` ≤ L, is represented by
a rectangular box with the height representing the number of processors used. A
parallel execution model is called homogeneous bulk synchronous if at any time in
the processor diagram, all tasks execute on the same level with the same number of
10
processors. Otherwise it is called heterogeneous bulk synchronous. The upper row of
Fig. 4.2 illustrates two examples of homogeneous bulk synchronous strategies, whereas
the lower row presents two heterogeneous strategies.
Level 2
Processors
Processors
synchronization point
Time
Time
synchronization point
synchronization point
Level 2
Processors
Processors
Level 2
Time
Level 0:
Level 1:
Level 2
Time
Fig. 4.2: Upper row: illustration of homogeneous bulk synchronous strategies; onelayer (left) and two-layer parallelism (right); Lower row: illustration of heterogeneous
bulk synchronous strategies. two-layer (left) and three-layer parallelism (right).
The one-layer homogeneous strategy, as shown in Fig. 4.2 (left), offers no flexPL PN`
max
ibility. The theoretical run-time is simply given by
), where
`=0
i=1 t(i, `, θl
max
θ`max is such that Pmax = 23`+θ` P0min . It guarantees perfect load balancing, but
will not lead to a good overall efficiency since on the coarser levels θ`max is typically
significantly larger than S. On the coarsest level we may even have M0 < Pmax , i.e.,
less grid points than processors. Thus we will not further consider this option.
5. Examples for scheduling strategies. Our focus is on scheduling algorithms
that are flexible with respect to the scalability window of the PDE solver and robust
up to a huge number of processors Pmax . To solve the optimization problems, we
will either impose additional assumptions that allow an exact solution, or we will use
meta-heuristic search algorithms such as, e.g., simulated annealing [38, 39]. Before we
introduce our scheduling approaches, we comment briefly on technical and practical
aspects that are important for the implementation.
Sub-communicators. To parallelize over samples as well as within samples, we
split the MPI COMM WORLD communicator via the MPI Comm split command
and provide each sample with its own MPI sub-communicator. This requires only
minimal changes to the multigrid algorithm and all MPI communication routines can
still be used. A similar approach, using the MPI group concept, is used in [35].
Random number generator. To generate the samples of the diffusion coefficient
k(x, ω) we use the approach described in Sect. 3.2. This requires suitable random
numbers for the definition of the white noise on the right hand side of (3.4). For large
scale MLMC computations we select the Ran [30] generator that has a period of ≈ 3.1·
1057 and is thus suitable even for 1012 realizations. It is parallelized straightforwardly
by choosing different seeds for each process, see, e.g., [24].
We consider now examples for the different classes of scheduling strategies.
11
Time
Processors
Processors
Processors
synchronization point
Time
Time
Fig. 5.1: Illustration of different homogeneous scheduling strategies. Left: sample synchronous homogeneous (SaSyHom); Centre: level synchronous homogeneous
(LeSyHom); Right: dynamic level synchronous homogeneous (DyLeSyHom, Sec. 5.3).
5.1. Sample synchronous and level synchronous homogeneous. Here, to
schedule the samples we assume that the run-time of the solver depends on the level
` and on the number of associated processors, but not on the particular sample Y`i ,
1 ≤ i ≤ N` . As the different levels are treated sequentially and each concurrent sample
is executed with the same number of processors, we can test all possible configurations.
Let 0 ≤ ` ≤ L be fixed. Then, for a fixed 0 ≤ θ ≤ S, the total time on level ` is
k`seq (θ)t`,θ . We select the largest index θ` ∈ {0, 1, . . . , S} such that
θ` = arg min k`seq (θ)t`,θ = arg max Eff` (θ)(1 − Imb` (θ)).
0≤θ≤S
0≤θ≤S
Thus the minimization of the run-time per level is equivalent to a maximization of
the total level efficiency. The computation of θ` is trivial provided t`,θ is known for
all θ. We can either set it to be the average of pre-computed timings of the solver on
level ` or we can use (4.3) with a fitted serial fraction parameter B. In that case,
θ` = arg min k`seq (θ)(B + 2−θ (1 − B)).
0≤θ≤S
The level ` only enters this formula implicitly, through N` and through the growth
factor 23` . Given θ` , we can group the processors accordingly and run k`seq (θ` ) sequential steps for each level `. Note that the actual value of t`,0 does not influence
the selection of θ` . It does of course influence the absolute run-time.
We consider two variants: (i) Sample synchronous homogeneous (SaSyHom) imposes a synchronization step after each sequential step (see Fig. 5.1, left). Here statistical quantities can be updated after each step. (ii) Level synchronous homogeneous
(LeSyHom), where each block of 23`+θ P0min processors executes all k`seq (θ` ) without
any synchronization (see Fig. 5.1, centre). Altogether k`seq (θ` )J` (θ` ) ≥ N` samples
are computed. When the run-time does not vary across samples, both strategies will
results in the same MLMC run-time. If it does vary then the LeSyHom strategy has
the advantage that fluctuations in the run-time t(i, `, θ) will be averaged and a shorter
overall MLMC run-time can be expected for sufficiently large k`seq (θ` ).
5.2. Run-time robust homogeneous. So far, we have assumed that the runtime is sample independent, which is idealistic (see Fig. 4.1, right). In the experiment
in Fig. 4.1 (right), 3 out of 2048 samples required a run-time of 50s on P0min = 512
processors. On a large machine with Pmax = 524288 and with θ0 = 0 we need only
k0seq (0) = 2 sequential steps on level 0. Therefore, the (empirical) probability that
the SaSyHom strategy leads
to a runtime of 100s is about 75%, while the theoretical
P2048
2
optimal run-time is 2048
i=1 ti ≈ 90s. Here, ti is the actual run-time of the ith
12
sample from Fig. 4.1 (right). The probability that the LeSyHom strategy leads to a
runtime of 100s is less than 1%; in all other cases, a run-time of ≤ 96s is achieved.
Let us now fix 0 ≤ ` ≤ L again and include run-time variations in the determination of θ` . Unfortunately, in general, run-time distribution functions are not known
analytically, and thus the expected run-time
!
k`seq (θ)
X
(5.1)
E`,θ := E max
t(ijk , `, θ)
1≤j≤J` (θ)
k=1
cannot be computed explicitly. Here, the samples are denoted by ijk with j =
1, . . . , J` (θ) and k = 1, . . . , k`seq (θ), related to their position in the time-processor
diagram. The expression in (5.1) yields the actual, expected run-time on level ` to
compute J` (θ)k`seq (θ) ≥ N` samples with 23`+θ P0min processors per sample when no
synchronization after the sequential steps is performed.
b`,θ for E`,θ , and then to
The main idea is now to compute an approximation E
b
b
b
minimise E`,θ for each `, i.e., to find θ` such that E`,θ` ≤ E`,θ for all 0 ≤ θ ≤ S. As a
first approximation, we replace t(ijk , `, θ) by the approximation C`,θ (ω ijk ) t`,θ in (4.1)
and assume that the stochastic cost factor distribution neither depends on the level
` nor on the scale parameter θ. Furthermore, we approximate the expected value by
an average over µ samples to obtain the approximation
seq
k` (θ)
µ
X
X
1
b`,θ (µ) :=
max
E
C0,0 (ω ijkm ) t`,θ ,
(5.2)
µ m=1 1≤j≤J` (θ)
k=1
where ijkm := ijk + (m − 1)J` (θ)k`seq (θ). If reliable data for t`,θ is available we define
b`,θ (µ)
θ` := arg min E
0≤θ≤S
Otherwise, we include a further approximation and replace t`,θ by B + 2−θ (1 − B) in
b`,θ (µ). Here, we still require an estimate for the
(5.2) before finding the minimum of E
serial fraction parameter B. To decide on the number of samples µ in (5.2), we keep
increasing µ until it is large enough so that µ and µ/2 yield the same θ` . For all our
test settings, we found that µ ≤ 500 is sufficient.
To evaluate (5.2), we need some information on the stochastic cost factor C0,0 (ω)
which was assumed to be constant across levels and across the scaling window. We
use a run-time histogram associated with level ` = 0. This information is either
available from past computations or can be built up adaptively within a MLMC
method. Having the run-times tk , 1 ≤ k ≤ K, of K samples on level ` = 0 at
hand, we emulate C0,0 (ω) by using a pseudo random integer generator from a uniform
discrete P
distribution ranging from one to K and replace the obtained value j ≤ K
K
by tj K/ k=1 tk . Having computed the value of θ` , we proceed as for LeSyHom and
call this strategy run-time robust homogeneous (RuRoHom). For constant run-times,
RuRoHom yields again the same run-times as LeSyHom and as SaSyHom.
5.3. Dynamic variants. So far, we have used pre-computed values for θ` and
k`seq (θ` ) in all variants, and each processor block carries out the computation for exactly k`seq (θ` ) samples. For large run-time variations, this will still lead to unnecessary
inefficiencies. Instead of assigning samples to each processor block a-priori, they can
also be assigned to the processor blocks dynamically at run-time. As soon as a block
13
terminates a computation, a new sample is assigned to it until the required number
N` is reached. This reduces over-sampling and can additionally reduce the total runtime on level `. However, on massively parallel architectures this will only be efficient
when the dynamic distribution of samples does not lead to a significant communication overhead. The dynamic strategy can be combined with either the LeSyHom
or the RuRoHom approach and we denote them dynamic level synchronous homogeneous (DyLeSyHom) and dynamic run-time robust homogeneous (DyRuRoHom),
respectively. Fig. 5.1 (right) illustrates the DLeSyHom strategy. Note specifically
that here not all processor blocks execute the same number of sequential steps.
In order to utilize the full machine, it is crucial that no processor is blocked by
actively waiting to coordinate the asynchronous execution. The necessary functionality may not be fully supported on all parallel systems. Here we use the MPI 2.0
standard that permits one-sided communication and thus allows a non-intrusive implementation. The one-sided communication is accomplished by remote direct memory
access (RDMA) using registered memory windows. In our implementation, we create
a window on one processor to synchronize the number of samples that are already
computed. Exclusive locks are performed on a get/accumulate combination to access
the number of samples.
5.4. Heterogeneous bulk synchronous scheduling. Heterogeneous strategies are clearly more flexible than homogeneous ones, but the number of scheduling
possibilities grows exponentially. Thus, we must first reduce the complexity of the
scheduling problem. In particular, we ignore again run-time variations and assume
t(i, `, θ) = t`,θ . We also assume that N` > 0 on all levels ` = 0, . . . , L. Within an
adaptive strategy, samples may only be required on some of the levels at certain times
and thus this condition has to hold true only on a subset of I := {0, . . . , L}.
In contrast to the homogeneous setups, we do not aim to find scaling parameters θ` that minimize the run-time on each level separately, but instead minimize the
total MLMC run-time. We formulate the minimization process as two constrained
minimization steps that are coupled only in one direction, where we have to identify
the number N`,θ ∈ N0 of samples on level ` which are carried out in parallel with
23`+θ P0min processors, as well as the number k`seq (θ) of associated sequential steps.
Firstly, assuming N`,θ to be given, for all 0 ≤ ` ≤ L, 0 ≤ θ ≤ S, we solve the
constrained minimization problem for k`seq (θ)
S
X
arg seqmin
max t`,θ k`seq (θ) ,
N`,θ k`seq (θ) ≥ N` .
k`
Secondly, having
(θ)∈N0
k`seq (θ)
0≤θ≤S
θ=0
at hand, we find values for N`,θ ∈ N0 such as to minimize
arg min
N`,θ ∈N0
max t`,θ k`seq (θ),
0≤θ≤S
0≤`≤L
the expected run-time, subject to the following inequality constraints
0 ≤ N`,θ ≤ 2−3` 2−θ Pmax /P0min ,
S
X
(5.3a)
N`,θ > 0, for ` ∈ I,
(5.3b)
N`,θ 23` 2θ P0min ≤ Pmax .
(5.3c)
θ=0
L X
S
X
`=0 θ=0
14
We apply integer encoding [32] for the initialization and for possible mutation
operators to guarantee that N`,θ ∈ N0 . Clearly, if N`,θ ∈ N0 then (5.3c) implies
(5.3a). However, even though it is redundant, (5.3a) is enforced explicitly to restrict
the search space in the meta-heuristic optimization algorithm. The condition in (5.3b)
that at least one sample is scheduled on each level at all times could also be relaxed.
However, this would require a redistribution of processors in the optimization problem
and can significantly increase the algorithmic and technical complexity. If (5.3b) is
violated on some level `, we set N`,0 = 1. Condition (5.3c), however, is a hard
constraint. The number of processors that are scheduled cannot be larger than Pmax .
If (5.3c) is violated, we enforce it by a repeated multiplication of N`,S , . . . , N`,0 by
1/2 until it holds. At first glance this possibly leads to an unbalanced work load, but
the applied meta-heuristic search strategy compensates for it. With the values of N`,θ
identified, the samples are distributed dynamically onto the machine, see also [34].
To illustrate the complexity of this optimization task, we consider the number
of different combinations for N`,θ that satisfy (5.3a) but not necessarily (5.3b) and
(5.3c). For example, for L = 3, S = 4, Pmax = 8 192 and P0min = 1, there are
O(1039 ) possible combinations. Even for the special case that the scalability window
degenerates, i.e., that S = 0, there are still O(1010 ) possibilities.
As an example for the following two subsections, we consider (N0 , N1 , N2 , N3 ) =
(4123, 688, 108, 16) and actual run-times from measurements in a set of numerical
experiments:
(t`,θ ) 0≤`≤3,
0≤θ≤4
167
171
=
177
179
83.84
86.28
90.40
91.61
42.30
44.53
47.07
48.27
21.63
23.13
24.21
24.86
11.60
12.41
.
12.97
13.63
(5.4)
5.4.1. The degenerate case S = 0 and a new auxiliary objective. For
S = 0, a cheap but non-optimal way to choose N`,0 is
%
$
Pmax N` t`,0
.
(5.5)
N`,0 = PL
3i min t
i,0
i=0 Ni 2 P0
The corresponding run-time is max`=0,...,L t`,0 dN` /N`,0 e. The total number of proPL
cessors is `=0 N`,0 23` P0min . This choice is acceptable when the workload is evenly
distributed across levels, which is one of the typical scenarios in MLMC. It also requires that the full machine can be exploited without any imbalance in the workload.
Using the first column of (5.4) in (5.5), we find as total run-time 716 s and the
distribution (N0,0 , N1,0 , N2,0 , N3,0 ) = (1314, 221, 36, 5), see the left of Fig. 5.2, while
on the right, the minimal run-time pattern with 684 s is illustrated. Due to weak
scaling effects, the lower levels tend to have a larger number of sequential steps than
the higher ones. We note that there exist many different configurations such that the
minimal run-time is reached. Using (5.5) as starting guess, and then performing a
local adaptive neighborhood search is much cheaper than an exhaustive search.
However for S > 0, we cannot define a good starting guess as easily and have
to resort to meta-heuristic strategies. We consider simulated annealing (SA) techniques, see, e.g., [38, 40], which provide a computationally feasible approach to solve
complex scheduling problems approximately. We start with S = 0. The following
experiments were performed with Python using inspyred 1 with minor modifications.
1 Garrett, A. (2012). inspyred (Version 1.0). Inspired Intelligence Initiative. Retrieved from
http://github.com
15
Level 3
Level 3
Level 2
Level 2
Level 1
Level 1
Level 0
Level 0
Fig. 5.2: Different scheduling patterns: Selection of N`,0 by (5.5) (left) and optimal
choice of N`,0
1050
1000
1000
Timew[seconds]
950
900
850
100
800
750
700
650
10
100
timew[w/owaux.wobj.]
200
400
Numberwofwiterations
timew[w/waux.wobj.]
nupw[w/owaux.wobj.]
Numberwofwunusedwprocessesw(nup)
The temperature parameter in the SA method is decreased using a geometric schedule
Tk+1 = 0.8 Tk . The initial temperature is chosen to be T0 = 103 which is of the order
of the initial changes of the objective function.
Here we choose a Gaussian mutation with distribution N (0, 0.1 Pmax /(23` P0min ))
and a mutation rate of 0.2 guaranteeing that roughly one gene per SA iteration is
changed. All runs were repeated ten times with different seeds and we report minimal (min), maximal (max), as well as the arithmetically averaged (avg) MLMC
run-times. Selecting 1 000 evaluations as the stopping criterion in SA, we obtain
t[min, avg, max] = [684, 691.2, 708] s. For comparison, a stopping criterion of 2 000
evaluations yields t[min, avg, max] = [684, 684, 684] s. Fig. 5.3 shows the evolution of
the average MLMC run-time between iteration 100 and 1 000 in the SA. (Please refer
to the curve labelled “time [w/o aux. obj.]”). Minimizing only the run-time in the SA
objective function, we observe that between iteration 250 and iteration 800 almost no
decrease in the average run-time is achieved. This is due to the rather flat structure
of the objective function in large parts of the search domain resulting from the fact
that many different possible combinations yield identical run-times.
800
nupw[w/waux.wobj.]
Fig. 5.3: Average MLMC run-time and number of unused processors (nup) with and
without auxiliary objective (w/ and w/o aux. obj.) w.r.t. the number of SA iterations.
To improve the performance of the SA scheduling optimizer, we introduce the
number of idle processors as a second auxiliary objective. This auxiliary objective is
only considered, if two candidates result in the same MLMC run-time. In this case the
candidate with the higher number of idle processors is selected. This choice is motivated by the observation that the probability to find a candidate with shorter run-time
is higher in the neighborhood of a candidate that has more idle processors. Fig. 5.3
shows clearly that the optimization can be accelerated by including the auxiliary ob16
jective. Here, we find a MLMC run-time of less than 700 s in less than 300 iterations.
The optimal run-time can be obtained with (N0,0 , N1,0 , N2,0 , N3,0 ) = (1031, 172, 36, 6)
and a total of 7783 processors used. From now on, we always include the number of
idle processors as a secondary objective in the SA optimization algorithm.
5.4.2. The highly scalable case S = 4 and new hybrid mutants. We use
the example data in (5.4) and compare five different mutation operators. In addition to
the already considered Gaussian mutation, we also use simpler and more sophisticated
strategies. Random reset mutation replaces a gene by a uniform randomly chosen
integer satisfying (5.3a). In the case of non-uniform mutation, see [27], a variation
is added to the selected gene and the mutation depends on the SA step. The initial
mutation strength is set to one and decreases with increasing iteration numbers. Tab.
5.1 shows that the Gaussian mutation is superior to both the random reset as well
as the non-uniform mutation. However, even for the Gaussian mutation, more than
50 000 SA iterations are necessary to find average run-times close to the optimal one.
Table 5.1: Comparison of obtained MLMC run-times (min, avg, max) for different
mutation operators and SA iteration numbers.
Mutation
Random reset
Non-uniform
Gaussian
Hybrid A
Hybrid B
1 000
627.6, 676.4, 772.4
641.2, 717.0, 774.6
612.0, 633.9, 641.2
624.6, 632.7, 641.2
603.9, 619.8, 627.3
4 000
624.6, 641.0, 676.8
627.6, 673.2, 717.1
605.2, 624.8, 635.5
603.9, 608.0, 612.0
603.9, 603.9, 603.9
16 000
627.6, 633.5, 659.0
627.6, 647.4, 716.0
603.9, 614.6, 624.6
604.5, 604.5, 604.5
603.9, 603.9, 603.9
64 000
603.9, 616.4, 632.8
627.6, 638.5, 641.2
603.9, 608.0, 612.0
604.5, 604.5, 604.5
603.9, 603.9, 603.9
Thus, new problem-adapted mutation operators in the SA are essential. We
propose two new hybrid variants. Both perform first a Gaussian mutation and then a
problem adapted mutation, taking into account the required processor numbers. The
mutation rate for both is set to 0.1.
Hybrid A: In each step, we select randomly two different “genes” N`1 ,θ1 and N`2 ,θ2 ,
0 ≤ `1 , `2 ≤ L, 0 ≤ θ1 , θ2 ≤ S, as well as a uniformly distributed random number
k ∈ [0, .., N`1 ,θ1 − 1]. Then we mutate
N`1 ,θ1 = N`1 ,θ1 − k
j
k
and N`2 ,θ2 = N`2 ,θ2 + k 2θ1 −θ2 23(`1 −`2 ) .
If the original values for N`1 ,θ1 and N`2 ,θ2 were admissible, satisfying the constraints
(5.3), then the mutated genes are also admissible. This type of mutation exploits
the scalability window of the solver as well as level parallelism. For the special case
S = 0, it reduces to balancing the workload on the different levels, by exploiting the
weak scalability of the solver.
Hybrid B. This variant is proposed for a PDE solver that has a large scalability
window. It follows the same steps, but keeps `1 = `2 fixed, therefore only exploiting
the strong and weak scaling properties of the solver, but not the MLMC hierarchy.
In Tab. 5.1, we see that Hybrid B shows the best performance. Compared to
Hybrid A it is less sensitive to the initial guess and robustly finds a very efficient
scheduling scheme in less than 4 000 SA iterations. Thus, we restrict ourselves to SA
with Hybrid B type mutations in the following examples. In the example considered
17
in this section, it leads to the schedule
0 443 73 0 0
1 98
0 0 0
(N`,θ ) 0≤`≤3, =
0
0≤θ≤4
0
3 3 0
6
0
0 0 0
seq
(k`,θ
) 0≤`≤3,
0≤θ≤4
0
3
=
0
3
7
7
0
0
14
0
12
0
0
0
24
0
0
0
. (5.6)
0
0
Comparing the two cases S = 0 and S = 4 shows how important the strong scalability
of the solver is to reach shorter MLMC run-times. It allows to reduce the runtime by more than 10%, and thus the parallel MLMC performance can be improved
significantly with such an advanced scheduling strategy. In the following, we call the
scheduling strategy StScHet, if strong scaling is included (S > 0). Otherwise, if no
strong scaling is included (S = 0), we call the scheduling strategy noStScHet.
To finish this section we summarise all the considered schedules in Tab. 5.2.
Table 5.2: Summary of parallel scheduling strategies.
Abbreviation
SaSyHom
LeSyHom
RuRoHom
DyLeSyHom
DyRuRoHom
StScHet
noStScHet
Schedule
Sample Synchronous Homogeneous
Level Synchronous Homogeneous
Run-Time Robust Homogeneous
Dynamic Level Synchronous Homogeneous
Dynamic Run-Time Robust Homogeneous
Heterogeneous with Strong-Scaling (S > 0)
Heterogeneous without Strong-Scaling (S = 0)
Defined in
Sec. 5.1
Sec. 5.1
Sec. 5.2
Sec. 5.3
Sec. 5.3
Sec. 5.4
Sec. 5.4
6. Scheduling comparison. In this section, we evaluate the sampling strategies
from the previous section and illustrate the influence of the serial fraction parameter
B, of the level-averaged number of sequential steps and of the run-time variation.
6.1. The influence of the number of sequential steps. The fact that processor and sample numbers have to be integer not only complicates the solution of
the optimization problem, it also strongly influences the amount of imbalance.
Let us start with some preliminary considerations and assume that there are no
run-time P
variations. Now, let ∆t ≥ 0 denote the relative difference ∆t between the
L
run-time `=0 k`seq (θ` )t`,θ` of the presented homogeneous strategies and the theoretically optimal run-time in (4.2) (with E(C`,0 ) = 1). Using the MLMC level efficiency
we can quantify ∆t as
PL
PL
N` 23` t`,0 (η` (θ` ))−1
Pmax `=0 k`seq (θ` )t`,θ`
− 1 = `=0PL
−1
∆t = min PL
3`
3`
P0
`=0 N` 2 t`,0
`=0 N` 2 t`,0
For the special case that Pmax /(23`+θ P0min ) ∈ N, we can further bound ∆t in terms of
PL
kseq := `=0 N` 23` P0min /Pmax . We assume that t0,0 ≤ t0,` ≤ t0,L , for all ` = 0, . . . , L,
which is typically the case. The ratio tL,0 /t0,0 reflects the weak scalability of the solver.
Peta-scale aware massively parallel codes have a factor close to one. Recall from (5.4)
that for our solver tL,0 /t0,0 = 179/167 ≈ 1.07. Since k`seq (θ` )t`,θ` ≤ k`seq (0)t`,0 and
P min
0
since k`seq (0) ≤ N` 23` Pmax
+ 1 we have
L+1
max0≤`≤L t`,0 Pmax
tL,0 L + 1
=
.
∆t ≤
P
3`
min0≤`≤L t`,0 P0min L
t0,0 kseq
`=0 N` 2
18
The larger kseq , the smaller the efficiency loss.
700
Time [seconds]
600
500
LeSyHom
400
noStScHet
300
StScHet
200
Optimal
100
0
0.2 0.4 0.6 0.8 0.9 1.1 1.3 1.5 1.7 1.9 2.1 2.3 2.5 2.6 2.8 3.0 3.2 3.4
Nseq
Fig. 6.1: Heterogeneous versus homogeneous scheduling for kseq ∈ [0.2, 3.4].
In Fig. 6.1 we compare LeSyHom, noStScHet, StScHet and increase kseq from 0.2
to 3.4. All strategies stay within the theoretically predicted upper bound. The two
scheduling strategies, LeSyHom and StScHet, that exploit the scaling properties of
the solver are significantly more robust with respect to kseq than the heterogeneous
strategy, noStScHet, for which we ignore the scalability window and set S = 0. This
observation is particularly relevant for adaptive MLMC strategies where N` may be
increased within any of the adaptive steps and then a new optimal scheduling pattern has to be identified. For noStScHet, we observe a staircase pattern that is a
direct consequence of the ceil operator. This effect can be easily counterbalanced by
exploiting the scalability window of the solver. Moreover the run-times for LeSyHom
and StScHet are larger than the optimal one by roughly a factor of 1.5 for kseq = 0.2,
but only by a factor of 1.15 for kseq = 3.4. Thus both these strategies are robust and
efficient with respect to variations in kseq .
6.2. The influence of solver scalability. The serial fraction parameter B
models the strong scaling of the solver, see Sec. 5. The higher B, the less beneficial
it is to increase θ. In Fig. 6.2, we consider the influence of B on the run-time for two
different values of kseq , namely 0.75 and 3, and compare LeSyHom and StScHet.
Time [seconds]
700
600
500
400
300
200
LeSyHom (kseq ≈ 3)
LeSyHom (kseq ≈ 0.75)
StScHet (kseq ≈ 3)
StScHet (kseq ≈ 0.75)
100
0
0.001
0.01
0.1
1
Serial fraction (B)
Fig. 6.2: Influence of the serial fraction parameter on the run-time.
First, we consider the case kseq ≈ 0.75. With a serial fraction parameter B ≤ 0.02,
there is almost no run-time difference between the two strategies. For B up to 0.1
the run-time difference is below 25%. However, for larger B, the run-time increases
significantly for LeSyHom. This can be explained by the fact that for large B, the
strong scalability property of the solver is too poor to obtain a robust scheduling
pattern, and only a heterogeneous strategy with its flexibility to schedule in parallel
samples on different levels can guarantee small run-times. Homogeneous strategies
19
provide enough flexibility to be efficient in the case of large scalability windows with
small values of B. For a small value of kseq , the run-time of StScHet depends only
very moderately on the serial fraction parameter B.
The situation is different for larger values of kseq . Then both strategies exhibit
roughly the same performance, but the total run-time is more sensitive to the size of
B. A good strong scaling of the PDE solver can improve the time to solution by up
to 27% for the homogeneous and up to 21% for the heterogeneous bulk synchronous
case. As expected, carrying out one synchronization step with kseq ≈ 3 is more
efficient than four steps with kseq ≈ 0.75. This observation is important for the
design of efficient adaptive strategies, i.e., they should not be too fine granular. For
highly performant multigrid solvers, i.e., B ≤ 0.05, the much simpler homogeneous
strategies are an excellent choice, in particular for kseq ≥ 1. On the other hand, when
the parallel performance of the solver is poorer, which is typically the case in the
peta-scale regime, i.e. near the strong scaling limit of the multigrid solver, the more
complex heterogeneous strategies lead to significantly better efficiency gains.
6.3. Robustness and efficiency with respect to the parameters. In this
subsection, we modify all three key parameters that we have discussed so far. We
assume again that the run-time variations C`,θ (·) are independent of ` and θ and
use a half-normal distribution to model C0,0 (·). More specifically, we assume that
C0,0 (·) − 1 follows a half-normal distribution with parameter Var, i.e. its mode is at 1.
The time t`,θ is chosen to be the run-time of the mode, as described in Sec. 4.1.
StScHet
SaSyHom
LeSyHom
DyLeSyHom
RuRoHom
DyRuRoHom
1
0.9
Parallel efficiency
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
kseq
B
V ar
1 5 24
0.01
1 5 24
0.1
0
1 5 24
1
1 5 24
0.01
1 5 24
0.1
0.5
1 5 24
1
1 5 24
0.01
1 5 24
0.1
2
1 5 24
1
Fig. 6.3: MLMC efficiencies for different values of the parameters kseq , B and Var,
and for all the different scheduling strategies.
Fig. 6.3 illustrates the parallel efficiencies of all the strategies developed above (cf.
Tab. 5.2), as well as their robustness with respect to the parameters kseq , B and Var,.
The parallel efficiency is calculated with respect to the theoretical, optimal run-times
given in (4.2) not with actual measured run-times. We choose B ∈ {0.01, 0.1, 1}
and Var ∈ {0, 0.5, 2}, and set the sample numbers on the different levels to be
(N0 , N1 , N2 , N3 ) = kseq (1366, 228, 36, 5), with kseq ∈ {0.98, 4.92, 23.60} (abbreviated
by {1, 5, 24} in Fig. 6.3).
20
We comment first on the case of no run-time variations, i.e., Var = 0, where our
numerical results confirm that all homogeneous strategies produce the same performance. For large numbers of sequential steps, the homogeneous variants are superior
to the heterogeneous ones. This is mainly due to the constraint (5.3b) which forces
us to consider all levels in parallel. As mentioned above, this constraint is not essential and dropping it might lead to more efficient heterogeneous strategies. This
will be the subject of future work. If variations in the run-time are included, then all
homogeneous strategies yield different results. The parallel efficiency of the simplest
one, SaSyHom, then drops to somewhere between 0.1 and 0.55. As expected, the
worst performance is observed for a small kseq , poor solver scalability, and high runtime variation. In that case, the dynamic variants can counterbalance the run-time
variations more readily and provide computationally inexpensive scheduling schemes
(provided the technical realization is feasible).
Secondly, we discuss the case of a small value of kseq . This typically occurs if
the machine is large or if the adaptive MLMC algorithm is used. Here, only the
heterogeneous strategies can guarantee acceptable parallel efficiencies for all values of
B. The homogeneous variants result in efficiencies below 0.7 and 0.4, for Var = 0 and
for B = 0.1 and B = 1, respectively.
In all considered cases, one of our strategies results in parallel efficiencies of
more than 0.5; in many cases even more than 0.7. For moderate run-time variations
and large enough kseq , the parallel efficiency of StScHet improves to more than 0.8.
StScHet is also the most robust strategy with respect to solver scalability. However,
for solvers with good scalability, i.e. B ≤ 0.05, the DyRoRuHom strategy is an attractive alternative, since it does not require any sophisticated meta-heuristic scheduling
algorithm and can dynamically adapt to run-time variations in the samples.
7. Numerical results for MLMC. In this section, the scheduling strategies
developed above are employed in a large-scale MLMC computation. We consider the
model problem in Sec. 3 with D = (0, 1)3 and f ≡ 1, discretised by piecewise linear
FEs. For the relevant problem sizes, the serial fraction parameter for our multigrid
PDE solver is B ≤ 0.02, and the fluctuations in run-time are < 2%. Only few timings
deviate substantially from the average (c.f. Fig. 4.1) so that we focus on investigating
strategies for that regime.
The following experiments were carried out on the peta-scale supercomputer
JUQUEEN, a 28 rack BlueGene/Q system located in Jülich, Germany2 . Each of
the 28 672 nodes has 16 GB main memory and 16 cores operating at a clock rate of
1.6 GHz. The compute nodes are connected via a five-dimensional torus network.
HHG is compiled by the IBM XL C/C++ Blue Gene/Q, V12.0 compiler suite with
MPICH2 that implements the MPI-2 standard and supports RDMA. Four hardware
threads can be used on each core to hide latencies. We always use 2 processes (threads)
per core to maximize the execution efficiency.
7.1. Static scheduling for scenarios with small run-time variations. We
choose four MLMC levels, i.e., L = 3, with a fine grid that has roughly 1.1 · 109
mesh nodes. The random coefficient is assumed to be lognormal with exponential
covariance, σ 2 = 1 and λ = 0.02. The quantity of interest is the PDE solution u
evaluated at the point x = (0.25, 0.25, 0.25). All samples are computed using a fixed
multigrid cycle structure with one FMG-2V(4,4) cycle, i.e., a full multigrid method
2 http://www.fz-juelich.de/ias/jsc/EN/Expertise/Supercomputers/JUQUEEN/JUQUEEN
21
node.html
(nested iteration) with two V-cycles per new level, as well as four pre- and four postsmoothing steps. In [19] it is shown that this multigrid method delivers the solution
of a scalar PDE with excellent numerical and parallel efficiency. In particular, the
example is designed such that after completing the FMG-2V(4,4) cycle for all samples, the minimal and maximal residual differ at most by a factor 1.5 within each
MLMC level. For the MLMC estimator, an a priori strategy is assumed, based on
pre-computed variance estimates, such that (N` )`=1,2,3,4 = (4 123, 688, 108, 16). We
first study the balance between sample and solver parallelism and thus the tradeoffs between the efficiency of the parallel solver and possible load imbalances in the
sampling strategy, as introduced in Sec. 4.1. We set P0min = 1, and consequently
P`min = 23` . The run-times to compute a single sample with P`min processors are
measured as (t`,0 )`=0,1,2,3 = (166, 168, 174, 177) seconds, showing only a moderate
increase in runtime and confirming the excellent performance of the multigrid solver.
A lower bound for the run-time of the parallel MLMC estimator of topt
mlmc = 520 s
is now provided by eq. (4.2). A static cost model is justified since the timings between
individual samples vary little. We therefore employ the level synchronous homogeneous (LeSyHom) scheduling strategy, as introduced in Sect. 5.1. This requires only a
few, cheap real time measurements to configure the MLMC scheduling strategy. The
smaller θ, the larger the solver efficiency Eff` (θ)while the larger θ, the smaller Imb` (θ).
To study this effect quantitatively, we measure the parallel solver efficiency. Here, we
do not use (4.3), but actual measured values for t`,θ instead, and we find that on level
` = 0 we have {Eff0 (0), Eff0 (1), Eff0 (2), Eff0 (3), Eff0 (4)} = {1, 0.99, 0.96, 0.92, 0.86}.
In Tab. 7.1, we present for each level the efficiency η` (θ), see (4.5), and the runtime as a function of θ. For θ = 0 and ` = 0, the runs are carried out with one thread,
and they go up to 8 192 hardware threads on 4 096 cores for θ = 4 and ` = 3. We see
a very good correlation between predicted efficiencies and actual measured times in
Tab. 7.1. The maximal efficiency and the minimal run-time on each level are marked
in boldface to highlight the best setting.
Table 7.1: Level and total run-time and efficiency of MLMC for fixed θ.
θ
0
1
2
3
4
time
167
168
127
108
104
η0 (θ)
0.50
0.50
0.64
0.74
0.77
time
171
173
134
139
136
η1 (θ)
0.67
0.67
0.86
0.83
0.84
time
177
181
188
169
181
η2 (θ)
0.84
0.84
0.81
0.89
0.81
time
179
183
193
199
218
η3 (θ)
1.00
0.99
0.96
0.92
0.86
time
694
704
642
615
640
η(θ)
0.75
0.74
0.81
0.85
0.83
Keeping θ fixed, the minimal run-time is 615. Our homogeneous scheduling strategies pick θ` for each level automatically and by doing so, a considerably shorter runtime of 586 is obtained, increasing the efficiency to 0.89. If it were possible to scale
the solver perfectly also to processor numbers that are not necessarily powers of 2, we
could reduce the compute time even further by about 11% from 586 to 520 seconds.
In summary, we see that exploiting the strong scaling of the PDE solver helps to avoid
load imbalances due to oversampling and improves the time to solution by about 15%
from 694 to 586 seconds. The cost is well distributed across all levels, although most
of the work is on the finest level which is typical for this model problem (cf. [4]).
We conclude this subsection with a strong scaling experiment, i.e., we increase
the number of processes in order to reduce the overall time to solution. Since we are
22
1024
1
Time][seconds]
0.8
256
0.7
128
0.6
0.5
64
0.4
32
0.3
0.2
16
MLMC]parallel]efficiency
0.9
512
0.1
8
0
2]048
4]096
Optimal
8]192
16]384
32]768
Number]of]processes]Pmax
MLMC
65]536
131]072
Parallel]efficiency
Fig. 7.1: Strong scaling of MLMC using homogeneous bulk synchronous scheduling.
interested to analyse the behavior for extremely large Pmax , we reduce the number
of samples to Nl = (1 031, 172, 27, 4). The time for one MLMC computation with an
increasing number of processes Pmax is presented in Fig. 7.1. The initial computation
employs Pmax = 2 048 which is large enough so that all fine-grid samples can be
computed concurrently. We scale the problem up to 131 072 processes. Increasing
Pmax while keepimg the number of samples per level fixed results in a decrease of
kseq . Thus the load imbalance increases and the total parallel efficiency decreases.
Nevertheles even with Pmax = 32768, we obtain a parallel efficiency over 60% while
for Pmax = 131072 the eficiency drops below 40%. This can be circumvented by an
increase of the size S of the sclability window. Overall the compute time for the
MLMC estimator can be reduced from 616 to 22 seconds. For each choice of Pmax ,
we select the optimal regime for θ` , ` = 0, . . . , 3, as discussed above. Together with
the excellent strong scaling behavior of the parallel HHG multigrid solver this leads
to the here demonstrated combined parallel efficiency of the MLMC implementation.
7.2. Adaptive MLMC. Finally, we consider an adaptive MLMC algorithm as
introduced in Sec. 2.3 in a weak scaling scenario, i.e., increasing the problem size
proportionally to the processor count. Each row in Tab. 7.2 summarizes one adaptive
MLMC computation. The MLMC method is initially executed on 2 048 cores and
P` is chosen as P`=1,2,3,4 = (2, 16, 128, 1 024), In each successive row of the table, the
number of unknowns on the finest level in MLMC and the number of processors on
each level is increased by a factor of eight. Moreover, the correlation length λ of the
coefficient field is reduced by a factor of two (σ 2 = 1 is kept fixed). This means that
the problems are actually getting more difficult as well. The quantity of interest is
defined as the flux across a separating plane Γ at x2 = 0.25, i.e.,
Z
∂u
Q(u, ω) =
k(x, ω) ds .
∂n
Γ
In all cases, the initial number of samples is set to N`=1,2,3,4 = (1 024, 256, 64, 16).
The final number of samples is then chosen adaptively by the MLMC algorithm. It
is listed in the table. As motivated in Sec. 2, the tolerance for the sampling error,
which is needed in (2.14) to adaptively estimate N` , is chosen as εs ≈ |E[QL − QL−1 ]|,
balancing the sampling error with the bias error. The estimates for the expected
values and for the variances of Q` and Y` , for a problem of size M` = 1 0243 and with
a correlation length of λ = 0.015, are plotted in Fig. 7.2. The expected values and the
23
Table 7.2: Weak scaling of an adaptive MLMC estimator.
Processes
4 096
32 768
262 144
Resolution
1 0243
2 0483
4 0963
No. Samples
Fine
Total
68 13 316
44 10 892
60 10 940
Runtime
5.0 · 103 s
3.9 · 103 s
5.2 · 103 s
Correlation
length
1.50E-02
7.50E-03
3.75E-03
Idle
time
3%
4%
5%
Table 7.3: Number of samples and over-samples for different levels for the largest run.
Level
0
1
2
3
No. partitions
2 048
256
32
4
No. Samples
Scheduled Calculated
7 506
8 192
2 111
2 304
382
384
57
60
1
1 × 10−3
1 × 10−1
1 × 10−4
1 × 10−2
1 × 10−5
1 × 10−3
1 × 10−6
1 × 10−4
1 × 10−7
1 × 10−5
No. Over-samples
Estimated Actual
3 726
686
429
193
15
2
3
3
V ar(Q)
1 × 10−8
E(Q)
E(Y )
1 × 10−6
Level 0
V ar(Y )
1 × 10−9
1
2
Level 0
3
1
2
3
Fig. 7.2: MLMC performance plots: expected value (left) and variance (right) of Q`
(red, solid) and Y` (blue, dashed) for λ = 0.015 and σ 2 = 1.
variances of Y` show the expected asymptotic behavior as ` increases, confirming the
benefits of the multilevel approach. The total number of samples that are computed
is 13 316, but only 68 of them on the finest grid. A standard Monte Carlo estimator
would require several thousand samples on level 3 and would be significantly more
costly. The idle time, in the last column of Tab. 7.2, accounts for the variation in
the number of V-cycles, required to achieve a residual reduction of 10−5 on each level
within each call to the FMG multigrid algorithm.
The largest adaptive MLMC computation shown in Tab. 7.2 involves a finest
grid with almost 7 × 1010 unknowns. Discrete systems of this size must be solved 60
times, together with more than 10 000 smaller problems, the smallest of which still
has more than 1.6 × 107 unknowns. With the methods developed here, a computation
of such magnitude requires a compute time of less than 1.5 hours when 131 072 cores
running 262 144 processes are employed. Additional details for this largest MLMC
computation are presented in Tab. 7.3. The table lists the number of partitions that
are used on each level for the respective problem sizes. The number of calculated
samples on each level is a multiple of the number of these partitions. As Tab. 7.3
illustrates, the number of scheduled samples is smaller and the difference indicates the
24
amount of oversampling. The number of unnecessary samples is presented explicitly
in the last column of the table to compare with the estimated number of unneeded
samples. This estimated number is significantly higher on each level, since it is the sum
of all the oversampled computations in all stages of the adaptive MLMC algorithm.
As we pointed out earlier, this is caused by a special feature of the adaptive MLMC
algorithm. Samples that were predicted to be redundant in an early stage of the
algorithm, may become necessary later in the computation. Thus at termination, the
actual oversampling is significantly less than predicted. This is a dynamic effect that
cannot be quantified easily in a static a priori fashion.
8. Conclusions. In this paper we have explored the use of multilevel Monte
Carlo methods on very large supercomputers. Three levels of parallelism must be coordinated, since it is not sufficient to just execute samples in parallel. The combination
of solver- and sample-parallelism leads to a non-trivial scheduling problem, where the
trade-off between solver scalability, oversampling, and additional efficiency losses due
to run-time variations must be balanced with care. This motivated the development
of scheduling strategies of increasing complexity, including advanced dynamic methods that rely on meta-heuristic search algorithms. These scheduling algorithms are
based on performance predictions for the individual tasks that can in turn be derived
from run-time measurements and performance models motivated by Amdahl’s law.
The success of the techniques and their scalability are demonstrated on a largescale model problem. The largest MLMC computation involves more than 10 000
samples and a fine grid resolution with almost 7 × 1010 unknowns. It is executed on
131 072 cores of a peta-scale class supercomputer in 1.5 hours of total compute time.
REFERENCES
[1] A. H. Baker, R. D. Falgout, T. Gamblin, T. V. Kolev, M. Schulz, and U. M. Yang. Scaling
algebraic multigrid solvers: On the road to exascale. In Competence in High Performance
Computing 2010, pages 215–226. Springer, 2012.
[2] A. Barth, C. Schwab, and N. Zollinger. Multi-level Monte Carlo finite element method for
elliptic PDE’s with stochastic coefficients. Numer. Math., 119:123–161, 2011.
[3] B. K. Bergen and F. Hülsemann. Hierarchical hybrid grids: data structures and core algorithms
for multigrid. Numer. Lin. Alg. Appl., 11(2-3):279–291, 2004.
[4] J. Charrier, R. Scheichl, and A. L. Teckentrup. Finite element error analysis of elliptic PDEs
with random coefficients and its application to multilevel Monte Carlo methods. SIAM
J. Numer. Anal., 51(1):322–352, 2013.
[5] E. Chow, R. D. Falgout, J. J. Hu, R. S. Tuminaro, and U. M. Yang. A survey of parallelization
techniques for multigrid solvers. In M. A. Heroux, P. Raghavan, and H. D. Simon, editors,
Parallel Processing for Scientific Computing, chapter 10, pages 179–201. SIAM, 2006.
[6] K. A. Cliffe, M. B. Giles, R. Scheichl, and A. L. Teckentrup. Multilevel Monte Carlo methods
and applications to elliptic PDEs with random coefficients. Comput. Visual. Sci., 14(1):3–
15, 2011.
[7] N. Collier, A. L. Haji-Ali, F. Nobile, E. von Schwerin, and R. Tempone. A continuation
multilevel Monte Carlo algorithm. BIT Numer. Math., 55(2):399–432, 2015.
[8] C. R. Dietrich and G. H. Newsam. Fast and exact simulation of stationary Gaussian processes
through circulant embedding of the covariance matrix. SIAM J. Sci. Comput., 18:1088–
1107, 1997.
[9] J. Dongarra.
Report on the Sunway TaihuLight system.
Technical report,
University of Tennessee, Oak Ridge National Laboratory, June 24, 2016.
http://www.netlib.org/utk/people/JackDongarra/PAPERS/sunway-report-2016.pdf.
[10] M. Drozdowski. Scheduling for Parallel Processing. Springer-Verlag, London, 2009.
[11] D. Elfverson, F. Hellman, and A. Målqvist. A multilevel Monte Carlo method for computing
failure probabilities. SIAM/ASA J. Uncertainty Quantification, 4(1):312–330, 2016.
[12] M. Garey and D. Johnson. Computers and Intractability; A Guide to the Theory of NPCompleteness. W. H. Freeman & Co, New York, 1990.
25
[13] R. G. Ghanem and P. Spanos. Stochastic Finite Elements: A Spectral Approach. Springer,
1991.
[14] M. B. Giles. Multilevel Monte Carlo path simulation. Operations Res., 56(3):981–986, 2008.
[15] M. B. Giles and B. J. Waterhouse. Multilevel quasi-Monte Carlo path simulation. Radon Series
Comp. Appl. Math., 8:1–18, 2009.
[16] B. Gmeiner, M. Huber, L. John, U. Rüde, and B. Wohlmuth. A quantitative performance study
for stokes solvers at the extreme scale. J. Comput. Sci., 2016. Accepted for publication.
[17] B. Gmeiner, H. Köstler, M. Stürmer, and U. Rüde. Parallel multigrid on hierarchical hybrid
grids: a performance study on current high performance computing clusters. Concurrency
and Computation: Practice and Experience, 26(1):217–240, 2014.
[18] B. Gmeiner, U. Rüde, H. Stengel, C. Waluga, and B. Wohlmuth. Performance and scalability of hierarchical hybrid multigrid solvers for Stokes systems. SIAM J. Sci. Comput.,
37(2):C143–C168, 2015.
[19] B. Gmeiner, U. Rüde, H. Stengel, C. Waluga, and B. Wohlmuth. Towards textbook efficiency
for parallel multigrid. Numer. Math. Theor. Meth. Appl., 8(01):22–46, 2015.
[20] I. G. Graham, F. Y. Kuo, D. Nuyens, R. Scheichl, and I. H. Sloan. Quasi-Monte Carlo methods
for elliptic PDEs with random coefficients and applications. J. Comput. Phys., 230:3668–
3694, 2011.
[21] G. Hager and G. Wellein. Introduction to high performance computing for scientists and
engineers. CRC Press, 2010.
[22] V. H. Hoang, C. Schwab, and A. M. Stuart. Complexity analysis of accelerated MCMC methods
for Bayesian inversion. Inverse Prob., 29:085010, 2013.
[23] J. K. Lenstra, A. H. G. Rinnooy Kan, and P. Brucker. Complexity of machine scheduling.
Annals of Discrete Mathematics, 1:343–362, 1977.
[24] J. L. Leva. A fast normal random number generator. ACM Transactions on Mathematical
Software (TOMS), 18(4):449–453, 1992.
[25] F. Lindgren, H. Rue, and J. Lindström. An explicit link between gaussian fields and gaussian
markov random fields: the stochastic partial differential equation approach. Journal of the
Royal Statistical Society: Series B (Statistical Methodology), 73(4):423–498, 2011.
[26] G. J. Lord, C. Powell, and T. Shardlow. An Introduction to Computational Stochastic PDEs.
Cambridge Texts in Applied Mathematics. Cambridge University Press, 2014.
[27] Z. Michalewicz. Genetic Algorithms + Data Structures = Evolution Programs. Springer, 1996.
[28] S. Mishra, C. Schwab, and J. Šukys. Multi-level Monte Carlo finite volume methods for shallow
water equations with uncertain topography in multi-dimensions. SIAM J. Sci. Comput.,
34:761–784, 2012.
[29] R. Potsepaev and C. L. Farmer. Application of stochastic partial differential equations to
reservoir property modelling. page B003. ECMOR XII, 2010.
[30] W. H. Press. Numerical Recipes 3rd Edition: The Art of Scientific Computing. Cambridge
University Press, 2007.
[31] D. Simpson, J. Illian, F. Lindgren, S. Sørbye, and H. Rue. Going off grid: Computationally
efficient inference for log-Gaussian Cox processes. Biometrika, 103:49–70, 2016.
[32] M. Srinivas and L. M. Patnaik. Genetic algorithms: A survey. Computer, 27(6):17–26, 1994.
[33] E. Strohmaier, H. W. Meuer, J. Dongarra, and H. D. Simon. The Top500 list and progress in
high-performance computing. Computer, 48(11):42–49, 2015.
[34] J. Šukys. Adaptive load balancing for massively parallel multi-level Monte Carlo solvers. In
Parallel Processing and Applied Mathematics, pages 47–56. Springer, 2014.
[35] J. Šukys, S. Mishra, and C. Schwab. Static load balancing for multi-level Monte Carlo finite
volume solvers. In Parallel Processing and Applied Maths, pages 245–254. Springer, 2012.
[36] A. L. Teckentrup, R. Scheichl, M. B. Giles, and E. Ullmann. Further analysis of multilevel MC
methods for elliptic PDEs with random coefficients. Numer. Math., 125(3):569–600, 2013.
[37] J. D. Ullman. NP-complete scheduling problems. J. Comput. System Sci., 10:384–393, 1975.
[38] P. J. M. Van Laarhoven and E. H. L. Aarts. Simulated annealing. Springer, 1987.
[39] P. J. M. Van Laarhoven, E. H. L. Aarts, and J. K. Lenstra. Job shop scheduling by simulated
annealing. Operations Res., 40(1):113–125, 1992.
[40] I. Wegener. Simulated annealing beats Metropolis in combinatorial optimization. In L. Caires
et al., editors, Automata, Languages and Programming, volume 3850 of Lecture Notes in
Computer Science, pages 589–601. Springer, 2005.
[41] P. Whittle. Stochastic processes in several dimensions. Bull. Inst. Internat. Stat., 40:974–994,
1963.
[42] D. Xiu and G. E. Karniadakis. The Wiener–Askey polynomial chaos for stochastic differential
equations. SIAM J. Sci. Comput., 24(2):619–644, 2002.
26
| 5cs.CE
|
Local Moufang sets and PSL2 over a local ring
Tom De Medts ([email protected])
Erik Rijcken∗ ([email protected])
arXiv:1510.03688v4 [math.GR] 15 Mar 2017
March 16, 2017
Abstract
We introduce local Moufang sets as a generalization of Moufang sets. We present a
method to construct local Moufang sets from only one root group and one permutation. We
use this to describe PSL2 over a local ring as a local Moufang set, and give necessary and
sufficient conditions for a local Moufang set to be of this form.
1
Introduction
A Moufang set is a permutation group G acting on a set X, along with a conjugacy class of
subgroups {Ux | x ∈ X}, such that each Ux fixes x and acts regularly on X \ {x}, and such that
Uxg = Uxg for all g ∈ G. These structures have been introduced by Jacques Tits in [Tit92]. The
easiest example of a Moufang set is PSL2 (k) acting on the projective line, where k is a field. In
[DW06], T. De Medts and R. Weiss gave some natural necessary and sufficient conditions for a
Moufang set to be isomorphic to the Moufang set of PSL2 (k) for some field k with char(k) 6= 2;
this result has been extended to fields of any characteristic by M. Grüninger [Grü10].
In this paper, we introduce a more general structure that encompasses permutation groups
such as PSL2 (R) for a local ring R. (All local rings appearing in this paper are assumed to be
commutative rings with 1.) We will call such a structure a local Moufang set. We have a set with
an equivalence relation (X, ∼), and for each x ∈ X a group Ux acting faithfully on this set and
preserving equivalence. These groups are called root groups, and the group generated by them
is the little projective group, denoted by G. We want this action to be sufficiently nice, similar
to the situation of Moufang sets, and hence impose a few natural axioms.
Section 2 consists of some main definitions and basic properties that will be used throughout
the paper, including the generalization of two important notions of Moufang sets: the µ-maps
and Hua maps. These are specific elements of the little projective group that respectively swap
and fix two chosen non-equivalent elements of X. Section 3 then aims to show that the two-point
stabilizer of those two chosen elements is in fact generated by the Hua maps.
To describe examples of local Moufang sets, it is convenient to have a simpler construction.
It turns out to be sufficient to have one root group and a permutation τ which swaps the fixed
point of the root group with a non-equivalent point, in order to reconstruct all the data for a
local Moufang set; some additional conditions are needed to show all axioms (Theorem 4.6).
Using this construction, it is then easy to describe the local Moufang set structure of PSL2 (R).
In the last section, we will define the notion of a special local Moufang set. From special
local Moufang sets with abelian root groups, satisfying some extra conditions, we can construct a
∗ PhD
Fellow of the Research Foundation - Flanders (Belgium) (F.W.O.-Vlaanderen)
1
2
Definition and basic properties of local Moufang sets
2
local ring R, and hence we expect the local Moufang set to be closely related to PSL2 (R) (Theorem 5.15). Using this construction, we find necessary and sufficient conditions for a local Moufang
set to be PSL2 (R) for a local ring with residue field of characteristic not 2 (Theorem 5.20).
Acknowledgment
The idea of axiomatizing groups such as PSL2 (R) as “non-primitive variant of Moufang sets”
arose from discussions with Pierre-Emmanuel Caprace about Moufang sets appearing as the set
of ends of a (locally finite) tree. We are grateful to him for sharing his insights with us. We also
thank the referee for a careful reading of the paper, which resulted in valuable suggestions and
minor corrections.
2
Definition and basic properties of local Moufang sets
2.1
Local Moufang sets
We first recall the notion of a Moufang set, and we refer the reader to [DS09] for a general
introduction to the theory (which we will not explicitly need in the current paper).
A Moufang set is defined as a set X together with a family of groups {Ux }x∈X acting faithfully
on the set, satisfying the following two properties:
(M1) For x ∈ X, Ux fixes x and acts sharply transitively on X \ {x}.
(M2) For x ∈ X and g ∈ hUy | y ∈ Xi, we have Uxg = Uxg .
We will generalize this to a family of groups acting on a set with an equivalence relation,
which will be the main objects we introduce and study.
Notation 2.1.
• If (X, ∼) is a set with an equivalence relation, we denote the equivalence class of x ∈ X
by x, and the set of equivalence classes by X.
• We denote the group of equivalence-preserving permutations of X by Sym(X, ∼).
• If g ∈ Sym(X, ∼), we will denote the corresponding element of Sym(X) by g.
• Our actions will always be on the right. The action of an element g on an element x will
be denoted by x · g or xg. Conjugation will correspondingly be g h = h−1 gh.
We are now ready to define our main objects, the local Moufang sets.
Definition 2.2. A local Moufang set M consists of a set with an equivalence relation (X, ∼) such
that |X| > 2, and a family of subgroups Ux 6 Sym(X, ∼) for all x ∈ X, called the root groups.
We denote Ux := Ux = Im(Ux → Sym(X)) for the permutation group induced by the action of
Ux on the set of equivalence classes. (This notation is justified by (LM1) below.) The group
generated by the root groups is called the little projective group, and will usually be denoted by
G := hUx | x ∈ Xi. Furthermore, we demand the following:
(LM1)
(LM2)
(LM2’)
(LM3)
If x ∼ y for x, y ∈ X, then Ux = Uy .
For x ∈ X, Ux fixes x and acts sharply transitively on X \ x.
For x ∈ X, Ux fixes x and acts sharply transitively on X \ {x}.
For x ∈ X and g ∈ G, we have Uxg = Uxg .
It is worth noting that from these axioms, we also get
(LM3’) For x ∈ X and g ∈ G, we have Uxg = Uxg ;
2
Definition and basic properties of local Moufang sets
3
this follows from (LM3) and the fact that we are working with the induced action. By (LM1)
the group Ux only depends on x, and (LM2’) and (LM3’) precisely state that (X, {Ux }x∈X ) is a
Moufang set.
Definition 2.3. Two local Moufang sets M and M′ are isomorphic, denoted M ∼
= M′ , if there is
′
′
a bijection ϕ : X → X and group isomorphisms θx : Ux → Uϕ(x) such that
• for all x, y ∈ X, we have x ∼ y ⇐⇒ ϕ(x) ∼′ ϕ(y);
• for all x, y ∈ X and u ∈ Uy , we have ϕ(x · u) = ϕ(x) · θy (u).
Before giving an example, we will first prove a basic property about local Moufang sets:
Proposition 2.4. Let M be a local Moufang set, and x, y ∈ X with x 6∼ y. Then hUx , Uy i = G.
Proof. It is sufficient to show that Uz ⊆ hUx , Uy i for any z ∈ X. Assume first that z ∼ x. Then
z 6∼ y, so by (LM2) there is a g ∈ Uy such that xg = z. By (LM3), Uz = Uxg ⊆ hUx , Uy i.
Similarly, if z 6∼ x, then there is a g ∈ Ux such that yg = z, so Uz = Uyg ⊆ hUx , Uy i.
This means that it is sufficient to give the set with its equivalence relation and two such root
groups to get all the data of a local Moufang set. We will use this to give an example of a local
Moufang set.
Example 2.5. The easiest example of a local Moufang set (with a non-trivial equivalence relation) is that of the projective special linear group acting on a projective line over a local ring R.
So let R be an arbitrary local ring, let m be its unique maximal ideal, and let R× = R \ m be
the set of invertible elements of R. First, we have to define the set and the equivalence relation,
i.e. the projective line over R. A vector line over R consists of all invertible multiples of a pair
(a, b) ∈ R2 such that aR + bR = R. We denote this vector line by
[a, b] := {(au, bu) | u ∈ R× }.
The projective line over R is then
P1 (R) := {[a, b] | aR + bR = R} .
Since R is a local ring, the condition is equivalent to saying at least one of a and b is invertible.
We can hence simplify our projective line to
P1 (R) = {[1, r] | r ∈ R} ∪ {[m, 1] | m ∈ m}.
Now the equivalence relation on P1 (R) can easily be defined by
[1, r] ∼ [1, s] ⇐⇒ r − s ∈ m ,
[m, 1] ∼ [n, 1] for all m, n ∈ m ,
(2.1)
and no other equivalences hold. We get two non-equivalent points on the projective line which
have a simple description, namely [1, 0] and [0, 1], for which we will define the root groups.
The root groups live inside
PSL2 (R) := A ∈ Mat2 (R) | det(A) = 1 / ( r0 0r ) | r ∈ R, r2 = 1 .
We will denote a matrix in this quotient by square brackets. We can now set
1 0
1 r
U[1,0] =
∈ PSL2 (R) r ∈ R
U[0,1] =
∈ PSL2 (R) r ∈ R ,
r 1
0 1
2
Definition and basic properties of local Moufang sets
4
which generates the entire little projective group, and hence defines all other root groups. We
denote this local Moufang set by M(R). Of course, we have not yet proven that this is a local
Moufang set. At this point, one could do the verifications, but we will later show sufficient
conditions which are easier to verify. What is important to note at this point, is that U[0,1]
indeed fixes [0, 1], preserves equivalence, and acts sharply transitively on {[1, r] | r ∈ R}.
We also observe that there is a natural bijection from P1 (R) to P1 (R/m), the projective line
over the residue field R/m of R. Furthermore, the Moufang set we get by the induced action on
P1 (R) is the standard Moufang set associated with PSL2 (R/m).
In the example, we chose two nice points to work with. We will also do this in general, but
we would first like to know to what extent the choice of the points is relevant. It turns out that
it is not, as long as we take points that are not equivalent.
Proposition 2.6. Let M be a local Moufang set. The little projective group G acts transitively
on {(x, y) ∈ X 2 | x 6∼ y}.
Proof. Let (x, y) and (x′ , y ′ ) be such pairs. We will first map x to x′ , for which we need two
cases:
x ∼ x′ : By (LM2), there is a g ∈ Uy mapping x to x′ , so we have (x, y) · g = (x′ , y).
x 6∼ x′ : Since |X| > 2, there is a z ∈ X such that x 6∼ z and x′ 6∼ z. Again by (LM2), there is
a g ∈ Uz s.t. x · g = x′ , and hence (x, y) · g = (x′ , y ′′ ) for some y ′′ 6∼ x′ .
So, after renaming, we have reduced the question to finding an element mapping (x, y) to (x, y ′ ).
Since y 6∼ x and y ′ 6∼ x, there is a g ∈ Ux mapping y to y ′ , hence (x, y) · g = (x, y ′ ).
Notation 2.7.
• In a local Moufang set, we now fix two points of X that are not equivalent, and we call
them 0 and ∞.
• For any x 6∼ ∞, by (LM2), there is a unique element of U∞ mapping 0 to x. We denote
this element by αx . In particular, α0 = 1.
−1
• For x 6∼ ∞, we set −x := 0 · α−1
x , so αx = α−x .
For many properties, we will have to restrict to those x ∈ X non-equivalent to both 0 and ∞.
Definition 2.8. In a local Moufang set, an element x ∈ X is a unit if x 6∼ 0 and x 6∼ ∞.
There are a few other ways of characterizing the units, based on their corresponding elements
of U∞ .
Proposition 2.9. Let M be a local Moufang set, and x ∈ X with x 6∼ ∞. Then the following
are equivalent:
(i) x is a unit;
(ii) αx does not fix 0;
(iii) αx does not fix any element of X \ ∞.
Proof.
(i) ⇔ (ii). We have x = 0αx ∈ 0αx , so
x is a unit ⇐⇒ x 6∼ 0 ⇐⇒ x 6∈ 0 ⇐⇒ 0 6= 0 · αx .
(ii) ⇔ (iii). The induced permutation αx is contained in U∞ . By (LM2’), this element fixes
either all elements or no elements of X \ {∞}.
2
Definition and basic properties of local Moufang sets
5
Corollary 2.10. Let M be a local Moufang set, and x ∈ X with x 6∼ ∞. Then x is a unit if and
only if −x is a unit.
Definition 2.11. Let M be a local Moufang set and x ∈ X. We define Ux◦ := Ker(Ux → Ux )
and Ux× := Ux \ Ux◦ .
×
These subsets are related to the units since x is a unit if and only if x ∈ U∞
. We have
◦
× g
×
= Uxg and (Ux ) = Uxg .
(Ux◦ )g
2.2
The µ-maps
Now that we have a fixed pair (0, ∞), we know that by Proposition 2.6 there must be an element
of G swapping these two. We first look at the double cosets U0 αx U0 , and indeed, there is often
an element switching our two points.
Proposition 2.12. For each unit x ∈ X, there is a unique element µx ∈ U0 αx U0 such that
0µx = ∞ and ∞µx = 0; it is called the µ-map corresponding to x.
Moreover, µx = gαx h ∈ U0× αx U0× , where g is the unique element of U0 mapping ∞ to −x
and h is the unique element of U0 mapping x to ∞.
Proof. Let gαx h be an element of U0 αx U0 , then the conditions translate to
∞ = 0gαx h = xh
−1
and ∞ = 0h−1 α−1
= (−x)g −1 ,
x g
so g is the unique element of U0 mapping ∞ to −x, and h is the unique element of U0 mapping
x to ∞. Since x 6∼ ∞, both g and h are in U0× , and since both g and h are unique, so is µx .
Notation 2.13. We fix one more object in a local Moufang set M: we pick one µ-map and call
it τ . Recall that |X| > 2, so there is at least one unit.
Lemma 2.14.
τ
(i) U0τ = U∞ and U∞
= U0 .
(ii) G = hU0 , U∞ i = hU∞ , τ i.
(iii) Let x ∈ X. Then x is a unit if and only if xτ is a unit.
Proof.
(i) This follows immediately from (LM3) and Proposition 2.12.
(ii) The fact that G = hU0 , U∞ i already follows from Proposition 2.4, and the second equality
hU0 , U∞ i = hU∞ , τ i then follows from (i).
(iii) Since τ preserves the equivalence and switches 0 and ∞, we have x ∼ 0 ⇐⇒ xτ ∼ ∞ and
x ∼ ∞ ⇐⇒ xτ ∼ 0.
Notation 2.15. For each x 6∼ ∞, we set γx := ατx ∈ U0 , which is the unique element of U0
mapping ∞ to xτ .
Example 2.16. In our main example of M(R), a point of the projective line is a unit if it can
be written as [1, r] with r ∈ R× . The two points we chose, [1, 0] and [0, 1], are usually denoted
by 0 and ∞ respectively. To find a µ-map corresponding to [1, r], we need the unique element of
U0 mapping ∞ to −[1, r], and the unique element of U0 mapping [1, r] to ∞.
−1
We first compute −[1, r]. Now α[1,r] = [ 10 r1 ], so −[1, r] = 0α−1
, 1]. The
[1,r] = [1, −r] = [−r
1 0
unique element of U0 mapping ∞ to −[1, r] is −r−1 1 ; the unique element of U0 mapping [1, r]
to ∞ is also −r1−1 01 . We find
1
0 1 r
1
0
0
r
µ[1,r] =
=
.
−r−1 1 0 1 −r−1 1
−r−1 0
2
Definition and basic properties of local Moufang sets
In the specific case where r = 1, we get
0 1
−1 0
6
, which is the usual choice for τ .
Many of the following identities will be crucial in later calculations.
Lemma 2.17. Let x be a unit, and ∼x := (−(xτ −1 ))τ . Then
(i)
(ii)
(iii)
(iv)
µx does not depend on the choice of τ ;
µx = ατ(−x)τ −1 αx ατ−(xτ −1 ) ;
µ−x = µ−1
x ;
µxτ = µτ−x ;
(v)
(vi)
(vii)
(viii)
µx = αx ατ−(xτ −1 ) α−∼x ;
∼x = −((−x)µx );
∼x does not depend on the choice of τ ;
µ−x = α−∼x µ−x αx µ−x α∼−x .
Proof.
(i) This follows from the definition of µx .
(ii) By Proposition 2.12, µx = gαx h, where g is the unique element of U0 mapping ∞ to −x
and h is the unique element of U0 mapping x to ∞. Hence g = γ(−x)τ −1 = ατ(−x)τ −1 and
−1
−τ
τ
h = γxτ
−1 = αxτ −1 = α−(xτ −1 ) .
−1
(iii) As µx swaps 0 and ∞, so does µ−1
x ∈ U0 αx U0 = U0 α−x U0 . Since µ−x is the unique such
−1
element in U0 α−x U0 , we must have µx = µ−x .
(iv) By (i), we can use (ii) with τ −1 in place of τ for the right-hand side, so we get
µxτ = µτ−x
−1
−1
⇐⇒
ατ(−(xτ ))τ −1 αxτ ατ−x = ατxτ α−x ατ−((−x)τ )
⇐⇒
ατ(−(xτ ))τ −1 αxτ ατ−x = αxτ ατ−x α−((−x)τ )
⇐⇒
−1 τ
τ
α−τ
−x αxτ α(−(xτ ))τ −1 αxτ α−x = α−((−x)τ )
⇐⇒
τ
−1 τ
α−τ
−x αxτ α−((−(xτ ))τ −1 ) αxτ α−x = α(−x)τ .
αxτ ατ−x
By (LM3), the left-hand side belongs to U0
is equal to αy for
τ
= U0·αxτ ατ−x = U∞ , so the left-hand side
−1 τ
τ
y = 0 · α−τ
−x αxτ α−((−(xτ ))τ −1 ) αxτ α−x
= −(xτ ) · τ −1 α−((−(xτ ))τ −1 ) τ αxτ ατ−x
= ∞ · ατ−x = (−x)τ.
Hence the last of the equivalent equalities holds, and indeed µxτ = µτ−x .
(v) By (ii), we have
µxτ = ατ(−(xτ ))τ −1 αxτ ατ−x ,
and hence by (iv),
µx = µ−τ
xτ
−1
−1
= αx ατ−xτ α−(−(xτ ))τ −1 .
Since µx does not depend on the choice of τ , we can replace τ by τ −1 in the right-hand
side, which gives the required identity.
(vi) When we apply both sides of the identity (v) to −x, we get
(−x) · µx = (−x) · αx ατ−(xτ −1 ) α−∼x = 0 · α−∼x = −∼x
which gives ∼x = −((−x)µx ).
(vii) Since µx does not depend on the choice of τ , and ∼x = −((−x)µx ), we conclude that ∼x
does not depend on the choice of τ either.
2
Definition and basic properties of local Moufang sets
7
(viii) The equation (v) does not depend on the choice of τ . If we substitute −x for x and µ−x
for τ , then we get, using (vi),
µ−x = α−x µx α−((−x)µx ) µ−x α−∼−x = α−x µx α∼x µ−x α−∼−x .
Moving all terms except µ−x from the right hand side to the left hand side gives the
result.
Proposition 2.18. Let x, y ∈ X be units such that x 6∼ y. Then z := xτ −1 α−(yτ −1 ) τ is
independent on the choice of τ . Furthermore, z = xα−y µy α∼y and ∼z = yα−x µx α∼x .
Proof. By Lemma 2.17(v) we get z = xατ−(yτ −1 ) = xα−y µy α∼y , so it does not depend on the
choice of τ . Now we have z = 0 · αxτ −1 α−(yτ −1 ) τ , so
∼z = 0 · (αxτ −1 α−(yτ −1 ) )−1 τ = 0 · αyτ −1 α−(xτ −1 ) τ.
This coincides with our definition of z with x and y interchanged, so ∼z = yα−x µx α∼x .
2.3
The Hua maps
Now that we have found a set of elements of G swapping 0 and ∞, we attempt to find the
elements fixing 0 and ∞.
Definition 2.19. The Hua map hx,τ corresponding to a unit x is the element
hx,τ = τ µx = τ αx τ −1 α−(xτ −1 ) τ α−∼x ∈ G .
Remark 2.20. The µ-maps did not depend on the choice of τ , so the Hua maps do, as is made
clear by the inclusion of τ in the notation hx,τ . When it is clear (or irrelevant) which τ is used,
we will omit this addition and simply write hx .
Some basic properties of the Hua maps are the following:
Lemma 2.21. Let x, y ∈ X be units. Then
(i) hx,τ −1 = h−1
xτ,τ ;
(iii) hxτ = hτ−x ;
(iv) hxhy = h−y h−1
xτ hy .
h
(ii) µxhy = µxy ;
Proof.
−1
(i) We have hx,τ −1 = h−1
µx = (τ µxτ )−1 , which holds by Lemma 2.17(iv).
xτ,τ if and only if τ
(ii) Applying Lemma 2.17(iv) twice (with τ and µy ), we get
−τ −µy
y
= µτxµy = µhxy .
µxhy = µxτ µy = µ−µ
xτ = (µx )
(iii) Using Lemma 2.17(iv), we get
hxτ = τ µxτ = τ µτ−x = (τ µ−x )τ = hτ−x .
(iv) Using Lemma 2.17(iv) twice, and inserting τ −1 τ , we get
−1 −1
τ µy = h−y h−1
hxhy = τ µxτ µy = τ µ−y µ−1
xτ µy = τ µ−y µxτ τ
xτ hy .
3
The Hua subgroup
8
Definition 2.22. The Hua subgroup is
H := hµx µy | x, y unitsi .
Since τ is chosen to be a µ-map, we also have
H = hhx | x a uniti.
Note that H 6 G0,∞ , where G0,∞ is the two-point stabilizer of 0 and ∞. In the next section,
we will show that, in fact, H = G0,∞ . The action of Hua maps on U∞ by conjugation behaves
well with respect to the action on X:
Lemma 2.23. Let x be a unit. Then for any y ∈ X \ ∞, we have αhy x = αyhx . In particular,
(αy αz )hx = αyhx αzhx for all y, z ∈ X \ ∞.
Proof. Since Hua maps normalize U∞ , we have αhy x ∈ U∞ . Since 0αhy x = 0h−1
x αy hx = yhx , the
first equality holds. The second identity now follows immediately.
Example
2.24. In Example 2.5, we calculated that for M(R) over a local ring R, we have
µ[1,r] = −r0−1 0r and we took τ = µ[1,1] . For the Hua maps and the Hua subgroup, this means
(using the fact that −10 −10 = [ 10 01 ]) that
h[1,r] =
3
3.1
−1
r
0
0
,
r
H=
r−1
0
0
r ∈ R× .
r
The Hua subgroup
Quasi-invertibility
Inspired by [Loo14, §4], we will introduce the notion of quasi-invertibility for local Moufang
sets. (This notion is, in turn, inspired by the notion of quasi-invertibility for Jordan algebras
and Jordan pairs.) This notion, and the identity in Proposition 3.3 that follows from it, will be
crucial to show that the Hua subgroup equals G0,∞ .
Definition 3.1. A pair (x, y) ∈ X 2 s.t. x 6∼ ∞ and y 6∼ ∞ is called quasi-invertible if either
xτ 6∼ −y, or x ∼ 0, or y ∼ 0.
Definition 3.2. Let (x, y) be quasi-invertible. Then we define the left quasi-inverse and right
quasi-inverse as
−1
x
y := (−y) · ατ−x
and
xy := −(x · ατy ) .
Note that the condition for quasi-invertibility ensures that the left and right quasi-inverse do
not lie in the ∞-branch. Furthermore, x ∼ 0 ⇐⇒ xy ∼ 0 and y ∼ 0 ⇐⇒ x y ∼ 0.
Proposition 3.3. Let (x, y) be quasi-invertible with y 6∼ 0 (so x 6∼ ∞ 6∼ y). Then
αx y ατx αy ατxy = µx y µy .
Proof. By the observation above, x y ∼
6 0, so the right-hand side is defined. By Lemma 2.17(v),
we have µx y = αx y ατ−(x yτ −1 ) α−∼ x y . Furthermore, by the definition of x y, we have
−(x yτ −1 ) = −(0 · α(−y)τ −1 α−x ) = 0 · αx α−(−y)τ −1 ,
3
The Hua subgroup
9
so α−(x yτ −1 ) = αx α−(−y)τ −1 . Plugging these into our equality gives
αx y ατx αy ατxy = µx y µy
⇐⇒
αy ατxy = ατ−(−y)τ −1 α−∼ x y µy
⇐⇒
α∼ x y ατ(−y)τ −1 αy ατxy = µy .
Now we use Lemma 2.17(ii) to find µy = ατ(−y)τ −1 αy ατ−(yτ −1 ) , which yields
αx y ατx αy ατxy = µx y µy
⇐⇒
α∼ x y ατ(−y)τ −1 αy ατxy = ατ(−y)τ −1 αy ατ−(yτ −1 )
⇐⇒
−τ
α∼ x y ατ(−y)τ −1 αy ατxy ατyτ −1 α−1
y α(−y)τ −1 = 1 .
−τ
α−1
y α
−τ
Now we have ατ(−y)τ −1 αy ατxy ατyτ −1 α−1
y α(−y)τ −1 ∈ U0
(−y)τ −1
−τ
= U0·α−1
y α
(−y)τ −1
(3.1)
by (LM3), and
−τ
−1
0 · α−1
α−(−y)τ −1 τ = (−y)τ −1 · α−(−y)τ −1 τ = 0τ = ∞ ,
y α(−y)τ −1 = (−y) · τ
so the left-hand side of (3.1) is an element of U∞ . To prove that it is the identity, it is now
sufficient to prove that it maps 0 to 0, by (LM2). Note first that ∼ x y = xα−1
(−y)τ −1 τ . We have
−τ
0 · α∼ x y ατ(−y)τ −1 αy ατxy ατyτ −1 α−1
y α(−y)τ −1
−τ
−1
= xα−1
α(−y)τ −1 τ αy ατxy ατyτ −1 α−1
y α(−y)τ −1
(−y)τ −1 τ τ
= x · ατy
−1
−τ
−1 −τ
τ
y
αxy τ ατyτ −1 α−1
y α(−y)τ −1 = (−x ) · αxy τ αyτ −1 αy α(−y)τ −1
= 0 · αyτ −1 τ α−y ατ−(−y)τ −1 = y · α−y ατ−(−y)τ −1
= 0 · ατ−(−y)τ −1 = 0 ,
and this finishes the proof.
Remark 3.4. In a similar fashion, one can show that if (x, y) is quasi-invertible with x 6∼ 0,
then
αx y ατx αy ατxy = µx µxy .
3.2
Bruhat decomposition of G
By refining the argument in Proposition 2.6, we will be able to obtain decompositions of the
little projective group G which resemble the Bruhat decomposition. This is based on a case
distinction depending on where the pair (0, ∞) is mapped to by a given element.
Proposition 3.5. The little projective group G can be decomposed into a disjoint union
G = U0 G0,∞ U∞ ∪ U0 G0,∞ τ U0◦ .
Moreover, the decomposition of an element is unique in each of the two cases.
Proof. Let g be in G and let (x, y) = (0, ∞) · g. We distinguish two mutually exclusive cases:
x ∼ ∞: By (LM2), there is a unique u0 ∈ U0 such that x · u0 = ∞. Since u0 fixes ∞, we have
u0 ∈ U0◦ . We get (x, y) · u0 τ −1 = (0, y ′ ) for some y ′ 6∼ 0. Hence there is a unique
u′0 ∈ U0 such that y ′ · u′0 = ∞, so we get
(0, ∞) · gu0 τ −1 u′0 = (x, y) · u0 τ −1 u′0 = (0, y ′ ) · u′0 = (0, ∞) ,
−1
◦
so gu0 τ −1 u′0 = h ∈ G0,∞ . Hence g = u′−h
hτ u−1
0
0 ∈ U0 G0,∞ τ U0 . Note that since u0
′
and u0 are unique, so is h, and hence the entire decomposition.
3
The Hua subgroup
10
x 6∼ ∞: In this case we find a unique u∞ ∈ U∞ such that x · u∞ = 0, so (x, y) · u∞ = (0, y ′ ).
Next, we find a unique u0 ∈ U0 such that y ′ · u0 = ∞, so
(0, ∞) · gu∞ u0 = (x, y) · u∞ u0 = (0, y ′ ) · u0 = (0, ∞) .
−1
This means gu∞ u0 = h ∈ G0,∞ , so g = u−h
hu−1
∞ ∈ U0 G0,∞ U∞ . Again, since u0 and
0
u∞ are unique, so is h and the entire decomposition.
By making similar case distinctions, one can get different decompositions, for example
G = U∞ G0,∞ U0◦ ∪ U∞ G0,∞ τ U∞ .
For this decomposition, we would separate two cases: ∞ · g ∼ ∞ or ∞ · g 6∼ ∞. In particular, we
can check that if we take g ∈ U0◦ U∞ , we always end up in the first component U∞ G0,∞ U0◦ , i.e.
U0◦ U∞ ⊆ U∞ G0,∞ U0◦ .
(3.2)
In section 3.3 below, we will show that G0,∞ = H. A first step towards this consists of showing
that the inclusion (3.2) holds with G0,∞ replaced by H. To obtain this, we will need the notion
of quasi-invertibility we introduced.
Proposition 3.6. In a local Moufang set, we have U0◦ U∞ ⊆ U∞ HU0◦ .
Proof. We will show this in two steps. First, we will prove that
×
×
U0◦ U∞
⊆ U∞
HU0◦ ,
(3.3)
and from that we will deduce the general inclusion.
×
So take an arbitrary element of U0◦ U∞
, and denote it by ατx αy . Then x ∼ 0 and ∞ 6∼ y 6∼ 0,
so (x, y) is quasi-invertible. By Proposition 3.3, we have
×
ατx αy = α− x y µx y µy ατ−xy ∈ U∞
HU0◦ ,
since 0 6∼ − x y 6∼ ∞ and −xy ∼ 0. This proves (3.3).
◦
For the general inclusion, we use the fact that if we have αy ∈ U∞
, we can split it up as
αy = αy′ αe , for units y ′ and e; indeed, units exist, and if e is a unit, then αy α−1
e does not fix 0,
so y ′ is also a unit. So we get
◦
× ×
×
×
×
×
× ×
U0◦ U∞
⊆ U0◦ U∞
U∞ ⊆ U∞
HU0◦ U∞
⊆ U∞
HU∞
HU0◦ = U∞
U∞ HU0◦ ⊆ U∞ HU0◦ ,
×
×
= U∞
H.
where we have used (3.3) twice, as well as the fact that HU∞
Putting these two inclusions together, we get
×
◦
×
U0◦ U∞ = U0◦ U∞
∪ U0◦ U∞
⊆ U∞
HU0◦ ∪ U∞ HU0◦ = U∞ HU0◦ .
3.3
The Hua subgroup is G0,∞
In the case of Moufang sets, one can use the Bruhat decomposition to prove that G0,∞ = H,
and as a consequence that the point stabilizer G0 = U0 H. In the case of local Moufang sets,
the additional U0◦ in the decomposition seems to cause further difficulties in the proof. However,
using Proposition 3.6, we will be able to resolve these difficulties, and we will again be able to
prove that the Hua subgroup coincides with the full 2-point stabilizer of 0 and ∞ in G.
4
Constructing local Moufang sets
11
Theorem 3.7. For a local Moufang set M, we have the decomposition G = U0 HU∞ ∪ U0 Hτ U0◦ ,
and hence G0 = U0 H and H = G0,∞ .
Proof. Let K = U0 H = HU0 . We will examine the set Q = KU∞ ∪ Kτ U0◦ ; our aim is to prove
that it equals G. More precisely, we will show that QhU∞ , τ i = QG ⊆ Q, from which Q = G
will follow immediately. We will do this for each of the two pieces of Q separately.
• We will first show that KU∞ G ⊆ Q. It is immediate that KU∞ U∞ = KU∞ ⊆ Q, so all
we need to prove is that KU∞ τ ⊆ Q, or equivalently, that Kαa τ ⊆ Q for all αa ∈ U∞ .
◦
Assume first that αa ∈ U∞
; then ατa ∈ U0◦ , so
Kαa τ = Kτ ατa ⊆ Kτ U0◦ ⊆ Q .
◦
×
Finally, when αa ∈ U∞ \U∞
= U∞
, we have µaτ ατa = ατ(−aτ )τ −1 αaτ by Lemma 2.17(ii).
−1
Since µaτ τ ∈ H ⊆ K, this implies
Kαa τ = Kτ ατa = Kµaτ τ −1 τ ατa = Kατ(−aτ )τ −1 αaτ ⊆ KU∞ .
◦
• We will now show that Kτ U0◦ G ⊆ Q. We have Kτ U0◦ τ = Kτ 2 U∞
⊆ KHU∞ = KU∞ ⊆ Q,
◦
so we need to prove that Kτ U0 U∞ ⊆ Q. We now invoke Proposition 3.6, and we get
Kτ U0◦ U∞ ⊆ Kτ U∞ HU0◦ = KU0 Hτ U0◦ = Kτ U0◦ ⊆ Q.
We conclude that QG ⊆ Q, and hence Q = G as claimed.
We now know that G = U0 HU∞ ∪ U0 Hτ U0◦ . If we take a g in the point stabilizer G0 and
look at the two possibilities of decomposing g, we get g ∈ U0 H, so G0 = U0 H. If we assume in
addition that g fixes ∞, we also see that the factor in U0 must be trivial, hence G0,∞ = H.
4
Constructing local Moufang sets
4.1
The construction M(U, τ )
We already know that, if we have a local Moufang set, then G = hU∞ , τ i. We now do the converse:
given a group U and a permutation τ , both acting faithfully on a set with an equivalence relation,
we will try to construct a local Moufang set. Of course, we will need additional conditions on U
and τ .
Construction A. The construction requires some data to start with. We need
• a set with an equivalence relation (X, ∼), such that |X| > 2;
• a group U 6 Sym(X, ∼), and an element τ ∈ Sym(X, ∼).
The action of U and τ will have to be sufficiently nice in order to do the construction.
(C1) U has a fixed point we call ∞, and acts sharply transitively on X \ ∞.
(C1’) The induced action of U on X is sharply transitive on X \ {∞}.
(C2) ∞τ 6∼ ∞ and ∞τ 2 = ∞. We write 0 := ∞τ .
In this construction, we now define the following objects:
• For x 6∼ ∞, we let αx be the unique element of U mapping 0 to x (by (C1) and (C2)).
• For x 6∼ ∞, we write γx := ατx , which then maps ∞ to xτ .
τ
. The other root groups are defined as
• We set U∞ := U and U0 := U∞
Ux := U0αx for x 6∼ ∞,
γ
Ux := U∞xτ −1 for x ∼ ∞.
4
Constructing local Moufang sets
12
• As in the definition of local Moufang sets, we write Ux for the induced action of Ux on X.
This gives us all the data that is needed for a local Moufang set; we denote the result of this
construction by M(U, τ ). Our goal in this section is to investigate when this is a local Moufang
set. This will require some additional definitions, which we have seen before for local Moufang
sets, but which we need to redefine in the current setup:
•
•
•
•
•
•
We call x ∈ X a unit if x 6∼ 0 and x 6∼ ∞.
For x 6∼ ∞, we set −x := 0α−1
x .
For a unit x, we define the µ-map µx := γ(−x)τ −1 αx γ−(xτ −1 ) .
For a unit x, we define the Hua map hx := τ αx τ −1 α−(xτ −1 ) τ α−(−(xτ −1 ))τ .
We set H := hµx µy | x, y unitsi.
We set U0◦ = {γx ∈ U0 | x ∼ 0}.
This construction automatically satisfies some of the axioms of a local Moufang set.
Proposition 4.1. The construction M(U, τ ) satisfies (LM1), (LM2) and (LM2’).
g
= g −1 U∞ g
Proof. By (C1), (LM2) holds for U∞ . Now, by definition, any other Ux is equal to U∞
for some g with ∞g = x. It follows that each Ux fixes x and acts sharply transitively on
(X \ ∞)g = X \ x, so (LM2) holds for all root groups.
Similarly, (C1’) implies (LM2’) for U∞ because U∞ fixes ∞, since U∞ fixes ∞. As before,
g
g
any Ux is equal to U∞
= g −1 U∞ g for some g with ∞g = x, so Ux is the induced action of U∞
on
X. This implies that Ux fixes ∞g = x and acts sharply transitively on (X \ {∞})g = X \ {x}.
α−1 α
We finally show (LM1). Let x ∼ y and suppose they are not equivalent to ∞. Then Ux x y =
Uy by definition. Now α−1
x αy is in U , and fixes x. By (C1’), this implies that the induced
permutation α−1
α
on
X
is trivial, so the induced action of Ux is the same as that of Uy . If
y
x
γ −1
γ
−1
−1
x ∼ y ∼ ∞, we have Ux xτ −1 yτ = Uy . Now γxτ
−1 γyτ −1 is in U0 , and fixes x. By (LM2’), the
−1
induced action of γxτ −1 γyτ −1 on X is trivial, so the induced action of Ux on X is identical to
that of Uy .
Example 4.2. In Example 2.5, we described M(R) for a local ring R using two root groups in
order to generate the entire little projective group. Alternatively, we can describe this example
with only one root group, and add τ to the setup, i.e., we define the local Moufang set M(R) as
M(U, τ ), with
1 r
0 1
U=
∈ PSL2 (R) r ∈ R ,
τ=
.
0 1
−1 0
We will now verify the three conditions (C1), (C1’) and (C2). For the first, we note that U fixes
[0, 1] =: ∞, andfor any
[r, 1] and [s, 1] in X \ ∞, there is a unique element of U mapping the first
:
. To show condition
to the second: 10 s−r
1
h i (C1’), we use the projection R → R/m r 7→ r. We
can see that [a, b] = [a, b] and [ 10 r1 ] =
1r
01
. The argument for condition (C1’) is now identical to
that for condition (C1), but using the residue field instead of the local ring. The last condition
is straightforward: ∞τ = [1, 0] 6∼ ∞, and ∞τ 2 = ∞.
By Proposition 4.1, we now know that M(R) satisfies the first three axioms of a local Moufang
set.
4.2
Conditions to satisfy (LM3)
To ensure that Construction A gives a local Moufang set, we will need more information about
the action of the Hua maps. We will first prove a few lemmas. Throughout this section, we let
M(U, τ ) be as in Construction A.
4
Constructing local Moufang sets
13
Lemma 4.3. Let x ∈ X be a unit. Then the following are equivalent:
hx
= U∞ ;
(i) U∞
γxτ −1
(ii) U∞
= Ux .
(iii) U0µx = U∞ ;
Proof.
(i)⇔(ii). We have
τ αx τ −1 α−(xτ −1 ) τ α−(−(xτ −1 ))τ
U∞
τ −1 α−1−1 τ
⇐⇒
Ux
xτ
αx τ −1 α−(xτ −1 ) τ
= U∞
⇐⇒
U0
= U∞
⇐⇒
Ux = U∞xτ −1 ,
α
= U∞(−(xτ
−1 ))τ
γ
where we only use the definitions of the root groups in M(U, τ ).
γ
(ii)⇔(iii). We have U0µx = U0 (−x)τ
−1 αx γ−(xτ −1 )
γ
= Ux −(xτ
−1 )
, so the equivalence follows.
Lemma 4.4. The following are equivalent:
(i)
(ii)
(iii)
(iv)
hx
= U∞ for all units x ∈ X;
U∞
γxτ −1
U∞
= Ux for all units x ∈ X.
U0µx = U∞ for all units x ∈ X;
µx
for all units x ∈ X;
U0 = U∞
Proof. The equivalence of (i)-(iii) is immediate from the previous lemma. The equivalence between (iii) and (iv) follows by replacing x with −x and noting that µ−x = µ−1
x .
Lemma 4.5. Assume that hx normalizes U for all units x ∈ X. Then
(i) U h = U for all h ∈ H;
(ii) U0◦ U∞ ⊆ U∞ HU0◦ .
Proof.
µx
(i) By Lemma 4.4, U0µx = U∞ and U0 = U∞
for all units x. Now H is generated by all
products of two µ-maps, which all normalize U∞ , so any element of H normalizes U∞ .
(ii) We can follow the proof of Proposition 3.6 mutatis mutandis. We point out that we used
(LM3) only twice: once in the proof of Lemma 2.17(iv), and once in Proposition 3.3. In the
αxτ ατ−x
= U∞ for any unit x. This
proof of Lemma 2.17(iv), we used (LM3) to deduce U0
also holds in the situation here since
αxτ ατ−x
U0
α−τ
γ
γx
⇐⇒ Uxτ = U∞(xτ )τ
= U∞ ⇐⇒ Uxτx = U∞ ⇐⇒ Uxτ = U∞
−1
.
These equivalences hold because of the definitions of the root groups in the construction,
and the final equality is true by the assumption and Lemma 4.4.
−τ
α−1
y α
In the proof of Proposition 3.3, we used (LM3) to deduce U0
y. In the situation here this also holds since
−τ
α−1
y α
U0
(−y)τ −1
ατ
= U∞ ⇐⇒ U−y = U∞(−y)τ
−1
(−y)τ −1
γ
= U∞ for any unit
⇐⇒ U−y = U∞(−y)τ
−1
,
where the equivalences are again by the definitions in the construction, and the final equality
holds by Lemma 4.4 again.
4
Constructing local Moufang sets
14
This additional assumption will also be sufficient to ensure that the construction is a local
Moufang set.
Theorem 4.6. Let M(U, τ ) be as in Construction A. Then M(U, τ ) is a local Moufang set if and
only if hx normalizes U for all units x.
Proof. Assume first that M(U, τ ) is a local Moufang set. By (LM3), all µ-maps send U0 to U∞ .
By Lemma 4.4, this implies that all Hua maps normalize U = U∞ .
For the converse, we have already shown that (LM1), (LM2) and (LM2’) hold, so what
remains is (LM3). Fix some unit e ∈ X and write µ = µe . Then, by our assumptions and by
Lemma 4.4, Ux ⊆ hU∞ , U0 i = hU∞ , µi for any x ∈ X, so in order to show that Uyg = Uy·g for all
g ∈ Ux and all y ∈ X, it is sufficient to show that Uyµ = Uy·µ and Uyαz = Uy·αz for any z 6∼ ∞.
We start by showing Uyαz = Uy·αz for all y ∈ X and z ∈ X s.t. z 6∼ ∞. We distinguish two
cases.
α
y 6∼ ∞: In this case, we have Uy = U0 y by definition. Since U∞ is a group, we have αy αz = αa
for some a, and by looking at the image of 0, we find a = yαz , so indeed Uyαz =
α α
α
U0 y z = U0 yαz = Uyαz .
γ
−1 αz
y ∼ ∞: By definition, Uyαz = U∞yτ
. Since γyτ −1 αz ∈ U0◦ U∞ , we have γyτ −1 αz = αa hγb for
some a 6∼ ∞, b ∼ 0 and h ∈ H. By calculating the image of ∞, we see bτ = yαz . So
we get
γ
Uyαz = U∞yτ
−1 αz
αa hγyαz τ −1
= U∞
hγ
= U∞ yαz τ
−1
γ
= U∞yαz τ
−1
= Uyαz ,
where we used Lemma 4.5(i).
Secondly, we need to show that Uyµ = Uy·µ for all y ∈ X; we make the same case distinction.
α µ
µαµ
αµ
y 6∼ ∞: We have Uyµ = U0 y = U0 y = U∞y . Now αµy ∈ U0 , so αµy = γa for some a, and by
checking the image of ∞ we find aτ = yµ. If yµ ∼ ∞, we can use the definition of
γ
αµ
−1
a root group to find U∞y = U∞yµτ = Uyµ . If yµ 6∼ ∞, we use y 6∼ ∞ to conclude
αµ
γ
yµ 6∼ 0, so yµ is a unit and by Lemma 4.4(ii), we also have U∞y = U∞yµτ
γ −1 µ
U∞yτ
µγ µ −1
yτ
−1
= Uyµ .
γ µ −1
yτ
µ
µ
y ∼ ∞: We have Uyµ =
= U∞
= U0
. Now γyτ
−1 ∈ U∞ , so γyτ −1 = αa for
some a, and by checking the image of 0 we find yµ = a. Notice that since y ∼ ∞,
γµ
we have yµ ∼ 0, so we can use the definition of the root group of yµ to get U0 yτ
α
U0 yµ = Uyµ .
−1
=
Remark 4.7. Let M(U, τ ) and M(U ′ , τ ′ ) be given by Construction A, with actions on (X, ∼)
and (X ′ , ∼′ ) respectively, and assume there is a bijection ϕ : X → X ′ and a group isomorphism
θ : U → U ′ such that
• for all x, y ∈ X, we have x ∼ y ⇐⇒ ϕ(x) ∼′ ϕ(y);
• for all x ∈ X and u ∈ U , we have ϕ(x · u) = ϕ(x) · θ(u);
• for all x ∈ X, we have ϕ(x · τ ) = ϕ(x) · τ ′ .
Then M(U, τ ) and M(U ′ , τ ′ ) are isomorphic.
Remark 4.8. Assume M is a local Moufang set, and take U = U∞ and τ to be any µ-map.
Then by (LM3), the root groups we get in Construction A are identical to the root groups of M.
Hence we can view any local Moufang set as M(U, τ ) for some U and τ .
Example 4.9. In Example 2.5, we have first introduced M(R) by giving two root groups. In
Example 4.2 in the previous section, we used construction A to define M(R). Up to now, we
5
Special local Moufang sets and PSL2 (R)
15
have not yet proven that what we described is indeed a local Moufang set. With Theorem 4.6
at hand, we can now easily do this.
For the construction, we used
1 r
0 1
U=
∈ PSL2 (R) r ∈ R ,
τ=
.
0 1
−1 0
One can check that [1, r]τ = [1, r]τ −1 = [1, −r−1 ] and −[1, r] = [1, −r] for [1, r] a unit. This
implies that
h[1,r] = τ α[1,r] τ −1 α−([1,r]τ −1 ) τ α−(−([1,r]τ −1 ))τ
0 −1 1 r 0 −1 1 r−1 0
=
1 0
0 1 1 0
0
1
1
−1 1
0
0
−1
r
r
=
1
0
0
.
r
Hence, for any unit [1, r],
U
h[1,r]
=
1 s
0 1
s∈R
h[1,r]
=
1
0
r2 s
1
s∈R
=U,
and this shows that M(R) is indeed a local Moufang set.
Special local Moufang sets and PSL2 (R)
5
5.1
Definition and first properties
We recall that a Moufang set is called special if ∼x = −x for all x ∈ U ∗ . This property was
introduced in the context of abstract rank one groups by F. Timmesfeld [Tim01, p. 2], and has
been thoroughly investigated for Moufang sets [DW06, DST08]. It is a difficult open problem
whether special Moufang sets always have abelian root groups. The converse, namely that proper
Moufang sets with abelian root groups are always special, has been shown by Y. Segev [Seg09].
We will now introduce the corresponding notion for local Moufang sets. We keep the notations
from before; in particular, τ is an arbitrary µ-map.
Definition 5.1. A local Moufang set M is called special if ∼x = −x for all units x ∈ X, or
equivalently, if (−x)τ = −(xτ ) for all units x ∈ X.
Some basic properties follow immediately from Proposition 2.17:
Lemma 5.2. Let x ∈ X be a unit in a special local Moufang set. Then
(i) (−y)µx = −(yµx ) for all units y ∈ X;
(ii) µx = αx ατ−xτ −1 αx ;
(iii) −x = xµx = xµ−x ;
µ
(iv) µx = αx αx±x αx .
(v) µ−x = αx µ−x αx µ−x αx .
Proof.
(i) This follows from Lemma 2.17(vii) and the definition of special.
(ii) This follows immediately from Lemma 2.17(v).
(iii) By Lemma 2.17(vi), we have −x = ∼x = −((−x)µx ), so x = (−x)µx and hence −x =
xµ−1
x = xµ−x . On the other hand, (i) implies x = (−x)µx = −(xµx ), hence xµx = −x.
(iv) Replace τ by µ±x in (ii), and use (iii).
Special local Moufang sets and PSL2 (R)
5
16
(v) This is a consequence of Lemma 2.17(viii) and the definition of special.
Lemma 5.3. If x, y ∈ X are units in a special local Moufang set, and xαy is a unit, then
xµxαy = (−y)α−x αxµy α−y .
Proof. Let x′ = xαy ; note that x′ 6∼ y since x 6∼ 0. Now let z = x′ τ −1 α−yτ −1 τ . By Proposition 2.18 and by specialness, we have
x′ α−y µy α−y = z = −∼z = −(yα−x′ µx′ α−x′ )
= 0 · (αyα−x′ µx′ α−x′ )−1 = x′ α−(yα−x′ µx′ )
= x′ α(−(yα−x′ ))µx′ ) = x′ αx′ α−y µx′
and hence x′ α−y µy = x′ αx′ α−y µx′ αy . Replacing x′ = xαy , we get
xµy = xαy αxµxαy αy
so
0 · αxµy = 0 · αx αy αxµxαy αy ,
hence αxµy = αx αy αxµxαy αy . By rearranging we get αxµxαy = α−y α−x αxµy α−y , which gives the
desired formula after applying it to 0.
Notation 5.4. For each x ∈ U , we write x · 2 := 0α2x = xαx . Hence α−(x·2) = (αx·2 )−1 = α−2
x =
α2−x , so −(x · 2) = (−x) · 2.
Lemma 5.5. Let x ∈ X be a unit in a special local Moufang set, and assume that also x · 2 is a
unit. Then
(i) there exists a unique y ∈ X such that y · 2 = x;
(ii) xτ · 2 ∼
6 0;
(iii) (x · 2)τ · 2 = xτ .
Proof. We will prove these three facts simultaneously. By (i) and (v) from Lemma 5.2, we have
−((x · 2)µ−x ) = ((−x) · 2)µ−x
= (−x)α−x αx µ−x αx µ−x αx
= (−x)µ−x αx µ−x αx
= xαx µ−x αx
= (x · 2)µ−x αx ,
and hence α−((x·2)µ−x ) = α(x·2)µ−x αx . Rearranging and applying to 0 gives
(x · 2)µ−x · 2 = −x .
Now µ−x τ and µx τ are Hua maps, so by Lemma 2.23 we get
(xτ ) · 2 = (−x)µ−x τ · 2 = (−x · 2)µ−x τ 6∼ 0
and
((x · 2)τ ) · 2 = ((x · 2)µ−x µx τ ) · 2 = ((x · 2)µ−x · 2)µx τ = (−x)µx τ = xτ .
This proves parts (ii) and (iii).
Now let y := ((−x) · 2)µx . Then by (iii), y · 2 = (−x)µx = x, which proves the existence in
part (i). It remains to show the uniqueness. So suppose there is another z such that z · 2 = x;
then
(−x) · 2 = (xµ−x ) · 2 = (z · 2)µ−x · 2 = zµ−x ,
by applying (iii) on z. Hence z = ((−x) · 2)µx = y, which proves uniqueness.
5
Special local Moufang sets and PSL2 (R)
17
If this holds for all units of a special local Moufang set and the root groups are abelian, we
get uniquely 2-divisible root groups.
Definition 5.6. A group U is (uniquely) 2-divisible if for every g ∈ U there is a (unique) h ∈ U
such that h2 = g. If this h is unique, we denote it by g/2.
Lemma 5.7. Let M be a special local Moufang set with U∞ abelian, such that each unit x, also
x · 2 is a unit. Then U∞ is uniquely 2-divisible.
Proof. Lemma 5.5(i) already shows that, if x is a unit, there is a unique y such that α2y = αx ;
therefore, it only remains to check the unique 2-divisibility for non-units. So suppose that x is
not a unit. Take any unit e; then αx = αxα−e αe . Now xα−e and e are units, so both αxα−e and
αe are uniquely 2-divisible, say with α2y = αxα−e and α2z = αe . Since U∞ is abelian, we get
(αy αz )2 = α2y α2z = αxα−e αe = αx .
To show uniqueness, suppose there are two elements u, u′ ∈ U∞ with u2 = u′2 = αx . Then
2
−2
−2 2 2
(α−1
y u) = αy αx = αy αy αz = αe ,
′ 2
−1 ′
−1
′
and similarly (α−1
y u ) = αe . By the uniqueness for units, we get αy u = αy u, so u = u .
Another property that holds when we have a special local Moufang set with abelian root
groups is the fact that µ-maps are automatically involutions.
Lemma 5.8. Let M be a special local Moufang set with U∞ abelian and |X| > 3. For any unit
x, we have hx = h−x and µx = µ−x . Hence µ2x = 1 for all units x.
Proof. We will show hx = h−x by proving yhx = yh−x for all y ∈ X. First assume y is a unit
with y 6∼ −x, then
(−y)µ−x = (−y)α−x τ −1 αxτ −1 τ α−x
= (−yαx )τ −1 αxτ −1 τ α−x
= (−yαx τ −1 )αxτ −1 τ α−x
(as −yαx is a unit)
= (−yαx τ −1 α−xτ −1 τ )α−x
= −yαx τ −1 α−xτ −1 τ αx
= −yµx ,
where we have repeatedly used commutativity of U∞ by
−1
−1
= −(y ′ αx′ ) .
(−y ′ )α−x′ = 0α−1
y ′ αx′ = 0(αy ′ αx′ )
From this, we get yµx = −(−yµx ) = −(−y)µ−x = yµ−x , so for any unit y with y 6∼ −x we have
yhx = yτ µx = yτ µ−x = yh−x .
Next, we look at the case where y ∼ −x. Take e a unit such that e 6∼ −x (such e exists as
|X| > 3), then yα−e is a unit and yα−e 6∼ −x, so we can use the previous case to get
yhx = yα−e αe hx
= (yα−e )hx αehx
= (yα−e )h−x αeh−x
= yα−e αe h−x
= yh−x .
(by Lemma 2.23)
(by the previous)
(by Lemma 2.23)
5
Special local Moufang sets and PSL2 (R)
18
Hence, we know yhx = yh−x for all units y.
Thirdly, we look at the case where y ∼ 0. Take any unit e, then yα−e is a unit, so we can
repeat the previous argument to get yhx = yh−x .
Finally, we need to cover the case y ∼ ∞. In this case,
yhx = yτ hτx τ −1
= yτ h−xτ τ −1
(by Lemma 2.21(iii))
−1
(by the previous case with yτ in place of y)
yτ hτ−x τ −1
(by Lemma 2.21(iii))
= yτ hxτ τ
=
= yh−x .
We now know that yhx = yh−x for all y ∈ X, so hx = h−x . It immediately follows that
µx = τ −1 hx = τ −1 h−x = µ−x .
To finish this section, we observe that our main example, M(R) for a local ring R, is indeed
a special local Moufang set.
Example 5.9. We have already mentioned that [1, r]τ = [1, r]τ −1 = [1, −r−1 ] for r ∈ R× , and
−[1, r] = [1, −r]. This means that
∼[1, r] = −([1, r]τ −1 ) τ = [1, r−1 ]τ = [1, −r] = −[1, r] ,
for any r ∈ R× , so M(R) is special.
5.2
Constructing a ring from a special local Moufang set
We have seen that M(R) is an example of a special local Moufang set. It is natural to ask what
conditions can be put on a local Moufang set to ensure that it is equal to M(R) for some local
ring R. With some additional assumptions, it is possible to recover the ring structure from the
local Moufang set, at least provided the characteristic of the residue field is different from 2. We
will use a method similar to the related result for Moufang sets [DW06, §6].
Construction B. Suppose that M is a local Moufang set satisfying the following conditions:
(R1)
(R2)
(R3)
(R4)
M is special;
U∞ is abelian;
the Hua subgroup H is abelian;
if x is a unit, then so is x · 2.
We consider the set R := X \ ∞, and define an addition and a multiplication. Note that there
is a bijection between R and U∞ by x 7→ αx . We define the addition on R as
x + y := 0 · αx αy
for all x, y ∈ R. This addition is simply the translation of the group composition in U∞ to the
set R. Since U∞ is an abelian group, so is (R, +). By Lemma 5.7, U∞ is uniquely 2-divisible,
hence also (R, +) is uniquely 2-divisible.
To define the multiplication, we first choose a fixed unit e ∈ R, which will be the identity
element of the multiplication. We will use the Hua maps corresponding to τ = µe , i.e. we use
hx = hx,µe = µe µx . Remark that by (R1), (R2) and (R4), the conditions of Lemma 5.8 are
5
Special local Moufang sets and PSL2 (R)
19
satisfied, so µe is an involution, and hence µe hx = µx . For any y ∈ R, we now define a map
Ry : X → X by
for y and y + e units,
(I)
x · he+y − x · hy − x
:=
− x · h−e+y + x + x · hy
for y a unit and y + e not a unit, (II)
xRy
x · h2e+y − x · he+y − x · h−2e + x for y not a unit.
(III)
(We will verify in the proof of Lemma 5.11 below that all Hua maps occurring in this definition
can indeed be defined.) Combining this with the unique 2-divisibility, we can now define the
multiplication on R by
xy := xRy /2
for all x, y ∈ R.
We will now prove that this structure is a local ring, so for the remainder of the subsection,
our local Moufang set satisfies (R1)–(R4). First, we observe that the action of any Hua map
on R is a group automorphism of (R, +):
Lemma 5.10. Let h be a Hua map. Then for any x, y ∈ R, we have (x + y)h = xh + yh. In
particular, (−x)h = −xh and (x/2)h = xh/2.
Proof. The first identity is Lemma 2.23 rewritten in terms of the addition in R. The second and
the third are immediate consequences, using the unique 2-divisibility.
Note that if h and h′ are Hua maps, we can define x(h + h′ ) := xh + xh′ , and the map h + h′
is also a group endomorphism of (R, +). In particular, each Rx is a group endomorphism of
(R, +).
Lemma 5.11. The structure (R, ·) is a commutative monoid with identity element e.
Proof. By the previous lemma and observation, we have (x/2)Ry = xRy /2. Furthermore, since
H is abelian, every two maps Rx and Ry commute. Hence
(xy)z = (xRy /2)Rz /2 = xRy Rz /4 = xRz Ry /4 = (xRz /2)Ry /2 = (xz)y
for all x, y, z ∈ R. This proves the identity
(xy)z = (xz)y
(5.1)
for all x, y, z ∈ R, which will be crucial for showing commutativity and associativity.
Next, we show that e is a left identity element for the multiplication, i.e. that ex = x for
all x ∈ R, and we will show this in each of the three cases, using Lemmas 5.2 and 5.3, which
translates to
xµx+y = −y − x + xµy − y for all x, y ∈ R such that x, y and x + y are units.
(I) In this case, x and x + e are units. We get
eRx = e · he+x − e · hx − e = eµe µe+x − eµe µx − e = (−e)µe+x − (−e)µx − e
= −eµe+x + eµx − e = −(−x − e + eµx − x) + eµx − e = 2x ,
so ex = eRx /2 = x.
5
Special local Moufang sets and PSL2 (R)
20
(II) In this case, x is a unit and x + e is not a unit. Hence x − e is a unit, or we would have
x − e ∼ 0 ∼ x + e, which would imply e ∼ −e, contradicting (R4). We get
eRx = −e · h−e+x + e + e · hx = −e · µe µ−e+x + e + e · µe µx
= −(−e)µ−e+x + e + (−e)µx = −(−x + e + (−e)µx − x) + e + (−e)µx = 2x ,
so ex = eRx /2 = x.
(III) In this case, x is not a unit, so x + e and x + 2e are units (since 2e 6∼ 0 and e is a unit).
We get
eRx = e · h2e+x − e · he+x − e · h−2e + e
= e · µe µ2e+x − e · µe µe+x − e · µe µ−2e + e
= −eµ2e+x + eµe+x − (−e)µ−2e + e
= −(−e − x − e + eµe+x − e − x) + eµe+x − (e + e + (−e)µe + e) + e = 2x ,
so ex = eRx /2 = x.
Substituting e for x in (5.1), we get yz = zy for all y, z ∈ R, so the multiplication is commutative.
In particular, xe = ex = x for all x ∈ R, so e is an identity element. Finally, we apply
commutativity to (5.1), and we get (yx)z = (xy)z = (xz)y = y(xz) for all x, y, z ∈ R, so the
multiplication is associative.
Theorem 5.12. The structure (R, +, ·) is a unital, commutative ring.
Proof. We know that (R, +) is an abelian group with identity element 0 (isomorphic to U ),
and that the multiplicative structure is a commutative monoid with identity element e. It only
remains to show the distributivity. We have seen before that the maps Rx , as linear combinations
of Hua maps, act linearly on (R, +). This means that (x + y)Rz = xRz + yRz , and so
(x + y)z = (xRz + yRz )/2 = xRz /2 + yRz /2 = xz + yz
for all x, y, z ∈ R. By commutativity, we also get x(y + z) = (y + z)x = yx + zx = xy + xz for
all x, y, z ∈ R. We conclude that R is indeed a unital, commutative ring.
Our next goal is to show that R is a local ring. To do this, we will identify the invertible
elements, and show that the non-invertible elements form an ideal.
Proposition 5.13. If x ∈ X is a unit, then x ∈ R is invertible with inverse x−1 := (−x)µe .
Proof. Note that we only need to show x−1 x = e, since by commutativity we will also get
xx−1 = e. We will again use Lemmas 5.2 and 5.3. Again, we need to proceed case by case, but
since x is a unit, only the two first cases occur.
(I) In this case, both x and x + e are units. We get
(−x)µe Rx = (−x)µe · he+x − (−x)µe · hx − (−x)µe = (−x)µe+x − (−x)µx − (−x)µe
= −xµx+e − x + xµe = −(−e − x + xµe − e) − x + xµe = 2e ,
so x−1 x = x−1 Rx /2 = e.
(II) In this case, x a unit and x + e not a unit (so x − e a unit), hence
(−x)µe Rx = −(−x)µe · h−e+x + (−x)µe + (−x)µe · hx
= −(−x)µ−e+x + (−x)µe + (−x)µx = xµx−e + (−x)µe + x
= (e − x + xµe + e) − xµe + x = 2e ,
so again x−1 x = x−1 Rx /2 = e.
5
Special local Moufang sets and PSL2 (R)
21
Observe that we have shown that the elements we called ‘units’ in the local Moufang set
correspond to the units in the ring R.
Proposition 5.14. The set m := 0 ⊆ R is an ideal in R.
Proof. Let x, y ∈ m; then x + y = xαy ∼ 0αy = y ∼ 0, so x + y ∼ 0 and x + y ∈ m. Also
−x = 0α−x ∼ xα−x = 0, so −x ∈ m. Next, we need to verify that m is closed under multiplication
with R. Take x ∈ m and r ∈ R. We get xr = xRr /2, but Rr is a linear combination of Hua
maps. Hence xRr is a sum of x · hri for some ri , and xhri ∼ 0hri = 0. Hence xRr ∼ 0, and (R4)
then implies that also xr ∼ 0. Hence m is an ideal.
We can now summarize our results of this section.
Theorem 5.15. Suppose that M is a local Moufang set satisfying (R1)–(R4). Then we can use
Construction B to get a local ring R from M, and 2 is invertible in R.
Proof. We have shown that m is an ideal in R. On the other hand, if r ∈ R\m, then r ∈ X\∞
is a unit, so m is exactly the set of non-invertible elements of R; it must therefore be the unique
maximal ideal of R. Since 2 = 2e is a unit, it is invertible in R.
5.3
Characterization of PSL2 (R) as a special local Moufang set
Our goal is to use Theorem 5.15 to characterize M(R) as a special local Moufang set satisfying
certain conditions. As a first step, we will apply Construction B on M(R); we will see that the
resulting local ring is indeed isomorphic to the ring R we started with.
Example 5.16. Let R be a local ring with 2 ∈ R× , and consider the special local Moufang set
M(R). Then U∞ is abelian and uniquely 2-divisible, as
2
1 r 1 s
1 r+s
1 s 1 r
1 r/2
1 r
=
=
and
=
0 1 0 1
0
1
0 1 0 1
0 1
0 1
for all r, s ∈ R. Secondly, the Hua
−1
r
0 s−1
0
r
0
subgroup is abelian, as
−1
0
(rs)−1 0
s
=
=
s
0
rs
0
0
s
−1
r
0
0
.
r
Finally, recall that [1, r] is a unit if and only if r ∈ R× , so since 2 is invertible, 2r ∈ R× and
[1, 2r] is a unit.
Hence we can perform Construction B on M(R) whenever 2 is invertible in R. We thus
get a new ring R′ := {[1, r] | r ∈ R}, and choose a unit [1, e]. The addition is given by
[1, r] + [1, s] = [1, r + s] for all r, s ∈ R, and a short computation shows that the multiplication
is given by
[1, r][1, s] = [1, re−1 s]
for all r, s ∈ R (in each of the three cases).
Theorem 5.17. If R is a local ring with residue field not of characteristic 2, then the ring R′
we get from M(R) using Construction B with unit [1, e] is isomorphic to R.
Proof. We define a bijection ϕ : R → R′ : r 7→ [1, re]. Then
ϕ(1) = [1, e] ,
ϕ(r + s) = [1, (r + s)e] = [1, re] + [1, se] = ϕ(r) + ϕ(s) ,
ϕ(rs) = [1, rse] = [1, ree−1 se] = [1, re][1, se] = ϕ(r)ϕ(s) ,
for all r, s ∈ R. We conclude that ϕ is a ring isomorphism.
5
Special local Moufang sets and PSL2 (R)
22
Remark 5.18. The ring R′ is, in fact, an isotope of R with new unit e, and we have simply
illustrated the (well known) fact that an isotope of an associative ring is always isomorphic to
the original ring.
Corollary 5.19. If M(R) is isomorphic to M(R′ ) for local rings R and R′ with residue field not
of characteristic 2, then R ∼
= R′ .
We will now characterize M(R) purely based on data from the local Moufang set. We will
need two extra assumptions on the local Moufang set, in addition to (R1)–(R4). Observe that
µ-maps are in this case involutions by Lemma 5.8.
Theorem 5.20. Let M be a local Moufang set satisfying (R1)–(R4). Let e and Rx be as in
Construction B. Assume furthermore that
xµe αy = yRx α−2e µe Rx µe
for all x ∼ 0 and y 6∼ ∞.
Then M is isomorphic to M(R), where R is the local ring obtained from Construction B.
Proof. We adopt the notations from Construction B for M, and we will denote the root group
U[0,1] of M(R) by U ′ . We will construct a bijection from X to P1 (R) preserving the equivalence,
a bijection from U∞ to U ′ , and an involution τ of M(R) such that the action of U∞ and µe on X
is permutationally equivalent with the action of U ′ and τ on P1 (R). By Remark 4.7 this will
show that M is indeed isomorphic to M(R).
By construction, R = X \ ∞, and we have ∞ = 0µe by the definition of µ-maps. So we define
(
[e, x]
if x ∈ R,
1
ϕ : X → P (R) : x 7→
[−xµe , e] if x ∈ ∞.
Note that in the second case xµe ∼
6 ∞, so this is indeed an element of the ring R. It is clear that
ϕ is a bijection; we claim that ϕ preserves the equivalence. First, if x, y ∈ R, then
x ∼ y ⇐⇒ xα−y ∼ 0 ⇐⇒ x − y ∼ 0 ⇐⇒ x − y ∈ m ⇐⇒ [e, x] ∼ [e, y] ,
where the last equivalence follows from (2.1) on p. 3. Secondly, if x ∼ y ∼ ∞, then xµe ∼ yµe ∼ 0,
so both are in m, and hence ϕ(x) ∼ ϕ(y). Finally, if x ∼ ∞ but y ∈ R, then again xµe ∼ 0, so
xµe ∈ m, hence
0 ϕ(x)
6∼ ϕ(y).
e
Let τ = −e
0 . It remains to show that the actions of U∞ and µe on X are permutationally
equivalent with the actions of U ′ and τ on P1 (R) via ϕ. For τ and µe , we compute, using µ2e = 1,
if xµe ∼ ∞ ⇐⇒ x ∼ 0 ,
[−x, e]
−1
ϕ(xµe ) = [e, xµe ] = [e, −x ]
if x ∈ R× ,
[e, xµe ]
if xµe ∼ 0 ⇐⇒ x ∼ ∞ ;
2
if x ∼ 0 ,
[e, x]τ = [−xe, e ] = [−x, e]
ϕ(x)τ = [−x, e] = [e, −x−1 ]
if x ∈ R× ,
2
[−xµe , e]τ = [−e , −xµe e] = [e, xµe ] if x ∼ ∞ ;
so ϕ(xµe ) = ϕ(x)τ . To show that the actions of U∞ and U ′ are the same, we first observe that
the map
θ : U∞ → U ′ : αx 7→ α[e,x]
References
23
for all x ∈ R, is a group isomorphism because
e x e y
e
θ(αx )θ(αy ) = α[e,x] α[e,y] =
=
0 e 0 e
0
for all x, y ∈ R. It only remains to show that
ϕ(xαy ) = ϕ(x)θ(αy )
x+y
= α[e,x+y] = θ(αx+y ) = θ(αx αy )
e
for all x ∈ X and y ∈ R.
We distinguish two cases: if x ∈ R, then
ϕ(xαy ) = ϕ(x + y) = [e, x + y] = [e, x]α[e,y] = ϕ(x)θ(αy ) .
′
If x ∼ ∞, we set x = xµ−1
e , which is then equivalent to 0, so
xαy = x′ µe αy = yRx′ α−2e µe Rx′ µe = (2yx′ − 2e)µe Rx′ µe
= −(2yx′ − 2e)−1 Rx′ µe = (−2−1 (yx′ − e)−1 )Rx′ µe
= (2(−2−1 (yx′ − e)−1 )x′ )µe = (−(yx′ − e)−1 x′ )µe ,
where we have used the fact that (2yx′ − 2e) is invertible because 2yx′ ∈ m. Since xαy ∼ ∞,
this implies
ϕ(xαy ) = [−(−(yx′ − e)−1 x′ )µe µe , e] = [(yx′ − e)−1 x′ , e] ,
where we use µ2e = 1. On the other hand,
ϕ(x)θ(αy ) = [−xµe , e]α[e,y] = [−x′ , e]
e y
0 e
= [−x′ e, −x′ y + e] = [(yx′ − e)−1 x′ , e] ,
and we conclude that ϕ(xαy ) = ϕ(x)θ(αy ) also in this case.
References
[DS09]
Tom De Medts and Yoav Segev. A course on Moufang sets. Innov. Incidence Geom.,
9:79–122, 2009. 2
[DST08] Tom De Medts, Yoav Segev, and Katrin Tent. Special Moufang sets, their root groups
and their µ-maps. Proc. Lond. Math. Soc. (3), 96(3):767–791, 2008. 15
[DW06] Tom De Medts and Richard M. Weiss. Moufang sets and Jordan division algebras.
Math. Ann., 335(2):415–433, 2006. 1, 15, 18
[Grü10] Matthias Grüninger. Special Moufang sets with abelian Hua subgroup. J. Algebra,
323(6):1797–1801, 2010. 1
[Loo14] Ottmar Loos. Rank one groups and division pairs. Bull. Belg. Math. Soc. Simon Stevin,
21(3):489–521, 2014. 8
[Seg09]
Yoav Segev. Proper Moufang sets with abelian root groups are special. J. Amer. Math.
Soc., 22(3):889–908, 2009. 15
[Tim01] Franz Georg Timmesfeld. Abstract root subgroups and simple groups of Lie type, volume 95 of Monographs in Mathematics. Birkhäuser Verlag, Basel, 2001. 15
[Tit92]
Jacques Tits. Twin buildings and groups of Kac-Moody type. In Groups, combinatorics
& geometry (Durham, 1990), volume 165 of London Math. Soc. Lecture Note Ser., pages
249–286. Cambridge Univ. Press, Cambridge, 1992. 1
| 4math.GR
|
Fast computation of all maximum acyclic agreement forests for
two rooted binary phylogenetic trees
Benjamin Albrecht
?
arXiv:1512.05656v1 [q-bio.PE] 17 Dec 2015
Institut für Informatik, Ludwig-Maximilians-Universität, Germany
Abstract. Evolutionary scenarios displaying reticulation events are often represented by rooted phylogenetic networks. Due to biological reasons, those events occur very rarely, and, thus, networks containing a minimum number of such events, so-called minimum hybridization networks, are of particular
interest for research. Moreover, to study reticulate evolution, biologist need not only a subset but all of
those networks. To achieve this goal, the less complex concept of rooted phylogenetic trees can be used
as building block. Here, as a first important step, the trees are disjoint into common parts, so-called
maximum acyclic agreement forests, which can then be turned into minimum hybridization networks
by applying further network building algorithms. In this paper, we present two modifications of the
first non-naive algorithm — called allMAAFs — computing all maximum acyclic agreement forests
for two rooted binary phylogenetic trees on the same set of taxa. By a simulation study, we indicate
that through these modifications the algorithm is on average 8 times faster than the original algorithm
making this algorithm accessible to larger input trees and, thus, to a wider range of biological problems.
Keywords Directed acyclic graphs · Hybridization · Maximum acyclic agreement forests · Bounded
search · Phylogenetics
?
Benjamin Albrecht
Institut für Informatik, Ludwig-Maximilians-Universität, Germany
Tel.: +49-89-2180-4069
E-mail: [email protected]
1
Introduction
Phylogenetic studies deal with the reconstruction of evolutionary histories, which are often represented by
rooted phylogenetic trees in which each node represents a certain speciation event. The representation of
reticulation events combining different evolutionary histories, however, requires the more general concept of
phylogenetic networks, which are more complex structures that can additionally consist of nodes of in-degree
larger than one. Hybridization, for instance, is an example for such a reticulation event merging a sizable
percentage of two genomes of two different species. Now, due to those reticulation events, for a set of species
there can exist incongruent gene trees, which are phylogenetic trees based on different genes sampled for this
particular set of species all representing different evolutionary histories. On should, however, keep in mind
that there can also exist other biological effects leading to incongruent gene trees, e.g. incomplete lineage
sorting, still allowing to model evolutionary histories by phylogenetic trees.
As it is the case for hybridization events, a reticulation event is in general expected to occur very
rarely. Thus, in order to study reticulate evolution, one is interested in the minimum number of such events
explaining a set of incongruent gene trees. A further more complicated but strongly connected problem is
the computation of rooted phylogenetic networks displaying evolutionary scenarios containing such events.
In such a network each node with in-degree one reflects a putative speciation event, whereas each node of
in-degree of at least two, a so-called reticulation node (or, in respect to hybridization, hybridization node)
represents a putative reticulation event.
Phylogenetic studies are often joint works between biologists, mathematicians and computer scientists;
usually biologists produce certain data, mathematicians develop a model for analyzing this data and computer
scientists implement programs, which enable the investigation of the data based on these models. Regarding
the investigation of reticulate evolution, the underlying problem is in general highly combinatorial, which
complicates the work of computer scientists as naive approaches are typically insufficient for analyzing
biologically relevant data.
In this work, we tackle this problem by focusing on the first non-naive algorithm allMAAFs [12] that
enables the computation of all maximum acyclic agreement forests for two rooted binary phylogenetic trees,
which can be considered, from a mathematical point of view, as an intermediate step in calculating all possible
rooted phylogenetic networks. This is the case, since those networks can then be computed by gluing the
components of such maximum acyclic agreement forests back together in a specific way which, for example,
has already been demonstrated by the algorithm HybridPhylogeny [4]. The computation of a maximum
acyclic agreement forest for two binary phylogenetic trees, however, is a well-known NP-hard problem [5]
and, thus, in order to apply this step to large input trees and thereby making the algorithm accessible
to a wider range of biological problems, we have worked out two modifications significantly improving the
practical running time of the algorithm allMAAFs.
Notice that so far there have been developed several approaches for analyzing reticulation events [1,
7, 8, 13]. Whereas some of these approaches just enable the computation of hybridization numbers, some
additionally can be used to generate a set of minimum hybridization networks each representing a particular
evolutionary scenario. Notice, however, that until now none of these methods is able to calculate all minimum
hybridization networks, which, as already emphasized, under a biological point of view, is an important
feature for studying reticulate evolution.
The paper is organized as follows. First, we give all basic notations and some well-known results of
phylogenetics research, which are used throughout this work. Next, we describe two modifications allMAAFs1 and allMAAFs2 of the algorithm allMAAFs by presenting its respective pseudo code. In a
subsequent step, the correctness of each of these modified algorithms is established by formal proofs and,
finally, in order to indicate the speedup obtained from applying these modifications, we present the results
of a simulation study comparing runtimes attained from its respective implementations. Lastly, we finish
this paper by briefly discussing the theoretical worst-case running time of both algorithms allMAAFs1 and
allMAAFs2 .
2
2
Preliminaries
In this section, we give all formal definitions that are used throughout this work for describing and discussing
the original algorithm allMAAFs as well as our two modified algorithms allMAAFs1 and allMAAFs2 .
These definitions correspond to those given in the work of Huson et. al. [10] and Scornavacca et. al. [12]. We
assume here that the reader is familiar with general graph-theoretic concepts.
Phylogenetic trees. A rooted phylogenetic X -tree T is a directed acyclic connected graph whose edges
are directed from the root to the leaves as defined in the following. There is exactly one node of in-degree 0,
namely the root of T . The set of nodes of out-degree 0 is called the leaf set of T and is labeled one-to-one by
a taxa set X , also denoted by L(T ). Here, the taxa set X usually consists of certain species or genes whose
evolution is outlined by T . The tree T is called binary if all of its nodes, except the root, provide an in-degree
of 1 and if all of its nodes, except all leaves (the so-called inner or internal nodes) provide an out-degree
of 2. Given a node v, the label set L(v) refers to all taxa that are contained in the subtree rooted at v. In
addition, given a set of rooted phylogenetic X -trees F, by L(F) we refer to the union of each leaf set of each
tree in F.
Now, given a set of taxa X 0 ⊆ L, the notation T (X 0 ) refers to the minimal connected subgraph of T
containing X 0 . A restriction of T to X 0 , shortly denoted by T |X 0 , is a rooted phylogenetic tree that is obtained
from T (X 0 ) by suppressing each node of both in- and out-degree 1. Moreover, given two phylogenetic trees
T1 and T2 in which X1 and X2 , with X1 ⊆ X2 , denotes the taxa set of T1 and T2 , respectively, we say T1 is
contained in T2 , shortly denoted by T1 ⊆ T2 , if T2 |X1 is isomorphic to T1 .
Next, the lowest common ancestor of a phylogenetic X -tree T in terms of a taxa set X 0 ⊆ X is the node v
in T with X 0 ⊆ L(v) such that there does not exist another node v 0 in T with X 0 ⊆ L(v 0 ) and L(v 0 ) ⊆ L(v).
In the following, such a node is shortly denoted by LCAT (X 0 ).
Lastly, given a set of phylogenetic trees F as well as an edge set E 0 that is contained in F such that for
each pair of edges e1 , e2 ∈ E 0 , with e1 6= e2 , e1 is not adjacent to e2 . Then, by F − E 0 we refer to the set
F 0 of trees that is obtained from F by cutting E 0 . More precisely, first each edge in E 0 is deleted in F and
then each node of both in- and out-degree 1 is suppressed. Notice that by deleting an edge of a tree F in
F, this tree is separated into two parts Fa and Fb so that the resulting forest equals F \ {F } ∪ {Fa , Fb }.
Consequently, after deleting E 0 from F the resulting set F 0 contains precisely |F| + |E 0 | trees.
During the algorithm allMAAFs, an agreement forests for two rooted binary phylogenetic X -trees T1
and T2 is calculated by cutting down one of both trees, say T2 , into several components such that each
component corresponds to a restricted subtree of T1 . In order to keep track of the component Fρ containing
the root, the root of T2 has to be a node that is marked in a specific way. Thus, throughout this paper, we
regard the root of each rooted binary phylogenetic X -tree as a node that is attached to its original root and
to a taxon ρ 6∈ X (cf. Fig. 1(a)).
Phylogenetic networks. A rooted phylogenetic network N on X is a rooted connected digraph whose
edges are directed from the root to the leaves as defined in the following. There is exactly one node of
in-degree 0, namely the root, and no nodes of both in- and out-degree 1. The set of nodes of out-degree 0 is
called the leaf set of N and is labeled one-to-one by the taxa set X , also denoted by L(N ). In contrast to
a phylogenetic tree, such a network may contain undirected but not any directed cycles. Consequently, N
can contain nodes of in-degree larger than or equal to 2, which are called reticulation nodes or hybridization
nodes. Moreover, all edges that are directed into such a reticulation node are called reticulation edges or
hybridization edges.
Hybridization Networks. A hybridization network N for two rooted binary phylogenetic X -trees T1
and T2 is a rooted phylogenetic network on X displaying T1 and T2 . More precisely, this means that for each
tree T ∈ {T1 , T2 } there exists a set E 0 ⊆ E(N ) of reticulation edges such that T can be derived from N by
conducting the following steps.
(1) Delete each edge from N that is not contained in E 0 .
3
(2) Remove each node whose corresponding taxon is not contained in X 0 .
(3) Remove each unlabeled node of out-degree 0 repeatedly.
(4) Suppress each node of both in- and out-degree 1.
Notice that, for determine E 0 , it suffices to select at most one in-edge of each hybridization node in N .
From a biological point of view, this means that N displays T if each speciation event of T is reflected by
N . Moreover, each internal node of in-degree 1 represents a speciation event and each internal node providing
an in-degree of at least 2 represents a reticulation event or, in terms of hybridization, a hybridization event.
This means, in particular, that such a latter node represents an individual whose genome is a chimaera of
several parents. Thus, such a node v of in-degree larger than or equal to 2 is called a hybridization node and
each edge directed into v is called a hybridization edge.
Now, based on those hybridization nodes, the reticulation number r(N ) of a hybridization network N is
defined by
X
r(N ) =
δ − (v) − 1 = |E| − |V | + 1,
(1)
v∈V :δ − (v)>0
where V denotes the node set and E the edge set of N . Next, based on the definition of the reticulation
number, for two rooted binary phylogenetic X -trees T1 and T2 the hybridization number h({T1 , T2 }) is
min{r(N ) : N is a hybridization network displaying T1 and T2 }.
(2)
Lastly, we call a hybridization network N for two rooted binary phylogenetic X -trees T1 and T2 a minimum
hybridization network if r(N ) = h({T1 , T2 }).
Forests. Let T be a rooted nonbinary phylogenetic X -tree T . Then, we call any set of rooted nonbinary
phylogenetic trees F = {F1 , . . . , Fk } with L(F) = X a forest on X , if we have for each pair of trees Fi and
Fj that L(Fi ) ∩ L(Fj ) = ∅. Moreover, if additionally for each component F in F the tree T |L(F ) equals
F , we say that F is a forest for T . Lastly, let F be a forest for a rooted binary phylogenetic X -tree T .
Then, by F we refer to the forest that is obtained from F by deleting each component only consisting of
an isolated node as well as the element containing the node labeled by taxon ρ if it contains at most one edge.
Agreement Forests. For technical purpose, the definition of agreement forests is based on two rooted
binary phylogenetic X -trees T1 and T2 whose roots are marked by a unique taxon ρ 6∈ X as follows. Let ri
be the root of the tree Ti with i ∈ {1, 2}. Then, we first create a new node vi labeled by a new taxon ρ 6∈ X
and then attach this node to ri by inserting the edge (vi , ri ). Notice that this case v1 and v2 is the new root
of T1 and T2 , respectively. Moreover, since we consider ρ as being a new taxon, the taxa set of both trees is
X ∪ {ρ}.
Now, given two such marked rooted binary phylogenetic trees T1 and T2 on X ∪ ρ, a set of components
F = {Fρ , F1 , . . . , Fk } is an agreement forest for T1 and T2 if the following three conditions are satisfied.
(1) Each component Fi with taxa set Xi equals T1 |Xi and T2 |Xi .
(2) There is exactly one component, denoted as Fρ , with ρ ∈ L(Fρ ).
(3) Let Xρ , X1 , . . . , Xk be the respective taxa sets of Fρ , F1 , . . . , Fk . All trees in {T1 (Xi )|i ∈ {ρ, 1, . . . , k}} as
well as {T2 (Xi )|i ∈ {ρ, 1, . . . , k}} are node disjoint subtrees of T1 and T2 , respectively.
An illustration of an agreement forest is given in Figure 1(b).
Lastly, an agreement forest for two rooted binary phylogenetic X -trees containing a minimum number
of components is called a maximum agreement forest. Now, based on maximum agreement forests, one can
compute a minimum hybridization network if an additional constraint is satisfied, which is presented in the
following. Notice that, from a biological point of view, this constraint is necessary since it prevents species
from inheriting genetic material from their own offspring.
Modified ancestor descendant graphs. Given two rooted binary phylogenetic X -trees T1 and T2 together with a forest F = {Fρ , F1 , . . . , Fk } for T1 (or T2 ), the modified ancestor descendant graph AG∗ (T1 , T2 , F)
with node set F contains a directed edge (Fi , Fj ), i 6= j, if
4
Fig. 1. (a) Two rooted binary phylogenetic X -trees T1 and T2 with taxa set X = {a, b, c, d, e, f, g, h, ρ}. (b) An acyclic
agreement forest F for T1 and T2 . (c) The directed graph AG∗ (T1 , T2 , F) not containing any directed cycles and,
thus, F is acyclic.
(1) the root of T1 (L(Fi )) is an ancestor of the root of T1 (L(Fj )),
(2) or the root of T2 (L(Fi )) is an ancestor of the root of T2 (L(Fj )).
An illustration of such an modified ancestor descendant graph is given in Figure 1(c).
Now, we say that F is an acyclic agreement forest if F satisfies the conditions of an agreement forest and if
the corresponding graph AG∗ (T1 , T2 , F) does not contain any directed cycles. Again, each acyclic agreement
forest of minimum size among all acyclic agreement forests is called a maximum acyclic agreement forest.
Note that the definition of an ancestor descendant graph given in both works Baroni et. al. [3] as well as
Scornavacca et. al. [12] is different to the one presented above. Regarding the definitions of those two works,
the set of components F has to be an agreement forest for T1 and T2 and, thus, this definition is more
strict than the one given for AG∗ (T1 , T2 , F) in which F just has to be a forest for T1 (or T2 ). For a better
illustration, in Figure 1 we give an example of an acyclic agreement forest together with its underlying
modified ancestor descendant graph.
Notice that the concept of maximum acyclic agreement forests for two rooted binary phylogenetic X -trees
is of interest, since its size minus one corresponds to the hybridization number of these two trees, which is
stated by Theorem 1 given in the paper of Baroni et. al. [3].
Theorem 1 ([3]). Let F be a maximum acyclic agreement forest of size k for two rooted binary phylogenetic
X -trees T1 and T2 , then h(T1 , T2 ) = k.
Cherries. For a rooted binary phylogenetic X -tree T we call two of its leaves a and c a cherry, denoted by
{L(a), L(c)}, if both nodes have the same parent. Moreover, let R and S be two rooted binary phylogenetic
trees on XR and XS , respectively, so that XR ⊆ XS , and let F be a forest for S. Then, a cherry {L(a), L(c)}
in R is called a common cherry of R and F, if there exists a cherry {L(a0 ), L(c0 )} in F with L(a0 ) = L(a) and
L(c0 ) = L(c). Otherwise, the cherry {L(a), L(c)} in R is called contradicting cherry of R and F. Moreover,
in order to ease reading, if there exists a component in F containing two leaves labeled by L(a) and L(c),
respectively, we write a ∼F c. Otherwise, if such a component does not exist, we write a 6∼F c.
5
Moreover, let F be a forest for a rooted binary phylogenetic X -tree and let {L(a), L(c)} be a cherry that
is contained in F in which e denotes the in-edge of its parent p. Then, if e exists, by F ÷ {L(a), L(c)} we
simply refer to the forest F − {e} and, otherwise, if p has in-degree 0, to F. Furthermore, let P be the path
connecting a and c in F. Then, the set of pendant edges for {L(a), L(c)} contains each edge (v, w) with
v ∈ V (P) \ {a, c} and w 6∈ V (P), where V (P) denotes the set of nodes in P. Notice that in general there
exist several, namely precisely |V (P)| − 3, edges satisfying the condition of such an edge.
Cherry Reductions. Let F be a forest for a rooted binary phylogenetic X -tree and let {L(a), L(c)} be
a cherry of a component Fi in F. Then, a cherry reduction, according to a cherry {L(a), L(c)} in one of its
components Fi , implies the following two operations.
(1) The parent of the two nodes a and c is labeled by {L(a), L(c)}.
(2) Both nodes a and c together with their adjacent edges are deleted from Fi .
Throughout the algorithm, such a reduction step of a common cherry {L(a), L(c)} in F is shortly denoted
by F[{L(a), L(c)} → L(a)∪L(c)]. Furthermore, the reverse notation F[L(a)∪L(c) → {L(a), L(c)}] describes
the insertion of two new taxa labeled by L(a) and L(c), respectively, together with removing the label
L(a) ∪ L(c) from its parent. Notice that, for applying such insertion steps, one has to keep track of the
preceding reduction steps.
Equivalently, for a cherry {L(a), L(c)} in a rooted binary phylogenetic X -tree R we write R[{L(a), L(c)} →
L(a) ∪ L(c)] to denote a cherry reduction of {L(a), L(c)} in R.
3
The Algorithm allMAAFs
In this section, we present the algorithm allMAAFs that was first published in the work of Scornavacca et
al. Scornavacca et. al. [12]. To increase its readability, we decided to split the original algorithm allMAAFs
into six parts (cf. Alg. 1–5). Note that, apart from its graphical representation, our presentation of the
algorithm allMAAFs together with its terminology in general adheres to the original algorithm.
Algorithm 1: allMAAFs(S, T , R, F, k, M )
1
2
3
4
5
6
Data: Two rooted binary phylogenetic X -trees S and T , a rooted binary phylogenetic tree R and a forest F such that
L(R) = L(F ) and L(T ) = L(F ), an integer k, and a list M that contains information of previously reduced cherries.
Result: A set F of forests for F and an integer. In particular, if F = T , R = S, M = ∅, and k ≥ h(S, T ) is the input to
allMAAFs, the output precisely consists of all maximum-acyclic-agreement forests for S and T and their respective
hybridization number.
if k < 0 then
return (∅, k − 1);
if |L(R)| = 0 then
F 0 ← cherryExpansion(F , M );
if AG(S, T, F 0 ) is acyclic then
return (F 0 , |F 0 | − 1);
else
7
return (∅, k − 1);
8
9
10
11
12
13
14
else
let {L(a), L(c)} be a cherry of R;
if {L(a), L(c)} is a common cherry of R and F then
return (ProcessCommonCherry(S, T , R, F , k, M , {L(a), L(c)}));
if k =
6 (|F | − 1) or {L(a), L(c)} is a contradicting cherry of R and F then
return (ProcessContradictingCherry(S, T , R, F , k, M , {L(a), L(c)}));
6
Algorithm 2: cherryExpansion(F, M )
3
while M is not empty do
M ← remove last element of M, say{L(a), L(c)};
F ← F [L(a) ∪ L(c) → {L(a), L(c)}];
4
return F
1
2
Algorithm 3: cherryReduction(R, F, M, {L(a), L(c)})
1
2
3
4
M 0 ← Add {L(a), L(c)} as last element of M ;
R0 ← R[{L(a), L(c)} → L(a) ∪ L(c)];
F 0 ← F [{L(a), L(c)} → L(a) ∪ L(c)];
return (R0 , F 0 , M 0 )
Algorithm 4: ProcessCommonCherry(S, T , R, F, k, M , {L(a), L(c)})
1
2
3
4
5
6
7
8
9
(R0 , F 0 , M 0 ) ← cherryReduction(R, F , M , {L(a), L(c)});
(Fr , kr ) ← allMAAFs(S, T , R0 |L(F 0 ) , F 0 , k, M 0 );
if Fr =
6 ∅ then
k ← min(k, kr );
if (k = |F | − 1) then return (Fr , k);
else
(Fa , ka ) ← allMAAFs(S, T , R|L(F −{e }) , F − {ea }, k − 1, M );
a
if Fa 6= ∅ then
k ← min(k, ka − 1);
if (ka − 1 = k) then F ← F ∪ Fa ;
(Fc , kc ) ← allMAAFs(S, T , R|L(F −{e }) , F − {ec }, k − 1, M );
c
if Fc 6= ∅ then
k ← min(k, kc − 1);
10
11
12
13
F ← ∅;
if (ka − 1 = k) then F ← Fa ;
if (kc − 1 = k) then F ← F ∪ Fc ;
if (kr = k) then F ← F ∪ Fr ;
return (F , k);
14
15
16
17
18
Algorithm 5: ProcessContradictingCherry(S, T , R, F, k, M , {L(a), L(c)})
1
2
3
4
5
6
(Fa , ka ) ← allMAAFs(S, T , R|L(F −{e }) , F − {ea }, k − 1, M );
a
if Fa 6= ∅ then
k ← min(k, ka − 1);
(Fc , kc ) ← allMAAFs(S, T , R|L(F −{e }) , F − {ec }, k − 1, M );
c
if Fc 6= ∅ then
k ← min(k, kc − 1);
11
F ← ∅;
if a F c then
if (ka − 1 = k) then F ← Fa ;
if (kc − 1 = k) then F ← F ∪ Fc ;
return (F , k);
12
else
7
8
9
10
13
(FB , kB ) ← allMAAFs(S, T , R|L(F −{e
14
if FB 6= ∅ then
k ← min(k, kB − 1);
15
16
17
18
19
B })
, F − {eB }, k − 1, M );
if (ka − 1 = k) then F ← Fa ;
if (kB − 1 = k) then F ← F ∪ FB ;
if (kc − 1 = k) then F ← F ∪ Fc ;
return (F , k);
7
4
Modifications to allMAAFs
In this section, we present two modifications of the algorithm allMAAFs that, on the one hand, do not
improve its theoretical runtime but, one the other hand, significantly improve its practical running time,
which will be indicated by a simulation study reported in Section 6.
The first modification improves the processing of a certain type of contradicting cherry whereas for the
other two modifications only the processing step of a common cherry is of interest.
Given a contradicting cherry {L(a), L(c)} for R and F with a ∼F c, the original algorithm conducts three
recursive calls. One by recursively calling the algorithm with F −{ea } and R|F −{ea } , one by recursively calling
the algorithm with F − {ec } and R|F −{ec } and one by recursively calling the algorithm with F − {eB } and
R|F −{eB } , in which ea and ec refers to the in-edge of leaf a and c, respectively, in F and eB refers to an
in-edge of a subtree lying on the path connecting a and c in F. Regarding the latter recursive call, in the
upcoming part of this work we will show that, in order to compute all maximum acyclic agreement forests,
instead of cutting just one in-edge eB one can cut all of those in-edges all at once.
Moreover, given a common cherry {L(a), L(c)} for R and F, the original algorithm allMAAFs always
branches into three new computational paths; one path corresponding to the cherry reduction of {L(a), L(c)}
and two corresponding to the deletion of both in-edges of the two leaves a and c (cf. Alg. 4, line 7–15). To
understand the sense of our two modifications, one has to take the necessity of these two additional edge
deletions into account.
Therefor, we demonstrate a specific scenario that is outlined in Figure 2 showing two phylogenetic trees
T1 and T2 as well as a maximum acyclic agreement F = {((a, b), ρ), (e, d), (g, h), c, f } for those two trees. By
running the algorithm allMAAFs for T1 and T2 without deleting in-edges of a common cherry, the given
maximum acyclic agreement forest F is never computed. This is due to the fact that, once the component
Fi = ((g, h), f ) occurs on any computational path, Fi will be part of the resulting agreement forest. Because
of Fi and the component (e, d), such a resulting agreement forest is never acyclic and, thus, does not satisfy
the conditions of an acyclic agreement forest. However, by cutting instead of contracting the common cherry
{(g, h), f }, the resulting agreement forest turns into the maximum acyclic agreement forest F. This example
implies that sometimes the deletion of in-edges corresponding to taxa of a common cherry is necessary, which
is, however, in practice not often the case, and, thus, the original algorithm allMAAFs usually produces a
lot of additional unnecessary computational steps compared with our second modification offering a different
solution for such a scenario.
4.1
The Algorithm allMAAFs1
Our first algorithm allMAAFs1 is a modification of the original algorithm allMAAFs improving the
processing of contradicting cherries. Let {L(a), L(c)} be a contradicting cherry of R and F such that a ∼F c
holds and let eB be an edge that is defined as follows. Let P be the path connecting a and c in F. Then, the
edge set EB contains each edge eB = (v, w) with v ∈ V (P) \ {a, c} and w 6∈ V (P), where V (P) denotes the
set of nodes of P. Note that in this case there exist precisely |V (P)−3| edges satisfying the conditions of such
an edge eB . Now, if such a cherry {L(a), L(c)} occurs, the original algorithm allMAAFs branches into a
computational path by cutting exactly one of those edges in EB . We will show, however, that in this case the
whole set EB can be cut from F all at once without having an impact on the computation of all maximum
acyclic agreement forests for both input trees. In Algorithm 6, we give a pseudo code of allMAAFs1 . Note
that, for the sake of clarity, we just present the modified part in respect of the original algorithm dealing
with the processing of contradicting cherries. The remaining parts are unmodified and can be looked up in
Section 3.
4.2
The Algorithm allMAAFs2
Our third algorithm allMAAFs2 is again a modification of our first algorithm allMAAFs1 and is based on
a tool turning agreement forests into acyclic agreement forests. This tool, published by Whidden et al. Whidden et. al. [14], is based on the concept of an expanded cycle graph refining cyclic agreement forest. Due to such
8
Fig. 2. (a) Two rooted binary phylogenetic X -trees T1 and T2 with taxa set X = {a, b, c, d, e, f, g, h, ρ}. (b) An acyclic
agreement forest F for T1 and T2 . (c) The directed graph AG∗ (T1 , T2 , F) not containing any directed cycles and,
thus, F is acyclic.
Algorithm 6: ProcessContradictingCherry1 (S, T , R, F, k, M , {L(a), L(c)})
1
2
3
4
5
6
(Fa , ka ) ← allMAAFs1 (S, T , R|L(F −{e }) , F − {ea }, k − 1, M );
a
if Fa 6= ∅ then
k ← min(k, ka − 1);
(Fc , kc ) ← allMAAFs1 (S, T , R|L(F −{e }) , F − {ec }, k − 1, M );
c
if Fc 6= ∅ then
k ← min(k, kc − 1);
11
F ← ∅;
if a F c then
if (ka − 1 = k) then F ← Fa ;
if (kc − 1 = k) then F ← F ∪ Fc ;
return (F , k);
12
else
7
8
9
10
13
14
15
16
17
18
19
(FB , kB ) ← allMAAFs1 (S, T , R|L(F −E
B)
, F − EB , k − |EB |, M );
if FB 6= ∅ then
k ← min(k, kB − 1);
if (ka − 1 = k) then F ← Fa ;
if (kB − 1 = k) then F ← F ∪ FB ;
if (kc − 1 = k) then F ← F ∪ Fc ;
return (F , k);
9
additional refinement steps, which are performed right after the computation of each maximum agreement
forest, both cutting steps for processing a common cherry can be omitted. The simulation study in Section 6
indicates that this refinement step is efficient enough so that this modification in general outperforms the
original algorithm allMAAFs as well as our first modification allMAAFs1 . In Algorithm 7 and 8, we
present a pseudo code describing the algorithm allMAAFs2 . Again, for the sake of clarity, we restrict the
presentation to only those parts that are modified in respect to allMAAFs1 . Moreover, we omit a description of the subroutine conducting the refinement step, denoted by RefineForest, as it can be looked up in
the work of Whidden et al. Whidden et. al. [14]. Notice that, due to this refinement step, an implementation
of this modification is quite more expensive compared with our first modified algorithm allMAAFs1 .
Algorithm 7: allMAAFs2 (S, T , R, F, k, M )
1
2
3
4
5
6
7
8
9
if k < 0 then
return (∅, k − 1);
if |L(R)| = 0 then
F 0 ← cherryExpansion(F , M );
F 00 ← RefineForest(F 0 );
F ← ∅;
foreach F 00 ∈ F 00 do
if |F 00 | = k then
F ← F ∪ F 00 ;
else if |F 00 | < k then
k ← |F 00 |;
F ← {F 00 };
10
11
12
return (F , k − 1);
13
14
15
16
17
18
19
else
let {L(a), L(c)} be a cherry of R;
if {L(a), L(c)} is a common cherry of R and F then
return (ProcessCommonCherry3 (S, T , R, F , k, M , {L(a), L(c)}));
if k =
6 (|F | − 1) or {L(a), L(c)} is a contradicting cherry of R and F then
return (ProcessContradictingCherry(S, T , R, F , k, M , {L(a), L(c)}));
Algorithm 8: ProcessCommonCherry3 (S, T , R, F, k, M , {L(a), L(c)})
1
2
3
4
(R0 , F 0 , M 0 ) ← cherryReduction(R, F , M , {L(a), L(c)});
(Fr , kr ) ← allMAAFs2 (S, T , R0 |L(F 0 ) , F 0 , k, M 0 );
if (kr = k) then return (Fr , k);
return (∅, k);
10
5
Proofs of Correctness
In this section, we give formal proofs showing the correctness of all two modified algorithms presented in
Section 4. In a first step, however, we give some further definitions that are crucial for what follows.
5.1
The algorithm processCherries
In the following, we introduce the algorithm processCherries, which has already been utilized in the
paper of Scornavacca et al. Scornavacca et. al. [12]. This algorithm is a simplified version of the algorithm
allMAAFs describing one of its computational paths by a list of cherry actions. Notice that this algorithm
is a major tool that will help us to establish the correctness of our two modified algorithms.
Cherry actions. Let R be a rooted binary phylogenetic X -tree and let F be a forest for R. Then, a
cherry action ∧= ({L(a), L(c)}, e) is a tuple containing a set {L(a), L(c)} of two taxa of two leaves a and c
as well as an edge e. We say that ∧ is a cherry action for R and F, if {L(a), L(c)} is a cherry of R and if,
additionally, one of the following three conditions is satisfied.
(1) Either {L(a), L(c)} is a common cherry of R and F and e ∈ {∅, ea , ec },
(2) or {L(a), L(c)} is a contradicting cherry of R and F with a 6∼F c and e ∈ {ea , ec },
(3) or {L(a), L(c)} is a contradicting cherry of R and F with a ∼F c and e ∈ {ea , eB , ec }.
In this context, ea and ec is an edge in F adjacent to both leaves a and c, respectively. Moreover, eB is
part of the set of pendant edges for {L(a), L(c)}.
V
Cherry lists. Now, given two rooted binary phylogenetic XV-trees T1 and T2 , we say that
is a cherry
list for T1 and T2 , if, while calling
processCherries(T
,
{T
},
),
in
the
i-th
iteration
∧
is
a
cherry
action
1
2
i
V
V
for Ri and Fi . Note that, if
is not a cherry list for T1 and T2 , calling processCherries(T1 , {T2 }, )
(cf. Alg. 9) returns the empty set.
Algorithm 9: processCherries(R, F,
1
2
3
4
5
6
7
8
else
F ← F − {ei };
R ← R|L(F ) ;
10
11
13
= (∧1 , . . . ,∧n ))
M ← ∅;
foreach i ∈ 1, . . . , n do
if ∧i is a cherry action for R and F then
({L(a), L(c)}, ei ) ←∧i ;
if ei = ∅ then
M ← Add {L(a), L(c)} as last element of M ;
R ← R[{L(a), L(c)} → L(a) ∪ L(c)];
F ← F [{L(a), L(c)} → L(a) ∪ L(c)];
9
12
V
else
return ∅;
16
while M is not empty do
M ← remove the last element, say {L(a), L(c)}, from M ;
F ← F [L(a) ∪ L(c) → {L(a), L(c)}];
17
return R, F , M
14
15
Further details as well as an example of processCherries can be found in the work of Scornavacca et
al. Scornavacca et. al. [12].
Next, we will introduce the algorithm processCherries1 . This algorithm is a simplified version of the
algorithm allMAAFs1 describing one of its computational paths by a list of extended cherry actions. Notice
11
that this algorithm is a major tool that will help us to establish the correctness of our first modified algorithm
allMAAFs1 .
Extended cherry action. Let R be a rooted binary phylogenetic X -tree and let F be a forest for R.
Then, an extended cherry action ∧= ({L(a), L(c)}, E 0 ) is a tuple containing a set {L(a), L(c)} of two taxa
of two leaves a and c as well as an edge set E 0 . We say that ∧ is an extended cherry action for R and F, if
{L(a), L(c)} is a contradicting cherry of R and F with a ∼F c and E 0 = EB , where EB refers to the set of
pendant edges for {L(a), L(c)} in F.
V
Extended cherry list. Now, given two rooted binary phylogenetic X -trees T1 and
T2 , we say that is
V
an extended cherry list for T1 and T2 , if, while calling processCherries1 (T1 , {T2 }, ),
Vin the i-th iteration
∧i is either a cherry action or an extended cherry action for RiVand Fi . Note that, if
is not an extended
cherry list for T1 and T2 , calling processCherries1 (T1 , {T2 }, ) (cf. Alg. 10) returns the empty set.
Algorithm 10: processCherries1 (R, F,
1
2
3
4
5
6
7
8
V
= (∧1 , . . . ,∧n ))
M ← ∅;
foreach i ∈ 1, . . . , n do
if ∧i is a cherry action for R and F then
({L(a), L(c)}, ei ) ←∧i ;
if ei = ∅ then
M ← Add {L(a), L(c)} as last element of M ;
R ← R[{L(a), L(c)} → L(a) ∪ L(c)];
F ← F [{L(a), L(c)} → L(a) ∪ L(c)];
else
9
F ← F − {ei };
R ← R|L(F ) ;
10
11
15
else if ∧i is an extended cherry action for R and F then
({L(a), L(c)}, Ei ) ←∧i ;
F ← F − Ei ;
R ← R|L(F ) ;
16
else
12
13
14
17
return ∅;
20
while M is not empty do
M ← remove the last element, say {L(a), L(c)}, from M ;
F ← F [L(a) ∪ L(c) → {L(a), L(c)}];
21
return R, F , M
18
19
V0
Lemma 1. Let
be an extended cherry
V0 list for two rooted binary phylogenetic
V X -trees T1 and T2 . Moreover,
let F be a forest for T2 calculated by . Then, there also exists a cherry list for T1 and T2 calculating F.
V0
so that the
Proof. To proof this Vlemma we will show how to replace each extended cherry action of
resulting cherry list
still computes F. Let ∧0i = ({L(a), L(c)}, EB ) be an extended cherry action with
EB = {e1 , e2 , . . . , ek }. Then, we can replace ∧0i through the sequence of cherry actions
({L(a), L(c)}, e1 ), ({L(a), L(c)}, e2 ), . . . , ({L(a), L(c)}, ek ).
0
Since by these cherry actions the same edges are cut from T2 as by ∧0i , the topology of Ri+1
equals Ri+k+1 ,
which directly implies that F is still computed.
5.2
Correctness of allMAAFs1
In this section, we will discuss the correctness of our first modified algorithm allMAAFs1 by establishing
the following theorem.
12
Theorem 2. Let T1 and T2 be two rooted binary phylogenetic X -trees and k ∈ N. Calling
allMAAFs1 (T1 , T2 , T1 , {T2 }, k, ∅)
returns all maximum acyclic agreement forests for T1 and T2 if and only if k ≥ h(T1 , T2 ).
V
Proof. Let T1 and T2 be two rooted binary phylogenetic X -trees and let
be a cherry list for T1 and T2
mimicking a computational path of allMAAFs calculating a maximum acyclicVagreement forest F for T1
and T2 . Then, in the following, we say a cherry action ∧i = ({L(a), L(c)}, ei ) in is a special cherry action,
if {L(a), L(c)} is a contradicting cherry of Ri and Fi and if ei is contained in the set of pendant edges for
{L(a), L(c)}. Note that, whereas such special cherry actions may occur in a computational path of allMAAFs, this is not the case for a computational path of allMAAFs1 . In the following, however, we will
show that the algorithm allMAAFs calculates a maximum acyclic agreement forest F for T1 and T2 if and
only if F is calculated by allMAAFs1 .
Lemma 2. Let T1 and T2 be two rooted binary phylogenetic X -trees and let F be a maximum acyclic agreement forest for T1 and T2 of size k. Then, F is calculated by calling allMAAFs(T1 , T2 , T1 , T2 , k, ∅) if and
only if F is calculated by calling allMAAFs1 (T1 , T2 , T1 , {T2 }, k, ∅).
Proof. ’=⇒’: From Lemma 1 we can directly deduce that if there exists a computational path of allMAAFs1 calculating F, then, there also exists a computational path in allMAAFs calculating F.
V
’⇐=’: Let ∧j = ({L(a), L(c)}, ej = ∅) be a cherry action of contracting both leaves a and c. Then, we
say a preceding special cherry action ∧i = ({L(x), L(y)}, ei ) (i < j) refers to ∧k if the following condition is
satisfied. Let ai and ci be the lowest common ancestor of L(a) and L(c) in Ri , respectively. Then, ei 6= ∅ has
to be a pendant edge lying on the path connecting both nodes ai and ci . In such a case, we say the special
cherry action is either of Type A or Type B (see definition below). Otherwise, if a special cherry action does
not refer to another cherry action, we say this cherry action is of Type C.
Now, based on the edge ei = (v, w), being part of the special action ∧i referring to the cherry action ∧j ,
we further distinguish whether ∧i is either of Type A or of Type B. Therefore, let T (w) be the subtree rooted
at w and let L(T (w)) be the set of taxa being contained in T (w). Now, we say ∧i is of Type A if there exists
a forest Fk and a tree Rk with i < k < j so that both of the following two conditions are satisfied.
(i) Each taxon in L(T (w)) is part of a taxa set of an isolated node. Notice that, as a direct consequence,
there exists a leaf w0 in Rk with label L(w).
(ii) The sibling s0 of w0 in Rk is a leaf. Notice that, as a direct consequence, {L(s0 ), L(w0 )} is a cherry of Rk .
Otherwise, if there does not exist such a forest Fk and such a tree Rk satisfying these two conditions, we
say ∧i is of Type B. An illustration of these two types is given in Figure 3
V
Notice that, due to the so chosen definition, each special cherry action in
has to be either of Type
V A,
of Type B, or of Type C. Next, based on this observation, we will show in three steps how to turn
into
another cherry list for T1 and T2 not containing any special cherry actions, but still calculating F, which
can be briefly summarized as follows.
V
In a first step, we will show how to modify
by replacing each special cherry action of Type A and of
Type C through a non-special cherry action so that the result is still a cherry list for T1 and T2 calculating
F. Next, during a second and a third step, we will show how to replace each set of special cherry actions
of Type B all referring to the same cherry action by a single extended cherry action so that the result is an
extended cherry list for T1 and T2 still calculating F.
Step 1. Let ∧i = ({L(x), L(y)}, ei = (v, w)) V
be a special cherry action of Type A or of Type C, and let
∧k with i < k < j be the first cherry action of
in which in Fk each taxon in L(T (w)) is contained inVa
0
taxa set of an isolated node so that in Rk the sibling s0 of the leaf labeled by L(w) is also leaf. Then, let
13
Fig. 3. An illustration of the cherry list (({a, c}, e1 ), ({a, c}), e2 ), ({a, c}), e3 ), ({a, c}), ∅)) in which ({a, c}, e1 ) is a
special cherry action of Type B and ({a, c}, e2 ) is a special cherry action of Type A. Note that the expanded cherry
list (({b, d}, e2 ), ({a, c}, {e1 , e3 }), ({a, c}), ∅)) applied to (i) yields the same scenario as depicted in (v) without making
use of any special cherry actions.
14
Fig. 4. An illustration of the scenario as described in Step 1. The figure shows a comparison between two sequences
of Ri -trees; one corresponding to the original cherry list and one corresponding to the modified cherry list, in which
a special cherry action of Type A is shifted from position i to k + 1.
V
be a cherry list that is obtained from
by first removingV∧i = ({L(x), L(y)}, ei ) and then by inserting the
0
cherry action ({L(w0 ), L(s0 )}, ew0 ) right after ∧k , so that
equals
(∧01 , . . . , ∧0n ) = (∧1 , . . . , ∧i−1 , ∧i+1 , . . . , ∧k , ∧0k , . . . , ∧0n ),
where ∧0k = ({L(w0 ), L(s0 )}, ew0 ) with ew0 being the in-edge of w0 .
0
Now, for the following, remember
that
V
V0 Rl (or Rl ) refers to the input 0tree occurring during iteration l
while processing the cherry list
(or
). Moreover, we write Ri =∧ V
Rj if both trees contain the same
set of cherries. Then, based on the position of the cherry action ∧l in , we can make the following five
observations (cf. Fig. 4).
– If l < i, then Rl0 equals Rl : This is the case because R1 equals R10 and through ∧l and ∧0l the same tree
operation is performed on Rl and Rl0 , respectively.
0
0
=∧ Rl : This is the case because Ri−1 equals Ri−1
, in Rl the node w cannot be part
– If l = i, then Rl−1
of a cherry (due to the definition of a special cherry action of Type A) and, thus, through ∧l (=∧i ) the
set of cherries in Ri−1 remains unchanged.
0
0
– If i < l < k + 1, then Rl−1
=∧ Rl : This is the case because Ri =∧ Ri−1
and again in Rl node w cannot
0
be part of a cherry. Thus, the two cherry actions ∧l and ∧l−1 produce the same set of cherries in Rl and
0
Rl−1
.
0
0
– If l = k + 1, then Rk+1
equals Rk+1 : This is the case because Rk =∧ Rk−1
and through ∧0k first node w is
0
cut and then removed from Rk−1 . Consequently, through the cherry actions ∧q and ∧0q , with 1 ≤ q < k+1,
0
the the same tree operations are applied to R1 and R10 and, thus, Rk+1 equals Rk+1
.
0
0
– If l > k + 1, then Rl equals Rl : This is the case because Rk+1 equals Rk+1 and through ∧l and ∧0l the
same tree operation is performed on Rl and Rl0 , respectively.
V0
Due to these observations
and since through
still the same edges are cut from T2 as through the
V V0
original cherry list ,
is a cherry list for T1 and T2 still calculating F. Now, by consecutively replacing
V
V(1)
all special cherry actions of Type A and Type C we can turn
into the cherry list
for T1 and T2 only
containing special cherry actions of Type B and still computing F. Next, we will show how to remove each
of those remaining special cherry actions.
V∗
V
V∗
| = k be a set of special cherry
actions of Type B all referring to a cherry
Step 2. Let
⊂
with |
V∗
action ∧j = ({L(a), L(c)}, ∅), and let i = mini0 {∧i0 :∧i0 ∈
} with ∧i = ({L(x), L(z)}, ei ). Moreover, let
V0
V(1)
V∗
be the cherry list that is obtained from
as follows. First
is
V∗ the cherry of each cherry action in
replaced by {L(a), L(c)} and then all those cherry actions in
are rearranged such that they are placed in
sequential order directly right before position j.
V0
This means, in particular, that
contains a sequence of special cherry actions
15
(1)
Fig. 5. An illustration of the scenario as described in Step 2. The figure shows a comparison of the two sequences of
Ri -trees; one corresponding to the original cherry list and one corresponding to the modified cherry list, in which a
sequence of k special cherry actions of Type B, beginning at position i, is shifted to position j − k.
..
.
∧j−k = ({L(a), L(c)}, ej−k = (vj−k , wj−k )),
∧j−k+1 = ({L(a), L(c)}, ej−k+1 = (vj−k+1 , wj−k+1 )),
..
.
∧j−1 = ({L(a), L(c)}, ej−1 = (vj−1 , wj−1 )),
∧j = ({L(a), L(c)}, ∅),
..
.
V∗
in which each edge of
is contained in the set of pendant edges for the cherry {L(a), L(c)}. Now,Vjust for
∗
convenience, in the following we assume, without loss of generality, that all those cherry actions in
occur
in sequential order beginning at position i. Moreover, for the following, let W be the set of target nodes of
V∗
V(1)
each
and let Rl (or Rl0 ) be the input tree of iteration l while processing the cherry list
(or
V0 edge in
). Additionally, again we write RiV=∧ Rj0 if both trees contain the same set of cherries. Then, based on
the position of a cherry action ∧l in , we can make the following five observations (cf. Fig. 5).
– If l < i, then Rl0 equals Rl : This is the case because R1 equals R10 and through ∧l and ∧0l the same tree
operation is performed on Rl and Rl0 , respectively.
0
0
– If i − 1 < l < i + k, then Rl =∧ Ri−1
: This is the case because Ri−1 equals Ri−1
and, since in Rl each
wi in W cannot be part of a cherry, through ∧l the set of cherries in Rl remains unchanged.
0
0
– If i + k − 1 < l < j, then Rl−k
=∧ Rl : This is the case because Ri+k−1 =∧ Ri−1
and again in Rl each
0
wi in W cannot be part of a cherry. Thus, the two cherry actions ∧l and ∧l−k produce the same set of
0
cherries in Rl and Rl−k
.
0
0
– If l = j, then Rl equals Rl : This is the case because Rj−1 =∧ Rj−k−1
and through each cherry action
V∗
0
in
each subtree T (wi ) is cut (and removed from Rj−k−1 , if wi is a leaf). Consequently, through the
cherry actions ∧q and ∧0q , with 1 ≤ q < j, the same tree operations are applied to R1 and R10 and, thus,
Rj equals Rj0 .
– If l > j, then Rl0 equals Rl : This is the case because Rj equals Rj0 and through ∧l and ∧0l the same tree
operation is performed on Rl and Rl0 , respectively.
V
Now, by consecutively rearranging all special cherry actions of Type B as described above, we can turn
V(2)
into the cherry list
for T1 and T2 in which all special cherry actions referring to the same cherry action
are located next to each other. Moreover, as a direct consequence of these observations and since through
V(2)
V(1) V(2)
still the same edges are cut from T2 as through
,
is still a cherry list for T1 and T2 calculating F.
16
V(2)
Step 3. Let
be the list that is obtained from applying Step 1 and Step 2 as described above. Then,
V(2)
V(3)
we can further modify the cherry list
to
by replacing each sequence
(∧i = ({L(a), L(c)}, ei ), . . . , ∧i+k−1 = ({L(a), L(c)}, ei+k−1 ))
of special cherry actions (i.e., cherry actions of Type B ) through a single extended cherry action ∧B =
({L(a), L(c)}, EB ), where EB denotes the pendant edge set for {L(a), L(c)}. As a consequence, since ∧B
V(3)
just summarizes all tree operations conducted by the replaced sequence of special cherry actions, by
the
V(2)
same edges are cut from T2 as it is the case for the cherry list
. As a direct consequence, the maximum
V(3)
acyclic agreement forest F is still calculated by the extended cherry list
.
V
In summary, as described by those three steps, we can consecutively replace all special cherry actions of
V(3)
so that the resulting list
satisfies all of the following three conditions.
V(3)
–
is an extended cherry list for T1 and T2 .
V(3)
–
does not contain any special cherry actions.
V(3)
calculates the maximum acyclic agreement forest F.
–
This directly implies that for each computational path in allMAAFs calculating a maximum acyclic
agreement forest F for both input trees there also exists a computational path in allMAAFs1 calculating
F.
Now, from Lemma 2 we can directly deduce the correctness of Theorem 2.
5.3
Correctness of allMAAFs2
The correctness of our third modified algorithm allMAAFs2 principally directly follows from the correctness
of the refinement step, which can be found in the work of Whidden et al. Whidden et. al. [14]. We still have to
show, however, that omitting both additional cutting steps when processing a common cherry still computes
all of those agreement forests from which all maximum acyclic agreement forest can be obtained by cutting
some of its edges.
Theorem 3. Let T1 and T2 be two rooted binary phylogenetic X -trees and k ∈ N. Calling
allMAAFs2 (T1 , T2 , T1 , {T2 }, k, ∅)
returns all maximum acyclic agreement forests for T1 and T2 if and only if k ≥ h(T1 , T2 ).
Proof. To proof Theorem 3 we first have to establish a further lemma. Here, we argument that by contracting common cherries instead of cutting one of its edges, results in an agreement forests containing less
components.
V
Lemma 3. Let
= (∧1 ,∧2 , . . . ,∧nV) be a cherry list for two rooted binary phylogenetic X -trees T1 and
T2 with ∧i = ({L(a), L(c)}, ec ) ∈
so that {L(a), L(c)} is a common cherry of Ri and Fi and ec the
in-edge of leaf c. Moreover,
let
F
be
the agreement forest for T1 and T2 that is calculated by calling
V
processCherries(R1 , F1 , ) such that |F| = k. Then, there exists an agreement forest F̂ with |F̂| = k − 1
that can be computed by calling processCherries(Ri , Fi , (∧0i , . . . ,∧0n )) in which ∧0i = ({L(a), L(c)}, ∅).
Proof. Let Fa and Fc be the two components in F that have been derived from expanding both components
containing a and c, respectively, and let Xa be the taxa set of the subtree rooted at a (i.e., Fa (a)). Then,
by attaching Fc back to Fa we can reduce the size of the agreement forest F by one. Here, depending on
whether LCAFa (Xa ) is the root of Fa or not, this re-attachment step can be done in two different ways.
17
– If LCAFa (Xa ) is not the root of Fa and, thus, has an in-going edge e = (u, w), first e is split into two
adjacent edges (u, v) and (v, w) and then Fc is re-attached to v by inserting a new edge (v, rc ), where rc
denotes the root of Fc .
– Otherwise, if LCAFa (Xa ) is the root of Fa , first a new node v is created and then v is connected to the
two roots of Fa and Fc .
In the following, we will denote the component that is obtained from attaching Fc back to Fa by Fr and
the resulting set of components by F̂. In Figure 6, we give an illustration of the two forests F and F̂.
Fig. 6. The two forests F and F̂ as defined in Lemma 3. Note that, since Fc is attached to Fa , the size of F̂ is |F | − 1.
Notice that weVcan calculate the agreement forest F̂ by calling the
V algorithm processCherries with a
0
specific cherry list
that can be derived from the original cherry list = (∧1 ,∧2 , . . . ,∧i = ({L(a), L(c)}, ec ), . . . ,∧n )
as follows.
(1) Replace ∧i by ({L(a), L(c)}, ∅).
(2) In each subsequent cherry action ∧i+1 , . . . ,∧n , replace L(a) through L(a) ∪ L(c).
0
0
Let Ri+1 and Fi+1 be computed by applying ({L(a), L(c)}, ec ) to Ri and Fi and let Ri+1
and Fi+1
be
computed by applying ({L(a), L(c)}, ∅) to Ri and Fi . Since {L(a), L(c)} is a common cherry of Ri and Fi ,
the only difference between the two topologies, disregarding node labels, is that Fi+1 contains an additional
component consisting of an isolated node labeled by L(c). Notice, however, that, since this component is
fully contracted, this component
V0 cannot have an impact on the subsequent cherry actions ∧i+1 , . . . ,∧n .
Consequently, the cherry list
calculates F̂.
Up to now, we have shown that there exists a set of components F̂ of size |F|−1,
V0 which can be computed by
calling the algorithm processCherries with a slightly modified cherry list . Now, to establish Lemma 3,
we still have to show that F̂ is an agreement forest for both input trees T1 and T2 . First, note that each
expanded forest to which allMAAFs applies
V0 the acyclic check (cf. Alg. 1, line 5) has to satisfy each condition
is imitating a computational path corresponding to the original
of an agreement forest [12, Lemma 3]. As
algorithm allMAAFs calculating F̂ , this directly implies that F̂ is an agreement forest for T1 and T2 .
Now, given two rooted binary phylogenetic X -trees T1 and T2 , assume that our first modified algorithm
allMAAFs1 contains a computational path calculating a maximum acyclic
agreement forest F of size
V
V k
by cutting instead of contracting a common cherry. More precisely, let
with ∧i = ({L(a), L(c)} ∈
be
the cherry list mimicking this computational path and, without loss of generality, let ∧i be the only cherry
18
action cutting instead of contracting aVcommon cherry. Then, as already discussed in the proof of Lemma 3,
0
there additionally exists a cherry list
calculating a specific agreement forest F̂ of size k − 1 containing a
component F̂ . More specifically, let Fa and Fc be the expanded component in F which is derived from the
contracted node a and c, respectively. Then, F̂ is obtained from F by re-attaching Fc back to Fa (cf. Fig. 6).
Thus, as a direct consequence, by cutting the in-edge corresponding to the root of F̂ (L(Fc )) the maximum
acyclic agreement forest F arises.
This implies that, if cutting instead of contracting a common cherry yields a maximum acyclic agreement
forest F, the algorithm allMAAFs2 guarantees the computation of an agreement forest F̂ from which F
can be obtained by cutting some of its edges. Note that the refinement step is always able to identify the
minimal number of such edges and, thus, as the only difference between the first modified algorithm and the
algorithm allMAAFs2 consists in the way a common cherry is processed, Theorem 3 is established.
6
Simulation Study
Our simulation study has been conducted on the same synthetic dataset as the one used for the simulation
study reported in the work of Albrecht et al. Albrecht et. al. [1]. It consists of binary phylogenetic tree pairs
that have been generated in respect to one combination of the following three parameters: the number of
taxa n = {20, 50, 100.200}, the executed number of rSPR-moves k = {5, 10, . . . , 50}, and the tangling degree
d = {3, 5, 10, 15, 20}. For each combination of those parameters 10 tree pairs have been generated, resulting
in 2000 tree pairs in total.
More precisely, a tree pair (T1 , T2 ) is computed as follows. In a first step, the first tree T1 with n leaves
is computed which is done initially by randomly selecting two nodes u and v of a specific set V consisting
of n nodes of both in- and out-degree 0. Then, those two selected nodes u and v are connected to a new
node w and, finally, V is updated by replacing u and v by its parent w. This process is repeated until V
consists only of one node corresponding to the root of T1 . In a subsequent step, T2 is computed by applying
k rSPR-moves to T1 each respecting tangling degree d.
When performing a sequence of rSPR-moves, one can undo or redo some of those moves and, thus, k
is only an upper bound of the real underlying rSPR-distance corresponding to both trees of a tree pair.
The tangling degree, as already described in the work of Albrecht et al. Albrecht et. al. [1], is an ad hoc
concept controlling the number of minimum common clusters during the construction of a tree pair. Since,
however, all the four implementations that are tested on the synthetic dataset perform a cluster reduction
and, thus, equally benefit from the number of minimum common clusters, this number is irrelevant for our
simulation study and, consequently, we do not give any further details about this parameter. Instead, we
refer the interested reader to the work of Albrecht et al. Albrecht et. al. [1].
In this section, the practical runtimes produced by the respective implementations of our two modifications are compared with the practical runtimes of an implementation of the original algorithm. More
precisely, by applying our synthetic dataset the practical runtime was measured for an implementation of
the original algorithm allMAAFs, the algorithm allMAAFs1 , and the algorithm allMAAFs2 . Each of
those algorithms has been integrated as a plug-in into the freely available Java software Hybroscale1 . The
simulation study has been conducted on a grid computer providing 16 cores and 40 GB RAM. In order to
receive a reasonable set of completed data sets within an appropriate time period, we decided to compute
only the hybridization number and omitted the computation of all maximum acyclic agreement forests.
Moreover, we set the maximum runtime of each tree pair to 20 minutes, which means that each tree pair
whose computation of the hybridization number could not be finished within 20 minutes was aborted. Notice
that depending on the runtime analysis, which is shown in the following, those aborted tree pairs were either
not taken into account or counted as being finished after 20 minutes.
The problem arising when computing all, instead of just one, maximum acyclic agreement forests for
certain tree sets is that, typically, there exists a large number of those agreement forests all being distributed
in a vast search space. Consequently, even if those tree sets are of low computational complexity, one has to
1
www.bio.ifi.lmu.de/softwareservices/hybroscale
19
investigate far more than 20 minutes, which means that one could not conduct the simulation study within
an appropriate time period. Otherwise, if we would choose a relatively small maximum running time, such
as 20 minutes, we could only process those tree sets of low computational complexity, which are not able to
indicate the speedup obtained from applying our three modifications.
In Figure 7, the mean average runtimes in terms of the computed hybridization numbers are shown. More
precisely, this plot was generated by first aggregating all tree pairs corresponding to the same hybridization
number and then by computing the mean average runtime of each of those aggregated subsets. Therefore,
each aborted tree pair whose computation would have taken longer than the given time limit of 20 minutes,
was treated as follows. If the tree pair could not be computed by both considered algorithms, this tree pair
was not taken into account. However, if at least one algorithm was able to compute the hybridization number
for a tree pair, the runtime of the other algorithm was set to 20 minutes if this algorithm was aborted in this
case. Regarding the Figure 7, the number assigned to each measurement denotes the number of tree pairs
whose hybridization number could be computed within the time limit by the corresponding two algorithm.
Figure 7 indicates that each of our two modified algorithms outperforms the original algorithm. Moreover,
this figure indicates that our third modified algorithm allMAAFs2 is significantly more efficient than our
first modified algorithm allMAAFs1 .
Figure 8 shows the distributions of the speedups obtained from our two modified algorithms via boxplots.
Each of those boxplots was generated by first selecting a relevant set D of tree pairs from our synthetic dataset
and then by computing the speedup of each of those tree pairs by taking the runtime obtained from the
corresponding two algorithms into account. More precisely, in a first step, we set the runtime of each tree
pair whose hybridization number could not be computed within the time limit to 20 minutes. Second, we
consider each tree pair d as relevant if at least one of both algorithms could process d within the time
limit of 20 minutes and if at least one computation referring to one of both algorithms took longer than 50
seconds. Consequently, the relevant set D excludes those tree pairs whose computational complexity is, on
the one hand, too low and, on the other hand, too high to reveal a difference between the runtimes of both
algorithms.
Figure 8 again indicates that both modified algorithms are more efficient than the original algorithm.
More specifically, for the considered set of tree pairs, allMAAFs1 and allMAAFs2 is on mean average
about 3.9 times and 11.9 times (median 2.3, 4.9 and 6.6), respectively, faster than allMAAFs. Moreover,
our second modified algorithm can significantly improve the practical runtime of our first modified algorithm.
More specifically, for the considered set of tree pairs, allMAAFs2 is on mean average 8.15 times (median
3.18), respectively, faster than allMAAFs1 .
In order to give a reason of the speedup obtained by our two modified algorithms, we measured the
number of recursive calls that have been performed for calculating the hybridization number of each tree
pair. To draw comparisons, we only took those tree pairs into account whose hybridization number could be
processed by both algorithms within the time limit of 20 minutes, which were in total 1302. Next, all these
tree pairs were grouped according to their hybridization number and, finally, the mean average number of
recursive calls for each group was computed. Regarding Figure 9, the numbers that are attached to each
measurement correspond to the number of tree pairs that have contributed to the mean average denoted at
the x-axis.
Figure 9 indicates that by applying our two modified algorithms there are significantly less recursive calls
necessary for the computation of a maximum acyclic agreement forest compared with the original algorithm.
Moreover, Figure 9 (c) shows that our second modified algorithm allMAAFs2 has to conduct significantly
less recursive calls than our first modified algorithm for computing hybridization numbers which, obviously,
compensates the effort of preventing additional recursive calls when processing a common cherry. Since the
computation of the hybridization number is just an intermediate step in computing all maximum acyclic
agreement forests, the difference between the number of recursive calls between the first and both the second
and the third modified algorithm is expected to be even larger in this case.
Finally, we finish this section by comparing the runtimes of the four algorithms via scatter-plots (cf. Fig. 10).
Again, all tree pairs whose hybridization number could not be computed within the time limit was set to
20
1000
●
allMAAFs1
allMAAFs
0
0
0
0
●
●
●
●
0
0
●
●
1
1
●
5
●
4
800
●
600
9
24
●
400
26 32
●
●
32 40
●
1
0
●
1
●
17
●
●
218
38
9 52 144 16 33 64 150 ● 84
●
43 51 17 24
●
●
●
●
●
●
●
●
219 85 38 32
9 52 144 16 33 64 150
0
21 24
●
●
●
22
●
4
17
●
20
5
●
●
21
2
11
●
2
22
48
200
Mean Runtime (s)
1200
(a)
12 28
25
33 29 30
31
27
33 42
10
1
1
7
20
30
40
Computed Hybridization Number
1000
●
1
allMAAFs3
allMAAFs
5
●
4
0
0
0
0
0
0
0
0
●
●
●
●
●
●
●
●
●
●
9
800
●
24
●
600
26
●
●
●
●
5
20
21
●
●
7
21
●
●
48
400
17
22
24
32
4
●
40
22
●
23
17
●
●
1
0
●
1
44
218
29
38
9 52 144 16 33 64 150 ● 84
55
●
●
●
●
●
●
●
●
●
63 17 25 39
51
9 52 144 16 33 64 150221 85 38 35
0
7
19
●
32
200
Mean Runtime (s)
1200
(c)
10
41 35 43
15 10
30
28
3
4
6
1
20
30
40
Computed Hybridization Number
(a)
1200
1000
1
5
0
2
2
1
4
800
11
12
7
600
5
30
400
42 33
43
200
Mean Runtime (s)
0
1
allMAAFs3
allMAAFs1
32
0
1
1
0
7
31
27
33
51
28 25
29
24
44
17
219
29
85 38
9 52 144 16 33 64 150
55
63 17 25 39
9 52 144 16 33 64 150221 85 38 35 51
10
4
23
41 35 43
7
19
15 10
30
28
20
3
4
6
1
30
40
Computed Hybridization Number
Fig. 7. Comparisons between the mean average runtimes in terms of the hybridization number of (a) allMAAFs and
allMAAFs1 ,(b) allMAAFs and allMAAFs2 , as well as (c) allMAAFs1 and allMAAFs2 . For each hybridization
number the corresponding number of tree pairs is given that could be computed within the time limit of 20 minutes
by the two respective algorithms and, thus, contributed to the mean average runtime denoted by the y-axis.
21
200.0
100.0
●
●
●
●
●
●
●
●
50.0
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
20.0
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
2
3
vs
al
lM
AA
Fs
1
Fs
Fs
AA
lM
al
al
lM
AA
Fs
3
vs
al
lM
AA
Fs
AA
lM
al
AA
Fs
lM
al
AA
Fs
lM
al
al
lM
AA
Fs
2
3
vs
vs
al
lM
al
lM
AA
Fs
Fs
AA
Fs
AA
al
lM
vs
2
1
vs
al
lM
AA
Fs
1
0.5
1.0
2.0
5.0
10.0
Speedup
●
Fig. 8. The distributions of the speedups, obtained from comparing algorithm A1 versus algorithm A2 , as denoted
at the x-axis, via boxplots. A tree pair d from our synthetic dataset was only considered if at least one of the two
algorithms could process d within the time limit of 20 minutes and if at least one algorithm took longer than 50
seconds for its computation. Notice that a base-10 log scale is used for the y-axis.
22
4
allMAAFs1
allMAAFs
●
●
6e+06
5
1
●
●
17
9
●
4e+06
●
20
22
2e+06
24 21
●
48
84 38
9 52 144 16 33 64 150218
●
●
1
0e+00
#Recursive Calls
8e+06
(a)
●
●
●
●
●
●
●
●
●
●
●
21
●
17 22 26 32
●
●
●
●
●
●
40
●
●
●
0
32
24
10
20
30
40
Computed Hybridization Number within 1200s
4
allMAAFs3
allMAAFs
●
●
6e+06
5
●
1
●
17
9
●
4e+06
●
22
2e+06
24 21
●
48
1
0e+00
#Recursive Calls
8e+06
(c)
●
84 38
9 52 144 16 33 64 150218
●
●
●
●
●
●
●
0
●
32
●
17 22 26 32
●
●
●
●
24
20
●
●
●
21
●
●
●
40
●
●
●
10
20
30
40
Computed Hybridization Number within 1200s
(a)
8e+06
6e+06
4
2
4e+06
5
2
2e+06
1
1
0e+00
#Recursive Calls
1e+07
1
allMAAFs1
allMAAFs3
0
24
32 43 51 17
9 52 144 16 33 64 150219 85 38
28 25 11
31
33 29 30
7
12
33 42
27
10
20
1
30
40
Computed Hybridization Number within 1200s
Fig. 9. Comparisons between the mean average number of recursive calls in terms of the hybridization number of
(a) allMAAFs and allMAAFs1 as well as (c) allMAAFs and allMAAFs2 . For each hybridization number the
corresponding number of tree pairs is given that could be computed within the time limit of 20 minutes by the two
respective algorithms and, thus, contributed to the mean average number of recursive calls denoted by the y-axis.
23
1200
1200
●
●
●
●
●
●
●
●
●
●
●
●
●
●
1000
1000
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
● ● ●
●
●
●
●
●
●●
●●
●●
●
●
●●
●
● ●
●
●
●
●
●
●
●●
● ●●
●
●●●
●●
●
● ●● ●
● ●●●●
●
●
●
●
●
●
●● ●
●
●●
●
●
●●●●
●
●
●
●
●●
●●
●
●
●●● ● ●● ●●
●
●
●
● ●●
●
●
●
●
●●● ●●●●●
●
●
●
●●
●●●
●
●
●
●
●● ●
●
●●
●
●
●●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●●
●
●
●
● ●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
200
400
●
●
●
0
●
●
600
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
800
800
●
1000
0
400
0
●
●●
600
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●● ●
●
●
●
●
● ●
●
●
●
●
●
● ●
●
●● ●
●
●●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
● ●
● ●● ●● ●
●
● ● ●
●
●
●
●●
●●
●
●●
●●
● ●●●
●
● ●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●● ● ● ●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
● ●●●●●● ●● ●
●
● ●● ● ● ●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
● ●●●
●●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●●
●
●●
●
200
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
allMAAFs3
600
●
400
800
●
●
●
●
200
allMAAFs1
●
●
●
1200
0
200
400
1000
800
●
●
1200
600
●
●
●
●
●
● ●
● ●●
●
●
●
●
●
●
●
●
●
200
400
●
●
● ●
●
●
●
● ●
●●
●
●
●
●
●
●●
● ●
●● ● ●
●
●
●
●
●
●
●
●
● ● ●● ● ●
● ●
●
●●●●
●
●
●
● ●
●●
●●●
●
●
● ● ●
●●●
●●● ●
●
●
●
●●
● ●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
● ● ●●●●
●●
●
●
●
●●
●
●
●●
●
●
●
● ●● ●● ●●
●
●
●
●
● ●● ●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●●● ●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
● ●●●●
●
●
●● ●●●
●
●
●
●
●
●
●●
●●
●
●
●● ●●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●
●
●●
●
●
●
●
●
●●
●
●
●
0
●
●
600
●
●
●
●
●
800
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
400
allMAAFs3
1000
●
●
●
●
●
●
200
800
●
●
●
●
0
600
allMAAFs
1200
allMAAFs
●
1000
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
● ●
1200
allMAAFs1
Fig. 10. Scatter-plots comparing the runtimes produced by the three different algorithms when applying our synthetic
dataset. Each dot refers to a tree set of this data set whose corresponding x- and y-value indicates the runtime attained
by the respective two algorithms.
24
20 minutes. The plots clearly show that our modified algorithms outperform the original algorithm whereat
the modified algorithm allMAAFs2 is more efficient than the modified algorithm allMAAFs1 .
7
Running time of allMAAFs1 and allMAAFs2
The theoretical worst-case runtime of allMAAFs1 is still the same as the one of the original algorithm
allMAAFs, which is O(3|X | |X |) as stated in the work of Scornavacca et al. [12, Theorem 3].
Theorem 4. Let T1 and T2 be two rooted phylogenetic X -trees and let k be an integer. The algorithm
allMAAFs1 (T1 , T2 , T1 , {T2 }, k) has a worst-case running time of O(3|X | |X |).
Proof. Let F = {Fρ , F1 , F2 , . . . , Fk−1 } be an agreement forest for T1 and T2 of size k. To obtain F from
T2 , one obviously has to cut k − 1 edges. Moreover, in order to reduce the size of leaf set X of R to 1, to
each component Fi in F we have to contract exactly |L(Fi )| − 1 cherries. Consequently, at most |X | cherry
contractions have to be performed in total. Thus, our algorithm has to perform at most k + |X | = O(|X |)
recursive calls for the computation of F. Now, as one of these recursive calls can at least branch into three
further recursive calls, O(3|X | ) is an upper bound for the total number of recursive calls that are performed
throughout the whole algorithm. Moreover, as each operation that is performed during a recursive call can
be performed in O(|X |) time, the worse-case running time of both algorithms is O(3|X | |X |).
When considering the theoretical worst-case runtime of allMAAFs2 , we have to take the running time
of a refinement step into account.
Theorem 5. Let T1 and T2 be two rooted phylogenetic X -trees and let k be an integer. The algorithm
allMAAFs2 (T1 , T2 , T1 , {T2 }, k) has a worst-case running time of O(3|X | 4k |X |).
Proof. As stated in Theorem 4, the algorithm has to conduct O(3|X | ) recursive calls. Potentially, for each
of these recursive calls we have to apply a refinement step whose theoretical running time is stated with
O(4k |X |) [14]. Moreover, as all other operations that are performed during a recursive call can be performed
in constant time, the worse-case running time of the algorithm is O(3|X | 4k |X |).
Even though both presented modifications do not improve the theoretical worst-case running time, our
simulation study indicates that in practice these modifications are significantly faster than the original
algorithm allMAAFs. Regarding our first modification allMAAFs1 , this is simply due to the fact that
the number of computational paths arising from processing a contradicting cherry is reduced. More precisely,
if there is a computational path calculating a maximum acyclic agreement forest in which the set of pendant
edges EB for a particular cherry is cut, the original algorithms produces |EB | − 1 redundant recursive calls.
Due to the refinement steps that are performed at the end of each recursive call, our third modification
allMAAFs2 always has to start only one recursive call for processing a common cherry. This refinement
step, however, has to be conducted for each maximum agreement forest F in order to transform F into all
maximum acyclic agreement forests. Nevertheless, as indicated by our simulation study, this post processing
step is still more efficient than always running two additional recursive calls when processing a common
cherry (as it is the case for the original algorithm allMAAFs).
In conclusion, due to the significant speedup that can be obtained from applying both presented modifications, we feel certain that this work describes a noticeable improvement to the original algorithm allMAAFs,
which will make this algorithm accessible for a wider range of biological real-world applications.
25
8
Conclusion
An important approach to study reticulate evolution is the reconciliation of incongruent gene trees into
certain kinds of phylogenetic networks, so-called hybridization networks. From a biological point of view, it is
important to calculate not only one but all of those networks as, once calculated all of these networks, one can
then apply certain filtering techniques testing certain hypothesis. As already discussed in Scornavacca et. al.
[12], the recently published algorithm allMAAFs calculating all maximum acyclic agreement forests for
two rooted binary phylogenetic X -trees, is considered to be an important step to achieve this goal.
In this paper, we have presented the two modifications allMAAFs1 and allMAAFs2 of the algorithm
allMAAFs still calculating all maximum acyclic agreement forests for both input trees as shown by formal
proofs. Moreover, as we have additionally indicated by a simulation study that both modifications significantly improve the practical running time of the original algorithm, we feel certain that this work will
facilitate the computation of hybridization networks for larger input trees and, thus, makes this approach
accessible to a wider range of biological problems.
Acknowledgements
We gratefully acknowledge Simone Linz for useful discussions.
26
Bibliography
[1] Albrecht, B.; Scornavacca, C.; Cenci, A.; Huson, D. H.: Fast computation of minimum hybridization
networks. Bioinformatics, 28(2): 191–197, 2011.
[2] Albrecht, B.: Computing Hybridization Networks for Multiple Rooted Binary Phylogenetic Trees by
Maximum Acyclic Agreement Forests. arXiv:1408.3044, 2014.
[3] Baroni, M.; Gruenewald, S.; Moulton, V.; Semple, C.: Bounding the number of hybridisation events for
a consisten evolutionary history. Mathematical Biology 51: 171–182, 2005.
[4] Baroni, M.; Semple, C.; Steel, M.: Hybrids in real time. Systematic Biology, 55: 46–56, 2006.
[5] Bordewich, M.; Semple, C.: Computing the minimum number of hybridization events for a consistent
evolutionary history. Discrete Applied Mathematics, 155: 914–928, 2007.
[6] Bordewich, M.; Semple, C.: Computing the Hybridization Number of Two Phylogenetic Trees Is FixedParameter Tractable. IEEE/ACM Trans. Comput. Biol. Bioinformatics, 4(3): 458–466, 2007.
[7] Chen, Z. Z.; Wang, L.: HybridNET: a tool for constructing hybridization networks. Bioinformatics, 26:
2912–2913, 2010.
[8] Collins, J.; Linz, S.; Semple, C: Quantifying hybridization in realistic time. Journal of Computational
Biology, 18(10): 1305–1318, 2011.
[9] Huson, D.; Rupp, R.; Scornavacca, C.: Phylogenetic Networks: Concepts, Algorithms and Applications.
Cambridge University Press, 2011.
[10] Huson, D.; Richter, D.; Rausch C.; Dezulian T.; Franz M.; Rupp R.: Dendroscope: An interactive viewer
for large phylogenetic trees. BMC Bioinformatics, 22;8(1):460, 2007.
[11] Mallet,J.: Hybrid speciation. Nature, 446, 279-283, 2007.
[12] Scornavacca, C.; Simone, L.; Albrecht, B.: A first step towards computing all hybridization networks
for two rooted binary phylogenetic trees. Journal of Computational Biology, 2012.
[13] Whidden, C; Zeh, N.: A unifying view on approximation and fpt of agreement forests. Algorithms in
Bioinformatics, LNCS Vol. 5724, 390–402, 2009.
[14] Whidden, C.; Beiko, R.; Zeh, N.: Fixed-Parameter and Approximation Algorithms for Maximum Agreement Forests. arXiv:1108.2664, 2011.
| 8cs.DS
|
An efficient algorithm for the calculation of reserves for non-unit
linked life policies
Mark Tucker* and J. Mark Bull
Edinburgh Parallel Computing Centre
arXiv:1401.1757v2 [q-fin.CP] 27 Jun 2014
Abstract
The underlying stochastic nature of the requirements for the Solvency II regulations has introduced significant challenges if the required calculations are to be performed correctly, without
resorting to excessive approximations, within practical timescales. It is generally acknowledged
by practising actuaries within UK life offices that it is currently impossible to correctly fulfil the requirements imposed by Solvency II using existing computational techniques based on
commercially available valuation packages.
Our work has already shown that it is possible to perform profitability calculations at a far
higher rate than is achievable using commercial packages. One of the key factors in achieving
these gains is to calculate reserves using recurrence relations that scale linearly with the number
of time steps.
Here, we present a general vector recurrence relation which can be used for a wide range
of non-unit linked policies that are covered by Solvency II; such contracts include annuities,
term assurances, and endowments. Our results suggest that by using an optimised parallel
implementation of this algorithm, on an affordable hardware platform, it is possible to perform
the ‘brute force’ approach to demonstrating solvency in a realistic timescale (of the order of a
few hours).
Keywords
Solvency II; Non-unit linked life reserves; Brute force; Parallel programming; OpenMP
Correspondence to: Mark Tucker, EPCC, The University of Edinburgh, King’s Buildings, Mayfield
Road, Edinburgh, EH9 3JZ, UK. email: [email protected]
1
Introduction
In the UK, life companies are required to demonstrate solvency on a regular basis; the calculations
for this are usually done using software created using commercial packages, such as Prophet, MoSes,
or Algo Financial Modeller. The requirements for Solvency II, if implemented naively, lead to the
need for a far greater volume of calculations. This volume of calculations is so vast that the use of
commercially available software to obtain the results is currently beyond contemplation.
Many of the currently available packages produce single-threaded executables although, in the
last few years, some packages have begun to venture into the world of multi-threaded programs.
Further, some of the commercial packages use naive summations to calculate reserves, while others
use an approach based on commutation functions. Typical implementations using either summations or commutation functions lead to computational complexity which is quadratic in the number
of time steps; Macdonald (2004) provides an overview of commutation functions from which it is
possible to deduce that commutation functions lead to quadratic computational complexity. The
combination of single-threadedness and inefficient methodologies means that the performance of
the programs produced by these packages is far below that required for demonstrations of solvency
under the new regulations.
1
Over the last two decades, conventional wisdom within life offices has been to improve computing
performance by obtaining more up-to-date PCs; it has been sufficient to rely on chip manufacturers
increasing processing power of computers roughly in line with Moore’s law. However, physical
considerations dictate that it is now no longer possible to rely on the speed of CPUs increasing
in order to improve the processing power of computers. Instead, over the past few years, chip
manufacturers have moved to placing more compute cores on each chip in order to increase their
processing power. This means that, to benefit from modern processors, it is necessary to embrace
programming paradigms which harness the power of the multi-core chips. Whilst some commercial
valuation packages do now have some parallel computation capability, these capabilities are often
limited.
A reasonably common alternative approach to achieving higher throughput has been to split the
data and calculations across large numbers of PCs to perform the calculations, and then combine
the results when the calculations are finished. The run times involved mean that using this approach
to produce solvency results from, say, 1000 simulations at each future time step is not practical.
Using recurrence relations instead of summations or commutation functions leads to computational complexity which is linear in the number of time steps; for a typical projection, using
monthly steps over 25 to 50 years, the saving is significant – around two orders of magnitude.
Our previous results (see Tucker and Bull, 2012) show that, using a 48-core Symmetric MultiProcessor (SMP), it is possible to perform profitability calculations for single-life annuities, using
monthly steps, at a rate of 3.9 × 105 policies per second; the commercial package we used as a
comparison processes these policies at 1 per second. The five orders of magnitude improvement
in performance comes from a combination of factors; two orders from a change of hardware and
parallelisation of the code using OpenMP (see OMP 3.1) to coordinate 48 threads, two orders from
changing to use the recurrence relation, and the last order from further manual optimisation of the
code.
Here we present a formulation of the recurrence approach in vector form, demonstrate how this
relates to a variety of non-unit linked policies, and state the assumptions which are required in
order that the recurrence relation holds. We provide motivation using annuities, since these are
the contracts which prompted this investigation; the vast numbers of policies and the associated
reserves make these policies of prime importance to practising actuaries. We develop the theory
for a very general case which can be used for a wide range of life policies, and then provide specific
in-depth examples for various policy types. Finally, we present some performance results which
demonstrate that an efficient implementation of the recurrence relation can lead to run times which
make the Brute Force approach to calculating solvency computationally tractable.
2
Motivation
Any company competing in the life assurance and pensions markets in the UK must demonstrate
solvency as part of the regulations. Throughout Europe, the Solvency II project (see Solvency II,
2008) is introducing tougher requirements on both the capital requirement (i.e. the demonstration
of solvency) and risk management aspects of the insurance industry; our work concentrates on the
calculation of liabilities, which is a major part of the calculation of the capital requirement.
The existing regulations require that a single, ‘best estimate’, calculation of solvency is performed using a set of assumptions chosen by actuaries. However, using industry-standard software,
the processing for a typical portfolio of annuity policies takes one-to-two days to complete on a
single modern PC. Under Solvency II, it is necessary for an insurer to hold sufficient capital such
that there is only a 1-in-200 chance of being unable to cover the liabilities.
Therefore, one widely-adopted approach to satisfying the new regulations is to replace the single
calculation of the liabilities by the generation of 1000 Monte Carlo scenarios at each future time
2
step; projections typically use monthly time steps and extend for about 60 years. From these
1000 scenarios generated, we seek the one which produces the 5th largest liability value since that
corresponds to a 0.5% chance of the actual liabilities being larger than the value reported. This is
computationally intractable using current commercial software and so it is necessary to seek efficient
algorithms, and implement those algorithms on hardware capable of highly parallel threading.
A significant amount of work has already been done on changing the way in which the stochastic
processes underlying the assets or liabilities are modelled, with those changed processes being
implemented on small scale parallel computers; see, for example, Corsaro et al. (2010) where the
focus is on the development of a parallel algorithm for the valuation of profit sharing policies. In
our work, we approach the problem differently; the generation of 1000 Monte Carlo scenarios at
each future time step has become colloquially known in the UK as the “Brute Force” approach.
We aim to show that it is possible to perform the full brute force calculation (with a full valuation
for each policy, for each scenario, for each step) within practical timescales using massively parallel
architectures. In order to achieve this, a fundamentally different approach to the calculations is
required; we present an algorithm which is linear rather than quadratic in time steps and hence
reduces the overall time required to perform the brute force calculations by a significant factor.
2.1
The Brute Force Approach
Within industry in the UK, current standard practice is that the calculations to value the liabilities
are performed using programs created on specialist software packages which are crafted for flexibility
and ease of use, rather than performance of the resulting programs. These packages require the
relationships between variables to be entered in a pseudo-code style language, and the package
then produces source code, compiles and links it, and launches the resulting executable. The clear
advantage of these packages is that quite complex programs can be developed by users who are not
trained programmers, allowing users who understand the intricacies of the policies to produce the
programs; this is clearly beneficial when complicated contracts are being modelled.
The fact that the users are usually not trained programmers is also one of the main disadvantages
of these packages; the art of programming is removed by the package and this usually results in the
executable having sub-optimal performance. It is currently not uncommon for life assurance offices
to have hundreds of PCs as dedicated slaves to perform these calculations; the problem with this
approach lies in the fact that once these machines have completed their calculations, the results
need to be collated and that renders this approach impractical.
It is generally acknowledged by practising actuaries that it is impossible to comply with the
incoming regulations using the current approach. The policies which have been chosen to base this
project on are life annuities; the payments made to a pensioner after retirement form a typical life
annuity. The profitability calculations required for these policies, using the packages outlined above
and hardware which the industry is comfortable with, currently run at a rate of approximately one
second per policy; it takes roughly 35 hours to produce one set of best estimate results for one
block of 160,000 policies.
Performing 1000 simulations at each future time step using this technology is infeasible; supposing that, on average, each policy has 60 years (i.e. 720 months) outstanding at the valuation
date, the estimated run time for this approach is
720 719 718
1
1000 ×
+
+
+ ··· +
× 35 hours ≈ 1.26 × 107 hours
720 720 720
720
≈ 1440 years.
Clearly, even with thousands of PCs, this is beyond contemplation.
Our previous work (see Tucker and Bull, 2012) shows that, by using straightforward optimisation techniques, it is possible to obtain significantly better performance than can be obtained
3
from the commercial packages. Firstly, using a good set of compiler options (including disabling
debugging information) gave a speedup of 3.8×. Secondly, manually optimising dominant routines
led to a further speedup of about 12.4×; these changes included unrolling nested loops, simplifying arithmetical steps, removing initialisation of arrays which hold results, and replacing power
calculations with successive multiplications. Finally, changing to more modern hardware led to a
speedup of about 1.87× so that, combining all of these optimisations, the overall improvement was
a speedup of about 88×. Even using this optimised code, the estimated run time for the brute
force approach is roughly 16 years, and this is still beyond contemplation. Using the recurrence
approach is critical in order to reduce this to practical timescales.
2.2
Actuarial Notation
As far as reasonably practical, we use standard International Actuarial Notation (see Green Tables),
so that
t px
= Pr[life aged precisely x survives until age x + t]
t qx
= Pr[life aged precisely x dies before reaching age x + t]
and
As usual, if t = 1, then the prefix is dropped so that px = 1 px and qx = 1 qx . Also following standard
notation, for t ≥ 0, lx+t is the number of lives expected to be alive at age x + t given that there
are lx lives at age x.
Sections 2.3 and 2.4 consider annuities; standard notation uses
- ax to represent the expected present value of an annuity where payments of amount 1 are
made at the end of a year to a life aged x at the time of valuation, so long as the life is alive
at the time of payment, and
- äx to represent the expected present value of an annuity where payments of amount 1 are
made at the start of a year to a life aged x at the time of valuation, so long as the life is alive
at the time of payment.
We use a0x to highlight that a general payment stream is assumed; i.e. the payments of amount
1 are made at some fraction f ∈ [0, 1] through the year, and hence ax and äx are simply special
cases of a0x with f = 1 and f = 0 respectively.
In a similar manner, standard notation uses ax|y and äx|y to represent two-life reversionary
annuities where the payments are made at the end or start of a year, respectively, to (y) after the
death of (x) (where (z) denotes a life aged exactly z). Here, we use a0x|y to indicate that payments
are made some fraction f ∈ [0, 1] through the year so that ax|y and äx|y are just special cases of
a0x|y .
Finally, we use v to denote the discount factor which applies over a period of unit length, so that
is the discount factor which applies from the point of valuation to a cash flow some fraction
f ∈ [0, 1] through the step from t to t + 1 for t ∈ Z+ .
v t+f
2.3
Annuities
At its most basic, an annuity is just a stream of payments, and a life annuity is an annuity where
the payments depend on the survival, or otherwise, of a pre-specified life, or lives. Section 2.3.1
considers single-life annuities certain; they depend on the survival, or death, of only one life, and
they will start to be paid. Other types of life annuity could depend on two or more lives, and yet
4
others may not even be start to be paid; for example, a child’s annuity which is a rider benefit to
a temporary assurance of the parent will only come into payment if that parent dies within the
period specified in the contract. Section 2.3.2 considers two-life reversionary annuities since this is
the contract which provided our motivation for developing a vector form of a recurrence relation.
2.3.1
Single-Life Annuity Certain
Annuities may be paid in arrears, or in advance, although many annuities are paid part way
through each period; obvious examples are pensions where the payments are often made on the
‘monthiversary’ of the policy inception date. In Section 2.2 it was noted that the advance and
arrear cases are just special cases of annuities where payments are made part way through each
period; it is therefore appropriate to we only consider cases where payments are made at a fraction
f ∈ [0, 1] through the step.
Level, single-life annuities serve as an introduction to several concepts which become useful
later, particularly when considering other types of annuity. For a general, single-life, level annuity
the summation formula for the reserve factor is
a0x
=
∞
X
t+f px
× v t+f
for f ∈ [0, 1]
(2.1)
t=0
The derivation of the recurrence relation is straightforward:
a0x =
∞
X
t+f px
× v t+f
t=0
∞
X
= f px × v f +
((s+1)+f ) px
(s+1)=1
∞
X
= f px v f + px v
(s+f ) px+1
× v ((s+1)+f )
× v (s+f )
s=0
and, by comparing the summation with Equation 2.1, the recurrence relationship is
a0x = f px v f + px v a0x+1
(2.2)
Note that this derivation assumes that both mortality rates and interest rates are constant in time;
in Section 3.7 these assumptions will be removed.
For an annuity payable annually in advance, this relation reduces to
äx = 1 + v px äx+1
Note that, when written in the form of Equation 2.2, the recurrence runs backwards in time;
this is convenient since natural boundary conditions exist (or may be assumed) at the end of the
policy, e.g. p120 = 0.
A useful side-effect of using the recurrence is that it removes the need to use the power function
to compute v t+f in Equation 2.1, and replaces it with multiplication, which is about 20 times
cheaper in modern hardware.
2.3.2
Two-Life Reversionary Annuity
Re-interpreting a standard definition (see Neill, 1989, Section 8.6), a reversionary annuity “becomes
payable on the failure of a specified status and remains payable during the continued existence of a
5
second status”. The simplest form of reversionary annuity is one which becomes payable to (y) on
the death of (x), if (y) is alive at that time, and remains payable until the death of (y); a common
example is a spouse’s pension which becomes payable to a pensioner’s spouse on the death of that
pensioner.
Suppose that the lives are aged x and y, and that they are independent. Assume, without loss
of generality, that the payment is made to (y) after the death of (x). Then
Pr[payment at time t is made] = Pr[x is dead at time t and y is alive at time t]
= Pr[x dies within t years] × Pr[y survives for t years]
= (1 − t px ) × t py
For the case where level payments are made part-way through an interval, at some fraction f
from the start of each interval, the summation formula for the reserve factor is
a0x|y
=
∞
X
(1 − t+f px ) × t+f py × v t+f
for f ∈ [0, 1]
(2.3)
t=0
Again, the recurrence relation may be derived straightforwardly:
a0x|y =
∞
X
(1 − t+f px ) · t+f py · v t+f
t=0
= f qx · f p y · v f
∞
X
+
(s+1)=1
∞
X
−
(s+1)+f py
· v (s+1)+f
(s+1)+f px
· (s+1)+f py · v (s+1)+f
(s+1)=1
= f qx · f p y · v
f
+ v py · a0y+1
∞
X
− v p x py
[1 − s+f qx+1 ] · s+f py+1 · v s+f
s=0
i.e.
a0x|y = f qx f py v f + qx py v a0y+1 + px py v a0x+1|y+1
(2.4)
In turn, the three components represent
1) the payment at time f which is made only if x has died but y is still alive,
2) the (single life) reserve factor at the start of the next step if x has died but y is still alive, and
3) the (reversionary) reserve factor at the start of the next step if both x and y are still alive.
2.4
Vector Recurrence Relation
The derivation of the vector form of the recurrence relation is based on the observation that the
recurrence relation for the reserve factor for the reversionary annuity in Equation 2.4 also involves
the reserve factor for the single life annuity.
6
2.4.1
Contract-Based Presentation
When expressed as a vector, the pair of recurrence relations in Equations 2.4 and 2.2 become
0
f
0
0
ax|y
f qx f py v + py qx v ay+1 + px py v ax+1|y+1
=
a0y
f py
which can be written as
0
ax|y
f qx f p y
= vf
0
a0y
v f + py v a0y+1
0
ax+1|y+1
1
p x p y p y qx
+ v
0
1
0
py
ay+1
f py
0
(2.5)
Hence, for a relatively simple contract, it is possible to find a vector expression for the recurrence
relation, where that vector expression brings together the relations for all contract types which may
be involved in the reserve calculations for that policy type.
2.4.2
Survival-Based Presentation
Although the two matrices in Equation 2.5 are the same shape, they are populated differently;
this results from the derivation being based on the type of the policy at each step. An alternative
presentation is to consider the survival state of each life at each step. Using binary indexing for
each life being either alive (state 0) or dead (state 1) leads to a simple representation of all the
possibilities of survival over the step; the states and labelling for two lives, currently aged x1 and
x2 , are therefore
State
0
1
2
3
Label
S00
S01
S10
S11
(x1 )
Alive
Alive
Dead
Dead
(x2 )
Alive
Dead
Alive
Dead
Using the ordering which results from this binary labelling, it is possible to construct a matrix
of probabilities of the lives surviving for time g, which could be either f (when considering the
probability of payment), or 1 (when considering the probability of requiring a reserve). For two
lives, the relevant Markov transition matrix is
g px g py g px g qy g qx g py g qx g qy
0
0
g px
g qx
0
0
g py
g qy
0
0
0
1
The states of the lives can be considered as a time-inhomogeneous Markov chain, where this is the
transition matrix if g = 1.
7
Using this matrix to combine Equations 2.4 and 2.2, and setting g = f or g = 1, leads to the
following recurrence for a reversionary annuity
0
ax+1|y+1
a0x|y
px py px qy qx py qx qy
0
f p x f p y f p x f qy f qx f p y f qx f qy
0
0
0
p
0
q
p
0
q
0
0
x
x
x
x
f
f
+ v
= vf
0
0
0
0
0
p
q
1
q
0
p
ay
ay+1
y
y
f y
f y
0
0
0
1
0
0
0
0
1
0
0
This vector expression for the reserve factors under consideration appears more complex than the
version in Equation 2.5; however, this expression involves only one matrix, which must be evaluated
at two points in time, rather than requiring two different matrices, one for each of the time points.
Therefore, although binary labelling using ‘zero for alive’ may initially seem counter-intuitive,
its use leads to a desirable property; since the number of dead people cannot decrease, this labelling
will naturally lead to the transition matrix being upper triangular.
2.5
Summary
This Section has provided our motivation; it has demonstrated that, for simple cases, it is possible
to produce a vector recurrence relation for successive reserve factors, using two instances of the
same matrix, where that matrix is expressed in terms of transitions between survival states of the
lives involved, rather than the contract types involved.
3
Derivation of Vector Relation
In this section we present a completely general vector recurrence relation which may be used for
different forms of insurance contract; after the initial presentation, we formally derive the relation
using a component-wise approach. The generic nature of the components in the recurrence relation
means that this vector relation can be used for any policy of arbitrary complexity, so long as
the cash flows and probabilities can be isolated and expressed in the form required by the vector
notation presented here.
3.1
General Form
The relations expressed in Equation 2.5 are a specific case of the general form which may be stated
as
rx,t = vtf Wx,t,f cx,t + vt Wx,t,1 rx+1,t
∀ t ∈ R+ f ∈ [0, 1]
(3.1)
where
- x + t is a vector representing the ages, at time t, of a collection of m lives,
- n is the number of states of survivorship which need to be considered for the m lives in x; all
the states must be considered in the formation of the complete vector recurrence relation,
- rx,t is a vector of n reserve factors, at time t ≥ 0, which may be required depending on the
survival state of x at that time,
n.b. since the relation involves rx,t and rx+1,t , the reserves are calculated using steps of unit
interval,
8
- f ∈ [0, 1] is the fraction through a step at which any cash flows occur,
- g ∈ {f, 1} is the fraction through the step at which events of importance occur; any cash
flows which occur happen at fraction f , and reserve factors relating to the end of a step of
unit length are required when g = 1,
- vtg , where g ∈ {f, 1}, is the discount factor which applies from time t either to any cash flows
which may be made (at time t + f ), or to the reserve factors relating to the end of a step of
unit length (so that t + 1); the underlying interest rate may vary with time,
- Wx,t,g is an n × n Markov transition matrix containing the probabilities of the lives surviving
from time t up to time t + g; the underlying mortality may vary with time,
- cx,t is a column vector of length n containing the nominal amounts of the cash flows which
may be made at time t + f , depending on the state of the lives at that time,
- units of time are arbitrary; in practice units are often months or years.
Since, in Equation 3.1, rx,t is expressed in terms of rx+1,t , numerical values for each element of
each vector may be obtained by backward substitution, starting from an appropriate boundary
condition.
Using a standard definition (see Neill, 1989, Section 4.2),
A (prospective) reserve is the present value of all future cash flows, allowing for discounting and the probability of those cash flows being made.
The vector approach presented here is confirmation of the intuitive interpretation of that definition,
i.e.
The reserve ‘now’ is the present value of ‘any cash flows which may occur during the
first period’ together with the present value of ‘any reserve which is required at the end
of the first period, so long as that reserve is then required’.
3.2
The Lives
In theoretical work, it is standard practice to assume that all lives involved in a policy are independent. For the most general case, there could be any number of lives, so we consider x + t,
which represents a collection of m independent lives aged precisely x1 + t, x2 + t, . . . , xm + t; this
notation reinforces the fact that the ages of the lives are a function of the time t since the start
of the projection, where t is a vector, of length m, whose elements are all equal to t. Using this
notation, the ages of a given set of lives are just a function of the duration from valuation to the
time of interest.
3.3
Cash Flows
We are interested in a sequence of cash flows which may be made at fixed future times, according
to the survival state of x at those future times. The nominal amount of each cash flow is fixed, but
the expected amount of each cash flow depends on the probability that the payment is made and
hence on the survival state, at the point of payment, of the collection of lives.
Let t ∈ R+ and let β ∈ R+ . Let cx,t,j be the nominal cash flow which happens at the fraction
f ∈ [0, 1] from time t, if x is in state j at that time. Then
the cash flow cx,t,j happens at time t + f if the lives are in state j at that time
⇒ the cash flow cx,t,j happens at time (t − β) + β + f if the lives are in state j at that time
⇒ the cash flow cx,t,j may also be denoted cx+β,t−β,j
9
i.e.
∀x
cx,t,j = cx+β,t−β,j
0<β<t
This is the mathematical formulation of the statement that “a payment of known nominal amount,
made at some time in the future, will be of the same nominal amount at that future date, irrespective
of how the time in the future is determined”.
Therefore, in particular, let β = 1 and let t = s + 1. Then
cx,s+1,j = cx+1,s,j
(3.2)
It should be noted that the cash flows are a completely general, arbitrary function of the state of
the lives at the time the payment is made; also see Section 3.7.1.
3.4
Discount Factors
In standard notation, if i is a constant annual rate of interest which applies over a period of t
years then the discount factor which applies for that period is v t = (1 + i)−t . There is an implicit
assumption that the period starts at time 0 and ends at at time t; if that is not the case then the
discount factor remains v t = (1 + i)−t wherever the difference in time is t. Hence, when the rate of
interest is constant, v t = v t1 v t−t1 for any 0 ≤ t1 ≤ t.
When the rate of interest varies with time, standard practice (see McCutcheon and Scott, 1991,
Section 2.4) is to consider the discount factor from 0 to t as
Z t
v(t) = exp −
δ(r) · dr
0
where δ(r) is the force of interest at time r.
Let 0 ≤ t1 ≤ t2 , and let δ(r) be the force of interest at time r ∈ [0, t2 ]. Let dx,t,t1 ,t2 be the
discount factor which applies from time t + t1 , and lasts for time t2 − t1 ; i.e. it applies from the
point at which the lives are aged x + t + t1 to the point the where lives are aged x + t + t2 . Then
Z t2
dx,t,t1 ,t2 = exp −
δ(r) · dr
t1
"Z 0
#!
Z
t
= exp −
t2
δ(r) dr +
Z
!
t0
= exp −
δ(r) dr
t0
t1
δ(r) dr
Z
× exp −
t2
δ(r) dr
t0
t1
i.e.
dx,t,t1 ,t2 = dx,t,t1 ,t0 · dx,t,t0 ,t2
∀x
t1 ≤ t0 ≤ t2
This is the mathematical formulation of the statement that “discounting an amount from t2 to t1
is the same as discounting that amount from t2 to an intermediate time t0 and then discounting
that (reduced) amount from t0 to t1 ”.
Therefore, in particular, let t1 = 0, t0 = 1 and t2 = s + 1 + f so that t2 − t1 = s + 1 + f ,
t0 − t1 = 1 and t2 − t0 = s + f . Then
dx,t,0,s+1+f = dx,t,0,1 · dx,t,1,s+f
(3.3)
Notice that these discount factors are independent of the state of the lives, a phenomenon which
is consistent with reality since mortality and investment returns are generally independent. Notice
also that the use of force of interest removes any assumption that the interest rate remains constant;
also see Section 3.7.2.
10
3.5
Survival Probabilities
In standard notation, t px is the probability that a life aged precisely x survives for t years. Hence,
the probability that two independent lives aged precisely x and y both survive for t years is denoted
t px · t py . By extension, the probability that a collection of m independent lives all survive for t
years is simply the product of the survival probabilities of each of the individual lives.
It is necessary to consider the probability of moving from any state to any other. Let wx,t,t1 ,t2 ,j,i
be the probability of the set lives x being in state i at time t + t2 , given that it is in state j at time
t + t1 for t1 ≤ t2 (so that the time for the possible transition between states is t2 − t1 ). Then, using
the Partition Theorem, and conditioning on the state at time t0 ∈ [0, t2 − t1 ],
X
wx,t,t1 ,t2 ,j,i =
wx,t,t1 ,t0 ,j,k · wx+t0 ,t0 ,t0 ,t2 ,k,i
k
where the sum is over all possible states to which the lives could migrate. This is the mathematical
formulation of the statement that “the probability of the set of lives moves from state j to state i
in time t is the same as the probability of moving from state j to any other state at time t0 and
then moving into state i in the remaining time”.
Therefore, in particular, let t1 = 0, t0 = 1 and t2 = s + 1 + f so that t2 − t1 = s + 1 + f ,
t0 − t1 = 1 and t2 − t0 = s + f . Then
X
wx,t,0,s+1+f,j,i =
wx,t,0,1,j,k · wx+1,t,1,s+f,k,i
(3.4)
k
Notice that there is no assumption of constant mortality; also see Section 3.7.3.
3.6
Recurrence Relation for Reserve Factors
Using the results in the preceding Sections it is possible to derive a general recurrence relation
where benefits may be payable, depending on the survival state of the lives. This is a two stage
process; first, a relation is derived for lives being in a particular state, and then the general relation
is obtained by considering all possible states that the lives may be in.
3.6.1
Reserve for Lives in a Particular State
From Section 3.3, any cash flows which happen in a step happen at some fraction f ∈ [0, 1] through
the step. Let rx,t,j be the reserve which must be held, at time t, for a set of lives x in state j at
that time. Then, because the lives could have migrated to state i by time s, the nominal amount
of the cash flow to be made in step t + s to the lives in x, if they are in state i at that time is
cx,t+s,i . Therefore, the expected amount of a cash flow to be made in step t + s to the lives in x,
given that they are currently in state j, and allowing for migration to any other state i, is
X
[wx,t,0,s+f,j,i · cx,t+s,i ]
i
Allowing for discounting to that time, the present value of such a cash flow is
X
dx,t,0,s+f
[wx,t,0,s+f,j,i · cx,t+s,i ]
i
Therefore, using an infinite sum to allow for possible cash flows in all future time steps, the reserve
required at time t, given that the lives x are in state j at that time, is
!
∞
X
X
rx,t,j =
dx,t,0,s+f
[wx,t,0,s+f,j,i · cx,t+s,i ]
(3.5)
s=0
i
for t = 0, 1, 2, · · · , ∞. For the brute force approach, the 1000 simulations at each time step t will
each be an instance of rx,t,j .
11
3.6.2
Relation for Lives in a Particular State
Starting from Equation 3.5, the recurrence relation for lives in a particular state may be derived in
a straightforward manner;
!
∞
X
X
rx,t,j =
dx,t,0,s+f
[wx,t,0,s+f,j,i · cx,t+s,i ]
(3.6)
s=0
i
which, splitting off the first term in the sum, is
= dx,t,0,f
+
X
i
∞
X
wx,t,0,f,j,i · cx,t,i
!
dx,t,0,s+f
X
s=1
wx,t,0,s+f,j,i · cx,t+s,i
i
and, shifting the index (so that s = s0 + 1), this is
= dx,t,0,f
+
X
i
∞
X
wx,t,0,f,j,i · cx,t,i
!
dx,t,0,s0 +1+f
X
s0 =0
wx,t,0,s0 +1+f,j,i · cx,t+s0 +1,i
i
which, using the results from Equations 3.2, 3.3, and 3.4, is
= dx,t,0,f
X
wx,t,0,f,j,i · cx,t,i
i
+
∞
X
#
!
"
X X
dx,t,0,1 · dx,t,1,s0 +f
wx,t,0,1,j,k · wx+1,t,1,s0 +f,k,i · cx+1,t+s0 ,i
s0 =0
i
k
and, reordering the summations and factoring terms which are sum independent, this is
= dx,t,0,f
X
wx,t,0,f,j,i · cx,t,i
i
"
+ dx,t,0,1
X
wx,t,0,1,j,k
∞
X
#
dx,t,1,s0 +f
s0 =0
k
X
wx+1,t,1,s0 +f,k,i · cx+1,t+s0 ,i
i
Finally, recognising that dx,t,1,s0 +f = dx+1,t,0,s0 +f , and comparing sum in square brackets to the
summation required for rx+1,t,k , gives
rx,t,j = dx,t,0,f
X
wx,t,0,f,j,i · cx,t,i + dx,t,0,1
i
X
wx,t,0,1,j,k · rx+1,t,k
(3.7)
k
where the i and k sums are over all possible states of the lives.
Various policy types could fit a particular instance of this formula just by changing the values
of cx,t,i which may or may not be zero, depending on the state j. Notice that a policy which has a
limited term trivially fits the use of an infinite sum by setting the cash flow amounts after the end
of the policy term to zero.
12
3.6.3
Relation Considering all Possible States
Equation 3.7 relates to the j th possibility of a set of possible states. Combining all of the possibilities
for rx,t,j into a vector, the relationship becomes
rx,t = dx,t,0,f Wx,t,f cx,t + dx,t,0,1 Wx,t,1 rx+1,t
where
- rx,t is a column vector, of length n, where the ith entry is rx,t,i ,
- dx,t,0,g is the discount factor from time t up to time t + g for g ∈ {f, 1},
- Wx,t,g is an n × n Markov transition matrix where the entries relate to the lives surviving
from time t to the time t + g for g ∈ {f, 1}; the (i, j)th entry of Wx,t,g is wx,t,0,g,j,i ,
- cx,t is a column vector, of length n, where the ith entry is cx,t,i .
Replacing the discount factors dx,t,0,g with vtg which is closer to the equivalent standard notation,
recognises that interest rates do not depend on the state of any lives, and explicitly allows the
possibility that the underlying interest rate varies through the projection, the equation becomes
rx,t = vtf Wx,t,f cx,t + vt Wx,t,1 rx+1,t
+
∀ x ∈ Rm
f ∈ [0, 1]
as presented in Equation 3.1.
Death is an absorbing state, so that if the lives are in state k at time 0 and still in state k at
time t > 0 then they must have been in that state at all intermediate times. Also, if there are
two or more changes of state in a step, then we are only interested in the overall change; e.g. if
a reversionary annuity changes to a single life annuity on the death of the first life, and becomes
ceased on the death of the second life within the same time step, then we are only interested in the
fact that it has gone from being a reversionary annuity to being ceased, and the fact that it was
temporarily a single life annuity is of no consequence.
3.6.4
Zero Reserve States
We call a state a ‘zero reserve state’ (ZRS) if it is a state for which all future cash flows are zero
and there is no path to a state which has any non-zero cash flows. There is no need to keep track of
ZRSs (since they do not contribute to the liabilities) and hence calculations which relate to ZRSs
are redundant and may be removed. For example, a reversionary annuity has a zero cash flow if
(x) is still alive, but the state where (x) and (y) are both alive is not a ZRS because there is a path
to a state where there are future cash flows, i.e. the path to the state where (x) dies before (y).
A ZRS is not necessarily a sink state for the Markov chain because it is possible to move out of
one ZRS into another ZRS; for example, a reversionary annuity has no future cash flows when (y)
dies so that the state where (y) dies first is a ZRS from which it is possible to move into another
ZRS (on the death of (x)). Allowing for ZRSs, Equation 3.1 may be written as
rx,t = vtf Wx,t,f cx,t + vt Wx,t,1 rx+1,t
where
- rx,t is rx,t with ZRSs removed,
- cx,t is cx,t with ZRSs removed, and
- Wx,t,g is Wx,t,g with rows and columns which correspond to ZRSs removed.
13
3.7
Assumptions
The derivations in the preceding sections require only the simplest of assumptions which, for practical purposes, are not particularly restrictive.
3.7.1
Cash Flows
The monetary amounts of all cash flows (whether they are premiums, benefits or expenses) must
be known in advance of their use. Also, more as a requirement to be able to use vector arithmetic
than an assumption, there must be a one-to-one correspondence between time steps and cash
flows (so that there can be a maximum of one cash flow of any particular type in each projection
step). Therefore, if a policy has monthly cash flows then monthly projection steps are required;
using monthly projection steps is fine for policies which have annual cash flows since 11 of any 12
consecutive steps will have a cash flow of amount zero.
3.7.2
Interest
The interest rate being used for a particular step must be known in advance of reaching the time
step being simulated. Since the derivation in Section 3.4 is based on the force of interest, there is
no need to assume that the interest rate is constant. There is, however, a requirement that the time
interval under consideration can be split appropriately and, for all practical purposes, this should
be possible. Treating inflation as ‘the other side of the interest coin’ requires that an equivalent
assumption applies to inflation of expenses.
There is a great body of research into modelling future interest rates; see for example Vasicek
(1977), Cox et al. (1985) and Chen (1996). Any interest rate model could be used to derive
the sequence of discount factors {vt } required for the recurrence relation in this paper because the
relation simply requires that interest rates are derivable, and available when required for use in the
calculations. Notice that {vt } is independent of the lives.
3.7.3
Mortality
The mortality table being used for a particular step must be known in advance of reaching the time
step being simulated. There is no need to assume that the underlying mortality cannot change,
provided that it is possible to derive a set of survival probabilities from whatever mortality model
is applicable throughout the period up to the point that the transition matrix is used.
There is a great body of work on the modelling of mortality rates; for example Cairns et
al. (2008) provide a comprehensive discussion on recent models. Much of this work shows that
mortality is currently improving over time, so that
‘the probability of a life currently aged 75 surviving for a year’
is less than
‘the probability of a life aged currently 65 surviving from age 75 to age 76, assuming he
first survives for 10 years to reach age 75’.
While it is noted that these improvements in mortality exist, they are not of fundamental
relevance to the workings of the recurrence relation derived in Section 3.5. Any mortality model
could be used to derive the sequence of survival probabilities {Wx,t,g } required for the recurrence
relation in this paper because the relation simply requires that mortality rates are derivable, and
available when required for use in the calculations. It should also be noted that the granularity
14
of the improvements in mortality might mean that the underlying table only needs to be changed
every twelfth step, effectively using annual improvements in a projection which uses monthly steps.
3.8
The Actual Algorithm
Explicitly outlining the algorithm highlights its computational complexity with respect to the number of steps for which the projection runs. Sample timing results which support these complexities
are given in Section 5.1.2.
3.8.1
Existing Summation Algorithm
In order to calculate the reserve at each future time using the standard summation approach, which
is a direct implementation of the formula in Equation 3.5, two forward loops are required. Let S
be the index of the maximum projection step, which may either be calculated from the data or set
as a parameter. Then the algorithm to calculate the reserves is
for t = 0, 1, . . . , S do
obtain cx,t , vt , Wx,t,f , Wx,t,1
end for
for j in states do
set rx,t,j = 0
end for
set vcash flow = 1
set vend step = 1
for t = 0, 1, . . . , S do
for s = t, . . . , S do
set vcash flow = vend step × vtf
set vend step = vend step × vt
for j in states do
!
Y
increment rx,t,j by vcash flow ×
Wx,r,1 × Wx,t,f × cx,s
r<t
end for
end for
end for
Hence, from the loop nest over t and s, it is clear that obtaining the sequence {rx,t } is O(S 2 ), i.e.
the computational complexity is quadratic in the number of projection steps.
3.8.2
Proposed Recurrence Algorithm
As stated in Section 3.1, calculating the reserve at each future time using the vectorised recurrence
approach allows the use of backward substitution, so that the nest of two forward loops is replaced
by a single backward loop. Again, let S be the index of the maximum projection step. Then the
algorithm to calculate the reserves is
for t = 0, 1, . . . , S do
obtain cx,t , vt , Wx,t,f , Wx,t,1
end for
for j in states do
set rx,S+1,j = 0
end for
calculate vt and vtf
for t = S, . . . , 0 do {descending}
15
for j in states do
set rx,t = vtf Wx,t,f cx,t + vt Wx,t,1 rx+1,t
end for
end for
Hence, obtaining the sequence {rx,t } is O(S), i.e. the computational complexity is linear in the
number of projection steps.
3.9
Summary
Having derived the vector form of the recurrence using an arbitrary, unstructured, case the recurrence relation should hold for all non-unit linked policy types where the policy type may change
(resulting from a change in the survival state of x) at undetermined future times, so long as the
nominal amount of each possible cash flow can be determined in advance.
It is important to notice that everything in this Section has been discussed in terms of unit
timescales. The derivation is independent of the scaling factor and therefore it applies directly to
projection steps of any length, in any investigation, without any need for ‘rescaling’; the assumption
in Section 3.7.1 can therefore applied without further adjustments.
4
Use Cases
Here we state the values with which the various components of Equation 3.1 must be populated in
order for the equation to be applicable to a small selection of contract types of interest.
4.1
Single-Life Contracts
For a single life, the simplest model has only two states, alive and dead; for this model, the states
to which it is possible to migrate can be tabulated as
Current State
0
1
which produces the transition matrix
p
Wx,t,g = g x
0
and hence the vector relation is
rx,t,0
p
= vtf f x
rx,t,1
0
4.1.1
g qx
x1
alive
dead
Possible Next State
01
1
1
cx,t,0
p
+ vt 1 x
1
cx,t,1
0
f qx
g ∈ {f, 1}
(4.1)
rx+1,t,0
1
rx+1,t,1
(4.2)
1 qx
Single Life Annuities
For single life annuities, Equation 4.2 can be intepreted directly as
0
0
ax
cx,t
px qx
ax+1
f px f qx
f
=v
+v
0
0
1
0
0 1
0
where the j index on the cash flow has been dropped since there is only one non-trivial cash flow.
Removing the ZRS leads to
a0x = v f f px cx,t + v px a0x+1
16
Different values of f lead to annuities where the timing differs, i.e. in advance, in arrear, or
part-way through the step. Varying cx,t leads to the recurrences for other differences in types of
annuity;
- for a level annuity, cx,t = θ where θ is constant,
- for an increasing annuity, cx,t = φt where φt increases in arithmetic progression,
- for an escalating annuity, cx,t = ωt where ωt increases in geometric progression,
- for a limited term annuity, cx,t = 0 for all time steps after the end of the policy term.
Hence, for a level annuity of amount θ = 1, payable in advance (so that f = 0), the relation is
äx = 1 + v px äx+1
as presented in Section 2.3.1.
For whole life annuities an appropriate boundary condition is s p120 = 0 for s > 0, and for
limited term annuities where the policy was effected by a life aged x and the original term was n
an appropriate boundary condition is a0x+n:0 = 0. Proceeding in this manner provides the annuity
factor which applies to a life aged x at time t.
4.1.2
Single Life Endowments
A pure endowment may be viewed as an annuity where all cash flows except one are zero, the nonzero cash flow being at the time the endowment is payable. Using this interpretation, the usual
T
value for the cash flow vector is cx,t = 0 0 , and the non-zero cash flow for a policy effected by a
T
life aged x at inception with original term n is cx+n,0 = 1 0 ; this non-zero cash flow is also the
boundary condition required to obtain the assurance factor as rx,0 while rx,1 is, again, redundant.
Therefore Equation 4.2 becomes
1
Ax+1:n−1
Ax:n1
px qx
0
f p x f qx
f
+v
=v
0
0 1
0
0
1
0
where f ∈ [0, 1] is arbitrary, and removing the ZRS leads to
1
Ax:n1 = v px Ax+1:n−1
with the boundary condition Ax+n:01 = 1.
4.1.3
Single Life Assurances
It is possible to show that assurances cannot use a Markov transition matrix with only two states.
This is because the assumption that the cash flows are computable as a function of x, t and state
does not hold; this results from benefit being payable on the transition from one state to another,
rather than the continuance in a particular state.
A straightforward solution is to introduce a third state, say ‘died in step’, from which the life
transfers to the dead state in the next step with probability 1. Under this construction, ternary
labelling must be used for the states, and the states to which it is possible to migrate can be
tabulated as
Current State
0
1
2
x1
alive
died in step
dead
17
Possible Next State
01
2
2
which produces the transition matrix
g p x g qx 0
0 1
Wx,t,g = 0
0
0 1
By setting the cash flow vector to cx,t =
becomes
rx,t,0
f p x f qx
rx,t,1 = v f 0
0
0
0
rx,t,2
g ∈ {f, 1}
T
0 1 0 for each step,
0
0
p x qx
1 1 + v 0 0
1
0
0 0
(4.3)
the vector form of the equation
0
rx+1,t,0
1 rx+1,t,1
1
rx+1,t,2
Even though another state has been introduced, the boundary conditions used in Section 4.1.1
are still appropriate; i.e. s p120 = 0 for s > 0 in the case of a whole life assurance, and rx+n,0 = 0
in the case of a term assurance. Removing the ZRSs leads to
rx,t,0 = v f f qx + v px rx+1,t,0
By setting f = 1, it is apparent that, depending on the boundary condition used, this is equivalent
to either a single-life whole life assurance, so that the relation is
Ax = v qx + v px Ax+1
or a single-life term assurance, where the relation is
1
1
Ax:n
= v qx + v px Ax+1:n−1
4.2
Two-Life Contracts
For two lives, using x and y rather than x1 and x2 , the states to which it is possible to migrate can
be tabulated as
Current State
0
1
2
3
Binary
00
01
10
11
and the complete transition matrix is
g p x g p y g p x g qy
0
g px
Wx,t,g =
0
0
0
0
4.2.1
x
alive
alive
dead
dead
y
alive
dead
alive
dead
Possible Future States
0123
13
23
3
g qx g p y
g qx g qy
0
p
g y
0
g qx
g qy
g ∈ {f, 1}
1
Two-Life Annuities
Many practising actuaries are only interested in a) single life annuities and b) reversionary annuities
(where the benefit is payable to a second life after the death of the first, so long as the survivorship
follows the ordering specified in the policy document). However, there are other forms of two-life
annuity, i.e. a) joint-life annuities (which are payable so long as both lives are alive at the time of
payment), and b) last survivor annuities (which are payable so long as at least one of the lives are
alive at the time of payment).
As with single life annuities, different values of f lead to annuities where the timing differs, and
varying cx,t leads to the relations for differences in types of annuity. However, in contrast to single
life annuities, the recurrence equation has different interpretations depending on how the cash flow
vector is populated;
18
T
T
a) when cx,t = 0 0 1 0 , rx,t can be interpreted as a0x|y 0 a0y 0
which is precisely
what is required to obtain the reserve factors for a reversionary annuity, per the derivation in
Section 2.4.2.
T
T
b) when cx,t = 1 0 0 0 , rx,t can be interpreted as a0x,y 0 0 0 which corresponds to
the recurrence required to obtain for the reserve factors for a joint life annuity.
T
T
0
0
c) when cx,t = 1 1 1 0 , rx,t can be interpreted as a0(LS)
which corresponds
a
a
0
x,y
x
y
to the recurrence required to obtain for the reserve factors for a last survivor annuity. The
probability that a payment on a last survivor annuity is made is usually given as 1 − f qx f qy ,
i.e. the probability that both are not dead; however, it is straightforward to show that
1 − f qx f qy = f px f py + f px f qy + f qx f py
so that the top row of Wx,t,f is, indeed, the required probability for payment on a last survivor
annuity when the cash flow vector is populated as given.
For whole life annuities which are payable on two lives, appropriate boundary conditions are
∗
∗
s p120 = 0 for s > 0. For limited term policies, an appropriate condition a = 0 where a is the
reserve factor at maturity of the policy.
4.2.2
Two-Life Endowments
Section 4.1.2 uses the observation that a single life endowment can be regarded as an annuity with
only one non-zero payment; a similar interpretation can be applied to two-life endowments. By
setting all cash flows except the one at maturity to zero, reserve factors for two-life endowments
T
T
can be obtained by setting cx,t = 0 0 0 0 in all steps and using cx+n,0 = 1 0 0 0 to
give Ax,y:n1 as the first element of rx,t , i.e. having removed ZRSs, the relation in standard notation
is
1
Ax,y:n1 = v g px g py Ax+1,y+1:n−1
subject to the boundary condition
Ax+n,y+n:01 = 1
4.2.3
Two-Life Assurances
For two-life assurances, the third
transition matrix for this model is
g px g py g px g qy
0
0
0
0
0
0
0
0
Wx,t,g =
0
0
0
0
0
0
0
0
state (died in step) must be applied to both lives and so the
0
g px
g px
0
0
0
0
0
0
g qx g p y
g qx g qy
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
g qx
g qx
0
0
0
0
0
0
0
0
0
g py
0
0
g py
0
0
0
0
0
g qy
0
0
g qy
0
0
0
0
0
0
1
1
0
1
1
g ∈ {f, 1}
As with other policy types described above, keeping cx,t fixed leads to level two-life assurances,
whereas having a varying cx,t leads to increasing assurances, escalating assurances, or fixed-term
assurances, depending on the flavour of the variation.
19
When the benefit is payable on the first death, cash flow vectors for level assurances should be
T
cx,t = 0 1 0 1 1 0 0 0 0
and when the benefit is payable on the second death, the cash flow vectors for level assurances
should be
T
cx,t = 0 0 0 0 1 1 0 1 0
4.3
More Than Two Lives
Although cases where there are more than two lives are uncommon, we include them for completeness.
The two-life models in Section 4.2 are straightforward extensions of the single life models in
Section 4.1. For larger numbers of lives, the transition matrices may be obtained by induction,
using tensor products.
m
Let Wx,t,g
be the transition matrix required for m lives. Then
a) for the two state model (used for annuities and endowments),
g p xm g q xm
m−1
m
Wx,t,g = Wx,t,g ⊗
0
1
m
0
has 4m elements, only 3m of which
= (1); there are 2m possible states so that Wx,t,g
with Wx,t,g
are non-zero.
b) for the three state model (used for assurances),
g pxm
m−1
m
Wx,t,g
= Wx,t,g
⊗ 0
0
g q xm
0
0
0
1
1
m
0
has 9m elements, only 4m of which
= (1); there are 3m possible states so that Wx,t,g
with Wx,t,g
are non-zero.
4.4
With-Profit Policies
The algorithm can be applied to with-profit policies so long as the cash flows can be determined
in advance of their use so that the assumption in Section 3.7.1 holds. This should not pose any
practical problems because the cash flows can be calculated in the forward loop and then used in
the backward loop. Even if the bonuses are applied on a two-tier basis, this algorithm can be used if
estimated bonus rates are available in advance of calculating the bonuses attaching at a particular
time.
4.5
Other Benefits
The algorithm can also be applied to other types of policies; for example Critical Illness, Permanent
Health Insurance and Unemployment Income policies may all be modelled using a Markov transition
matrix derived from the relevant state transition diagram. Given the reversible nature of some of
the transitions (say, between the ‘able’ and ‘ill’ states of a P.H.I. policy), the Markov transition
matrix might not be upper triangular.
20
4.6
Step Lengths
In all of the cases considered in Sections 4.1 and 4.2, the relationships are based on time steps of
unit length; in theoretical work, it is usually assumed that the default length of a step is a year.
Cases where cash flows happen more frequently need to be adjusted to allow for the frequency of
payments.
Using this vector form of the recurrence there is no need for such adjustments; the sequence
{cx,t } indicates the nominal amount of each cash flow. Hence, in the case of escalating payments,
for example, the stream of payments which populate the sequence {cx,t } must already include
the allowance for escalation and the timing of the cash flows. Similarly, if a projection is being
performed using monthly steps and the cash flows occur yearly, then the sequence {cx,t } would
have a non-zero value for every twelfth step only.
4.7
Summary
The vector form of the recurrence automatically allows for all possible combinations of lengths of
time steps, and variability in cash flow, simply by populating the sequence of cash flow vectors
appropriately, and populating Wx,t,g accordingly. For each of the two-state and three-state models
presented here, there is only one transition matrix for m lives; how each matrix is applied will vary
depending on the type of policy under consideration.
The examples presented in this section have an equivalent interpretation to relationships which
may be derived from first principles using the relevant summation for a particular type of policy;
the derivations are straightforward, if laborious.
The collection of recurrences described above indicates that this vector approach can be applied
to a wide range of contracts; the collection of policy types included here is not intended to be
exhaustive, but it is meant to demonstrate that any potential restrictions which appear to apply to
non-unit linked policies can be overcome, and that the relation can be used in the form presented
here.
5
Performance Implications
This paper is mainly theoretical, and is intended as the basis for future implementations; the
timing results included here illustrate the potential performance benefits of our recurrence relation
approach.
To test the performance of our implementation, we created some synthetic data which has
properties which are representative of a cohort of people who retired recently, either at normal
retirement age, or slightly early. The main characteristics of these data are; the date of birth
is uniformly distributed so that the age at the valuation date is between 57 and 67; the policy
inception date is uniformly distributed over the calendar year prior to the valuation date, so that
these policies represent a cohort of new business; roughly 75% of the policyholders are male; roughly
80% of the policies have payments made monthly, the remainder having payments made annually;
the amount paid at each payment has a log-normal distribution with a mean of roughly 5.0 and
standard deviation of about 1.5; approximately 95% of the policies have no escalation and the
remainder escalate at either 3%, 4.25% or 5% on each policy anniversary. For two-life policies,
both lives were drawn from the same uniform distribution of ages, with no regard as to which of
the lives was older.
All timings were obtained using a laptop with an Intel i5-2450 2.5GHz dual core CPU running
Windows 7, and using Intel ifort 12.1. Although this platform is primarily used for development,
it gives an indication of performance available within industry where PCs with i5 and i7 CPUs are
21
widely available.
5.1
5.1.1
Effect of Changing the Algorithm
Measuring Performance
Our starting point was a Fortran 90 code which closely mirrored the code produced by one of the
commercial actuarial valuation packages in that a) the financial results produced by our program
agreed to those from the commercial system to the 7 significant figures output when the same
demographic and economic assumptions were used, and b) the performance of our original code was
almost identical to that obtainable from the commercial package. The code went through several
stages of optimisation including implementation of the vector version of the recurrence approach;
this was not particularly difficult, and led to significant speedup in its own right. The code was
then parallelised using OpenMP, with a decomposition of policies across available threads, and a
dynamic loop schedule to overcome any potential load imbalance resulting from different policies
requiring different numbers of time steps. A more complete discussion of that optimisation is given
in our previous work (see Tucker and Bull, 2012).
5.1.2
Performance Improvement
Figure 1 shows the average time to process single-life annuity policies using the two algorithms in
Section 3.8; the implementations of the algorithms had the same level of optimisation, so that the
only difference between the runs was the algorithm. It is clear that the shapes of the curves in that
graph match the complexities stated in Sections 3.8.1 and 3.8.2.
Also, on the scale presented to demonstrate the quadratic nature of the summation approach,
the times for the recurrence approach are almost indistinguishable from the axis; this indicates
that as well as having better scaling than the summation approach, the recurrence approach is far
faster than the summation approach; in fact, recursion is two about orders of magnitude faster
than summation.
5.2
Scalability of the Algorithm
A highly optimised ab-initio implementation of our recurrence relation approach was produced in
Fortran 90, and its performance was measured for annuity contracts.
5.2.1
Timing Methodology
To allow for operating system daemons, each timing run was performed seven times; the minimum
of those runs is reported. This method produces a smooth progression because the times reported
approach those which are the best which might be obtained if nothing else was running on the
system.
5.2.2
Performance Results
Figure 2 shows the average time to process single-life and reversionary annuity policies using the the
recurrence approach to calculating reserves; it is clear that using the recurrence approach creates
linear scaling with number of steps for both types of annuity; this agrees with the complexities
derived in Section 3.8.2. The fact that the slope of the line for the reversionary annuity policies is
roughly twice the slope for the single-life annuity policies is to be expected because, for a reversionary annuity, there are twice as many lives for which {lx } needs to be obtained and, at each step,
two reserve factors are required, per Equation 2.5.
22
5.3
Predicted Performance
5.3.1
Representative Portfolio
The information obtained from placing timers in our optimised implementation of the recurrence
algorithm is shown in Table 1.
The fact that calculating {lx+t } takes twice as long for two-life policies as for single life policies
is to be expected since, on average, twice as many lx calculations are required, and this is simply
a result of the fact that there twice as many lives, and in our synthetic data they are drawn from
the same uniform distribution of ages at the valuation date.
5.3.2
Predicted Results
Table 2 shows the composition of a representative portfolio of 500,000 annuity policies.
Using the results from Table 1, the wall-clock time to process that representative portfolio using
both cores of the dual core laptop is estimated to be 5.7 seconds. Since the ages at the valuation
date have a uniform distribution, the average age (at the valuation date) of the representative
portfolio is 62. Therefore, assuming a limiting age of 120, the average outstanding projection term
is 58 years (i.e. 696 months).
Hence, using the same reasoning as in Section 2.1, the time to process a ‘brute-force’ solvency
calculation, using 1000 scenarios at each future monthly step, is estimated to be
696 695 694
1
1000 ×
+
+
+ ··· +
× 5.7 sec ≈ 550 hr.
696 696 696
696
This is the time estimate using one dual-core laptop; clearly, it compares extremely favourably
with the estimate of about 16 years obtained in Section 2.1. However, Section 6.2 discusses the
possibility that this time could still be significantly reduced by using different hardware.
6
6.1
Conclusion
Overview
This work has demonstrated that a vector form of recurrence relations can be found for many nonunit linked life assurance contracts of arbitrary complexity, and has provided specific recurrence
relations for several exemplar contracts. It has demonstrated that efficient implementations of the
recurrence relations have scalability which is particularly favourable.
6.2
Future Work
As part of our investigation into the use of High Performance Computing techniques, this work
forms the foundation of an implementation of the brute force approach to calculating liabilities for
use in demonstrating solvency. Work to transfer vectorised code to a 48-core SMP, using OpenMP
to coordinate the threads, has already begun. It is expected that the scalability of this method is
particularly good; given the results of our previous work, we expect to achieve over 95% efficiency
on the highest numbers of cores.
Following that, we plan to implement a version of the calculations on Graphics Processing Units
(GPUs). GPUs are ideal for problems which require a large amount of processing on a relatively
small amount of data, thereby overcoming the cost of transferring the data to the device. For a
projection of N policies over S steps, the amount of data required is O(N ), and the amount of
23
computation (using a recurrence approach) is O(N S), making these solvency calculations especially
well-suited to such an implementation.
GPUs can process data roughly 50 to 60 times more quickly than CPUs; for the representative
portfolio in Section 5.3, a single GPU might be able to process 1000 scenarios at each future step in
roughly 10 hours, or ‘over night’, and a cluster of, say, 20 GPUs would be able to do the processing
‘over lunch’. This brings a full brute force calculation of liabilities for the solvency calculation
within the realms of practicality.
Acknowledgements
We would like to thank Gavin Conn for his comments on drafts of Sections 2, 3 and 4 of this paper.
References
Cairns, A. J. G., Blake, D. and Dowd, K. (2008). Modelling and Management of Mortality Risk:
A Review. Scandinavian Actuarial Journal, Volume 108, 2008, No. 2, pp.79-113.
Chen, L. (1996). Stochastic Mean and Stochastic Volatility - A Three-Factor Model of the Term
Structure of Interest Rates and Its Application to the Pricing of Interest Rate Derivatives. Interest
Rate Dynamics, Derivatives Pricing, and Risk Management. Lecture Notes in Economics and
Mathematical Systems, 435. Springer. ISBN 978-3-540-60814-1.
Corsaro, S., de Angelis, P.L., Marino, Z., Perla, F. and Zanetti, P. (2010). On parallel asset-liability
management in life insurance: a forward risk-neutral approach. Parallel Computing, Volume 36:7,
pp.390-402.
Cox, J.C., Ingersoll, J.E. and Ross, S.A. (1985). A Theory of the Term Structure of Interest Rates.
Econometrica 53: pp.385-407.
Green Tables (1980). Formulae and Tables for Actuarial Examinations. Institute Of Actuaries and
Faculty Of Actuaries.
Macdonald, A.S. (2004). Commutation Functions. In Encyclopedia Of Actuarial Science, Volume
1, pp.300-302.
McCutcheon, J.J. and Scott, W.F. (1991). An Introduction to The Mathematics Of Finance.
Butterworth Heinemann, Oxford.
Neill, A. (1989). Life Contingencies. Institute Of Actuaries and Faculty Of Actuaries.
OMP 3.1. (2011) OpenMP Architecture Review Board. OpenMP Application Program Interface,
Version 3.1. www.openmp.org/mp-documents/OpenMP3.1.pdf [Accessed September 2013].
Solvency II (2008) Commission of the European Communities. Amended Proposal for a Directive
of the European Parliament and the Council on the Taking-up and Pursuit of the Business of Insurance and Reinsurance [SOLVENCY II]. COM/2008/0119 final/2 - COD 2007/0143. Brussels,
2008.
Tucker, M. and Bull, J.M. (2012). The Application of High Performance Computing to Solvency and
Profitability Calculations for Life Assurance Contracts. In proceedings of 5th Workshop on High
Performance Computational Finance, SuperComputing, 2012. (Published in High Performance
Computing, Networking, Storage and Analysis 2012 (SC Companion)).
Vasicek, O. (1977). An Equilibrium Characterisation of the Term Structure. Journal of Financial
Economics 5 (2): pp.177-188.
24
Annuity Type
Number of lives
Number of reserve factors
Time to calculate {lx+t }
Time to calculate the reserve factors
Total processing time
Single
Life
1
1
2.67
2.54
6.73
Joint
Life
2
1
5.32
3.18
10.39
Reversion
2
2
5.32
5.11
12.37
Last
Survivor
2
3
5.19
8.77
15.90
Table 1: Time (seconds) to process 400,000 annuities of each type (using 2 threads on both cores
of an Intel i5-2450 2.5GHz dual core CPU). Note that these are ‘total computing’ times, so that
wall-clock times would be about 50% of the figures given.
25
Policy Type
Number of policies
Single
Life
300,000
Joint
Life
50,000
Reversion
100,000
Last
Survivor
50,000
Table 2: Composition of a representative portfolio of 500,000 annuity policies.
26
1.50E-02
1.25E-02
Average Time per Policy (seconds)
1.00E-02
7.50E-03
5.00E-03
2.50E-03
0.00E+00
360
420
480
540
600 660 720 780 840
Number of Projection Steps
Summation
900
960 1020
Recurrence
Figure 1: Time to calculate single life annuity reserves, averaged over 60,000 policies, using naive
summation and recurrence algorithms on an Intel i5-2450 2.5GHz dual core CPU.
27
5.0E-05
Average Time per Policy (seconds)
4.0E-05
3.0E-05
2.0E-05
1.0E-05
0.0E+00
360
420
480
540
600 660 720 780 840
Number of Projection Steps
Single Life
900
960 1020
Reversionary
Figure 2: Time to calculate single life and reversionary annuity reserves, averaged over 60,000
policies, using recurrence algorithms on an Intel i5-2450 2.5GHz dual core CPU.
28
| 5cs.CE
|
Vulnerability of linear systems against sensor attacks–a system’s
security index
arXiv:1602.06594v1 [cs.SY] 21 Feb 2016
Michelle S. Chong1 and Margreta Kuijper2
Abstract— The ‘security index’ of a discrete-time LTI system
under sensor attacks is introduced as a quantitative measure
on the security of an observable system. We derive ideas from
error control coding theory to provide sufficient conditions for
attack detection and correction.
I. I NTRODUCTION
The security of control systems against adversarial attacks
is a challenge to maintain when the adversary knows the
workings of any component of the system and has garnered
access, with the malicious intent of causing disruption. This
has lead to a proliferation of works in tackling this issue,
in particular in detecting the occurrence of an attack [9],
[8], [10], or in designing resilient control or estimation
algorithms, see [3], [11], [2], [1], [4] and many more.
In this paper, we concentrate on LTI systems where the
sensing component has been compromised by the attacker,
who has full knowledge of the system. The vulnerability of
the sensors is modelled by an additive attack signal to the
sensor measurements, which is non-zero when the particular
sensor is compromised. Inspired by ideas in coding theory,
we introduce the notion of the ‘security index’ for linear systems, a quantitative measure of the vulnerability of a system
to sensor attacks. While ideas from error control coding have
already been employed to this context in recent literature [3],
our aim is to further strengthen this link. Our notion of a
‘security index’ is formulated based on the measurement time
series from all sensors and is analogous to the notion of the
‘minimum distance’ of a code in error control coding theory.
We demonstrate that by using ideas from coding theory, the
formulation simplifies the approach in [3], leading to new
results. Particularly, we express the ‘security index’ of a
system in terms of different representations of the system
concerned.
Previous works in state estimation for systems under
sensor attacks include [2], [3], [1], [9], [10], [7]. There is
a consensus with [3] and [2] that the states of an LTI system
can only be reconstructed if strictly less than half of the
sensors are under attack. We will see in this paper that this
condition is also derived when approached with ideas from
coding theory. Other related works are [5], [4] which focus
on power networks. It is in this specialised setting that the
authors of [4] introduce the terminology ‘security index’,
1 Michelle
S. Chong is with the Department of Automatic Control, Lund University, SE-221 00 Lund, Sweden.
[email protected]
2 Margreta Kuijper is
Electronic
Engineering,
with the Department of Electrical and
University
of
Melbourne,
Australia.
[email protected]
which we adopted for a broader context in this paper. The
presence of measurement noise has been considered in [7],
which we do not consider, but is the subject of further work.
Notation: We denote the set of integers and complex numbers
as Z and C, respectively. The notation Z+ is used to denote
the set of positive integers including 0.
II. P ROBLEM
FORMULATION
We consider a discrete-time, observable linear timeinvariant (LTI) system Σ given by a n × n state matrix A
and a N × n observation matrix C, defined as follows:
x(t + 1) =
y(t) =
Ax(t)
(1)
Cx(t).
(2)
The behavior B of the system is defined as the set of all
possible output trajectories y : Z+ 7→ CN that satisfy the
system’s equations for some initial condition x(0) ∈ Cn .
Due to the time-invariant finite dimensional nature of the
underlying system, the behavior B has the following two
properties:
• B is left-shift invariant: if y ∈ B then σy ∈ B, where
the shift operator σ is defined via σy(t) := y(t + 1).
• B is autonomous: there exists T ∈ Z+ such that for any
y ∈ B and ỹ ∈ B we have that y|[0,T ] = ỹ|[0,T ] implies
that y = ỹ.
We assess the vulnerability of an LTI system via its measurable outputs, which may have been compromised by an
attacker. While the usual assumption for many applications
is that the matrix C has full row rank, we do not make
such an assumption in our setting. This setting occurs in
the case where each sensor is measuring a local part of the
system, such as in sensor networks implemented in a large
geographical location. To aid in the introduction of a measure
of a system’s security against sensor attacks, we define the
following for a system’s trajectory.
Definition 2.1: The support of a trajectory y : Z+ 7→ CN ,
denoted by supp (y), is defined as the set of indices i
in {0, 1, 2, . . . , N } such that its component trajectory yi :
Z+ 7→ C is not the zero trajectory.
Definition 2.2: The weight of a trajectory y, denoted by
kyk, is defined as |supp (y)|, i.e., the number of components
of y that are not the zero trajectory.
We now introduce a concept that is central to this paper:
Definition 2.3: The security index of the system Σ is
defined as
δ(Σ) := min kyk.
06=y∈B
This notion plays a paramount role in our investigation into
the resilience of the system under adversarial attack. More
precisely, we consider attacks on the system Σ that result in
the system Σ′ given by:
Σ′ : x(t + 1) = Ax(t)
r(t)
= Cx(t) + η(t),
where η : Z+ 7→ CN is the unknown attack signal and
r : Z+ 7→ CN is the known received signal. Thus we are
focusing exclusively on scenarios where the system’s outputs
(= sensors) are attacked. The behavior B ′ of the system Σ′
is defined as the set of all possible trajectories r that satisfy
the above equations for some initial condition x(0) and some
attack signal η. We consider the following two problems:
Problem 1 (attack detection): Given received signal
r ∈ B ′ , detect that r ∈
/ B.
Problem 2 (attack correction): Given received signal r ∈ B ′ , find y ∈ B such that kr−yk is minimal.
In the sections that follow, we derive conditions such that
the problems above are solvable in a tractable manner.
III. C ONDITIONS
FOR ATTACK DETECTION AND
CORRECTION USING A SYSTEM ’ S SECURITY INDEX
The first question that arises is: under which conditions
on the attack signal η are these problems solvable? We have
the following results.
Theorem 3.1: Suppose the received signal r ∈ B ′ corresponds to an attack η with kηk an unknown non-zero value
< δ(Σ). Then r ∈
/ B, i.e., attack detection is possible.
Proof: Let r and η be as stated in the theorem and let
y ∈ B be such that r = y + η. Then r − y ∈
/ B because
of Definition 2.3 and the assumption that 0 6= kηk < δ(Σ).
Since B is linear it then follows that r ∈
/ B.
Consequently, we can interpret the security index δ(Σ) as
the minimum number of sensors that an attacker needs to
compromise without being detected. We call the system Σ
maximally secure if δ(Σ) = N . It is easily seen that generically, systems described by equations (1)-(2) are maximally
secure.
Theorem 3.2: Suppose the received signal r ∈ B ′ corresponds to an attack η with kηk an unknown value < δ(Σ)/2.
Then there exists a unique y ∈ B such that kr − yk is
minimal, i.e. unique attack correction is possible.
Proof: Let r be as stated in the theorem and let y1 ∈ B
and η 1 be such that r = y1 + η 1 with kη1 k < δ(Σ)/2.
Suppose that there also exist y2 ∈ B and η 2 such that r =
y2 + η 2 with kη2 k < δ(Σ)/2. Clearly kη1 − η 2 k < δ(Σ),
so that y1 − y2 = η 1 − η 2 is a trajectory in B of weight
< δ(Σ). Definition 2.3 now implies that y1 − y2 is the zero
trajectory, in other words, y1 = y2 is unique. It follows that
kr − y1 k = kη 1 k < δ(Σ)/2 is minimal.
We have formulated system Σ’s security index δ(Σ) and
results in therms of the output trajectory y, instead of Σ’s
initial condition x(0) to transparently draw an analogy with
error control coding. In fact, this choice is natural because the
recovery of x(0) is equivalent to having y ∈ B during attack
correction, due to the observability assumption on system Σ.
IV. C OMPUTING A SYSTEM ’ S SECURITY INDEX
In this section, we show how system Σ’s security index
δ(Σ) can be computed. To this end, we introduce the coding
matrix G of the system Σ defined as follows
Ci
G1
Ci A
G2
2
(3)
G = . , where Gi := Ci A
..
..
.
GN
Ci An−1
with Ci defined as the i’th row of C for i = 1, . . . , N .
We call the above matrix G the coding matrix of the
system as it exhibits the link with error control coding [6].
Note however in contrast to error control coding, the coding
matrix G cannot be chosen freely—instead, it is fixed and
given by the system Σ. In particular, the number of sensors N
is fixed. In the following theorem, we use GJ to denote the
matrix that is obtained by stacking the matrices Gi defined
in (3), for i ∈ J ⊆ {1, . . . , N }.
Theorem 4.1:
δ(Σ) = N − L,
where L is the largest integer in {0, 1, . . . , N } for which
there exists a subset J of {1, . . . , N } of cardinality L such
that ker GJ 6= {0}.
Proof: Let J be a subset of {0, 1, . . . , N } of cardinality
L such that ker GJ 6= {0}. Let x(0) 6= 0 be such that
GJ x(0) = 0. Then Gx(0) 6= 0 because of left invertibility
of G. Furthermore, the trajectory y that corresponds to initial
condition x(0) satisfies yJ = 0 (use Cayley-Hamilton).
Thus y is a trajectory in B of weight ≤ N − L, so that
δ(Σ) ≤ N − L. Further, by definition of L, trajectories in
B cannot have more than L of their component trajectories
equal to the zero trajectory. Therefore δ(Σ) = N − L.
Corollary 4.2: The system Σ given by equations (1)-(2)
is maximally secure if and only if each matrix Gi , as defined
in (3), has full column rank (i = 1, 2, . . . , N ).
Not surprisingly, a system is maximally secure if and only
if the system is observable via each sensor. In this case, we
obtain from Theorem 3.2 that the sufficient condition for the
attack on the system to be correctable is that the number
of compromised sensors is strictly less than half of the total
number of sensors. This conforms with the results in [3]
and [2] for discrete-time and continuous-time LTI systems,
respectively.
We further provide ways of computing a system’s security
index δ(Σ). Since we assume that the LTI system is observable and hence the nN × n matrix G given by (3) has full
rank, there exists a full rank (n(N − 1) × nN ) matrix H,
written as
(4)
H = H1 H 2 · · · H N ,
such that HG = 0; mindful of the analogous coding theoretic
terminology, in this paper we call such a matrix H a check
matrix of the system. In the next theorem HJ denotes
the matrix that is obtained by juxtaposing the matrices Hi
defined in (4), for i ∈ J ⊂ {1, . . . , N }.
Lemma 4.6: Let the system Σ be given by equations (1)(2). Let x(0) be a linear combination of eigenvectors of A
that correspond to different eigenvalues, so
Theorem 4.3:
δ(Σ) = spark (H),
where spark (H) is defined as the smallest integer L in
{1, . . . , N } for which there exists a subset J of {1, . . . , N }
of cardinality L such that ker HJ 6= {0}.
Proof: Let J be a subset of {1, . . . , N } of cardinality
L such that ker HJ 6= {0}. Let yJ be a non-zero trajectory
such that HJ yJ = 0. Let y be the trajectory that coincides
with yJ at the appropriate locations and that has zero
component trajectories at all other locations. Then y is a
trajectory in B of weight ≤ spark (H), so that δ(Σ) ≤
spark (H). Further, by definition of spark (H), trajectories
in B have weight ≥ spark (H) (use Cayley-Hamilton) and
therefore δ(Σ) = spark (H).
The terminology ”spark” stems from the compressed sensing literature [12].
Corollary 4.4: The system Σ given by equations (1)-(2)
is maximally secure if and only if all square n(N − 1) ×
n(N − 1) submatrices of H of the form HJ , where J is a
subset of {1, . . . , N } of cardinality N − 1, are nonsingular.
An alternative representation of the system Σ is given by
a set of N difference equations
R(σ)y = 0,
(5)
where R(ξ) is a N × N polynomial matrix and σ represents
the left shift, as before. In the special case where R(ξ)
corresponds to a minimal lag representation, its N row
degrees are the observability indices of the system Σ.
Recall that a square polynomial matrix is called unimodular if it has a polynomial inverse; a nonsquare polynomial
matrix is called left unimodular if it has a polynomial left
inverse. Two polynomial matrices R(ξ) and Q(ξ) of the same
size are called left unimodularly equivalent if there exists a
unimodular matrix U (ξ) such that Q(ξ) = U (ξ)R(ξ).
Theorem 4.5: Let the system Σ be given by (5). Then its
security index δ(Σ) is given by the smallest integer L in
{1, . . . , N } for which there exists a subset J of {1, . . . , N }
of cardinality L such that RJ (ξ) is not left unimodular (here
RJ (ξ) denotes the matrix that consists of the ith columns
of R(ξ) where i ∈ J ).
Proof: Let J be a subset of {1, . . . , N } of cardinality
L such that RJ (ξ) is not left unimodular. Then there exists
yJ 6= 0 such that RJ yJ = 0. Let y be the trajectory that
coincides with yJ at the appropriate locations and that has
zero component trajectories at all other locations. Then y
is a trajectory in B of weight ≤ L, so that δ(Σ) ≤ L. By
definition of L we must also have δ(Σ) ≥ L which proves
the theorem.
In the remainder of this section, we show that the system
Σ’s security index δ(Σ) can be computed more easily by
exploiting the system’s special structure. Below, supp (y)
denotes the support of a vector y ∈ CN , i.e. the set of
indices i in {0, 1, 2, . . . , N } such that yi 6= 0. As mentioned
in Definition 2.1, supp (y) denotes the support of a trajectory
y, i.e. the set of indices i in {0, 1, 2, . . . , N } such that yi is
not the zero trajectory.
x(0) = α1 v1 + α2 v2 + · · · + αm vm ,
(6)
where 0 6= αj ∈ C, Avj = λj vj for j = 1, 2, . . . , m and
λj 6= λi for j 6= i. Let y be the trajectory in the behavior B
of Σ that corresponds to x(0). Then
supp (y) = ∪m
(7)
j=1 supp (Cvj ).
Proof: It follows immediately from the definition of y
that
supp (y) ⊂ ∪m
j=1 supp (Cvj ).
(8)
To prove the reverse inclusion, let ℓ ∈ ∪m
j=1 supp (Cvj ).
Then there exists i ∈ {1, 2, . . . , m} such that Cℓ vi 6= 0.
Since αi is assumed non-zero it follows that αi Cℓ vi 6= 0.
Then
yℓ (0)
1
1 ···
1
α1 Cℓ v1
yℓ (1) λ1 λ2 · · · λm α2 Cℓ v2
yℓ (2) = λ21 λ22 · · · λ2m
..
.
..
..
..
..
..
.
.
.
.
α Cv
.
m
ℓ m
is non-zero because of the Vandermonde structure and the
fact that all λi ’s are distinct. Thus yℓ is not the zero trajectory
so that ℓ ∈ supp (y). This implies that
∪m
j=1 supp (Cvj ) ⊂ supp (y).
(9)
From (8) and (9) we conclude that (7) holds.
A special consequence of the above lemma is that the
system’s security index δ(Σ) is determined by the weight
of the special trajectories that have initial conditions in an
eigenspace Vi of A:
Theorem 4.7: Let the system Σ be given by equations (1)(2). Let λ1 , λ2 , . . . , λm be the m distinct eigenvalues of the
matrix A. Let Cn = V1 ∪ V2 ∪ . . . ∪ Vm , where Vi is the
subspace spanned by all eigenvectors corresponding to λi
for i = 1, 2, . . . , m. Then
δ(Σ) =
min
min |supp (Cx)|,
j∈{1,2,...,m} x∈Vj
In particular, if m = n, meaning that all n eigenvalues of A
are distinct, then
δ(Σ) =
min
j∈{1,2,...,n}
|supp (Cvj )|,
where v1 , v2 , . . . , vn is a basis of eigenvectors of A.
Proof: Let y be an arbitrary trajectory in B, corresponding to initial condition x(0), written as in (6). It follows from
Lemma 4.6 that
kyk = |supp (y)| ≥
min
j∈{1,2,...,m}
|supp (Cvj )|.
This implies that
δ(Σ) ≥
min
min |supp (Cx)|.
j∈{1,2,...,m} x∈Vj
To prove the reverse inequality, let j ∈ {1, 2, . . . , m} and
x ∈ Vj be such that |supp (Cx)| is minimal. Let y ∈ B be
the system’s trajectory that corresponds to initial condition
x(0) := x. From Ax = λj x it follows immediately that
kyk = |supp (Cx)|. Therefore
δ(Σ) ≤
min
min |supp (Cx)|.
j∈{1,2,...,m} x∈Vj
which proves the theorem.
From the theorem above, we see that the computation of
the security index is straightforward for a system that has
n distinct eigenvalues : simply use a diagonal A-matrix; the
security index of the system is then given by the size of
the support of the sparsest column of the corresponding Cmatrix. In particular, the system is maximally secure if and
only if this C-matrix has no zero values. We now illustrate
the computation of system Σ’s security index through an
example.
Example 4.8: Let N = 2 and consider
1 1
λ1 0
, C=
A=
.
1 −1
0 λ2
Then it follows from Theorem 4.7 that for λ1 6= λ2 the
system’s security index is 2 so that by Theorem 3.1 detection
of attacks on one output is possible. Minimal lag equations
for this system are given by
σ − λ1
σ − λ1
y1
R(σ)y =
= 0.
σ − λ2 −(σ − λ2 )
y2
We observe that when λ1 6= λ2 , every column of R(σ) is
left unimodular. Hence, by Theorem 4.5, we obtain L = N
and thus, the security index δ(Σ) equals N .
V. ATTACK
DETECTION , CORRECTION AND OTHER
FUTURE WORK
Using the results developed in the previous sections, we
now discuss how they can be employed in detecting the
occurrence of an attack and how correction can be performed.
To start with detection, we first recall that the observability
assumption on our system implies that a check matrix
H (defined in (4)) exists. Attack detection is most easily
formulated in terms of H as follows.
Attack detection rule: Given received signal r ∈
B ′ , compute the ‘syndrome trajectory’ s := Hr
and conclude that an attack has taken place (meaning r ∈
/ B) if and only if s 6= 0.
Alternatively, we can use the description of the form
R(σ)y = 0 as the main ingredient of our decision rule to
detect/correct.
Attack detection rule: Given received signal r ∈
B ′ , compute the ‘syndrome trajectory’ s := R(σ)r
and conclude that an attack has taken place (meaning r ∈
/ B) if and only if s 6= 0.
In this paper, we choose to formulate all fundamental notions
and results in terms of the trajectory y rather than in terms of
the initial condition x(0) so as to have a clear fundamental
theory that exhibits the link with error control coding in a
transparent way. Indeed, we are able to make this choice because for attack correction the recovery of x(0) is equivalent
to the recovery of y ∈ B due to the observability assumption
on the system Σ. It is however important to note that in
practical situations the recovery of x(0) from the received
trajectory r is the main objective. In fact, we only need to
consider attack scenarios where attacks happen in the first n
time instances (otherwise we can simply reconstruct x(0)
from (r(0), r(1), · · · , r(n − 1)) = (y(0), y(1), · · · , y(n))
because of the observability of Σ).
For attack correction, the above syndrome trajectory s
should be used to identify the attack locations. Once the
attack locations have been found, x(0) can be computed
on the basis of the remaining attack-free components of y.
How to identify the attack locations from s in terms of our
knowledge of the system matrices A and C is a topic of
future research. Another topic of future research is the effect
of feedback control on the security index of an LTI system
with inputs.
R EFERENCES
[1] Y. Chen, S. Kar, and J.M.F. Moura. Cyber-physical systems: dynamic sensor attacks and strong observability. In Proc. of the 40th
International Conference on Acoustics, Speech and Signal Processing
(ICASSP), Brisbane, Australia, April 2015.
[2] M.S. Chong, M. Wakaiki, and J.P. Hespanha. Observability of linear
systems under adversarial attacks. In Proc. of the 2015 Amer. Contr.
Conf., Chicago, USA, July 2015.
[3] H. Fawzi, P. Tabuada, and S.N. Diggavi. Secure estimation and
control for cyber-physical systems under adversarial attacks. IEEE
Transactions of Automatic Control, 59(6):1454–1467, 2014.
[4] A. Teixeira H. Sandberg and K.H. Johansson. On security indices for
state estimators in power networks. In Proceedings First Workshop on
Secure Control Systems, CPSWEEK, Stockholm, Sweden, pages 1–6,
2010.
[5] J.M. Hendrickx, K.H. Johansson, R.M. Jungers, H. Sandberg, and K.C.
Sou. Efficient computations of a security index for false data attacks
in power networks. IEEE Trans. Automat. Contr., 59(12):3194–3208,
2014.
[6] C. Lee, H. Shim, and Y. Eun. Secure and robust state estimation
under sensor attacks, measurement noises, and process disturbances:
Observer-based combinatorial approach. In Proc. of the 2015 European Control Conference, Linz, Austria, July 2015.
[7] M. Pajic, P. Tabuada, I. Lee, and G.J. Pappas. Attack-resilitent state
estimation in the presence of noise. In The proceedings of the 54th
Conference on Decision and Control (CDC), 2015.
[8] F. Pasqualetti, F. Dörfler, and F. Bullo. Attack detection and identification in cyber-physical systems–part ii: Centralized and distributed
monitor design. arXiv preprint arXiv:1202.6049, 2012.
[9] F. Pasqualetti, F. Dorfler, and F. Bullo. Attack detection and identification in cyber-physical systems. IEEE Transactions on Automatic
Control, PP(99):1–1, 2013.
[10] F. Pasqualetti, F. Dorfler, and F. Bullo. A divide-and-conquer approach
to distributed attack identification. In The proceedings of the 54th IEEE
Conference on Decision and Control, 2015.
[11] Y. Shoukry and P. Tabuada. Event-triggered state observers for sparse
sensor noise/attacks. arXiv preprint arXiv:1309.3511, 2013.
[12] A.M. Tillmann and M.E. Pfetsch. The computational complexity of
the restricted isometry property, the nullspace property, and related
concepts in compressed sensing. IEEE Transactions on Information
Theory, 60(2):1248–1259, 2014.
| 3cs.SY
|
arXiv:1605.08283v1 [cs.LG] 26 May 2016
Discrete Deep Feature Extraction: A Theory and New Architectures
Thomas Wiatowski1
Michael Tschannen1
Aleksandar Stanić1
Philipp Grohs2
Helmut Bölcskei1
1
Dept. IT & EE, ETH Zurich, Switzerland
2
Dept. Math., University of Vienna, Austria
Abstract
First steps towards a mathematical theory of deep
convolutional neural networks for feature extraction were made—for the continuous-time case—
in Mallat, 2012, and Wiatowski and Bölcskei,
2015. This paper considers the discrete case,
introduces new convolutional neural network architectures, and proposes a mathematical framework for their analysis. Specifically, we establish deformation and translation sensitivity results of local and global nature, and we investigate how certain structural properties of the
input signal are reflected in the corresponding feature vectors. Our theory applies to
general filters and general Lipschitz-continuous
non-linearities and pooling operators. Experiments on handwritten digit classification and facial landmark detection—including feature importance evaluation—complement the theoretical findings.
1. Introduction
Deep convolutional neural networks (DCNNs) have proven
tremendously successful in a wide range of machine learning tasks (Bengio et al., 2013; LeCun et al., 2015). Such
networks are composed of multiple layers, each of which
computes convolutional transforms followed by the application of non-linearities and pooling operators.
DCNNs are typically distinguished according to (i)
whether the filters employed are learned (in a supervised
(LeCun et al., 1998; Huang & LeCun, 2006; Jarrett et al.,
2009) or unsupervised (Ranzato et al., 2006; 2007; JarProceedings of the 33 rd International Conference on Machine
Learning, New York, NY, USA, 2016. JMLR: W&CP volume
48. Copyright 2016 by the author(s).
WITHOMAS @ NARI . EE . ETHZ . CH
MICHAELT @ NARI . EE . ETHZ . CH
ASTANIC @ STUDENT. ETHZ . CH
PHILIPP. GROHS @ UNIVIE . AC . AT
BOELCSKEI @ NARI . EE . ETHZ . CH
rett et al., 2009) fashion) or pre-specified (and structured,
such as, e.g., wavelets (Serre et al., 2005; Mutch & Lowe,
2006; Mallat, 2012), or unstructured, such as random filters (Ranzato et al., 2007; Jarrett et al., 2009)), (ii) the
non-linearities used (e.g., logistic sigmoid, hyperbolic tangent, modulus, or rectified linear unit), and (iii) the pooling operator employed (e.g., sub-sampling, average pooling, or max-pooling). While a given choice of filters, nonlinearities, and pooling operators will lead to vastly different performance results across datasets, it is remarkable
that the overall DCNN architecture allows for impressive
classification results across an extraordinarily broad range
of applications. It is therefore of significant interest to understand the mechanisms underlying this universality.
First steps towards addressing this question and developing a mathematical theory of DCNNs for feature extraction were made—for the continuous-time case—in (Mallat,
2012; Wiatowski & Bölcskei, 2015). Specifically, (Mallat, 2012) analyzed so-called scattering networks, where
signals are propagated through layers that employ directional wavelet filters and modulus non-linearities but
no intra-layer pooling. The resulting wavelet-modulus
feature extractor is horizontally (i.e., in every network
layer) translation-invariant (accomplished by letting the
wavelet scale parameter go to infinity) and deformationstable, both properties of significance in practical feature extraction applications. Recently, (Wiatowski &
Bölcskei, 2015) considered Mallat-type networks with arbitrary filters (that may be learned or pre-specified), general Lipschitz-continuous non-linearities (e.g., rectified linear unit, shifted logistic sigmoid, hyperbolic tangent, and
the modulus function), and a continuous-time pooling operator that amounts to a dilation. The essence of the results in (Wiatowski & Bölcskei, 2015) is that vertical (i.e.,
asymptotically in the network depth) translation invariance
and Lipschitz continuity of the feature extractor are induced by the network structure per se rather than the specific choice of filters and non-linearities. For band-limited
Discrete Deep Feature Extraction: A Theory and New Architectures
signals (Wiatowski & Bölcskei, 2015), cartoon functions
(Grohs et al., 2016), and Lipschitz-continuous functions
(Grohs et al., 2016), Lipschitz continuity of the feature extractor automatically leads to bounds on deformation sensitivity.
A discrete-time setup for wavelet-modulus scattering networks (referred to as ScatNets) was considered in (Bruna
& Mallat, 2013).
We write (Tm f )[n] := f [n − m], m ∈ Z, for the cyclic
translation operator. The supremum norm of a continuoustime function c : R → C is kck∞ := supx∈R |c(x)|. The
indicator function of an interval [a, b] ⊆ R is defined as
1[a,b] (x) := 1, for x ∈ [a, b], and 1[a,b] (x) := 0, for
x ∈ R\[a, b]. The cardinality of the set A is denoted by
card(A).
2. The basic building block
Contributions. The purpose of the present paper is to develop a theory of discrete DCNNs for feature extraction.
Specifically, we follow the philosophy put forward in (Wiatowski & Bölcskei, 2015; Grohs et al., 2016). Our theory incorporates general filters, Lipschitz non-linearities,
and Lipschitz pooling operators. In addition, we introduce
and analyze a wide variety of new network architectures
which build the feature vector from subsets of the layers.
This leads us to the notions of local and global feature
vector properties with globality pertaining to characteristics brought out by the union of features across all network
layers, and locality identifying attributes made explicit in
individual layers.
Besides providing analytical performance results of general validity, we also investigate how certain structural properties of the input signal are reflected in the corresponding
feature vectors. Specifically, we analyze the (local and
global) deformation and translation sensitivity properties of
feature vectors corresponding to sampled cartoon functions
(Donoho, 2001). For simplicity of exposition, we focus on
the 1-D case throughout the paper, noting that the extension to the higher-dimensional case does not pose any significant difficulties.
Our theoretical results are complemented by extensive numerical studies on facial landmark detection and handwritten digit classification. Specifically, we elucidate the role of
local feature vector properties through a feature relevance
study.
Notation. The complex conjugate of z ∈ C is denoted by
z. We write Re(z) for the real, and Im(z) for the imaginary
part of z ∈ C. We let HN := {f : Z → C | f [n] = f [n +
N ], ∀ n ∈ Z} be the set of N -periodic discrete-time signals1 , and set IN := {0, 1, . . . , N − 1}. The delta function
δ ∈ HN is δ[n] := 1, for n = kN , k ∈ Z,
Pand δ[n] := 0,
else. For f, g ∈ HN , we set hf, gi :=
f [k]g[k],
P
P k∈IN
2 1/2
kf k1 :=
|f
[n]|,
kf
k
:=
(
|f
[n]|
) ,
2
n∈IN
n∈IN
and kf k∞ := supn∈IN |f [n]|. We denote the discrete Fourier transform (DFT) of f ∈ HN by fˆ[k] :=
P
−2πikn/N
. The circularPconvolution of f ∈
n∈IN f [n]e
HN and g ∈ HN is (f ∗ g)[n] := k∈IN f [k]g[n − k].
1
We note that HN is isometrically isomorphic to CN , but we
prefer to work with HN for the sake of expositional simplicity.
The basic building block of a DCNN, described in this section, consists of a convolutional transform followed by a
non-linearity and a pooling operation.
2.1. Convolutional transform
A convolutional transform is made up of a set of filters ΨΛ = {gλ }λ∈Λ . The finite index set Λ can be
thought of as labeling a collection of scales, directions, or
frequency-shifts. The filters gλ —referred to as atoms—
may be learned (in a supervised or unsupervised fashion),
pre-specified and unstructured such as random filters, or
pre-specified and structured such as wavelets, curvelets,
shearlets, or Weyl-Heisenberg functions.
Definition 1. Let Λ be a finite index set. The collection
ΨΛ = {gλ }λ∈Λ ⊆ HN is called a convolutional set with
Bessel bound B ≥ 0 if
X
kf ∗ gλ k22 ≤ Bkf k22 ,
∀f ∈ HN .
(1)
λ∈Λ
Condition (1) is equivalent to
X
|c
gλ [k]|2 ≤ B,
∀k ∈ IN ,
(2)
λ∈Λ
and hence, every finite set {gλ }λ∈Λ isP
a convolutional set
with Bessel bound B ∗ := maxk∈IN λ∈Λ |c
gλ [k]|2 . As
(f ∗ gλ )[n] = f, gλ [n − ·] , n ∈ IN , λ ∈ Λ, the outputs
of the filters gλ may be interpreted as inner products of the
input signal f with translates of the atoms gλ . Frame theory
(Daubechies, 1992) therefore tells us that the existence of a
lower bound A > 0 in (2) according to
X
A≤
|c
gλ [k]|2 ≤ B,
∀k ∈ IN ,
(3)
λ∈Λ
implies that every element in HN can be written
as a linear combination of elements in the set
gλ [n − ·] n∈I ,λ∈Λ (or in more technical parlance, the
N
set gλ [n − ·] n∈I ,λ∈Λ is complete for HN ). The abN
sence of a lower bound A > 0 may therefore result in ΨΛ
failing to extract essential features of the signal f . We note,
however, that even learned filters are likely to satisfy (3) as
all that is needed is, for each k ∈ IN , to have gc
λ [k] 6= 0 for
at least one λ ∈ Λ. As we shall see below, the existence of
a lower bound A > 0 in (3) is, however, not needed for our
theory to apply.
Discrete Deep Feature Extraction: A Theory and New Architectures
Examples of structured convolutional sets with A = B = 1
include, in the 1-D case, wavelets (Daubechies, 1992) and
Weyl-Heisenberg functions (Bölcskei & Hlawatsch, 1997),
and in the 2-D case, tensorized wavelets (Mallat, 2009),
curvelets (Candès et al., 2006), and shearlets (Kutyniok &
Labate, 2012a).
2.2. Non-linearities
The non-linearities ρ : C → C we consider are all pointwise and satisfy the Lipschitz property |ρ(x) − ρ(y)| ≤
L|x − y|, ∀x, y ∈ C, for some L > 0.
2.2.1. E XAMPLE NON - LINEARITIES
• The hyperbolic tangent non-linearity, defined as
ρ(x) = tanh(Re(x)) + i tanh(Im(x)), where
x
−x
tanh(x) = eex −e
+e−x , has Lipschitz constant L = 2.
• The rectified linear unit non-linearity is given by
ρ(x) = max{0, Re(x)} + i max{0, Im(x)}, and has
Lipschitz constant L = 2.
• The modulus non-linearity is ρ(x) = |x|, and has Lipschitz constant L = 1.
• The logistic sigmoid non-linearity is defined as
ρ(x) = sig(Re(x)) + i sig(Im(x)), where sig(x) =
1
1+e−x , and has Lipschitz constant L = 1/2.
We refer the reader to (Wiatowski & Bölcskei, 2015) for
proofs of the Lipschitz properties of these example nonlinearities.
2.3. Pooling operators
The essence of pooling is to reduce signal dimensionality
in the individual network layers and to ensure robustness of
the feature vector w.r.t. deformations and translations.
The theory developed in this paper applies to general pooling operators P : HN → HN/S , where N, S ∈ N with
N/S ∈ N, that satisfy the Lipschitz property kP f −
P gk2 ≤ Rkf − gk, ∀f, g ∈ HN , for some R > 0. The integer S will be referred to as pooling factor, and determines
the “size” of the neighborhood values are combined in.
2.3.1. E XAMPLE POOLING OPERATORS
• Sub-sampling, defined as P : HN → HN/S ,
(P f )[n] = f [Sn], n ∈ IN/S , has Lipschitz constant
R = 1. For S = 1, P is the identity operator which
amounts to “no pooling”.
• Averaging, defined as P : HN → HN/S , (P f )[n] =
PSn+S−1
αk−Sn f [k], n ∈ IN/S , has Lipschitz conk=Sn
stant R = S 1/2 maxk∈{0,...,S−1} |αk |. The weights
{αk }S−1
k=0 can be learned (LeCun et al., 1998) or prespecified (Pinto et al., 2008) (e.g., uniform pooling
corresponds to αk = S1 , for k ∈ {0, . . . , S − 1}).
• Maximization, defined as P : HN → HN/S ,
(P f )[n] = maxk∈{Sn,...,Sn+S−1} |f [k]|, n ∈ IN/S ,
has Lipschitz constant R = 1.
We refer to Appendix B in the Supplement for proofs of the
Lipschitz property of these three example pooling operators
along with the derivations of the corresponding Lipschitz
constants.
3. The network architecture
The architecture we consider is flexible in the following
sense. In each layer, we can feed into the feature vector
either the signals propagated down to that layer (i.e., the
feature maps), filtered versions thereof, or we can decide
not to have that layer contribute to the feature vector.
The basic building blocks of our network are the triplets
(Ψd , ρd , Pd ) of filters, non-linearities, and pooling operators associated with the d-th network layer and referred to
as modules. We emphasize that these triplets are allowed to
be different across layers.
Definition 2. For network layers d, 1 ≤ d ≤ D, let Ψd =
{gλd }λd ∈Λd ⊆ HNd be a convolutional set, ρd : C → C
a point-wise Lipschitz-continuous non-linearity, and Pd :
HNd → HNd+1 a Lipschitz-continuous pooling operator
d
with Nd+1 = N
Sd , where Sd ∈ N denotes the pooling factor
in the d-th layer. Then, the sequence of triplets
Ω := (Ψd , ρd , Pd )
1≤d≤D
is called a module-sequence.
Note that the dimensions of the spaces HNd satisfy
N1 ≥ N2 ≥ . . . ≥ ND . Associated with the module
(Ψd , ρd , Pd ), we define the operator
(Ud [λd ]f ) := Pd (ρd (f ∗ gλd ))
(4)
and extend it to paths on index sets
q = (λ1 , λ2 , . . . , λd ) ∈ Λ1 × Λ2 × · · · × Λd := Λd1 ,
for 1 ≤ d ≤ D, according to
U [q]f = U [(λ1 , λ2 , . . . , λd )]f
:= Ud [λd ] · · · U2 [λ2 ]U1 [λ1 ]f.
(5)
For the empty path e := ∅ we set Λ01 := {e} and let
U [e]f := f , for all f ∈ HN1 .
The network output in the d-th layer is given by (U [q]f ) ∗
χd , q ∈ Λd1 , where χd ∈ HNd+1 is referred to as outputgenerating atom. Specifically, we let χd be (i) the delta
Discrete Deep Feature Extraction: A Theory and New Architectures
(j) (l) (m)
U λ1 , λ2 , λ3
f
(p) (r) (s)
U λ1 , λ2 , λ3
f
(j) (l)
U λ1 , λ2 f
(p) (r)
U λ1 , λ2
f
(j) (l)
U λ1 , λ2 f ∗ χ2
(p) (r)
U λ1 , λ2
f ∗ χ2
(j)
U λ1 f
(j)
U λ1 f ∗ χ1
(p)
U λ1 f
U [e]f = f
(p)
U λ1 f ∗ χ1
f ∗ χ0
Figure 1. Network architecture underlying the feature extractor
(k)
(6). The index λd corresponds to the k-th atom gλ(k) of the
d
convolutional set Ψd associated with the d-th network layer. The
function χd is the output-generating atom of the d-th layer. The
root of the network corresponds to d = 0.
function δ[n], n ∈ INd+1 , if we want the output to equal
the unfiltered features U [q]f , q ∈ Λd1 , propagated down
to layer d, or (ii) any other signal of length Nd+1 , or (iii)
χd = 0 if we do not want layer d to contribute to the feature
vector. From now on we formally add χd to the set Ψd+1 =
{gλd+1 }λd+1 ∈Λd+1 , noting that {gλd+1 }λd+1 ∈Λd+1 ∪ {χd }
0
forms a convolutional set Ψ0d+1 with Bessel bound Bd+1
≤
2
Bd+1 + maxk∈INd+1 |b
χd [k]| . We emphasize that the
atoms of the augmented set {gλd+1 }λd+1 ∈Λd+1 ∪ {χd }
are employed across two consecutive layers in the sense
of χd generating the output in the d-th layer according
to (U [q]f ) ∗ χd , q ∈ Λd1 , and the remaining atoms
{gλd+1 }λd+1 ∈Λd+1 propagating the signals U [q]f , q ∈ Λd1 ,
from the d-th layer down to the (d + 1)-st layer according
to (4), see Fig. 1. With slight abuse of notation, we shall
henceforth write Ψd for Ψ0d and Bd for Bd0 as well.
We are now ready to define the feature extractor ΦΩ based
on the module-sequence Ω.
Definition 3. Let Ω = (Ψd , ρd , Pd ) 1≤d≤D be a modulesequence. The feature extractor ΦΩ based on Ω maps f ∈
HN1 to its features
D−1
[
ΦΩ (f ) :=
ΦdΩ (f ),
(6)
d=0
where ΦdΩ (f ) := {(U [q]f ) ∗ χd }q∈Λd1 is the collection of
features generated in the d-th network layer (see Fig. 1).
The dimension of the feature vector ΦΩ (f ) is given by
PD−1
Qd
ε0 N1 + d=1 εd Nd+1
k=1 card(Λk ) , where εd = 1,
if an output is generated (either filtered or unfiltered) in the
d
d-th network layer, and εd = 0, else. As Nd+1 = N
Sd =
1
· · · = S1N···S
, for d ≥ 1, the dimension of the overall fead
ture vector is determined by the pooling factors Sk and, of
course, the layers that contribute to the feature vector.
Remark 1. It was argued in (Bruna & Mallat, 2013;
Andén & Mallat, 2014; Oyallon & Mallat, 2014) that the
features Φ1Ω (f ) when generated by wavelet filters, modulus non-linearities, without intra-layer pooling, and by employing output-generating atoms with low-pass characteristics, describe mel frequency cepstral coefficients (Davis
& Mermelstein, 1980) in 1-D, and SIFT-descriptors (Lowe,
2004; Tola et al., 2010) in 2-D.
4. Sampled cartoon functions
While our main results hold for general signals f , we can
provide a refined analysis for the class of sampled cartoon
functions. This allows to understand how certain structural properties of the input signal, such as the presence
of sharp edges, are reflected in the feature vector. Cartoon
functions—as introduced in continuous time in (Donoho,
2001)—are piecewise “smooth” apart from curved discontinuities along Lipschitz-continuous hypersurfaces. They
hence provide a good model for natural images (see Fig. 2,
left) such as those in the Caltech-256 (Griffin et al., 2007)
and the CIFAR-100 (Krizhevsky, 2009) datasets, for images of handwritten digits (LeCun & Cortes, 1998) (see Fig.
2, middle), and for images of geometric objects of different
shapes, sizes, and colors as in the Baby AI School dataset2 .
Bounds on deformation sensitivity for cartoon functions in
continuous-time DCNNs were recently reported in (Grohs
et al., 2016). Here, we analyze deformation sensitivity
for sampled cartoon functions passed through discrete DCNNs.
Definition 4. The function c : R → C is referred to as a
cartoon function if it can be written as c = c1 + 1[a,b] c2 ,
where [a, b] ⊆ [0, 1] is a closed interval, and ci : R → C,
i = 1, 2, satisfies the Lipschitz property
|ci (x) − ci (y)| ≤ C|x − y|,
∀x, y ∈ R,
for some C > 0. Furthermore, we denote by
K
CCART
:= {c1 + 1[a,b] c2 | |ci (x) − ci (y)| ≤ K|x − y|,
∀ x, y ∈ R, i = 1, 2, kc2 k∞ ≤ K}
the class of cartoon functions of variation K > 0, and by
n
N,K
CCART
:= f [n] = c(n/N ), n ∈ {0, 1, . . . , N − 1}
K
c = (c1 + 1[a,b] c2 ) ∈ CCART
with
n 1
N − 1 oo
a, b ∈
/ 0, , . . . ,
N
N
the class of sampled cartoon functions of length N and
variation K > 0.
We note that excluding the boundary points a, b of the interval [a, b] from being sampling points n/N in the def2
http://www.iro.umontreal.ca/%7Elisa/
twiki/bin/view.cgi/Public/BabyAISchool
Discrete Deep Feature Extraction: A Theory and New Architectures
tisfy ρd (0) = 0 and Pd (0) = 0 (as all non-linearities
and pooling operators in Sections 2.2.1 and 2.3.1,
apart from the logistic sigmoid non-linearity, do), then
|||ΦΩ (f )||| ≤ kf k2 ,
∀f ∈ HN1 .
(10)
iii) For every variation K > 0 and deformation Fτ of the
form
Figure 2. Left: A natural image (image credit: (Kutyniok & Labate, 2012b)) is typically governed by areas of little variation,
with the individual areas separated by edges that can be modeled
as curved singularities. Middle: Image of a handwritten digit.
Right: Pixel values corresponding to the dashed row in the middle image.
N,K
inition of CCART
is of conceptual importance (see Remark D.1 in the Supplement). Moreover, our results can
N,K
easily be generalized to classes CCART
consisting of functions f [n] = c(n/N ) with c containing multiple “1-D
edges” (i.e.,
PLmultiple discontinuity points) according to
c = c1 + l=1 1[al ,bl ] c2 with ∩L
l=1 [al , bl ] = ∅. We also
N,K
note that CCART reduces to the class of sampled Lipschitzcontinuous functions upon setting c2 = 0.
A sampled cartoon function in 2-D models, e.g., an image
acquired by a digital camera (see Fig. 2, middle); in 1N,K
can be thought of as the pixels in a row
D, f ∈ CCART
or column of this image (see Fig. 2 right, which shows a
cartoon function with 6 discontinuity points).
5. Analytical results
We analyze global and local feature vector properties with
globality pertaining to characteristics brought out by the
union of features across all network layers, and locality
identifying attributes made explicit in individual layers.
5.1. Global properties
Theorem 1. Let Ω = (Ψd , ρd , Pd ) 1≤d≤D be a modulesequence. Assume that the Bessel bounds Bd > 0, the Lipschitz constants Ld > 0 of the non-linearities ρd , and the
Lipschitz constants Rd > 0 of the pooling operators Pd
satisfy
max max{Bd , Bd Rd2 L2d } ≤ 1.
(7)
1≤d≤D
i) The feature extractor ΦΩ is Lipschitz-continuous with
Lipschitz constant LΩ = 1, i.e.,
|||ΦΩ (f ) − ΦΩ (h)||| ≤ kf − hk2 ,
(8)
for all f, h ∈ HN1 , where the feature space norm is
defined as
D−1
X X
|||ΦΩ (f )|||2 :=
||(U [q]f ) ∗ χd ||22 . (9)
d=0 q∈Λd
1
ii) If, in addition to (7), for all d ∈ {1, . . . , D − 1} the
non-linearities ρd and the pooling operators Pd sa-
(Fτ f )[n] : = c(n/N1 − τ (n/N1 )),
n ∈ IN1 , (11)
where τ : R → [−1, 1], the deformation sensitivity is
bounded according to
1/2
|||ΦΩ (Fτ f ) − ΦΩ (f )||| ≤ 4KN1 kτ k1/2
∞ ,
(12)
N1 ,K
for all f ∈ CCART
.
Proof. See Appendix C in the Supplement.
The Lipschitz continuity (8) guarantees that pairwise distances of input signals do not increase through feature extraction. As an immediate implication of the Lipschitz continuity we get robustness of the feature extractor w.r.t. additive bounded noise η ∈ HN1 in the sense of
|||ΦΩ (f + η) − ΦΩ (f )||| ≤ kηk2 ,
for all f ∈ HN1 .
Remark 2. As detailed in the proof of Theorem 1, the Lipschitz continuity (8) combined with the deformation sensitivity bound (see Proposition D.1 in the Supplement) for the
signal class under consideration, namely sampled cartoon
functions, establishes the deformation sensitivity bound
(12) for the feature extractor. This insight has important
practical ramifications as it shows that whenever we have
deformation sensitivity bounds for a signal class, we automatically get deformation sensitivity guarantees for the
corresponding feature extractor.
From (12) we can deduce a statement on the sensitivity of
ΦΩ w.r.t. translations on R. To this end, we first note that
setting τt (x) = t, x ∈ R, for t ∈ [−1, 1], (11) becomes
(Fτt f )[n] = c(n/N1 − t),
n ∈ IN1 .
Particularizing (12) accordingly, we obtain
1/2
|||ΦΩ (Fτt f ) − ΦΩ (f )||| ≤ 4KN1 |t|1/2 ,
(13)
which shows that small translations |t| of the underlying
analog signal c(x), x ∈ R, lead to small changes in the feature vector obtained by passing the resulting sampled signal
through a discrete DCNN. We shall say that (13) is a translation sensitivity bound. Analyzing the impact of deformations and translations over R on the discrete feature vector generated by the sampled analog signal closely models
real-world phenomena (e.g., the jittered acquisition of an
analog signal with a digital camera, where different values
of N1 in (11) correspond to different camera resolutions).
Discrete Deep Feature Extraction: A Theory and New Architectures
We note that, while iii) in Theorem 1 is specific to cartoon
functions, i) and ii) apply to all signals in HN1 .
The strength of the results in Theorem 1 derives itself
from the fact that condition (7) on the underlying modulesequence Ω is easily met in practice. To see this, we first
note that Bd is determined by the convolutional set Ψd , Ld
by the non-linearity ρd , and Rd by the pooling operator Pd .
Condition (7) is met if
Bd ≤ min{1, Rd−2 L−2
d },
∀ d ∈ {1, 2, . . . , D},
(14)
which, if not satisfied by default, can be enforced simply by normalizing the elements in Ψd . Specifically, for
e d := {C −1/2 gλ }λ ∈Λ
Cd := max{Bd , Rd2 L2d } the set Ψ
d
d
d
d
fd = Bd and hence satisfies (14). While
has Bessel bound B
Cd
this normalization does not have an impact on the results in
Theorem 1, there exists, however, a tradeoff between energy preservation and deformation (respectively translation)
sensitivity in ΦdΩ as detailed in the next section.
5.2. Local properties
Theorem 2. Let Ω = (Ψd , ρd , Pd ) 1≤d≤D be a modulesequence with corresponding Bessel bounds Bd > 0, Lipschitz constants Ld > 0 of the non-linearities ρd , Lipschitz
constants Rd > 0 of the pooling operators Pd , and outputgenerating atoms χd . Let further L0Ω := kχ0 k1 and 3
d
Y
1/2
LdΩ := kχd k1
Bk L2k Rk2
,
d ≥ 1.
(15)
k=1
i) The features generated in the d-th network layer are
Lipschitz-continuous with Lipschitz constant LdΩ , i.e.,
|||ΦdΩ (f ) − ΦdΩ (h)||| ≤ LdΩ kf − hk2 ,
(16)
2
d
:=
for
P all f, h ∈ HN21 , where |||ΦΩ (f )|||
||(U
[q]f
)
∗
χ
||
.
d 2
q∈Λd
1
ii) If the non-linearities ρk and the pooling operators Pk
satisfy ρk (0) = 0 and Pk (0) = 0, respectively, for all
k ∈ {1, . . . , d}, then
|||ΦdΩ (f )||| ≤ LdΩ kf k2 ,
∀f ∈ HN1 .
(17)
iii) For all K > 0 and all τ : R → [−1, 1], the features
generated in the d-th network layer satisfy
|||ΦdΩ (Fτ f )
−
ΦdΩ (f )|||
≤
4LdΩ KN 1/2 kτ k1/2
∞ ,
(18)
N1 ,K
for all f ∈ CCART
, where Fτ f is defined in (11).
iv) If the module-sequence employs sub-sampling, average pooling, or max-pooling with corresponding
pooling factors Sd ∈ N, then
ΦdΩ (Tm f ) = T S
m
1 ...Sd
ΦdΩ (f ),
(19)
3
We note that kχd k1 in (15) can be upper-bounded (and hence
substituted) by Bd+1 , see Remark E.1 in the Supplement.
m
for all f ∈ HN1 and all m ∈ Z with S1 ...S
∈ Z.
d
d
Here, Tm ΦΩ (f ) refers to element-wise application of
Tm , i.e., Tm ΦdΩ (f ) := {Tm h | ∀h ∈ ΦdΩ (f )}.
Proof. See Appendix E in the Supplement.
One may be tempted to infer the global results (8), (10),
and (12) in Theorem 1 from the corresponding local results
in Theorem 2, e.g., the energy bound in (10) from (17)
P
1/2
D−1
d
2
according to |||ΦΩ (f )||| =
≤
d=0 |||ΦΩ (f )|||
√
d
Dkf k2 , where we employed LΩ ≤ 1 owing to (7). This
would, however, lead to the “global” Lipschitz constant
√
LΩ = 1 in (8), (10), and (12) to be replaced by LΩ = D
and thereby render the corresponding results much weaker.
Again, we emphasize that, while iii) in Theorem 2 is specific to cartoon functions, i), ii), and iv) apply to all signals
in HN1 .
For a fixed network layer d, the “local” Lipschitz constant
LdΩ determines the noise sensitivity of the features ΦdΩ (f )
according to
|||ΦdΩ (f + η) − ΦdΩ (f )||| ≤ LdΩ kηk2 ,
(20)
where (20) follows from (16). Moreover, LdΩ via (18) also
quantifies the impact of deformations (or translations when
τt (x) = t, x ∈ R, for t ∈ [−1, 1]) on the feature vector. In practice, it may be desirable to have the features ΦdΩ
become more robust to additive noise and less deformationsensitive (respectively, translation-sensitive) as we progress
deeper into the network. Formally, this vertical sensitivity
reduction can be induced by ensuring that Ld+1
< LdΩ .
Ω
kχ k B
1/2
L R
d 1 d
d d
Thanks to LdΩ =
Ld−1
Ω , this can be ackχd−1 k1
complished by choosing the module-sequence such that
1/2
kχd k1 Bd Ld Rd < kχd−1 k1 . Note, however, that owing to (17) this will also reduce the signal energy contained in the features ΦdΩ (f ). We therefore have a tradeoff
between deformation (respectively translation) sensitivity
and energy preservation. Having control over this tradeoff
through the choice of the module-sequence Ω may come in
handy in practice.
For average pooling with uniform weights αkd = S1d ,
k = 0, . . . , Sd − 1 (noting that the corresponding Lip−1/2
schitz constant is Rd = Sd
, see Section 2.3.1), we
Q
1/2
Bk L2k
d
d
, which illustrates that
get LΩ = kχd k1
k=1 Sk
pooling can have an impact on the sensitivity and energy
properties of ΦdΩ .
We finally turn to interpreting the translation covariance rem
sult (19). Owing to the condition S1 ...S
∈ Z, we get transd
lation covariance only on the rough grid induced by the
product of the pooling factors. In the absence of pooling,
Discrete Deep Feature Extraction: A Theory and New Architectures
i.e., Sk = 1, for k ∈ {1, . . . , d}, we obtain translation covariance w.r.t. the fine grid the input signal f ∈ HN1 lives
on.
Remark 3. We note that ScatNets (Bruna & Mallat, 2013)
are translation-covariant on the rough grid induced by the
factor 2J corresponding to the coarsest wavelet scale. Our
result in (19) is hence in the spirit of (Bruna & Mallat,
2013) with the difference that the grid in our case is induced by the pooling factors Sk .
6. Experiments4
We consider the problem of handwritten digit classification
and evaluate the performance of the feature extractor ΦΩ in
combination with a support vector machine (SVM). The results we obtain are competitive with the state-of-the-art in
the literature. The second line of experiments we perform
assesses the importance of the features extracted by ΦΩ in
facial landmark detection and in handwritten digit classification, using random forests (RF) for regression and classification, respectively. Our results are based on a DCNN
with different non-linearities and pooling operators, and
with tensorized (i.e., separable) wavelets as filters, sensitive
to 3 directions (horizontal, vertical, and diagonal). Furthermore, we generate outputs in all layers through low-pass
filtering. Circular convolutions with the 1-D filters underlying the tensorized wavelets are efficiently implemented
using the algorithme à trous (Holschneider et al., 1989).
To reduce the dimension of the feature vector, we compute features along frequency decreasing paths only (Bruna
& Mallat, 2013), i.e., for every node U [q]f , q ∈
Λd−1
, we retain only those child nodes Ud [λd ]U [q]f =
1
Pd ρd ((U [q]f )∗gλd ) that correspond to wavelets gλd with
scales larger than the maximum scale of the wavelets used
to get U [q]f . We refer to (Bruna & Mallat, 2013) for a detailed justification of this approach for scattering networks.
6.1. Handwritten digit classification
We use the MNIST dataset of handwritten digits (LeCun &
Cortes, 1998) which comprises 60,000 training and 10,000
test images of size 28×28. We set D = 3, and compare different network configurations, each defined by a single module (i.e., we use the same filters, non-linearity, and pooling operator in all layers). Specifically, we consider Haar
wavelets and reverse biorthogonal 2.2 (RBIO2.2) wavelets
(Mallat, 2009), both with J = 3 scales, the non-linearities
described in Section 2.2.1, and the pooling operators described in Section 2.3.1 (with S1 = 1 and S2 = 2). We
use a SVM with radial basis function (RBF) kernel for
classification. To reduce the dimension of the feature vec4
Code available at http://www.nari.ee.ethz.ch/
commth/research/
Haar
RBIO2.2
abs ReLU tanh LogSig abs ReLU tanh LogSig
n.p. 0.55 0.57 1.41 1.49
0.50 0.54 1.01 1.18
0.59 0.62 1.04 1.13
sub. 0.60 0.58 1.25 1.45
max. 0.61 0.60 0.68 0.76
0.55 0.56 0.71 0.75
avg. 0.57 0.58 1.26 1.44
0.51 0.60 1.04 1.18
Table 1. Classification error in percent for handwritten digit classification using different configurations of wavelet filters, nonlinearities, and pooling operators (sub.: sub-sampling; max.:
max-pooling; avg.: average-pooling; n.p.: no pooling).
tors from 18,424 (or 50,176, for the configurations without
pooling) down to 1000, we employ the supervised orthogonal least squares feature selection procedure described in
(Oyallon & Mallat, 2014). The penalty parameter of the
SVM and the localization parameter of the RBF kernel are
selected via 10-fold cross-validation for each combination
of wavelet filter, non-linearity, and pooling operator.
Table 1 shows the resulting classification errors on the test
set (obtained for the SVM trained on the full training set).
Configurations employing RBIO2.2 wavelets tend to yield
a marginally lower classification error than those using
Haar wavelets. For the tanh and LogSig non-linearities,
max-pooling leads to a considerably lower classification
error than other pooling operators. The configurations
involving the modulus and ReLU non-linearities achieve
classification accuracy competitive with the state-of-theart (Bruna & Mallat, 2013) (class. err.: 0.43%), which
is based on directional non-separable wavelets with 6 directions without intra-layer pooling. This is interesting as
the separable wavelet filters employed here can be implemented more efficiently.
6.2. Feature importance evaluation
In this experiment, we investigate the “importance” of the
features generated by ΦΩ corresponding to different layers, wavelet scales, and directions in two different learning
tasks, namely, facial landmark detection and handwritten
digit classification. The primary goal of this experiment
is to illustrate the practical relevance of the notion of local properties of ΦΩ as established in Section 5.2. For facial landmark detection we employ a RF regressor and for
handwritten digit classification a RF classifier (Breiman,
2001). In both cases, we fix the number of trees to 30
and select the tree depth using out-of-bag error estimates
(noting that increasing the number of trees does not significantly increase the accuracy). The impurity measure
used for learning the node tests is the mean square error for facial landmark detection and the Gini impurity for
handwritten digit classification. In both cases, feature importance is assessed using the Gini importance (Breiman
et al., 1984), averaged over all trees. The Gini importance
I(θ, T ) of feature θ in the (trained) tree T is defined as
Discrete Deep Feature Extraction: A Theory and New Architectures
P
n
n
`
I(θ, T ) = `∈T : ϕ(`)=θ nntot
(ı̂` − n``L ı̂`L − n``R ı̂`R ), where
ϕ(`) denotes the feature determined in the training phase
for the test at node `, n` is the number
of training samples
P
passed through node `, ntot =
`∈T n` , ı̂` is the impurity at node `, and `L and `R denote the left and right child
node, respectively, of node `. For the feature extractor ΦΩ
we set D = 4, employ Haar wavelets with J = 3 scales
and the modulus non-linearity in every network layer, no
pooling in the first layer and average pooling with uniform
weights 1/Sd2 , Sd = 2, in layers d = 2, 3.
Facial landmark detection. We use the Caltech 10,000
Web Faces data base (Angelova et al., 2005). Each of the
7092 images in the data base depicts one or more faces in
different contexts (e.g., portrait images, groups of people).
The data base contains annotations of the positions of eyes,
nose, and mouth for at least one face per image. The learning task is to estimate the positions of these facial landmarks. The annotations serve as ground truth for training
and testing. We preprocess the data set as follows. The
patches containing the faces are extracted from the images
using the Viola-Jones face detector (Viola & Jones, 2004).
After discarding false positives, the patches are converted
to grayscale and resampled to size 120 × 120 (using linear interpolation), before feeding them to the feature extractor ΦΩ . This procedure yields a dataset containing a
total of 8776 face images. We select 80% of the images
uniformly at random to form a training set and use the
remaining images for testing. We train a separate RF for
each facial landmark. Following (Dantone et al., 2012) we
report the localization error, i.e., the `2 -distance between
the estimated and the ground truth landmark positions, on
the test set as a fraction of the (true) inter-ocular distance.
The errors obtained are: left eye: 0.062; right eye: 0.064;
nose; 0.080, mouth: 0.095. As an aside, we note that these
values are comparable with the ones reported in (Dantone
et al., 2012) for a conditional RF using patch comparison
features (evaluated on a different dataset and a larger set of
facial landmarks).
Handwritten digit classification. For this experiment,
we again rely on the MNIST dataset. The training set is
obtained by sampling uniformly at random 1, 000 images
per digit from the MNIST training dataset and we use the
complete MNIST test set. We train two RFs, one based on
unmodified images, and the other one based on images subject to a random uniform displacement of at most 4 pixels in
(positive and negative) x and y direction to study the impact
of offsets on feature importance. The resulting RFs achieve
a classification error of 4.2% and 9.6%, respectively.
Discussion. Figure 3 shows the cumulative feature importance (per triplet of layer index, wavelet scale, and direction, averaged over all trees in the respective RF) in
handwritten digit classification and in facial landmark detection. Table 2 shows the corresponding cumulative fea-
left eye
nose
j =0
j =1
0.2
j =0
j =1
0.2
j =2
0.1
j =2
0.1
0.0
0.0
0/- 1/0 1/1 1/2 2/0 2/1 2/2 3/0 3/1 3/2
0/- 1/0 1/1 1/2 2/0 2/1 2/2 3/0 3/1 3/2
digits
displaced digits
j =0
j =1
0.2
j =0
0.2
j =2
j =1
j =2
0.1
0.1
0.0
0.0
0/- 1/0 1/1 1/2 2/0 2/1 2/2 3/0 3/1 3/2
0/- 1/0 1/1 1/2 2/0 2/1 2/2 3/0 3/1 3/2
Figure 3. Average cumulative feature importance and standard
error for facial landmark detection and handwritten digit classification. The labels on the horizontal axis indicate layer index
d/wavelet direction (0: horizontal, 1: vertical, 2: diagonal).
left eye right eye nose mouth digits disp. digits
Layer 0
Layer 1
Layer 2
Layer 3
0.020
0.629
0.261
0.090
0.023
0.646
0.236
0.095
0.016
0.576
0.298
0.110
0.014
0.490
0.388
0.108
0.046
0.426
0.337
0.192
0.004
0.094
0.280
0.622
Table 2. Cumulative feature importance per layer. Columns 1–4:
facial landmark detection. Columns 5 and 6: handwritten digit
classification.
ture importance for each layer.
For facial landmark detection, the features in layer 1 clearly
have the highest importance, and the feature importance
decreases with increasing layer index d. For handwritten
digit classification using the unshifted MNIST images, the
cumulative importance of the features in the second/third
layer relative to those in the first layer is considerably
higher than in facial landmark detection (see Table 2). For
the translated MNIST images, the importance of the features in the second/third layer is significantly higher than
those in the 0-th and in the first layer. An explanation for
this observation could be as follows: In a classification task
small sensitivity to translations is beneficial. Now, according to our theory (see Section 5.2) translation sensitivity,
indeed, decreases with increasing layer index for average
pooling as used here. For localization of landmarks, on the
other hand, the RF needs features that are covariant on the
fine grid of the input image thus favoring features in the
layers closer to the root.
Discrete Deep Feature Extraction: A Theory and New Architectures
Acknowledgments
The authors would like to thank C. Geiger for preliminary
work on the experiments in Section 6.2 and M. Lerjen for
help with computational issues.
References
Andén, J. and Mallat, S. Deep scattering spectrum. IEEE
Trans. Sig. Process., 62(16):4114–4128, 2014.
Angelova, A., Abu-Mostafa, Y., and Perona, P. Pruning
training sets for learning of object categories. In Proc.
of IEEE Conf. Comp. Vision Pattern Recog. (CVPR), pp.
494–501, 2005.
Bengio, Y., Courville, A., and Vincent, P. Representation
learning: A review and new perspectives. IEEE Trans.
Pattern Anal. Mach. Intell., 35(8):1798–1828, 2013.
Bölcskei, H. and Hlawatsch, F. Discrete Zak transforms,
polyphase transforms, and applications. IEEE Trans.
Sig. Process., 45(4):851–866, 1997.
Breiman, L. Random forests. Machine learning, 45(1):
5–32, 2001.
Breiman, L., Friedman, J., Stone, C. J., and Olshen, R. A.
Classification and regression trees. CRC Press, 1984.
Bruna, J. and Mallat, S. Invariant scattering convolution
networks. IEEE Trans. Pattern Anal. Mach. Intell., 35
(8):1872–1886, 2013.
Candès, E. J., Demanet, L., Donoho, D., and Ying, L. Fast
discrete curvelet transforms. Multiscale Modeling and
Simulation, 5(3):861–899, 2006.
Dantone, M., Gall, J., Fanelli, G., and Van Gool, L. Realtime facial feature detection using conditional regression
forests. In Proc. of IEEE Conf. Comp. Vision Pattern
Recog. (CVPR), pp. 2578–2585, 2012.
Daubechies, I. Ten lectures on wavelets. Society for Industrial and Applied Mathematics, 1992.
Davis, S. and Mermelstein, P. Comparison of parametric representations for monosyllabic word recognition
in continuously spoken sentences. IEEE Trans. Acoust.,
Speech, and Signal Process., 28(4):357–366, 1980.
Donoho, D. Sparse components of images and optimal
atomic decompositions. Constructive Approximation, 17
(3):353–382, 2001.
Folland, G. B. A course in abstract harmonic analysis,
volume 29. CRC Press, 2015.
Golub, G. H. and Van Loan, C. F. Matrix computations.
Johns Hopkins University Press, 2013.
Griffin, G., Holub, A., and Perona, P. Caltech-256 object category dataset. http://authors.library.
caltech.edu/7694/, 2007.
Grohs, P., Wiatowski, T., and Bölcskei, H. Deep convolutional neural networks on cartoon functions. In Proc.
of IEEE Int. Symp. on Inform. Theory (ISIT), to appear.
2016.
Holschneider, M., Kronland-Martinet, R., Morlet, J., and
Tchamitchian, P. A real-time algorithm for signal analysis with the help of the wavelet transform. In Wavelets,
pp. 286–297. Springer, 1989.
Huang, F. J. and LeCun, Y. Large-scale learning with
SVM and convolutional nets for generic object categorization. In Proc. of IEEE Conf. Comp. Vision Pattern
Recog. (CVPR), pp. 284–291, 2006.
Jarrett, K., Kavukcuoglu, K., Ranzato, M. A., and LeCun,
Y. What is the best multi-stage architecture for object
recognition? In Proc. of IEEE Int. Conf. on Computer
Vision (ICCV), pp. 2146–2153, 2009.
Krizhevsky, A. Learning multiple layers of features from
tiny images. MS thesis, University of Toronto, 2009.
Kutyniok, G. and Labate, D. (eds.). Shearlets: Multiscale
analysis for multivariate data. Birkhäuser, 2012a.
Kutyniok, G. and Labate, D. Introduction to shearlets. In
Shearlets: Multiscale analysis for multivariate data, pp.
1–38. Birkhäuser, 2012b.
LeCun, Y. and Cortes, C. The MNIST database of handwritten digits. http://yann.lecun.com/exdb/
mnist/, 1998.
LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. Gradientbased learning applied to document recognition. In Proc.
of the IEEE, pp. 2278–2324, 1998.
LeCun, Y., Bengio, Y., and Hinton, G. Deep learning. Nature, 521:436–444, 2015.
Lowe, D. G. Distinctive image features from scaleinvariant keypoints. International Journal of Computer
Vision, 60(2):91–110, 2004.
Mallat, S. A wavelet tour of signal processing: The sparse
way. Academic Press, 3rd edition, 2009.
Mallat, S. Group invariant scattering. Comm. Pure Appl.
Math., 65(10):1331–1398, 2012.
Mutch, J. and Lowe, D. G. Multiclass object recognition
with sparse, localized features. In Proc. of IEEE Conf.
Comp. Vision Pattern Recog. (CVPR), pp. 11–18, 2006.
Discrete Deep Feature Extraction: A Theory and New Architectures
Oyallon, E. and Mallat, S. Deep roto-translation scattering
for object classification. arXiv:1412.8659, 2014.
Pinto, N., Cox, D. D., and DiCarlo, J. J. Why is real-world
visual object recognition hard. PLoS Computational Biology, 4(1):151–156, 2008.
Ranzato, M., Poultney, C., Chopra, S., and LeCun, Y. Efficient learning of sparse representations with an energybased model. In Proc. of Int. Conf. on Neural Information Processing Systems (NIPS), pp. 1137–1144, 2006.
Ranzato, M. A., Huang, F. J., Boureau, Y. L., and LeCun,
Y. Unsupervised learning of invariant feature hierarchies with applications to object recognition. In Proc.
of IEEE Conf. Comp. Vision Pattern Recog. (CVPR), pp.
1–8, 2007.
Serre, T., Wolf, L., and Poggio, T. Object recognition with
features inspired by visual cortex. In Proc. of IEEE
Conf. Comp. Vision Pattern Recog. (CVPR), pp. 994–
1000, 2005.
Tola, E., Lepetit, V., and Fua, P. Daisy: An efficient dense
descriptor applied to wide-baseline stereo. IEEE Trans.
Pattern Anal. Mach. Intell., 5(815–830), 2010.
Viola, P. and Jones, M. J. Robust real-time face detection. International Journal of Computer Vision, 57(2):
137–154, 2004.
Wiatowski, T. and Bölcskei, H. A mathematical theory of
deep convolutional neural networks for feature extraction. arXiv:1512.06293, 2015.
Discrete Deep Feature Extraction: A Theory and New Architectures
A. Appendix: Additional numerical results
A.1. Handwritten digit classification
For the handwritten digit classification experiment described in Section 6.1, Table 3 shows the classification error
for Daubechies wavelets with 2 vanishing moments (DB2).
DB2
abs ReLU tanh LogSig
n.p. 0.54 0.51 1.29 1.40
sub. 0.60 0.58 1.16 1.34
max. 0.57 0.57 0.75 0.67
avg. 0.52 0.61 1.16 1.27
right eye
j =0
0.25
j =1
j =2
Table 3. Classification errors in percent for handwritten digit
classification using DB2 wavelet filters, different non-linearities,
and different pooling operators (sub.: sub-sampling; max.: maxpooling; avg.: average-pooling; n.p.: no pooling).
0.20
0.15
0.10
A.2. Feature importance evaluation
0.05
For the feature importance experiment described in Section
6.2, Figure 4 shows the cumulative feature importance (per
triplet of layer index, wavelet scale, and direction, averaged
over all trees in the respective RF) in facial landmark detection (right eye and mouth).
0.00
0/-
1/0
1/1
1/2
2/0
2/1
2/2
3/0
3/1
3/2
mouth
B. Appendix: Lipschitz continuity of pooling
operators
j =0
0.25
j =1
j =2
We verify the Lipschitz property
0.20
kP (f ) − P (h)k2 ≤ Rkf − hk2 ,
∀f, h ∈ HN ,
0.15
for the pooling operators in Section 2.3.1.
0.10
Sub-sampling: Pooling by sub-sampling is defined as
0.05
P : HN → HN/S ,
P (f )[n] = f [Sn],
n ∈ IN/S ,
where N/S ∈ N. Lipschitz continuity with R = 1 follows
from
X
kP (f ) − P (h)k22 =
|f [Sn] − h[Sn]|2
n∈IN/S
≤
X
|f [n] − h[n]|2 = kf − hk22 ,
∀f, h ∈ HN .
n∈IN
Averaging: Pooling by averaging is defined as
P : HN → HN/S ,
P (f )[n] =
Sn+S−1
X
αk−Sn f [k],
k=Sn
for n ∈ IN/S , where N/S ∈ N. We start by setting α0 :=
0.00
0/-
1/0
1/1
1/2
2/0
2/1
2/2
3/0
3/1
3/2
Figure 4. Average cumulative feature importance and standard error for facial landmark detection. The labels on the horizontal axis
indicate layer index d/wavelet direction (0: horizontal, 1: vertical,
2: diagonal).
Discrete Deep Feature Extraction: A Theory and New Architectures
maxk∈{0,...,S−1} |αk |. Then,
The key step is then to show that ad can be upper-bounded
according to
kP (f ) − P (h)k22
=
≤
Sn+S−1
X
X
n∈IN/S
k=Sn
X
Sn+S−1
X
n∈IN/S
k=Sn
≤α S
n∈IN/S
X
= α02 S
ad ≤ bd − bd+1 ,
P
with bd :=
note that
2
0
α |f [k] − h[k]|
Sn+S−1
X
X
02
2
αk−Sn (f [k] − h[k])
D−1
X
2
f [k] − h[k]
(B.1)
q∈Λd
1
ad ≤
d=0
k=Sn
2
f [k] − h[k]
=
(C.1)
kfq − hq k22 , for d = 0, . . . , D, and to
D−1
X
d=0
= α02 Skf − hk22 ,
d = 0, . . . , D − 1,
(bd − bd+1 ) = b0 − bD ≤ b0
|{z}
X
≥0
kfq −
hq k22
= kf − hk22 ,
q∈Λ01
n∈IN
P
where we used k∈IS |f [k]−h[k]| ≤ S 1/2 kf −hk2 , f, h ∈
HS , to get (B.1), see, e.g., (Golub & Van Loan, 2013).
which then yields (8). Writing out (C.1), it follows that we
need to establish
Maximization: Pooling by maximization is defined as
X
P : HN → HN/S ,
P (f )[n] =
|f [k]|,
max
q∈Λd
1
k∈{Sn,...,Sn+S−1}
≤
n∈IN/S
≤
f [k] − h[k]
max
k∈{Sn,...,Sn+S−1}
X S−1
X
|f [Sn + k] − h[Sn + k]|2
2
q̃ ∈ Λd+1
= Λ1 × · · · × Λd ×Λd+1
1
{z
}
|
(B.2)
(B.3)
n∈IN/S k=0
= kf − hk22 ,
where we employed the reverse triangle inequality
kf k∞ − khk∞ ≤ kf − hk∞ , f, h ∈ HS , to get (B.2),
and in (B.3) we used kf k∞ ≤ kf k2 , f ∈ HS , see, e.g.,
(Golub & Van Loan, 2013).
C. Appendix: Proof of Theorem 1
We start by proving i). The key idea of the proof is—
similarly to the proof of Proposition 4 in (Wiatowski &
Bölcskei, 2015)—to employ telescoping series arguments.
For ease of notation, we let fq := U [q]f and hq := U [q]h,
for f, h ∈ HN1 , q ∈ Λd1 . With (9) we have
=Λd
1
of length d + 1 can be decomposed into a path q ∈ Λd1
of length d and an index λd+1 ∈ Λd+1 according to q̃ =
(q, λd+1 ). Thanks to (5) we have U [q̃] = U [(q, λd+1 )] =
Ud+1 [λd+1 ]U [q], which yields
X
|||ΦΩ (f ) − ΦΩ (h)|||2 =
X
||(fq − hq ) ∗ χd ||22 .
d=0 q∈Λd
1
|
{z
=:ad
}
kfq̃ − hq̃ k22 =
X
X
kUd+1 [λd+1 ]fq
λ
∈Λd+1
q∈Λd
1 d+1
q̃∈Λd+1
1
− Ud+1 [λd+1 ]hq k22 .
(C.3)
Substituting (C.3) into (C.2) and rearranging terms, we obtain
X
k(fq − hq ) ∗ χd k22
(C.4)
q∈Λd
1
+
D−1
X
(C.2)
2
|h[k]|
k∈{Sn,...,Sn+S−1}
X
d = 0, . . . , D − 1.
We start by examining the second sum on the right-hand
side (RHS) in (C.2). Every path
|f [k]|
k∈{Sn,...,Sn+S−1}
max
kfq − hq k22 ,
q∈Λd+1
1
kP (f ) − P (h)k22
X
=
max
−
||fq − hq k22
q∈Λd
1
X
−
for n ∈ IN/S , where N/S ∈ N. We have
n∈IN/S
X
k(fq − hq ) ∗ χd k22 ≤
X
kUd+1 [λd+1 ]fq − Ud+1 [λd+1 ]hq k22
(C.5)
λd+1 ∈Λd+1
≤
X
q∈Λd
1
||fq − hq k22 ,
d = 0, . . . , D − 1.
(C.6)
Discrete Deep Feature Extraction: A Theory and New Architectures
We next note that the sum over the index set Λd+1 inside
the brackets in (C.4)-(C.5) satisfies
X
kUd+1 [λd+1 ]fq − Ud+1 [λd+1 ]hq k22
d ∈ {1, . . . , D − 1}, the feature extractor ΦΩ satisfies
ΦΩ (0) = 0, and to employ (8) with h = 0 which yields
|||Φ(f )||| ≤ kf k,
λd+1 ∈Λd+1
X
=
kPd+1 ρd+1 (fq ∗ gλd+1 )
λd+1 ∈Λd+1
− Pd+1 ρd+1 (hq ∗ gλd+1 ) k22
X
kρd+1 (fq ∗ gλd+1 )
2
≤ Rd+1
(C.7)
λd+1 ∈Λd+1
− ρd+1 (hq ∗ gλd+1 )k22
X
2
≤ Rd+1
L2d+1
k(fq − hq ) ∗ gλd+1 k22 ,
(C.8)
=0
(C.9)
λd+1 ∈Λd+1
where we employed the Lipschitz continuity of Pd+1 in
(C.7)-(C.8) and the Lipschitz continuity of ρd+1 in (C.9).
Substituting the sum over the index set Λd+1 inside the
brackets in (C.4)-(C.5) by the upper bound (C.9) yields
X
k(fq − hq ) ∗ χd k22
q∈Λd
1
X
+
kUd+1 [λd+1 ]fq − Ud+1 [λd+1 ]hq k22
λd+1 ∈Λd+1
X
≤
2
max{1, Rd+1
L2d+1 } k(fq − hq ) ∗ χd k22 (C.10)
q∈Λd
1
X
+
k(fq − hq ) ∗ gλd+1 k22 ,
(C.11)
λd+1 ∈Λd+1
for d = 0, . . . , D − 1. As {gλd+1 }λd+1 ∈Λd+1 ∪ {χd } are
atoms of the convolutional set Ψd+1 , and fq , hq ∈ HNd+1 ,
we have
X
k(fq − hq ) ∗ χd k22 +
k(fq − hq ) ∗ gλd+1 k22
λd+1 ∈Λd+1
≤ Bd+1 kfq − hq k22 ,
which, when used in (C.10)-(C.11) yields
X
k(fq − hq ) ∗ χd k22
+
kUd+1 [λd+1 ]fq −
Ud+1 [λd+1 ]hq k22
λd+1 ∈Λd+1
≤
X
2
max{Bd+1 , Bd+1 Rd+1
L2d+1 }kfq
for λd ∈ Λd , by assumption. With the definition of U [q] in
(5) this then yields (U [q]h) = 0 for h = 0 and all q ∈ Λd1 .
ΦΩ (0) = 0 finally follows from
ΦΩ (h) =
D−1
[
d=0
U [q]h ∗ χd
{z
}
|
q∈Λd
1
= 0.
(C.13)
=0
We proceed to iii). The proof of the deformation sensitivity
bound (12) is based on two key ingredients. The first one
is the Lipschitz continuity result stated in (8). The second
ingredient, stated in Proposition D.1 in Appendix D, is an
upper bound on the deformation error kf − Fτ f k2 given
by
1/2
kf − Fτ f k2 ≤ 4KN1 kτ k1/2
(C.14)
∞ ,
N1 ,K
where f ∈ CCART
. We now show how (8) and (C.14) can
be combined to establish (12). To this end, we first apply
(8) with h := (Fτ f ) to get
|||ΦΩ (f ) − ΦΩ (Fτ f )||| ≤ kf − Fτ f k2 ,
(C.15)
N1 ,K
for f ∈ CCART
⊆ HN1 , N1 ∈ N, and K > 0, and then replace the RHS of (C.15) by the RHS of (C.14). This completes the proof of iii).
D. Appendix: Proposition D.1
q∈Λd
1
X
for f ∈ HN1 . It remains to prove that ΦΩ (h) = 0 for
h = 0. For h = 0, the operator Ud , d ∈ {1, 2, . . . , D},
defined in (4) satisfies
(Ud [λd ]h) = Pd ρd (h ∗ gλd ) ,
| {z }
=0
| {z }
=0
|
{z
}
−
hq k22 ,
q∈Λd
1
(C.12)
for d = 0, . . . , D − 1. Finally, invoking (7) in (C.12) we
get (C.4)-(C.6) and hence (C.1). This completes the proof
of i).
We continue with ii). The key step in establishing (10)
is to show that for ρd (0) = 0 and Pd (0) = 0, for
Proposition D.1. For every N ∈ N, every K > 0, and
every τ : R → [−1, 1], we have
kf − Fτ f k2 ≤ 4KN 1/2 kτ k1/2
∞ ,
(D.1)
N,K
for all f ∈ CCART
.
Remark D.1. As already mentioned at the end of Section 4,
excluding the interval boundary points a, b in the definition
N,K
of sampled cartoon functions CCART
(see Definition 4) is
necessary for technical reasons. Specifically, without imposing this exclusion, we can not expect to get deformation
sensitivity results of the form (D.1). This can be seen as
follows. Let us assume that we seek a bound of the form
Discrete Deep Feature Extraction: A Theory and New Architectures
kf − Fτ f k2 ≤ CN,K kτ kα
∞ , for some CN,K > 0 and
some α > 0, that applies to all f [n] = c(n/N ), n ∈ IN ,
K
with c ∈ CCART
. Take τ (x) = 1/N , in which case the
deformation (Fτ f )[n] = c(n/N − 1/N ) amounts to a
simple translation by 1/N and kτ k∞ = 1/N ≤ 1. Let
K
c(x) = 1[0,2/N ] (x). Then c ∈ CCART
for K = 1 and
√
kf − Fτ f k2 = 2, which obviously does not decay with
−α
kτ kα
for some α > 0. We note that this phe∞ = N
nomenon occurs only in the discrete case.
Proof. The proof of (D.1) is based on judiciously combining deformation sensitivity bounds for the sampled components c1 (n/N ), c2 (n/N ), n ∈ IN , in (c1 + 1[a,b] c2 ) ∈
K
CCART
, and the sampled indicator function 1[a,b] (n/N ),
n ∈ IN . The first bound, stated in Lemma D.1 below,
reads
kf − Fτ f k2 ≤ CN 1/2 kτ k∞ ,
(D.2)
and applies to discrete-time signals f [n] = f (n/N ), n ∈
IN , with f : R → C satisfying the Lipschitz property with
Lipschitz constant C. The second bound we need, stated in
Lemma D.2 below, is given by
k1N
[a,b]
−
Fτ 1N
[a,b] k2
≤ 2N
1/2
kτ k1/2
∞ ,
It remains to establish (D.2) and (D.3).
Lemma D.1. Let c : R → C be Lipschitz-continuous with
Lipschitz constant C. Let further f [n] := c(n/N ), n ∈ IN .
Then,
kf − Fτ f k2 ≤ CN 1/2 kτ k∞ .
Proof. Invoking the Lipschitz property of c according to
X
kf − Fτ f k22 =
|f [n] − (Fτ f )[n]|2
n∈IN
=
N
+ k(1N
[a,b] − Fτ 1[a,b] )(Fτ f2 )k2
X
|c(n/N ) − c(n/N − τ (n/N ))|2
n∈IN
X
|τ (n/N )|2 ≤ C 2 N kτ k2∞
n∈IN
completes the proof.
We continue with a deformation sensitivity result for sampled indicator functions 1[a,b] (x).
Lemma D.2. Let [a, b] ⊆ [0, 1] and set 1N
[a,b] [n] :=
1
N −1
1[a,b] (n/N ), n ∈ IN , with a, b ∈
/ {0, N , . . . , N }. Then,
we have
N
1/2
k1N
kτ k1/2
∞ .
[a,b] − Fτ 1[a,b] k2 ≤ 2N
kf − Fτ f k2 ≤ kf1 − Fτ f1 k2 + k1N
[a,b] (f2 − Fτ f2 )k2
(D.4)
≤ kf1 − Fτ f1 k2 + kf2 − Fτ f2 k2
−
1/2
where in the last step we used kτ k∞ ≤ kτ k∞ , which is
thanks to the assumption kτ k∞ ≤ 1. This completes the
proof of (D.1).
n ∈ IN ,
we have
+
≤ 4KN 1/2 kτ k1/2
∞ ,
(D.3)
f [n] = c1 (n/N ) + 1[a,b] (n/N )c2 (n/N )
k1N
[a,b]
kf − Fτ f k2 ≤ 2KN 1/2 kτ k∞ + 2KN 1/2 kτ k1/2
∞
≤ C2
and applies to sampled indicator functions 1N
[a,b] [n] :=
1
1[a,b] (n/N ), n ∈ IN , with a, b ∈
/ {0, N , . . . , NN−1 }. We
now show how (D.2) and (D.3) can be combined to estabN,K
lish (D.1). For a sampled cartoon function f ∈ CCART
,
i.e.,
=: f1 [n] + 1N
[a,b] [n]f2 [n],
supn∈IN |c2 (n/N − τ (n/N ))| ≤ supx∈R |c2 (x)| =
kc2 k∞ ≤ K), this yields
Fτ 1N
[a,b] k2 kFτ f2 k∞ ,
where in (D.4) we used
Fτ (1N
[a,b] f2 ) [n] = (1[a,b] c2 )(n/N − τ (n/N ))
= 1[a,b] (n/N − τ (n/N ))c2 ((n/N − τ (n/N )))
= (Fτ 1N
[a,b] )[n](Fτ f2 )[n].
With the upper bounds (D.2) and (D.3), invoking properN,K
ties of CCART
(namely, (i) c1 , c2 satisfy the Lipschitz property with Lipschitz constant C = K and hence f1 [n] =
c1 (n/N ), f2 [n] = c2 (n/N ), n ∈ IN , satisfy (D.2) with
C = K, and (ii) kFτ f2 k∞ = supn∈IN |(Fτ f2 )[n]| =
Proof. In order to upper-bound
X
N
2
N
2
k1N
|1N
[a,b] − Fτ 1[a,b] k2 =
[a,b] [n] − (Fτ 1[a,b] )[n]|
n∈IN
=
X
|1[a,b] (n/N ) − 1[a,b] (n/N − τ (n/N ))|2 ,
n∈IN
we first note that the summand h(n) := |1[a,b] (n/N ) −
1[a,b] (n/N − τ (n/N ))|2 satisfies h(n) = 1, for n ∈ S,
where
n
o
n
n
n
∈ [a, b] and
−τ
∈
/ [a, b]
S := n ∈ IN
N
N
n
N
o
n
n
n
∪ n ∈ IN
∈
/ [a, b] and
−τ
∈ [a, b] ,
N
N
N
and h(n) = 0, for n ∈ IN \S. Thanks to a, b ∈
/
{0, N1 , . . . , NN−1 }, we have S ⊆ Σ, where
n
o
n
Σ := n ∈ Z
− a < kτ k∞
N
n
o
n
∪ n∈Z
− b < kτ k∞ .
N
Discrete Deep Feature Extraction: A Theory and New Architectures
The cardinality of the set Σ can be upper-bounded by
k∞
2 2kτ
1/N , which then yields
N
2
k1N
[a,b] − Fτ 1[a,b] k2 =
X
and to note that
ad ≤ (Bd L2d Rd2 )ad−1 ≤ · · · ≤
|h(n)|2
X
1≤
n∈S
X
1 ≤ 4N kτ k∞ .
(D.5)
=
Remark D.2. For general a, b ∈ [0, 1], i.e., when we drop
the assumption a, b ∈
/ {0, N1 , . . . , NN−1 }, it follows that
0
S ⊆ Σ , where
o
n
n
Σ0 := n ∈ Z
− a ≤ kτ k∞
N
n
o
n
∪ n∈Z
− b ≤ kτ k∞ .
N
Noting that the cardinality of Σ0 can be upper-bounded by
k∞
2 2kτ
1/N + 1 = 4N kτ k∞ + 2, this then yields (similarly
to (D.5))
X
N
2
k1N
1 ≤ 4N kτ k∞ + 2,
[a,b] − Fτ 1[a,b] k2 ≤
which shows that the deformation error—for general a, b ∈
[0, 1]—does not decay with kτ kα
∞ for some α > 0 (see also
the example in Remark D.1).
q∈Λd
1
||(fq − hq )||22 ,
(E.2)
q∈Λd
1
}
where (E.2) follows by Young’s inequality (Folland, 2015).
Remark E.1. We P
emphasize that (E.1) can also be upperbounded by Bd+1 q∈Λd ||(fq −hq )||22 , which follows from
1
the fact that {gλd+1 }λd+1 ∈Λd+1 ∪ {χd } are atoms of the
convolutional set Ψd+1 with Bessel bound Bd+1 . Hence,
one can substitute kχd k1 in (15) by Bd+1 .
The key step is then to show that ad can be upper-bounded
according to
ak ≤ (Bk L2k Rk2 )ak−1 ,
Bk L2k Rk2 kf − hk22 ,
k=1
which yields (16). We now establish (E.3). Every path
q̃ ∈ Λk1 = Λ1 × · · · × Λk−1 ×Λk
{z
}
|
=Λk−1
1
of length k can be decomposed into a path q ∈ Λk−1
1
of length k − 1 and an index λk ∈ Λk according to
q̃ = (q, λk ). Thanks to (5) we have U [q̃] = U [(q, λk )] =
Uk [λk ]U [q], which yields
X
X X
kUk [λk ]fq
kfq̃ − hq̃ k22 =
q̃∈Λk
1
λk ∈Λk
q∈Λk−1
1
(E.4)
We next note that the term inside the sums on the RHS in
(E.4) satisfies
≤ L2k Rk2 k(fq − hq ) ∗ gλk k22 ,
We start by establishing i). For ease of notation, again, we
let fq := U [q]f and hq := U [q]h, for f, h ∈ HN1 , q ∈ Λd1 .
We have
X
|||ΦdΩ (f ) − ΦdΩ (h)|||2 =
||(fq − hq ) ∗ χd ||22 (E.1)
{z
d
Y
kfq − hq k22
q∈Λ01
kUk [λk ]fq − Uk [λk ]hq k22
= kPk ρk (fq ∗ gλk ) − Pk ρk (hq ∗ gλk ) k22
E. Appendix: Theorem 2
=:ad
Bk L2k Rk2
X
− Uk [λk ]hq k22 .
n∈Σ
|
d
Y
k=1
This completes the proof.
≤ kχd k21
=
n∈Σ
X
Bk L2k Rk2 a0
k=1
n∈IN
=
d
Y
k = 1, . . . , d,
(E.3)
(E.5)
where we used the Lipschitz continuity of Pk and ρk with
Lipschitz constants Rk > 0 and Lk > 0, respectively. As
{gλk }λk ∈Λk ∪ {χk−1 } are the atoms of the convolutional
set Ψk , and fq , hq ∈ HNk by (5), we have
X
k(fq − hq ) ∗ gλk k22 ≤ Bk kfq − hq k22 ,
λk ∈Λk
which, when used in (E.5) together with (E.4), yields
X
X
kfq̃ − hq̃ k22 ≤ Bk L2k Rk2
kfq − hq k22 ,
q̃∈Λk
1
q∈Λk−1
1
and hence establishes (E.3), thereby completing the proof
of i).
We now turn to ii). The proof of (17) follows—as in the
proof of ii) in Theorem 1 in Appendix C—from (16) together with ΦdΩ (h) = {(U [q]h) ∗ χd }q∈Λd1 = 0 for h = 0,
see (C.13).
We continue with iii). The proof of the deformation sensitivity bound (18) is based on two key ingredients. The
Discrete Deep Feature Extraction: A Theory and New Architectures
first one is the Lipschitz continuity result in (16). The second ingredient is, again, the deformation sensitivity bound
(D.1) stated in Proposition D.1 in Appendix D. Combining
(16) and (D.1)—as in the proof of iii) in Theorem 1 in Appendix C—then establishes (18) and completes the proof
of iii).
We proceed to iv). For ease of notation, again, we let fq :=
U [q]f , for f ∈ HN1 , q ∈ Λd1 . Thanks to (5), we have
fq ∈ HNd+1 , for q ∈ Λd1 . The key step in establishing (19)
is to show that the operator Uk , k ∈ {1, 2, . . . , d}, defined
in (4) satisfies the relation
(Uk [λk ]Tm f ) = Tm/Sk (Uk [λk ]f ),
(E.6)
for f ∈ HNk , m ∈ Z with Smk ∈ Z, and λk ∈ Λk . With the
definition of U [q] in (5) this then yields
(U [q]Tm f ) = Tm/(S1 ···Sd ) (U [q]f ),
(E.7)
m
for f ∈ HN1 , m ∈ Z with S1 ...S
∈ Z, and q ∈ Λd1 . The
d
identity (19) is then a direct consequence of (E.7) and the
translation-covariance of the circular convolution operator
m
(which holds thanks to S1 ...S
∈ Z):
d
ΦdΩ (Tm f ) =
=
U [q]Tm f ∗ χd
(Pk0 Tm f )[n] =
n+S
k −1
X
αl−n f [l − m]
l=n
(n−m)+Sk −1
X
=
αl−(n−m) f [l0 ]
l0 =(n−m)
= (Tm Pk0 f )[n],
and
for
(iii)
max-pooling
maxl∈{n,...,n+Sk −1} |f [l]| with
(Pk0 Tm f )[n] =
=
=
max
l∈{n,...,n+Sk −1}
(Pk0 f )[n]
=
|f [l − m]|
max
(l−m)∈{n−m,...,(n−m)+Sk −1}
max
l0 ∈{(n−m),...,(n−m)+Sk −1}
|f [l − m]|
|f [l0 ]|
= (Tm Pk0 f )[n],
in all three cases for f ∈ HNk , n ∈ INk , and m ∈ Z. This
then yields
(Pk Tm f )[n] = (Pk0 Tm f )[Sk n] = (Tm Pk0 f )[Sk n]
= Pk0 (f )[Sk n − m]
q∈Λd
1
Tm/(S1 ···Sd ) U [q]f ∗ χd
(Tm f )[n] = f [n − m] = (Tm Pk0 f )[n], (ii) average pooling
Pn+S −1
(Pk0 f )[n] = l=n k αl−n f [l] with
= Tm/(S1 ···Sd ) (U [q]f ) ∗ χd
q∈Λd
1
q∈Λd
1
= Tm/(S1 ···Sd ) ΦdΩ (f ),
m
for f ∈ HN1 and m ∈ Z with S1 ...S
∈ Z. It remains to
d
establish (E.6):
(Uk [λk ]Tm f ) = Pk ρk ((Tm f ) ∗ gλk )
= Pk ρk (Tm (f ∗ gλk ))
(E.8)
= Pk Tm (ρk (f ∗ gλk )) ,
(E.9)
where in (E.8) we used the translation covariance of the circular convolution operator (which holds thanks to m ∈ Z),
and in (E.9) we used the fact that point-wise non-linearities
commute with the translation operator thanks to
(ρk Tm f )[n] = ρk ((Tm f )[n])
= ρk (f [n − m]) = (Tm ρk f )[n],
for f ∈ HNk , n ∈ INk , and m ∈ Z. Next, we note
that the pooling operators Pk in Section 2.3.1 (namely,
sub-sampling, average pooling, and max-pooling) can all
be written as (Pk f )[n] = (Pk0 f )[Sk n], for some Pk0 that
commutes with the translation operator, namely, for (i)
sub-sampling (Pk0 f )[n] = f [n], with (Pk0 Tm f )[n] =
= Pk0 (f )[Sk (n − Sk−1 m)]
= Pk (f )[n − Sk−1 m]
= (Tm/Sk Pk f )[n],
(E.10)
for f ∈ HNk and n ∈ INk+1 . Here, we used m/Sk ∈
Z, which is by assumption. Substituting (E.10) into (E.9)
finally yields
(Uk [λk ]Tm f ) = Tm/Sk Uk [λk ]f,
for f ∈ HNk , m ∈ Z with Smk ∈ Z, and λk ∈ Λk . This
completes the proof of (E.6) and hence establishes (19).
| 9cs.NE
|
Computer Algebra and Material Design
Akihito Kikuchi
arXiv:1612.02275v4 [cond-mat.mtrl-sci] 17 Feb 2018
Canon Inc, Tokyo, Japan∗
1
Abstract
This article is intended to an introductory lecture in material physics, in which the
modern computational group theory and the electronic structure calculation are in collaboration. The effort of mathematicians in the field of group theory have ripened as a new
trend, called “computer algebra”, outcomes of which now can be available as handy computational packages, and would also be useful to physicists with practical purposes. This
article, in the former part, explains how to use the computer algebra for the applications in
the solid-state simulation, by means of one of the computer algebra package, the GAP system. The computer algebra enables us to obtain various group theoretical properties with
ease, such as the representations, the character tables, the subgroups, etc. Furthermore,
it would grant us a new perspective of material design, which could be executed in mathematically rigorous and systematic way. Some technical details and some computations
which require the knowledge of a little higher mathematics (but computable easily by the
computer algebra) are also given. The selected topics will provide the reader with some
insights toward the dominating role of the symmetry in crystal, or, the “mathematical
first principles” in it. In the latter part of the article, we analyze the relation between the
structural symmetry and the electronic structure in C60 (as an example to the system without periodicity). The principal object of the study is to illustrate the hierarchical change
of the quantum-physical properties of the molecule, in accordance with the reduction of
the symmetry (as it descends down in the ladder of subgroups). As an application, this
article also presents the computation of the vibrational modes of the C60 by means of the
computer algebra. In order to serve the common interest of the researchers, the details
of the computations (the required initial data and the small programs developed for the
purpose) are explained as minutely as possible.
∗
akihito [email protected] (The corresponding author)
2
I.
INTRODUCTION
In the history of physics, the cooperation of physicists and mathematicians has
yielded a great harvest in the first half of the twentieth century, as the installation
of group theory in quantum mechanics[1]. One of the main applications is realized
in the field of solid-state physics[2–4]. In subsequent years, however, such productive relationship between physics and mathematics became enfeebled, as physicists
and mathematicians were pursuing their own interests separately. In material science, the typical tool of study has turned into “first principle electronic structure
computation”, in which rapid computers are intensively used so that the quantitative simulation could be achieved. In contrast, the group theoretical view in the
quantum physics is rather a qualitative one, which could explain the likeness in similar material structures but could not illuminate the origin of subtle but distinct
differences. The standpoints of group theoretical analysis, and of first principles
simulation, are located at cross-purposes. This is one of the reasons which brought
about the breaking-off between the group theory and the first principles electronic
structure computation. There is another hardship which hinders the collaboration
of mathematics and physics. The computation of the “representation” in group
theory[5–9], useful in physical applications, requires special arts, which is recondite
to non-experts. Thus, for the purpose of general use, the character tables are listed
up in literature and textbooks (recently, moreover in databases). Traditionally we
are obliged to consult with the non-electric data, which is prone to troublesome errors. Such a circumstance has little affinity to the modern style of computational
physics, where necessary data should easily be accessible on the computer or computed anew by the researcher. Nevertheless, it must be stressed here that such a
cumbersome situation has already become surmountable. The development of the
effective computer architectures and the program packages for computational group
3
theory in recent decades have benefited mathematicians so well that they could actually obtain definite symbolic or numeric solutions to the problems of their own, not
only in proving the existence of the solution[10, 11]. These computational tools are
not restricted to mathematicians, but available to every scientific researcher. Powerful outcomes by mathematicians, also, will confer a modernized viewpoint to the
material science, so that the conventional computational tools could be assisted and
improved, although, in the present state, most of the material scientists seem to be
content in having old-fashioned knowledge of mathematics. This article takes several
examples in the group theoretical analysis in the electronic structure calculation and
presents the details of the computation so that the readers would know how to use
modern mathematical packages of computational group theory. The necessary group
theoretical data are computed by the desktop computer, without references to other
resources, and are applicable to the analysis of the quantum physics in materials
which is principally governed by the symmetry of the system. The article furthermore expounds the possible style of the systematic material design, by which the
electronic structure might be controlled artificially from the viewpoint of the structural symmetry. From these topics, the potentiality of the cooperation of computer
algebra and first-principles electronic structure (which are legitimate heirs of group
theory and quantum mechanics) may be observed.
To be fair, there are a lot of works of physicists who make use of computer algebra,
although the works of pedagogical purpose are not so many. My article is only one
of them. And here I list up some of such works for the benefit of readers.
The vibrational spectrum of C60 is analyzed in the thesis of Mooij[13]. The representation of the group is computed by the computer algebra system Maple; the
vibrational modes are computed by the reduction of the size of the dynamical matrix (of the empirical potential model); the parameters of these models are fit against
the experimental data for the optically active modes. My article also treats the vibra4
tional modes of this molecule, but it does not arrive at the quantitative computation
of vibrational frequencies. My article, by means of GAP, classifies geometrically possible deformations with respect to irreducible representations and generates the basis
set of deformations in the symbolic formulas, although the obtained results could be
applied to more quantitative studies.
The book by El-Batanouny and Wooten treats the wide range of topics related
to the symmetry in the condensed matter[14]. The interest of the research is almost
the same as my article. One of the principal difference to my article is this: the
authors of this book developed the programs for the group theoretical study by
means of Mathematica, while I adopt the GAP system. The authors of this book
developed Mathematica programs which compute the characters of groups by means
of Dixon’s method and applied them toward various topics. (I should say this:
the GAP package computes the characters of groups by the same method, so we can
directly step into applications in physics, as end-users of mathematics.) I recommend
the readers to read the book along with my article, if possible, in order to solidify
their understanding.
The book by Stauffer et al treats the general topics on the computer simulation
and the computer algebra[15]. The adopted computer algebra system is ”reduce”, for
the purpose of symbolic computations, and the application to the physics is rather
introductory.
Hergert and coworkers made the group-theoretical analysis of electronic and photonic band structure[16, 17]. The symmetry properties of Schrödinger equation and
Maxwell equation are investigated. The computer algebra system Mathematica enables them to simplify group theoretical study.
As for the introduction to the computational physics, the book by Thijssen will
be useful[18].
In the study of quantum chemistry, Sakiyama and Waki utilized the computer
5
algebra system GAP[19]. The possible conformations (for a hexakis-methylamine
nickel(ii) complex cation) are generated by the programs built on GAP, and the
stability of possible formations is computed and compared through the quantitative
electronic structure computation.
In the thesis of Rykhlinskaya, the author developed a computer algebra package
”Bethe” in the framework of Maple, and applied it to the symmetric molecular
geometry and symmetries, the vibrational analysis of the molecule, and the analysis
of the atomic behavior and the atoms in the crystal field[20].The applications of
Bethe are demonstrated in the article of Fritzshche[21].
Barnett gave us the review article in the application of computer algebra, not only
in the life science but in the material science[22]. The interest of the author is centered
on the symbolic computation (especially in chemistry), not of the computer group
theory. However, we can find useful references there, if we extend our researches
beyond group theory.
“Bilbao Crystallographic Server” is an online database of crystallography[23–25].
From this server, we can find data of crystallographic point groups and space groups
and useful programs.
Also, there are studies of the group theoretical property of C60 by mathematicians[26,
27]. The authors of these articles (Chung, Kostant, and Sternberg) discuss the embedding of the icosahedral group (of C60 symmetry) into a larger group, which is
a realization of the statement of Galois. And they analyze the energy spectrum of
the molecule by the representation theory, with the hope that their results will find
applications in the physical properties of the molecule.
Eick and Souvignier give us a survey of the algorithms for space groups and
crystallographic groups available in the computer algebra system GAP and in the
software packages Carat and Cryst[28].
It seems to me, that the application of the GAP system to the group theoretical
6
analysis in the material science is still undeveloped, although a great deal of the
products in the pure mathematics can be attainable easily from GAP. Some of these
related works explain to us how to use the computational packages developed on
some computer algebra system, and provide us with those packages, but they are
not kind enough to explain their implementation. So in my article, I intend to
present ”the raw programs” and ”the raw computed results” so that the readers may
have experience on the actual computations by GAP. Again, I must say this: the
examples taken in the following chapters are not novel, and a lot of group theoretical
analysis is already done, without the aid of computer algebra. I only show an easier
way of these analyses by means of a desktop PC. If readers thought that some of the
methods are unfamiliar or novel, they are not of my invention, in the strict sense:
they would be well-known things in the field of pure mathematics, and they are only
interpretations in the viewpoint of physics.
II.
COMPUTATION OF GROUP THEORETICAL PROPERTIES USING
“GAP”
The computational discrete algebra package GAP is one of the powerful tools for
the computation in the group theory[12]. It can be applied to the determination of the
group theoretical properties which are necessary for solid-state electronic structure
computation. For this purpose, the point group of the crystal must be determined.
The group operations are to be listed up; the multiplication table is to be prepared;
through which the character table is computed.
At first, the concept of the character of the group is explained here. Let us consider
an equilateral triangle as in Fig. 1. There are operations in the plane, such as, the
rotation by 60 degrees at the center of mass, and the reflection at an axis which con7
nects the center of mass and one of the vertexes. These operations exchange vertexes
among themselves but fix the triangle unmoved. In the mathematical terminology,
these operations consist a group C3v . These operations are the permutation of the
vertexes, numbered as 1,2, and 3 in the clockwise direction.
FIG. 1. A triangle, as the object of the symmetric operation. The symmetric operation
can be through the permutations of three symbols, 1,2,3, or the rotations and reflections
in the x-y plane.
They are expressed as
E=
B=
D=
1 2 3
1 2 3
1 2 3
1 3 2
1 2 3
3 1 2
,A =
,C =
,F =
1 2 3
2 1 3
1 2 3
3 2 1
1 2 3
2 3 1
,
,
.
In the above the top and the bottom low denote the initial and the final arrangements
8
of the vertexes respectively.
These operations can also be expressed by six rotation matrices in the x-y plane:
E=
B=
D=
1 0
0 1
,A =
√
3
2
− 12
√
3
2
1 0
,
0 −1
√
3
1
− − 23
2
,
, C = √2
3
1
1
−
2
2
2
√
√
3
1
− − 23
2
.
, F = √2
3
1
− 21
−
2
2
√
− 12
−
The multiplication table of these elements is written in table 1.
E
A
B
C
D
F
E
E
A
B
C
D
F
A
A
E
F
D
C
B
B
B
D
E
F
A
C
C
C
F
D
E
B
A
D
D
B
C
A
F
E
F
F
C
A
B
E
D
TABLE 1. The multiplication table. The multiplications xy between the elements in the
left column x and in the top row y are shown.
Let us examine these maps:
Γ1 : E, A, B, C, D, F → 1
Γ10 : E, D, F → 1; A, B, C → −1 (Determinant of matrices)
Γ2 : E, A, B, C, D, F → E, A, B, C, D, F (Identity map as matrices)
(1)
One can see that the multiplication table is kept unaltered by these maps with the
replacement of the six symbols to corresponding targets. The geometrical operations
9
in the triangle, forming a group, are represented by these maps, even if not always
faithfully; the viewpoint can be switched from geometrical one to numerical one.
This is an example of the representation of the group; all elements in the group are
represented by proxies of scalars or matrices; the multiplications among them are
subject to the same rule as the original group elements. The trace of the matrix
representation, such as Γ2 , is called “character”, having certain properties favorable
to the application of quantum physics, as will be seen later. The character table of
C3v is given in table 2.
E
Γ1 1
Γ10 1
Γ2 2
A
1
-1
0
B
1
-1
0
C
1
-1
0
D
1
1
-1
F
1
1
-1
TABLE 2. The character table of C3v .
As can be seen in the multiplication table, the group elements are divided into
subsets such as
Cl(a) = {g ∈ G| there exists x ∈ G such as g = xax−1 }.
(2)
These subsets are called conjugacy classes. The elements in the same conjugacy
class have a common value of the character; it is customary for character tables to
be labeled by conjugacy classes, not by each group elements. The conjugacy classes
for C3v are given by three subsets: {E}, {A, B, C} and {D, F}.
The gap computation for this example proceeds as follows. The group is defined
by the minimal set of generators including a rotation and a reflection:
gap> G:=Group((1,2,3),(1,2));
Group([ (1,2,3), (1,2) ])
10
The program returns the results in the shorthand notation preferred by mathematicians.
gap> Elements(G);
[ (), (2,3), (1,2), (1,2,3), (1,3,2), (1,3) ]
gap> M:=MultiplicationTable(G);
[ [ 1, 2, 3, 4, 5, 6 ], [ 2, 1, 4, 3, 6, 5 ], [ 3, 5, 1, 6, 2, 4 ],
[ 4, 6, 2, 5, 1, 3 ], [ 5, 3, 6, 1, 4, 2 ], [ 6, 4, 5, 2, 3, 1 ] ]
The multiplication table is given by a matrix, which is represented by a doubly
nested list; each of the inner lists should be read as one of the lows in the matrix.
The entry Mij shows the result of the multiplication gi · gj between elements gi and
gj .
Subgroups and generators are computed by:
gap> AllSubgroups(G);
[ Group(()), Group([ (2,3) ]), Group([ (1,2) ]), Group([ (1,3) ]),
Group([ (1,2,3) ]), Group([ (1,2,3), (2,3) ]) ]
gap> GeneratorsOfGroup(G);
[ (1,2,3), (1,2) ]
gap> G.1;
(1,2,3)
gap> G.2;
(1,2)
gap> S:=Subgroup(G,[(1,2)]);
Subgroup(G,[(1,2)]);
gap> Elements(S);
[ (), (1,2) ]
gap> A:=Group((1,2,3));
11
Group([ (1,2,3) ])
gap> IsSubgroup(G,A);
true
gap> IsSubgroup(A,G);
false
gap> A:=Group((1,2,3,4));
Group([ (1,2,3,4) ])
gap> IsSubgroup(G,A);
false
The group can also be defined from the multiplication table as
gap> G2:=GroupByMultiplicationTable(M);
<group of size 6 with 6 generators>
gap> Elements(G2);
[ m1, m2, m3, m4, m5, m6 ]
The elements in the group constructed from the multiplication tables are represented
by abstract symbols m1,m2,...,m6.
In order to construct the group from the rotation in the Euclidean space, the
generating set of the rotation matrices should be given.
gap> M1:=[[-1/2,ER(3)/2],[-ER(3)/2,-1/2]];
[ [ -1/2, -1/2*E(12)^7+1/2*E(12)^11 ],
[ 1/2*E(12)^7-1/2*E(12)^11, -1/2 ] ]
gap> M2:=[[1,0],[0,-1]];
[ [ 1, 0 ], [ 0, -1 ] ]
gap> G3:=Group(M1,M2);
Group(
12
[
[ [ -1/2, -1/2*E(12)^7+1/2*E(12)^11 ],
[ 1/2*E(12)^7-1/2*E(12)^11, -1/2 ] ], [ [ 1, 0 ], [ 0, -1 ] ]
])
Here ER(n) is
√
n and E(n) is the primitive n-th root of unity exp(2πi/n).
Let us inquire of the GAP package whether the three different definitions should
generate equivalent groups.
gap> G=G;
true
gap> G2=G;
false
gap> G3=G;
false
The GAP program decides that the three groups are not identical in the strict sense,
because they are composed from different resources, i.e. permutations, abstract
symbols, and matrices. However, we can construct isomorphisms among them. The
following GAP command returns the isomorphism on the generators of the groups.
gap> IsomorphismGroups(G,G);
[ (1,2,3), (1,2) ] -> [ (1,2,3), (1,2) ]
gap> IsomorphismGroups(G2,G);
[ m1, m2, m3, m4, m5, m6 ] -> [ (), (2,3), (1,2), (1,2,3), (1,3,2),
(1,3) ]
gap> IsomorphismGroups(G,G2);
[ (1,2,3), (1,2) ] -> [ m4, m3 ]
gap> G4:=Group((1,2,3,4))
13
gap> IsomorphismGroups(G,G4);
fail
The existence of the isomorphism between G and G3 can also be verified in this way.
The conjugacy classes are computed by this command:
gap> ConjugacyClasses(G);
[ ()^G, (2,3)^G, (1,2,3)^G ]
To access entries in the conjugacy classes, a special method is needed. The following cannot work well:
gap> Elements(ConjugacyClasses(G));
[ ()^G, (2,3)^G, (1,2,3)^G ]
Instead, using “List” command and anonymous functions in λ calculus, we can
access each entry in the list:
gap> List(Elements(G),x->x);
[ (), (2,3), (1,2), (1,2,3), (1,3,2), (1,3) ]
gap> List(Elements(G),x->x^-1);
[ (), (2,3), (1,2), (1,3,2), (1,2,3), (1,3) ]
gap> List(ConjugacyClasses(G),Elements);
[ [ () ], [ (2,3), (1,2), (1,3) ], [ (1,2,3), (1,3,2) ] ]
gap> List(ConjugacyClasses(G),Representative);
[ (), (2,3), (1,2,3) ]
The character table is computed by this command:
gap> irrg:=Irr(G);
[ Character( CharacterTable( Sym( [ 1 .. 3 ] ) ), [ 1, 1, 1 ] ),
Character( CharacterTable( Sym( [ 1 .. 3 ] ) ), [ 1, -1, 1 ] ),
Character( CharacterTable( Sym( [ 1 .. 3 ] ) ), [ 2, 0, -1 ] ) ]
14
The character table can be accessible as a numeral list, or as a map on the group
elements:
gap> List([1,2,3],y->List([1,2,3],x->irrg[y][x]));
[ [ 1, 1, 1 ], [ 1, -1, 1 ], [ 2, 0, -1 ] ]
gap> List([1,2,3],y->List(Elements(G),x->x^irrg[y]));
[ [ 1, 1, 1, 1, 1, 1 ], [ 1, -1, -1, 1, 1, -1 ],
[ 2, 0, 0, -1, -1, 0 ] ]
The irreducible representation is computed by
gap> rep:=IrreducibleRepresentations(G);
[ Pcgs([ (2,3), (1,2,3) ]) -> [ [ [ 1 ] ], [ [ 1 ] ] ],
Pcgs([ (2,3), (1,2,3) ]) -> [ [ [ -1 ] ], [ [ 1 ] ] ],
Pcgs([ (2,3), (1,2,3) ]) ->
[ [ [ 0, 1 ], [ 1, 0 ] ], [ [ E(3), 0 ], [ 0, E(3)^2 ] ] ] ]
gap> List(Elements(G),x->x^rep[3]);
[ [ [ 1, 0 ], [ 0, 1 ] ], [ [ 0, 1 ], [ 1, 0 ] ],
[ [ 0, E(3)^2 ], [ E(3), 0 ] ], [ [ E(3), 0 ], [ 0, E(3)^2 ] ],
[ [ E(3)^2, 0 ], [ 0, E(3) ] ], [ [ 0, E(3) ], [ E(3)^2, 0 ] ] ]
This command returns the irreducible representations, giving the relations between
the generators of the group and the matrix representation. The characters are computed by taking traces of the matrix representation.
The character table is also computed by
gap> tbl:=CharacterTable(G);
CharacterTable( Sym( [ 1 .. 3 ] ) )
gap> Display(tbl);
CT1
15
2
1
1
.
3
1
.
1
1a 2a 3a
2P 1a 1a 3a
3P 1a 2a 1a
X.1
1
1
1
X.2
1 -1
1
X.3
2
. -1
To make use of the character table computed in this way, various subsidiary commands are prepared in GAP system. (N.B. The orderings of representations or
characters by these commands do not always coincide with each other.)
gap> ConjugacyClasses(tbl);
[ ()^G, (2,3)^G, (1,2,3)^G ]
gap> Irr(tbl);
[ Character( CharacterTable( Sym( [ 1 .. 3 ] ) ), [ 1, -1, 1 ] ),
Character( CharacterTable( Sym( [ 1 .. 3 ] ) ), [ 2, 0, -1 ] ),
Character( CharacterTable( Sym( [ 1 .. 3 ] ) ), [ 1, 1, 1 ] ) ]
Remark. There are various algorithms for the computation of characters. The simplest one, the Burnside algorithm, starts from the counting in the elements in the
conjugacy classes, sets up a matrix, and the characters are computed through the
eigenvalue problem. The Dixon (or Dixon-Schneider) algorithm goes in a more advanced way, employing the computation in the prime fields and the other techniques
16
of the symbolic computation, but it is effective in the large groups[29–31]. The latter
algorithm is adopted in GAP.
III.
SOME PRELIMINARIES
A.
Projection operator
Once the characters of the group are obtained, “projection operators” are constructed and assigned to each of the irreducible representation. The definition is as
follows:
P (p) =
lp X (p)∗
χ
· OT ,
|G| T ∈G T
(3)
where lp is the dimension of the irreducible representation,|G| is the order of the
(p)
group, and χT is the character of the group element T , which is allowed to be
complex-valued. OT is the operator allotted to each T in the space in which the
group action is defined. This operator acts on functions in the Euclidean space as
follows:
OT f (r) = f (T −1 (r)).
(4)
By this construction, we can distinguish whether a function belongs to the corresponding irreducible representation or not. If the projector is applied to the basis
functions of the corresponding irreducible representations, these functions are invariant as a set: on the other hand, the projector is applied to the basis functions of
different irreducible representations, the result of the projection goes to zero. (The
epithet “invariant” should be understood in the following way: in many cases, plural
basis functions belong to one irreducible representation. Through the group operations, these basis functions are interchanged with each other or re-expressed as the
linear combination of them; moreover, they are never transformed into functions be17
longing to other representations. In a word, they are invariant as a subspace; if the
subspace corresponding to a representation is minimal, no more being divisible, it is
called irreducible. The whole of representations and characters of the arbitrary finite
group could be computed exactly, without omission, by the principle of the group
theory.)
(p)
If we know the explicit forms of the irreducible matrix representation Dij (R),
the projection operator of the following form is useful:
(p)
Pkl =
lp X (p)∗
D (OT ) · OT .
|G| T ∈G kl
(5)
Owing to the definition of the partner functions
OT |p; li =
X
(p)
|p; jiDjl (T )
(6)
j
and the orthogonality theorem of the representation
X
(p)∗
(q)
Dik (T )Dlm (T ) =
T
|G|
δpq δil δkm
lp
(7)
this projector transforms one basis vector into another basis in one irreducible representation:
(p)
Pkl |p; li = |p; ki.
(8)
Moreover it can project out k-th partner function of the irreducible representation
P P (q)
from an arbitrary function F = q j fj |q; ji as
(p)
(p)
Pkk F = fk |p; ki.
(9)
At this junction, the connection between the group theory and the energy spec18
trum of the Schrodinger equations arises. Let the potential term be invariant under
some symmetry operations. With this potential term, certain eigenfunctions should
exist. When the symmetry operations are applied to the Schrodinger equation, the
Laplacian, the potential, and the energy spectrum are invariant: the changeable is
only the wavefunction, which is expressed by the basis functions in the irreducible
representation. If the irreducible representation is one-dimensional, the wavefunction will be invariant up to a certain phase factor by the symmetry operation: on
the other-hand, if the representation is multi-dimensional, the symmetry operation
may generate different wave-functions which are other solutions to the same equation with the same eigenvalue. Such circumstances lead to the degeneracy of the
energy spectrum. The identification of each wavefunction to its proper irreducible
representation enables us to clarify the relation between the energy spectrum and the
symmetry in the system, which is helpful in the analysis of the electronic properties
of materials.
B.
Space group
An n-dimensional space group S is a discrete subgroup in a group of Euclidean
motions in Rn , such that the subgroup T in S, composed of pure translations(without
rotations or reflections) is a free Abelian subgroup of rank n, having finite index in
S. Thus an exact sequence of groups exists as
0 → T → S → P → 1,
(10)
in which S acts on T by means of the conjugation action t → s · t · s−1 , which
represents the action of the point group P , defined as a factor group S/T , on the
lattice translations.
19
In a more intuitive expression, the space group S is the set of symmetry operations
in a periodic system; it is composed of rotations and reflections which fix components
in the unit cell, combined with parallel movements along crystal axes; therefore S
is of infinite order; the parallel movements, by themselves, compose the discrete
subgroup T of infinite order, in which the direction and the stride of the movements
are determined by the primitive lattice vectors. Meanwhile, the point group P is the
set of the reminders of operations in S, taken modulo of the crystalline periodicity, i.e.
with the parallel movements in T being nullified; it is a finite group and may include
fractional translations, which are represented by linear combinations of fractions of
primitive lattice vectors.
C.
Crystallographic group
In the case of crystallographic group, the symmetry operation is the affine mapping, composed of the linear parts RT , including the reflection, rotation, and inversion, and the translation parts τT [28]. The definition is
OT = {RT |τT } : r → RT · r + τT ,
(11)
OT · OS = {RT · RS |τT + RT · τs },
(12)
OT−1 = {RT−1 | − RT−1 · τT }.
(13)
The operation on the plane wave is
OT exp(ikr) = exp(ik(RT−1 r − RT−1 τ )) = exp(i(RT · k)(r − τ )).
(14)
The affine mapping gives rise to the phase factor exp(−i(RT · k)τ )), which is not
20
unity at a general k-point. On the other hand, the matrix representation of the affine
mapping is subject to the following relation:
D(p,k) ({Rk |fRk + τn }) = e−ik·τn D(p,k) ({Rk |fR })
(15)
The trace of the matrix D(p,k) is the character: therefore the phase-factors in the character and the wave-number part in the Bloch-type wavefunction (eikr in eikr uk (r) )
are canceled with each other as complex conjugates; in the general k-point, the phase
shift in the character can actually be negligible in the projection operator.(Indeed
there are textbooks or lectures omitting these terms.) However, if the Bloch-type
P
wavefunction is expressed by plane-wave expansion
CG ei(k+G)r , one should be
cautious. The phase shift in the periodic part is eiRk (G)τn , which remains without
cancellation in the projection operation.
D.
Theoretical set-up for Wyckoff positions
The mathematical definition of the Wyckoff positions is stated as follows.
Let G be a space group, T be its translation lattice, K be its point group, tg be
fractional translations. Each element in G has a form of {k|tk + t} for k ∈ K and
t ∈ T , and acts on a vector v in V = Rn as
{k|tk + y}(v) = k · v + tk + t.
(16)
The stabilizer of v ∈ V under this action is denoted as StabG (v)={g ∈ G|g(v) =
v}. Let v ∈ V . An equivalence relations ∼ is set up if StabG (v) is conjugate
to StabG (w),i.e. StabG (w)=StabG (g(v))=g StabG (w)g −1 for some g ∈ G. This
equivalence classes of ∼ is called the Wyckoff positions of G.
21
This definition, given in the mathematicians’ terminology, is a little recondite for
physicists. For a more intuitive understanding, it can be restated as:
The Wyckoff positions are a set of coordinate points, composed from two
subsets:
[Type 1] The coordinate points which may be fixed by a certain symmetry
operation of the point group. They are regarded as generators of the
Wyckoff positions.
[Type 2] The coordinate points generated from those in Type 1 subset
by all of the symmetry operations.
The command in GAP “WyckoffPositions(S)” returns the type 1 set for a space
group S; the equivalent coordinate points to a generator “W” in type 1 set by the
symmetry operations are computed from an another command “WyckoffOrbit(W)”.
E.
Units in the computation
Throughout the computations of quantum physics in this article, we mainly use
atomic units, abbreviated as a.u. if necessary; for the length, the Bohr radius(a0 u
5.292 × 10−11 m); for the energy, electron volt unit (eV) or the Hartree unit(Eh u
27.211eV). However, some computations adopt arbitrary unit in the energy.
IV.
APPLICATION 1: IDENTIFICATION OF WAVEFUNCTIONS TO IR-
REDUCIBLE REPRESENTATIONS
The application in this section is the classical example of group theory in quantum
physics.
22
A.
The simplest case: at Γ point
In this section, the group-theoretical analysis of the wavefunction is exemplified.
The wavefunctions at the Γ point in the diamond crystal are classified to corresponding irreducible representations. The treatment for the general k-point (k 6= 0) shall
be discussed later.(The knowledge of the distinction between symmorphic or nonsymmorphic crystal is necessary; the existence of these two types of crystal makes
the discussion not a little complicated.)
At first, the character table should be computed. The symmetric operations in the
diamond structure, whose unit cell is the minimal one, including two carbon atoms,
are given in the appendix, as well as the multiplication table of these operations.
When one uses the GAP packages, there are several options for the preparation
of the point group; of which three types can be used.
First: the symmetry operations are given as a set of the three-dimensional matrix.
gap> MT[
1]:=[[
1,
0,
0],[
0,
1,
0],[
0,
0,
1]];;
gap> MT[
2]:=[[
1,
0,
0],[
0, -1,
0],[
0,
0, -1]];;
0,
0],[ -1,
..........................
gap> MT[ 48]:=[[
0,
0,
1],[
1,
0,
0]];;
gap> G:=Group(MT);
Second: the multiplication table in the group is supplied.
gap> M:=[[1,2,3,...,48],[...],...,[...]];;
gap> G:=GroupByMultiplicationTable(M);
gap> Elements(G);
[ m1, m2, m3, m4,............,m45, m46, m47, m48 ]
In this case, the group elements are denoted by the symbolical way, m1,m2,...,m48.
23
Actually, all of the group elements are not necessarily provided to define a group:
it is enough to give the smallest generating set, from which other elements are constructed. The group should be remade by means of the smallest generating set. (In
the present implementation of GAP, especially when the multiplication table is supplied, the following tendency in the computation is observed: the computations for
groups, constructed from the minimal generating set, are much quicker than those for
groups where the all elements are stored as generators.) For the case of the diamond,
the computation goes as:
GS:=SmallGeneratingSet(G);
[ m36, m48 ]
G:=Group(GS);
<group with 2 generators>
The command “SmallGeneratingSet()” yields a reasonably small generating set.
In the cases of finite solvable groups (a typical example of this is the crystal point
group) and of finitely generated nilpotent groups, “MinimalGeneratingSet()” command is also available to get minimal generators, but the computation is timeconsuming. In addition, each element “elm” in the group “G“ can be expressed
by generators, by means of Factorization(G,elm):
gap> List(Elements(G),x->Factorization(G,x));
[ <identity ...>, x2^-1*x1*x2*x1, x2^2, (x2*x1)^2, x2^2*x1^2,
x1*x2^2*x1, x1^2, x2^-1*x1^-1*x2^-1*x1, x1^-1*x2^-1*x1*x2,
x2^-1*x1^2*x2, x2^-1*x1*x2*x1^-1, x1^-2, x2^-1*x1, x1^-1*x2^-1,
x2*x1, x2*x1*x2^2, x2^2*x1*x2, x2*x1^-1, x1*x2, x2^-1*x1^-1,
x1^-1*x2^-1*x1^2, x2^-1*x1^3, x2*x1*x2^2*x1^2, x2*x1^3, x1^3,
x2^-1*x1^-1*x2^-1*x1^2, x2^2*x1^3, x1*x2^2*x1^2, x2^2*x1^-1,
24
x2*x1*x2, x1^-1, x2^-1*x1*x2, x1*x2^2, x2^2*x1, x2^-1*x1^-1*x2^-1,
x1, x1^2*x2, x1*x2*x1^-1, x2*x1^-2, x1*x2^2*x1*x2, x2*x1*x2^2*x1,
x2*x1^2, x1^-1*x2^-1*x1, x2^-1*x1^2, x1*x2*x1, x2^-1,
x2*(x2*x1)^2, x2 ]
Third: the crystallographic groups are defined by four-dimensional augmented
matrices, in which both of the point group operations and the translations are inscribed. The notations for the augmented can be given in the following form
A v
0 1
(17)
acting on column vectors (x, 1) from the right or, alternatively,
T
A
0
v 1
(18)
acting on low vectors (x, 1) from the left, which represent the affine mapping x →
A · x + v.(Crystallographers prefer the later notation.)
Once the crystallographic group is defined, the point group can easily be deduced.
As for the diamond case, in the GAP computation, the crystallographic group is
defined as follows. (The minimal generating set is used for simplicity.)
gap> M1:=[[0,0,1,0],[1,0,0,0],[0,-1,0,0],[1/4,1/4,1/4,1]];;
gap> M2:=[[0,0,-1,0],[0,-1,0,0],[1,0,0,0],[0,0,0,1]];;
gap> S:=AffineCrystGroup([M1,M2]);
<matrix group with 2 generators>
gap> P:=PointGroup(S);
Group([ [ [ 0, 0, 1 ], [ 1, 0, 0 ], [ 0, -1, 0 ] ],
25
[ [ 0, 0, -1 ], [ 0, -1, 0 ], [ 1, 0, 0 ] ] ])
By these preparations the conjugacy classes and the character table of the diamond crystal are computed as tables 3 and 4.(The group elements for the symmetry
operation are given in a table in the appendix. The entries in the conjugacy classes
are given by the numbering of the table of the group elements.)
Class
C.1
C.2
C.3
C.4
C.5
C.6
C.7
C.8
C.9
C.10
Group Elements
1 (The identity)
2,3,4
5 , 6 , 7 , 8 , 9 , 10 , 11 , 12
13 , 16 , 17 , 18 , 21 , 23
14 , 15 , 19 , 20 , 22 , 24
25
26 , 27 , 28
29 , 30 , 31 , 32 , 33 , 34 , 35 , 36
37 , 40 , 41 , 42 , 45 , 47
38 , 39 , 43 , 44 , 46 , 48
TABLE 3. Conjugacy classes of the point group in the diamond structure. The elements
are indexed in the same way as the table of the symmetry operations. )
x.1
x.2
x.3
x.4
x.5
x.6
x.7
x.8
x.9
x.10
C.1
1
1
1
1
2
2
3
3
3
3
C.2 C.3 C.4 C.5 C.6 C.7 C.8 C.9 C.10
1 1 1 1 1 1 1 1
1
1 1 -1 -1 -1 -1 -1 1
1
1 1 -1 -1 1 1 1 -1 -1
1 1 1 1 -1 -1 -1 -1 -1
2 -1 0 0 -2 -2 1 0
0
2 -1 0 0 2 2 -1 0
0
-1 0 -1 1 -3 1 0 1 -1
-1 0 -1 1 3 -1 0 -1
1
-1 0 1 -1 -3 1 0 -1
1
-1 0 1 -1 3 -1 0 1 -1
TABLE 4. The character table of the basic diamond structure at Γ point.
We can apportion the wavefunction at the Γ point to each irreducible representations by applying the projection operator. (It is enough to check whether the
26
numerical result of the projection remains almost as it is, or it nearly vanishes to
zero.) The identification in the example of the diamond is given in table 5. In the
numerical expansion coefficients of the wavefunctions, we can imagine the existence
of a hidden mechanism, as is suggested by the succession of zeros, or the interchanges
of specific values with the alteration of signs. The hierarchy and the symmetry in
them can be unraveled and classified by the group theoretical analysis.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Ene.
Rep.
k1
0
-1
1
0
0
0
0
1
-1
-1
1
0
0
1
-1
k2
0
-1
1
-1
1
0
0
0
0
-1
1
-1
1
0
0
k3
0
-1
1
0
0
1
-1
0
0
0
0
-1
1
1
-1
kGk2
0
1.6243
1.6243
1.6243
1.6243
1.6243
1.6243
1.6243
1.6243
1.8756
1.8756
1.8756
1.8756
1.8756
1.8756
1
-0.9765
0.0742
0.0742
-0.0742
-0.0742
-0.0742
-0.0742
-0.0742
-0.0742
0
0
0
0
0
0
-2.1091
x.1
2
0
0.3519
0.3519
-0.1488
-0.1488
0.3159
0.3159
0.1848
0.1848
0.0285
0.0285
0.1325
0.1325
0.3968
0.3968
19.5799
x.10
3
0
-0.0007
-0.0007
0.1173
0.1173
0.301
0.301
-0.419
-0.419
-0.2391
-0.2391
0.3315
0.3315
-0.0935
-0.0935
19.5799
x.10
4
0
0.2934
0.2934
0.4172
0.4172
-0.1397
-0.1397
0.0159
0.0159
0.3432
0.3432
0.2199
0.2199
-0.0981
-0.0981
19.5799
x.10
5
0
-0.3562
0.3562
-0.0785
0.0785
0.084
-0.084
-0.5186
0.5186
0
0
0
0
0
0
25.1394
x.7
6
0
-0.1637
0.1637
0.5376
-0.5376
0.295
-0.295
0.0788
-0.0788
0
0
0
0
0
0
25.1394
x.7
7
0
-0.3913
0.3913
0.1078
-0.1078
-0.4612
0.4612
0.1777
-0.1777
0
0
0
0
0
0
25.1394
x.7
8
0
-0.3212
0.3212
-0.3212
0.3212
0.3212
-0.3212
0.3212
-0.3212
-0.1671
0.1671
-0.1671
0.1671
0.1671
-0.1671
33.1443
x.2
TABLE 5. The identification between the irreducible representation and the wavefunctions
in diamond crystal at Γ point. Wavefunctions are expressed by coefficients (Ci , k1 , k2 , k3 )
in the plane wave expansion Ci exp(iGr); G = k1 G1 + k2 G2 + k3 G3 , where G1 , G2 , G3 are
reciprocal vectors. The left columns (from the second to the fifth ) indicate (k1 , k2 , k3 , |G|2 )
and the remaining parts give the coefficients Ci . The energies and the irreducible representations are given in last two columns indicated by “Ene.” and “Rep.”, respectively. The
carbon atoms are placed at ±(1/8 · a1 + 1/8 · a2 + 1/8 · a3 ) so that the coefficients of the
wavefunctions in the plane wave expansion are real-valued.
27
B.
Character table computation in super-cell
It is a standard way to use the minimal unit cell. In the diamond structure, the
minimal primitive cell is chosen to be the same as the face-centered one, in which
two carbon atoms are located with the point-group G(I × O, C2 × S4 ). Meanwhile,
one can choose the cubic unit cell; the volume of which is four times as large as the
minimal cell, because the cubic cell is constructed from four of the simplest unit cell
stuck together obliquely. The unit cell enlarged in a similar way, (if possible, in which
extra atoms are embedded) is called super-cell. The point groups of the supercells
are inevitably altered by the presence of extra fractional translations. The group
theoretical computations for such super-cells will be difficult for bare human powers
since the point group becomes larger and more complicated. Nevertheless, if one
can use the computer algebra system, the character tables are easily computed. We
should note that it is more advisable to compute characters of groups by ourselves if
we can. Although the possible super-cells are infinite, the character tables available
in textbooks and databases are limited, possibly only for the minimal unit cell.
For the super-cell, actually, the point group of the minimal crystal can be extended
by the semidirect product with a finite Abelian group, by which extra fractional
translations in the enlarged cell are represented. (The construction of the semidirect
product will be stated elsewhere in the following section of this article.) The easiest
and straightest way for this is to remake the multiplication table. This can be done
in the following way: first, apply possible fractional translations to group elements
in symmetry operations of the minimal cell, so that all symmetry operations in the
supercell are prepared in the following form:
{E|τi }{Rj |σj }.
28
(19)
By means of these new set of operations, we must construct the multiplication
table in the super-cell:
{Rj |σj + τi }{Rk |σk + τl } = {Rp |σp + τq }.
(20)
In the evaluation, we can make use of the multiplication table in the minimal cell:
{Rj |σj }{Rk |σk } = {Rj · Rk |σj + Rj σk } ≡ {Rp |σp }
(21)
There is a relation:
{Rj |σj + τi }{Rk |σk + τl } = {Rj · Rk |σj + τi + Rj σk + Rj τl } = {Rp |σp + τq }. (22)
From the equations (21) and (22), it leads that
τq ≡ σj + τi + Rj σk + Rj τl − σp (modulo translational vectors)
(23)
Using this τq , the extended multiplication table can be constructed with ease.
The exemplary calculation for the diamond structure shall be executed in this section. The basic unit cell in the diamond structure is the face-centered one, including
two atoms, The lattice vectors are defined as
a1 = (0, 1/2, 1/2), a2 = (1/2, 0, 1/2), a3 = (1/2, 1/2, 0)
(24)
This basic unit cell could be extended four-fold, in which atomic coordinates are
given by translations such as x → x + τi (i = 0, 1, 2, 3). The vectors τi are defined
as:
τ0 = (0, 0, 0), τ1 = (0, 1/2, 1/2), τ2 = (1/2, 0, 1/2), τ3 = (1/2, 1/2, 0)
29
(25)
The extended unit cell is the cubic one, including eight atoms, to which the lattice
vectors are defined as:
a01 = (1, 0, 0), a02 = (0, 1, 0), a03 = (0, 0, 1)
(26)
Let F be the group composed of the four fractional translation vectors τi (i =
0, 1, 2, 3). And let T be the group of primitive lattice translations, composed of
a01 , a02 , a03 and the zero vector. The factor group N(:=F/T) is a finite group, whose
generators τ̂ are given by the projection of τi onto N. This factor group is the Klein
four group, or Z2 ⊕ Z2 . The generators satisfy the following relations.
2 · τ̂i = τ̂0
(27)
τ̂1 + τ̂2 + τ̂3 = τ̂0
The point-group G0 in the cubic unit cell is the semidirect product N o G, between
the group of the fractional translations N and the point-group G in the basic unit
cell. The new group includes 192 (= 48 × 4) elements. The textbooks or databases,
however, only show the point group of the basic, minimal unit cells, but not those of
the extended super-cells. In the latter cases, the point group is made of the semidirect
product; one can compute its character tables by straightforwardly constructing the
multiplication table, or, in a more complicated and elegant way, i.e. by making
use of the induced representation of the group, when the group N is Abelian. The
computation of the induced representation will be explained later.
The group elements in the extended group are renumbered as follows:
{E|τi }{Rj |sj } → (i × 48 + j)th element of the extended group
30
(28)
The conjugacy classes and the character tables for the extended group are given
in table 6 and 7.
Class Group Elements
Ĉ.1 [0,1]
Ĉ.2 [0,2],[0,3],[0,4],[1,2],[2,3],[3,4]
[0,5],[0,6],[0,7],[0,8],[0,9],[0,10],[0,11],[0,12],
[1,5],[1,6],[1,7],[1,8],[1,9],[1,10],[1,11],[1,12],
Ĉ.3
[2,5],[2,6],[2,7],[2,8],[2,9],[2,10],[2,11],[2,12],
[3,5],[3,6],[3,7],[3,8],[3,9],[3,10],[3,11],[3,12]
[0,13],[0,17],[0,21],[1,16],[1,17],[1,23],
Ĉ.4
[2,16],[2,18],[2,21],[3,13],[3,18],[3,23]
[0,14],[0,15],[0,19],[0,20],[0,22],[0,24],
[1,14],[1,15],[1,19],[1,20],[1,22],[1,24],
Ĉ.5
[2,14],[2,15],[2,19],[2,20],[2,22],[2,24],
[3,14],[3,15],[3,19],[3,20],[3,22],[3,24]
[0,16],[0,18],[0,23],[1,13],[1,18],[1,21],
Ĉ.6
[2,13],[2,17],[2,23],[3,16],[3,17],[3,21]
Ĉ.7 [0,25],[1,25],[2,25],[3,25]
[0,26],[0,27],[0,28],[1,26],[1,27],[1,28],
Ĉ.8
[2,26],[2,27],[2,28],[3,26],[3,27],[3,28]
[0,29],[0,30],[0,31],[0,32],[0,33],[0,34],[0,35],[0,36],
[1,29],[1,30],[1,31],[1,32],[1,33],[1,34],[1,35],[1,36],
Ĉ.9
[2,29],[2,30],[2,31],[2,32],[2,33],[2,34],[2,35],[2,36],
[3,29],[3,30],[3,31],[3,32],[3,33],[3,34],[3,35],[3,36]
[0,37],[0,40],[0,41],[0,42],[0,45],[0,47],
Ĉ.10
31
[1,41],[1,42],[2,45],[2,47],[3,37],[3,40]
[0,38],[0,39],[0,43],[0,44],[0,46],[0,48],
[1,38],[1,39],[1,43],[1,44],[1,46],[1,48],
Ĉ.11
[2,38],[2,39],[2,43],[2,44],[2,46],[2,48],
[3,38],[3,39],[3,43],[3,44],[3,46],[3,48]
Ĉ.12 [1,1],[2,1],[3,1]
Ĉ.13 [1,3],[1,4],[2,2],[2,4],[3,2],[3,3]
[1,37],[1,40],[1,45],[1,47],[2,37],[2,40],[2,41],[2,42],
Ĉ.14
[3,41],[3,42],[3,45],[3,47]
TABLE 6: Conjugacy classes for the cubic diamond structure. The element {1|τj }{Ri |σi } is denoted as [j,i].
y.1
y.2
y.3
y.4
y.5
y.6
y.7
y.8
y.9
y.10
y.11
y.12
y.13
y.14
Ĉ.1
1
1
1
1
2
2
3
3
3
3
6
6
6
6
Ĉ.2 Ĉ.3 Ĉ.4 Ĉ.5 Ĉ.6 Ĉ.7 Ĉ.8
1 1 1 1 1 1 1
1 1 -1 -1 -1 -1 -1
1 1 -1 -1 -1 1 1
1 1 1 1 1 -1 -1
2 -1 0 0 0 -2 -2
2 -1 0 0 0 2 2
-1 0 -1 1 -1 -3 1
-1 0 -1 1 -1 3 -1
-1 0 1 -1 1 -3 1
-1 0 1 -1 1 3 -1
-2 0 -2 0 2 0 0
-2 0 2 0 -2 0 0
2 0 0 0 0 0 0
2 0 0 0 0 0 0
Ĉ.9 Ĉ.10 Ĉ.11 Ĉ.12 Ĉ.13 Ĉ.14
1
1
1
1
1
1
-1
1
1
1
1
1
1 -1 -1
1
1 -1
-1 -1 -1
1
1 -1
1
0
0
2
2
0
-1
0
0
2
2
0
0
1 -1
3 -1
1
0 -1
1
3 -1 -1
0 -1
1
3 -1 -1
0
1 -1
3 -1
1
0
0
0 -2
2
0
0
0
0 -2
2
0
0 -2
0 -2 -2
2
0
2
0 -2 -2 -2
TABLE 7. The character table of the cubic diamond unit cell at Γ point.
The eight lowest eigenvalues in energy spectra in the minimal and cubic diamond
cell are shown in table 8. There are two sets of threefold degeneracy in the former
32
case. On the other hand, in the latter, there is one sixfold degeneracy, which is
attributable to the overlap of the spectra at the three X points in the Brillouin zone
of the minimal unit cell. However, the character tables of the Γ point in the minimal
diamond cell could not assign them in suitable representation; the character tables
rebuilt for the cubic cell can do this.
Minimal
-2.11 19.58 19.58 19.58 25.14 25.14 25.14 33.14
Representation x.1 x.10 x.10 x.10 x.7 x.7 x.7 x.2
Cubic
-2.11 6.74 6.74 6.74 6.74 6.74 6.74 13.03
Representation y.1 y.14 y.14 y.14 y.14 y.14 y.14 y.12
TABLE 8. The energy spectra at the Γ point in the minimal and cubic diamond unit cell
(unit in eV). The assigned irreducible representations are also shown.
V.
APPLICATION 2: A SYSTEMATIC WAY OF THE MATERIAL DE-
SIGNING
A.
How to manage quantum-dynamics in crystal?
The group theoretical view will be able to bring about a systematic way of material design. In this section, although in the level of table-top exercise, a practicable
form of material design is presented. In a nutshell, through the systematic reduction of the crystal symmetry, the band-structure could tactically be modulated; the
degeneracy of the energy spectrum can be artificially split into different levels and
the new energy-gap can be opened. Toward the systematic reduction of the crystal
symmetry, the crystallographic concept, called Wyckoff positions, is utilized[28, 33].
This mathematical idea is not an abstract one, but actually computable by means
of the computer algebra.
33
One should note this: the degeneracy of the energy spectrum is governed by the
symmetry of the crystal. From the mathematical viewpoint, the point group of the
crystal system is the manifestation of the symmetry; the subgroups of the point
group describe the possible style with the reduced symmetry.(It is well known that
a material has general tendency to reduce its own symmetry in order to release itself
from the energetically unstable situation owing to the existence of the degeneracy
of occupied energy levels.) By tracing the tree structure of the inclusion relations
of the subgroup of the finite groups, one can enumerate every possibility of reduced
symmetry. For each of reduced symmetries, one can compute group representations.
The comparison of the character tables in the subgroups and the parent groups (by
means of analyzing compatibility relations) one can rummage the possibility of the
breaking-down of the energy spectrum degeneracy. The opening of the energy-gaps
in the reduced symmetry will be speculated, sorted and classified with their group
theoretical origins. This plan, therefore, will lead to a kind of systematic way of
material design. It must be admitted that such a tactic still lacks in the sufficiently
quantitative argument, since the only possible existence of energy-gap, together with
their relationship with crystal symmetry reduction, is discussed. In order to evaluate
the stability of the reduced structure or the actual possibility of its existence, one
must employ massive first-principles computation. However, the group theoretical
inspection has its own merit; it can be achieved by a light and quick computation,
frugal in hardware usage; maybe simple computations in the level of elementary
perturbation theory will do. Hence it can be applied to screening purposes for picking
up candidates before serious massive first-principles computation.
As an example, consider the crystallographic group of a cube, including following
operations.
• Three rotations along x, y, or z axes by 90 degrees.
34
FIG. 2. The paradigm of the crystal symmetry, or the subgroup lattice.
• Exchange of x, y, and z axes ( Rotations along the diagonal in the cube by 120
degrees).
• The inversion and reflections are omitted for simplicity.
These rotations generate 24 operations in the cubic lattice to construct the point
group. Using GAP, The relationship of the inclusion in subgroups (the lattice subgroup) can be inquired and visualized(Fig.2). The point group of the cubic lattice
is located at the topmost node of the tree. We can trace the subgroups of lower
symmetries in descending along the branches and finally arrive at the trivial group
(the identity group), which is located at the root of the tree. This means that we
can explore and predict the alternation of the band structure (viz, the splitting of
the degenerated energy levels, or the opening of new band gaps) systematically, by
consulting with the “paradigm” of the crystal symmetry.
Now there arise following questions.
Question 1. How to make up actual crystal structures in correspondence with the
35
subgroups of lower symmetries?
Answer 1. The reduction of the symmetry may be occasioned, for example, by the
presence of alien atoms, or the occurrence of point defects in the perfect crystal. The
change in the electronic structure will be investigated by simple perturbation theory
with sufficient accuracy.
Question 2. In what positions alien atoms or point defects can be located? It is easy
to cause the great change the symmetry, or, to abolish it completely. However, for the
case with the small change in the symmetry, for example, that between neighboring
connected entries in the subgroup lattice, can we set up favorable atomic dispositions?
Answer 2. Choose an arbitrary point. Apply all of the operations of the subgroup
in consideration to this coordinate. The number of the generated points is equal to
the order of the subgroup. The set of these points is invariant with respect to the
operation of the subgroup. By putting alien atoms or defects in this set of points,
we can reduce the symmetry of the crystal, and, by continuing this way, descend
through the subgroup lattice.
Question 3. It is impossible for us to put alien atoms at arbitrary points in the
crystal. With the view of the material design, a certain restriction should be imposed
on the number of alien atoms. We would like to know the set of atomic sites, with a
limitation in number, and invariant by the subgroup operation.
Answer 3. Make use of the concept of the Wyckoff position.
The mathematical definition is given in the appendix. In a familiar word of
material sciences, the Wyckoff positions are the set of equivalent points in the unit
cell, which are transformed among themselves by symmetry operations. In general,
the total number of such points is the same as that of the symmetry operations (the
36
order of the point group). However, if the generating point of the Wyckoff position
remains fixed by certain operations of the point group, the total number of Wyckoff
positions will be fewer, which will be a factor of the order of the point group. In a
mathematical word, it is equal to the order of some factor group. In GAP program,
the Wyckoff position is easily computed from a “Cryst” package.
In this example of the cubic lattice, the Wyckoff positions are classified as follows:
• A discrete point (0,0,0).
• A discrete point (1/2,1/2,1/2).
• The centers of the edges of the cube:(1/2,0,0),(0,1/2,0),(0,0,1/2).
• The centers of the faces:(1/2,1/2,0) (1/2,0,1/2),(0,1/2,1/2)
• A segment (x,0,0) (0<x<1),invariant by 90-degrees rotation along the x axis.
• A segment (0,y,0) (0<y<1), invariant by 90-degrees rotation along the y axis.
• A segment (0,0,z) (0<z<1), invariant by 90-degrees rotation along the z axis.
• A segment (x,1/2,1/2) (0<x<1),invariant by 90-degrees rotation along the x
axis.
• A segment (1/2,y,1/2) (0<y<1), invariant by 90-degrees rotation along the y
axis.
• A segment (1/2,1/2,z) (0<z<1), invariant 90-degrees rotation along the z axis.
• A segment (t,t,t), invariant 120-degrees rotations which exchange three axes.(In
a similar way, all of the diagonal lines of the cube belong to this type.)
• Several segments (not discreet points), connected to the vertexes, the centers
of the edges.
37
• Arbitrary points in the unit cell.
In table 9, the Wyckoff positions of the group of full symmetry of the cubic cell
and those of one of the subgroups are compared. These two groups have common
generators of the Wyckoff positions. However, in some of them, the number of
equivalent positions differs. By making use of such positions and stationing atoms in
them, we can switch the crystal symmetry. For example, take the Wyckoff positions
given in the 7-th low. They are in the orbit of the generating point (t, 0, 1/2).
The group of No.11 and No.10 are isomorphic to the symmetric group S4 and the
alternative group A4 respectively. In the full symmetry, given by the group of No.11,
the orbit is composed of 12 points. On the other hand, in the reduced symmetry of
the group No.10, these 12 points are split into two orbits, each of which includes 6
points. If atoms (say, of type W) are positioned in all of 12 points in the cubic lattice,
the crystal symmetry remains unaltered as that of the group of No.11. However, if,
in the half part of them, at the Wyckoff positions of the Group No.10, the atoms are
replaced by those of type B, the crystal symmetry is reduced to the group of No.10.
Those Wyckoff positions are illustrated in Fig. 3, where two split orbits are shown
in B(black) and W(white) atoms.
B.
Control of band gaps by means of the Wyckoff positions
In this section, the example of the material design discussed in the previous section
is presented; in which ab-initio electronic structure calculations are executed, where
the symmetry of the crystal is artificially altered and the band gaps are controlled.
The exemplary calculation goes in the following way: in order to alter the crystal symmetry, certain coordinate points are chosen; and at these localities, we put
impurities, which interact to electrons with short-ranged potentials; these impurities
38
+--------------+
|
|
|
W
|
|
|
| B
TOP
B |
|
|
|
W
|
|
|
+--------------+--------------+--------------+--------------+
|
|
|
|
|
|
W
|
B
|
W
|
B
|
|
|
|
|
|
| B LEFT
B | W FRONT W | B RIGHT B | W REAR
W |
|
|
|
|
|
|
W
|
B
|
W
|
B
|
|
|
|
|
|
+--------------+--------------+--------------+--------------+
|
|
|
W
|
|
|
| B BOTTOM B |
|
|
|
W
|
|
|
+--------------+
FIG. 3. An illustration of split Wyckoff Orbit in the six faces of the cube.
are assumed to be neutral; therefore they do not alter the net charge. The change in
the symmetry of the crystal will be caused by the presence of impurities, the effects
of which are expressed by model potentials with adjustable parameters. We use the
following form of the scattering potential. It is given in the real space as
1
V (r) = C exp −
2
39
r
rloc
2 !
.
(29)
Group NO.11
(#G=24)
[0,0,0]
[1/2,1/2,1/2]
[0,1/2,1/2]
[1/2,0,0]
[t,0,0]
[t,1/2,1/2]
Size
1
1
3
3
6
6
[t,0,1/2]
12
[t,t,0]
[t,t,1/2]
[t,t,t]
[t1,t2,t3]
12
12
8
24
Group NO.10
(#G=12)
[0,0,0]
[1/2,1/2,1/2]
[0,1/2,1/2]
[1/2,0,0]
[t,0,0]
[t,1/2,1/2]
[0,t,1/2]
[1/2,t,0]
(Split in two)
not existent
not existent
[t,t,t]
[t1,t2,t3]
Size
1
1
3
3
6
6
6
6
0
0
4
12
TABLE 9. The Wyckoff positions in the cubic lattice. The left three columns gives the
generators and the right two show the number of equivalent points.
The reciprocal space representation is
1
1
2
V (K) = Z exp − (rloc K) ,
Ω
2
(30)
where Ω is the volume, required in the normalization of the plane wave √1Ω exp (iKr).
p
3
and Z = C (2π)3 (rloc
). There are two parameters: Z is the strength of the potential
and rloc is the effective range.
A simple example is taken from diamond band structure. The coordinate of the
carbon atoms are
(0 · V1 + 0 · V3 + 0 · V3 )
(1/4 · V1 + 1/4 · V3 + 1/4 · V3 ).
40
(31)
The lattice vectors are
a a
V1 =
, ,0 ,
2 2
a
a
V2 =
, 0,
,
2
2 a a
V3 = 0, ,
.
2 2
(32)
In this structure, the lowest eight eigenvalues at Γ point are shown in table 10.
There are two set of threefold degeneracy above and below the band gap, the energy
Γ
-2.11 19.58 19.58 19.58 25.14 25.14 25.14 33.14
Representation x.1 x.10 x.10 x.10 x.7 x.7 x.7 x.2
TABLE 10. The lowest eight eigenvalues at Γ point in the diamond structure (unit in eV).
levels of which are at 19.58 eV and 25.14 eV. (The irreducible representations to
them are x.10 and x.7 respectively, as given in the previous section.)
The band gap (in exact terms, the degeneracy of the eigenvalues) is modified by
the insertion of a single impurity in the crystal. Its position is chosen at (1/8 · V1 +
1/8 · V3 + 1/8 · V3 ). The point group of the genuine diamond structure ( of the order
of 48) is reduced to its subgroup (of the order of 12.) The symmetry operations of
the subgroup are shown in table 11.
The character table of this subgroup is given in table 12. The rows S.1,..., S.6
show the character values of the subgroup, and the rows x.7 and x.10 shows those of
the group of the full diamond symmetry (without impurity).
In the rows of the table, there are two relations: x.7=S.2+S.5 and x.10=S.1+S.6;
which are called matching relations.(In some mathematical context, they are referred
as branching rules.) These relations indicate the threefold degeneracy in the energy
levels in the full diamond symmetry split into two irreducible representations S.2 and
S.5, or S.1 and S.6 with the presence of an impurity. The character values of the
41
TABLE 11. The symmetry operations with the presence of an impurity. The rows (from the
second to the bottom ) show the rotation matrix R (R11, R12,..., R33) and the translation
T (T1, T2, T3).
1
2
3
4
5
6
7
8
9
10
11
12
R11 R12 R13 R21 R22 R23 R31 R32 R33 T1 T2 T3
1
0
0
0
1
0
0
0
1 0 0 0
0
1
0
0
0
1
1
0
0 0 0 0
0
0
1
1
0
0
0
1
0 0 0 0
0 -1
0 -1
0
0
0
0 -1 1/4 1/4 1/4
-1
0
0
0
0 -1
0 -1
0 1/4 1/4 1/4
0
0 -1
0 -1
0 -1
0
0 1/4 1/4 1/4
-1
0
0
0 -1
0
0
0 -1 1/4 1/4 1/4
0 -1
0
0
0 -1 -1
0
0 1/4 1/4 1/4
0
0 -1 -1
0
0
0 -1
0 1/4 1/4 1/4
0
1
0
1
0
0
0
0
1 0 0 0
1
0
0
0
0
1
0
1
0 0 0 0
0
0
1
0
1
0
1
0
0 0 0 0
TABLE 12. The character table of the diamond structure with an impunity. The characters
for each operation are shown here.
S.1
S.2
S.3
S.4
S.5
S.6
x.7
x.10
1
1
1
1
1
2
2
3
3
2
1
1
1
1
-1
-1
0
0
3
1
1
1
1
-1
-1
0
0
4
1
-1
-1
1
0
0
-1
1
5
1
-1
-1
1
0
0
-1
1
6
1
-1
-1
1
0
0
-1
1
7
1
-1
1
-1
-2
2
-3
3
8
1
-1
1
-1
1
-1
0
0
9
1
-1
1
-1
1
-1
0
0
10
1
1
-1
-1
0
0
1
1
11
1
1
-1
-1
0
0
1
1
12
1
1
-1
-1
0
0
1
1
identity element, shown in the second columns from the left, are the degeneracy of
the energy levels; so we can predict the threefold-degeneracy split into two different
levels, viz. twofold and single levels.
To inspect this, the actual computation is executed; where the impurity potential
is set to be a weak Gaussian form. For example, set the parameters in Eq.30 as
Z = −0.1a.u., rloc = 0.1a.u. The lowest eight eigenvalues are given in table 13. The
42
prediction by group theory is realized there.
Γ -2.18 18.86 19.67 19.67 25.04 25.18 25.18 33.13
TABLE 13. The lowest eight eigenvalues at Γ point in the diamond structure with an
impurity (unit in eV).
In the practice of the artificial material design, for example, the following demand
will arise:
Demand: How to substantiate the situation, in which the band gap is
altered, but the degeneracy of the energy levels is unaffected?
Let us proceed with the example of the diamond. In the diamond structure, there
are plural possible Wyckoff positions.(The computation can be done by GAP.) For
example, chose the position of (1/8V1 +1/8V2 +1/8V3 ). By the symmetry operations,
it is transformed into three other points such as: (5/8V1 + 1/8V2 + 1/8V3 ), (1/8V1 +
5/8V2 + 1/8V3 ), (1/8V1 + 1/8V2 + 5/8V3 ). If impurities are stationed at these four
points, there is no change in the point group; thus the degeneracy of the energy
spectrum remains unchanged. The computed energy spectrum is given in table
14, which guarantees the validity of the group theory. The band dispersions along
Γ(0, 0, 0) − X(2π/a, 0, 0) are shown in Fig. 4, in which that of the basic diamond
structure, and that with single impurity, and that with four impurities are compared.
Γ -2.41 18.83 18.83 18.83 25.11 25.11 25.11 33.07
TABLE 14. The lowest eight eigenvalues at Γ point in the diamond structure with four
impurities (unit in eV).
There is a certain noteworthy point in these two cases (of the single impurity,
and of the four impurities). In the former, the single impurity is located at(1/8V1 +
1/8V2 + 1/8V3 ); this point is one of the Wyckoff positions in the point group of the
43
crystal with the deteriorated symmetry; it remains fixed by any symmetry operation
and not transformed into other positions; however, this point also belongs to the
Wyckoff position of the perfect diamond crystal and there transformed to four points.
We can interpret this circumstance as follows: the former example corresponds to a
situation, where the only one of the four Wyckoff position in the latter case is chosen
so that the symmetry of the crystal could be reduced. In a similar way, (but in a
more general and liberal way, by inserting extra atoms or excavating vacancy site) we
can reduce the crystal symmetry for the purpose of controlling the band structure;
by doing this we can cleave into the degeneracy of the energy spectra and open extra
energy gaps.
C.
A possible plan of material design in the super-cell
Now we have seen that the band structure is controllable by the presence of impurities or defects, although the presented example is that of instructive purpose,
and not realistic, as an only model case for the sake of illustrating a possible form
of theoretical material design. The diamond structure is compressed in such a compact cell that the supernumerary atoms could not be actually stationed at arbitrary
places, such as at Wyckoff positions. Meanwhile, in recent decades, crystalline materials which have a large unit cell such as In3 O2 have been put into practical use.
At a glance, In3 O2 seems to have a huge complex structure. However, actually, it
is a super-cell composed from the stacking of body-centered lattice of In-O, with
oxygen vacancies. The band dispersion is almost free-electron-like; its behavior under the presence of doped Sn, viz. impurities in the cell, can be understood by the
group theoretical view[34]. Furthermore, the modern material composing technologies, such as epitaxial growth, or micro atom-scale probes, may enable us to realize
such microscopic disposition of atoms in the supercell structure, which will lead to
44
FIG. 4. The band dispersions in diamond structures altered by the presence of impurities,
along Γ-L direction. The left figure is that without impurity: the center, with a single
impurity: the right, with four impurities.
an actual phase of material design. In the super-cell, there is a more broad freedom
in the atomic arrangement, in which certain atoms could be replaced or removed so
that the periodical arrangement of impurities or defects should be permitted.
Now a principle of material design based on group theoretical view can be proposed.
45
1 Chose a certain crystal as a starting point: we may start from simple atomicand electronic- structures, such that we can execute the electronic structure
computation easily, as the primitive module of the design.
2 Consider an enlarged cell, composed of the stacking of the minimal unit cell.
(The crystal axes may be cut newly and a supercell should be created.)
3 We can determine the band structure and the band gap of the enlarged cell,
simply by folding the band structure of the minimal cell.(The folding of the
band structure is to re-distribute k-points of the old Brillouin zone to the new
one in accordance with the new periodicity and to redraw the band dispersion
by means of new k-points.)
4 Suppose a situation where the symmetry is reduced by the formation of the
supercell with atomic displacements. Check the possibility of the opening of
the extra band gaps (or the splitting of the degeneracy) at k-points with high
symmetries, such as at Γ points and the boundary of the Brillouin zone. The
group theoretical view, such as the character theory, the compatible relation,
the subgroup lattices, or the Wyckoff positions, should be fully made use of.
The change in the band structure may be quantitatively evaluated by simple
perturbation theory. These sorts of trials-and-errors, by tracing the sequence of
crystal symmetry, will help us to search crystal structures with more desirable
band structure.
VI.
TECHNICAL DETAILS
In the preceding sections of the article, the basics of the usage of computer algebra, concerning the computation of group theoretical properties, are explained and
applied to the analysis of the wave-functions. Moreover, the possible scheme for
46
controlling band gaps, by means of tracing the sequence of subgroups, which corresponds to the reduction of crystal symmetry, is suggested through simple electronic
structure computations. Meanwhile, in this part, the topics are rather promiscuous
and technical, such as the computation of characters of the group of the supercell
by means of “semi-direct product” which is an effective way to extend groups or the
special treatment for non-symmorphic crystal, and so on. Moreover, “mathematical
first principle” of crystal formation is demonstrated based on the cohomology theory,
which shall explain to us how specific types of crystals, finite in number, can exist
in accordance with the actions of certain point groups. The topics might arrive at
somewhat higher mathematics, that may be not well-known to physicists. However,
with the aid of computer algebra, one can conquer such hardships so that he might
actually execute necessary computations. As for the fundamental knowledge of the
group theory and its application, the reader should consult with standard references,
such as Refs.[1–6] for the group theory, such as Refs.[7–9] for the representation theory, such as Refs.[10, 11] for the algorithm in the computer algebra. We make use of
the computer algebra package GAP[12], as well as in the previous sections.
A.
Determination of crystal symmetry
It is essential for us to determine the crystal symmetry in order to execute the
group theoretical analysis. The symmetry operations which fix the crystal axes and
atomic coordinates must be listed to construct the point group. The determination
algorithm of the point group is composed of the following two stages.
Stage 1. Determination of the point group of the lattice.
Let A = (A1 , A2 , A3 ) be the lattice vectors. There are two large categories of the crystal system; the one of which is hexagonal, and the other
47
is cubic. In the former, rotations along the crystal axis by 60 degrees are
allowed; in the latter, the rotations of this kind are composed of those
by 90 degrees. The point group of arbitrary crystal system is given as
a subgroup which is reduced from the full hexagonal or cubic symmetries. The full set of the symmetry operations (i.e. rotation matrices) for
the hexagonal or cubic symmetry should be prepared, and from which
admissible operations are extracted.
For each operation R of the full symmetry group, compute the rotation
of the lattice vectors:
R · Ai (i = 1, 2, 3).
(33)
If the vector R · Ai protrudes out of the unit cell, it must be pulled
back according to the crystal periodicity. Hereafter the pulling back of
a vector x into the unit cell is denoted as P (x). For example, R · Ai =
l · A1 + m · A2 + n · A3 for integers l, m, n, then exactly P (R · Ai ) = 0.
Then store the operation R such as
kP (R · Ai )k ≤
for all Ai (=1,2,3), with a threshold value ( This is an attention needed
only in the floating-point computation.) From the stored set of admissible
operations {R}C , the point group of the crystal system is determined; in
this stage, however, the symmetry preservation in the atomic positions
is not yet checked.
Stage 2. Determination of the point group of the crystal.
The symmetry operations which fix atomic coordinates should be ex48
tracted now. The atom transformation table, the fractional translations
associated to each rotation, and the multiplication table should be provided at the same time. Let X and Y be entries in the set of atomic
positions {X}, and let T(X) and T(Y) be the corresponding type of
atoms.
For each R ∈ {R}L , and for each X ∈ {X}, compute R · X; and chose
every Y ∈ {X} such as T(X) = T(Y). Now the vector τc = P (R · X − Y )
is a candidate for the fractional translation associated to R. If τc is the
admissible fractional translation to R, the following conditions should be
satisfied.
Condition: for all X1 in the atomic coordinates, there exist a Y1
such that
kτc − P (R · X1 − Y1 )k ≤ ,
and
T (X1 ) = T (Y1 ).
If this condition is satisfied, store R and τc as a symmetry operation
{R|τc } in the point group of the crystal.
This algorithm works well if the unit cell is minimal, in the sense that the fractional
translation is uniquely determined; i.e. with respect to one symmetry operation of
{R|t}, other operations, such as {E|s}{R|t} = {R|t + s}, do not coexist in the point
group operations. If the unit cell is not minimal, in other words, if it is a supercell
(the accumulation of the minimal unit cell), the existence of plural τ for one R cannot
be negligible. The readers will be able to find and utilize subroutines or functions
purposed for the identification of the crystal symmetry (possibly with more efficacy
than the algorithm in this section)in the band calculation packages of one’s own.
49
B.
Symmetry operations in the reciprocal space
As for the plane wave, the symmetry operation in the real-space is given as
{RT |τT } exp(ikr) = exp ik RT−1 r − RT−1 τT
= exp (i(RT · k) · (r − τT )) .
(34)
Consequently, it is more convenient to give operations with respect to reciprocal
lattice vectors in the wave number space. If the wave-number-vector k is expressed
as k1 · G1 + k2 · G2 + k3 · G3 , the rotation matrix for coefficients in the rotated vector
RT · k is given as
0
R11
0
R12
0
R13
k
1
0
0
0
R21 R22
k2 .
R23
0
0
0
k3
R31 R31 R33
(35)
Thus the rotation matrix for the coefficients is computed from that in the Euclidean
space as:
R0 := (a1 , a2 , a3 )−1 RT (a1 , a2 , a3 )
T
(36)
where a1 , a2 , a3 are primitive lattice vectors.
For the case of the face-centered unit cell, the gap computation is done as:
gap> A:=[ [ 0, 1/2, 1/2 ], [ 1/2, 0, 1/2 ], [ 1/2, 1/2, 0 ] ];;
gap> RM:=[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ];;
gap> TransposedMat(A^-1*TransposedMat(RM)*A);
[ [ -1, 0, 0 ], [ -1, 0, 1 ], [ -1, 1, 0 ] ]
The rotation matrices in the reciprocal space, as well as those in the real space,
are given in the appendix for cubic and hexagonal symmetries.
50
C.
The computation of the compatibility relation
The irreducible representation of the diamond structure shows branching if the
k-point moves from Γ point to non-zero k-point. The situation of the branching (the
compatibility relation) can be checked by GAP. For this purpose, the computation
goes as follows:
• Compute the character table for the small group. Then, for each representation,
evaluate the characters of elements in the small group, and store them in the
lists S1 , S2 , ..., SN . (Let N be the number of the representations of the Small
group.)
• Compute character values of elements in the small group, using the irreducible
representation at the Γ point, and store them in the lists G1 , G2 , ..., GM .(Let
M be the number of the representations of the point group.)
• If the j-th irreducible representation (G.j) branches into the composition of
those in the small (S.i) as follows
Gj =
N
X
Cji × Si
(37)
i=1
the coefficients are computed by the inner product of the lists as
Cji = Gj · Si / |#K|
where |#K| is the order of the small group.
For this purpose, this function can be used:
GetCompati:=function(A,B)
51
(38)
local repA,repB,AV,BV,CV;
repA:=Irr(A);
repB:=Irr(B);
if ( false=IsSubgroup(A,B) ) then
Print("A < B ; Inverse A and B."); return;
fi;
AV:=List(repA,y->List(Elements(B),x->x^y));
BV:=List(repB,y->List(Elements(B),x->x^y));
CV:=AV*TransposedMat(BV)/Size(B);
return CV;
end;
This function computes the characters of the group elements in the group “B”,
in two ways, those by the irreducible representation of group “A” and group “B”.
The results are stored in two lists “AV” and “BV”. Each row of these lists contains
the characters of each irreducible representation. By the computation of the inner
product between these lists, the ramification of the representations is computed.
Symmetry operations defined in wave-number space are used here. These matrices
are three-dimensional ones given in Table 45.(They are stored in the list “MT[ ]” in
the following computation.) The coordinate of the K point along Γ − X point is
given, as the coefficients of reciprocal lattice vectors.
gap> kx:=[0,1,1];
[ 0, 1, 1 ]
The symmetry operations which fix the k-point are extracted, and from them, the
small group is generated.
gap> FL:=Filtered(MT,i->i*kx=kx);
52
[ [ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ -1, 0, 0 ], [ -1, 0, 1 ], [ -1, 1, 0 ] ],
[ [ 0, -1, 1 ], [ 0, 0, 1 ], [ -1, 0, 1 ] ],
[ [ 0, 1, -1 ], [ -1, 1, 0 ], [ 0, 1, 0 ] ],
[ [ 0, 1, -1 ], [ 0, 1, 0 ], [ -1, 1, 0 ] ],
[ [ 0, -1, 1 ], [ -1, 0, 1 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 0, 1 ], [ 0, 1, 0 ] ],
[ [ -1, 0, 0 ], [ -1, 1, 0 ], [ -1, 0, 1 ] ] ]
We can directly construct the small group from the above list of the output. For
the purpose of the comparison of the representations, however, the point group and
the small group must be expressed by common abstract symbols such as “m1,m2,...”,
which are employed to make up the multiplication table.(There is an alternative: the
point group could be constructed from the 48 matrices of symmetry operations in
wave-number space.)
gap> FL:=List(FL,i->Position(MT,i));
[ 1, 2, 19, 20, 27, 28, 41, 42 ]
gap> FL:=List(FL,i->Elements(G)[i]);
[ m1, m2, m19, m20, m27, m28, m41, m42 ]
gap> GS:=Group(FL);
<group with 8 generators>
The character table of the full point group GN is given as:
gap> ConjugacyClasses(GN);
[ m1^G, m2^G, m5^G, m13^G, m14^G, m25^G, m26^G, m29^G,
m37^G, m38^G
]
gap> Display(Irr(GN));
53
[ [
1,
1,
1,
1,
1,
1,
1,
1,
1,
1 ],
[
1,
1,
1,
-1,
-1,
-1,
-1,
-1,
1,
1 ],
[
1,
1,
1,
-1,
-1,
1,
1,
1,
-1,
-1 ],
[
1,
1,
1,
1,
1,
-1,
-1,
-1,
-1,
-1 ],
[
2,
2,
-1,
0,
0,
-2,
-2,
1,
0,
0 ],
[
2,
2,
-1,
0,
0,
2,
2,
-1,
0,
0 ],
[
3,
-1,
0,
-1,
1,
-3,
1,
0,
1,
-1 ],
[
3,
-1,
0,
-1,
1,
3,
-1,
0,
-1,
1 ],
[
3,
-1,
0,
1,
-1,
-3,
1,
0,
-1,
1 ],
[
3,
-1,
0,
1,
-1,
3,
-1,
0,
1,
-1 ] ]
The character table of the small group is as follows:
gap> List(ConjugacyClasses(GS),Elements);
[ [ m1 ], [ m2 ], [ m19, m20 ], [ m27, m28 ], [ m41, m42 ] ]
gap> Display(Irr(GS));
[ [
1,
1,
1,
1,
1 ],
[
1,
1,
-1,
-1,
1 ],
[
1,
1,
-1,
1,
-1 ],
[
1,
1,
1,
-1,
-1 ],
[
2,
-2,
0,
0,
0 ] ]
The branching relations are computed now:
gap> Display(GetCompati(GN,GS));
[ [
1,
0,
0,
0,
0 ],
[
0,
1,
0,
0,
0 ],
[
0,
0,
1,
0,
0 ],
[
0,
0,
0,
1,
0 ],
54
[
0,
1,
0,
1,
0 ],
[
1,
0,
1,
0,
0 ],
[
1,
0,
0,
0,
1 ],
[
0,
0,
0,
1,
1 ],
[
0,
0,
1,
0,
1 ],
[
0,
1,
0,
0,
1 ] ]
This output list should be read as the indication of the following branching relation:
POINT GROUP
SMALL GROUP
G.1
->
S.1
G.2
->
S.2
G.3
->
S.3
G.4
->
S.4
G.5
->
S.2 + S.4
G.6
->
S.1 + S.3
G.7
->
S.1 + S.5
G.8
->
S.4 + S.5
G.9
->
S.3 + S.5
G.10
->
S.2 + S.5
D.
The character table at the boundary of the Brillouin zone
1.
The necessity of the special treatment for the non-symmorphic crystal
In general non-zero k-points, the wave-functions are apportioned to representations different to that of the Γ point. The new representations are deduced from
subgroups of the point group; the subgroup admissible to each k-point must be
55
composed of operations which fix that k-point. The invariance on the k-point by
symmetry operations should be kept up to the periodicity of the wave-space. Thus
at the boundary of the Brillouin zone, different but equivalent k-points are mingled
together in the representation. In general, it is sufficient to choose the subgroup
which connects equivalent k-points and compute the representation in the same way
as in Γ point. However, in special cases, this naive approach inevitably fails, owing
to the existence of two different kinds of crystal structures, i.e. the distinction between the symmorphic and non-symmorphic ones. Let us review this troublesome
situation.
The matrix representation could be written as
Dk ({α|τ }) = exp(−ik · τ )Γ(α).
(39)
Then the product is
Dk ({α1 |τ1 })Dk ({α2 |τ2 }) = exp(−ik·(τ1 +τ2 ))Γ(α1 )Γ(α2 ) = exp(−ik·(τ1 +τ2 ))Γ(α1 α2 ).
(40)
The product should be equal to this:
Dk ({α1 · α2 |α1 · τ2 + τ1 }) = exp(−ik · (α1 · τ2 + τ1 ))Γ(α1 α2 ).
(41)
These equations lead to the relation:
exp(i(α1−1 · k − k) · τ2 ) = 1.
(42)
Certain conditions are required so that the above relation should be valid:
1 The k point is inside the Brillouin zone; α1−1 does not move k (i.e. α1 · k = k).
56
At the Γ point, this condition is always satisfied.
2 The k point is located at the boundary of the Brillouin zone: the vector α1 ·k−k
may coincide with a certain reciprocal vector due to the periodicity on the wave
space. The subgroups of the operations of these kinds are called small groups.
In addition, the τ2 should be zero or a lattice translation vector in the real
space.
The crystal structure, in which the second condition always holds, is said to be
symmorphic. In certain types of crystals, however, there exist affine mappings with
non-zero τ2 , which are fractions of some lattice translational vectors. This type
of the crystal is called non-symmorphic. Indeed, the diamond structure is nonsymmorphic one, as can be seen in the symmetry operations for this crystal. For
the non-symmorphic case, therefore, the relation to be satisfied with respect to the
product of the matrix representation does not hold at the boundary of the Brillouin
zone. Thus it is necessary to employ special methods to treat this case. One of these
treatments was proposed by Herring[32]. The intuitive interpretation of this method
could be given in the following way. One should bear in mind these two points: First,
at the Γ point, there is no problem in the representation even in the non-symmorphic
case. Secondly, the k-point at the boundary of the Brillouin zone can be pulled back
at the Γ point, if the unit cell is extended. For example, in a one-dimensional crystal
with the lattice constant of a, the k-point k = π/a is at the boundary of the Brillouin
zone. However, when the doubled unit cell (with the lattice constant 2a) is assumed,
the k-point k = π/a agrees with the new primitive reciprocal lattice vector; hence
it is equivalent to the Γ point. The irreducible representation at π/a in the small
original cell can be deduced from that at the Γ point in the enlarged cell. Thus the
difficulty in non-symmorphic representation can be eluded. We shall examine the
actual computation of this method in the next section.
57
2.
Herring Method: An example
FIG. 5. Symmorphic and non-symmorphic one-dimensional crystals.
Assume a double array of atoms, as in the Fig. 5. This figure shows two types of
one-dimensional periodic lattice, consisted by double lines of atoms. In case I, two
rows forms one dimensional periodic system of squares. In case II, the two lows are
shifted by half of the lattice constants. The latter is of non-symmorphic type. The
58
symmetry operations which fix the k-point at k = π/a for the latter case is given as
e = {E|0}
(43)
c = {C2 |0} A rotation by 180 degrees along the origin
m = {σx |a/2} A reflection over x axis
g = {σy |a/2} A reflection over y axis
From the aforementioned reasons, the irreducible representation of this group is not
the proper one to represent the small group at k = π/a.
According to the Herring method, instead of this, a factor group G/I should be
made use of, where G is a group of infinite order, composed of
e0 = {E|2na},
(44)
c0 = {C2 |2na},
m0 = {σx |a/2 + 2na},
g 0 = {σy |a/2 + 2na},
ē0 = {E|a + 2na},
c̄0 = {C2 |a + 2na},
m̄0 = {σx |3a/2 + 2na},
ḡ 0 = {σy |3a/2 + 2na},
and I is the normal subgroup composed of
e0 = {E|2na}.
(45)
This factor group is the point group in the unit cell extended twofold in the x
direction. By extending the unit cell, the k-points at the boundary of the original
59
Brillouin zone are pulled back at the Γ point in the new Brillouin zone. As the newly
allotted Γ-point is the zero vector, the mathematical difficulty in the representation
can be avoided. This situation should rather be interpreted as follows: the wavefunction at k = π/a has the periodicity of 2a, that is two times as large as the original
lattice width. On the other hand, the wavefunction at k = 0 in the original Brillouin
zone also matches into the periodicity of 2a.(Because originally it has periodicity
of a) In the representation at the new Γ point, those in the old Γ point and the
old boundary point are mingled with each other. These two different contributions,
however, can be distinguishable. Let us check the character of the elements of following types {E|τ }, viz. of the simple fractional translations in the extended cell, in
which the linear operation is the identity. With respect to the representation of the
boundary of the Brillouin zone of the original smaller cell, the representation matrix
should be represented as Dk=π/a ({E|a}) = exp(−i πa a)I d = −I d ; character of which
is equal to −d ( d is the dimensionality of the representation). From Table 16 we
can see the proper irreducible representation at k = π/a is given by Γ5 alone.
e0
ē0
c0
c̄0
m0
m̄0
g0
ḡ 0
e0
e0
ē0
c0
c̄0
m0
m̄0
g0
ḡ 0
ē0
ē0
e0
c̄0
c0
m̄0
m0
ḡ 0
g0
c0
c0
c̄0
e0
ē0
g
ḡ 0
m0
m̄0
c̄0
c̄0
c0
ē0
e0
ḡ 0
g0
m̄0
m0
m0
m0
m̄0
ḡ 0
g0
ē0
e0
c0
c̄0
m̄0
m̄0
m0
g0
ḡ 0
e0
ē0
c̄0
c0
g0
g0
ḡ 0
m̄0
m0
c̄0
c0
e0
ē0
ḡ 0
ḡ 0
g0
m0
m̄0
c0
c̄0
ē0
e0
TABLE 15. The Multiplication table of G/I. The first column and the top row show the
representatives in G of the elements in G/I. The elements in the table are the products
between the element a in the first column and the element b in the top row, a · b.
The prediction by the character table can be verified. The electronic structure
calculation is executed to the mismatched double allay of Na atoms: The two atoms
60
Γ1
Γ2
Γ3
Γ4
Γ5
e0
1
1
1
1
2
ē0 [c0 ,c̄0 ] [m0 ,m̄0 ] [g’, ḡ 0 ]
1 1
1
1
1 1
-1
-1
1 -1
1
-1
1 -1
-1
-1
-2 0
0
0
TABLE 16. The character table. The top row shows the elements in the conjugacy classes.
The rows below shows the characters in the five irreducible representations.
are placed at ±(1/4ax +1/8ay ) in the unit cell with ax = (a, 0, 0), ay = (0, 4a, 0), az =
(0, 0, 5a) and a = 8.1au. This computation affirms that every pair of eigenvalues at
the boundary of the Brillouin zone shows the double-fold degeneracy.
k=0 -2.775 -2.124 -0.506 -0.479 0.139 0.753 0.780 1.054
k= π/a -0.717 -0.717 -0.072 -0.072 1.501 1.501 1.528 1.528
TABLE 17. Energy spectrum at k=0 and π/a in the mismatched double allay of Na atoms
(unit in eV).
The above example is a simple one-dimensional model. For a non-symmorphic
three-dimensional crystal, the irreducible representation for the k-point at the boundary of the Brillouin zone can be computed in a similar way. For this purpose we should
set up a supercell structure; its periodicity should be equal to that of the wavefunction at the boundary k-point in consideration; the symmetry operations to be taken
into account of must be confined to ones which fix the boundary k-point up to the periodicity of the old Brillouin zone of the small cell. (To consider the full symmetry of
the new super-cell may be superfluous); the irreducible representation at the Γ point
in the supercell is the mixture of the wavefunctions at the Γ point and the k-point
in consideration, but we can distinguish them by the comparison of the character
of the pure fractional translation. The example of the three-dimensional case is the
irreducible representation of diamond structure in the face-centered cell at X point
61
kX = [0, 2π/a, 0](= 1/2G2 + 1/2G3 ). The necessary super-cell is the cubic unit cell.
We can check this: the primitive lattice vectors in the cubic cell is expressed by those
in the face-centered cell as â1 = a1 + a2 − a3 , â2 = a1 − a2 + a3 and â3 = a1 + a2 − a3 ;
so there are relations as â1 · kX = 0,â2 · kX = 0 and â1 · kX = 2π. These relations
indicate that the wavefunction at kX in the face-centered cell fits itself into the cubic
supercell, in the sense that the translations by the primitive lattice vector in the
cubic cell causes no phase shift in the wavefunction. The symmetry operations for
the small group at kX are constructed from those which fix kX (kX → kX ) and those
which invert kX (kX → −kX ), due to the periodicity in the wave-number space. In
contrast, the full symmetry operations in the cubic cell couples six wavefunctions at
[±2π/a, 0, 0],[0, ±2π/a, 0],and [0, 0, ±2π/a]. As the wave-vector kX cannot be fixed,
the corresponding irreducible representation for the full symmetry is not adequate
for the small group at kX . (Elsewhere in this article, the point group of the full
symmetry operations in the cubic unit cell is computed; from which the small group
at kX can be derived as a subgroup.)
E.
The semidirect product and the induced representation
In the previous section of this article, concerning the computation of the character
tables for the supercell structures, we have to remake the multiplication table of the
extended group by the presence of the fractional translation. However, if we make use
of the mathematical idea of the semidirect product, we can treat such a problem more
effectively. In this section, this concept is explained and the possible computation
GAP is presented.
The semidirect product A n G is obtained by extending a group G by an another
group A in the following way. The elements in the new group are given as the
62
Cartesian product of each group:
{(a, g); a ∈ A, g ∈ G}.
(46)
The multiplication is defined as
(a1 , g1 ) · (a2 , g2 ) = (a1 g1 (a2 ), g1 g2 ),
(47)
in which the notation g1 (a2 ) means the automorphism in A acting on a2 , induced by
g1 in G. The typical example of this is the affine mapping in crystallographic group.
The definition is given as follows:
OT = {RT |τT } : r → RT · r + τT (Affine mapping),
(48)
OT · OS = {RT · RS |τT + RT · τS }.
(49)
If the group A is Abelian, the representation of the semidirect product is given by
χV (a, g) =
1
|Gx |
X
x(h(a))χU (hgh−1 ).
(50)
h∈G;hgh−1 ∈Gx
Here Gx is the set of elements in G, such that the character x in A is fixed by the
automorphism induced by elements in Gx . All of the irreducible representations of the
semidirect product are exhausted by this formula through possible combinations of
the character x in A and the irreducible representation χU of Gx .(For the theoretical
background, consult references of pure mathematics[8, 35].)
One of the simplest example is C3v = Group((1, 2, 3), (1, 2)) ≡ S3 . One can check
the validity of this formula, taking C3 = Group((1, 2, 3)) as A and I = Group((1, 2))
63
as G. Let us follow this computation.
The character tables of those three groups are given as
[()] [(1,2),(2,3),(3,1)] [(1,2,3),(1,3,2)]
χ1 1
1
1
χ2 1
-1
1
χ3 2
0
1
TABLE 18. The character table of C3v .
[eA =()] [aA =(1,2,3)] [a−1
A =(1,3,2)]
x1
1
1
1
1
ω
ω2
x2
2
x3
1
ω
ω
TABLE 19. The character table of C3 , where ω = exp(2π/3).
y1
y2
[eG =()] [gG =(1,2)]
1
1
1
-1
TABLE 20. The character table of I.
The automorphism on A by G is defined as:
eG (eA ) = eA ,
eG (aA ) = aA ,
−1
eG (a−1
A ) = aA ,
gG (eA ) = eA ,
gG (aA ) = a−1
A ,
gG (a−1
A ) = aA .
64
(51)
As a trivial case, take x1 . Since Gx1 = G, y1 can be chosen as χU . Since G is
Abelian, i.e. hgh−1 = g, the formula is
χV (a, g) =
=
1
|Gx1 |
1
2
X
x1 (h(a))y1 (hgh−1 )
(52)
h∈G;hgh−1 ∈G
X
h∈G;hgh−1 ∈Gy1
1
x1 (h(a))y1 (g) = (1 + 1) = 1,
2
for all elements in the semidirect product.
Again take x1 . Since Gx1 = G,y2 can also be chosen as χU . The formula is
χV (a, g) =
1
2
X
x1 (h(a))y2 (hgh−1 ).
(53)
h∈G;hgh(−1) ∈Gx1
It gives
1
χV (eA , eG ) = (1 + 1) = 1,
2
(54)
1
χV (aA , eG ) = χV (a−1
A , eG ) = (1 + 1) = 1,
2
1
χV (eA , gG ) = (−1 − 1) = −1,
2
1
χV (aA , gG ) = χV (a−1
A , gG ) = (−1 − 1) = −1.
2
Take x2 . Since Gx2 = hIdi (the trivial group of the identity), it must be χU = 1.
The formula is
χV (a, g) =
X
x2 (h(a))χU (hgh−1 )
h∈G;hgh−1 ∈G
=
x2
X
x2 (h(a)) · 1.
h∈G;hgh−1 ∈G
65
x2
(55)
It gives
χV (eA , eG ) = (1 + 1) = 2,
(56)
2
χV (aA , eG ) = χV (a−1
A , eG ) = (ω + ω ) = −1,
χV (eA , gG ) = 0,
χV (aA , gG ) = χV (a−1
A , gG ) = 0.
The last two characters are zero, since for any h, hgh−1 = g ∈
/ hIdi, no terms are
summed up.
(N.B. When we take x1 , it seems plausible that we could also choose Gx1 =
hIdi.Then the formula is
X
χV (a, g) =
x1 (h(a))χU (hgh−1 )
h∈G;hgh−1 ∈G
x1
X
=
(57)
x1 (h(a)) · 1.
h∈G;hgh−1 ∈<Id>
It gives
χV (eA , eG ) = 2,
(58)
χV (aA , eG ) = χV (a−1
A , eG ) = 2,
χV (eA , gG ) = 0,
χV (aA , gG ) = χV (a−1
A , gG ) = 0.
However, this result is not irreducible. It is essential to use the maximal set of Gx .)
There are other combinations of characters including x3 of C3v . However, they
only produce the redundant results.
The aforementioned computations are tedious and complicated; a naive imple66
mentation by means of GAP programming language could be given as in appendix
G.
In GAP, the semidirect product can be constructed by a built-in function.
gap> A=Group((1,2,3));
Alt( [ 1 .. 3 ] )
gap> G=Group((1,2));
Sym( [ 1 .. 2 ] )
The automorphism in A is evaluated as follows:
gap> au:=AutomorphismGroup(A);
<group of size 2 with 1 generators>
gap> auE:=Elements(au);
[ [ (1,2,3) ] -> [ (1,2,3) ], [ (1,2,3) ] -> [ (1,3,2) ] ]
gap> gens:=GeneratorsOfGroup(G);
[ (1,2) ]
The automorphism induced by G is defined as follows: (The generator (1,2) in G
causes the automorphism in A as (1,2,3)->(1,3,2). Indeed (1,2,3)*(1,3,2)=1: this is
the inversion.)
gap> hom:=GroupHomomorphismByImages(G,au,gens,[auE[2]]);
[ (1,2) ] -> [ [ (1,2,3) ] -> [ (1,3,2) ] ]
gap> p:=SemidirectProduct(G,hom,A);
Group([ (3,4,5), (1,2)(4,5) ])
gap> Irr(p);
[ Character( CharacterTable( Group([ (3,4,5), (1,2)(4,5) ]) ),
[ 1, 1, 1 ] ), Character( CharacterTable( Group([ (3,4,5), (1,2)
(4,5) ]) ), [ 1, 1, -1 ] ), Character( CharacterTable( Group([ (3,
4,5), (1,2)(4,5) ]) ), [ 2, -1, 0 ] ) ]
67
F.
The computation of the character table in the cubic diamond by GAP
In this section, the character table in the cubic diamond structure is computed by
GAP command for the semidirect product. The point group G of the basic diamond
structure is extended by the Abelian group of fractional translations N.
The point group G is given in the table in the appendix of this article, as the set of
48 three-dimensional matrices.(Although in the table accompanying fractional transformations are listed, the point group G is constructed solely from three-dimensional
rotation matrices; the fractional translations are omitted in the point group construction; they are included in the Abelian group N.) The group N is the factor group
F/T, where generators of F are
τ1 := (1/4, 1/4, 1/4), τ2 := (1/4, −1/4, −1/4), τ3 := (−1/4, −1/4, 1/4)
(59)
and those of T are
(1, 0, 0), (0, 1, 0), (0, 0, 1).
(60)
Using these generators, the group A and T are given as the vector space (or module)
over integers.
For the convenience, the group F/T can be expressed as another factor group
A/B, where
A := Group((1, 2, 3, 4), (5, 6, 7, 8), (9, 10, 11, 12));
(61)
and
B : = Group((1, 2, 3, 4)∧ 2 ∗ (5, 6, 7, 8)∧ 2, (5, 6, 7, 8)∧ 2 ∗ (9, 10, 11, 12)∧ 2,
(62)
(9, 10, 11, 12)∧ 2 ∗ (1, 2, 3, 4)∧ 2)
( : = Group((1, 3)(2, 4)(5, 7)(6, 8), (5, 7)(6, 8)(9, 11)(10, 12), (1, 3)(2, 4)(9, 11)(10, 12));
68
This definition corresponds to the relations of the fractional translations in F/I;
4 ∗ τ̄i := (0, 0, 0) [Identity]
(63)
2 ∗ (τ̄1 + τ̄2 ) = 2 ∗ (τ̄2 + τ̄3 ) = 2 ∗ (τ̄3 + τ̄1 ) ≡ (0, 0, 0) [Identity]
(64)
and
The gap computation goes as follows:
gap> A:=Group((1,2,3,4),(5,6,7,8),(9,10,11,12));
Group([ (1,2,3,4), (5,6,7,8), (9,10,11,12) ])
gap> B:=Group((1,2,3,4)^2*(5,6,7,8)^2,(5,6,7,8)^2*(9,10,11,12)^2,
(9,10,11,12)^2*(1,2,3,4)^2);
Group([ (1,3)(2,4)(5,7)(6,8), (5,7)(6,8)(9,11)(10,12), (1,3)(2,4)
(9,11)(10,12) ])
gap> N:=FactorGroup(A,B);
Group([ f1, f2, f3, f4 ])
gap> NaturalHomomorphismByNormalSubgroup(A,B);
[ (1,2,3,4), (5,6,7,8), (9,10,11,12) ] -> [ f3, f2, f1 ]
The last command returns the homomorphism from A to another group whose
kernel is B: in other words, gives the mapping of A onto A/B. The tree generators
(1,2,3,4),(5,6,7,8),(9,10,11,12) in A are mapped to f3,f2,f1 in A/B.
gap> N.1;
f1
gap> N.2;
f2
gap> N.3;
69
f3
gap> N.4;
f4
There is a newcomer f4; we must identify it. For this purpose, it is enough to
compute the multiplication table of for generators.
gap> NV:=[N.1,N.2,N.3,N.4];
[ f1, f2, f3, f4 ]
gap> TransposedMat([NV])*[NV];
[ [ f4, f1*f2, f1*f3, f1*f4 ], [ f1*f2, f4, f2*f3, f2*f4 ],
[ f1*f3, f2*f3, f4, f3*f4 ],
[ f1*f4, f2*f4, f3*f4, <identity> of ... ] ]
We can see the relation f1*f1=f2*f2=f3*f3=f4, which means that the element f4
corresponds to elements (1,2,3,4)^2, (5,6,7,8)^2, (9,10,11,12)^2, viz. the
fractional transformation (1/2,1/2,1/2).
The size of the fractional group is given as
gap> Size(C);
16
Concerning this, the volume of the parallelepiped spanned by τi is
gap> AbsoluteValue(Determinant(A));
1/16
Thus, in the unit cubic cell, the translations by τi generates 16 discrete points, the
number of which is equal to the order of F/I.
70
It is sufficient to define the automorphisms in N induced by G only for the generators of G: the candidates of which are two matrices, those of No.36 and No.48 in
the list of symmetry operations.
0 0 1
0 0 −1
M36 := 1 0 0 , M48 := 0 −1 0
0 −1 0
1 0 0
Let us check the operations of these two matrices on τi .
gap> A:=[[1,1,1],[-1,-1,1],[-1,1,-1]]/4;
[ [ 1/4, 1/4, 1/4 ], [ -1/4, -1/4, 1/4 ], [ -1/4, 1/4, -1/4 ] ]
gap> A:=TransposedMat(A);
[ [ 1/4, -1/4, -1/4 ], [ 1/4, -1/4, 1/4 ], [ 1/4, 1/4, -1/4 ] ]
gap> M36:=[[0,0,1],[1,0,0],[0,-1,0]];
[ [ 0, 0, 1 ], [ 1, 0, 0 ], [ 0, -1, 0 ] ]
gap> M48:=[[0,0,-1],[0,-1,0],[1,0,0]];
[ [ 0, 0, -1 ], [ 0, -1, 0 ], [ 1, 0, 0 ] ]
The generators are set to be the transposed forms of above matrices.
gap> TM36:=TransposedMat(M36);
[ [ 0, 1, 0 ], [ 0, 0, -1 ], [ 1, 0, 0 ] ]
gap> TM48:=TransposedMat(M48);
[ [ 0, 0, 1 ], [ 0, -1, 0 ], [ -1, 0, 0 ] ]
gap> (A*TM36*A^-1);
[ [ 0, -1, 0 ], [ 0, 0, -1 ], [ -1, 0, 0 ] ]
gap> (A*TM48*A^-1);
[ [ 0, 1, 0 ], [ 0, 0, 1 ], [ -1, -1, -1 ] ]
71
(65)
The last two computations (such as A · M · A−1 ) give the transformations on the
row vectors (τi ) in A.
The operations by M36 and M48 are given as:
M 36 · τ1 = −τ2
(66)
M 36 · τ2 = −τ3
M 36 · τ3 = −τ1
M 48 · τ1 = τ2
M 48 · τ2 = τ3
M 48 · τ3 = −τ1 − τ2 − τ3
Remark. In the definition of GAP, the semidirect product of a group N with a group
G (acting on N via a homomorphism α from G into the automorphism group of N) is
α
the Cartesian product G × N with the multiplication (g, n)·(h, m) = (gh, nh m), since
in GAP all groups act from the right. On the other hand, the familiar style for the
physicists is {G|n} · {H|m} = {G · H|n + G ∗ m}, as in the crystallography; the point
group acts from the left. However, these two different definitions are equivalent. This
is based on this: the crystallography, the symmetry operation {M |t} is represented by
an augmented matrix, which can take two forms. The one is the matrix of the form
[ M 0 ]
[ t 1 ]
acting from the right on row vectors (x, 1). The another is
[ M t ]
[ 0 1 ]
acting from the left on column vectors (x, 1). These two definitions are convertible by
72
transposition and one should interpret the GAP computations by the former definition. The generators used here, (M36, M48), should be replaced with their transposed
forms.
We can identify τi to three permutations:
τ1 := (9, 10, 11, 12) ≡ f1
(67)
τ2 := (5, 6, 7, 8) ≡ f2
τ3 := (1, 2, 3, 4) ≡ f3
Other identifications will do. But we use the identification consistent with the
result of the “NaturalHomomorphismByNormalSubgroup” command.
It suffices that the automorphism in N induced by G is defined at generators of
groups.
gap> gens:=GeneratorsOfGroup(N);;
gap> CC1:=[gens[2]^-1,gens[3]^-1,gens[1]^-1,gens[4]];;
gap> CC2:=[gens[2],gens[3],(gens[1]*gens[2]*gens[3])^-1,gens[4]];;
gap> hom1:=GroupHomomorphismByImages(N,N,gens,CC1);;
gap> hom2:=GroupHomomorphismByImages(N,N,gens,CC2);;
The command “GroupHomomorphismByImage(G, H, gens, imgs)” defines a homomorphism from the group G to the group H through the mapping of the list “gens”
of the generators of G to the list “imgs” of the images in H. If the definition of the
mapping is irrelevant, the program returns a “fail” message.
gap> CC3:=[gens[3]^-1,gens[1]*gens[3]^-1,gens[2]*gens[3]^-1];;
gap> hom3:=GroupHomomorphismByImages(N,N,[gens[1],gens[2],gens[3]],CC3);
fail
73
Let G be the point group of the minimal unit cell, generated from the matrices
TM36 and TM48. By those preparations, we can make the semidirect product GnN .
gap> au:=AutomorphismGroup(N);;
gap> G:=Group(TM36,TM48);;
gap> GTON:=GroupHomomorphismByImages(G,au,GeneratorsOfGroup(G),[hom1,hom2]);;
gap> GPN:=SemidirectProduct(G,GTON,N);;
gap> irrep:=Irr(SemidirectProduct(G,GTON,N));;
Once the semidirect product G n N is computed, the representation of the point
group in the cubic diamond can be deduced from a suitable subgroup of this. For
this purpose let us see the embedding of G and N into the semidirect product. The
command “Embedding” creates the embedding from G and N into the semidirect
product, and the command “Projection” creates the projection from the semidirect
product to G.
gap> embd1:=Embedding(GPN,1);
CompositionMapping( [ f1, f2, f3, f4, f5 ] -> [ f1, f2, f3, f4, f5 ],
CompositionMapping( Pcgs([ (5,8)(6,7), (1,2)(3,5,7,4,6,8), (3,7,6)(4,8,5),
(1,6)(2,5)(3,7)(4,8), (1,3)(2,4)(5,8)(6,7) ]) -> [ f1, f2, f3, f4, f5 ],
<action isomorphism> ) )
gap> embd2:=Embedding(GPN,2);
[ f1, f2, f3, f4 ] -> [ f6, f7, f8, f9 ]
gap> GeneratorsOfGroup(G);
[ [ [ 0, 1, 0 ], [ 0, 0, -1 ], [ 1, 0, 0 ] ],
[ [ 0, 0, 1 ], [ 0, -1, 0 ], [ -1, 0, 0 ] ] ]
gap> List(GeneratorsOfGroup(G),g->Image(embd1,g));
[ f2, f1*f3*f5 ]
gap> GeneratorsOfGroup(N);
74
[ f1, f2, f3, f4 ]
gap> List(GeneratorsOfGroup(N),g->Image(embd2,g));
[ f6, f7, f8, f9 ]
gap> Projection(GPN);
[ f1, f2, f3, f4, f5, f6, f7, f8, f9 ] ->
[ [ [ 1, 0, 0 ], [ 0, 0, 1 ], [ 0, 1, 0 ] ],
[ [ 0, 1, 0 ], [ 0, 0, -1 ], [ 1, 0, 0 ] ],
[ [ 0, 0, -1 ], [ -1, 0, 0 ], [ 0, 1, 0 ] ],
[ [ -1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, -1 ] ],
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ] ]
The correspondence of the generators is given in table21.
Elements in G and N
Elements in G n N
TM36(in G)
f2
TM48(in G)
f1*f3*f5
f1(in N) (The image of (9,10,11,12))
f6
f2(in N) (The image of (5,6,7,8))
f7
f3(in N) (The image of (1,2,3,4))
f8
f4(in N) (The image of (1,3)(2,4) or (5,7)(6,8) or (1,3)(2,4))
f9
TABLE 21. The correspondence of the generators in groups G, N and G n N .
The generators of the point group of the cubic diamond structure are these elements:
{m36|τi }
{m48|0}
75
(68)
(τi can be any of three alternatives. The addition of pure fractional translations
{E|τi } is superfluous in this case.)
The corresponding subgroup can be constructed in this way:
Subgroup(GPN,[GPN.2*GPN.6,GPN.1*GPN.3*GPN.5])
or
Subgroup(GPN,[GPN.2*GPN.7,GPN.1*GPN.3*GPN.5])
or
Subgroup(GPN,[GPN.2*GPN.8,GPN.1*GPN.3*GPN.5])
These three definitions generate the same group so that any choice can be used.
(They are the subgroups of GPN, generated by the elements in the square brackets
of the second argument.)
G.
1.
Crystal lattice and Cohomology
A brief introduction to group cohomology
In the modern style, the crystal group can be described by the theory of the
cohomology, which is actually computable by the computer algebra.
At first, we construct the one-cohomology. Let K ∈ GL(n, Z) and define V :=
Rn /Zn . The one-cocycle is defined as
Z 1 (K, V ) := {ρ; k → V |ρ(1) = 0, ρ(gh) = ρ(g) + gρ(h) for all g, h ∈ K}.
(69)
The one-coboundary is defined as
B 1 (K, V ) := {ρ ∈ Z 1 (K, V |ρ(g) = gv − v for some v ∈ V }.
76
(70)
The one-cohomology is defined as their quotient:
H 1 (K, V ) := Z 1 (K, V )/B 1 (K, V )
(71)
We can see the vector system in the space group, which satisfies the condition
τ (gh) ≡ τ (g) + g · τ (h) mod Zn (mod lattice translations)
(72)
is a kind of cocycle. Furthermore, the one-coboundary is interpreted as the shift of
origin; we can observe this from the conjugation operation:
{1|s}−1 {g|t}{1|s} = {g|(gs − s) + t}
(73)
Thus, the vector system could be understood in the context of one-cohomology.
The above definition of the one-cohomology is derived from the general theory of
group cohomology. Let G be a finite group and M be a module. Let C n (G, M ) be
the set functions from Gn = (g1 , g2 , . . . , gn ) to M . The differential (dn : C n (G, M ) →
C n+1 (G, M ) is defined as
dn (f )(g0 , g1 , . . . , gn ) = g0 · f (g1 , . . . , gn )
n
X
+
(−1)j f (g0 , . . . , gj−2 , gj−1 gj , gj+1 , . . . , gn )
(74)
j
+(−1)n+1 f (g0 , g1 , . . . , gn−1 ).
It satisfies
dn ◦ dn−1 = 0.
(75)
The n-cocycle Z i (G, M ) is defined as Ker(dn ), and the n-coboundary is as Im(dn−1 ).
The n-th cohomology group is defined as H i (G, M ) := Z i (G, M )/B i (G, M ).
77
2.
The possible fractional translations with one point group
We make use of following matrices (as explained elsewhere in this article) in order
to define the crystallographic group in the diamond crystal lattice:
gap>M1:=[[0,0,1,1/4],[1,0,0,1/4],[0,-1,0,1/4],[0,0,0,1]];
gap>M2:=[[0,0,-1,0],[0,-1,0,0],[1,0,0,0],[0,0,0,1]];
gap>S:=AffineCrystGroupOnLeft([M1,M2]);
gap>P:=PointGroup(S);
In this definition, the fractional translation is given a priori. With regard to
this, there is a question: what is the possible fractional translation compatible to
the point group of the diamond crystal? Is it uniquely determined? Let us inspect
this problem. For example, take the face-centered lattice. For simplicity in the
following computation, the symmetry operations are represented by the “lattice axes
coordinate”, by means of the vector system of the primitive lattice vectors. For this
new coordinate, the new point group Q is made from the group P by the conjugation
of the generators. (The symbol ^ is used for the conjugation in GAP.)
gap> A:=[[0,1/2,1/2],[1/2,0,1/2],[1/2,1/2,0]];
gap> Q:=P^A;
gap> Q.1;Q.2;
[ [ -1, 0, 0 ], [ 0, 0, -1 ], [ 1, 1, 1 ] ]
[ [ 0, 1, 0 ], [ 0, 0, 1 ], [ -1, -1, -1 ] ]
The generators of the space groups are defined now by means of unknown variables.(For the heuristic purpose, we venture to employ generators different to the
first definition.)
a:=Indeterminate(Rationals,1);
78
b:=Indeterminate(Rationals,2);
c:=Indeterminate(Rationals,3);
N1:=[[-1,0,0,0],[0,0,-1,0],[1,1,1,0],[0,0,0,1]];
N2:=[[0,1,0,a],[0,0,1,b],[-1,-1,-1,c],[0,0,0,1]];
Let us consider multiplications of two elements gi and gj in the point group, which
yields the identity. If gi and gj are represented by the products of the generators, the
multiplication gives a relation of the generators both in the point group and in the
space group. For example, take (Q.2*Q.1)^4=I. This relation is lifted up to the space
group as (N2*N1)^4≡ {I|T }, where T is a translation. The linear part of (N2*N1)^4
is the identity matrix; and the translation part [2*a+2*b,2*a+2*b,-2*a-2*b] must
be one of translation vectors. In general, the translation part in the relation of the
generators in the space group gives three linear equations with respect to unknown
variables “a”,“b”, and “c”. These equations should be equivalent to zero, modulo
translation vectors; viz. they should be integer-valued.
The relations in generators in Q shall be set up now. For this purpose, each elements in the group and its inverse are represented by generators respectively, so that
the product of them gives the relation in generators which yields the unity. In the
following, the symbols x1 and x2 indicate the two generators of the group Q. For simplicity, the trivial relation only representing the unity ( returned as <identity ...>
in the GAP result) is eliminated in the list.
gap> MQ:=Elements(Q);;
gap> rel:=List(MQ,i->Factorization(Q,i)*Factorization(Q,i^-1));;
gap> IDENTITY:=Factorization(Q,One(Q));;
gap> rel:=Filtered(rel,i->i<>IDENTITY);
[ (x2^-1*x1*x2*x1)^2, x1*(x2*x1^-1)^2*x1^-1, (x2*x1^2)^2,
(x2^-1*x1)^2, x1*x2^2*x1*x2^-1*x1*x2*x1^-1, x1^6,
79
(x2*x1*x2^2)^2, (x2^2*x1*x2)^2, (x2*x1*x2^2*x1^2)^2,
(x2^-1*x1)^2, (x1*x2^2*x1^2)^2, x2^4, x2*x1^3*x2^-1*x1^3,
x1*x2^4*x1^-1, x1^-1*(x2^-1*x1)^2*x1, x2^-1*x1*x2^3*x1,
(x2^2*x1^3)^2, (x2*x1^-1)^2, x2^4,
x2^-1*x1^-1*(x2^-1*x1)^2*x1*x2, (x1*x2*x1)^2, (x2^-1*x1)^2,
x1^-1*x2^-1*x1*x2^3*x1^2, x2^-1*x1*x2^3*x1,
x1^-1*(x2^-1*x1)^2*x1, x2^2*x1*x2^-1*x1*x2, (x2*x1)^4,
(x1^2*x2)^2, (x2*(x2*x1)^2)^2, (x2*x1*x2^2*x1)^2,
(x1*x2^2*x1*x2)^2, x2^-1*x1^3*x2*x1^3, x2^2*x1*x2^-1*x1*x2,
(x2*x1^-1)^2, (x2^-1*x1^-1*x2^-1*x1^2)^2 ]
The relations of the point group generators are lifted to the space group. The
matrix relations are stored in the list “relMat”; and the fractional translations are
taken out in the list “relv”. Now, if the fractional translation [a,b,c] is admissible,
the list “relv” must be integer-valued by the substitutions to the indeterminates.
(Apropos of this, we can use symbols of unknowns as “a”,“b”,“c” for the input,
whereas the GAP results are represented by “x_1”,“x_2”,“x_3”. Be careful not to
confuse indeterminates and generators of the group.) The GAP program cannot
invert the symbolically represented matrices N1 and N2, so we must prepare the
inverses for them (N1I and N2I) by the power products.
N1I:=N1^5;
N2I:=N2^3;
relMat:=[ (N2I^1*N1*N2*N1)^2, N1*(N2*N1I^1)^2*N1I^1, (N2*N1^2)^2,
(N2I^1*N1)^2, N1*N2^2*N1*N2I^1*N1*N2*N1I^1, N1^6,
(N2*N1*N2^2)^2, (N2^2*N1*N2)^2, (N2*N1*N2^2*N1^2)^2, (N2I^1*N1)^2,
(N1*N2^2*N1^2)^2, N2^4, N2*N1^3*N2I^1*N1^3,
N1*N2^4*N1I^1, N1I^1*(N2I^1*N1)^2*N1, N2I^1*N1*N2^3*N1,
80
(N2^2*N1^3)^2, (N2*N1I^1)^2, N2^4,
N2I^1*N1I^1*(N2I^1*N1)^2*N1*N2, (N1*N2*N1)^2, (N2I^1*N1)^2,
N1I^1*N2I^1*N1*N2^3*N1^2, N2I^1*N1*N2^3*N1,
N1I^1*(N2I^1*N1)^2*N1, N2^2*N1*N2I^1*N1*N2, (N2*N1)^4,
(N1^2*N2)^2, (N2*(N2*N1)^2)^2, (N2*N1*N2^2*N1)^2,
(N1*N2^2*N1*N2)^2, N2I^1*N1^3*N2*N1^3, N2^2*N1*N2I^1*N1*N2,
(N2*N1I^1)^2, (N2I^1*N1I^1*N2I^1*N1^2)^2 ];
relv:=List(relMat,i->[i[1][4],i[2][4],i[3][4]]);
[ [ 2*x_1+x_2+x_3, -2*x_1-x_2-x_3, -2*x_1-x_2-x_3 ],
[ -2*x_1-x_2-x_3, 0, 2*x_1+x_2+x_3 ], [ -x_2-x_3, 2*x_2, 2*x_3 ],
[ 2*x_1+x_2+x_3, -2*x_1-x_2-x_3, 0 ], [ 0, 2*x_1+x_2+x_3, 0 ],
[ 0, 0, 0 ], [ 0, 0, 2*x_1+x_2+x_3 ],
[ 0, 2*x_1+x_2+x_3, -2*x_1-x_2-x_3 ], [ 0, 2*x_1+4*x_2+2*x_3, 0 ],
[ 2*x_1+x_2+x_3, -2*x_1-x_2-x_3, 0 ],
[ -2*x_1-2*x_2, 2*x_1+2*x_2, 2*x_2+2*x_3 ], [ 0, 0, 0 ],
[ 2*x_1, 2*x_2, 2*x_3 ], [ 0, 0, 0 ],
[ -2*x_1-x_2-x_3, 0, 2*x_1+x_2+x_3 ],
[ 2*x_1+x_2+x_3, -2*x_1-x_2-x_3, 0 ],
[ 2*x_1+2*x_2, 2*x_2+2*x_3, -2*x_1-2*x_2 ],
[ 2*x_1+x_2+x_3, 0, 0 ], [ 0, 0, 0 ], [ 0, -2*x_1-x_2-x_3, 0 ],
[ x_2+x_3, -2*x_3, x_2+x_3 ], [ 2*x_1+x_2+x_3, -2*x_1-x_2-x_3, 0 ]
, [ -2*x_1-x_2-x_3, 0, 2*x_1+x_2+x_3 ],
[ 2*x_1+x_2+x_3, -2*x_1-x_2-x_3, 0 ],
[ -2*x_1-x_2-x_3, 0, 2*x_1+x_2+x_3 ],
[ 0, 2*x_1+x_2+x_3, -2*x_1-x_2-x_3 ],
[ 2*x_1+2*x_2, 2*x_1+2*x_2, -2*x_1-2*x_2 ],
[ -x_2-x_3, -x_2-x_3, 2*x_2 ],
81
[ 4*x_1+4*x_2+2*x_3, -x_2-x_3, -4*x_1-2*x_2 ],
[ 4*x_1+2*x_2, x_2+x_3, x_2+x_3 ],
[ -4*x_1-4*x_2-2*x_3, 4*x_1+2*x_2, x_2+x_3 ],
[ 2*x_1+2*x_2+2*x_3, -2*x_1, -2*x_2 ],
[ 0, 2*x_1+x_2+x_3, -2*x_1-x_2-x_3 ], [ 2*x_1+x_2+x_3, 0, 0 ],
[ 2*x_2+2*x_3, -2*x_1-2*x_3, 2*x_1+2*x_3 ] ]
Since there are duplicates in the linear equations of the fractional translation, the
unique ones should be extracted by
gap> Unique(Flat(relv*a^0));
[ 2*x_1+x_2+x_3, -2*x_1-x_2-x_3, 0, -x_2-x_3, 2*x_2, 2*x_3,
2*x_1+4*x_2+2*x_3, -2*x_1-2*x_2, 2*x_1+2*x_2, 2*x_2+2*x_3, 2*x_1,
x_2+x_3, -2*x_3, 4*x_1+4*x_2+2*x_3, -4*x_1-2*x_2, 4*x_1+2*x_2,
-4*x_1-4*x_2-2*x_3, 2*x_1+2*x_2+2*x_3, -2*x_1, -2*x_2,
-2*x_1-2*x_3, 2*x_1+2*x_3 ]
It is necessary to use an input such as “relv*a^0”, so that all entries in the list
should be polynomials, for the convenience of sorting, by dint of the fact a^0 is the
unity as a polynomial. Now we have following four equations to be integer-valued:
x_2+x_3
2*x_1
2*x_2
2*x_3
In Q/Z, there are four representatives of the fractional translations, as are given
below, with the primitive lattice vectors accompanied to them. (The translation
vectors are computed by the substitution in the list “relv”, from which the basis
vectors are extracted by the GAP built-in command “BaseIntMat(mat)”.)
82
Fractional translation
[ 0, 1/2, 1/2 ]
[ [ 1,
[ [ 1,
[ 1/2, 0, 0 ]
[ 1/2, 1/2, 1/2 ]
[ [ 1,
Translation vectors
0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ]
0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ]
1, 1 ], [ 0, 2, 0 ], [ 0, 0, 2 ] ]
TABLE 22. The fractional translations and the translation vectors.
To inspect the validity, by means of GAP built-in command, the possible space
groups with a given point group can be computed. In order to check the computed
fractional translations, it is enough to see the generators of the space group and the
primitive translation vectors.
gap> SS:=SpaceGroupsByPointGroupOnLeft(Q);;
gap> Display(List(SS,GeneratorsOfGroup));
[ [ [ [ -1, 0, 0, 0 ], [ 0, 0, -1, 0 ], [ 1, 1, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ -1, -1, -1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ],
[ [ [ -1, 0, 0, 0 ], [ 0, 0, -1, 0 ], [ 1, 1, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 0, 1, 0, 0 ], [ 0, 0, 1, 1/2 ], [ -1, -1, -1, 1/2 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ],
[ [ [ -1, 0, 0, 0 ], [ 0, 0, -1, 0 ], [ 1, 1, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 0, 1, 0, 1/2 ], [ 0, 0, 1, 0 ], [ -1, -1, -1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ],
[ [ [ -1, 0, 0, 0 ], [ 0, 0, -1, 0 ], [ 1, 1, 1, 0 ], [ 0, 0, 0, 1 ] ],
83
[ [ 0, 1, 0, 1/2 ], [ 0, 0, 1, 1/2 ], [ -1, -1, -1, 1/2 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ] ]
gap> Display(List(SS,TranslationBasis));
[ [ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ] ]
There are four space groups with the point group Q. The last three of the GAP
result is the same fractional translation in Table 22. However, their translational
vectors are different: the latter has larger translational vectors. This is because the
GAP computation implicitly includes the primitive translations of the unit cube in
generators of the space groups.
Readers may question about the relationship between the fractional translation
computed now and those of the definition of the space group given in the first part of
this section. The generators defined earlier are rewritten in the lattice axes coordinate
as:
gap> MM1:=[[-1,0,0,1/4],[0,0,-1,1/4],[1,1,1,1/4],[0,0,0,1]];
gap> MM2:=[[0,1,0,0],[0,0,1,0],[-1,-1,-1,0],[0,0,0,1]];
gap> S2:=AffineCrystGroupOnLeft([MM1,MM2]);
We can apply the shift of the origin to this space group by means of the conjugation
{E| − s}{R|t}{E|s} = {R|t + (R − I) · s}. If the matrix R − I is normal, we can annihilate the fractional translation. Let us apply such a conjugation to the generators.
(The command M^T is the abbreviation of the conjugation T−1 · M · T.)
84
gap> T:=[[1,0,0,1/8],[0,1,0,-3/8],[0,0,1,5/8],[0,0,0,1]];
gap> MM1^T;
[ [ -1, 0, 0, 0 ], [ 0, 0, -1, 0 ], [ 1, 1, 1, 0 ], [ 0, 0, 0, 1 ] ]
gap> MM2^T;
[ [ 0, 1, 0, -1/2 ], [ 0, 0, 1, 1 ], [ -1, -1, -1, -1 ], [ 0, 0, 0, 1 ] ]
Thus we can see that the shift of the origin generates the space group, which is equivalent to that composed by matrices N1 and N2 with [a,b,c]=[ -1/2, 1, -1 ]∼
=[1/2,0,0].
Incidentally, let us generate other representation of generators:
N1:=[[-1,0,0,a],[0,0,-1,b],[1,1,1,c],[0,0,0,1]];
N2:=[[0,1,0,0],[0,0,1,0],[-1,-1,-1,0],[0,0,0,1]];
The linear equations for the fractional translation are given as
[ 0, 4*x_3, -2*x_2-2*x_3, -2*x_2+2*x_3, -x_2+x_3, x_2-x_3,
2*x_2+2*x_3, -4*x_1-4*x_2-4*x_3, -2*x_1-2*x_3, -2*x_1+2*x_3,
-2*x_1-4*x_2-2*x_3, -2*x_1-2*x_2, -2*x_1-2*x_2-4*x_3,
-2*x_1-x_2-x_3, -2*x_1+2*x_2, 2*x_1-2*x_3, 2*x_1+2*x_3,
2*x_1-2*x_2, 2*x_1+x_2+x_3, 2*x_1+2*x_2+4*x_3,
2*x_1+4*x_2+2*x_3, 4*x_1+4*x_2+4*x_3 ]
From them, seven possible vectors of [a,b,c] (=[x_1,x_2,x_3]) are obtained as
[ 0, 1/2, 1/2 ], [ 1/4, 1/4, 1/4 ], [ 1/4, 3/4, 3/4 ],
[ 1/2, 0, 0 ], [ 1/2, 1/2, 1/2 ], [ 3/4, 1/4, 1/4 ],
[ 3/4, 3/4, 3/4 ] ]
We can see one of the fractional translations is [1/4,1/4,1/4].
Whereas the GAP computation goes as
85
gap> Q2:=Group(Q.1,Q.2);
Group([ [ [ -1, 0, 0 ], [ 0, 0, -1 ], [ 1, 1, 1 ] ],
[ [ 0, 1, 0 ], [ 0, 0, 1 ], [ -1, -1, -1 ] ] ])
gap> Q2=Q;
true
gap> SS2:=SpaceGroupsByPointGroupOnLeft(Q2);
[ <matrix group with 5 generators>, <matrix group with 5 generators>
, <matrix group with 5 generators>,
<matrix group with 5 generators> ]
gap> List(SS2,GeneratorsOfGroup);
[ [ [ [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ -1, -1, -1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ -1, 0, 0, 0 ], [ 0, 0, -1, 0 ], [ 1, 1, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ],
[ [ [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ -1, -1, -1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ -1, 0, 0, 0 ], [ 0, 0, -1, 1/2 ], [ 1, 1, 1, 1/2 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ],
[ [ [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ -1, -1, -1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ -1, 0, 0, 1/4 ], [ 0, 0, -1, 1/4 ], [ 1, 1, 1, 1/4 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ],
[ [ [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ -1, -1, -1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ -1, 0, 0, 1/4 ], [ 0, 0, -1, 3/4 ], [ 1, 1, 1, 3/4 ], [ 0, 0, 0, 1 ] ],
86
[ [ 1, 0, 0, 1 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 1 ], [ 0, 0, 1, 0 ], [ 0, 0, 0, 1 ] ],
[ [ 1, 0, 0, 0 ], [ 0, 1, 0, 0 ], [ 0, 0, 1, 1 ], [ 0, 0, 0, 1 ] ] ] ]
Four space groups are generated by GAP: non-zero fractional translations are three
in number:
[ 0, 1/2, 1/2 ], [ 1/4, 1/4, 1/4 ], [ 1/4, 3/4, 3/4 ]
Apparently the GAP result is lacking in some of fractional translations computed just
before by means of linear equations. The GAP result should rather be interpreted as
the basis vectors for the possible fractional translations. The other possible fractional
translations can be generated from this basis set.
It is well to remind readers of the point that the setting up of the admissible
fractional translations (which is called “vector system” in the mathematical context)
is nothing other than the computation of the one-cohomology. In the above example,
the fractional translation in the generator N1 is chosen to be zero from the outset.
It is possible to assume the non-zero translation in N1. However, a shift of the
origin can cancel a nonzero fractional translation as we have seen. The shift of the
origin could be caused by the conjugation, which corresponds to the choice of the
representative of the elements in the map of the one-cohomology.
3.
Group extension
The space group of the crystal is determined by the following short exact sequence
(or the group extension):
π
1→K→E→
− G→1
(76)
with the condition that there exists a normal subgroup of N in E which is isomorphic
to K and that E/K is isomorphic to G. The group E is called an extension of G by K.
87
The equivalence classes of extension of E are computed by a concept of the cohomological algebra, i.e. two-cohomology (denoted as H 2 (G, K) = Z 2 (G, K)/B 2 (G, K)).
In order to understand this concept, however, one should get acquainted with the
language of commutative algebra, especially concerning homological algebra, which
would necessitate long explanation. Thus if the reader would get interested in these
topics, he should try advanced mathematical textbooks. However, the viewpoint of
cohomological algebra might be useful in material designing, especially in studying
the algebraic aspects of crystallography. For example, the Bravais lattice in two- or
three-dimensional crystals can be classified by two-cohomology.
Remark. The construction of the two-cohomology is briefly stated here. For every
g ∈ G there are elements eg ∈ E with π(eg ) = g. However eg is not unique: another
element e0g behaves similarly, if e0g (eg )−1 ∈ ker(π) = K. Now we define a map
ϕg : k → eg ke−1
g , as an automorphism in K. The two-cocycle is defined by
f (g, h) = eg eh e−1
gh .
For each g, let us give an element kg ∈ K (as a homomorphism k(∗) from G to K).
Then the two-coboundary is defined as
−1
b(g, h) = kg ϕg (kh )kgh
.
(It will be represented as k(g) + g · k(h) − k(gh) when K is a module.)
Then we have
f (g, h)f (gh, k) = ϕg (f (h, k))f (g, hk)
and for another choice (e0g = kg eg ),
f 0 (g, h) = e0g e0h (e0gh )−1 = b(h, g)f (g, h).
88
In two-cohomology f (∗, ∗) and f 0 (∗, ∗) represent the same class. The elements of
E is uniquely written in the form aeg for some a ∈ K and g ∈ G. The product is
given as (aeg )(beh ) = aϕg (b)f (g, h)egh , which is the usual semidrect product when
f (g, h) = 1 for all g, h.
2
2
For simplicity, consider the two-dimensional
case.
hx, y|x2 =
Let P =
y = (xy) = 1i,
−1 0
1 0
and My =
. The action
which acts on T = Z2 through Mx =
0 1
0 −1
1
−1
x
y
y
and
on T is written as {x e1 = e−1
1 , e2 = e2 , e1 = e1 , e2 = e2 }. (e1 =
0
0
e2 = .
1
Let us extend this group, and, if possible, let us construct space groups which
have this group as their point groups. (This example is taken from the lecture notes
of Webb[36]. For theoretical details, consult with that article and references therein.)
There is a general procedure (Zassenhauss algorithm) to compute two-cohomology,
which is based on the following theorem.
Theorem 1. Let P be a finite group given by the representation
P = hx1 , x2 , ..., xd |r1 , r2 , ..., rt i .
Let n be the dimension of the space, and T be a module. From the relators ri , a
matrix Λ ∈ Mnt,nd (Z) is computed, and the elements Λij are defined as
ri − 1 =
d
X
Λij (xj − 1) (i = 1, ..., t).
j=1
89
Then the two-cohomology is given by
H 2 (P, T ) ∼
= {x ∈ (RT )d |Λ · x ∈ T t }/(Ker Λ + T d ).
(77)
This definition says that the representative elements in two-cohomology are given as
(v1 , · · · , vd ) , the concatenation of the vectors in RT n , (from which the component
in Ker Λ + T d should be factored out), so as to be
v
1
..
Λ . ∈ T t.
vd
(78)
The integer matrix Λ can be converted into its Smith normal form
b1
b2
..
.
bl
0
..
.
(79)
0
with bi 6= 0, dividing bj for i < j. (A matrix A ∈ Mm,n (Z) can be converted into the
diagonal form as this, by means matrix operation P · A · Q with P ∈ GL(m, Z) and
Q ∈ GL(n, Z).)
Then H 2 (P, T ) ∼
= Z/b1 Z ⊕ · · · ⊕ Z/bl Z. Now the relators in P are lifted in the
90
extension of P by T as
r1
r2
= Λ · v (∀v ∈ H 2 (P, T )).
..
.
rt
(80)
The vector systems are given as {xi |vi } (i = 1, ..., d) in the crystallography.
Remark. This algorithm is a refined style of the computation of one-cohomology
given in the previous section. The relation such as xyxy = 1( in G) is lifted to
exyxy = 1E , as the relation in E. The latter is rewritten by two-cocycles
1E = exyxy = f (x, yxy)−1 ϕx (f (y, xy)−1 )ϕx (ϕy (f (x, y)−1 ))ex ey ex ey
It is furthermore rewritten as
ex ey ex ey = ϕx (ϕy (f (x, y)))ϕx (f (y, xy))f (x, yxy).
The right side of this equation lies in K. The formula of Λ̂(xy)2 in the algorithm
represents the translational part in ex ey ex ey by means of those of ex and ey , and it
must be in K(= T ). To find a v such as Λ · v ∈ T is to fix a representative of cocycle
f (∗, ∗) in two-cohomology.
91
In the aforesaid question, the matrix Λ is computed as
2
2
Λx ,x Λx ,y
Λ = Λy2 ,x Λy2 ,y
Λ(xy)2 ,x Λ(xy)2 ,y
0 0 0 0
0
x
+
1
0
0
= 0
y+1 =
0
xy + 1 xyx + x
0
2 0 0
0 2 0
0 0 0
0 0 0
(81)
0 0 0 0
From this, it can be concluded that H 2 (P, T ) = Z/2Z⊕Z/2Z. The homomorphism
representing the two-cohomology, which determines the extension of the group, is
generated from these vectors:
0
0
1e
1/2 0
0
2
2
,
=
,
1
0 1/2
0
e
1
2
0
0
(82)
The linear combination of them (with integer coefficients, of 0 or 1 ) admits four
possible mappings. The
in the extension E are obtained from the entries in
relators
2
Λ̂x
Λ(c1 · e1 + c2 · e2 ) = Λ̂y2 . They are listed as: They generate four different
Λ̂(xy)2
Λ̂x2
Λ̂y2
Λ̂(xy)2
type 1
type 2
type 3
type 4
0 · e1 + 0 · e2 0 · e1 + 1 · e2 0 · e1 + 0 · e2 0 · e1 + 1 · e2
0 · e1 + 0 · e2 0 · e1 + 0 · e2 1 · e1 + 0 · e2 1 · e1 + 1 · e2
0 · e1 + 0 · e2 0 · e1 + 0 · e2 0 · e1 + 0 · e2 0 · e1 + 0 · e2
TABLE 23. Four possible types of relators in the extensions.
92
extensions in table 24. The extensions of type 2 and type 3 are isomorphic, since
they are converted with each other by interchanging x and y, and e1 and e2 .
type
type
type
type
−1
x
y
y
1
x, y, e1 , e2 |x2 = y 2 = (xy)2 = [e1 , e2 ] = 1, x e1 = e−1
1 , e2 = e2 , e1 = e1 , e2 = e2
−1 x
2
2
2
x
y
y
2 x, y, e1 , e2 |x = e2 , y = (xy) = [e1 , e2 ] = 1, e1 = e1 , e2 = e2 , e1 = e1 , e2 = e−1
2
−1
x
y
y
,
e
=
e
,
e
=
e
,
e
=
e
3 x, y, e1 , e2 |y 2 = e1 , x2 = (xy)2 = [e1 , e2 ] = 1, x e1 = e−1
2
2
1
1
2
1
2
−1
4 x, y, e1 , e2 |x2 = e2 , y 2 = e1 , (xy)2 = [e1 , e2 ] = 1, x e1 = e1 , x e2 = e2 , y e1 = e1 , y e2 = e−1
2
TABLE 24. Four types of the extension. e1 and e2 should be assumed to be translations
{E|e1 } and {E|e2 }, rather than vectors.
We can see the relators of the extensions provide the relations which determine
the vector system in the non-symmorphic crystal, with the point group P .(To obtain
the vector system, it is enough to re-express x,y as the augmented matrix forms
of the space groups, with indeterminate fractional translations. The e1 and e2 are
also represented by matrices as translations. The equivalence relations in the entries
of the matrix products in the group multiplications will give the vector systems,
as was demonstrated elsewhere in this article. In that demonstration, the vector
system was also derived from cohomological algebra, not from two-cohomology, but
from one-cohomology.) From this simple example, we can observe some importance
of homological algebra, which provides us the criterion of the existence of the nonsymmorphic crystal, in accordance with the non-triviality of two-cohomology. On
the other hand, if the two-cohomology is trivial, or of zero-mapping, the extension
of the group simply becomes a semidirect product, which is always a symmorphic
crystal. In this case of zero two-cohomology, the exact sequence splits, in the sense
that there is a homomorphism s : G → H such as π ◦ s = idG . It can be proved that
an extension (or, the short exact sequence as above) is split if and only if the group
H is the semidirect product of N and G. (We should remember this: elsewhere in this
article, the point group of the non-symmorphic crystal is constructed by means of the
semidirect product. In that computation, the point group of the non-symmorphic
93
crystal is not directly obtained as a semidirect product but extracted as one of its
subgroups. So there is no contradiction against the above statement.)
Now the relationship is established among the symmorphic properties of crystals,
the splitting extension of a group, and the two-cohomology, as is shown in the table
below.
type
type
type
type
1 {Mx |0}, {My |0}, {E|e1 }, {E|e2 }
2 {Mx | 12 e2 }, {My |0}, {E|e1 }, {E|e2 }
3 {Mx |0}, {My | 21 e1 }, {E|e1 }, {E|e2 }
4 {Mx | 12 e2 }, {My | 21 e1 }, {E|e1 }, {E|e2 }
TABLE 25. The generators of the four types of the space groups classified by the twocohomology.
In getting through this section, two necessary points for the computations in GAP
are given. The first is how to derive the relators, or the finite representation: the
second is the computation of the Smith Normal form. (The two-cohomology can be
computed by GAP built-in command, for the case of a finite group G and a module
M , over a finite field. Because of this assumption, the GAP built-in command is not
applicable to the above example, in which the module M is Zn .)
A definition is now needed: a group G is polycyclic if there is a series of subgroups
G = C1 > C2 > · · · > Cn > Cn+1 = 1,
where Ci+1 is normal in Ci and Ci /Ci+1 is a cyclic group. The generators of G is
given by a sequence G := (g1 , . . . , gn ) such that Ci = hCi+1 , gi i. As a finite solvable
group, the point group of the crystal is polycyclic, which is composed from the
polycyclic generating system.(The point group of the crystal is always solvable, by
means of a famous theorem of Burnside, which relates the order of the group to the
solvability.) We have defined the point groups by means of matrix generators, or
94
the multiplication table. These groups can be converted into isomorphic polycyclic
groups, and then, polycyclic groups are converted into finitely presented groups with
relators. The conversion goes as follows:
gap> A:=[ [ 0, 0, 1 ], [ 1, 0, 0 ], [ 0, -1, 0 ] ];
gap> B:=[ [ 0, 0, -1 ], [ 0, -1, 0 ], [ 1, 0, 0 ] ];
gap> G:=Group(A,B);
Group([ [ [ 0, 0, 1 ], [ 1, 0, 0 ], [ 0, -1, 0 ] ],
[ [ 0, 0, -1 ], [ 0, -1, 0 ], [ 1, 0, 0 ] ] ])
gap> p:=Pcgs(G);
Pcgs([ [ [ 0, 0, 1 ], [ 0, 1, 0 ], [ -1, 0, 0 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, -1 ] ],
[ [ 0, 0, -1 ], [ -1, 0, 0 ], [ 0, 1, 0 ] ],
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ],
[ [ -1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, 1 ] ] ])
gap> iso:=IsomorphismFpGroupByPcgs(p,"g");
CompositionMapping( [ (1,4,6,8)(2,3,5,7), (1,4)(2,3)(5,7)(6,8),
(3,6,7)(4,5,8), (1,7)(2,8)(3,6)(4,5), (1,3)(2,4)(5,8)(6,7) ] ->
[ g1, g2, g3, g4, g5 ], <action isomorphism> )
gap> fp:=Image(iso);
<fp group of size 48 on the generators [ g1, g2, g3, g4, g5 ]>
gap> RelatorsOfFpGroup(fp);
[ g1^2*g5^-1*g4^-1, g2^-1*g1^-1*g2*g1*g5^-1*g4^-1,
g3^-1*g1^-1*g3*g1*g5^-1*g4^-1*g3^-1, g4^-1*g1^-1*g4*g1*g5^-1*g4^-1,
g5^-1*g1^-1*g5*g1*g5^-1*g4^-1, g2^2, g3^-1*g2^-1*g3*g2*g4^-1,
g4^-1*g2^-1*g4*g2, g5^-1*g2^-1*g5*g2, g3^3, g4^-1*g3^-1*g4*g3*g5^-1*g4^-1,
g5^-1*g3^-1*g5*g3*g4^-1, g4^2, g5^-1*g4^-1*g5*g4, g5^2 ]
95
The command “Pcgs(G)” returns the polycyclic generating system of G. (In
this example, the generators are accessible through the list p[1],...p[5].) The
command “IsomorphismOfFpGroupByPcgs(p,”g”)” gives the map which converts
the polycyclic generating system into a finitely represented group. As the image of
this isomorphism, the polycyclic group is given. However, the polycyclic generating
system is not minimal. To generate the minimal one, we should do as this:
gap> iso:=IsomorphismFpGroupByGenerators(G,GeneratorsOfGroup(G));
<composed isomorphism:[ [ [ 0, 0, 1 ], [ 1, 0, 0 ], [ 0, -1, 0 ] ],
[ [ 0, 0, -1 ], [ 0, -1, 0 ], [ 1, 0, 0 ] ] ]->[ F1, F2 ]>
gap> fp:=Image(iso);
<fp group of size 48 on the generators [ F1, F2 ]>
gap> RelatorsOfFpGroup(fp);
[ F2^4, (F1*F2^-1)^2, F1^6, (F1*F2*F1)^2 ]
gap> [G.2^4,(G.1*G.2^-1)^2,G.1^6,(G.1*G.2*G.1)^2];
[ [ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ],
[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ] ]
The smith normal form is also computed by GAP:
gap> M:=[[0,0,0,0],[0,2,0,0],[0,0,0,0],[0,0,2,0],[0,0,0,0],[0,0,0,0]];
gap> SmithNormalFormIntegerMat(M);
[ [ 2, 0, 0, 0 ], [ 0, 2, 0, 0 ], [ 0, 0, 0, 0 ], [ 0, 0, 0, 0 ],
[ 0, 0, 0, 0 ], [ 0, 0, 0, 0 ] ]
96
H.
Remarks to this section
In this section, the technical details which would be important in the actual application of the computational group theory in the field of solid-state physics are
explained. The chosen topics are as follows: the algorithm to determine the symmetry of the crystal; the computation of the compatibility relation, which is useful
in the discussion of the splitting of the energy spectra with the reduction of the
structural symmetry; the concept of the semidirect product, which is applicable to
the preparation of the character table of the supercell structure; the concept of the
group extension, by means of cohomology, which can be utilized to design the crystal lattice. The author hopes that the readers of this article would get familiar with
these technical or advanced topics in order to deepen the knowledge of group theory,
applicable to concrete problems in material science. The small programs presented in
this article can be executed on a desktop PC. They are composed in accordance with
the elementary formulas and definitions in the references and textbooks so that they
should be as readable as possible, thus they are not are always effective in speed, and
some part of them could be replaced by built-in functions in GAP. In the preceding sections of this article, the topics are taken from the solid-state physics. In the
forthcoming sections, however, the application to the quantum chemistry (applicable
to the isolated system, equipped with a certain symmetry, without periodicity) shall
be discussed in a parallel way as the solid state physics, where the importance of
the symmetry in the quantum dynamics of the molecule shall be illustrated in the
different light.
97
VII.
SYMMETRY IN C60
The C60 molecule is characterized by the peculiar structure, that of soccer-ball,
composed of 60 atoms, 90 bonds, and 12 pentagons, and 60 hexagons. The group
theoretical symmetry is of the icosahedral type, which contains 120 symmetric operations and 10 irreducible representations, according to which, the eigenstates are
classified and the origins of degeneracies in energy spectrum are explained. We begin
the construction of the symmetric operations by means of permutations of atoms,
generate the icosahedral group by GAP, compute the irreducible representations, and
compose the projection operators. We execute the electronic structure calculations of
the molecule and analyze the wave-functions from the viewpoint of their irreducible
representation. Furthermore, some concepts of symmetries (to which physicists pays
little notice) are presented also by the analysis of the computer algebra. Throughout
the study, it is necessary that whole subgroups of icosahedral symmetry should be
extracted and classified by equivalence, and we will see that the utilization of the
computer algebra is effective for this purpose.
A.
The initial set-up
The C60 molecule (in Fig. 6) is projected on a plane as in the Fig. 7. The
vertexes are numbered from 1 to 60, where we follow the following rule: the two
atoms numbered as I and I+30 (I=1,...,30) are located in the opposite poles in the
molecular sphere. The pentagons are also numbered from 1 to 12 (with under-bars),
and two pentagons numbered as J and J+6 (J=1,...,6) are opposed with each other.
There are two types of bonds, single and double, having slightly different lengths,
distinguished by broad and broken segments in the figure. The icosahedral group are
generated by three basic operations and represented by the permutations of atoms as
98
FIG. 6. A C60 molecule. The numbering is given on some atoms and pentagons, in
correspondence with the projected images of the molecule, in Fig. 7.
follows. The group is the direct product of the fifth alternating group “A5” (the 5-th
alternating group) and the group of the inversion “I”. The generators are represented
by permutations:
gap> g1:=(1, 14, 20)(2, 15, 16)(3, 11, 17)(4, 12, 18)(5, 13, 19)
(6, 58, 22)(7, 57, 23)( 8, 56, 24)( 9, 60, 25)( 10, 59, 21)
( 31, 44, 50)( 32, 45, 46)( 33, 41, 47)( 34, 42, 48)( 35, 43, 49)
( 36, 28, 52)( 37, 27, 53)( 38, 26, 54)( 39, 30, 55)( 40, 29, 51);
gap> g2:=( 1, 5, 4, 3, 2)( 6,26,21,16,11)( 7,27,22,17,12)( 8,28,23,18,13)
( 9,29,24,19,14)(10,30,25,20,15)(31,35,34,33,32)(36,56,51,46,41)
(37,57,52,47,42)(38,58,53,48,43)(39,59,54,49,44)(40,60,55,50,45);
gap> g3:=( 1, 31)( 2, 32)( 3, 33)( 4, 34)( 5, 35)( 6, 36)( 7, 37)( 8, 38)
( 9, 39)( 10, 40)( 11, 41)( 12, 42)( 13, 43)( 14, 44)( 15, 45)( 16, 46)
( 17, 47)( 18, 48)( 19, 49)( 20, 50)( 21, 51)( 22, 52)( 23, 53)( 24, 54)
99
FIG. 7. The projected image of C60 on a plane. The pentagon in the center of this figure
(indexed as 1) is the one located in the lower half in Fig. 6. (The pentagon at No.7, not
shown in the projected image, is formed by five vertex at 31,32,33,34,35. )
( 25, 55)( 26, 56)( 27, 57)( 28, 58)( 29, 59)( 30, 60);
The elements g1 (the rotation by 120 degrees) and g2 (that by 72 degrees) generate
“A5“ and g3 generates “I”. The permutation, denoted by (a, b, c, d,...,x), should be
interpreted as the replacement of atoms in the following way: the atom located at the
100
site “a” is replaced by that at the site “b”, that at “b” by that at “c”, then in turn and
finally, the last entry “x” by the first “a”. The usage of sixty symbols for the definition
of the group is superfluous, and in fact, twelve symbols (representing the 12 vertexes
of the icosahedron, or the 12 pentagons in the molecule) will do. If it is unnecessary
to use the geometrical image, one can use 7 symbols, since, the icosahedral group is
the direct product of the 5-th alternating group A5 and the inversion I, as the former
requires five symbols “1,2,3,4,5” and the latter two symbols “6,7”. The equivalence
of these different definitions is assured by constructing isomorphisms among them,
which is an easy task for the computer algebra.
Remark. This correspondence (between the group elements and the atoms) is consistent to the operation of the permutation in GAP. For example, the permutation
(3,4,5) operates on the list [1,2,3,4,5] as follows:
gap> ListPerm((3,4,5));
[1,2,4,5,3];
Here the entries 3,4,5 are replaced by 4,5,3, as our interpretation assumes.
The group is constructed as
gap> gp:=Group(g1,g2,g3);
For the later utilization, it is more convenient to generate the group by means of
matrices (which represents the permutations on sixty vertexes). We use the built-in
function “ListPerm()” which computes the action on the permutation on the list
[1,2,...,60].
gap> pl1:=ListPerm(g1);
[ 14, 15, 11, 12, 13, 58, 57, 56, 60, 59, 17, 18, 19, 20, 16, 2, 3, 4, 5, 1,
10, 6, 7, 8, 9, 54, 53, 52, 51, 55, 44, 45, 41, 42, 43, 28, 27, 26, 30, 29,
101
47, 48, 49, 50, 46, 32, 33, 34, 35, 31, 40, 36, 37, 38, 39, 24, 23, 22, 21,
25 ]
gap> pl2:=ListPerm(g2);;
gap> pl3:=ListPerm(g3);;
These three lists are converted into three matrices by means of this user-defined
function. (The computed matrix element h[i][j] is nonzero when the symbol “i” goes
to “j” by permutation given in the list “l”.)
ListToMat:=function(l)
local size,h,i,j;
size:=Size(l);
h:=List([1..size],i->List([1..size],j->0));
for i in [1..size] do
h[i][l[i]]:=1;
od;
return h;
end;
gap> m1:=ListToMat(pl1);;
gap> m2:=ListToMat(pl2);;
gap> m3:=ListToMat(pl3);;
The irreducible representations are computed now.
gap> g:=Group(m1,m2,m3);
gap> irr:=Irr(g);
gap> Size(irr);
10
102
Remark. The computed irreducible representations by the group generated by the
permutations and that by the matrices are listed by different orders (at least, in
the result by GAP of the version 4.7.7.) This will cause some trouble in doing
computations. In this article, to avoid confusion, the groups generated by different
resources are given by different names even if they are isomorphic, and the generators
are presented in each time of the definition.
The characters for each element can be listed as numeral tables.
gap> elm:=Elements(g);
gap> A1:=List(elm,g->g^irr[1]);;
gap> A2:=List(elm,g->g^irr[2]);;
...............................
gap> A10:=List(elm,g->g^irr[10]);;
The notation g^irr[X] means the computation of the character value of the element
“g” by the X-th representation in this context.
The characters of the irreducible representations are orthogonal and normalized.
gap> A1*A1;
120
gap> A2*A2;
120
gap> A1*A2;
0
The virtual character is decomposed into the direct sum.
gap> (4*A1+2*A2)*TransposedMat([A1,A2,A3,A4,A5,A6,A7,A8,A9,A10])/Size(elm);
[4, 2, 0, 0, 0, 0, 0, 0, 0, 0]
103
I.1
I.2
I.3
I.4
I.5
I.6
I.7
I.8
I.9
I.10
I.Tr
CL1
1
1
1
3
3
3
3
4
4
5
5
60
CL2
12
1
-1
-E(5)-E(5)^4
-E(5)^2-E(5)^3
E(5)^2+E(5)^3
E(5)+E(5)^4
-1
1
0
0
0
CL3
12
1
1
-E(5)^2-E(5)^3
-E(5)-E(5)^4
-E(5)-E(5)^4
-E(5)^2-E(5)^3
-1
-1
0
0
0
CL4
20
1
-1
0
0
0
0
1
-1
-1
1
0
CL5
20
1
1
0
0
0
0
1
1
-1
-1
0
CL6
15
1
-1
-1
-1
1
1
0
0
1
-1
4
CL7
12
1
-1
-E(5)^2-E(5)^3
-E(5)-E(5)^4
E(5)+E(5)^4
E(5)^2+E(5)^3
-1
1
0
0
0
CL8
1
1
-1
3
3
-3
-3
4
-4
5
-5
0
CL9
CL10
12
15
1
1
1
1
-E(5)-E(5)^4
-1
-E(5)^2-E(5)^3 -1
-E(5)^2-E(5)^3 -1
-E(5)-E(5)^4
-1
-1
0
-1
0
0
1
0
1
0
0
TABLE 26. The table of characters of the irreducible representation of the icosahedral
group. Each row shows the characters of 10 conjugacy classes (CL1,...,CL10).
There are 10 conjugacy classes and 10 irreducible representations. They are shown
in the tables 26 and 27. In table 26, the second row shows the number of elements in
each conjugacy classes (CL1,...,CL10), and the ten rows bellow are characters for each
of the ten irreducible representations (I.1,...,I.10). The last row (I.Tr) shows the trace
of the group elements as matrices, which is a virtual character (not irreducible). The
√
notation E(5) means a complex number, the fifth root of the unity exp(2 −1π/5),
thus the characters are real-valued. Simply the computed result by GAP is tabulated
here, as it is. The projector on the p-th irreducible representation is computed by
the formula
P (p) =
l(p) X (p)∗
χ (T ) · O(T )
|G| T ∈G
by summing up the elements T of the group G, where l(p) is the dimension of the
p-th irreducible representation, g the order of the group, χ(p) the character, O(T )
the operation of the group (which would be represented as matrices and applied to
the vector representation of the wave function. The irreducible representation of the
104
The representatives
CL1
<identity ...>
x1^-1*x2^-2*x3
CL2
CL3
x2^-1*x1*x2*x1
x3*x2*x1*x2^2
CL4
CL5 x2^-2*x1*x2*x1 (also including x1)
CL6
x3*x2^2*x1*x2^2
CL7
x3*x2^2
CL8
x3
CL9 x1^-1*x2^-2 (also including x2)
CL10
x1*x2*x1
TABLE 27. The representatives in the conjugacy classes of the irreducible representations
of the icosahedral group. The generators of the groups are represented by symbols x1...xN,
as are numbered by GAP. In this case x1=m1,x2=m2,x3=m3.
wave function is determined numerically, by the condition that the application of the
projector should keep the wave function as it is, or delete it.)
B.
The analysis of the eigenstates
Now that we have the irreducible representations and the symmetry operations,
we can construct the projection operators.
At first, we assume a model of C60 , where, in each vertex, one s or π orbital is
located. The inter-atomic interactions are limited to those among three different
neighboring sites, as are signified by single and double bonds. We permit two different strengths of interactions in the Fock matrix H, according to two types of the
bonding. We neglect the overlap matrix elements between different sites, the overlapping matrix S being the unity. A computation is executed by the above assumption,
with the values of Hsingle =-2 (on the single bonds in the pentagons), Hdouble =-1 (on
the double bonds in the hexagons).
Before the electronic structure computation, we can deduce a prediction from
105
the group theory. The group elements given by the matrices which represent the
replacement of the 60 vertexes provide us the trace representation I.Tr. The virtual
representation is decomposed by the irreducible representations (as the sum of row
vectors in the table):
I.Tr = 1×I.1+0×I.2+1×I.3+1×I.4+2×I.5+2×I.6+2×I.7+2×I.8+3×I.9+2×I.10.
This relation tells us that the 60-dimensional vector space (on which the matrix group
acts) is divided into the ten irreducible components, given as the ten irreducible
representations of the group. The multiplicity of X-th component is equal to the
coefficient to “I.X”. One of these components, belonging to “I.2”, is a null space in
this case. The wave functions, computed as the eigenvectors of the Fock matrix, are
also subject to this division of the vector space.
Let us verify this. The eigenstates and the corresponding irreducible representations are shown in the following list. The abbreviations as “I.”,“D.”,“E” denote
the index of the irreducible representation, the degeneracy of the energy, and the
energies. The values of the energies are listed in the bracket. If the bracket is blank,
the eigenstates in correspondence with the irreducible representations do not exist.
The number of different eigenvalues in each irreducible representation is equal to the
multiplicity in the above relation between I.Tr and I.1,..., I.10.
I.1
D.1 E.[ -5.0000 ]
I.2
D.1 E.[ None ]
I.3
D.3 E.[ 4.2361 ]
I.4
D.3 E.[ -0.2361 ]
I.5
D.3 E.[ -4.6542, -0.5819 ]
I.6
D.3 E.[ -3.6765, 2.9126 ]
I.7
D.4 E.[ -1.4495, 3.4495 ]
106
I.8
D.4 E.[ -2.1623, 4.1623 ]
I.9
D.5 E.[ -4.0642, -1.6946, 2.7588 ]
I.10 D.5 E.[ -1.0000, 3.0000 ]
For a realistic quantum chemistry computation of C60 , the similar analysis is possible. For simplicity, one can make use of the amplitudes of 1s orbitals, because
those orbitals are transformed among themselves, without mixing with other types,
2s,2p,3s,3p. Except for one special case where the amplitudes of 1s orbital are null,
the correspondence between the eigenstates and the irreducible representations can
be well established. The result is listed in the following list. (The last entry of the list
I.0 is for the case of the failures in the identification.) Even in the case of the failure
of the analysis, the degeneracy is single, and one can conclude the corresponding
irreducible representation is I.1 or I.2, that of 1-dimension. The numbers of the different eigenvalues in the each irreducible representations are predictable by means of
the decomposition of the trace representation of the character of the transformation
matrices of the 1s,2s,2px,2py,2pz,3s,3px,3py,3pz orbitals. (The p orbitals are transformed by the same matrices prepared for the vibrational mode of the molecule, as
will be discussed later.)
I.1
D.1
E.[
7 different eigenvalues ]
I.2
D.1
E.[ ? ]
I.3
D.3
E.[ 11 different eigenvalues ]
I.4
D.3
E.[ 11 different eigenvalues ]
I.5
D.3
E.[ 16 different eigenvalues ]
I.6
D.3
E.[ 16 different eigenvalues ]
I.7
D.4
E.[ 24 different eigenvalues ]
I.8
D.4
E.[ 24 different eigenvalues ]
I.9
D.5
E.[ 25 different eigenvalues ]
107
I.10 D.5
E.[ 20 different eigenvalues ]
I.0
E.[
C.
D.1
2 different eigenvalues ]
The analysis again, against the failure in the identification
As the degeneracy of the energy is one, one can see the levels denoted as I.0 in
the previous section belong to the representation I.1 or I.2. The characters of the
one-dimensional representation in table 26 are shown here again (in table 28).
CL1 CL2 CL3 CL4 CL5 CL6 CL7 CL8 CL9 CL10
I.1 1
1
1
1
1
1
1
1
1
1
I.2 1 -1 1 -1 1 -1 -1 -1 1
1
TABLE 28. The one-dimensional irreducible representations.
To distinguish the representation I.1 and I.2 with ease, one may apply the element
in the conjugacy class 8 (the inversion). By the inversion, one vertex moves to the
opposite site with respect to the center of the geometry, and the coefficients of porbitals change signs. The amplitudes of the wave-functions (in the two atoms located
in the antipodes positions) in I.0 are typically obtained as in table 29.
CA 1S
2S
2PX
2PY
2PZ
CB 1S
2S
2PX
2PY
2PZ
0
0
X
Y
Z
0
0
X
Y
Z
TABLE 29. The amplitudes of the wave-function (coefficients for each orbital) in two
opposing atoms.
108
The inversion changes the sign of non-zero coefficients. We must recall that the
wave-functions are basis functions of some irreducible representations, and the traces
of their transformation matrices are the characters. Since the traces are -1 in the
above cases, the wave-functions belong to the representation I.2. The reason of the
failure in the analysis by the projection operators applied on s orbitals are fundamental one: the projection operator of I.2 over s orbitals is exactly the zero-matrix,
which gives rise to exact zero amplitudes in s orbitals (i.e. the null space.) The
reason of the existence of the orbitals of the I.2 is that the actual C60 molecule has
the symmetry of the inversion and the carbon atom has p-orbitals.
D.
The perturbation and the symmetry
We often have to compute matrix elements in the perturbation theory such as
hψi |Vp |ψj i = [Ii |Ip |Ij ] = [Ii |Ip ∗ Ij ].
(83)
The product of the perturbation potential and the wave function belongs to the
direct product of the representation, which is decomposed into the direct sum
Ip ∗ Ij =
X
Ap,j
k Ik .
(84)
k
We can determine beforehand whether this matrix element would be zero or not, only
from the group theoretical consideration: if the two irreducible representations in the
inner product are the same, the matrix element will be non-zero. In the system with
the high symmetry, the group theoretical constraint is so strong that it annihilates
many matrix elements. However, the symmetry is reduced, there is some possibility
where the matrix element will become non-zero. Indeed there is no symmetry, there
109
is no group theoretical constraint which annihilates the matrix elements. And if
one uses the group theory, one can see the least reduction of the symmetry where
the matrix element will become non-zero. The irreducible representation of group g
ramifies in the subgroups as follows.
(g,s)
(g,s)
(g,s)
Igp = Ci,1 Is1 + Ci,2 Is2 + · · · + Ci,M IsM .
(85)
One can find the subgroup where the non-zero matrix element will exist, by tracing
the sequence of subgroups (in the subgroup lattice) and checking the possible combinations of the ramified irreducible representations. In this stage, the concrete style
of the structural deformation is not discussed yet.
One of the important perturbation potentials is that of optical transition, such as
E ·r. Now we determine the irreducible representation to which the perturbation E ·r
(or the spatial vector r) belongs. We now construct the icosahedral group from the
rotation matrices in three-dimensional Euclidean space, acting on the twelve vertexes
of one icosahedron.
The twelve vertexes in one icosahedron are given by
1
[ 0, 0, 1 ]
2
[ 2/5*E(5)-2/5*E(5)^2-2/5*E(5)^3+2/5*E(5)^4, 0,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4 ]
3
[ -3/5*E(5)-2/5*E(5)^2-2/5*E(5)^3-3/5*E(5)^4,
-1/5*E(20)+1/5*E(20)^9+2/5*E(20)^13-2/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4 ]
4
[ 2/5*E(5)+3/5*E(5)^2+3/5*E(5)^3+2/5*E(5)^4,
-2/5*E(20)+2/5*E(20)^9-1/5*E(20)^13+1/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4 ]
5
[ 2/5*E(5)+3/5*E(5)^2+3/5*E(5)^3+2/5*E(5)^4,
110
2/5*E(20)-2/5*E(20)^9+1/5*E(20)^13-1/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4 ],
6
[ -3/5*E(5)-2/5*E(5)^2-2/5*E(5)^3-3/5*E(5)^4,
1/5*E(20)-1/5*E(20)^9-2/5*E(20)^13+2/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4 ]
7
[ 0, 0, -1 ]
8
[ -2/5*E(5)+2/5*E(5)^2+2/5*E(5)^3-2/5*E(5)^4, 0,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4 ]
9
[ 3/5*E(5)+2/5*E(5)^2+2/5*E(5)^3+3/5*E(5)^4,
1/5*E(20)-1/5*E(20)^9-2/5*E(20)^13+2/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4 ]
10
[ -2/5*E(5)-3/5*E(5)^2-3/5*E(5)^3-2/5*E(5)^4,
2/5*E(20)-2/5*E(20)^9+1/5*E(20)^13-1/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4 ]
11
[ -2/5*E(5)-3/5*E(5)^2-3/5*E(5)^3-2/5*E(5)^4,
-2/5*E(20)+2/5*E(20)^9-1/5*E(20)^13+1/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4 ]
12
[ 3/5*E(5)+2/5*E(5)^2+2/5*E(5)^3+3/5*E(5)^4,
-1/5*E(20)+1/5*E(20)^9+2/5*E(20)^13-2/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4 ]
The three generators of the group are as follows:
gap> f1:=(1,2,3)(4,6,11)(5,10,12)(7,8,9);;
gap> f2:=(2,6,5,4,3)(8,12,11,10,9);;
gap> f3:=(1,7)(2,8)(3,9)(4,10)(5,11)(6,12);;
These permutations are also represented by the rotation matrices as
111
gap> f1:=[ [ 3/10*E(5)+1/5*E(5)^2+1/5*E(5)^3+3/10*E(5)^4,
1/10*E(20)-1/10*E(20)^9-1/5*E(20)^13+1/5*E(20)^17,
2/5*E(5)-2/5*E(5)^2-2/5*E(5)^3+2/5*E(5)^4 ],
[ -1/2*E(20)+1/2*E(20)^9, -1/2*E(5)-1/2*E(5)^4, 0 ],
[ -3/5*E(5)-2/5*E(5)^2-2/5*E(5)^3-3/5*E(5)^4,
-1/5*E(20)+1/5*E(20)^9+2/5*E(20)^13-2/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4 ] ];
gap> f2:=[ [ 1/2*E(5)+1/2*E(5)^4, -1/2*E(20)+1/2*E(20)^9, 0 ],
[ 1/2*E(20)-1/2*E(20)^9, 1/2*E(5)+1/2*E(5)^4, 0 ], [ 0, 0, 1 ] ];;
gap> f3:=[ [ -1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ];;
N.1
N.2
N.3
N.4
N.5
N.6
N.7
N.8
N.9
N.10
Dimension
1
1
3
3
3
3
4
4
5
5
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
0,
0,
0,
0,
1,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0
0
0
0
0
0
0
0
0
0
],
],
],
],
],
],
],
],
],
],
Operator
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 1, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[
[
[
[
[
[
[
[
[
[
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0
0
0
0
1
0
0
0
0
0
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
The characters of [f1,f2,f3]
[ 1, 1, 1 ]
[ 1, 1, -1 ]
[ 0,-E(5)^2-E(5)^3, 3 ]
(I.4 in table 26)
[ 0,-E(5)-E(5)^4, 3 ]
(I.3 in 26)
[ 0,-E(5)^2-E(5)^3, -3 ]
[ 0,-E(5)-E(5)^4, -3 ]
[ 1, -1, 4 ]
[ 1, -1, -4 ]
[ -1, 0, 5 ]
[ -1, 0, -5 ]
TABLE 30. The projection operators. Be careful that the computation by GAP returns the representations (N.1,...,N.10) in the slightly different order from those in table
26(I.1,...,I.10).
The ten projection operators are computed as in table 30. The operators, except
that of the fifth representation, are exactly zero. The symmetry operations on one
vector, which is subject to the icosahedral symmetry, are only representable in the
fifth irreducible representation, and the perturbation term E · r is represented by
the fifth representation. (The fifth representation can be distinguished from others,
112
as such having characters [0,-E(5)^2-E(5)^3,-3] for three generators. Especially,
the rotation of 72 degrees should have the character value -E(5)^2-E(5)^3. The
rotation by 72 degrees and its inverse (f2 and f2^-1) have the same character value,
because they are conjugate.)
Remark. In order to see whether the elements x and y are conjugate in group G,
try the command.
gap> RepresentativeAction(G,x,y);
E.
The possible deformation in the reduced symmetry
We can construct molecular geometry which is admissible in each subgroup. If the
vertexes are dislocated as V(i) → V(i)+d(i), the set of d(i) should satisfy a certain
geometrical constraint in the symmetry operation in the subgroup. We should take
note of this : since the icosahedral group which includes 120 operations and acts on
60 vertexes in one C60 , there are two operations, both of which equally transfers one
atomic site into same, but another site. If the dislocations of atoms are still subject
to the icosahedral symmetry, there is a constraint. When the atom at the site V[i]
moves to V[j] by the operation R, it must be
R · V[i] = V[j],
and
R · (V[i] + d[i]) = V[j] + d[j],
i.e.
R · d[i] = d[j].
If there are two possible transformations R and Q, it must be
R · d[i] = Q · d[i].
113
If we chose T=R−1 Q, then T·d[i]=d[i], and the possible dislocation is the solution of
the equation (T − E) · d[i] = 0 (E: the unit matrix). In the case of the icosahedral
symmetry in C60 , such a T is the reflection at the plain in which one double bond
and its antipodes are located.
In general, the dislocations d[i] on all of the vertex are generated as follows. For
example, take V[1], and find the operation R, Q, T as above, and then find a possible
solution d of (T − E) · d = 0 as the generator. The application of the 120 operations
on V[1]+d shall generate 60 sites, which are represented by V[i]+d[i], and the set
of d[i] is obtained. (This is equivalent to the projection on the trivial irreducible
representation, in which the basis set is invariant under the group action.) In a
subgroup, the symmetric operations do not act transitively, and the “orbit” of one
vertex by the symmetry operations (which covers all of the vertexes in C60 by the
icosahedral symmetry) splits into several subsets, as we will see later. For each
subset, we construct independent d, the generator of the dislocation in each subset.
If the subgroups contain the reflection, the dislocations on the atoms have similar
constraints to d (being fixed on one of the reflectional planes) as in the case of
the icosahedral symmetry. On the contrary, when the reflection is missing in the
subgroup, we can choose arbitrary generator of dislocations in each sub-orbits. One
should note that the center of mass will not be kept since d[1]+d[2]+...+d[60] is not
equal to 0 in general.
Concerning this topic, we can classify the vibrational mode of C60 according to
the group theory. The procedure is to form the direct product of the virtual representation with respect to the replacement of atoms and the irreducible representation
of the symmetry operations on the space vector, and then to decompose it into the
direct sum. One will find that the vibrational mode in the icosahedral symmetry
includes the duplicated trivial irreducible representation, to which the degrees of the
freedom is two. The geometrical interpretation of this is that the possible distortion
114
which keeps the icosahedral symmetry is confined in one plane as discussed above.
F.
The reduction of the symmetry, from the viewpoint of the orbit in the
permutation
The full symmetry of C60 is the icosahedral group, generated by “g1”, “g2”, “g3”
in the previous section, and it has 164 subgroups. All of the subgroups are computed
as
gap> gr:=Group(g1,g2,g3);
gap> AS:=AllSubgroups(gr);;
The list of the subgroups is given in the descending order of the sizes of the
subgroups, and the largest of them (No.164) is the icosahedral group.
gap> irrA:=Irr(AS[164]);
Take a subgroup No.162. This group is of the order 24, and the irreducible
representation is also computed.
gap> irrB:=Irr(AS[162]);;
The irreducible representations of these groups are given in tables 31,32,33, and
34.
The branching relation from the group of No.164 to that of No.162 is computed
now. The characters of the elements of the group of No.162 are computed and listed
in two ways: the one is by the irreducible representations of the group of No.164,
and another by those of the group of No.162. The representations in the first list are
decomposed as the summation of the entries in the second list, and the result is the
branching relation.
115
No.1
No.2
No.3
No.4
No.5
No.6
No.7
No.8
No.9
No.10
CL1
1
1
3
3
3
3
4
4
5
5
CL2
1
-1
-1
-1
1
1
0
0
1
-1
CL3
1
1
-E(5)-E(5)^4
-E(5)^2-E(5)^3
-E(5)-E(5)^4
-E(5)^2-E(5)^3
-1
-1
0
0
CL4
1
1
-E(5)^2-E(5)^3
-E(5)-E(5)^4
-E(5)^2-E(5)^3
-E(5)-E(5)^4
-1
-1
0
0
CL5
1
1
-1
-1
-1
-1
0
0
1
1
CL6
1
-1
-E(5)^2-E(5)^3
-E(5)-E(5)^4
E(5)^2+E(5)^3
E(5)+E(5)^4
-1
1
0
0
CL7
1
1
0
0
0
0
1
1
-1
-1
CL8
1
-1
0
0
0
0
1
-1
-1
1
CL9
1
-1
3
3
-3
-3
4
-4
5
-5
CL10
1
-1
-E(5)-E(5)^4
-E(5)^2-E(5)^3
E(5)+E(5)^4
E(5)^2+E(5)^3
-1
1
0
0
TABLE 31. The character table of the group AS[164].
CL1
CL2
CL3
CL4
CL5
CL6
CL7
CL8
CL9
CL10
Representative
<identity ...>
x1^-1*x2^-2*x3*x1*x2*x1
x2^-1
x2^-2
x2^-2*x1*x2
x2^2*x1*x2^2*x1^-1*x3
x2^2*x1*x2
x2^-1*x1^-1*x2^-2*x1^-1*x3
x3
x2^-1*x3
Size
1
15
12
12
15
12
20 (including g1)
20
1
12
TABLE 32. The conjugacy classes of the group AS[164].
gap> LA:=List(irrA,a->List(Elements(AS[162]),g->g^a));
[ [ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ],
.............................................................................
.............................................................................
[ 5, -1, 1, -1, -1, 1, -1, 1, 1, -1, 1, -1, -5, 1, -1, 1, 1, -1, 1, -1, -1,
1, -1, 1 ] ]
gap> LB:=List(irrB,a->List(Elements(AS[162]),g->g^a));
[ [ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ],
116
No.1
No.2
No.3
No.4
No.5
No.6
No.7
No.8
CL1
1
1
1
1
1
1
3
3
CL2
1
-1
-1
-1
1
1
1
-1
CL3
1
-1
-E(3)
-E(3)^2
E(3)^2
E(3)
0
0
CL4
1
1
E(3)
E(3)^2
E(3)^2
E(3)
0
0
CL5
1
1
E(3)^2
E(3)
E(3)
E(3)^2
0
0
CL6
1
-1
-E(3)^2
-E(3)
E(3)
E(3)^2
0
0
CL7
1
1
1
1
1
1
-1
-1
CL8
1
-1
-1
-1
1
1
-3
3
TABLE 33. The character table of the group AS[162].
CL1
CL2
CL3
CL4
CL5
CL6
CL7
CL8
Representative
<identity ...>
x2^-1*x1^-1*x2^-2*x3*x1*x2*x1
x1^-1*x2^-2*x1^-1*x3
x2^-1*x1*x2
x2^-2*x1
x1*x2^2*x1^-1*x2^-1*x3
x2*x1^-1
x3
Size
1
3
4
4
4
3
3
1
TABLE 34. The conjugacy classes of the group AS[162].
.............................................................................
.............................................................................
0 ], [ 3, -1, 0, 0, 0, 0, -1, -1, 0, 0, 0, 0, 3, -1, 0, 0, 0, 0, -1,
-1, 0, 0, 0, 0 ] ]
gap> LA*TransposedMat(LB)/Size(AS[162]);
The result returns as the matrix in the bellow: the irreducible representations are
ramified as
No.164 -> No.162
[ [ 1, 0, 0, 0, 0, 0, 0, 0 ],
1
-> 1
[ 0, 1, 0, 0, 0, 0, 0, 0 ],
2
-> 2
117
[ 0, 0, 0, 0, 0, 0, 0, 1 ],
3
-> 8
[ 0, 0, 0, 0, 0, 0, 0, 1 ],
4
-> 8
[ 0, 0, 0, 0, 0, 0, 1, 0 ],
5
-> 7
[ 0, 0, 0, 0, 0, 0, 1, 0 ],
6
-> 7
[ 1, 0, 0, 0, 0, 0, 0, 1 ],
7
-> 1 + 8
[ 0, 1, 0, 0, 0, 0, 1, 0 ],
8
-> 2 + 7
[ 0, 0, 0, 0, 1, 1, 0, 1 ],
9
-> 5 + 6 + 8
[ 0, 0, 1, 1, 0, 0, 1, 0 ] ]
10
-> 3 + 4 + 7
One can see the change in the geometry by checking the “orbit” of the vertex “1”
and the bond “[1,6]”. In the icosahedral symmetry, the vertex is transferred to other
sixty vertexes transitively, and this double bond is transferred to all other thirty
double bonds.
gap> List(Elements(AS[164]),g->OnPoints(1,g));
[ 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12,
12, 13, 13, 14, 14, 15, 15, 16, 16, 17, 17, 18, 18, 19, 19, 20, 20, 21, 21,
22, 22, 23, 23, 24, 24, 25, 25, 26, 26, 27, 27, 28, 28, 29, 29, 30, 30, 31,
31, 32, 32, 33, 33, 34, 34, 35, 35, 36, 36, 37, 37, 38, 38, 39, 39, 40, 40,
41, 41, 42, 42, 43, 43, 44, 44, 45, 45, 46, 46, 47, 47, 48, 48, 49, 49, 50,
50, 51, 51, 52, 52, 53, 53, 54, 54, 55, 55, 56, 56, 57, 57, 58, 58, 59, 59,
60, 60 ]
gap> BONDA:=Unique(List(Elements(AS[164]),g->OnSets([1,6],g)));
[ [ 1, 6 ], [ 2, 11 ], [ 3, 16 ], [ 4, 21 ], [ 5, 26 ], [ 7, 30 ], [ 8, 49 ],
[ 9, 53 ], [ 10, 12 ], [ 13, 54 ], [ 14, 58 ], [ 15, 17 ], [ 18, 59 ],
[ 19, 38 ], [ 20, 22 ], [ 23, 39 ], [ 24, 43 ], [ 25, 27 ], [ 28, 44 ],
[ 29, 48 ], [ 31, 36 ], [ 32, 41 ], [ 33, 46 ], [ 34, 51 ], [ 35, 56 ],
[ 37, 60 ], [ 40, 42 ], [ 45, 47 ], [ 50, 52 ], [ 55, 57 ] ]
118
In the symmetry of the subgroup No.162, in contrast, the vertex “1” is transferred
only to 24 other vertexes. The orbit of single vertexes (which covers the whole system
in the case of the subgroup 164) splits into three subsets. Similarly the bond “[1,6]”
is not transferred transitively, only to 24 of double bonds. The remaining six double
bonds are transferred among themselves. This situation can be interpreted that
one of the double bonds (one of these connecting one pentagon with five neighbors)
slightly changes the length, while those of remaining four bonds are unchanged.
gap> List(Elements(AS[162]),g->OnPoints(1,g));
[ 1, 5, 9, 10, 11, 15, 18, 19, 21, 22, 27, 28, 31, 35, 39, 40, 41, 45, 48,
49, 51, 52, 57, 58 ]
gap> List(Elements(AS[162]),g->OnPoints(2,g));
[ 2, 4, 8, 6, 12, 14, 17, 20, 25, 23, 26, 29, 32, 34, 38, 36, 42, 44, 47, 50,
55, 53, 56, 59 ]
gap> List(Elements(AS[162]),g->OnPoints(3,g));
[ 3, 3, 7, 7, 13, 13, 16, 16, 24, 24, 30, 30, 33, 33, 37, 37, 43, 43, 46, 46,
54, 54, 60, 60 ]
gap> BONDB:=Unique(List(Elements(AS[162]),g->OnSets([1,6],g)));
gap> Sort(BONDB);
gap> BONDB;
[ [ 1, 6 ], [ 2, 11 ], [ 4, 21 ], [ 5, 26 ], [ 8, 49 ], [ 9, 53 ],
[ 10, 12 ], [ 14, 58 ], [ 15, 17 ], [ 18, 59 ], [ 19, 38 ], [ 20, 22 ],
[ 23, 39 ], [ 25, 27 ], [ 28, 44 ], [ 29, 48 ], [ 31, 36 ], [ 32, 41 ],
[ 34, 51 ], [ 35, 56 ], [ 40, 42 ], [ 45, 47 ], [ 50, 52 ], [ 55, 57 ] ]
The splitting of the energy degeneracy with the symmetry reduction is verified
numerically, for example, by a model computation, in which the matrix elements of
these two split orbits of double bonds take slightly different values, in correspon119
dence with the change in the symmetry. In this computation, the elements of the
Fock matrix which belong to the one of the split orbit (including six bonds, the complementary part of BONDB in BONDA) are enhanced slightly, multiplied by 1.005.
The energy spectra at HOMO and LUMO are computed as follows.(The positions of
HOMO and LUMO are ambiguous in this case. We presently assume that the half of
the 60 eigenstates are occupied by 60 electrons. The left and right columns show the
spectra of the full and the reduced symmetries. The HOMO and LUMO belong to
the irreducible representations of No.10 and No.5 of the character table of the group
AS[164], and the splitting of the spectra is deduced from the above relation.)
HOMO : Split from one fivefold degeneracy
to one threefold degeneracy and two single degeneracies.
26
-1.0000000
->
-1.0000021
27
-1.0000000
->
-1.0000021
28
-1.0000000
->
-1.0000021
29
-1.0000000
->
-1.0000000
30
-1.0000000
->
-1.0000000
In this case, two single states still have the same energy levels. This is due to
the fact that the two single states belong to two of the one-dimensional irreducible
representation which are complex-conjugates with each other. As the eigenvalue
problem is the real-valued one, these two eigenvalues should be equal owing to the
invariance by the complex conjugation acting on it. It could also be interpreted as
of the time-reversal symmetry in the context of the quantum physics.
LUMO: Without splitting.
31 -0.5818861
->
-0.5810887
32 -0.5818861
->
-0.5810887
33 -0.5818861
->
-0.5810887
120
The symmetry reduction from the icosahedral group to one of the subgroups of
order 24 can be realized clearly, as supposed in this computation. On the other hand,
the difference between the icosahedral group and the maximal subgroup AS[163] (the
5-th alternating group: A5) is more subtle. The splitting of the “orbits” by the group
action does not occur in the transposition of the bonds between nearest or secondnearest neighbors. The length of the single or double bonds is also unique in A5
symmetry, and the change in the symmetry could be seen in the “orbit” of the bonds
among more distant neighbors. Take the bond of [1,11].
gap> lg:=List(Elements(AS[164]),i->OnSets([1,11],i));
[ [ 1, 11 ], [ 1, 26 ], [ 2, 6 ], [ 2, 16 ], [ 3, 11 ], [ 3, 21 ], [ 4, 16 ],
[ 4, 26 ], [ 5, 6 ], [ 5, 21 ], [ 6, 30 ], [ 6, 12 ], [ 1, 7 ], [ 7, 49 ],
[ 8, 30 ], [ 8, 53 ], [ 9, 49 ], [ 9, 12 ], [ 1, 10 ], [ 10, 53 ],
[ 10, 11 ], [ 11, 17 ], [ 2, 12 ], [ 12, 54 ], [ 10, 13 ], [ 13, 58 ],
[ 14, 54 ], [ 14, 17 ], [ 2, 15 ], [ 15, 58 ], [ 15, 16 ], [ 16, 22 ],
[ 3, 17 ], [ 17, 59 ], [ 15, 18 ], [ 18, 38 ], [ 19, 59 ], [ 19, 22 ],
[ 3, 20 ], [ 20, 38 ], [ 20, 21 ], [ 21, 27 ], [ 4, 22 ], [ 22, 39 ],
[ 20, 23 ], [ 23, 43 ], [ 24, 39 ], [ 24, 27 ], [ 4, 25 ], [ 25, 43 ],
[ 25, 26 ], [ 7, 26 ], [ 5, 27 ], [ 27, 44 ], [ 25, 28 ], [ 28, 48 ],
[ 29, 44 ], [ 7, 29 ], [ 5, 30 ], [ 30, 48 ], [ 31, 41 ], [ 31, 56 ],
[ 32, 36 ], [ 32, 46 ], [ 33, 41 ], [ 33, 51 ], [ 34, 46 ], [ 34, 56 ],
[ 35, 36 ], [ 35, 51 ], [ 36, 60 ], [ 36, 42 ], [ 31, 37 ], [ 19, 37 ],
[ 38, 60 ], [ 23, 38 ], [ 19, 39 ], [ 39, 42 ], [ 31, 40 ], [ 23, 40 ],
[ 40, 41 ], [ 41, 47 ], [ 32, 42 ], [ 24, 42 ], [ 40, 43 ], [ 28, 43 ],
[ 24, 44 ], [ 44, 47 ], [ 32, 45 ], [ 28, 45 ], [ 45, 46 ], [ 46, 52 ],
[ 33, 47 ], [ 29, 47 ], [ 45, 48 ], [ 8, 48 ], [ 29, 49 ], [ 49, 52 ],
[ 33, 50 ], [ 8, 50 ], [ 50, 51 ], [ 51, 57 ], [ 34, 52 ], [ 9, 52 ],
121
[ 50, 53 ], [ 13, 53 ], [ 9, 54 ], [ 54, 57 ], [ 34, 55 ], [ 13, 55 ],
[ 55, 56 ], [ 37, 56 ], [ 35, 57 ], [ 14, 57 ], [ 55, 58 ], [ 18, 58 ],
[ 14, 59 ], [ 37, 59 ], [ 35, 60 ], [ 18, 60 ] ]
On the other hand, the “orbit” by A5 includes only 60 bonds.
gap> la5:=List(Elements(AS[163]),i->OnSets([1,11],i));
[ [ 1, 11 ], [ 2, 16 ], [ 3, 21 ], [ 4, 26 ], [ 5, 6 ], [ 6, 30 ], [ 7, 49 ],
[ 8, 53 ], [ 9, 12 ], [ 1, 10 ], [ 10, 11 ], [ 12, 54 ], [ 13, 58 ],
[ 14, 17 ], [ 2, 15 ], [ 15, 16 ], [ 17, 59 ], [ 18, 38 ], [ 19, 22 ],
[ 3, 20 ], [ 20, 21 ], [ 22, 39 ], [ 23, 43 ], [ 24, 27 ], [ 4, 25 ],
[ 25, 26 ], [ 27, 44 ], [ 28, 48 ], [ 7, 29 ], [ 5, 30 ], [ 31, 56 ],
[ 32, 36 ], [ 33, 41 ], [ 34, 46 ], [ 35, 51 ], [ 36, 42 ], [ 31, 37 ],
[ 38, 60 ], [ 19, 39 ], [ 23, 40 ], [ 41, 47 ], [ 32, 42 ], [ 40, 43 ],
[ 24, 44 ], [ 28, 45 ], [ 46, 52 ], [ 33, 47 ], [ 45, 48 ], [ 29, 49 ],
[ 8, 50 ], [ 51, 57 ], [ 34, 52 ], [ 50, 53 ], [ 9, 54 ], [ 13, 55 ],
[ 37, 56 ], [ 35, 57 ], [ 55, 58 ], [ 14, 59 ], [ 18, 60 ] ]
The remaining 60 bonds are located in the separated orbit now.
gap> la5_2:=List(Elements(AS[163]),i->OnSets([2,6],i));
[ [ 2, 6 ], [ 3, 11 ], [ 4, 16 ], [ 5, 21 ], [ 1, 26 ], [ 1, 7 ], [ 8, 30 ],
[ 9, 49 ], [ 10, 53 ], [ 6, 12 ], [ 2, 12 ], [ 10, 13 ], [ 14, 54 ],
[ 15, 58 ], [ 11, 17 ], [ 3, 17 ], [ 15, 18 ], [ 19, 59 ], [ 20, 38 ],
[ 16, 22 ], [ 4, 22 ], [ 20, 23 ], [ 24, 39 ], [ 25, 43 ], [ 21, 27 ],
[ 5, 27 ], [ 25, 28 ], [ 29, 44 ], [ 30, 48 ], [ 7, 26 ], [ 35, 36 ],
[ 31, 41 ], [ 32, 46 ], [ 33, 51 ], [ 34, 56 ], [ 31, 40 ], [ 36, 60 ],
[ 19, 37 ], [ 23, 38 ], [ 39, 42 ], [ 32, 45 ], [ 40, 41 ], [ 24, 42 ],
[ 28, 43 ], [ 44, 47 ], [ 33, 50 ], [ 45, 46 ], [ 29, 47 ], [ 8, 48 ],
122
[ 49, 52 ], [ 34, 55 ], [ 50, 51 ], [ 9, 52 ], [ 13, 53 ], [ 54, 57 ],
[ 35, 60 ], [ 55, 56 ], [ 14, 57 ], [ 18, 58 ], [ 37, 59 ] ]
The bonds, equivalent to “[1,11]”, with the unique length in the icosahedral symmetry, are permitted to have two different lengths in the A5 symmetry. The deformation from the icosahedral symmetry to that of A5 is realized not in the bond
structures between the nearest or the second-nearest neighbors, but those between
remoter atomic sites. By such deformation into A5 symmetry, the two pentagons,
located in the antipodes positions, can rotate themselves in the opposite directions
with each other.
G.
The symmetry in the eigenstates (hidden one)
The irreducible representations of the same dimension show a certain similarity
in the icosahedral group. For example, in table 26, the representations of I.3 and I.4
or I.5 and I.6 seem to be interchangeable, if we replace E(5), E(5)^2, E(5)^3, E(5)^4
among themselves. This suggests a certain kind of affinity of the basis functions
of the wave functions. This is the symmetry in the context of the Galois theory,
especially of the cyclotomic field, and manifest itself in the secular equation of the
energy spectrum.
I.3
I.4
I.5
I.6
CL1
3
3
3
3
CL2
-E(5)-E(5)^4
-E(5)^2-E(5)^3
E(5)^2+E(5)^3
E(5)+E(5)^4
CL3
-E(5)^2-E(5)^3
-E(5)-E(5)^4
-E(5)-E(5)^4
-E(5)^2-E(5)^3
CL4
0
0
0
0
CL5
0
0
0
0
CL6
-1
-1
1
1
CL7
-E(5)^2-E(5)^3
-E(5)-E(5)^4
E(5)+E(5)^4
E(5)^2+E(5)^3
CL8
3
3
-3
-3
CL9
CL10
-E(5)-E(5)^4
-1
-E(5)^2-E(5)^3 -1
-E(5)^2-E(5)^3 -1
-E(5)-E(5)^4
-1
TABLE 35. The three dimensional irreducible representations in the icosahedral group.
Let us start again at the model computation of the electronic structure of C60 . The
eigenvalue problem is solved by one of the functions in the GAP package. By this,
123
the secular equation is decomposed into polynomials with coefficients over the specified field, and the “generalized” eigenstates for these polynomials are obtained.(The
eigenvalues are computed as the set of polynomials, each of which has a part of
eigenvalues as the roots. The eigenstates of the matrix H are the solutions of the
matrix equation p(H)·x=0, where p(t) is one of these polynomials.)
Let “hh” be the Fock matrix, used in the model computation of the electronic
structure of C60 , with the same parameter as in the previous section. The generators
of the group g1,g2,g3 are the same as defined previously. The eigenvalue problem is
solved now. The last part of the following computation is to extract the basis set of
the eigenspaces. (The secular equation is representable as the product of polynomials
equipped the parameter for the transfer integrals. See the appendix H.)
Hgen:=function(GP,a)
local i,h,tuples;
h:=List([1..60],i->List([1..60],j->0));
tuples:=Orbit(GP,a,OnTuples);
for i in tuples
do
h[i[1]][i[2]]:=1;
od;
return h;
end;
This function sets non-zero matrix elements hij , when the symmetry of the molecule
demand the bond between site “i” and “j”. To do this, the bond defined in the
list “a”, in the argument of this function, is transferred to equivalent bonds in the
symmetry of the group “GP”.
gap> G:=Group(g1,g2,g3);
124
No. Polynomial
1
2
3
4
5
6
7
8
Eigenspace
dimension
x_1-3
5
x_1+1
5
x_1+5
1
x_1^2-4*x_1-1
6
x_1^2-2*x_1-9
8
x_1^2-2*x_1-5
8
x_1^3+3*x_1^2-9*x_1-19
15
x_1^4+6*x_1^3-4*x_1^2-54*x_1-29 12
Degeneracy
of one eigenvalue
5
5
1
3
4
4
3
3
Irreducible
representation
I.10
I.10
I.1
I.3,I.4
I.8
I.7
I.9
I.5,I.6
TABLE 36. The classification of the solutions of the generalized eigenvalue problem. The
irreducible representations are those presented in table 26.
gap> h1:=Hgen(G,[1,2]);
gap> h2:=Hgen(G,[1,6]);
gap> hh:=-2*h1-h2;
gap> ge:=GeneralizedEigenvalues(Rationals,hh);
[ x_1-3,
x_1+1,
x_1^2-2*x_1-5,
x_1+5,
x_1^2-4*x_1-1,
x_1^2-2*x_1-9,
x_1^3+3*x_1^2-9*x_1-19,
x_1^4+6*x_1^3-4*x_1^2-54*x_1-29 ]
gap> gs:=GeneralizedEigenspaces(Rationals,hh);;
gap> gsbas:=List(gs,i->Basis(i));;
The eight polynomials of the eigenvalues, decomposed up to the field of rational
number are given as the table 36, and the basis vectors in the eigenspace are allotted
into the irreducible representations.
The representations of I.3 and I.4 or I.5 and I.6 belong to the same eigenspace,
specified by the same polynomial.
The automorphism of the cyclotomic field
(E(5)+E(5)^4→E(5)^2 + E(5)^3), which forms a group, interchanges these pairs
of the irreducible representation, and consequently, also interchanges the vector solutions x of the equation p(H)·x=0. From the viewpoint of the Galois theory, Let
125
us review the situation. Take the last polynomial of the four degrees. The 8-th
polynomial in the above table has C4 as the Galois group, and the four roots are
transported with each other by the operation of this group( as 1 → 2 → 3 → 4 → 1,
by a certain sequential order of the roots.)
Remark. The Galois group of the polynomial is computed by GAP.
gap> ge[8];
x_1^4+6*x_1^3-4*x_1^2-54*x_1-29
gap> GaloisType(ge[8]);
1
gap> TransitiveGroup(4,1);
C(4) = 4
The command Galoistype(f ) computes the Galois Group of the polynomial f, and
returns an integer i if the Galois group is isomorphic to TransitiveGroup(n, i), where
n is the degree of f.
Let us represent this Galois group concretely as the operation of the eigenvectors.
The polynomial splits in four polynomials of one degree in the 15-th cyclotomic
field Q15 . The generalized eigenvalue problem is solved again in this field, for which
CF(15) is the notation in GAP.
gap> ge2:=GeneralizedEigenvalues(CF(15),hh);
[ x_1-3, x_1+1, x_1+5, x_1+(E(5)+3*E(5)^2+3*E(5)^3+E(5)^4),
x_1+(3*E(5)+E(5)^2+E(5)^3+3*E(5)^4),
x_1+(E(15)+E(15)^2+E(15)^4+3*E(15)^7+3*E(15)^8+E(15)^11+E(15)^13+E(15)^14),
x_1+(E(15)+E(15)^2+3*E(15)^4+E(15)^7+E(15)^8+3*E(15)^11+E(15)^13+E(15)^14),
x_1+(E(15)+3*E(15)^2+E(15)^4+E(15)^7+E(15)^8+E(15)^11+3*E(15)^13+E(15)^14),
x_1+(3*E(15)+E(15)^2+E(15)^4+E(15)^7+E(15)^8+E(15)^11+E(15)^13+3*E(15)^14),
126
No. Polynomial
1
2
3
4
5
6
x_1-3
x_1+1
x_1+5
x_1+(E(5)+3*E(5)^2+3*E(5)^3+E(5)^4)
x_1+(3*E(5)+E(5)^2+E(5)^3+3*E(5)^4)
x_1+(E(15)+E(15)^2+E(15)^4+3*E(15)^7
+3*E(15)^8+E(15)^11+E(15)^13+E(15)^14)
7 x_1+(E(15)+E(15)^2+3*E(15)^4+E(15)^7
+E(15)^8+3*E(15)^11+E(15)^13+E(15)^14)
8 x_1+(E(15)+3*E(15)^2+E(15)^4+E(15)^7
+E(15)^8+E(15)^11+3*E(15)^13+E(15)^14)
9 x_1+(3*E(15)+E(15)^2+E(15)^4+E(15)^7
+E(15)^8+E(15)^11+E(15)^13+3*E(15)^14)
10 x_1^2-2*x_1-9
11 x_1^2-2*x_1-5
12 x_1^3+3*x_1^2-9*x_1-19
Eigenspace
dimension
5
5
1
3
3
3
Irreducible
representation
10
10
1
3
4
6
3
5
3
6
3
5
8
8
15
8
7
9
TABLE 37. The classification of the solutions of the generalized eigenvalue problem in
Q15 .
x_1^2-2*x_1-9, x_1^2-2*x_1-5, x_1^3+3*x_1^2-9*x_1-19 ]
gap> gs2:=GeneralizedEigenspaces(CF(15),hh);;
gap> gsbas2:=List(gs2,i->Basis(i));;
Remark. In fact, the splitting field of the polynomial x_1^4+6*x_1^3-4*x_1^2-54*x_1-29
is not Q15 ( of the extension of eight degrees to Q), but Q(a) (of the extension of four
degrees to Q), the latter of which is obtained by adding one of the solution “a” of
this polynomial to Q. It is equal to Q(E(15)+E(15)^14) (denoted as NF(15,[1,14])
in GAP). To be exact, one should discuss in this field.
The Galois group of Q15 is computed as.
gap> GaloisGroup(CF(15));
127
<group with 2 generators>
gap> g:=GaloisGroup(CF(15));
<group with 2 generators>
gap> g.1;
ANFAutomorphism( CF(15), 11 )
gap> g.2;
ANFAutomorphism( CF(15), 7 )
The generator g.1 is the map, replacing E(15)→ E(15)^11, of order 2, and g.2,
E(15)→ E(15)^7, of order 4. The latter generator, in table 37, transports the 6th eigenspace to the 7-th, the 7-th to the 8-th, in turn, and the 10-th to the 6-th,
generating the group C4 ( the forth cyclic group.), realizing the Galois group of
the polynomial x_1^4+6*x_1^3-4*x_1^2-54*x_1-29 as the following computation
verifies.
gap> ge2[6]*ge2[7]*ge2[8]*ge2[9]=ge[8];
true
gap> for j in [1..3] do;
> Print(RootsOfPolynomial(ge2[6])[1]^g.2=RootsOfPolynomial(ge2[7])[1]);
> Print(RootsOfPolynomial(ge2[7])[1]^g.2=RootsOfPolynomial(ge2[8])[1]);
> Print(RootsOfPolynomial(ge2[8])[1]^g.2=RootsOfPolynomial(ge2[9])[1]);
> Print(RootsOfPolynomial(ge2[9])[1]^g.2=RootsOfPolynomial(ge2[6])[1]);
>od;
truetruetruetruetruetruetruetruetruetruetruetrue
gap> for j in [1..3] do;
> Print(List(gsbas2[6][j],x->x^g.2)=gsbas2[7][j]);
128
> Print(List(gsbas2[7][j],x->x^g.2)=gsbas2[8][j]);
> Print(List(gsbas2[8][j],x->x^g.2)=gsbas2[9][j]);
> Print(List(gsbas2[9][j],x->x^g.2)=gsbas2[6][j]);
>od;
truetruetruetruetruetruetruetruetruetruetruetrue
The first part of the above computation shows that the polynomial ge[8] (the 8-th
solution in the field of rational numbers) splits into ge2[6],...,ge2[9], (the four solutions, from the 6-th to 9-th, in Q15 ). The second part verifies the C4 group property
in the transportation of the eigenvalues (the root of ge2[6],...,ge2[9]) , and the
third part verifies it in the transportation of the eigenvectors, gsbas2[6],...,gsbas2[9],
each of which includes three basis (gsbas2[X][1],gsbas2[X][2],gsbas2[X][3]).
The irreducible representations are transported as the following sequence by g.2,
I.5 → I.6 → I.5 → I.6. In this sequence, the same types of the irreducible representations appear two times, but due to the duplicity of these representations in
the whole eigenspace (as we have seen the decomposition of the trace representation I.Tr in the previous section) they are two distinct subspaces which belong to
the same type of the irreducible representation. Therefore the Galois group has
the order 4, running through the four distinct subspaces. The replacement of the
eigenspaces by this group is given in table38, from which one could extract the Galois
correspondence between sub-fields and subgroup.
This example may be regarded as a hidden symmetry in the sense that the wave
functions are transformed into each other by the operations of a group, which is
different from that of the structural symmetry of the molecule.
We can approach the hidden symmetry in another direction: there is an automorphism on the icosahedral group which interchanges these two irreducible representations. For simplicity, we reduce the problem in A5. (The icosahedral group is
129
<identity ...> 1 2 3 4 5 6 7
x2*x1
1235496
x2^2
1234589
x2
1235478
x2^-1*x1
1235478
x1
1234589
x2^-1
1235496
x2^2*x1
1234567
8
7
6
9
9
6
7
8
9
8
7
6
6
7
8
9
10
10
10
10
10
10
10
10
11
11
11
11
11
11
11
11
12
12
12
12
12
12
12
12
TABLE 38. The replacement of 12 eigenspaces by the automorphism in CF(15). The x1
and x2 are the generators of the Galois group, given as g.1 and g.2 in the article.
constructed from this by the direct product with the inversion.) The generators of
A5 are g1 and g2, (given as the permutations in the previous section.)
gap> G:=Group(g1,g2);;
gap> Irr(G);;
The character table is as follows (written briefly by a list in GAP).
[[ 1, 1, 1, 1, 1 ],
[ 3, -E(5)-E(5)^4, -E(5)^2-E(5)^3, -1, 0 ],
[ 3, -E(5)^2-E(5)^3, -E(5)-E(5)^4, -1, 0 ],
[ 4, -1, -1, 0, 1 ],
[ 5, 0, 0, 1, -1 ] ]
The conjugacy classes are computed now. (As the list of the classes is given as
the permutations, it is transformed to the representations by generators, to make the
list as brief as possible. The symbols x1 and x2 represent g1 and g2 respectively.)
gap> A:=List(ConjugacyClasses(G),Elements);;
gap> B:=List(A,i->List(i,j->Factorization(G,j)));
[ [ <identity ...> ],
130
[ x2^-1, x2, x2*x1*x2, x1*x2, x2^2*x1, x2*x1, x2^-1*x1^-1, x2^-2*x1^-1,
x2^-1*x1^-1*x2^-1, x1^-1*x2^-1, x1*x2^2, x1^-1*x2^-2 ],
[ x2^-2, x2^2, x2^2*x1*x2^2*x1^-1, x1*x2^2*x1^-1, (x1*x2^2)^2,
(x2^-1*x1^-1*x2^-1)^2, (x2*x1*x2)^2, x2^-1*x1^-1*x2^-2*x1*x2,
(x2^2*x1)^2, x1^-1*x2^-2*x1, (x2*x1)^2, x2^-1*x1*x2*x1 ],
[ x2^-2*x1*x2, x2^-1*x1, x2*x1^-1, x2^2*x1^-1*x2^-1, x2^2*x1*x2^2,
x1^-1*x2^-2*x1*x2*x1, x1*x2*(x2*x1)^2, (x2*x1*x2)^2*x1,
(x2^2*x1)^2*x2*x1, x2^-1*x1^-1*x2^-2*x1*x2*x1, x2*x1*x2^2*x1^-1,
(x2*x1*x2)^2*x2, (x2^2*x1)^2*x2, x2^-1*x1^-1*x2^-2*x1, x1*x2*x1 ],
[ x2^2*x1*x2, x2^-1*x1*x2, x2^-2*x1, x1, x2^2*x1^-1, x1^-1,
x2^-2*x1^-1*x2^-1, x2*x1^-1*x2^-1, x2^-1*x1^-1*x2^-2, x2*x1*x2^2,
x2^-1*x1^-1*x2^-2*x1^-1, x1^-1*x2^-2*x1^-1, x1^-1*x2^-2*x1^-1*x2^-1,
x1*x2^2*x1^-1*x2^-1, x1*x2^2*x1*x2, x1^-1*x2^-2*x1*x2, x1*x2^2*x1,
x2*x1*x2^2*x1, x2*(x2*x1)^2, x2^-2*x1*x2*x1 ] ]
The group of all automorphism is computed as
gap> AUALL:=AutomorphismGroup(G);
Take first of the generator, AUALL.1.
gap> AUALL.1;
[ (1,14,20)(2,15,16)(3,11,17)(4,12,18)(5,13,19)(6,58,22)(7,57,23)
(8,56,24)(9,60,25)(10,59,21)(26,54,38)(27,53,37)(28,52,36)(29,51,40)
(30,55,39)(31,44,50)(32,45,46)(33,41,47)(34,42,48)(35,43,49),
(1,2,3,4,5)(6,11,16,21,26)(7,12,17,22,27)(8,13,18,23,28)(9,14,19,24,29)
(10,15,20,25,30)(31,32,33,34,35)(36,41,46,51,56)(37,42,47,52,57)
(38,43,48,53,58)(39,44,49,54,59)(40,45,50,55,60) ] ->
[ (1,57,45)(2,56,44)(3,60,43)(4,59,42)(5,58,41)(6,55,47)(7,54,46)(8,53,50)
131
(9,52,49)(10,51,48)(11,35,28)(12,34,29)(13,33,30)(14,32,26)
(15,31,27)(16,37,24)(17,36,25)(18,40,21)(19,39,22)(20,38,23),
(1,59,9,16,55)(2,58,10,17,54)(3,57,6,18,53)(4,56,7,19,52)
(5,60,8,20,51)(11,14,12,15,13)(21,35,30,38,50)(22,34,26,37,49)
(23,33,27,36,48)(24,32,28,40,47)(25,31,29,39,46)(41,44,42,45,43) ]
This automorphism maps g1 and g2 in the following way.
gap> Factorization(G,Image(AUALL.1,g1));
x2*(x2*x1)^2
gap> Factorization(G,Image(AUALL.1,g2));
x1^-1*x2^-2*x1
The two conjugacy classes (of the second and of the third) are interchanged with
each other, while the other classes are unchanged. This causes the exchange of the
character values and that of the two irreducible representation (of No.2 and No.3).
Let us verify this. The image of the automorphism on the elements of the conjugacy
class 2 is computed and sorted as
gap> imgCl2:=List(A[2], a ->Image(AUALL.1,a));;
gap> SortedList(List(imgCl2,i->Factorization(G,i)));
[ x2^-2, x2^2, x1^-1*x2^-2*x1, x1*x2^2*x1^-1, x2^-1*x1*x2*x1, (x2*x1)^2,
(x1*x2^2)^2, (x2^-1*x1^-1*x2^-1)^2, x2^-1*x1^-1*x2^-2*x1*x2, (x2*x1*x2)^2,
x2^2*x1*x2^2*x1^-1, (x2^2*x1)^2 ]
The sorted list of the conjugacy class 3 is:
gap> SortedList(B[3]);
[ x2^-2, x2^2, x1^-1*x2^-2*x1, x1*x2^2*x1^-1, x2^-1*x1*x2*x1, (x2*x1)^2,
(x1*x2^2)^2, (x2^-1*x1^-1*x2^-1)^2, x2^-1*x1^-1*x2^-2*x1*x2, (x2*x1*x2)^2,
x2^2*x1*x2^2*x1^-1, (x2^2*x1)^2 ]
132
These two lists are identical. As the automorphism AUALL.1 has the order 2, the
conjugacy classes 2 and 3, or the irreducible representation (of No.2 and No.3), are
interchanged with each other by this automorphism.
As the each element of the group A5 has a correspondence with one of the vertexes
of the C60 molecule, this automorphism can be interpreted as a geometrical operation,
that of the replacement of vertex. But this replacement of atoms does not conserve
the bonding structure between nearest neighbors. This means that this replacement
of atoms is not the symmetry operation on the C60 .
H.
The super-symmetries
In this section, we will see that we can construct a symmetry higher than the
icosahedral one. The icosahedral group is the direct product of A5 and I. If one of
the components, the group A5, is extended to a larger group, the direct product is a
super-group of the icosahedral one. The bonding structure of the C60 is representable
by the pairs of the sequential numbers of the two vertexes, and it can be interpreted
as the “orbit” of these pairs by the action of the icosahedral group. Then the orbit
of the pairs of vertex, generated by the action of this new larger group, provides us
with the bonding structure in the higher symmetry. The guiding principle is based
the following observation: in the previous section, we have seen the existence of a
certain group automorphism which interchanges two of the irreducible representation.
If a group, including A5 and this group automorphism, could be constructed, the
two three-dimensional irreducible representations will be merged into one. And if
a molecule obeys this higher symmetry, the electronic structure will show different
features, such as sixfold degeneracies. Now, concerning the extension of a group, the
concept of the semi-direct product (with the automorphism group) is usable. For this
purpose, we should construct a group composed from all of the automorphisms on
133
the group A5. Taking this automorphism group or its subgroups, we can construct
the semi-direct product with the group A5. The necessary computation is executed
in the following way. As the full automorphism group of A5 is too large, we take a
small subgroup of this (a subgroup generated by one of the generators of this group).
gap> G:=Group(g1,g2);
gap> AUALL:=AutomorphismGroup(G);
gap> p:=SemidirectProduct(Group(AUALL.1),G);
The newly created group “p” is larger than A5 (the group “G”), and contains A5
as a subgroup. The three generators are given as
gap> p.1;
(3,16,22)(4,17,18)(5,13,19)(6,14,20)(7,15,21)(8,60,24)(9,59,25)(10,58,26)
(11,62,27)(12,61,23)(28,56,40)(29,55,39)(30,54,38)(31,53,42)(32,57,41)
(33,46,52)(34,47,48)(35,43,49)(36,44,50)(37,45,51)
gap> p.2;
(3,7,6,5,4)(8,28,23,18,13)(9,29,24,19,14)(10,30,25,20,15)(11,31,26,21,16)(12,
32,27,22,17)(33,37,36,35,34)(38,58,53,48,43)(39,59,54,49,44)(40,60,55,50,
45)(41,61,56,51,46)(42,62,57,52,47)
gap> p.3;
(1,2)(4,61)(5,11)(6,18)(7,57)(8,21)(9,28)(10,42)(12,46)(13,54)(14,39)
(15,49)(16,59)(17,44)(19,51)(20,26)(22,47)(23,58)(25,34)(27,45)(29,36)
(31,55)(32,43)(33,50)(37,38)(40,48)(41,60)(52,53)
The first and the second generators are inherited from A5, and the third is the new
one. They are represented by 62 symbols. The symbols from 3 to 62 correspond to
the 60 vertexes in one C60 . The extension of the group introduces extra symbols “1”
and “2”. We can give a geometrical interpretation to this group in the following way.
134
From the 62 symbols, we can construct a bonding system which contains 120 atoms,
located in the double C60 molecules. The atoms will be indexed as [i,j] (i=3,..,62,
and j=1,2) by the duplication of single C60 . This structure (C120 ) is regulated by the
symmetry of the new extended group. The operation p.1 and p.2 are the rotations
by 120 and 72 degrees. The operation p.3 interchanges 60 vertexes between two C60
shells. To see this, we have only to compute the orbit of the vertex at [I, J] by
these operations. The permutation (1,2), applied to the latter element of the pairs,
represents the exchange of vertexes between the two C60 .
By means of a function defined below, the bonding structure in this new symmetry
can be set up in matrices (of 120 × 120 dimension). If one bond is bridged between
the vertex [i[1],i[2]] and [i[3],i[4]], the corresponding element in the networking matrix
is set to be the unity.
listforh:=function(A)
local h,i1,i2,i3,i4,k,l,i;
h:=List([1..120],i->List([1..120],j->0));
for i in A do
i1:=i[1];
i2:=i[2];
i3:=i[3];
i4:=i[4];
k:=(i1-2)+60*(i2-1);
l:=(i3-2)+60*(i4-1);
h[k][l]:=1;
h[l][k]:=1;
od;
return h;
135
end;
The orbits of the double bond [3,1,4,1] (read as from [3,1] to [4,1]), the single
bond [3,1,8,1], and the bond between the first and the second C60 ( such as [3,1,4,2],
a possible example, taken par hazard) are computed into the lists “A1”,“B1”,“C1”
as bellow. The bonding structures generated from them are represented by three
matrices “a1”,“b1”,“c1”.
gap> A1:=Orbit(p,[3,1,4,1],OnTuples);
[ [ 3, 1, 4, 1 ], [ 16, 1, 17, 1 ], [ 4, 1, 5, 1 ], [ 3, 2, 61, 2 ],
[ 22, 1, 18, 1 ], [ 21, 1, 22, 1 ], [ 59, 2, 44, 2 ], [ 17, 1, 13, 1 ],
[ 5, 1, 6, 1 ], [ 61, 2, 11, 2 ], [ 16, 2, 23, 2 ], [ 4, 2, 41, 2 ],
...(omitted hereafter)...
gap> B1:=Orbit(p,[3,1,8,1],OnTuples);;
gap> C1:=Orbit(p,[3,1,4,2],OnTuples);;
gap> a1:=listforh(A1);;
gap> b1:=listforh(B1);;
gap> c1:=listforh(C1);;
In the C60 , the symmetry operation of the icosahedral symmetry includes the
inversion. The inversion itself is not included in the new group, but we can furnish
this operation with first sixty vertexes (from the third to the sixty-second) in the
same manner as the icosahedral case. And the inversion in the second 60 vertexes can
be written in the following way. The second sixty vertex are transported from the
first sixty, by the operation p.3. If the inversions on the first and second 60 vertexes
are denoted as I1, I2, they are related by the conjugation I2 := (p.3) · I1 · (p.3)−1
(symbolically written here). The inversion on the total system is written by their
136
union as
I1 0
0 I2
where the each entry is assumed to be 60×60-dimensional matrices.
The destinations of the vertexes by the inversion “g3” in one C60 shell are stored
in the list bellow.
gap> ListInv:=List([1..60],i->OnPoints(i,g3));
The destinations of the vertexes in the first C60 shell to the second one, by the
operation “p.3”, is stored in the list bellow.
gap> mconjg:=List([3..62],i->OnPoints(i,p.3))-2;
(In the last computation, we must extract 2 from the result, due to the numbering
in the atoms, which are numbered from 3 to 62.)
The inversions in the first and the second C60 shell are composed now (as matrices
“invrs”, “invrs2”, and by uniting them, the inversion in the full system is computed,
as in the matrix “ivfull”.
gap> invrs:=List([1..60],i->List([1..60],j->0));
gap> mc:=List([1..60],i->List([1..60],j->0));
gap> for i in [1..60] do
invrs[i][ListInv[i]]:=1;
mc[i][mconjg[i]]:=1;
od;
gap> invrs2:=mc*invrs*mc^-1;
gap> ivfull:=List([1..120],i->List([1..120],j->0));
gap> for i in [1..60] do
137
for j in [1..60] do
ivfull[i][j]:=invrs[i][j];
ivfull[i+60][j+60]:=invrs2[i][j];
od;
od;
The generators of the group “p”, as the symmetry operations of the 120 atoms,
are also represented by matrices. At first, we prepare the lists to represent the
permutation.
gap> l1:=List([1..2],j->List([3..62],i->(OnPoints(i,p.1)+60*(j-1))));
gap> l1:=Flat(l1);
gap> l2:=List([1..2],j->List([3..62],i->(OnPoints(i,p.2)+60*(j-1))));
gap> l2:=Flat(l2);
gap> l3:=List([1..2],j->List([3..62],i->(OnPoints(i,p.3)+60*(OnPoints(j,p.3)-1))));
gap> l3:=Flat(l3);
The permutations are transformed to the matrix forms by this function.
ListToMat:=function(l)
local size,h,i,j;
size:=Size(l);
h:=List([1..size],i->List([1..size],j->0));
for i in [1..size] do
h[i][l[i]]:=1;
od;
return h;
end;
138
gap> P1:=ListToMat(l1-2);
gap> P2:=ListToMat(l2-2);
gap> P3:=ListToMat(l3-2);
We must examine the congruence of the computed data, by checking whether
the matrices representing the bonding structure are subject to the symmetry of the
extended group. For this purpose, we can compute the commutators between these
matrices (“a1”,“b1”,“c1”) and the generators of this group (P1, P2, P3 and ivfull):
the results should be zero. The matrices “a1” and “b1” pass the test, but “c1”
not. As the matrix “c1” is not conserved by the matrix of the inversion “ivfull”
(or not commutable), the symmetrization is necessary for the construction of the
Fock matrix. One of the Fock matrices is given as, fixing the values of the transfer
integrals,
gap> hh:=-2*a1-b1-1/2*(c1+ivfull*c1*ivfull);
The generalized eigenvalue problems is solved up to in the field of rational numbers.
gap> gev:=GeneralisedEigenvalues(Rationals,hh);
[ x_1-3, x_1+1, x_1+3, x_1+7, x_1^2-4*x_1-13/4, x_1^2-4*x_1-2,
x_1^2-2*x_1-81/4, x_1^2-2*x_1-9/4, x_1^2-12, x_1^2-21/4,
x_1^3+5*x_1^2-4*x_1-24, x_1^4+6*x_1^3-6*x_1^2-60*x_1-36 ]
gap> ges:=GeneralisedEigenspaces(Rationals,hh);
[ <vector space over Rationals, with 10 generators>,
<vector space over Rationals, with 15 generators>,
<vector space over Rationals, with 1 generators>,
<vector space over Rationals, with 1 generators>,
<vector space over Rationals, with 8 generators>,
139
<vector space over Rationals, with 12 generators>,
<vector space over Rationals, with 8 generators>,
<vector space over Rationals, with 8 generators>,
<vector space over Rationals, with 10 generators>,
<vector space over Rationals, with 8 generators>,
<vector space over Rationals, with 15 generators>,
<vector space over Rationals, with 24 generators> ]
The group is composed and the irreducible representation is computed now.
gap> irrg:=Irr(Group(P1,P2,P3,ivfull));
[ Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 1, -1, -1, -1, -1, -1, 1, 1, 1, 1, 1, 1, 1, -1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 1, -1, -1, -1, 1, 1, 1, -1, 1, -1, 1, -1, -1, 1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 1, 1, 1, 1, -1, -1, 1, -1, 1, -1, 1, -1, -1, -1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 4, -2, 1, 0, 1, -2, 0, -1, -1, 1, 1, 0, 4, 0 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 4, 2, -1, 0, -1, 2, 0, -1, -1, 1, 1, 0, 4, 0 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 4, -2, 1, 0, -1, 2, 0, 1, -1, -1, 1, 0, -4, 0 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 4, 2, -1, 0, 1, -2, 0, 1, -1, -1, 1, 0, -4, 0 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
140
[ 5, 1, 1, -1, 1, 1, 1, 0, 0, -1, -1, 1, 5, -1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 5, -1, -1, 1, -1, -1, 1, 0, 0, -1, -1, 1, 5, 1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 5, 1, 1, -1, -1, -1, 1, 0, 0, 1, -1, -1, -5, 1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 5, -1, -1, 1, 1, 1, 1, 0, 0, 1, -1, -1, -5, -1 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 6, 0, 0, 0, 0, 0, -2, 1, 1, 0, 0, -2, 6, 0 ] ),
Character( CharacterTable( <matrix group of size 240 with 4 generators> ),
[ 6, 0, 0, 0, 0, 0, -2, -1, 1, 0, 0, 2, -6, 0 ] ) ]
Solution
x_1-3
x_1+1
x_1+3
x_1+7
x_1^2-4*x_1-13/4
x_1^2-4*x_1-2
x_1^2-2*x_1-81/4
x_1^2-2*x_1-9/4
x_1^2-12
x_1^2-21/4
x_1^3+5*x_1^2-4*x_1-24
x_1^4+6*x_1^3-6*x_1^2-60*x_1-36
Representation
[ 11, 12 ]
[ 10, 11, 12 ]
[2]
[1]
[6]
[ 13 ]
[8]
[7]
[ 10 ]
[5]
[9]
[ 14 ]
Dimension
5
5
1
1
4
6
4
4
5
4
5
6
TABLE 39. The correspondence between the solutions of the generalized eigenvalue problem and the irreducible representations. The left column shows the polynomial, the center
the corresponding irreducible representation by sequential numbers, and the right the dimension of the irreducible representation.
The four eigenvalues, as the solutions of “x_1^4+6*x_1^3-6*x_1^2-60*x_1-36”,
composed by the <vector space over Rationals, with 24 generators>, are of
141
the sixfold degeneracies, allotted to the 14th irreducible representation, which we
have expected to see in this structure. Although this example is rather an imaginary one, it suggests us the importance of the concept of “group extension” in the
artificial design of the electronic structure in molecular systems of high symmetries.
Usually, we argue the electronic structure from the viewpoint of the splitting of the
degeneracies in accordance with the reduction of the symmetry, but we will be able
to proceed in the inverse direction by actually composing the “super-group”.
I.
Remarks to this section
In this section, with the viewpoint of the group theory, we analyze the electronic
properties of C60 . The group theoretical properties of the icosahedral symmetry of
C60 are computed by a computer algebra program package GAP, and the obtained
results are utilized to the analysis. We started in a traditional way in order to classify
the eigenstates into irreducible representations. However, the possible applications of
the computer algebra are not limited to this sort of the rather classical analysis, since
we can construct the group and extract the group theoretical properties more freely
and adroitly by the computer algebra than by the utilization of the group theoretical
data given in some references. We can construct the all of the subgroups of the
icosahedral symmetry. By means of the “orbitals” of the movement of the vertex by
the symmetry operations, the correspondence between the subgroup and the possible
deformation of the molecule can be established. The change in the energy spectra
and the existence of the zero or nonzero transition moments could be discussed
qualitatively, without massive computation. Furthermore, through the construction
of the semi-direct group of the icosahedral group, the structure equipped with a
symmetry higher than that of the icosahedral symmetry, can be generated. The
interest with relation to the “molecular design” is as follows: the higher symmetry
142
in the semidirect product merges some of the irreducible representations into one
and gives rise to a higher degeneracy of the energy spectra. These examples suggest
the importance of the group theoretical viewpoint in the material design, re-armored
(or revived) by the modern computational tools, developed as “computer algebra”
among mathematicians.
VIII.
ANALYSIS OF VIBRATIONAL MODE IN C60
In this section, the computation of the vibrational modes of the C60 by means
of the computer algebra is demonstrated. The C60 molecule has the unique symmetry, comparable to one soccer-ball (see Fig. 6). The symmetry of the molecule is
described by the icosahedral group, and this group is represented by the permutation of 60 vertexes or the group of rotation matrices operating on them. Owing to
this high symmetry, the molecule shows peculiar electronic and dynamic properties,
both of scientific and industrial interest, and the vibrational mode might be one of
the principal phenomena in governing the quantum dynamics in it. As a tool of
the computation, we utilize the software GAP, developed in the field of the pure
mathematics[12]. This software can construct the symmetry group and compute the
irreducible representations and other group-theoretical properties and enable us to
put into practice the application in the material science, possibly in a modernized
way. The computation presented hereafter will be useful to the working physicists
and the students, who want to deepen the understanding of the group theory that is
given in the textbooks[2, 4, 5, 8, 9].
143
A.
The vibrational modes of C60
The atoms and the pentagons in the molecule are provided with sequential numbers as in Fig. 7 in the preceding section. The symmetry operations on the spatial
vectors on the vertexes of C60 are made by the direct products of the two types
of matrices. The first type Ri represents the replacement of 60 vertexes and the
second type Gi the rotation and the reflection of the spatial vectors. The symmetry
operations on the spatial vectors placed into 60 vertexes are represented as this:
Oi = Rif ⊗ Gi ,
(86)
where Ri and Gi should represent the equivalent operation, and the superscript f
is the group automorphism in the group R. The matrices generates the icosahedral
group, and the trace of Oi provides us with the trace representation χtr (Oi )(not irreducible). The decomposition of the trace representation by means of the irreducible
representations indicates the distinct modes of the vibration, and the projection operators composed by Oi can generate the basis vectors for them. And to obtain the
vibrational mode, the degrees of the freedoms of the translation and the rotation of
the molecule should be removed:
χtr − χT rans − χRot .
(87)
The presence of the automorphism f is inevitable. Consider the operation r =
(1, 2, 3, 4, 5) · · · (· · · ). This operation, as the rotation G, acts on the dislocated ver144
texes (V ([j] → V [j] + d[j]):
G(V [1] + d[1]) = V [2] + G · d[1]
G(V [2] + d[2]) = V [3] + G · d[2]
G(V [3] + d[3]) = V [4] + G · d[3]
G(V [4] + d[4]) = V [5] + G · d[4]
G(V [5] + d[5]) = V [1] + G · d[5]
····················· .
The operation on d[i] should be represented as
0 0 0 0 G ···
G
0
0
0
..
.
d[1]
d[2]
0 0 0 0 ···
G 0 0 0 · · · d[3]
d[4]
0 G 0 0 ···
0 0 G 0 · · · d[5]
.. .. .. ..
..
. . . .
.
so that the rotation to d[j] should be located at the correct position, while the
permutation r is represented in the form which is the transposition of that operation:
0 1 0 0 0 ···
0
0
0
1
..
.
0 1 0 0 ···
0 0 1 0 ···
0 0 0 1 ···
0 0 0 0 ···
.. .. .. ..
. . . .
145
It happens that the permutation r should act on d[j] as its transposition, or its
inverse. And the product of the operations is given by (R1 , G1 ) ◦ (R2 , G2 ) = (R2 ·
R1 , G1 · G2 ), with inverse directions for Ri and Gi . This observation shows the
inadequacy of the simple definition of the direct product Ri ⊗ Gi as the symmetry
operation. However, the naive redefinition, such as Ri−1 ⊗ Gi for all i, does not
work well. Now we can make use the fact the conjugation (A → M · A · M−1 ) is an
automorphism which preserves the sequential order of the group multiplication; we
should find the elements which inverts the generators of the group by the conjugation
in order to define the exact symmetry operations, the products of which go from
right to left both for Rif and Gi . Indeed this can be done in the final stage of the
computation. So, before stepping into it, let us execute the naive computation (with
the direct product Ri ⊗ Gj ). Concerning the dimension counting of the basis set,
such a computation does not fail.
The icosahedral group is defined in several ways. We prepare the twelve points,
located at the centers of the pentagons in the molecule (or the vertexes in the icosahedron) as follows. E(5) is the fifth root of the unity, exp(2πi/5), and the symbolically
presented data are all real. (The data and results are given by the expressions in
GAP.)
1
[0,0,1],
2
[2/5*E(5)-2/5*E(5)^2-2/5*E(5)^3+2/5*E(5)^4,
0,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4],
3
[-3/5*E(5)-2/5*E(5)^2-2/5*E(5)^3-3/5*E(5)^4,
1/5*E(20)-1/5*E(20)^9-2/5*E(20)^13+2/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4],
4
[2/5*E(5)+3/5*E(5)^2+3/5*E(5)^3+2/5*E(5)^4,
146
2/5*E(20)-2/5*E(20)^9+1/5*E(20)^13-1/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4],
5
[2/5*E(5)+3/5*E(5)^2+3/5*E(5)^3+2/5*E(5)^4,
-2/5*E(20)+2/5*E(20)^9-1/5*E(20)^13+1/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4],
6
[-3/5*E(5)-2/5*E(5)^2-2/5*E(5)^3-3/5*E(5)^4,
-1/5*E(20)+1/5*E(20)^9+2/5*E(20)^13-2/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4],
7
[0,0,-1],
8
[-2/5*E(5)+2/5*E(5)^2+2/5*E(5)^3-2/5*E(5)^4,
0,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4],
9
[3/5*E(5)+2/5*E(5)^2+2/5*E(5)^3+3/5*E(5)^4,
-1/5*E(20)+1/5*E(20)^9+2/5*E(20)^13-2/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4],
10 [-2/5*E(5)-3/5*E(5)^2-3/5*E(5)^3-2/5*E(5)^4,
-2/5*E(20)+2/5*E(20)^9-1/5*E(20)^13+1/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4],
11
[-2/5*E(5)-3/5*E(5)^2-3/5*E(5)^3-2/5*E(5)^4,
2/5*E(20)-2/5*E(20)^9+1/5*E(20)^13-1/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4],
12
[3/5*E(5)+2/5*E(5)^2+2/5*E(5)^3+3/5*E(5)^4,
1/5*E(20)-1/5*E(20)^9-2/5*E(20)^13+2/5*E(20)^17,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4]
These twelve points are transported among themselves by these three operations
(g1,g2,g3). The three operations (g1,g2,g3) induces the three permutations (f1,f2,f3)
147
on the 12 vertexes of the icosahedron, and(r1,r2,r3) on 60 vertexes in the C60 .
gap> g1:=[ [ 3/10*E(5)+1/5*E(5)^2+1/5*E(5)^3+3/10*E(5)^4,
1/10*E(20)-1/10*E(20)^9-1/5*E(20)^13+1/5*E(20)^17,
2/5*E(5)-2/5*E(5)^2-2/5*E(5)^3+2/5*E(5)^4 ],
[ -1/2*E(20)+1/2*E(20)^9, -1/2*E(5)-1/2*E(5)^4, 0 ],
[ -3/5*E(5)-2/5*E(5)^2-2/5*E(5)^3-3/5*E(5)^4,
-1/5*E(20)+1/5*E(20)^9+2/5*E(20)^13-2/5*E(20)^17,
1/5*E(5)-1/5*E(5)^2-1/5*E(5)^3+1/5*E(5)^4 ] ];
gap> g2:=[ [ 1/2*E(5)+1/2*E(5)^4, 1/2*E(20)-1/2*E(20)^9, 0 ],
[ -1/2*E(20)+1/2*E(20)^9, 1/2*E(5)+1/2*E(5)^4, 0 ], [ 0, 0, 1 ] ];
gap> g3:=[ [ -1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ];
gap> G:=Group(g1,g2,g3);
gap> f1:=(1,2,3)(4,6,11)(5,10,12)(7,8,9);
gap> f2:=(2,3,4,5,6)(8,9,10,11,12);
gap> f3:=(1,7)(2,8)(3,9)(4,10)(5,11)(6,12);
gap> F:=Group(f1,f2,f3);
gap> r1:=(
(
1, 14, 20)(
6, 58, 22)(
2, 15, 16)(
7, 57, 23)(
3, 11, 17)(
8, 56, 24)(
4, 12, 18)(
5, 13, 19)
9, 60, 25)( 10, 59, 21)
( 31, 44, 50)( 32, 45, 46)( 33, 41, 47)( 34, 42, 48)( 35, 43, 49)
( 36, 28, 52)( 37, 27, 53)( 38, 26, 54)( 39, 30, 55)( 40, 29, 51);
gap> r2:=(
1,
2,
3,
4,
5)(
6, 11, 16, 21, 26)(
148
7, 12, 17, 22, 27)
(
8, 13, 18, 23, 28)(
9, 14, 19, 24, 29)( 10, 15, 20, 25, 30)( 31, 32, 33, 34, 35)
( 36, 41, 46, 51, 56)( 37, 42, 47, 52, 57)( 38, 43, 48, 53, 58)( 39, 44, 49, 54, 59)
( 40, 45, 50, 55, 60);
gap> r3:=(
(
1, 31)(
2, 32)(
3, 33)(
4, 34)(
5, 35)(
6, 36)(
7, 37)(
8, 38)
9, 39)( 10, 40)( 11, 41)( 12, 42)( 13, 43)( 14, 44)( 15, 45)( 16, 46)( 17, 47)
( 18, 48)( 19, 49)( 20, 50)( 21, 51)( 22, 52)( 23, 53)( 24, 54)( 25, 55)( 26, 56)
( 27, 57)( 28, 58)( 29, 59)( 30, 60);
gap> R:=Group(r1,r2,r3);
gap> Irr(R);
The command “Irr(R)” returns the character table of R, a part of which is listed
in table 40:
Rep
N.1
N.2
N.3
N.4
N.5
N.6
N.7
N.8
N.9
N.10
Dim
1
1
3
3
3
3
4
4
5
5
[f1,f2,f3]
[ 1, 1, 1 ]
[ 1, 1, -1 ]
[ 0,-E(5)-E(5)^4, 3 ]
[ 0,-E(5)^2-E(5)^3, 3 ]
[ 0,-E(5)-E(5)^4, -3 ]
[ 0,-E(5)^2-E(5)^3, -3 ]
[ 1, -1, 4 ]
[ 1, -1, -4 ]
[ -1, 0, 5 ]
[ -1, 0, -5 ]
Ag
Au
F2g
F1g
F2u
F1u
Hg
Hu
Gg
Gu
TABLE 40. The characters of the generators in ten irreducible representations of the
icosahedral group.
We use the group R as the standard. The isomorphism from R to G is set up at
the three generators (from [R.1,R.2,R.3] to [G.1,G.2,G.3].)
gap> hom:=GroupHomomorphismByImages(R,G,[R.1,R.2,R.3],[G.1,G.2,G.3]);
149
It is necessary to judge the irreducible representations to which the translation belongs. The projectors (applicable to three-dimensional spatial vectors) are computed
now for this purpose. They are defined as
P (p) =
l(p) X (p)∗
χ (T ) · O(T )
|G| T ∈G
by summing up the elements T of the group G, where l(p) is the dimension of the
p-th irreducible representation, |G| the order of the group, χ(p) the character, O(T )
the operation of the group. The gap script is given as:
Projopr:=function(irr,Em,hom)
#
# irr: Irreducible representation of the group.
# Em: Elements in the group.
# hom: Homomorphism of the permutation group to the matrix group
#
return List(irr,i->Sum(List(Em,j->j^i*Image(hom,j)))/Size(Em)*i[1]);
end;
In this function, the elements “j” (in the list “Em”) in the given group are mapped
to another group by the homomorphism “hom”. The characters are computed in the
first group (with the irreducible representation “irrep”) and multiplied to the images
of “j” in the second group. The summation over them produces the projector. In
this definition, the projectors of the matrix group G are composed through the
isomorphism from the permutation group R to G, not directly from G. This is a
precaution to avoid a possible confusion, since the character tables of R and G,
computed by GAP, might be presented in the slightly different orders, even if these
two groups are isomorphic. The result is given in table 41. The non-zero one exists
150
only in the representation N.6, so we conclude the spatial translation belongs to this
representation.
N.1
N.2
N.3
N.4
N.5
N.6
N.7
N.8
N.9
N.10
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
[
0,
0,
0,
0,
0,
1,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0
0
0
0
0
0
0
0
0
0
],
],
],
],
],
],
],
],
],
],
Operator
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 1, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[ 0, 0, 0 ],
[
[
[
[
[
[
[
[
[
[
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0
0
0
0
0
1
0
0
0
0
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
]
TABLE 41. The projection operators for the spatial vectors.
.
There is another necessity to judge the representation of the spatial rotations.
The group operation T on the rotation matrix M is defined as the product from both
sides, T · M · Tt . This operation should be treated with care. This product is also
a rotation matrix, but the summation of such products not. Thus the projection
operator defined as above will not generate the proper rotation matrix if it is naively
applied to the rotation matrices. Concerning this, the exponential map is useful,
which is defined as
M = exp(A),
T · M · Tt = exp(T · A · Tt ).
The pre-image A of the rotation M is the skew-symmetric matrix. The projection
operator should be applied to the linear space of these pre-images. The gap computation for the determination of the irreducible representation is now presented. The
151
projectors are computed by the formula
P (p) (A) =
l(p) X (p)∗
χ (T )T · A · T t
|G| T ∈G
by means of the function
ProjoprR:=function(irr,Em,MM,hom)
return List(irr,i->Sum(List(Em,j->j^i*Image(hom,j)
*MM*TransposedMat(Image(hom,j))))/Size(Em)*i[1]);
end;
The computation is done now.
gap> x:=Indeterminate(Rationals,"x");
gap> y:=Indeterminate(Rationals,"y");
gap> z:=Indeterminate(Rationals,"z");
gap> MM:=[[0*x,x,y],[-x,0*x,z],[-y,-z,0*x]];
gap> ProjoprR(Irr(R),Elements(R),MM,hom);
Remark. The expressions such as 0 and 0*x signify different objects. The former is
the zero as a number, and the latter the zero monomial. If one promiscuously uses
0 and 0*x, it might cause some errors.
The non-zero projector is the fourth one. The spatial rotations belong to the
fourth representation (N.4).
Remark. In the basis vectors of the vibrational mode computed hereafter, those which
152
N.1
N.2
N.3
N.4
N.5
N.6
N.7
N.8
N.9
N.10
Operator
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, x, y ], [-x, 0, z ], [-y, -z, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
[ [ 0, 0, 0 ], [ 0, 0, 0 ], [ 0, 0, 0 ] ]
TABLE 42. The projection operators for the rotations.
.
belong to the fourth irreducible representation, include the elements such as
Ii
0
0
.
..
0 0 ···
Ii 0 · · ·
·v
0 Ii · · ·
.. ..
. .
where v are the basis vector in the trivial irreducible representation (invariant by the
symmetry operation), and Ii (i = 1, 2, 3) are
0 0 0
0 0 1
0 1 0
I1 = −1 0 0 , I2 = 0 0 0 , I3 = 0 0 1 .
0 −1 0
−1 0 0
0 0 0
These vectors represent the degrees of freedom of the rotation. They are set of 60
of three-dimensional vectors, defined on each vertex, and located in the tangent plane
to the sphere where the 60 vertexes are placed.
Now we construct the symmetry operation (Oi = Ri ⊗ Gi ) on the spatial vectors
153
on the 60 vertexes. The following function converts a permutation Ri to an integer
matrix which represents the replacement of vertexes.
ListToMat:=function(l)
local size,h,i,j;
size:=Size(l);
h:=List([1..size],i->List([1..size],j->0));
for i in [1..size] do
h[i][l[i]]:=1;
od;
return h;
end;
The following function constructs the direct product of Ri and Gi . The isomorphism “iso” from the group R to G convert an element “rr” (in R) to the equivalence
in G exactly. In the generated matrix, the non-zero entries in Ri are replaced by Gi
(the rotation), as this.
0
1
0
·
·
·
0
0 Gi 0 · · ·
1 0 0 ··· 0
Gi 0 0 · · ·
Ri = .
→
R
⊗
G
=
i
i
..
.
.
.
0 0 0 ··· 1
0 0 0 ···
0
0
..
.
Gi
LargeMatrix:=function(rr,iso)
local l,rmat,rrimg,i,i1,i2,n,m;
l:=ListPerm(rr,60);
rmat:=List([1..Size(l)*3],i->List([1..Size(l)*3],j->0));
rrimg:=Image(iso,rr);
for i in [1..Size(l)]
154
do
i1:=i-1;i2:=l[i]-1;
for n in [1..3]
do
for m in [1..3]
do
rmat[3*i1+n][3*i2+m]:=rrimg[n][m];
od;
od;
od;
return rmat;
end;
This function at first prepares a 180×180 matrix, in order to represent the symmetry
operation on the vectors at the vertex of C60 . The symmetry operation of the group
element “rr” is transformed by the isomorphism “iso” to Gi and placed in the large
matrix, by means of the list “l” which stores the data of the permutation Ri .
The projector function, composed from the direct product, is defined as:
ProjoprE:=function(irr,Em,hom)
#
# irr:Irreducible representation.
# Em :Elements of the group R.
# hom:Isomorphism from the group R to the group G.
#
return List(irr,i->Sum(List(Em,j->j^i*LargeMatrix(j,hom)))/Size(Em)*i[1]);
end;
We can make now the basis set of the mode of deformation. The projectors are
155
placed on the list of ten matrices (of dimension 180) (PRJE[1]...PRJE[10]), and each
of them does the projection on each irreducible representations (N.1...N.10). The
basis vectors for the each vibrational mode are generated from the column vectors of
the corresponding projector. In the computation, the list “BASSIZE” gives us the
dimensions of each mode.
gap> PRJE:=ProjoprE(Irr(R),Elements(R),hom);;
gap> BAS:=List(PRJE,i->BaseMat(TransposedMat(i)));;
gap> BASSIZE:=List(BAS,Size);
[ 2, 1, 12, 12, 15, 15, 24, 24, 40, 35 ]
gap> BAS;
[ [ [ 1, E(20)-E(20)^9+E(20)^13-E(20)^17, 0, 2*E(5)+E(5)^2+E(5)^3+2*E(5)^4,
-E(20)^13+E(20)^17, 0, -2*E(5)-2*E(5)^4, 0, 0,
2*E(5)+E(5)^2+E(5)^3+2*E(5)^4, E(20)^13-E(20)^17, 0, 1,
-E(20)+E(20)^9-E(20)^13+E(20)^17, 0,
-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4,
1/5*E(20)-1/5*E(20)^9+3/5*E(20)^13-3/5*E(20)^17,
12/5*E(5)+8/5*E(5)^2+8/5*E(5)^3+12/5*E(5)^4,
-6/5*E(5)-4/5*E(5)^2-4/5*E(5)^3-6/5*E(5)^4,
...........................................
...........................................
Remark. One might demand numerical basis as well as symbolic one. The latter can
be easily modified into quasi-Fortran form. For example, by replacements of strings,
a formula such as
f:=-1/5*E(5)+1/5*E(5)^2+1/5*E(5)^3-1/5*E(5)^4;
is rewritten as
156
FUNCTION E(I)
INTEGER I
COMPLEX*16 E
REAL*8 TWOPI
TWOPI=8.0D0*ATAN(1.0D0)
E=DCMPLX(COS(TWOPI/I),SIN(TWOPI/I))
RETURN
END
SUBROUTINE(f)
COMPLEX*16 f,E
f=-1./5*E(5)+1./5*E(5)**2+1./5*E(5)**3-1./5*E(5)**4
RETURN
END
At present projectors “PRJE[i]” are defined by the “bare” forms, using the direct
product Ri ⊗ Gi . and they should be modified to the correct ones, such as Mf ·
f
PRJE[i] · M−1
f , composed by the direct product Ri ⊗ Gi , where f is a conjugation
which inverts the generators of the permutation group R. There are two candidates
for such automorphisms.
gap>RPC:=Filtered(Elements(R),x->List([R.1,R.2,R.3],i->i^x)=[R.1^-1,R.2^-1,R.3^-1]);
[ (1,4)(2,3)(6,21)(7,25)(8,24)(9,23)(10,22)(11,16)(12,20)(13,19)(14,18)(15,
17)(27,30)(28,29)(31,34)(32,33)(36,51)(37,55)(38,54)(39,53)(40,52)(41,
46)(42,50)(43,49)(44,48)(45,47)(57,60)(58,59),
(1,34)(2,33)(3,32)(4,31)(5,35)(6,51)(7,55)(8,54)(9,53)(10,52)(11,46)(12,
50)(13,49)(14,48)(15,47)(16,41)(17,45)(18,44)(19,43)(20,42)(21,36)(22,
40)(23,39)(24,38)(25,37)(26,56)(27,60)(28,59)(29,58)(30,57) ]
157
gap>
RPC[1]*r3=RPC[2];
true
The conjugation is computed as a matrix through this function.
ConjugMatrix:=function(rr)
local l,rmat,vmat,i,i1,i2,n,m;
l:=ListPerm(rr,60);
vmat:=List([1..Size(l)*3],i->List([1..Size(l)*3],j->0));
for i in [1..Size(l)]
do
i1:=i-1;i2:=l[i]-1;
for n in [1..3]
do
vmat[3*i1+n][3*i2+n]:=1;
od;
od;
return vmat;
end;
gap> PCM:=ConjugMatrix(RPC[1]);
gap> PRJEC:=List(PRJE,i->i^PCM);;
gap> BASC:=List(PRJEC,i->BaseMat(TransposedMat(i)));;
gap> BASCSIZE:=List(BASC,Size);
[ 2, 1, 12, 12, 15, 15, 24, 24, 40, 35 ]
In fact, in order to obtain the correct vibrational modes, it is enough to multiply the
matrix “PCM” for the conjugation to the basis vectors “BAS” computed by the bare
158
projectors. If we neglect such corrections, it happens that the “breathing mode”
would not belong to the trivial representation.
To check the validity, the characters of χT r and those by the irreducible representations are computed in the two lists “CTRC” and “CIRR”. The irreducible
representation of χV ec is located at the sixth component of CIRR, and that of χRot
is at the fourth.
CTRC:=List(Elements(R),i->Trace(ListToMat(ListPerm(i,60))));
CIRR:=List(Irr(R),r->List(Elements(R),i->i^r));
The χT r × χV ec is computed and decomposed.
gap> List([1..Size(R)],i->CTRC[i]*CIRR[6][i])*TransposedMat(CIRR)/Size(R);
[ 2, 1, 4, 4, 5, 5, 6, 6, 8, 7 ]
gap> List([1..10],i->BASSIZE[i]/Irr(R)[i][1]);
[ 2, 1, 4, 4, 5, 5, 6, 6, 8, 7 ]
These results are the multiplicities to each irreducible representation (in theory and
in computation) and shows the validity of the computation. The direct product of
the characters (i.e. the vibrational mode plus extra terms) is decomposed as
χtr ×χV ec = 2·N.1+1·N.2+4·N.3+4·N.4+5·N.5+5·N.6+6·N.7+6·N.8+8·N.9+7·N.10.
B.
One technical problem
The modes of vibrations are decomposed into irreducible representations, some
of which appear with multiplicities. Take one of the basis vectors, computed in the
previous section, and apply the symmetry operations on it: for example, one of those
which belong to the tenth irreducible representation. In the following computation
159
the two lists “EMR” and “AR” contains the group elements in R and the whole of
the symmetry operation, “vs” is the set of the generated vectors, “vsb”, the basis set
to it, and the command RankMat(...) computes the dimension of the vector space.
gap> EMR:=Elements(R) ;;
gap> AR:=List(Elements(R),j->LargeMatrix(j,hom));;
gap> vs:=List(AR,x->x*BAS[10][1]);
gab> vsb:=BaseMat(vs);
gap> RankMat(vs);
25
The chosen vector generates the vector space of dimension 25 by the symmetry
operation.(The basis vectors generated in this way are “vertex-wise” orthogonal to
those in the trivial representation: the basis vectors are the set of 60 small threedimensional vectors, and the small vectors of the former and the latter are orthogonal
at each vertex. The vectors, “vertex-wise” orthogonal to those in the trivial representation, make the subspace that includes several copies of the proper irreducible
space.) But the tenth irreducible representation is of dimension 5. The generated
vectors span the copies of this irreducible space, not confined to one of them. For
the practical purpose, we should remove the degrees of freedoms which correspond
to those of the translation and the rotation. And it is enough to the reset the dislocated origin or the disorientated axes at the proper positions when we make the
structures altered by vibrations. Thus the technical problem of this kind might not
be of importance. However, we can divide the multiplied irreducible spaces, so that
the each component is proper minimal one, separated with each other, kept invariant
under the symmetry operation. In this section, two kinds of such constructions are
demonstrated.
(p)
The first is to use the projectors composed from matrix representations Dkl (OT ),
160
not from characters, as was explained in the previous section. The formula is given
again:
(p)
Pkl =
lp X (p)∗
D (OT ) · OT .
|G| T ∈G kl
The basis vectors are transformed by the matrix representation, forming the subspace
which has the proper dimension of the irreducible representation. We can make the
(p)
plural irreducible spaces from the column vectors of the projector one of the Pkk ,
the diagonal part of the projectors. Unfortunately, even in this construction, the
generated irreducible spaces overlap with each other, not being separated. Thus we
should make the separation, projecting out overlapping subspaces. At first the matrix
representation is computed in the following way. (The GAP, in each run, returns the
different results concerning the matrix representation, since we can determine them
up to the isomorphism, but not furthermore; there are plural possibilities.)
gap> irrR:=Irr(R);;
gap> irrRM:=IrreducibleRepresentations(R);;
gap> EMR:=Elements(R);;
gap> AM:=List(EMR,j->LargeMatrix(j,hom));;
(p)
The projector Pij is defined as a function
PRJR:=function(i,j,p)
local C,REP,dim;
REP:=irrRM[p];
dim:=Trace(()^REP);
C:=List([1..Size(AM)],k->AsList(EMR[k]^REP)[i][j]*AM[k]);
return ComplexConjugate(Sum(C))/Size(AM)*dim;
end;
161
(10)
Take the P1,1 component of the projector. The basis vectors are generated from
column vectors of this. As the GAP software defines the vector set by row vectors,
The basis sets are computed from the transposed matrix. The basis set of the vector
(10)
(10)
space is listed in “vtb”. The projectors P1,1 , ..., P5,1 are stored now for the afterward
usage.(As for the matrix representations, see Appendix I.)
gap> vt:=TransposedMat(PRJR(1,1,10));;
gap> vtb:=BaseMat(vt);;
gap> PRJRS:=List([1..5],i->PRJR(i,1,10));;
The validity of the transformation of the partner functions is checked now.
gap> PRJR(2,2,10)*PRJR(2,1,10)*vtb[1]=PRJR(2,1,10)*vtb[1];
true
This function at first puts p-th irreducible matrix representation in “REP”. (The irreducible matrix representations are already stored in “irrRM”.) For a group element
“x”, the matrix representation D(p) (x) is computed by “x∧ REP”. The dimension of
the irreducible representation “dim” is computed from the trace of the matrix representation of the unit of the group ( represented by “( )”). Each group element “x”
(p)
(stored in the list “EMR”) is transformed to Dij (x), and multiplied to the equivalent matrix operation on the spatial vectors on the vertex of C60 (stored in the list
“AM”). The summation over them, through the complex conjugation, makes the
projector.
Each of the basis vector x generates the five-dimensional, irreducible vector spaces
(10)
(10)
under the symmetry operation, and the vector space is composed of P1,1 ·x, ..., P5,1 ·
x. In the following computations, the irreducible vector space for each of basis vector
is stored in the list “zz” (in the seven component zz[1],...,zz[7]). However, these
162
irreducible spaces are overlapped with each other. The computations of the inner
products show this.
gap> zz:=List(vtb,y->List(List([1..5],i->PRJRS[i]*y)));;
gap> List(zz,a->List(zz,b->RankMat(a*TransposedMat(b))));
[ [ 5, 0, 5, 0, 0, 0, 0 ], [ 0, 5, 0, 0, 0, 0, 0 ], [ 5, 0, 5, 5, 0, 0, 0 ],
[ 0, 0, 5, 5, 0, 5, 0 ], [ 0, 0, 0, 0, 5, 0, 0 ], [ 0, 0, 0, 5, 0, 5, 5 ],
[ 0, 0, 0, 0, 0, 5, 5 ] ]
We now project out the overlapping between irreducible subspaces. We start from
the one of the irreducible space, say zz[1], and construct the orthogonal space to it.
(10)
One of the basis vectors in the latter space, if the projector P1,1 is applied to it,
(10)
belongs to the irreducible space invariant by P1,1 . This vector is chosen as the
generator of another irreducible space. Similarly, we construct the orthogonal space
to the set of the irreducible spaces obtained until now, from this, we extract the
generator of another irreducible space and proceed so on.
gap> B:=ShallowCopy(zz[1]);;
gap> D:=[ShallowCopy(zz[1])];;
gap> for n in [2..Size(zz)]
do
OV:=BaseOrthogonalSpaceMat(B);;
xx:=BaseMat(List(OV,x->PRJRS[1]*x))[1];
vv:=List(List([1..5],i->PRJRS[i]*xx));
Append(B,vv);
Append(D,[vv]);
od;
We generated seven separated irreducible spaces; let us check this (by computing
the inner products of vectors):
163
gap> List(D,a->List(D,b->RankMat(a*TransposedMat(b))));;
[ [ 5, 0, 0, 0, 0, 0, 0 ], [ 0, 5, 0, 0, 0, 0, 0 ], [ 0, 0, 5, 0, 0, 0, 0 ],
[ 0, 0, 0, 5, 0, 0, 0 ], [ 0, 0, 0, 0, 5, 0, 0 ], [ 0, 0, 0, 0, 0, 5, 0 ],
[ 0, 0, 0, 0, 0, 0, 5 ] ]
The vector space generated by the symmetry operations from one of the separated
irreducible spaces should be orthogonal to others; let us check this (using the list
“AM” of symmetry operations prepared in the preceding computation):
gap> OO:=List(AM,o->o*D[1][1]);;
gap> List(D,d->RankMat(OO*TransposedMat(d)));
[ 5, 0, 0, 0, 0, 0, 0 ]
gap> OO:=List(AM,o->o*D[2][1]);;
gap> List(D,d->RankMat(OO*TransposedMat(d)));
[ 0, 5, 0, 0, 0, 0, 0 ]
. . . etc . . .
We can see that all is done well.
The second method is to divide the multiplied irreducible spaces W (as computed
in the previous section, by the projectors defined by the characters) into proper
minimal components, invariant under the symmetry operation. This method utilizes
the coset decomposition of the group G by the subgroup H, G/H. In general, one
vector v1 in W, under the operation of the elements gi in G, generates the set of
vectors, the total number of which is equal to the order of G. Now let us take
another vector w1 , also in W, generated by the operation of the elements hi in H,
P
such as w1 = i hi · v1 . Under the symmetry operation of G, it generates the set
of vectors, the total number of which is equal to the order of G/H. The number of
the generated vectors decreased in the latter case. We might expect the dimension
of the basis vectors of the generated vectors would also be diminished in the latter
164
case. And if w1 would generate the vector space of the proper dimension of the
irreducible representation, we adopt it, as the generator of the separated, proper,
irreducible space. (If not, with another chose of H, this step is repeated.) Then we
project out the vector space generated by w1 from W. We take another vector v2
from W, make an H-invariant vector w2 with the suitable H (if necessary, H should
be chosen again) to generate the another proper irreducible space. We proceed so
on until W is exhausted. This method will be applicable to the present problem in
order to separate each irreducible component, even if it involves the trial-and-error,
in the choice of the suitable subgroup H. At the start, the vector v1 is taken from the
multiplied irreducible spaces. The candidate vector wi is generated by the projection
of vi into the trivial irreducible representation of the subgroup H. (Indeed, it could
also be prepared by the projection of v into other irreducible representations of H,
for the purpose of diminishing the number of the generated vectors by the symmetry
operation.)
In order to execute this computation, the function “Cutout” (given in appendix
J) is prepared. The function cuts out the subspace (as minimal as possible) from the
given vector space, making use of the coset decomposition of R by the subgroups H.
In projecting out the proper irreducible component from the multiplied copies, we
employ slightly different step from the above description. Let W be the initial starting
place, that is, the multiplied components of the irreducible representations, from
which the separated proper irreducible components W1 , ..., Wi are computed. The
next candidate of the generator vi+1 is not taken directly from the complementary
space (W \ (W1 ∪ W2 ∪ · · · ∪ Wi )); instead, in the whole vector space S (of 180
dimension in this case), the orthogonal space to W1 ∪ W2 ∪ · · · ∪ Wi , biz, S \ (W1 ∪
W2 ∪ · · · ∪ Wi ) is constructed. The basis set in the latter space is projected into the
irreducible representation under the consideration, and the candidate vi+1 is taken
from this projected space (the renewed W). This treatment is to quicken the symbolic
165
computation.
In our problem, the latter method works well, too. We can separate out seven
distinct irreducible spaces. In the following computation the function “Cutout”
tries to divide out the vector space “BAS[10]” with respect to the tenth irreducible
representation. In this vector space seven copies of irreducible spaces exist. The
divided subspaces are given in list “H1”. Each subspace is of dimension 5 and
orthogonal to others. The symmetry operation (given in list “AM”, in the preceding
computation) to each vector in each subspace generates the 5 dimensional vector
space, and the generated vector space exits exactly in the same subspace of that
vector. (As for the other irreducible representations, it is successful, with the suitable
choice of the coset decomposition. For example, the subgroups of order 5 operate well
at the first, third, fourth, fifth, sixth, ninth and tenth irreducible representations;
that of order 8 does at the seventh; that of order 4 does at the eighth. )
gap> H1:=Cutout(BAS[10],PRJE[10],R,hom,5,5);;
gap> List(H1,a->List(H1,b->RankMat(a*TransposedMat(b))));
[ [ 5, 0, 0, 0, 0, 0, 0 ], [ 0, 5, 0, 0, 0, 0, 0 ], [ 0, 0, 5, 0, 0, 0, 0 ],
[ 0, 0, 0, 5, 0, 0, 0 ], [ 0, 0, 0, 0, 5, 0, 0 ], [ 0, 0, 0, 0, 0, 5, 0 ],
[ 0, 0, 0, 0, 0, 0, 5 ] ]
gap> OO:=List(AM,o->o*H1[1][1]);;
gap> List(H1,d->RankMat(OO*TransposedMat(d)));
[ 5, 0, 0, 0, 0, 0, 0 ]
gap> OO:=List(AM,o->o*H1[2][1]);;
gap> List(H1,d->RankMat(OO*TransposedMat(d)));
[ 0, 5, 0, 0, 0, 0, 0 ]
. . . etc . . .
A warning is given: in the above discussion, the symmetry operations are defined
166
in the bare forms, using the direct product Ri ⊗ Gi . To be precise one should use
Rif ⊗ Gi , even if the argument on the dimension counting on the vector space is not
affected. To obtain the proper vector system one should dutifully operate Mf ( the
matrix prepared in the previous section for the conjugation) to any vector v, such as
Mf · v.
C.
Remark to this section
In this section, the computation of the vibrational mode of C60 by means of the
computer algebra system GAP is demonstrated. The computation can be executed
by a standard desktop personal computer. Though the topic in the present article
is limited to the C60 molecule, the way of the computation shall be applicable to
other systems, such as crystal. The small programs in the article are composed as
brief as possible and as comprehensible as possible but without considerations of
the optimization, such as memory usage, or the avoidance of the re-computation of
same objects. So there is the possibility to improve the efficiency, although they
are computable even by the standard desktop personal computer. The computed
basis of the vibrational modes are obtained as the symbolic formulas, and they are
convertible to the numerical program functions simply by the rewriting of strings.
Some of the formulas might be so lengthy that the generated program functions might
be beyond the capacity of the compiler. However, the computations presented in this
article can be done by the floating-point computation, started from the numerical
representation of the rotation matrices, with a certain toleration to the numerical
error under the threshold.
I must add this: there are various studies on the electronic structure and the
deformation of C60 in the quantitative electronic structure computations (not of
the group theoretical view) such as in the references[37–39], or in the empirical
167
model (of the group theoretical view)[13]. The computations in this article, of this
and the previous section, only classifies geometrically possible deformations with
respect to irreducible representations and generates the basis set of deformations in
the symbolic formulas. The obtained results, however, could be applied to more
quantitative studies as those.
IX.
FINAL REMARKS
The theme of this article is how to simulate the property of the material by
means of electronic structure calculations with the aid of computer algebra. The
collaboration of the electronic structure calculation and the computer algebra is not
novel one; rather being regarded as the modernization of the union of the quantum
dynamics and the group theory, the success of which was brilliantly displayed in the
former half of last century. In this article, the treated topics in the group theory are
rather classical ones concerning the character theory. However, there can be various
applications of the myriad topics of mathematics. As for the combinatorics, the Polya
counting theory is a typical example; it is applicable to determine the number of
configurations, equivalent by some symmetry operations, for example, to determine
the total number of possible molecular formations indicated by a single chemical
formula[40]. This type of computations in combinatorics can also be executed by the
computer algebra. As for the application of combinatorics in quantum physics, a sort
of extension of Polya counting theorem can be useful in computing the spin weight in
molecules[41–43]. In a recent study, it is shown that the framework of the electronic
structure could be reconstructed by means of a concept in the algebraic geometry,
so-called Groebner basis theory, through the utilization of symbolic computation[44].
In my article, I have omitted several important application of the group theory
and the algebra. One of this is the treatment of magnetic or color groups. Also, I
168
omitted the treatment of Lie algebra, important in the spin system or the study of
angular momentum. As for the former, a package of color groups is implemented
in GAP[12]. As for the latter, the Lie algebra is computable by GAP, although it
is not of the digested style, suitable for the practical usage of physicists. (Some of
the implemented GAP packages are intended to study the model of spin systems,
proposed in the statistical- or mathematical physics.) I must admit that I have
not yet developed my skills to compute them through GAP. But I believe that the
readers if they are acquainted with GAP through my article, can construct their own
application in these topics.
In the present circumstances of the material physics, indeed, there is an awkward
inclination in which one solely relies on computer powers; only by trial and errors,
or, only by varying the input data, one tries to execute “material design”, without
any effective guiding principle. We must admit that the computational packages
have developed so complex and so vast that a common researcher, if not being an
expert, would be obliged to utilize them as black boxes, without knowing the detail
of the programs; fortunately, those packages are so powerful and effective that one
can obtain reliable results with ease. The blind reliance on these packages, however,
should be avoided; the mere enumeration of results of computations, with the slight
modification of input data, will be the waste of computational resources. Certainly,
by gathering and comparing computed results we could surmise vague tendencies,
but may not go anymore, if we do not know the physical or mathematical origin.
Instead, it is necessary for us to have another utensil, which will compensate for
such a shortcoming; the group theory, or its modernization, the computer algebra,
will be of great use for this purpose. With the aid of computer algebra, advanced
concepts in group theory are easily accessible and applicable, so that we can execute
systematic research of material design from the mathematical first principles, as the
group theoretical view illuminates for us the mathematical principles lurking behind
169
observable phenomena. I hope that the computer algebra will bring forth the blessed
reunion of physics and mathematics in the research of materials; the available tool
for us is not limited to the classics such as Courant-Hilbert[45], although they had
been powerful in the early stages of the development of quantum dynamics. The
mathematics and physics have progressed greatly since then; the forefront of these
developments seems to be so esoteric that non-experts feel dismayed when urged to
put it into practice. Nevertheless, one should not be timid; for the sake of ardent
researchers, the computer algebra will be a most suitable guide for their efforts, and
there they will find a most speedily passable bridge over the ever-deepening chasm
between physics and mathematics.
Appendix A: Symmetry operations in the diamond structure (real space)
The symmetry operations in the diamond structure (in Cartesian coordinates)
are given in the table. The operation on the atomic coordinate is given by x →
R · x + T 1 · a1 + T 2 · a2 + T 3 · a3 , where (a1 , a2 , a3 ) are primitive translations.There are
two types of fractional translations, according to the location of the origin of the unit
cell. The first, denoted as (T1,T2,T3) corresponds to the case of atomic positions
of ±(1/8 · a1 + 1/8 · a3 + 1/8 · a3 ):the second, (T1(*),T2(*),T3(*)) corresponds to
positions of (0 · a1 + 0 · a3 + 0 · a3 ) and (1/4 · a1 + 1/4 · a3 + 1/4 · a3 ). Although the
latter case is more conventional, the former has a merit such that the coefficients of
the plane-wave-expression of the wavefunctions are real-valued.
R11 R12 R13 R21 R22 R23 R31 R32 R33 T1 T2 T3 T1(*) T2(*) T3(*)
1
1
0
0
0
1
0
0
0
2
1
0
0
0
-1
0
0
3
-1
0
0
0
1
0
0
0
0
0
0
0
0
0
-1 0.5
0
0
0
0
0
0
-1
0 0.5
0
0
0
0
170
1
4
-1
0
0
0
-1
0
0
0
1
0
0 0.5
0
0
0
5
0
1
0
0
0
1
1
0
0
0
0
0
0
0
0
6
0
1
0
0
0
-1
-1
0
0 0.5
0
0
0
0
0
7
0
-1
0
0
0
1
-1
0
0
0 0.5
0
0
0
0
8
0
-1
0
0
0
-1
1
0
0
0
0 0.5
0
0
0
9
0
0
1
1
0
0
0
1
0
0
0
0
0
0
0
10
0
0
1
-1
0
0
0
-1
0 0.5
0
0
0
0
0
11
0
0
-1
1
0
0
0
-1
0
0 0.5
0
0
0
0
12
0
0
-1
-1
0
0
0
1
0
0
0 0.5
0
0
0
13
0
-1
0
-1
0
0
0
0
-1
0
0
0
0.25
0.25
0.25
14
0
-1
0
1
0
0
0
0
1 -0.5
0
0
0.25
0.25
0.25
15
0
1
0
-1
0
0
0
0
1
0 -0.5
0
0.25
0.25
0.25
16
0
1
0
1
0
0
0
0
-1
0
0 -0.5
0.25
0.25
0.25
17
-1
0
0
0
0
-1
0
-1
0
0
0
0
0.25
0.25
0.25
18
-1
0
0
0
0
1
0
1
0 -0.5
0
0
0.25
0.25
0.25
19
1
0
0
0
0
-1
0
1
0
0 -0.5
0
0.25
0.25
0.25
20
1
0
0
0
0
1
0
-1
0
0
0 -0.5
0.25
0.25
0.25
21
0
0
-1
0
-1
0
-1
0
0
0
0
0
0.25
0.25
0.25
22
0
0
-1
0
1
0
1
0
0 -0.5
0
0
0.25
0.25
0.25
23
0
0
1
0
-1
0
1
0
0
0 -0.5
0
0.25
0.25
0.25
24
0
0
1
0
1
0
-1
0
0
0
0 -0.5
0.25
0.25
0.25
25
-1
0
0
0
-1
0
0
0
-1
0
0
0
0.25
0.25
0.25
26
-1
0
0
0
1
0
0
0
1 -0.5
0
0
0.25
0.25
0.25
27
1
0
0
0
-1
0
0
0
1
0 -0.5
0
0.25
0.25
0.25
28
1
0
0
0
1
0
0
0
-1
0 -0.5
0.25
0.25
0.25
171
0
29
0
-1
0
0
0
-1
-1
0
0
0
0
0
0.25
0.25
0.25
30
0
-1
0
0
0
1
1
0
0 -0.5
0
0
0.25
0.25
0.25
31
0
1
0
0
0
-1
1
0
0
0 -0.5
0
0.25
0.25
0.25
32
0
1
0
0
0
1
-1
0
0
0
0 -0.5
0.25
0.25
0.25
33
0
0
-1
-1
0
0
0
-1
0
0
0
0
0.25
0.25
0.25
34
0
0
-1
1
0
0
0
1
0 -0.5
0
0
0.25
0.25
0.25
35
0
0
1
-1
0
0
0
1
0
0 -0.5
0
0.25
0.25
0.25
36
0
0
1
1
0
0
0
-1
0
0
0 -0.5
0.25
0.25
0.25
37
0
1
0
1
0
0
0
0
1
0
0
0
0
0
0
38
0
1
0
-1
0
0
0
0
-1 0.5
0
0
0
0
0
39
0
-1
0
1
0
0
0
0
-1
0 0.5
0
0
0
0
40
0
-1
0
-1
0
0
0
0
1
0
0 0.5
0
0
0
41
1
0
0
0
0
1
0
1
0
0
0
0
0
0
0
42
1
0
0
0
0
-1
0
-1
0 0.5
0
0
0
0
0
43
-1
0
0
0
0
1
0
-1
0
0 0.5
0
0
0
0
44
-1
0
0
0
0
-1
0
1
0
0
0 0.5
0
0
0
45
0
0
1
0
1
0
1
0
0
0
0
0
0
0
0
46
0
0
1
0
-1
0
-1
0
0 0.5
0
0
0
0
0
47
0
0
-1
0
1
0
-1
0
0
0 0.5
0
0
0
0
48
0
0
-1
0
-1
0
1
0
0
0
0 0.5
0
0
0
TABLE 43: Symmetry operations in the point group in the
diamond unit cell. If fractional translations are set to be zero,
these operations are applicable to the cubic lattice.
172
Appendix B: Symmetry operations in the hexagonal lattice (real space)
The symmetry operations in graphite (in Cartesian coordinates)are given in the
table:
R11
1
R12 R13
1
0
√
0
1
0
0
0
1
0
0
3/2 0.5
0
0
0
1
0
0 -0.5
3/2 -0.5
0
0
0
1
0
0
-1
0
0
0
1
0
0 -0.5
-0.5
0
0
0
1
0
0
0.5
0
0
0
1
0
0 -0.5
0.5
0
0
0
-1
0
0 -0.5
-0.5
0
0
0
-1
0
0
-1
0
0
0
-1
0
0 -0.5
3/2 -0.5
0
0
0
-1
0
0
3/2 0.5
0
0
0
-1
0
0 -0.5
1
0
0
0
-1
0
0
0 -1
√
0 - 3/2 -0.5
√
0 - 3/2 0.5
0
0
0
-1
0
0 -0.5
0
0
0
-1
0
0
0
0
0
-1
0
0 -0.5
0
1
0
0
0
-1
0
0
3/2 0.5
0
0
0
-1
0
0 -0.5
3/2 -0.5
0
0
0
-1
0
0
0
3/2 -0.5
0
0
0
1
0
0
0
3/2 0.5
0
0
0
1
0
0 -0.5
0
0
0
1
0
0
2 0.5 - 3/2
√
3 -0.5 - 3/2
0
4
0
0
3/2
0
3/2
0
7 -0.5 - 3/2
√
8 0.5 - 3/2
0
9
0
-1
5 -0.5
6 0.5
√
√
√
1
10 0.5
11 -0.5
12
-1
13
-1
14 -0.5
15 0.5
16
0
√
√
√
√
1
3/2
0
0
0
3/2
0
0
0
0
0
3/2
3/2
0
√
17 0.5 - 3/2
√
18 -0.5 - 3/2
√
19 0.5
3/2
√
20 -0.5
3/2
0
21
0
-1
0
R21 R22 R23 R31 R32 R33 T1 T2 T3
0
0
0
0
√
√
0
√
- 3/2
√
- 3/2
√
- 3/2
√
- 3/2
0
√
√
0
0
√
√
√
√
0
1
173
0
0
0
0
0
0
0
0
0
√
22 -0.5 - 3/2
√
23 0.5 - 3/2
√
0 - 3/2 0.5
√
0 - 3/2 -0.5
24
0
1
0
0
-1
0
0
0
1
0
0 -0.5
0
0
0
1
0
0
0
0
0
1
0
0 -0.5
0
TABLE 44: The symmetry operations in graphite in real
space.
Appendix C: Symmetry operations in the diamond structure (wave-number
space)
The symmetry operations in the diamond structure, as the rotation matrices in
the reciprocal spaces are given in the table:
R’11 R’12 R’13 R’21 R’22 R’23 R’31 R’32 R’33 T1 T2 T3
1
1
0
0
0
1
0
0
0
1
0
0
0
2
-1
0
0
-1
0
1
-1
1
0 0.5
0
0
3
0
-1
1
0
-1
0
1
-1
0
0 0.5
0
4
0
1
-1
1
0
-1
0
0
-1
0
0 0.5
5
0
1
0
0
0
1
1
0
0
0
0
0
6
0
-1
0
1
-1
0
0
-1
1 0.5
0
0
7
1
0
-1
0
0
-1
0
1
-1
0 0.5
0
8
-1
0
1
-1
1
0
-1
0
0
0
0 0.5
9
0
0
1
1
0
0
0
1
0
0
0
0
10
0
0
-1
0
1
-1
1
0
-1 0.5
0
0
11
-1
1
0
-1
0
0
-1
0
1
0 0.5
0
12
1
-1
0
0
-1
1
0
-1
0
0
0 0.5
13
0
-1
0
-1
0
0
0
0
-1
0
0
174
0
14
0
1
0
0
1
-1
-1
1
15
1
0
-1
1
0
0
1
-1
16
-1
0
1
0
-1
1
0
17
-1
0
0
0
0
-1
18
1
0
0
1
-1
19
0
-1
1
0
20
0
1
-1
21
0
0
22
0
23
0
0
0
0 -0.5
0
0
1
0
0 -0.5
0
-1
0
0
0
0
0
1
0
-1 -0.5
0
0
0
1
-1
0
1
0 -0.5
0
-1
1
0
0
1
0
0
0 -0.5
-1
0
-1
0
-1
0
0
0
0
0
0
1
-1
0
1
0
-1
1 -0.5
0
0
-1
1
0
0
1
0
0
1
-1
0 -0.5
0
24
1
-1
0
1
0
-1
1
0
0
0
0 -0.5
25
-1
0
0
0
-1
0
0
0
-1
0
0
0
26
1
0
0
1
0
-1
1
-1
0 -0.5
0
0
27
0
1
-1
0
1
0
-1
1
0
0 -0.5
0
28
0
-1
1
-1
0
1
0
0
1
0
0 -0.5
29
0
-1
0
0
0
-1
-1
0
0
0
0
0
30
0
1
0
-1
1
0
0
1
-1 -0.5
0
0
31
-1
0
1
0
0
1
0
-1
1
0 -0.5
0
32
1
0
-1
1
-1
0
1
0
0
0
0 -0.5
33
0
0
-1
-1
0
0
0
-1
0
0
0
0
34
0
0
1
0
-1
1
-1
0
1 -0.5
0
0
35
1
-1
0
1
0
0
1
0
-1
0 -0.5
0
36
-1
1
0
0
1
-1
0
1
0
0
0 -0.5
37
0
1
0
1
0
0
0
0
1
0
0
0
38
0
-1
0
0
-1
1
1
-1
0 0.5
0
0
175
0 -0.5
39
-1
0
1
-1
0
0
-1
1
0
0 0.5
0
40
1
0
-1
0
1
-1
0
0
-1
0
0 0.5
41
1
0
0
0
0
1
0
1
0
0
0
0
42
-1
0
0
-1
1
0
-1
0
1 0.5
0
0
43
0
1
-1
0
0
-1
1
0
-1
0 0.5
0
44
0
-1
1
1
-1
0
0
-1
0
0
0 0.5
45
0
0
1
0
1
0
1
0
0
0
0
0
46
0
0
-1
1
0
-1
0
1
-1 0.5
0
0
47
1
-1
0
0
-1
0
0
-1
1
0 0.5
0
48
-1
1
0
-1
0
1
-1
0
0
0
0 0.5
TABLE 45: Symmetry operations in the diamond crystal in
the reciprocal space.
Appendix D: Symmetry operations in the hexagonal lattice (wave-number
space)
The crystal axes are chosen to be a1 = (1, 0, 0), a2 = (−1/2,
p
(3)/2, 0), a3 =
(0, 0, c). Then The symmetry operations as the operations in the wave-number space
are given as:
R’11 R’12 R’13 R’21 R’22 R’23 R’31 R’32 R’33 T1 T2 T3
1
1
0
0
0
1
0
0
0
1
0
0
2
0
-1
0
1
1
0
0
0
1
0
0 -0.5
3
-1
-1
0
1
0
0
0
0
1
0
0
4
-1
0
0
0
-1
0
0
0
1
0
0 -0.5
5
0
1
0
-1
-1
0
0
0
1
0
0
176
0
0
0
6
1
1
0
-1
0
0
0
0
1
0
0 -0.5
7
-1
-1
0
0
1
0
0
0
-1
0
0 -0.5
8
0
-1
0
-1
0
0
0
0
-1
0
0
9
1
0
0
-1
-1
0
0
0
-1
0
0 -0.5
10
1
1
0
0
-1
0
0
0
-1
0
0
11
0
1
0
1
0
0
0
0
-1
0
0 -0.5
12
-1
0
0
1
1
0
0
0
-1
0
0
13
-1
0
0
0
-1
0
0
0
-1
0
0 -0.5
14
0
1
0
-1
-1
0
0
0
-1
0
0
15
1
1
0
-1
0
0
0
0
-1
0
0 -0.5
16
1
0
0
0
1
0
0
0
-1
0
0
17
0
-1
0
1
1
0
0
0
-1
0
0 -0.5
18
-1
-1
0
1
0
0
0
0
-1
0
0
0
19
1
1
0
0
-1
0
0
0
1
0
0
0
20
0
1
0
1
0
0
0
0
1
0
0 -0.5
21
-1
0
0
1
1
0
0
0
1
0
0
22
-1
-1
0
0
1
0
0
0
1
0
0 -0.5
23
0
-1
0
-1
0
0
0
0
1
0
0
24
1
0
0
-1
-1
0
0
0
1
0
0 -0.5
TABLE 46: The symmetry operations in graphite in reciprocal space.
177
0
0
0
0
0
0
0
Appendix E: The computation of Wyckoff positions by GAP
Computations for space groups can be executed by means of “Cryst” package in
GAP. Though the manual of this package is provided, the description is too curt and
the computed result is too unfriendly for beginners to utilize. So in this section, the
usage of this package is illustrated.
The crystallographic group is defined at first (in this example, with minimal generators):
gap> M1:=[[0,0,1,0],[1,0,0,0],[0,-1,0,0],[1/4,1/4,1/4,1]];
gap> M2:=[[0,0,-1,0],[0,-1,0,0],[1,0,0,0],[0,0,0,1]];
gap> S:=AffineCrystGroup([M1,M2]);
The point group is obtained by:
gap> P:=PointGroup(S);
The Wyckoff positions are computed now:
gap> W:=WyckoffPositions(S);
[ < Wyckoff position, point group 4, translation := [ 1/8, 3/8, 7/8 ],
basis := [
] >
, < Wyckoff position, point group 4, translation := [ 1/8, 3/8, 3/8 ],
basis := [
] >
, < Wyckoff position, point group 5, translation := [ 1/4, 1/4, 3/4 ],
basis := [
] >
, < Wyckoff position, point group 5, translation := [ 1/4, 1/4, 1/4 ],
basis := [
] >
, < Wyckoff position, point group 3, translation := [ 0, 0, 0 ],
178
basis := [ [ 1, -1, -1 ] ] >
, < Wyckoff position, point group 6, translation := [ 1/4, 1/4, 1/2 ],
basis := [ [ 0, 0, 1 ] ] >
, < Wyckoff position, point group 7, translation := [ 0, 1/4, 7/8 ],
basis := [ [ 1/2, 1/2, 0 ] ] >
, < Wyckoff position, point group 2, translation := [ 0, 0, 0 ],
basis := [ [ 1/2, -1/2, 0 ], [ 0, 0, 1 ] ] >
, < Wyckoff position, point group 1, translation := [ 0, 0, 0 ],
basis := [ [ 1/2, 0, 1/2 ], [ 0, 1/2, 1/2 ], [ 0, 0, 1 ] ] >
]
The returned result is the list of the generating set of the Wyckoff positions. The
expression such as
, < Wyckoff position, point group 3, translation := [ 0, 0, 0 ],
basis := [ [ 1, -1, -1 ] ] >
indicates the information of a generator for the Wyckoff positions, which is located
at the segment [0,0,0]+t[1,-1,-1], where t is arbitrary real number inasmuch that the
segment is confined in the unit cell. Now we get nine generators. The computed
result, therefore, should be read as,
[ 1/8, 3/8, 7/8 ]
[ 1/8, 3/8, 3/8 ]
[ 1/4, 1/4, 3/4 ]
[ 1/4, 1/4, 1/4 ]
[ 0, 0, 0 ]+t[ 1, -1, -1 ]
[ 1/4, 1/4, 1/2 ]+t[ 0, 0, 1 ]
[ 0, 1/4, 7/8 ]+t[ 1/2, 1/2, 0 ]
179
[ 0, 0, 0 ]+t1[ 1/2, -1/2, 0 ]+t2[ 0, 0, 1 ]
[ 0, 0, 0 ]+t1[ 1/2, 0, 1/2 ]+t2[ 0, 1/2, 1/2 ]+t3[ 0, 0, 1 ]
The first four are single points, but the remaining five are one-dimensional segments, two-dimensional planes, or, a three-dimensional shape. (The last of them is
the arbitrary position in the unit cell.) Let us pay attention to the fourth of them. It
is the position of one carbon atom in the minimal diamond unit cell; which is moved
to [0,0,0] (the position of another Carbon atom) by means of symmetry operations.
In order to obtain whole of the Wyckoff positions, the equivalent coordinate points
(orbits by the symmetry operations) can be evaluated in the following way.
The total numbers of the equivalent coordinate points are given by
gap> List(WyckoffPositions(S),x->Size(WyckoffOrbit(x)));
[ 4, 4, 2, 2, 8, 12, 24, 24, 48 ]
For each generating points, the equivalents coordinates can be computed separately:
gap> WyckoffOrbit(WyckoffPositions(S)[1]);
[ < Wyckoff position, point group 4, translation := [ 1/8, 3/8, 7/8 ],
basis := [
] >
, < Wyckoff position, point group 4, translation := [ 3/8, 1/8, 7/8 ],
basis := [
] >
, < Wyckoff position, point group 4, translation := [ 3/8, 3/8, 5/8 ],
basis := [
] >
, < Wyckoff position, point group 4, translation := [ 1/8, 1/8, 5/8 ],
basis := [
] >
]
gap> WyckoffOrbit(WyckoffPositions(S)[2]);
180
[ < Wyckoff position, point group 4, translation := [ 1/8, 3/8, 3/8 ],
basis := [
] >
, < Wyckoff position, point group 4, translation := [ 3/8, 1/8, 3/8 ],
basis := [
] >
, < Wyckoff position, point group 4, translation := [ 3/8, 3/8, 1/8 ],
basis := [
] >
, < Wyckoff position, point group 4, translation := [ 1/8, 1/8, 1/8 ],
basis := [
] >
]
gap> ...................................................................
gap> WyckoffOrbit(WyckoffPositions(S)[8]);
[ < Wyckoff position, point group 2, translation := [ 0, 0, 0 ],
basis := [ [ 1/2, -1/2, 0 ], [ 0, 0, 1 ] ] >
, < Wyckoff position, point group 2, translation := [ 1/4, 1/4, 1/4 ],
basis := [ [ -1/2, 0, 1/2 ], [ 0, -1, 0 ] ] >
, < Wyckoff position, point group 2, translation := [ 0, 0, 0 ],
..................................................................
..................................................................
]
Appendix F: Subgroup lattice
In this article, the sequence of crystalline symmetry reduction is displayed by
means of the mathematical concept of the lattice of subgroups, in which the inclusion
relations among them are illustrated as a graph. The GAP computation can be done
as follows:
gap> M1:=[[0,1,0],[0,0,1],[1,0,0]];
181
[ [ 0, 1, 0 ], [ 0, 0, 1 ], [ 1, 0, 0 ] ]
gap> M2:=[[0,1,0],[-1,0,0],[0,0,1]];
[ [ 0, 1, 0 ], [ -1, 0, 0 ], [ 0, 0, 1 ] ]
gap> G:=Group(M1,M2);
Group([ [ [ 0, 1, 0 ], [ 0, 0, 1 ], [ 1, 0, 0 ] ],
[ [ 0, 1, 0 ], [ -1, 0, 0 ], [ 0, 0, 1 ] ] ])
gap> Size(G);
24
gap> l:=LatticeSubgroups(G);
<subgroup lattice of Group(
[ [ [ 0, 1, 0 ], [ 0, 0, 1 ], [ 1, 0, 0 ] ],
[ [ 0, 1, 0 ], [ -1, 0, 0 ], [ 0, 0, 1 ] ] ]), 11 classes,
30 subgroups>
gap> ConjugacyClassesSubgroups(l);
[ Group([],[ [ 1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, 1 ] ])^G,
Group([ [ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ] ])^G,
Group([ [ [ -1, 0, 0 ], [ 0, 0, -1 ], [ 0, -1, 0 ] ] ])^G,
Group([ [ [ 0, 1, 0 ], [ 0, 0, 1 ], [ 1, 0, 0 ] ] ])^G,
Group([ [ [ -1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, -1 ] ],
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ] ])^G,
Group([ [ [ -1, 0, 0 ], [ 0, 0, -1 ], [ 0, -1, 0 ] ],
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ] ])^G,
Group([ [ [ 1, 0, 0 ], [ 0, 0, 1 ], [ 0, -1, 0 ] ],
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ] ])^G,
Group([ [ [ -1, 0, 0 ], [ 0, 0, -1 ], [ 0, -1, 0 ] ],
[ [ 0, 1, 0 ], [ 0, 0, 1 ], [ 1, 0, 0 ] ] ])^G,
Group([ [ [ -1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, -1 ] ],
182
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ],
[ [ -1, 0, 0 ], [ 0, 0, -1 ], [ 0, -1, 0 ] ] ])^G,
Group([ [ [ -1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, -1 ] ],
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ],
[ [ 0, 1, 0 ], [ 0, 0, 1 ], [ 1, 0, 0 ] ] ])^G,
Group([ [ [ -1, 0, 0 ], [ 0, 1, 0 ], [ 0, 0, -1 ] ],
[ [ 1, 0, 0 ], [ 0, -1, 0 ], [ 0, 0, -1 ] ],
[ [ 0, 1, 0 ], [ 0, 0, 1 ], [ 1, 0, 0 ] ],
[ [ -1, 0, 0 ], [ 0, 0, -1 ], [ 0, -1, 0 ] ] ])^G ]
gap> DotFileLatticeSubgroups(l,"paradigm.dot");
The last command “DotfileLatticeSubgroups” is used to generate an output file
for the visualization.
Appendix G: A short program to compute the semidirect product of groups
In the main article, the characters of C3v = Group((1, 2, 3), (1, 2)) ≡ S3 are
computed by the semidirect product between C3 = Group((1, 2, 3)) and I =
Group((1, 2)). The computations are tedious and complicated; a naive implementation by means of GAP programming language could be given as follows:
# The definition of the groups G and A is done.
# The irreducible representations and the lists
# of group elements are prepared.
G:=Group((1,2));
A:=Group((1,2,3));
repG:=Irr(G);
183
repA:=Irr(A);
Gmem:=Elements(G);
Amem:=Elements(A);
#
# The definition of the automorphism on A induced by G
# is given in list "au".
# The elements of G and A are indexed
#
#
#
G[1],G[2],...
and
A[1],A[2],....
# The elements in A are mapped as
#
#
A[i]^G[j]=A[au[i,j]];
#
au:=[[1,1],[2,3],[3,2]];
#
This list now means the following automorphism:
#
x(in A)^E(in G)
: E -> E, a -> a,
#
x(in A)^g(in_G)
: E -> E, a -> a^-1, a^-1 -> a
#
Here a=(1,2,3),
#
and the elements in A are listed [e, a, a^-1].
#
The following is the list of elements
#
in the semidirect product [A,G],
#
given by the Cartesian products of sequential numbers.
a^-1 -> a^-1
#
Elist:=[[1,1],[1,2],[2,2],[3,2],[2,1],[3,1]];
184
GAction:=function(a,g)
#
# A function to compute the automorphism
(a^g)
# on an element "a" in A
# induced by an element "g" in G.
# This function returns the group element,
# according to the list "au".
#
local apos,gpos,ha;
apos:=Position(Amem,a);
gpos:=Position(Gmem,g);
ha:=au[apos][gpos];
return Amem[ha];
end;
InvarianceCHK:=function(iA,g,au)
#
# This function checks whether the character x in A
# (that by the iA-th representation) is
# fixed by the automorphism induced by the element g in G.
#
# iA: index to the irreducible representation of A
# au: the automorphism on A by G, defined above.
#
local ol,tl;
#
185
# The list of characters, before the automorphism.
ol:=List(Amem,x->x^repA[iA]);
# The list of characters, after the automorphism.
tl:=List(Amem,x->Amem[au[Position(Amem,x)][Position(Gmem,g)]]);
tl:=List(tl,x->x^repA[iA]);
return ol=tl;
end;
GetGx:=function(iA,au)
#
# Gx is the set of elements in G,
# such that the character x in A
# (that by the iA-th representation)
# is fixed by the automorphism induced
# by the elements in Gx.
#
# iA: the index to the irreducible representation of A
# au: the automorphism on A by G, defined above.
#
#
local Gx;
Gx:=Filtered(G,g->InvarianceCHK(iA,g,au));
return Group(Gx);
end;
ProSumma:=function(a,g,Gx,iA,iG)
# This is a subordinate function to compute the character
186
# of the elements (a,g),
# when the representations in group A and G,
# and Gx (defined above) are specified.
#
# iA: index to the irreducible representations of A
# iG: index to the irreducible representations of G
#
local k,sum,l1,l2,l3,l4,repGx;
l1:=List(Gmem,x->x*g*x^-1);
l2:=List(Gmem,x->GAction(a,x));
l3:=List(l2,x->x^repA[iA]);
repGx:=Irr(Gx);
sum:=0;
for k in [1..Size(l1)]
do
if l1[k] in Gx then
l4:=l1[k]^repGx[iG];
sum:=sum+l3[k]*l4;
fi;
od;
sum:=sum/Size(Gx);
return sum;
end;
ProChr:=function(au,iA)
#
# au
: the automorphism, defined in the list as above.
187
# iA
: the index to the irreducible representations in A
#
# This function computes the characters in the semidirect product
# (for each element in Elist, defined above).
#
local Gx,repGx,index,retTBL;
Gx:=GetGx(iA,au);
repGx:=Irr(Gx);
retTBL:=[];
for index in [1..Size(repGx)]
do;
Append(retTBL,
[List(Elist,x->ProSumma(Amem[x[1]],Gmem[x[2]],Gx,iA,index))]
) ;
od;
return retTBL;
end;
gap> for iA in [1..Size(repA)]
do;
Display(ProChr(au,iA));
od;
[ [
1,
1,
1,
1,
1,
1 ],
[
1,
-1,
-1,
-1,
1,
1 ] ]
[ [
2,
0,
0,
0,
-1,
-1 ] ]
[ [
2,
0,
0,
0,
-1,
-1 ] ]
188
Appendix H: The analytical representation of the energy spectra of C60 .
In the model of C60 in section VII B, if we chose the transfer integral as Hsingle =1 (on the single bonds in the pentagons) and Hdouble =-y (on the double bonds in
the hexagons), the secular equation is decomposed into the polynomial with some
multiplicity in each representation, as in table 47. The irreducible representations
I.1,...,I.10 are the same as given in the main article, at table 26. The energy spectra
of in the main article can be constructed from this table, setting E = 2x and y=-1/2.
I.1
I.2
I.3
I.4
I.5
I.6
I.7
I.8
I.9
I.10
Polynomial
[ x+y+2 ]
[ ]
[ x-y+(E(5)^2+E(5)^3) ]
[ x-y+(E(5)+E(5)^4) ]
[ x^2-y^2+(-E(5)-2*E(5)^2-2*E(5)^3-E(5)^4)*x
+(-E(5)-E(5)^4)*y+(2*E(5)+2*E(5)^4) ]
[ x^2-y^2+(-2*E(5)-E(5)^2-E(5)^3-2*E(5)^4)*x
+(-E(5)^2-E(5)^3)*y+(2*E(5)^2+2*E(5)^3) ]
[ x^2-y^2-x-1 ]
[ x^2-y^2-x-2*y-1 ]
[ x^3+x^2*y-x*y^2-y^3+x^2+2*x*y+y^2-3*x-y-2 ]
[ x^2-y^2-x+y-1 ]
multiplicity
1
3
3
3
3
4
4
5
5
TABLE 47. The decomposition of the secular equation.
Appendix I: Comment on the irreducible representations in the symmetry
group of C60
In section VIII we have computed the irreducible matrix representations of the
symmetry group of C60 . Since the tenth irreducible matrix representation is a real
representation, the generated bases are real and we can work in the real number.
Meanwhile, the four-dimensional irreducible matrix representations (computed by
189
GAP) are complex and the generated basis are complex. However, the complex
irreducible representation, in this case, can be remade into real one. The complexvalued matrix representation of a generator (Mi ) can be replaced by a real matrix:
Re(Mi ) −Im(Mi )
Im(Mi ) Re(Mi )
.
This representation is reducible, including two copies of the irreducible representation, but as the matrix group, it is isomorphic to the original group. We extract
the basis set from the columns of the projector of this matrix group and divide it
into two independent irreducible vector spaces by means of the procedure explained
thereafter in the article. Each of the separated vector space provides us the realvalued basis set and the real-valued matrix representation. One set of such real
matrix representations (for the generators r1 and r2) is given as follows:
r1:=[[1,
3/2*E(5)+1/2*E(5)^2+1/2*E(5)^3+3/2*E(5)^4,
-3*E(5)-7*E(5)^2-7*E(5)^3-3*E(5)^4,
-1081/2131*E(5)-1797/2131*E(5)^2-1797/2131*E(5)^3-1081/2131*E(5)^4],
[0,
-1/2*E(5)-E(5)^2-E(5)^3-1/2*E(5)^4,
-5*E(5)-14*E(5)^2-14*E(5)^3-5*E(5)^4,
-1681/2131*E(5)-3510/2131*E(5)^2-3510/2131*E(5)^3-1681/2131*E(5)^4],
[0,
-2*E(5)-3/2*E(5)^2-3/2*E(5)^3-2*E(5)^4,
-2*E(5)-7*E(5)^2-7*E(5)^3-2*E(5)^4,
-600/2131*E(5)-1713/2131*E(5)^2-1713/2131*E(5)^3-600/2131*E(5)^4],
[0,
190
113/44*E(5)+92/11*E(5)^2+92/11*E(5)^3+113/44*E(5)^4,
623/22*E(5)+807/11*E(5)^2+807/11*E(5)^3+623/22*E(5)^4,
5/2*E(5)+8*E(5)^2+8*E(5)^3+5/2*E(5)^4]];
r2:=[[-4/11*E(5)-3/11*E(5)^2-3/11*E(5)^3-4/11*E(5)^4,
7/22*E(5)+19/22*E(5)^2+19/22*E(5)^3+7/22*E(5)^4,
36/11*E(5)+104/11*E(5)^2+104/11*E(5)^3+36/11*E(5)^4,
383/2131*E(5)+2127/2131*E(5)^2+2127/2131*E(5)^3+383/2131*E(5)^4],
[12/11*E(5)-2/11*E(5)^2-2/11*E(5)^3+12/11*E(5)^4,
-16/11*E(5)-1/11*E(5)^2-1/11*E(5)^3-16/11*E(5)^4,
57/11*E(5)+128/11*E(5)^2+128/11*E(5)^3+57/11*E(5)^4,
-1516/2131*E(5)+1596/2131*E(5)^2+1596/2131*E(5)^3-1516/2131*E(5)^4],
[32/11*E(5)+13/11*E(5)^2+13/11*E(5)^3+32/11*E(5)^4,
-39/11*E(5)-31/22*E(5)^2-31/22*E(5)^3-39/11*E(5)^4,
53/11*E(5)+92/11*E(5)^2+92/11*E(5)^3+53/11*E(5)^4,
-1784/2131*E(5)+703/2131*E(5)^2+703/2131*E(5)^3-1784/2131*E(5)^4],
[0,0,
-623/22*E(5)-807/11*E(5)^2-807/11*E(5)^3-623/22*E(5)^4,
-2*E(5)-7*E(5)^2-7*E(5)^3-2*E(5)^4]];
We can use them in common in the two four-dimensional representations, with r3 :=
±I (the four dimensional unit matrix).
Appendix J: The GAP script in the computation of deformation in C60
As is pointed in the main article, some of the modes of deformation in C60 computed by the projectors includes several copies of proper irreducible representations.
To decompose them into separated irreducible spaces, the following function is pre191
pared. The function tries to find the irreducible subspace of dimension “REPSIZ”
in the operation of the group “R”. The function cuts out the subspace (as minimal
as possible) from the given vector space, making use of the coset decomposition of
R by the subgroups H of order “NMULTI”. In projecting out the proper irreducible
component from the multiplied copies, we employ these steps. Let W be the initial
starting place, that is, the multiplied components of the irreducible representations,
from which the separated proper irreducible components W1 , ..., Wi are computed.
The next candidate of the generator vi+1 is not taken directly from the complementary space (W \ (W1 ∪ W2 ∪ · · · ∪ Wi )); instead, in the whole vector space S
(of 180 dimension in this case), the orthogonal space to W1 ∪ W2 ∪ · · · ∪ Wi , biz,
S \ (W1 ∪ W2 ∪ · · · ∪ Wi ) is constructed. The basis set in the latter space is projected
into the irreducible representation under the consideration, and the candidate vi+1
is taken from this projected space (the renewed W). This treatment is to quicken the
symbolic computation.
gopr:=function(O,v,i,hom)
#
# When i=1, This function applies the symmetry operations of Group O
# to a vector v. And the generated vectors are summed up.
# (In fact, this function is the operation of the projector to the vector v.)
#
local irr;
irr:=Irr(O);
return Sum(List(Elements(O),j->j^irr[i]*LargeMatrix(j,hom)*v));
end;
Cutout:=function(BAS,PROJ,R,hom,NMULTI,REPSIZ)
192
#
# BAS
: THE VECTOR SPACE INCLUDING THE COPIES OF IRREDUCIBLE SPACES
# PROJ
: THE PROJECTORS
# R
: THE GROUP IN THE PROBLEM.
# hom
: THE HOMOMORPHISM FROM <R> TO THE MATRIX GROUP.
# NMULTI : THE SIZE OF THE SUBGROUP TO FORM THE COSET.
# REPSIZ : THE SIZE OF THE IRREDUCIBLE REPRESENTATION.
#
local C,D,BASN,EA,O,l,P,p,B,OV,SI,AS,OREP,ww,OT,ith,ifind;
C:=[];
D:=[];
BASN:=BAS;
EA:=List(Elements(R),j->LargeMatrix(j,hom));;
#
# The subgroups with the order of "NMULTI" are extracted.
#
AS:=AllSubgroups(R);
while (BASN<>[])
do
Print("\n",Size(BASN),"\n");
O:=Filtered(AS,g->Order(g)=NMULTI);;
Print(O);
#
# The subgroups which generate the vector space of
# the correct dimension of the irreducible representation
# are chosen.
#
193
if (O=[]) then
O:=[Group(One(R))];
fi;
l:=List([1..Size(O)],i->0);
ifind:=0;
ith:=1;
while (ifind=0 and ith<=Size(O))
do
ww:=gopr(O[ith],BASN[1],1,hom);
l[ith]:=RankMat(List(EA,g->g*ww));
Print(l);
if (l[ith]=REPSIZ) then
ifind:=1;
OT:=O[ith];
fi;
ith:=ith+1;
od;
if (ifind=0) then
Print("CHANGE THE SUPPOSITION!\n");
return 0;
fi;
#
#
By means of the chosen subgroup, the irreducible
#
vector space "B" is generated.
#
It is stored in the vector space "D".
#
B:=BaseMat(List(EA,g->g*gopr(OT,BASN[1],1,hom)));;
194
Append(C,[B]);
Append(D,B);
#
# The orthogonal space "OV" to the vector space "D" is prepared.
#
OV:=BaseOrthogonalSpaceMat(D);
#
# The vector space "BASN" to be processed is renewed now,
# by the projection "PROJ" on the vectors in "OV".
#
BASN:=BaseMat(List(OV,o->PROJ*o));
#
# Instead, it is possible that "BASN" is renewed
# as the intersection between "BASN" and "OV".
#
#
SI:=SumIntersectionMat(BASN,OV);
#
BASN:=SI[2];
#
od;
#
# Return C, the separated subspaces.
#
return C;
end;
195
[1] H. Weyl, The Theory of Groups and Quantum Mechanics,(Dover,1950).
[2] T. Inui, Y. Tanabe and Y. Onodera, Group Theory and Its applications in Physics
(Springer-Verlag,1990).
[3] M. Tinkham, Group Theory and Quantum Dynamics (Dover Books on Chemistry,2003).
[4] M. S. Dresselhaus, G. Dresselhaus, and A. Jorio, Group Theory: Application to the
Physics of Condensed Matter, (Springer-Verlag, 2008).
[5] W. Burnside, Theory of groups of finite order,(Cambridge University Press,1897).[Now
available at Project Gutenberg, <http://www.gutenberg.org/ebooks/40395>.]
[6] J. W. Thomas, Abstract Algebra: Theory and applications, <http://abstract.ups.
edu/>
[7] T. Y. Lam, Representations of finite groups: a hundred years, Notices of the AMS
(American Mathematical Society) 45(3,4):361–372 (Part I),465–474(Part II) (1998).
[8] Jean-Pierre Serre, Linear Representations of Finite Groups, (Springer-Verlag,1977).
[9] W. Fulton, J. Harris, Representation theory. A first course, Graduate Texts in Mathematics, Readings in Mathematics 129, (Springer-Verlag,1991).
[10] D. Holt, B. Eick and E. O’Brien,Handbook of Computational Group Theory, (Chapman
& Hall/CRC Press,2005)
[11] A. Hulpke, Note On Computational Group Theory, <http://www.math.colostate.
edu/~hulpke/CGT/cgtnotes.pdf>.
[12] “GAP - Groups, Algorithms, Programming - a System for Computational Discrete
Algebra”, <http://www.gap-system.org/>.
[13] J. Mooij, The vibrational spectrum of Buckminsterfullerene - An application of symmetry reduction and computer algebra, Master’s Thesis in Mathematics, Katholieke
196
Universiteit Nijmegen (2003).
[14] Michael El-Batanouny and F. Wooten, Symmetry and condensed matter physics: a
computational approach, Cambridge University Press, 2008.
[15] D. Stauffer, F.W.Hehl, N.Ito, V.Winkelmann, J.G.Zabolitzky, Computer simulation
and of principle or computer algebra: lectures for beginners, (Springer Science & Business Media, 2012).
[16] W.Hergert, M. Däne, “Group theoretical investigations of photonic Band Structures”,
Phys. Stat. Sol. (a) 197, No. 3, 620–634 (2003).
[17] W.Hergert, M. Däne, and D. Ködderitzsch, Symmetry Properties of Electronic and
Photonic Band Structures, Computational Materials Science, volume 642 of the series
Lecture Notes in Physics, pp 103–125, (Springer Science & Business Media, 2004).
[18] J.Thijssen,Computational physics, (Cambridge university press, 2007).
[19] H.Sakiyama, K.Waki, “Conformational Analysis of Hexagon’s-Thiamine Nickel(II)
Complex on the Basis of Computational Group Theory and Density Functional Theory”, J. Computed. Chem. Jpn., Vol. 13, No. 4, pp. 223–228 (2014).
[20] E. Rykhlinskaya, “A Computer-algebraic Approach to the Study of the Symmetry
Properties of Molecules and Clusters”. Doctoral thesis, Kassel University, 2006.
[21] S.Fritzshche, “Application of Point-Group Symmetries in Chemistry and Physics: A
Computer-Algebraic Approach”, International Journal of Quantum Chemistry, Vol
106, 98–129 (2006).
[22] M. P. Barnett,“Computer algebra in the life sciences”,Communications in Computer
Algebra, Vol 36, No.4, 5–32, 2002.
[23] M. I. Aroyo, J. M. Perez-Mato, D. Orobengoa, E. Tasci, G. de la Flor, A. Kirov, “Crystallography online: Bilbao Crystallographic Server”, Bulg. Chem. Commun. 43(2)
183-197 (2011).
[24] M. I. Aroyo, J. M. Perez-Mato, C. Capillas, E. Kroumova, S. Ivantchev, G.
197
Madariaga, A. Kirov and H. Wondratschek “Bilbao Crystallographic Server I:
Databases and crystallographic computing programs” Z. Krist. 221, 1, 15–27 (2006).
doi:10.1524/zkri.2006.221.1.15.
[25] M. I. Aroyo, A. Kirov, C. Capillas, J. M. Perez-Mato and H. Wondratschek “Bilbao
Crystallographic Server II: Representations of crystallographic point groups and space
groups” Acta Cryst. A62, 115–128 (2006). doi:10.1107/S0108767305040286.
[26] B. Kostant, “The Graph of the Truncated Icosahedron and the Last Letter of Galois”,
Notice of AMS, Volume 42, No. 9, 959–968, (1995).
[27] F.R.K.Chung, B. Kostant, B., S. Sternberg, “Groups and the buckyball”, in Lie Theory and geometry, Birkhauser, Boston, pp 97–126,(1994).
[28] B. Eick and B. Souvignier, “Algorithms for crystallographic groups”, International
Journal of Quantum Chemistry 106, 316–343 (2006).
[29] J. D. Dixon, “High speed computation of group characters”, Numer. Math. 10,446–
450, 1099, (1967).
[30] G. J. A. Schneider,“ Dixon’s character table algorithm revisited”, J. Symbolic Comput.9(5-6):601–606, (1990).
[31] A. Hulpke, “Zur Berechnung von Charaktertafeln”, Diplomarbeit, Lehrstuhl D für
Mathematik, Rheinisch Westfälische Technische Hochschule, (1993).
[32] C. Herring, “Character tables for two space groups”, J. Franklin, Inst. 233, 525–543
(1942).
[33] R. W. Wyckoff, The Analytical Expression of the Results of the theory of Space Groups,
Carnegie Institute: Washington, DC,(1922).
[34] O. N. Nyrasov and A. J. Freeman, “Electronic band structure of indium tin oxide and
criteria for transparent conducting behavior”, Phys.Rev.B 64,233111-1–3(2001).
[35] P. Etingof, O. Golberg, S. Hensel, T. Liu, A. Schwendner, D. Vaintrob, and E.
Yudovina, Introduction to representation theory, <http://math.mit.edu/~etingof/
198
replect.pdf>, p76–77.
[36] P. J. Webb, “An Introduction to the Cohomology of Groups”, <http://www.math.
umn.edu/~webb/oldteaching/Year2010-11/8246CohomologyNotes.pdf>.
[37] R.C.Haddon, L.E.Brus. K.Raghavachari,“Electronic structure and binding in icosahedral C60 , Chem. Phys. Lett, 125, issues 5–6, 459–464, (1986).
[38] W. H. Green, Jr, S. M. Gorun, G. Fitzgerald, P. W. Fowler, A. Ceulemans and B.
Titeca, “Electronic Structures and Geometries of C60 Anions via Density Functional
Calculations”, J.Phys.Chem, 100, 14892–14898, (1996).
[39] V. de Coulon, J.L.Martins, F.Reuse, “Electronic structure of neutral and charged C60
clusters”, Phys. Rev. B 45, No. 23, 13671–13675, (1992).
[40] G.
und
Polya,
“Kombinatorische
chemische
Verbindungen”,
Anzahlbestimmungen
Acta
Mathematica
für
Gruppen,
68(1):
Graphen
145–254(1937),
doi:10.1007/BF02546665.
[41] R. Schmied, K. K. Lehmann, “Computer-generated character tables and nuclear spin
statistical weights: Application to benzene dimer and methane dimer”,Journal of
Molecular Spectroscopy 226 201–202 (2004), DOI:10.1016/j.jms.2004.04.003.
[42] D. M. Jonas,“Spin statistics: An error in Landau and Lifschitz’ Quantum Mechanics”,
J.Chem.Phys.90(10),5563–5565 (1989).
[43] L. D. Landau and E. M. Lifschitz, Quantum Mechanics, 3rd ed. Sec.105 (Pergamon,1977).
[44] A. Kikuchi, “An approach to first principles electronic structure computation by
symbolic-numeric computation”, QScience Connect 2013:14 (2013), <http://dx.doi.
org/10.5339/connect.2013.14>.
[45] R. Courant and D. Hilbert, Method of mathematical physics, Vol.1–2,(WileyVCH,1989).
199
SUPPLEMENT: MATRIX GENERATORS OF THE POINT GROUP
In this supplement I give the matrix generators, in order to define the point group.
They are written in the GAP language.
#
# THE MATRIX IN THE POINT GROUP OF CUBIC LATTICE.
#
# THE OPERATIONS IN THE REAL SPACE.
#
CMR:=[[[1,0,0],[0,1,0],[0,0,1]],[[1,0,0],[0,-1,0],[0,0,-1]],
[[-1,0,0],[0,1,0],[0,0,-1]],[[-1,0,0],[0,-1,0],[0,0,1]],
[[0,1,0],[0,0,1],[1,0,0]],[[0,1,0],[0,0,-1],[-1,0,0]],
[[0,-1,0],[0,0,1],[-1,0,0]],[[0,-1,0],[0,0,-1],[1,0,0]],
[[0,0,1],[1,0,0],[0,1,0]],[[0,0,1],[-1,0,0],[0,-1,0]],
[[0,0,-1],[1,0,0],[0,-1,0]],[[0,0,-1],[-1,0,0],[0,1,0]],
[[0,-1,0],[-1,0,0],[0,0,-1]],[[0,-1,0],[1,0,0],[0,0,1]],
[[0,1,0],[-1,0,0],[0,0,1]],[[0,1,0],[1,0,0],[0,0,-1]],
[[-1,0,0],[0,0,-1],[0,-1,0]],[[-1,0,0],[0,0,1],[0,1,0]],
[[1,0,0],[0,0,-1],[0,1,0]],[[1,0,0],[0,0,1],[0,-1,0]],
[[0,0,-1],[0,-1,0],[-1,0,0]],[[0,0,-1],[0,1,0],[1,0,0]],
[[0,0,1],[0,-1,0],[1,0,0]],[[0,0,1],[0,1,0],[-1,0,0]],
[[-1,0,0],[0,-1,0],[0,0,-1]],[[-1,0,0],[0,1,0],[0,0,1]],
[[1,0,0],[0,-1,0],[0,0,1]],[[1,0,0],[0,1,0],[0,0,-1]],
[[0,-1,0],[0,0,-1],[-1,0,0]],[[0,-1,0],[0,0,1],[1,0,0]],
[[0,1,0],[0,0,-1],[1,0,0]],[[0,1,0],[0,0,1],[-1,0,0]],
[[0,0,-1],[-1,0,0],[0,-1,0]],[[0,0,-1],[1,0,0],[0,1,0]],
[[0,0,1],[-1,0,0],[0,1,0]],[[0,0,1],[1,0,0],[0,-1,0]],
200
[[0,1,0],[1,0,0],[0,0,1]],[[0,1,0],[-1,0,0],[0,0,-1]],
[[0,-1,0],[1,0,0],[0,0,-1]],[[0,-1,0],[-1,0,0],[0,0,1]],
[[1,0,0],[0,0,1],[0,1,0]],[[1,0,0],[0,0,-1],[0,-1,0]],
[[-1,0,0],[0,0,1],[0,-1,0]],[[-1,0,0],[0,0,-1],[0,1,0]],
[[0,0,1],[0,1,0],[1,0,0]],[[0,0,1],[0,-1,0],[-1,0,0]],
[[0,0,-1],[0,1,0],[-1,0,0]],[[0,0,-1],[0,-1,0],[1,0,0]]];;
#
# THE OPERATIONS IN THE RECIPROCAL SPACE.
#
CMT:=[[[1,0,0],[0,1,0],[0,0,1]],[[-1,0,0],[-1,0,1],[-1,1,0]],
[[0,-1,1],[0,-1,0],[1,-1,0]],[[0,1,-1],[1,0,-1],[0,0,-1]],
[[0,1,0],[0,0,1],[1,0,0]],[[0,-1,0],[1,-1,0],[0,-1,1]],
[[1,0,-1],[0,0,-1],[0,1,-1]],[[-1,0,1],[-1,1,0],[-1,0,0]],
[[0,0,1],[1,0,0],[0,1,0]],[[0,0,-1],[0,1,-1],[1,0,-1]],
[[-1,1,0],[-1,0,0],[-1,0,1]],[[1,-1,0],[0,-1,1],[0,-1,0]],
[[0,-1,0],[-1,0,0],[0,0,-1]],[[0,1,0],[0,1,-1],[-1,1,0]],
[[1,0,-1],[1,0,0],[1,-1,0]],[[-1,0,1],[0,-1,1],[0,0,1]],
[[-1,0,0],[0,0,-1],[0,-1,0]],[[1,0,0],[1,-1,0],[1,0,-1]],
[[0,-1,1],[0,0,1],[-1,0,1]],[[0,1,-1],[-1,1,0],[0,1,0]],
[[0,0,-1],[0,-1,0],[-1,0,0]],[[0,0,1],[-1,0,1],[0,-1,1]],
[[-1,1,0],[0,1,0],[0,1,-1]],[[1,-1,0],[1,0,-1],[1,0,0]],
[[-1,0,0],[0,-1,0],[0,0,-1]],[[1,0,0],[1,0,-1],[1,-1,0]],
[[0,1,-1],[0,1,0],[-1,1,0]],[[0,-1,1],[-1,0,1],[0,0,1]],
[[0,-1,0],[0,0,-1],[-1,0,0]],[[0,1,0],[-1,1,0],[0,1,-1]],
[[-1,0,1],[0,0,1],[0,-1,1]],[[1,0,-1],[1,-1,0],[1,0,0]],
[[0,0,-1],[-1,0,0],[0,-1,0]],[[0,0,1],[0,-1,1],[-1,0,1]],
[[1,-1,0],[1,0,0],[1,0,-1]],[[-1,1,0],[0,1,-1],[0,1,0]],
201
[[0,1,0],[1,0,0],[0,0,1]],[[0,-1,0],[0,-1,1],[1,-1,0]],
[[-1,0,1],[-1,0,0],[-1,1,0]],[[1,0,-1],[0,1,-1],[0,0,-1]],
[[1,0,0],[0,0,1],[0,1,0]],[[-1,0,0],[-1,1,0],[-1,0,1]],
[[0,1,-1],[0,0,-1],[1,0,-1]],[[0,-1,1],[1,-1,0],[0,-1,0]],
[[0,0,1],[0,1,0],[1,0,0]],[[0,0,-1],[1,0,-1],[0,1,-1]],
[[1,-1,0],[0,-1,0],[0,-1,1]],[[-1,1,0],[-1,0,1],[-1,0,0]]];;
#
# THE PRIMITIVE TRANSLATIONS.
#
A:=[[0,1,1],[1,0,1],[1,1,0]]/2;;
#
# TRANSFORMATION FROM MR TO MT;
#
CMT=List(CMR,m->TransposedMat(A^-1*TransposedMat(m)*A));
#
# THE MULTIPLICATION TABLE.
#
TBL:=List(CMR,i->List(CMR,j->Position(CMR,i*j)));
#
#
# THE MATRIX IN THE POINT GROUP OF HEXAGONAL LATTICE.
#
# THE PRIMITIVE TRANSLATIONS HA:=[A1,A2,A3];
#
[ Sqrt(3)/2=-1/2*E(12)^7+1/2*E(12)^11 ]
#
HA:=[[1,-1/2,0],[0,-1/2*E(12)^7+1/2*E(12)^11,0],[0,0,1]];;
#
202
# THE OPERATIONS IN THE REAL SPACE.
#
HMR:=[[[1,0,0],[0,1,0],[0,0,1]],
[[1/2,1/2*E(12)^7-1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,1/2,0],[0,0,1]],
[[-1/2,1/2*E(12)^7-1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,-1/2,0],[0,0,1]],
[[-1,0,0],[0,-1,0],[0,0,1]],
[[-1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,-1/2,0],[0,0,1]],
[[1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,1/2,0],[0,0,1]],
[[-1/2,1/2*E(12)^7-1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,1/2,0],[0,0,-1]],
[[1/2,1/2*E(12)^7-1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,-1/2,0],[0,0,-1]],
[[1,0,0],[0,-1,0],[0,0,-1]],
[[1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,-1/2,0],[0,0,-1]],
[[-1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,1/2,0],[0,0,-1]],
[[-1,0,0],[0,1,0],[0,0,-1]],[[-1,0,0],[0,-1,0],[0,0,-1]],
[[-1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,-1/2,0],[0,0,-1]],
[[1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,1/2,0],[0,0,-1]],
[[1,0,0],[0,1,0],[0,0,-1]],
[[1/2,1/2*E(12)^7-1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,1/2,0],[0,0,-1]],
[[-1/2,1/2*E(12)^7-1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,-1/2,0],[0,0,-1]],
[[1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,-1/2,0],[0,0,1]],
[[-1/2,-1/2*E(12)^7+1/2*E(12)^11,0],[-1/2*E(12)^7+1/2*E(12)^11,1/2,0],[0,0,1]],
[[-1,0,0],[0,1,0],[0,0,1]],
[[-1/2,1/2*E(12)^7-1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,1/2,0],[0,0,1]],
[[1/2,1/2*E(12)^7-1/2*E(12)^11,0],[1/2*E(12)^7-1/2*E(12)^11,-1/2,0],[0,0,1]],
[[1,0,0],[0,-1,0],[0,0,1]]];;
#
#
203
# THE OPERATIONS IN THE RECIPROCAL SPACE.
#
HMT:=[[[1,0,0],[0,1,0],[0,0,1]],[[0,-1,0],[1,1,0],[0,0,1]],
[[-1,-1,0],[1,0,0],[0,0,1]],[[-1,0,0],[0,-1,0],[0,0,1]],
[[0,1,0],[-1,-1,0],[0,0,1]],[[1,1,0],[-1,0,0],[0,0,1]],
[[-1,-1,0],[0,1,0],[0,0,-1]],[[0,-1,0],[-1,0,0],[0,0,-1]],
[[1,0,0],[-1,-1,0],[0,0,-1]],[[1,1,0],[0,-1,0],[0,0,-1]],
[[0,1,0],[1,0,0],[0,0,-1]],[[-1,0,0],[1,1,0],[0,0,-1]],
[[-1,0,0],[0,-1,0],[0,0,-1]],[[0,1,0],[-1,-1,0],[0,0,-1]],
[[1,1,0],[-1,0,0],[0,0,-1]],[[1,0,0],[0,1,0],[0,0,-1]],
[[0,-1,0],[1,1,0],[0,0,-1]],[[-1,-1,0],[1,0,0],[0,0,-1]],
[[1,1,0],[0,-1,0],[0,0,1]],[[0,1,0],[1,0,0],[0,0,1]],
[[-1,0,0],[1,1,0],[0,0,1]],[[-1,-1,0],[0,1,0],[0,0,1]],
[[0,-1,0],[-1,0,0],[0,0,1]],[[1,0,0],[-1,-1,0],[0,0,1]]];;
#
#
# THE OPERATIONS (ROTATIONS AND TRANSLATIONS),
# AS ARE GIVEN IN THE TABLES IN THE ARTICLE.
#
# CUBIC LATTICE.(DIAMOND)
#
OPRCUB:=[[1,1,0,0,0,1,0,0,0,1,0,0,0,0,0,0],
[2,1,0,0,0,-1,0,0,0,-1,1/2,0,0,0,0,0],
[3,-1,0,0,0,1,0,0,0,-1,0,1/2,0,0,0,0],
[4,-1,0,0,0,-1,0,0,0,1,0,0,1/2,0,0,0],
[5,0,1,0,0,0,1,1,0,0,0,0,0,0,0,0],
[6,0,1,0,0,0,-1,-1,0,0,1/2,0,0,0,0,0],
204
[7,0,-1,0,0,0,1,-1,0,0,0,1/2,0,0,0,0],
[8,0,-1,0,0,0,-1,1,0,0,0,0,1/2,0,0,0],
[9,0,0,1,1,0,0,0,1,0,0,0,0,0,0,0],
[10,0,0,1,-1,0,0,0,-1,0,1/2,0,0,0,0,0],
[11,0,0,-1,1,0,0,0,-1,0,0,1/2,0,0,0,0],
[12,0,0,-1,-1,0,0,0,1,0,0,0,1/2,0,0,0],
[13,0,-1,0,-1,0,0,0,0,-1,0,0,0,1/4,1/4,1/4],
[14,0,-1,0,1,0,0,0,0,1,-1/2,0,0,1/4,1/4,1/4],
[15,0,1,0,-1,0,0,0,0,1,0,-1/2,0,1/4,1/4,1/4],
[16,0,1,0,1,0,0,0,0,-1,0,0,-1/2,1/4,1/4,1/4],
[17,-1,0,0,0,0,-1,0,-1,0,0,0,0,1/4,1/4,1/4],
[18,-1,0,0,0,0,1,0,1,0,-1/2,0,0,1/4,1/4,1/4],
[19,1,0,0,0,0,-1,0,1,0,0,-1/2,0,1/4,1/4,1/4],
[20,1,0,0,0,0,1,0,-1,0,0,0,-1/2,1/4,1/4,1/4],
[21,0,0,-1,0,-1,0,-1,0,0,0,0,0,1/4,1/4,1/4],
[22,0,0,-1,0,1,0,1,0,0,-1/2,0,0,1/4,1/4,1/4],
[23,0,0,1,0,-1,0,1,0,0,0,-1/2,0,1/4,1/4,1/4],
[24,0,0,1,0,1,0,-1,0,0,0,0,-1/2,1/4,1/4,1/4],
[25,-1,0,0,0,-1,0,0,0,-1,0,0,0,1/4,1/4,1/4],
[26,-1,0,0,0,1,0,0,0,1,-1/2,0,0,1/4,1/4,1/4],
[27,1,0,0,0,-1,0,0,0,1,0,-1/2,0,1/4,1/4,1/4],
[28,1,0,0,0,1,0,0,0,-1,0,0,-1/2,1/4,1/4,1/4],
[29,0,-1,0,0,0,-1,-1,0,0,0,0,0,1/4,1/4,1/4],
[30,0,-1,0,0,0,1,1,0,0,-1/2,0,0,1/4,1/4,1/4],
[31,0,1,0,0,0,-1,1,0,0,0,-1/2,0,1/4,1/4,1/4],
[32,0,1,0,0,0,1,-1,0,0,0,0,-1/2,1/4,1/4,1/4],
[33,0,0,-1,-1,0,0,0,-1,0,0,0,0,1/4,1/4,1/4],
205
[34,0,0,-1,1,0,0,0,1,0,-1/2,0,0,1/4,1/4,1/4],
[35,0,0,1,-1,0,0,0,1,0,0,-1/2,0,1/4,1/4,1/4],
[36,0,0,1,1,0,0,0,-1,0,0,0,-1/2,1/4,1/4,1/4],
[37,0,1,0,1,0,0,0,0,1,0,0,0,0,0,0],
[38,0,1,0,-1,0,0,0,0,-1,1/2,0,0,0,0,0],
[39,0,-1,0,1,0,0,0,0,-1,0,1/2,0,0,0,0],
[40,0,-1,0,-1,0,0,0,0,1,0,0,1/2,0,0,0],
[41,1,0,0,0,0,1,0,1,0,0,0,0,0,0,0],
[42,1,0,0,0,0,-1,0,-1,0,1/2,0,0,0,0,0],
[43,-1,0,0,0,0,1,0,-1,0,0,1/2,0,0,0,0],
[44,-1,0,0,0,0,-1,0,1,0,0,0,1/2,0,0,0],
[45,0,0,1,0,1,0,1,0,0,0,0,0,0,0,0],
[46,0,0,1,0,-1,0,-1,0,0,1/2,0,0,0,0,0],
[47,0,0,-1,0,1,0,-1,0,0,0,1/2,0,0,0,0],
[48,0,0,-1,0,-1,0,1,0,0,0,0,1/2,0,0,0]];;
#
#
# HEXAGONAL LATTICE.
#
OPRHEXA:=[[1,1,0,0,0,1,0,0,0,1,0,0,0],
[2,1/2,-Sqrt(3)/2,0,Sqrt(3)/2,1/2,0,0,0,1,0,0,-1/2],
[3,-1/2,-Sqrt(3)/2,0,Sqrt(3)/2,-1/2,0,0,0,1,0,0,0],
[4,-1,0,0,0,-1,0,0,0,1,0,0,-1/2],
[5,-1/2,Sqrt(3)/2,0,-Sqrt(3)/2,-1/2,0,0,0,1,0,0,0],
[6,1/2,Sqrt(3)/2,0,-Sqrt(3)/2,1/2,0,0,0,1,0,0,-1/2],
[7,-1/2,-Sqrt(3)/2,0,-Sqrt(3)/2,1/2,0,0,0,-1,0,0,-1/2],
[8,1/2,-Sqrt(3)/2,0,-Sqrt(3)/2,-1/2,0,0,0,-1,0,0,0],
206
[9,1,0,0,0,-1,0,0,0,-1,0,0,-1/2],
[10,1/2,Sqrt(3)/2,0,Sqrt(3)/2,-1/2,0,0,0,-1,0,0,0],
[11,-1/2,Sqrt(3)/2,0,Sqrt(3)/2,1/2,0,0,0,-1,0,0,-1/2],
[12,-1,0,0,0,1,0,0,0,-1,0,0,0],
[13,-1,0,0,0,-1,0,0,0,-1,0,0,-1/2],
[14,-1/2,Sqrt(3)/2,0,-Sqrt(3)/2,-1/2,0,0,0,-1,0,0,0],
[15,1/2,Sqrt(3)/2,0,-Sqrt(3)/2,1/2,0,0,0,-1,0,0,-1/2],
[16,1,0,0,0,1,0,0,0,-1,0,0,0],
[17,1/2,-Sqrt(3)/2,0,Sqrt(3)/2,1/2,0,0,0,-1,0,0,-1/2],
[18,-1/2,-Sqrt(3)/2,0,Sqrt(3)/2,-1/2,0,0,0,-1,0,0,0],
[19,1/2,Sqrt(3)/2,0,Sqrt(3)/2,-1/2,0,0,0,1,0,0,0],
[20,-1/2,Sqrt(3)/2,0,Sqrt(3)/2,1/2,0,0,0,1,0,0,-1/2],
[21,-1,0,0,0,1,0,0,0,1,0,0,0],
[22,-1/2,-Sqrt(3)/2,0,-Sqrt(3)/2,1/2,0,0,0,1,0,0,-1/2],
[23,1/2,-Sqrt(3)/2,0,-Sqrt(3)/2,-1/2,0,0,0,1,0,0,0],
[24,1,0,0,0,-1,0,0,0,1,0,0,-1/2]];;
SUPPLEMENT:SOME FUNCTIONS USED IN THE COMPUTATIONS
AT ”SYMMETRY IN C60 ”
#
# This file contains some functions used in the computations
# in VII. "SYMMETRY IN C60".
#
listtoh:=function(A)
207
local h,i1,i2,i3,i4,k,l,i;
h:=List([1..120],i->List([1..120],j->0));
for i in A do
i1:=i[1];
i2:=i[2];
i3:=i[3];
i4:=i[4];
k:=(i1-2)+60*(i2-1);
l:=(i3-2)+60*(i4-1);
h[k][l]:=1;
h[l][k]:=1;
od;
return h;
end;
#
# We construct an extended group by the direct product
# between A5 and one of the automorphism group.
#
# The new group contains 120 elements, represented by 62 symbols.
# The symbols from the 3rd to the 62nd are descendants from A5.
# The symbols of the first and the second are the newcomers
# in the extension.
#
# From the 62 symbols, we construct a bonding system
# which contains 120 atoms, regulated by the symmetry of the
# new group. (C120)
208
# The atoms will be indexed as (i,j) (i=3,..,62, and j=1,2)
# by the duplication of single C60.
#
#
# The generators of the group. (A5 and I)
#
g1:=(
(
1, 14, 20)(
6, 58, 22)(
2, 15, 16)(
7, 57, 23)(
3, 11, 17)(
8, 56, 24)(
4, 12, 18)(
5, 13, 19)
9, 60, 25)( 10, 59, 21)
( 31, 44, 50)( 32, 45, 46)( 33, 41, 47)( 34, 42, 48)( 35, 43, 49)
( 36, 28, 52)( 37, 27, 53)( 38, 26, 54)( 39, 30, 55)( 40, 29, 51);
g2:=(
(
1,
2,
3,
4,
8, 13, 18, 23, 28)(
5)(
6, 11, 16, 21, 26)(
7, 12, 17, 22, 27)
9, 14, 19, 24, 29)( 10, 15, 20, 25, 30)
( 31, 32, 33, 34, 35)( 36, 41, 46, 51, 56)( 37, 42, 47, 52, 57)
( 38, 43, 48, 53, 58)( 39, 44, 49, 54, 59)( 40, 45, 50, 55, 60);
g3:=(
(
1, 31)(
8, 38)(
2, 32)(
3, 33)(
4, 34)(
5, 35)(
6, 36)(
7, 37)
9, 39)( 10, 40)( 11, 41)( 12, 42)( 13, 43)( 14, 44)
( 15, 45)( 16, 46)( 17, 47)( 18, 48)( 19, 49)( 20, 50)( 21, 51)
( 22, 52)( 23, 53)( 24, 54)( 25, 55)( 26, 56)( 27, 57)( 28, 58)
( 29, 59)( 30, 60);
# I define g2 in this way in the article.
g2:=g2^-1;
#
209
G:=Group(g1,g2);
AUALL:=AutomorphismGroup(G);
p:=SemidirectProduct(Group(AUALL.1),G);
N:=Image(Embedding(p,2));
#
# The orbits of the double bond [(3,1),(4,1)]
# and the single bonds [3,1,8,1]
# and the bond between the first and the second C60
# are computed as A1,B1,C1
#
# The bonding structures from them are represented by
# three matrices a1,b1,c1.
#
A1:=Orbit(p,[3,1,4,1],OnTuples);
B1:=Orbit(p,[3,1,8,1],OnTuples);
C1:=Orbit(p,[3,1,4,2],OnTuples);
a1:=listtoh(A1);
b1:=listtoh(B1);
c1:=listtoh(C1);
#
# The symmetry operations of the 120 atoms are written as:
#
l1:=List([1..2],j->List([3..62],i->(OnPoints(i,p.1)+60*(j-1))));
l1:=Flat(l1);
210
l2:=List([1..2],j->List([3..62],i->(OnPoints(i,p.2)+60*(j-1))));
l2:=Flat(l2);
l3:=List([1..2],j->List([3..62],i->(OnPoints(i,p.3)+60*(OnPoints(j,p.3)-1))));
l3:=Flat(l3);
#
#
inv1:=List([1..60],i->OnPoints(i,g3)+2);
inv2:=inv1+60;
inv:=Flat([inv1,inv2]);
ListToMat:=function(l)
local size,h,i,j;
size:=Size(l);
h:=List([1..size],i->List([1..size],j->0));
for i in [1..size] do
h[i][l[i]]:=1;
od;
return h;
end;
#
# Now generators (p.1,p.2,p.3) are represented by matrix P1,P2,P3.
#
P1:=ListToMat(l1-2);
P2:=ListToMat(l2-2);
211
P3:=ListToMat(l3-2);
#
# In C60, of the icosahedral symmetry, the symmetry operation also includes
# the inversion. The inversion itself is not includes in the new group,
# but we can write the operation of this operation on first 60 vertexes
# (from the 3rd to the 62nd)
# The inversion in the second 60 vertexes can be written in the following way.
# The second 60 vertexes are transmuted from the first 60, by the operation p.3.
#
# If the inversions on the first and second 60 vertexes are denoted as I1,I2,
# they are related by the conjugation I2:= (p.3)*I1*(p.3)^(-1).
# The inversion on the total system is written as
# [[ I1,
#
[
0]
0, I2]].
#
#
# ListInv: the operation of "g3" in the first shell.
ListInv:=List([1..60],i->OnPoints(i,g3));
# mconjg : the operation of "p.3" from the first to the second shell.
mconjg:=List([3..62],i->OnPoints(i,p.3))-2;
#
# These operations are put into matrix, "invrs", "mc".
#
invrs:=List([1..60],i->List([1..60],j->0));
212
mc:=List([1..60],i->List([1..60],j->0));
for i in [1..60] do
invrs[i][ListInv[i]]:=1;
mc[i][mconjg[i]]:=1;
od;
invrs2:=mc*invrs*mc;
# "invrs": the inversion in the first shell.
# "invrs2": the inversion in the second shell.
#
# Set the inversions in two shells in a matrix "ivfull".
#
ivfull:=List([1..120],i->List([1..120],j->0));
for i in [1..60] do
for j in [1..60] do
ivfull[i][j]:=invrs[i][j];
ivfull[i+60][j+60]:=invrs2[i][j];
od;
od;
#
# As the matrix c1 are not conserved by the inversion "ivfull",
# (as ivfull*c1*ivfull^-1 is not equal to c1)
# it is modified to have inversion symmetry.
#
hh:=-2*a1-b1-1/2*(c1+ivfull*c1*ivfull);
213
#
# The generalized eigenvalue problems are solved in
# the field of rational numbers.
#
gev:=GeneralisedEigenvalues(Rationals,hh);
ges:=GeneralisedEigenspaces(Rationals,hh);
#
# The irreducible representation is computed.
#
# "irrg" is the group of the symmetry in the composed structure.
# "irrg2" is the subgroup.
#
irrg:=Irr(Group(P1,P2,P3,ivfull));
irrg2:=Irr(Group(P1,P2,ivfull));
Em:=Elements(Group(P1,P2,P3,ivfull));;
Em2:=Elements(Group(P1,P2,ivfull));;
Projopr:=function(irr,Em)
#
# The projector, by means of characters.
#
return List(irr,i->Sum(List(Em,j->j^i*j))/Size(Em)*i[1]);
end;
prjs:=Projopr(irrg,Em);;
214
prjs2:=Projopr(irrg2,Em2);;
#
# The basis vectors are computed and allotted to the irreducible representation.
#
bases:=List(ges,i->Basis(i));
for b in bases do;Display(List(b,j->List( prjs,p->(p*j)^2 ))); od;
for b in bases do;Display(List(b,j->List(prjs2,p->(p*j)^2 )));od;
SUPPLEMENT:SOME FUNCTIONS USED IN THE COMPUTATIONS
AT ”ANALISYS OF VIBRATIONAL MODE IN C60 ”
#
# THIS FILE CONTAINS SMALL FUNCTIONS FOR THE COMPUTATION CONCERNING C60.
# (VIII."ANALYSIS OF VIBRATIONAL MODE IN C60")
#
cc:=(E(5)+E(5)^4)/2;
ss:=(E(5)-E(5)^4)/2/E(4);
cc10:=(E(10)+E(10)^4)/2;
ss10:=(E(10)-E(10)^4)/2/E(4);
x:=Indeterminate(Rationals,"x");
z:=Indeterminate(Rationals,"z");
y:=Indeterminate(Rationals,"y");
215
#
# MATRIX OF ROTATION OF 72 DEGREES (A GENERATOR OF GROUP A5).
#
M72D:=[[cc,-ss,0],[ss,cc,0],[0,0,1]];
#
#
# HEAFTER, VERY TRIVIAL COMPUTATIONS ARE DONE
# IN ORDER TO GIVE 12 VERTEX ON ICOSAHEDRON.
#
v:=[[0,0,0],[0,0,0],[0,0,0]];
v[1]:=[0*x,0*x,x^0];
v[2]:=[x,0*x,z];
v[3]:=M72D^-1*v[2];
v21:=v[2]-v[1];
v32:=v[3]-v[2];
L21:=v21*v21;
L32:=v32*v32;
B:=[L21-L32,v[2]*v[2]-1];
GB:=GroebnerBasis(B,MonomialLexOrdering());
RTPS:=RootsOfPolynomial(GB[3]);
zval:=RTPS[2];
xval:=Sqrt(1-zval*zval);
vval:=List(v,i->List(i,j->Value(j,[x,z],[xval,zval])));
216
#
# NOW 12 VERTEX ARE PREPARED.
#
sommets:=List([1..12],i->0);
sommets[1]:=vval[1];
sommets[2]:=
vval[2];
sommets[3]:=M72D^-1*vval[2];
sommets[4]:=M72D^-2*vval[2];
sommets[5]:=M72D^-3*vval[2];
sommets[6]:=M72D^-4*vval[2];
sommets[7]:=-vval[1];
sommets[8]:=-
vval[2];
sommets[9]:=-M72D^-1*vval[2];
sommets[10]:=-M72D^-2*vval[2];
sommets[11]:=-M72D^-3*vval[2];
sommets[12]:=-M72D^-4*vval[2];
#
Tv:=TransposedMat(vval);
Tv2:=TransposedMat([vval[2],vval[3],vval[1]]);
Inv:=[[-1,0,0],[0,-1,0],[0,0,-1]];
#
# THE THREE VECTORS, STORED IN THE COLUMNS in "Tv" AND "Tv2"
# ARE RELATED BY THE ROTATION, WHICH IS OTHER GENERATOR OF A5.
#
# THE GROUP (A5 x I) , DEFINED IN 3D-EUCLID SPACE.(THE GROUP G)
217
#
# THE GENERATORS OF THIS GROUP (G.1,G.2,G.3) ARE (g1,g2,g3) IN THE ARTICLE.
#
G:=Group(Tv2*Tv^-1,M72D^-1,Inv);
irrg:=Irr(G);
elm:=Elements(G);
#
# THE PROJECTORS, EXPLANED IN THE ARTICLE.
#
Projopr:=function(irr,Em,hom)
#
return List(irr,i->Sum(List(Em,j->j^i*Image(hom,j)))/Size(Em)*i[1]);
end;
ProjoprR:=function(irr,Em,MM,hom)
#
return
List(irr,i->Sum(List(Em,j->j^i*Image(hom,j)*MM*TransposedMat(Image(hom,j))))
/Size(Em)*i[1]);
end;
#
# THE DEFINITION OF GROUP (A5 x I) BY THE PERMUTATION OF 12 VERTEX IN ICOSAHEDRON
# (GROUP F)
218
#
f1:=(1,2,3)(4,6,11)(5,10,12)(7,8,9);
f2:=(2,3,4,5,6)(8,9,10,11,12);
f3:=(1,7)(2,8)(3,9)(4,10)(5,11)(6,12);
F:=Group(f1,f2,f3);
#
# THE DEFINITION OF GROUP (A5 x I) BY THE PERMUTATION OF 60 VERTEX IN C60
# (GROUP R)
#
r1:=(
(
1, 14, 20)(
6, 58, 22)(
2, 15, 16)(
7, 57, 23)(
3, 11, 17)(
8, 56, 24)(
4, 12, 18)(
5, 13, 19)
9, 60, 25)( 10, 59, 21)
( 31, 44, 50)( 32, 45, 46)( 33, 41, 47)( 34, 42, 48)( 35, 43, 49)
( 36, 28, 52)( 37, 27, 53)( 38, 26, 54)( 39, 30, 55)( 40, 29, 51);
r2:=(
(
1,
2,
3,
4,
8, 13, 18, 23, 28)(
5)(
6, 11, 16, 21, 26)(
7, 12, 17, 22, 27)
9, 14, 19, 24, 29)( 10, 15, 20, 25, 30)
( 31, 32, 33, 34, 35)( 36, 41, 46, 51, 56)( 37, 42, 47, 52, 57)
( 38, 43, 48, 53, 58)( 39, 44, 49, 54, 59)( 40, 45, 50, 55, 60);
r3:=(
(
1, 31)(
8, 38)(
2, 32)(
3, 33)(
4, 34)(
5, 35)(
6, 36)(
7, 37)
9, 39)( 10, 40)( 11, 41)( 12, 42)( 13, 43)( 14, 44)
( 15, 45)( 16, 46)( 17, 47)( 18, 48)( 19, 49)( 20, 50)( 21, 51)
( 22, 52)( 23, 53)( 24, 54)( 25, 55)( 26, 56)( 27, 57)( 28, 58)
( 29, 59)( 30, 60);
R:=Group(r1,r2,r3);
#
# THE ISOMORPHISM BETWEEN GROUPS
219
#
iso:=IsomorphismGroups(F,R);
iso2:=IsomorphismGroups(F,G);
hom:=GroupHomomorphismByImages(F,G,[F.1,F.2,F.3],[G.1,G.2,G.3]);
hom2:=GroupHomomorphismByImages(R,G,[R.1,R.2,R.3],[G.1,G.2,G.3]);
Display(List(Irr(R),r->List([R.1,R.2,R.3],i->i^r)));
MM:=[[0*x,x,y],[-x,0*x,z],[-y,-z,0*x]];
PRJ:=Projopr(Irr(R),Elements(R),hom2);
PRJR:=ProjoprR(Irr(R),Elements(R),MM,hom2);
#
# THE FUNCTIONS DEFINED AND EXPLAINED IN THE ARTICLE.
#
#
ListToMat:=function(l)
local size,h,i,j;
size:=Size(l);
h:=List([1..size],i->List([1..size],j->0));
for i in [1..size] do
h[i][l[i]]:=1;
od;
return h;
end;
LargeMatrix:=function(rr,iso)
220
local l,rmat,rrimg,i,i1,i2,n,m;
l:=ListPerm(rr,60);
rmat:=List([1..Size(l)*3],i->List([1..Size(l)*3],j->0));
rrimg:=Image(iso,rr);
for i in [1..Size(l)]
do
i1:=i-1;i2:=l[i]-1;
for n in [1..3]
do
for m in [1..3]
do
rmat[3*i1+n][3*i2+m]:=rrimg[n][m];
od;
od;
od;
return rmat;
end;
#
# THE DEFINITION OF PROJECTOR, DEFINED IN THE ARTICLE
# IN ORDER TO COMPUTE THE MODE OF DEFORMATION IN C60.
#
ProjoprE:=function(irr,Em,hom2)
#
local A,B;
#A:=List(Em,j->LargeMatrix(j,hom2));
221
#B:=List(irr,i->List(Em,j->j^i));
return List(irr,i->Sum(List(Em,j->j^i*LargeMatrix(j,hom2)))/Size(Em)*i[1]);
end;
PRJE:=ProjoprE(Irr(R),Elements(R),hom2);
BAS:=List(PRJE,i->BaseMat(TransposedMat(i)));
BASSIZE:=List(BAS,Size);
CROT:=List(Elements(R),i->Trace(ListToMat(ListPerm(i,60))));
CIRR:=List(Irr(R),r->List(Elements(R),i->i^r));
# In the trace representation the basis set composed from column vectors
# of the projector may include the multiplicated spaces of
# the representation. The basis vectors might be transmuted
# with each other transitively, not confined in the one of
# the component of the multiplicated irreducible spaces.
# The basis set of the decomposion into the irreduciple
# representations and the conjugacy classes are prepared.
cng:=ConjugacyClasses(R);;
rcng:=List(cng,Representative);;
rcng2:=List(rcng,j->LargeMatrix(j,hom2));;
charact:=function(g,B)
222
return Trace(B*g*TransposedMat(B)*(B*TransposedMat(B))^-1);
end;
gopr:=function(O,v,i,hom)
local irr;
irr:=Irr(O);
return Sum(List(Elements(O),j->j^irr[i]*LargeMatrix(j,hom)*v));
end;
Cutout:=function(BAS,PROJ,R,hom,NMULTI,REPSIZ)
#
# BAS
: THE VECTOR SPACE INCLUDING THE COPIES OF IRREDUCIBLE SPACES
# PROJ
: THE PROJECTORS
# R
: THE GROUP IN THE PROBLEM.
# hom
: THE HOMOMORPHISM FROM <R> TO THE MATRIX GROUP.
# NMULTI : THE SIZE OF THE SUBGROUP TO FORM THE COSET.
# REPSIZ : THE SIZE OF THE IRREDUCIBLE REPRESENTATION.
#
local C,D,BASN,EA,O,l,P,p,B,OV,SI,AS,OREP,ww,OT,ith,ifind;
C:=[];
D:=[];
BASN:=BAS;
EA:=List(Elements(R),j->LargeMatrix(j,hom));;
#
# The subgroups with the order of "NMULTI" are extracted.
#
223
AS:=AllSubgroups(R);
while (BASN<>[])
do
Print("\n",Size(BASN),"\n");
O:=Filtered(AS,g->Order(g)=NMULTI);;
Print(O);
#
# The subgroups which generate the vector space of
# the correct dimension of the irreducible representation
# are chosen.
#
if (O=[]) then
O:=[Group(One(R))];
fi;
l:=List([1..Size(O)],i->0);
ifind:=0;
ith:=1;
while (ifind=0 and ith<=Size(O))
do
ww:=gopr(O[ith],BASN[1],1,hom);
l[ith]:=RankMat(List(EA,g->g*ww));
Print(l);
if (l[ith]=REPSIZ) then
ifind:=1;
OT:=O[ith];
fi;
ith:=ith+1;
224
od;
if (ifind=0) then
Print("CHANGE THE SUPPOSITION!\n");
return 0;
fi;
#
#
By means of the chosen subgroup, the irreducible
#
vector space "B" is generated.
#
It is stored in the vector space "D".
#
B:=BaseMat(List(EA,g->g*gopr(OT,BASN[1],1,hom)));;
Append(C,[B]);
Append(D,B);
#
# The orthogonal space "OV" to the vector space "D" is prepared.
#
OV:=BaseOrthogonalSpaceMat(D);
#
# The vector space "BASN" to be processed is renewed now,
# by the projection "PROJ" on the vectors in "OV".
#
BASN:=BaseMat(List(OV,o->PROJ*o));
#
# Instead, it is possible that "BASN" is renewed
# as the intersection between "BASN" and "OV".
#
#
SI:=SumIntersectionMat(BASN,OV);
225
#
BASN:=SI[2];
#
od;
#
# Return C, the separated subspaces.
#
return C;
end;
H1:=Cutout(BAS[9],PRJE[9],R,hom2,5,5);
#
# TO MAKE THE MATRIX NEEDED IN THE COMPUTATION OF MATRIX CONJUGATION
# EXPLAINED IN THE ARTICLE.
#
ConjugMatrix:=function(rr)
local l,rmat,vmat,i,i1,i2,n,m;
l:=ListPerm(rr,60);
vmat:=List([1..Size(l)*3],i->List([1..Size(l)*3],j->0));
for i in [1..Size(l)]
do
i1:=i-1;i2:=l[i]-1;
for n in [1..3]
do
vmat[3*i1+n][3*i2+n]:=1;
od;
226
od;
return vmat;
end;
227
| 4math.GR
|
arXiv:1110.5865v1 [q-bio.MN] 26 Oct 2011
Cancer Networks
A general theoretical and computational framework for
understanding cancer
Eric Werner ∗
University of Oxford
Department of Physiology, Anatomy and Genetics,
and Department of Computer Science,
Le Gros Clark Building, South Parks Road, Oxford OX1 3QX
email: [email protected]
Abstract
We present a general computational theory of cancer and its developmental dynamics. The theory is based on a theory of the architecture and function of developmental control networks which guide the formation of multicellular organisms.
Cancer networks are special cases of developmental control networks. Cancer
results from transformations of normal developmental networks. Our theory generates a natural classification of all possible cancers based on their network architecture. Each cancer network has a unique topology and semantics and developmental dynamics that result in distinct clinical tumor phenotypes. We apply this
new theory with a series of proof of concept cases for all the basic cancer types.
These cases have been computationally modeled, their behavior simulated and
mathematically described using a multicellular systems biology approach. There
are fascinating correspondences between the dynamic developmental phenotype
of computationally modeled in silico cancers and natural in vivo cancers. The
theory lays the foundation for a new research paradigm for understanding and
investigating cancer. The theory of cancer networks implies that new diagnostic
methods and new treatments to cure cancer will become possible.
Key words: cancer networks, cene, cenome, developmental control networks, stem cells,
stem cell networks, cancer stem cells, stochastic stem cell networks, stochastic cancer
stem cell networks, metastases hierarchy, linear networks, exponential networks, geometric cancer networks, cell signaling, cell communication networks, cancer communication
∗
Balliol Graduate Centre, Oxford Advanced Research Foundation (http://oarf.org), Cellnomica, Inc.
(http://cellnomica.com). We gratefully acknowledge the use of Cellnomica’s Software Suite to construct the cancer and stem cell networks used to model and simulate all the multicellular processes that generated the in silico
cancers described and illustrated in this paper. ©Werner 2006-2011. All rights reserved.
Eric Werner, A general theory of cancer
i
networks, systems biology, computational biology, multiagent systems, muticellular modeling, simulation, cancer modeling, cancer simulation
Contents
1
Introduction
1.1 Plan . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 What is cancer? . . . . . . . . . . . . . . . . . . . . .
1.3 A theoretical framework for understanding cancer . .
1.4 Developmental networks and cancer . . . . . . . . . .
1.5 Implementation of developmental control networks .
1.6 Organizational information is not reducible to its parts
1.7 Information and control of multicellular development
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
4
5
5
5
6
6
6
2
Developmental Control Networks (Cenes), the Cenome, and IES
7
2.1 Self-sustaining cellular control networks versus developmental control networks 8
2.2 Pitchers, pots and catchers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
2.3 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
2.4 Simple cell division into two distinct cell types . . . . . . . . . . . . . . . . . . 10
2.5 Molecular implementation of developmental networks . . . . . . . . . . . . . . 11
3
Architecture of developmental control networks
3.1 Notation and Definitions . . . . . . . . . . . . . . .
3.2 A small terminal developmental network . . . . .
3.3 Normal Networks of type NN . . . . . . . . . . . .
3.4 Networks that produce identical cell types . . . . .
3.5 Stem cell networks . . . . . . . . . . . . . . . . . .
3.6 Orthogonal categories of developmental networks
3.7 Development and cancer . . . . . . . . . . . . . . .
3.7.1 Cell interactions and cancer . . . . . . . .
3.7.2 Self serving interactions in cancer . . . . .
.
.
.
.
.
.
.
.
.
12
12
13
13
14
15
15
16
16
17
4
Classification of cancers by their network architecture
4.1 Infinite cancer loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1 An example of simple infinite looping . . . . . . . . . . . . . . . . . .
4.2 Basic types of cancer networks . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
17
18
18
5
Linear cancers
5.1 Linear cancer network NL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 An interactive, conditional linear cancer network NIL . . . . . . . . . . . . . .
5.3 An interactive linear and stochastic exponential cancer network . . . . . . . . .
5.4 An almost invariant transformation of a liner network Network NLJ . . . . . .
5.5 Linear mixed cell network NLM: Linear mixed cell cancer with different cell
types and structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
20
21
22
23
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
25
Eric Werner, A general theory of cancer
5.6
5.7
5.8
6
7
8
9
ii
Linear single teratoma network NLTs . . . . . . . . . . . . . . . . . . . . . . .
A linear double teratoma network NLTd: . . . . . . . . . . . . . . . . . . . . .
Multi linear cancer networks NLIk of identical cells . . . . . . . . . . . . . . .
Exponential cancers
6.1 Exponential aggressive growth cancer network NX . . . . . . . .
6.2 Exponential mixed cell cancer network NXMa . . . . . . . . . .
6.2.1 Ideal treatment that would stop exponential cancers . . .
6.3 Exponential single teratoma cancer: Network NXTs . . . . . . .
6.4 Exponential double teratoma cancer network NXTd . . . . . . .
6.5 Hyper cancer networks NXH . . . . . . . . . . . . . . . . . . . .
6.6 Simple Hyper Cancer NXH2 . . . . . . . . . . . . . . . . . . . .
6.7 Identical cells with one loop back is still exponential but delayed
6.8 Invariant transformation of exponential networks . . . . . . . . .
6.9 Exponential cancer network with only one loop NXo . . . . . . .
6.10 n-Node Hyper Cancer Network NXHn . . . . . . . . . . . . . . .
6.11 An almost equivalent Hyper Cancer Network with a jump . . . .
6.12 Exponential hyper teratoma cancer NXHTs . . . . . . . . . . . .
6.13 Growth rates for Hyper Cancer Networks . . . . . . . . . . . . .
6.14 Multi-exponential cancers NiXn . . . . . . . . . . . . . . . . . . .
Geometric Cancer Networks
7.1 1st-Order Geometric Cancer Networks G1 . . . . . . . .
7.2 2nd-Order Geometric Cancer Networks G2 . . . . . . .
7.3 Meta-stem cells . . . . . . . . . . . . . . . . . . . . . . .
7.4 Geometric cancer networks NGk with k loops . . . . . .
7.5 Mathematical properties of geometric cancer networks .
Stem cell networks and cancer stem cells
8.1 First order stem cells . . . . . . . . . . . . .
8.2 Second order stem cells or meta-stem cells
8.3 Higher order stem cells . . . . . . . . . . . .
8.4 Cancer susceptibility of stem cells . . . . .
8.5 Normal stem cells and cancer stem cells . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
26
28
29
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
31
32
33
33
34
35
36
36
37
37
38
38
39
39
40
41
.
.
.
.
.
42
43
43
44
44
44
.
.
.
.
.
47
47
48
49
49
49
Stochastic differentiation in stem cell networks
9.1 Historical background: A network that generates Till’s stochastic stem cell model
9.2 The Till stochastic network extended to generate progenitor cells . . . . . . . .
9.3 A flexible stochastic network architecture with exponential and linear potential
9.4 Linear stochastic stem cell network LSSC . . . . . . . . . . . . . . . . . . . . .
9.5 1st-Order Geometric Cancer Networks with open stochastic dedifferentiation .
9.6 1st-Order Geometric Cancer Networks with closed stochastic dedifferentiation
9.7 Closed exponential stem cell network with stochastic delays . . . . . . . . . .
50
51
54
56
58
59
59
60
Eric Werner, A general theory of cancer
9.8
9.9
9.10
9.11
9.12
9.13
A 1st order geometric/exponential stochastic stem cell network with stochastic
progenitor cell dedifferentiation . . . . . . . . . . . . . . . . . . . . . . . . . . .
A broad spectrum Linear or 2nd order geometric or exponential stochastic stem
cell network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2nd Order Geometric stochastic stem cell network . . . . . . . . . . . . . . . .
Geometric stochastic stem cell network with dedifferentiation . . . . . . . . . .
Modeling deterministic stem cell networks with stochastic networks . . . . . .
Transformations of probability distributions change stochastic stem cell behavior
iii
61
62
63
64
65
65
10 Conditional cancers
65
11 Reactive communication networks
11.1 A reactive signal based differentiation network architecture with linear and exponential potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.2 A reactive signal based geometric network architecture with 1st and 2nd order
geometric potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.3 A pure reactive signal based geometric network architecture with 1st and 2nd
order geometric potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.4 A hybrid signal stochastic network architecture with exponential and linear
potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
12 Communicating cancer networks in systems of cooperating social cells
12.1 Interactive signaling mono generative networks . . . . . . . . . . . . . . . . . .
12.2 Interactive signaling mono-linear network architecture . . . . . . . . . . . . . .
12.3 Interactive signaling mono-exponential network with linear growth . . . . . .
12.3.1 Preconditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.3.2 Signal differentiation determines linear or exponential growth . . . . .
12.3.3 Social cell proliferation as systemic and not cell autonomous . . . . .
12.3.4 Signal identity changes can transform linear to exponential growth . .
12.4 Dual cancer signaling networks . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.5 A interactive signaling dual linear network architecture . . . . . . . . . . . . .
12.6 Interactive signaling network with exponential and linear subnetworks . . . . .
12.7 A doubly exponential signaling network with linear dynamics . . . . . . . . .
12.7.1 Linear development with communicating dual exponential networks .
12.8 Self-signaling cancer networks . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.8.1 A self-signaling network architecture with an exponential subnetwork
12.8.2 Two self-signaling networks with linear versus exponential subnetworks
70
71
71
72
73
73
74
74
74
76
77
77
79
79
81
81
13 Bilateral cancers
13.1 Apparent unilateral cancers generated by submerged bilateral networks . . . .
13.2 Temporally displaced bilateral tumors . . . . . . . . . . . . . . . . . . . . . . .
82
83
84
14 Vacuous and implicit non growing "cancers"
14.1 Vacuous non-cytogenic networks with cyclic differentiation . . . . . . . . . . .
84
84
67
68
69
70
Eric Werner, A general theory of cancer
iv
14.2 Vacuous cancer sub-networks skipped by the developmental controlling network 85
15 Metastases
15.1 Metastatic potential of cancer network types . . . . . . . . . . . . . . . .
15.2 Metastasis via signaling . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15.3 Moving signal-based metastases . . . . . . . . . . . . . . . . . . . . . . .
15.4 Context dependence of signal based metastasis . . . . . . . . . . . . . . .
15.5 Stochastic signal induced metastasis . . . . . . . . . . . . . . . . . . . . .
15.6 Stopping signal induction of metastases can have side effects . . . . . . .
15.7 A Hierarchy of metastases formed by geometric cancer networks . . . . .
15.7.1 Secondary and tertiary metastases generated by 3rd order cancer
works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15.7.2 Relating metastatic phenotype with geometric cancer networks .
15.7.3 Treatment options for metastases generated by geometric cancer
works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15.8 Stochastic reactivation of terminal cells in geometric networks . . . . . .
15.9 Metastasizing signal-autonomous exponential cancer networks . . . . . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
net. . .
. . .
net. . .
. . .
. . .
16 Network competence and performance in real continuous physical systems
16.1 Ideal developmental networks in real contexts . . . . . . . . . . . . . . . .
16.2 Cell physics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16.3 Invasive cancers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16.4 Tumor vasculature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16.5 Multiple parallel networks . . . . . . . . . . . . . . . . . . . . . . . . . . .
16.6 Relations between gene-based and cene-based networks . . . . . . . . . .
16.7 The addressing architecture and control . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
85
85
86
87
87
88
88
88
90
91
91
92
92
.
.
.
.
.
.
.
93
94
94
94
94
95
95
96
17 Network classification of cancers
17.1 A higher dimensional table of cancer types . . . . . . . . . . . . . . . . . . . .
17.2 Request for comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
96
96
99
18 Network theory and the underlying cause of cancer
18.1 Network theory and gene mutations . . . . . . . . . . . . . .
18.2 How mutations lead to cancer . . . . . . . . . . . . . . . . . .
18.3 Viruses and cancer . . . . . . . . . . . . . . . . . . . . . . . .
18.4 Aneuploidy and network theory . . . . . . . . . . . . . . . . .
18.5 Telomeres, aging and network theory . . . . . . . . . . . . . .
18.5.1 Aging, Cancer Networks and Regenerative Medicine
19 Curing cancer
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
100
100
100
100
101
102
103
103
20 Comparison of Traditional Theories with our Control Network Theory
104
20.1 The cell cycle and cancer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
20.1.1 Higher level developmental networks control the cell cycle . . . . . . . 104
Eric Werner, A general theory of cancer
20.2 Problems with the traditional gene-centered theory of cancer . .
20.2.1 Knudson’s two hit model . . . . . . . . . . . . . . . . . .
20.2.2 Repressor genes and cancer . . . . . . . . . . . . . . . .
20.3 How does our theory differ from the classical account of cancer?
20.4 The explanatory power of our theory and diagnostics . . . . . . .
20.5 Cancer growth rates . . . . . . . . . . . . . . . . . . . . . . . . .
v
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
104
105
106
106
107
108
21 What cancer is not
21.1 Physics-based models: Cancer is not just physics . . . . . . . . . . . . . . . . .
21.1.1 Physics-based models fail to explain the ontogeny of morphology . . .
21.1.2 Physics-based models fail to explain complexity in development . . .
21.2 Rate-based models: Cancer is not just about rates of cell division . . . . . . . .
21.3 Systemic models: Cancer is not the result of self-organization based on physical emergent properties of cells . . . . . . . . . . . . . . . . . . . . . . . . . . .
21.4 Stochastic models: Stochasticity is not a sufficient explanation of cancer dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21.4.1 Stochasticity, competence and performance in developmental networks
21.5 Gene-based models: Genes alone are not the cause of cancer . . . . . . . . . .
21.5.1 Gene-centered view of development and cancer is a delusion . . . . .
21.5.2 Cancer genes: Cancer networks, not cancer genes, cause cancer . . . .
21.5.3 Cell cycle models: An apparently uncontrolled cell cycle is side-effect
and not the original cause of cancer . . . . . . . . . . . . . . . . . . . .
21.6 Agent-based models: Agents controlled the rules of local interaction are not
sufficient to explain cancer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21.7 Hybrid models: Cancer is not explained by hybrid rate-based, genetic models
21.8 Evolutionary models: Cancer is not just an evolutionary process . . . . . . . .
21.9 Development as complex social physical system . . . . . . . . . . . . . . . . .
109
109
110
110
110
111
112
113
113
113
113
114
114
114
114
115
22 Conclusion
116
22.1 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
22.2 Open Questions and New Research Problems . . . . . . . . . . . . . . . . . . . 118
22.3 Beyond Cancer Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
23 Materials and methods
119
24 Request for comments
119
References
120
Eric Werner, A general theory of cancer
1
1
Introduction
We present a unified computational theory of cancer based on a theory of developmental control
networks that, starting from a single cell, control the ontogenesis of multicellular organisms.
We investigate the space or range of all possible cancer networks. We model and simulate the
major types of cancer networks. We illustrate and discuss the result of simulations of instances
of the major cancer network categories. We relate the network types to their dynamic, developmental tumor phenotype. The relevance to diagnosis, treatment and cure for the different
cancer classes is indicated. The goal is to present a new paradigm for cancer research that
gives a deeper understanding of all cancers and will ultimately lead to a cure for cancer.
While the research on cancer is enormous in scope and financing, a cure for all cancers is
still nowhere in sight. The nature of cancer is still elusive. There is no real understanding of
how cancer works, of its etiology, ontogeny and developmental dynamics. Models of cancer
often display just so many pictures giving no details. Up to now, mathematical models have
been very limited in scope and offer limited insight into the nature of cancer. Furthermore, the
models are piecemeal, varying with each cancer. There seem to be as many molecular models
of cancer as there are types of cancer.
At present there is no unifying theory of cancer that relates the many different types of cancer.
At fault is the current paradigm about the nature of cancer. It is a gene-centered theory where
so called cancer genes are responsible for cancer.
The dominant paradigm for understanding cancer is that cancer is caused by mutated, misbehaving genes. When cancer genes are muted, and there is a failure of cell death, it leads to
uncontrolled cell growth. Cancer is seen as progressive, where multiple mutations are necessary to convert a normal cell into a cancerous cell. What this account does not explain is how
cancer is controlled. It does not explain how the mutation in a gene results in the development
of a particular tumor. It does not explain the way cancer cells and tumors develop, differentiate
and de-differentiate.
We propose instead that pathological developmental control networks cause cancer. Therefore,
we further propose that to understand cancer fully, we must understand the architecture of
control networks that govern cancer development. These genomic and epigenomic control
networks are interpreted and executed by the cell. From this perspective, cancer is a special
case of multicellular developmental processes. The following table Table 1 compares and sums
up the basic differences between the gene-centered theory of cancer and the control network
theory of cancer.
Eric Werner, A general theory of cancer
2
Comparing Traditional Theories and Our Network Theory of Cancer
Traditional Theories of Cancer
Control Network Theory of Cancer
Cause
Mutations in genes cause cancer
Mutations in developmental networks
cause cancer
Process
Cancer is uncontrolled growth
Cancer is highly regulated process, controlled by developmental networks
Likelihood
Multiple mutations make cancers more
likely
Multiple mutations in specific networks
make cancers more likely
Severity
Multiple mutations may make cancers
more severe
Multiple mutations in specific networks
determine the architecture and metastatic
potential of cancer networks, and, thereby,
the potential severity of cancers
...
Networks, their topology and locality are
key to understanding cancer
Not defined or confused with cell phenotype
Control states linked in networks specify
development and differ from phenotype
Based on cell phenotype not theory
Based on network architecture and topology as related to cell phenotype
Gene centred classification: Classification is based on which gene is mutated
Networked centred classification: Classification is based on which area is mutated
and how that transforms the network
Gene expression / protein levels
Cell phenotype is only indirectly related to
network activation. Protein and RNA profiles are supportive evidence for network
activation states
Networks
Control States
Classification
Dynamics
...
A systematic relationship exists between
cancer dynamics and genome network
architecture
Metastases
Caused by cell invasive properties.
Metastatic types and dynamics not in the
conceptual framework
Different cancer networks have different
metastatic potential, dynamics and distinct
hierarchies of metastatic tumors
Diagnostics
Diagnostics is based on cell and tissue phenotype, gene expression levels, gene mutations
Calls for new diagnostics methods and
technologies to determine local genome
network architecture. Traditional diagnostics still apply and can point to cancer network types.
Table 1: Comparison of the gene centered and network view of cancer
While research into the molecular and genetic mechanisms underlying cancers is vast in scope,
Eric Werner, A general theory of cancer
3
its focus has been so detailed as to miss the most fundamental unifying properties underlying
all cancers. If cancer were a forest then research has been looking not even at just single
trees, but only at the bark of those trees. No matter how detailed our knowledge of the bark
of individual trees, it will give no real insight into the organization of the forest. Without a
deeper understanding of cancer, we cannot expect to adequately treat or cure cancer. Hence,
it should be no surprise that current treatment of cancer is piecemeal, arbitrary with many
unwanted side effects. It includes carcinogenic agents and procedures some of which can lead
to future cancers in the same patient. To cure cancer we need to understand cancer. We need a
comprehensive, unifying theory of cancer.
The object of this work is to present a theory of cancer that unifies all cancers showing their
commonalities and how they differ. Our theory has fundamental implications for diagnostic
methods, drug discovery and approaches to treatment. Most importantly, the theory implies
that it is in principle possible to change any cancer cell into a normal cell. This can be achieved
by inverse operators that reverse the antecedent cancerous network transformations, thereby
changing the cancer network into its normal precursor. Therefore, the theory provides the
foundation for a general strategy of how to cure cancer.
The theory presented here may be considered by some as revolutionary in that it views cancer
in a very different way from the dominant paradigm. Contrary to the current view of cancer as
uncontrolled cell growth (Hanahan [14]), we view cancer as a highly regulated developmental,
multicellular process (Werner [38, 39]). Our theory envisions cancer at higher levels of organization than the molecular level. At the same time our theory relates cancer to the cellular,
epigenetic, genomic and molecular levels.
Cancer is a multilevel phenomenon. At its core are cytogenic networks that generate and
control cell growth and proliferation.
If correct, this paradigm shift will generate whole new methodologies and approaches for understanding and treating cancer. Importantly, unlike many other models, our theory is testable,
and leads to precise predictions on how to control and cure cancers. It is a sustainable new
paradigm that offers scientists a rich new framework for posing novel questions, hypotheses, explanations, and mechanisms, as well as guiding and opening new vistas for their research.
Historical note: The theory was motivated by the discovery of cancers in computationally
simulated embryonic, multicellular systems. Starting from a single cell, with an artificial
genome, these embryo-like systems normally generated simple artificial multicellular organisms (MCOs) on the computer. The surprise was that with certain types of mutations these
multicellular systems developed what appeared to be tumors. The static and dynamic phenomenological properties of these computationally simulated in silico cancers were strikingly
similar to those exhibited by naturally occurring in vivo cancers. This led the author to search
for a common theoretical framework that would explain both the computational in silico cancers and the natural in vivo cancers. The first in silico cancer was discovered in a Pensione in
Rome, Italy in 1992 while the author was a visiting scientist at the Italian National Research
Council (CNR) in Rome, Italy in March of 1992.
Eric Werner, A general theory of cancer
1.1
4
Plan
We hypothesize that the development of embryos and, indeed, the generation of all multicellular life from a single cell is a highly controlled process using interlinked control information in
the genome in the form of developmental control networks. We call such developmental control networks cenes for control genes. The global developmental control network underlying
the ontogeny of multicellular organisms we call it the cenome. It is contained in the genome
(the cell’s DNA) but it is at a higher level of control than genes. We distinguish cenes from
genes. Protein coding genes are largely parts-genes used to generate the parts of the cell, its
structural and functional units. Multicellular development requires additional information to
that contained in genes. Cenes control genes and cenes control the cell’s actions.
We view cancer as the result of mutated or transformed normal developmental networks. After
introducing the theory of normal developmental control networks, we apply the theory to cancer. We then describe all possible cancer networks. We start with slow growing linear cancer
networks, then develop the theory of exponential cancer networks that result in extremely
fast growth. Next we look at geometric cancer networks lying in between the linear and exponential cancers. Linear cancer networks are a special instance of geometric cancer networks.
We show that geometric cancer networks have growth properties related to the coefficients of
Pascal’s Triangle and these coefficients are related to the geometric numbers. Hence, the name
geometric cancer networks.
After that we develop a theory of stem cell networks and related them to geometric cancer
networks. The theory of stem cell networks helps explain both the properties of stem cells and
the properties of cancer stem cells.
Next, we introduce stochastic cancer networks which allow for probabilistic alternative paths
in the multicellular development. We look at the properties of stochastic cancer networks for
linear, exponential and geometric networks showing the deterministic cancer networks can be
approximated by stochastic cancer networks by adjusting probabilities.
We then look at reactive communication cancer networks that react to cellular and environmental signals but do not send signals. Next we look at interactive communication cancer
networks that receive and send signals to form communicating social systems of cells. We show
that communication can both exacerbate and ameliorate the effects of a cancer network. For
instance, we show that, under certain conditions inter-cellular communication, an exponential
cancer network need not lead to exponential growth.
Next we describe particularly interesting cases of cancer that can be explained by our theory,
these include bilateral cancers such as those that occur in rare forms of breast cancer where
the same tumor forms symmetrically in both breasts. There is also the case of vacuous cancer
networks that do not result in any cancer but lie hidden in the genome until activated by some
circumstance.
We develop a theory of metastases (cancers that spread to other parts of the body) showing that
different cancer networks have distinct metastatic phenotypes. We show that geometric cancer
Eric Werner, A general theory of cancer
5
networks generate a hierarchy of types of metastases with distinct dynamic and proliferative
properties. While some cancers can form dangerous metastases, other cancer networks form
relatively harmless metastases. We explore all the major types of possible metastases.
We then discuss important factors that influence growth such as cell physics. We discuss the relationship to treatment. We relate our network theory of cancer with the previous gene centered
theories of cancer.
1.2
What is cancer?
A careful review of the scientific literature leaves the deeper aspects of this question unanswered. We know there are many types of cancers with differences in ontogeny and phenotype.
We know too that genes, so called oncogenes or cancer genes, are involved. We know too that
mutations of oncogenes can lead to cancer. In addition, we know that in the standard multiple
hit model of cancer, more than one mutation appears to be necessary for cancer to develop
(Knudson [16]). While this model indicates that genes play a causal role in the development
of cancer, it is not an explanation of the functional cause of cancer. It gives no account of
the functional, developmental dynamics of cancer. We have no explanation as to why cancer
develops as it does, why some cells in a tumor are cancerous and others not, why one cancer
is dynamically different in phenotype and ontogeny than another, why some cancers are fast
growing and others not.
1.3
A theoretical framework for understanding cancer
We propose a general theoretical framework for understanding the dynamics of development
of cancers. Our theory is surprisingly simple and at the same time has extensive explanatory
power. In our view cancer is a special case of developmental multicellular processes generally.
The better we understand the development of organisms the better we will understand cancer.
To model dynamics of cancer we need to relate the genome and cell architecture of cancerous
cells with the dynamic development and differentiation of cancerous multicellular systems. So
too, since cell signaling is inherent to some cancers, we need to account for the relationship
between cells, signals, genomes and cancer dynamics. Our theory will leave many open questions both theoretical and experimental. But that is as it should be. Any new paradigm should
propose new questions and problems with new approaches to answering and solving them.
It is hoped, therefore, that this paper will be a stimulus to new research based on this new
paradigm.
1.4
Developmental networks and cancer
We propose that pathological developmental control networks cause cancer. By a developmental control network we mean a control system of interlinked commands that changes the state
of a cell or that of its offsprings. The network topology underlying all cancers contains at least
Eric Werner, A general theory of cancer
6
one loop that leads to the original state that starts the cancer cascade. We can best illustrate the
theory by a series of examples of typical cancers or cases. For each case we provide the general
structure of the genomic control network that underlies the cancer, we model what happens as
the cancer develops, we make predictions of the expected clinical phenotype and we propose
methods for differential diagnosis that distinguish the different types.
1.5
Implementation of developmental control networks
Each of these cases of cancer networks are abstract and will require instantiation to particular
in vivo cases. The instantiation ultimately will also involve a molecular implementation of
these networks. However, the general principles presented here that govern the dynamics and
development of a general cancer type based on its network properties will apply to the special
cases as well.
For any abstract network many implementations are possible. For example, the networks may
be implemented using protein transcription factors, but the same network could also be implemented using RNA. The abstract properties of the developmental control network will be
the same regardless of its implementation. Thus nature had many different options for evolving developmental control networks. Indeed, we hypothesize that there was a switch in the
implementation of genome control necessary for multicellular life to evolve. Even after the
first primitive multicellular life evolved there may again have been new implementations of
developmental networks necessary to achieve the increasing complexity of multicellular organisms.
1.6
Organizational information is not reducible to its parts
The information in an organization such as a network is not, in general, contained in the parts
used to implement or construct it. This is because the properties of the components of the
instantiation, be it using protein transcription factors or RNA or some other molecular mechanism, do not carry the information contained in the network composed out of those components. The properties of the parts of a structure do not in general carry the information that
is contained in that organized structure. For example, the letters A, C, G, T do not carry the
information contained in a sequence of these letters. The interaction of the nucleic acids represented by A, C, G, T can generate a random sequence of such units, but the information in
the parts is not sufficient to generate any particular sequence. Organization is not, in general,
reducible to the information in its parts nor to the interaction of those parts.
1.7
Information and control of multicellular development
Essential to understanding cancer, and the development of organisms generally, is to understand how information functions to control what cells do. We must understand not only how
Eric Werner, A general theory of cancer
7
information can control a single cell, but also how it controls a dynamically developing, growing and dividing multicellular system of cells. Thus, we assume that there is control information within a system of cells that controls the behavior and development of those cells. Where
is this control information located? I have argued elsewhere that the primary source and locus
of this control information is in the genome and not primarily in the non-genomic portion of
the cell (Werner [40]). While control information is distributed throughout a cell, its primary
source is the genome. However, a great deal of this control information can be transferred to
the cell from the genome. Furthermore, signals from the cell itself, other cells and the environment interact with the genome’s control networks leading to activation of new control
networks and altered control states. In effect, the genome control architecture and the cell
control architecture cooperate in process of development.
2
Developmental Control Networks (Cenes), the Cenome, and IES
We use the words cene, developmental network, generative network, cytogenic network, or
proliferative network equivalently for a control network that contains directive, control information that when interpreted and executed by the cell induces, controls and directs one or more
cellular divisions resulting in a multicellular system. We call developmental control networks
cenes (for control genes) to distinguish them from genes that code for proteins. Cenes are more
like Mendel’s original notion of a gene as the unit of inheritance that specifies some phenotypic
property of the developing and mature multicellular organism. ’Gene’, in its contemporary use,
codes for protein parts of the cell. To emphasize their role as parts, we call such genes parts
genes. Many organism share parts genes. Many parts genes are evolutionarily conserved between distantly related species. It has turned out that the genes in humans and chimps are 99%
identical, yet the organisms are patently very different. Parts genes cannot account for the vast
differences between organisms. Just as a house and bridge may be built of the same parts yet
have very different architectures, so the parts genes of two organisms may be the same yet they
undergo different embryonic development resulting in very different phenotypes.
We hypothesize that the difference in the development and morphology of multicellular organisms lies in their developmental control networks or cenes and not mainly in their parts genes.
Cenes can be combined and linked in many ways to form larger, more complex cenes. Indeed,
we hypothesize that the rapid evolution of multicellular organisms involved the duplication,
modification and combination of cenes to produce more complex cenes. We define cenome to
be the complete developmental control network encoded in the genome of an organism. Thus,
the cenome is just another cene that consists of numerous linked developmental subnetworks
or sub-cenes. Cenes and the cenome are the bases of the evolution and ontogeny of embryos
and multicellular organisms.
How can information in the genome control the actions of the cell? The cell and the genome
form a cooperative system. Since the information in the genome is primarily passive, the cell
has to interpret and execute the control information in the genome. We call this complex
system the Interpretive Executive System or IES. The IES of the cell interprets and executes the
directives encoded in cenes. Cell communication also is handled and interpreted by the IES.
Eric Werner, A general theory of cancer
8
The IES interprets and executes both the internal and the external signals the cell receives from
its external environment, including other cells. Such signaling systems interact by way of the
IES with the genome to activate yet further control networks in the genome.
In addition, genome control is self reflexive. The genome controls itself with the cooperation
of the IES. The genome interacts by way of the IES with itself producing a network of cascades
of controlling information. Implicit, therefore, in genomes are developmental control networks
or cenes.
2.1
Self-sustaining cellular control networks versus developmental control networks
It is essential not to confuse developmental cenes (developmental control networks) with selfsustaining networks such as metabolic gene networks and other non-developmental cellular
molecular networks and pathways. The proteins produced by parts genes are not passive, but
active molecules with agent like properties. These molecular agents interact and self-assemble
to produce complex cellular organization and networks of interacting molecular pathways that
perform the essential life sustaining functions of the cell. This cellular organization including
its networks of pathways is also the basis of sophisticated cellular behavior and strategic action vis a vis the environment. Furthermore, the cell and the networks of proteins produced
by parts genes form and constitute the system (the IES) that interprets and executes developmental control networks (cenes). Thus, developmental control networks or cenes are a distinct
level of control information in addition to the self-sustaining, intra-generational cellular control
networks of the cell.
2.2
Pitchers, pots and catchers
When a parent cell divides into two daughter cells that each differentiate into a different control
state than that of the parent, something about the control state inherited from the parent cell
has to change in the daughter cells. This change we hypothesize is mediated by transcription
factors or something like them that activate different areas of DNA in the daughter cells than
were active in the parent cell. Each daughter cell gets a different activation unit from the
parent. In baseball a pitcher throws a ball to the catcher. The event of throwing the ball is the
pitch. We use the pitcher-pitch-ball-catcher metaphor from baseball to describe the process of
intergenerational control of cell states:
A potential pitcher is an area of DNA in a parent cell that when executed throws a potential
pitch or pot to a catcher or cat in the daughter cell. A catcher or cat is an area of DNA in
the daughter cell that catches the pot that was thrown by the parent cell. When the pot is
caught by a catcher in the daughter cell, the pot binds to the corresponding area of DNA in
the daughter cell. These pots are inter-generational in that they pass from one generation to
the next generation of cells. We therefore call them potential pitches or pots because they are
not active until after cell division when they are eventually caught by a catcher on the daughter
Eric Werner, A general theory of cancer
9
cell DNA. Thus, we distinguish intra-generational pitches, called jumps and which are caught
within the lifetime of a cell, from inter-generational potential pitches (pots) which are only
caught by a daughter cells after the parent cell divides (Fig. 1).
We use the ball-arrow notation A B to indicate a potential pitch or pot is thrown from a
potential pitcher at area A in the parent cell to a catcher at area B in a daughter cell. When the
potential pitch is caught by a corresponding catcher at B in a daughter cell, that daughter cell
enters control state B. When it is clear from the context, we use the term pot to refer either to
the potential pitcher or to the potential pitch itself.
c
A
B
C
D
b
Fig. 1: Network NDiv: A basic cell division network. The developmental control network of
the parent cell type A divides into two terminal daughter cells B and C. Two potential pitches
pot1 with address c activates control state C in the first daughter cell while in parallel pot2 with
address b activates control state B in the second daughter cell.
When two daughter cells enter distinct control states (and/or phenotypic differentiation states
mediated by control states), the parent cell must have thrown or sent each daughter cell a
different pot. By convention, we refer to theses possibly distinct pots as pot1 and pot2, where
it is specified prior to cell division which daughter cell receives pot1 and which daughter cell
receives pot2. Two daughter cells enter identical control states when they receive identical
pots, i.e., when pot1 = pot2.
2.3
Notation
By convention, we use upper case letters A, B, C, . . . for DNA potentially active control areas,
lower case letters a, b, c, . . . for their respective addresses. a∧ is a potential pitcher that throws a
potential pitch a○ with address a. a∨ is a corresponding catcher with matching address a. Thus,
a○ is the pot thrown by pitcher a∧ in the parent cell and caught by a∨ in the daughter cell. We
say a catcher a∨ is loaded if the pot has been caught, in symbols a ∨. A catcher is blocked if it
−
cannot be loaded, in symbols a ∨.
In the graphical representation of developmental control networks (cenes), we usually leave
out the details of the addressing structure of pitchers, pots and catchers since these are evident
in the visual representation of the network links. We show only the links, such as A B (or
Eric Werner, A general theory of cancer
10
b
at most A B), to indicate that an area A throws a potential pitch to area B. In more detail,
b
A B means that area A has a pot pitcher b∧ that when activated throws a potential pitch b○
with address b to an area B with catcher b∨ with matching address b. By convention, pot1 links
are usually drawn on top of the genome and pot2 on the bottom (see Fig. 1).
2.4
Simple cell division into two distinct cell types
If we need more detail, c∧1 is a pot1 pitcher that throws a potential pot1 pitch c○1. b∧2 is a pot2
pitcher that throws a pot2 pitch b∧2. Combining pot1 and pot2 into synchronized throws to two
∧
daughter cells, an area Abc ∧12 that throws a pot1 pitch c○1 with address c to the first daughter cell
C with catcher c∨ and a pot2 pitch b○2 with address b to the second daughter cell B with catcher
b∨. Once the catchers are loaded, this activates the first daughter cell to state C and the second
daughter cell to state B.
○1
c
Φ(σ, p)a∨Abc ∧12
b∨B
∧
b
c∨C
D
∧2
Fig. 2: Network NDiv: Cell division network. The developmental control network of the
parent cell type A divides into two terminal daughter cells B and C. The ball-arrow notation
links a parent cell with one of its daughter cells. It indicates that when the parent cell’s
execution of the network reaches the control area on DNA that is next to the ball at beginning of
the arrow then the parent cell divides and the linked resulting daughter cell activates the control
area pointed to by the end of arrow. In the above network, the cell in control state A divides
into daughter cells that differentiate into control states B and C. C is terminal. The notation
∧
Acb∧12 means that the area A encodes pot1 with address c and pot2 with address b. The notation
b∨B means that B encodes a catcher with address b that will catch the pitch b○1 thrown by pot2
( b∧2 ) of A. The prefix Φ(σ, p) before the catcher a∨ denotes further possible pre-activation
conditions such as cell signaling, environmental, and other conditions (more on this below).
Thus, each pot pitcher α∧ and its corresponding thrown pot α○ contains an address α that is
matched by its corresponding catcher α∨. We leave open how these addressing systems are implemented in DNA and we leave open how the intra and inter-generational transfer mechanisms
are molecularly implemented.
An illustration of one possible scenario of how the pots might be transferred to their respective
daughter cells. While it may be helpful in understanding the process, this is just one of several
possible implementations of the cell control inheritance. Here the pots might attach to the
Eric Werner, A general theory of cancer
11
opposite Centrioles or the opposite cell walls. In another scenario, the pots might directly
attach to the chromosomes to be carried over to their intended daughter cells. Yet another
scenario would have the replicated DNA itself marked (transformed by a viral like pot-vector)
prior to division to be carried over and become active. The point here is that any mechanism
must in some way insure that different control pointers or pots are transferred to different
daughter cells if they are to differentiate and develop differently.
b○1
b
b○1
B
b
c
A
c
c○2
C
○2
c
Fig. 3: A cell A divides into two daughter cells B and C. This developmental control network
regulates the division of a cell A into two daughter cells B and C. At control state A a two
pitchers b∧ and c∧ throws two pot pitches pot1 = b○1 and pot2 = c○2, respectively. Prior to
division these two pot types, pot1 and pot2, are separated so that they will be transferred in
opposite directions along the axis of division to different daughter cells, B and C. After division
these pots are transferred and caught by their catchers with matching addresses. pot1 = b○1 is
caught by the catcher b∨ and pot2 = c○2 is caught by c∨. Once the catchers are loaded the
daughter cells enters new control states, B and C. We leave open how prior and post cell
division separation and transfer systems are implemented in the cell.
2.5
Molecular implementation of developmental networks
We use this special terminology of pots and their catchers for several reasons: First, we want
to avoid confusion with normal transcription factors which are often intra-generational jumps.
Second, we purposely abstract away from the details of the biological mechanisms that control
cell division and differentiation. Third, as a consequence, we leave open how pots are biologically instantiated or molecularly implemented. Fourth, the theory is a guide to developing and
designing experiments to discover the actual biological implementations in multicellular animals and plants. Fifth, in particular, pots may or may not, in some or all cases, be implemented
molecularly by protein transcription factor mechanisms.
Thus, we want to leave the specific molecular implementation open because pots may be implemented by RNA in combination with protein transcription factors or other mechanisms as
Eric Werner, A general theory of cancer
12
yet undiscovered. Ultimately, we want to discover the syntax, semantics and pragmatics of the
hidden source code of life underlying and controlling multicellular development. By abstracting from the details of molecular implementation, we can focus on the universal properties
of developing multicellular systems, and the corresponding universal properties of the control
code and its meaning.
3
Architecture of developmental control networks
By definition, development will refer throughout the text as the generation of a multicellular
system from one or more founder cells. We hypothesize that cancer is a form of development.
A developmental control network or cene controls the development and ontogeny a multicellular system or organism. Developmental control networks have an architecture of interlinked
control areas in genomes that control the states of cells. The architecture of cenes is based on
the architecture of their addressing systems (Werner [45]). Developmental networks (cenes)
can be interlinked to form larger developmental networks (cenes). The nodes of developmental
networks are built out of and linked to further more basic networks that control cell processes.
Thus we distinguish cell differentiation networks that control cell type from cenes which control multicellular development. Developmental networks or cenes are executed in parallel in a
vast multicellular developing system.
3.1
Notation and Definitions
A cene is a developmental control network. Cene stands for control gene because they are
more like Mendel’s original concept than protein coding genes (Werner [43]). The cenome
is the global developmental control network that controls the multicellular development, i.e.,
the embryogenesis and ontogeny of an multicellular organism. An MCO is a multicellular
organism or system. An MCS is a multicellular system. Every MCO is an MCS. We propose
that when an MCS develops from a single cell it does so by means of a developmental control
network or cene.
We use the phrases cenes, developmental control networks and developmental networks interchangeably. However, because of the lack of familiarity with the term ’cene’, we will often
refer to cenes as developmental control networks, or simply developmental networks. There is
a danger of confusing developmental networks with metabolic networks, cell cycle networks,
or other cellular networks that control cell differentiation. Cell cycle networks and cell differentiation networks are subsumed under (meta-controlled by) developmental control networks
(cenes).
Development of a multicellular system is mosaic if that development is independent of outside
communication. It develops the same way, in an inflexible context independent way. A network is mosaic if the development it generates is mosaic. Development that is not mosaic, but
Eric Werner, A general theory of cancer
13
dependent on cell signaling or environmental conditions is called regulatory development. A
network is regulatory if the development it controls is regulatory. Most networks will not be
purely mosaic but instead are regulatory containing regulatory elements that respond conditionally to stimuli and signals. Development is terminal if cell proliferation eventually stops under
all (normal) conditions. A network is terminal if the development it generates is terminal. This
occurs if all possible developmental paths in the network terminate in a terminal node that halts
with no further network controlled cell division. Networks that contain antecedent activation
conditions on nodes are called conditional networks. Regulatory networks are conditional on
signal inputs. A developmental path in a network is nonterminal if it contains a cycle or loop.
A network is conditionally nonterminal if all nonterminal paths have antecedent conditions for
their activation.
3.2
A small terminal developmental network
Some normal developmental networks are terminal where all developmental paths result in
terminal cells that no longer divide as seen in Fig. 4.
b
A
d
B
f
C
c
D
e
E
F
e
Fig. 4: NBMC: A normal bounded multicellular network. In the above network, the cell
in control state A divides into daughter cells that differentiate into control states B and C. C
is terminal. B divides into D and E. D then divides into E and F both of which are terminal
control states. The phenotype of the multicellular system produced by this network starting at
A consists of five terminal cells: One each of type C, D, F and two of type E.
3.3
Normal Networks of type NN
A normal network NN (Fig. 5) is a noncancerous developmental genomic network with possible signal inputs. It is a non-cancerous network. It may contain loops that do housekeeping
chores that are continually active; however these do not generate abnormal cells. It may contain
conditional loops that induce some rounds of cellular development but only if there are signals
that indicate some special situation such as a wound or other circumstance. We will see that any
normal developmental sub-network can become cancerous given the right mutations.
In exponential cell division a cell divides into two daughter cells that both divide. Normal
developmental processes may be exponential for a time in that the daughter cells of a cell also
Eric Werner, A general theory of cancer
14
generate two daughter cells that divide. However, as we will see, geometric networks show
that this does not always lead to exponential growth.
ny
A
Nx
Ny
D
nx
Fig. 5: Network NDev: Normal Developmental Network A divides into a cell controlled by
a bounded network N x and a cell controlled by network Ny . The daughter cell’s developmental
network N x may or may not generate further multicellular development and structure with
possibly multiple cell types.
3.4
Networks that produce identical cell types
Some networks produce identical cell types. In this case, the daughter cell networks may or
may not generate further multicellular development and structure with possibly multiple cell
types (Figures 6 and 7).
nx
A
Nx
Ny
D
nx
Fig. 6: Network NI1 : Normal Identical Cell Developmental Network A divides into two
cells both controlled by a bounded network N x . Both daughter cell networks may or may not
generate further multicellular development and structure with possibly multiple cell types.
Eric Werner, A general theory of cancer
a1
A0
an
a2
A1
a1
15
A2
a2
An − 1
An
an
Fig. 7: Network NIn produces 2n identical cells in n steps: This network generates n identical
divisions to produce 2n identical daughter from one founder cell.
3.5
Stem cell networks
Another basic class of developmental control networks are stem cell networks. Stem cell networks are nonterminal. Some stem cell networks are unconditionally nonterminal while most
are conditionally nonterminal. We will discuss these in detail when we discuss cancer stem
cell networks (see Sec. 8). Cancer stem cell networks and normal stem cell networks share
many of their topological and dynamic properties. Crucially, normal stem cell networks have
preconditions for their activation as well as normal cell differentiation patterns. We will discuss
stem cell networks in greater detail in Sec. 8.
3.6
Orthogonal categories of developmental networks
While the networks determine ideal growth rate they also determine cell differentiation and
specify organism phenotype in space and time. Differentiation and organization architectures
are orthogonal categories of developmental networks, forming extra dimensions in addition to
the network proliferative architecture. There are several interrelated properties that can be used
to classify developmental networks:
1. Network Architecture: The network architecture as given by the global interlinking of
the developmental control network is one dimension that influences phenotype of developing multicellular systems and organisms. This determines its proliferative capacity, its
developmental dynamics and its complexity.
2. Differentiation: Cell differentiation into various cell types is controlled by differentiation control networks that are subsumed under the more global developmental control
networks. Developmental networks can be subclassed according to the kind and number
of cell types they generate and, thereby, implicitly according to the cell differentiation
sub-networks they activate.
3. Conditionality: In addition to an activating site or catcher, a network may have further
activation conditions that must hold before some of its nodes are activated and executed.
Eric Werner, A general theory of cancer
16
Networks that contain activating conditions are called conditional networks.
4. Cell interactivity: Cell interactivity is closely related to conditionality of networks.
(a) Cell communication: Cell to cell signaling protocols and environmental conditions can activate developmental networks including cancer networks.
(b) Cell physical interactions: Cell cohesion and cohesion cell strategies can give a
cell the capacity to invade other tissue.
5. Regulatory versus mosaic development: How autonomous is the developmental network from cell communication?
6. Stochasticity: Is the network deterministic or does it have stochastic elements?
7. Complexity: How complex is the network and how complex is the organism it generates? What is the relationship between the complexity of the cell and the complexity of
the genome?
8. Physics: How dependent is the generated multicellular dynamics and development dependent on the physics of interacting cells?
9. Multi-Cellular System (MCS) spatial and temporal organization:
Another orthogonal category of a network is the MCS-phenotypic organization (the
space-time development and morphology) of a multicellular cancer generated by the
network.
In general, nonterminal and terminal, normal and stem cell, conditional and non-conditional,
deterministic and stochastic, communicative and non-communicative networks will be mixed
together in one large eclectic, evolutionary network.
3.7
Development and cancer
The space-time developmental phenotype of a multicellular system such as cancer is reflected
in the cancer network that controls it. On this view, cancers are pathological multicellular
systems controlled by cancer networks that, apart from their deviant architecture, are similar
to the networks that control the development of healthy multicellular systems. In other words,
cancer is a type of multicellular development similar to normal embryonic development.
3.7.1
Cell interactions and cancer
The effect of cancer control networks will be more or less severe to the extent that cellular interaction can lead to novel outbreaks of cancers spawned by such interactions. For simplicity,
we assume that the cancers do not have pathogenic interactions with other cell types unless
specifically stated in the example case. In other words, we assume that cancer is monoclonal,
Eric Werner, A general theory of cancer
17
i.e., arising from a single cell. This is usually taken to imply what is actually an additional assumption, namely, that cancer cells cannot influence non-cancerous cells to become cancerous.
However, there are situations of cell communication in which this latter implicit assumption
need not hold.
3.7.2
Self serving interactions in cancer
We also assume that a cancer can engage in self serving interactions in order to gain resources
for growth from helper cell systems such as blood vessels. While these are necessary conditions
for the growth of cancers they are tertiary in terms of its conceptual and causal essence even
if many of the adverse side effects and spread of cancer in the body depend on such tertiary
conditions.
4
Classification of cancers by their network architecture
We will now go into more detail about the nature of the networks that lie at the origin of
all cancers. Cancer networks can be classified into basic types whose network architecture
and topology determines whether the cancers are linear, geometric, or exponential in their
proliferative dynamics, as well as determining their general clinical phenotype. Each general
class of cancer can be further divided into vast variety of subtypes. Each cancer subtype may
differ diagnostically, clinically and dynamically according to the particular architecture of its
regulatory cancer network.
Each of these subtypes of cancer has been modeled and simulated with software. The mathematical equations that describe particular ideal growth rates of cancer types have been computationally verified. The computationally predicted, static and dynamic phenotypes of these
artificial in silico cancers show strong correlations with the clinical, static and dynamic phenotypes of in vivo, naturally occurring cancers.
4.1
Infinite cancer loops
Fundamental to all cancer is simple repetition of action. A cell divides and results in at least
one daughter cell that is also a cancer cell. If both daughter cells were normal then the so
called cancer cell would proliferate normally being finitely bounded and all its offspring would
be non-cancerous. Hence, no cancer. Thus, cancer cells must generate cancer cells if the cancer
is to continue. What must the state of the cancer cell be like to generate another cell that is also
cancerous?
Eric Werner, A general theory of cancer
4.1.1
18
An example of simple infinite looping
You are told to follow the following instructions in their proper order starting at 1 then 2 then
3, etc. And don’t stop until you are told to do so by the instructions.
1. Open the door.
2. Close the door.
3. Go to instruction 1 and follow it.
Clearly, if you follow the instructions, you will open the door; then you will close the door.
Then you will read instruction 3 and jump to instruction 1, which tells you to open the door,
which you do. Next, you read instruction 2, and close the door; next you read instruction 3,
and so on. Hence, if you follow the instructions you are in an endless loop and you never
stop.
This sort of program is quite common in computer science. Often it appears in introductory
courses when students write a program that "hangs", because it never stops precisely because
the student has unknowingly written an infinite loop into the program. However, such loops
also have their uses in control programs like operating systems which use endless loops to
continuously operate the system. The computer has the ability for outside input to stop these
loops, for which students are quite thankful. In the case of cancer, each proliferative loop
generates new cells and thereby causing endless growth. Here too there are functional proliferative loops where some cells, such as stem cells, continuously regenerate cells, e.g., skin
cells, hair or fingernails. Hence, what is defined as a cancer will depend on not just on the
existence of proliferative loops, but on the nature of the proliferative network and its functional
context.
4.2
Basic types of cancer networks
We hypothesize that all developmental multicellular processes are controlled by developmental
control networks. We distinguish normal networks from cancerous control networks. Keep in
mind, there are however cases where the control network may not be the distinguishing factor,
when, for example, a stem cell gains the capacity to invade other tissues. These are lower level
control strategies of the cell that interact with the developmental control networks. More on
this later.
According to our network theory, cancer is a regulated developmental multicellular process
resulting from the transformation of a normal developmental control network. All cancers can
be grouped into three broad classes according to the architecture of their control networks and
their corresponding developmental dynamics.
1. Linear cancers NL (see Sec. 5)
Linear cancer networks generate cancers that proliferate linearly in time. Linear cancers
include 1st order stem cells Fig. 34.
Eric Werner, A general theory of cancer
19
2. Exponential cancers NX (see Sec. 6)
Exponential cancer networks generate cancers that proliferate exponentially as a function
of time.
3. Geometric cancers NG (see Sec. 7)
A category of cancer networks lying between linear and exponential cancer networks are
the geometric cancer networks whose ideal competence is to grow at a rate in accordance
with one of the geometric numbers. We will see that they are related to coefficients of
the binomial theorem and Pascal’s Triangle (see subsection 7.5). Geometric networks
are related to the concept of a meta-stem cell, i.e., a stem cell that produces stem cells.
4. Cancer stem cell networks (see Sec. 8)
Cancer stem cell networks are a subset of geometric cancer networks. They are closely
related to normal stem cell networks. We will describe:
(a) Deterministic stem cell networks follow a well defined, determined developmental path (see subsection 8.1).
(b) Stochastic stem cell networks (Sec. 9) follow developmental paths probabilistically. Depending on their network topology and their developmental probability
distribution, stochastic stem cell networks can emulate many of the other cancer
networks. Thus, some stochastic networks can jump out of their category (e.g., linear or geometric) into another cancer class (e.g., exponential) depending on their
developmental topology and probability distribution.
5. Reactive signaling networks react to signals that determine their topology (see Sec. 11).
6. Communicating cancer networks (see Sec. 12)
Many cancers involve communication with other cells and the environment. Each of
the above cancer networks can combined with or implemented by cell signaling communication protocols resulting in more flexible multicellular development generally and
specifically, more flexible cancer networks. Indeed, many cancers result from mutational
transformations of the cell signaling protocols.
7. Complex cancer networks in context of developmental networks
All these cancer networks are part of developmental control networks (cenes), that can
be combined with other normal and cancer developmental networks (resulting in more
complex cenes) to produce complex cancer dynamics together with complex cellular
and multicellular, organizational, structural phenotypes. Moreover, depending on their
position in the developmental network hierarchy, and hence, their position and activation
in the spatial-temporal development of the organism, the very same cancer network can
exhibit very different cellular and multicellular dynamics and phenotypes (e.g., bilateral
cancers, complex teratomas, fetus in fetu).
Eric Werner, A general theory of cancer
5
20
Linear cancers
The most basic type of cancer network generates linear cancers, e.g., basal cell carcinoma or
grade I glioma of the brain. In a linear cancer the number of cells increases as a linear function
of the number of given cells at any given time.
General Phenotype: Clinically linear cancers are slow growing when compared to geometric
or exponential cancers. With linear cancers no additional cancer cells are generated by the
cancer network. In deterministic, non-stochastic linear cancer networks, number of cancer cells
in the tumor stays constant. Each execution of the cancer loop will result in the generation of
new cells, but they are not cancerous in the sense that the cancer network is not active.
While a linear cancer cell does not produce active cancer cells (other than one daughter cell
with the same control state as itself), linear cancer cells produce potential cancer cells that have
inherited the linear cancer network. If, by some means, this network is activated in the passive
daughter cell then we have another linear cancer cell.
Linear cancers have network architectures similar to stem cell networks (see below Fig. 34).
Linear control networks are also common in non-cancerous tissue, examples include warts,
finger nails, and hair growth.
5.1
Linear cancer network NL
a
A
B
b
Fig. 8: Network NL: Linear cancer network. The developmental control network of the
parent cell type A self-loops giving a daughter cell of the same cell type A as the parent. The
other daughter cell differentiates to type B. The network generates a slow growing cancer
containing only one cancer cell A. All the generated cells contain the cancer network. The
difference is that in cells of type B the cancer network that generated them is not active.
The simplest cancer network NL generates only one cancer cell. NL controls a cancer that
exhibits linear growth and is possibly benign. The network NL controls cell division where a
cell of type A divides into a cell of type B and a cell of type A.
Eric Werner, A general theory of cancer
21
Phenotype: This network will result in a tumor whose outward phenotype consists of two
main cell types, A and B. There will be only one single cell type A that is cancerous. The
rest will be all cells of type B or, if B is non-terminal, cells derived from type B. The growth
rate will depend on the rate of cell division. The complexity (maturity) of the cell type A
may influence how quickly it can divide. More complex cells may take longer to duplicate all
their parts making division more complex. However, whether the cells simple or complex, the
growth rate is linear with no increase in tumor cells. Irrespective of growth rate, such tumors
are benign in the sense that their source is one or more single cells controlled by a linear cancer
network while in all the other cells in the tumor are non-cancerous because the linear cancer
network is inactive.
For example, if the given number of cells is g at time t it will tend to increase by no more than
some constant number c to give g + c after the next round through the cancer loop. c represents
the number of active linear cancer cells in the tumor. The units of time are not necessarily
constant units such as seconds, hours, days or months. But they are determined by how long it
takes for each of the cancer cells to go through its developmental cancer cycle. Thus, after n
rounds of the cancer loop the number of cells will be no more than:
Cells(n) = n ∗ c + g
(5.1)
Treatment: Since such tumors have only a few cancer cells that are actually proliferating, the
prime objective is to remove these active cancer cells by some method. However, the passive
cells generated by the active linear cancer cells still have the mutation, i.e., they have inherited
the cancer network even if it is inactive. Thus, if the network is a stochastic cancer stem cell
network and there is a nonzero probability of reactivation then these cancer nets could become
active again forming new cancer stem cells (e.g., see the section on linear stochastic caner
stem cells Fig. 41). Hence, depending on the probability of spontaneous reactivation, removal
of such inactive tumor cells might still be indicated. Radiation therapy would be relatively
contraindicated because of the danger of increasing the likelihood of reactivation of existing
passive cancer cells, transforming linear cancer stem cells into exponential cancer stem cells,
or generating new cancer cells.
The network NL results in a benign cancer to the extent that the cells of type A are minimally
interactive with other cells leading to no cancerous induction of other cells. Thus, cell interactivity is an independent component that is relevant in assessing phenotypes of the dynamics of
cancer. Just one network mutation can convert a linear cancer into an aggressive exponential
cancer (see Fig. 18).
5.2
An interactive, conditional linear cancer network NIL
An interactive or conditional linear cancer network is just like the linear cancer network except
that it must be activated by some internal or external signal σ with some property Φ. In this
case a passive cell B generated by the network, can revert into a cancer cell of type A given
the right signaling context. This may be the context of other cells. If, for example, the cell B
Eric Werner, A general theory of cancer
22
invades other tissue with the appropriate intercellular signaling protocol then B may become
cancerous in the new multicellular context.
σ
c
Φ(σ)B
A
C
b
Fig. 9: Network NIL: Interactive linear cancer network. The developmental control network of the parent cell type A divides into daughter cells B and C. Given a context that sends
a signal σ with property Φ, then the network generates a slow growing linear cancer. If the
tumor is generated from stem cell A then the tumor contains only one cancer stem cell A which
becomes the signal dependent cancer cell B after division. The rest are terminal cells of type
C. Thus, cells of type B have the potential to be cancerous given a signal σ with property Φ.
Phenotype: The multicellular phenotype of this cancer network is just like that of the linear
network NL except that it contains three cell types A, B and C where A and B transform into
each other. Cell A differentiates to B after division. Cell B transforms/dedifferentiates into A
given a signal σ with property Φ. Thus, in different contexts B cells may revert to type A cancer
cells. This may appear to be stochastic, but is actually determined by some environmental or
cellular signal.
While we will not describe most of them explicitly, all the cancer networks described below
can have interactive variants.
5.3
An interactive linear and stochastic exponential cancer network
An interactive linear cancer stochastic exponential network behaves like a linear cancer but
turns exponential, with propability p, in a signaling context. Given an internal or external
signal σ with some property Φ, a passive cell B generated by the network from a cell of type A,
can with probability p revert into that parent cancer cell A. Thus, the cellular communicative
context can make a passive cell into a cancer cell. If, for example, the cell B invades other
tissue with the appropriate intercellular signaling protocol then B may become cancerous in
the new multicellular context. In this network the proliferation turns exponential in a signaling
context, but remains linear when no signaling occurs.
Eric Werner, A general theory of cancer
23
σ
a
A
Φ(σ)B
J
p
b
p-1
Fig. 10: Network NISXL: Interactive linear cancer network. The developmental control
network of the parent cell type A self-loops giving a daughter cell of the same cell type A as
the parent. The other daughter cell differentiates to type B. The cell type B has receptors for
a signal σ. B jumps to state J on receipt of a signal σ with property Φ. Depending on the
probability p, a cell in state J dedifferentiates either to its grandparent A or its parent B. If
it dedifferentiates to A it makes a step in the direction of exponential growth. Otherwise, it
just cycles back to B. Thus, cells of type B can become cancerous with probability p given a
signal σ with property Φ. Interactive cancers are more dangerous since the passive B cells can
spontaneously become cancerous in some new multicellular or environmental context.
Phenotype: The multicellular phenotype of this interactive stochastic cancer network is just
like that of a linear cancer network containing only two cell types A and B. However, in different cellular and environmental signaling contexts some of the B cells may stochastically revert
to type A cancer cells. Unlike the deterministic communication protocol in the interactive linear
network this only activates A with some probability p. Hence, even if the right communication
context is present, the lower the probability the lower the chance of generating more cancer
cells. As the probability p approaches one and there an ample source of signals σ this network
approximates an exponential network. As p approaches zero the signals σ have no effect and
there is no dedifferentiation from B to A making the network linear. When p = 1/2 and there
is sufficient signaling, half the B cells will dedifferentiate to A cells. If B is in a signal σ free
context, the tumor behaves linearly. Once the B cells are in a σ signaling context, the tumor
behaves exponentially with probability p.
Note, when B loops back to A by way of J, the two loops do not form a second order geometric
network, because both loops go back to the same parent cell A.
5.4
An almost invariant transformation of a liner network Network NLJ
Certain transformations on particular networks leave those networks invariant with respect to
the rate of cell growth. For example, a linear pragmatically equivalent control network to NL
contains a control jump that dedifferentiates cell type C back into its parent cell type A.
Eric Werner, A general theory of cancer
24
a
c
A
B
C
b
Fig. 11: Network NLJ an almost invariant transform of the linear cancer network NL
An almost equivalent linear cancer with one loop and two cell types but three control states.
The dashed arrow represents a jump within the cell cycle and does not involve cell division.
Cell A divides into two daughter cells with control states B and C. Daughter cell C jumps
to the control state of its parent cell A. The functional result is a linear proliferative network
practically equivalent to net NL Fig. 8 above.
Phenotype: Even though the networks NL and NLJ are almost equivalent, the tumor phenotype of NLJ may show three distinct cell types (A, B and C) and just the two (A and B) of
NL.
Eric Werner, A general theory of cancer
5.5
25
Linear mixed cell network NLM: Linear mixed cell cancer with different
cell types and structures
a
b
A
d
B
c
C
D
e
E
F
e
Fig. 12: NLM: Linear mixed cell cancer. The ball-arrow notation links a parent cell
with one of its daughter cells. It indicates that when the parent cell’s execution of the network
reaches the control area on DNA that is next to the ball at beginning of the arrow then the
parent cell divides and the linked resulting daughter cell activates the control area pointed to
by the end of arrow. In the above network, the cell in control state A divides into daughter
cells that differentiate into control states B and C. C is terminal. B divides into D and E. E
is terminal. A cell in control state D divides into E and A. The link from D to A creates an
infinite control loop that, given no inhibiting preconditions, results in endless cellular division.
This sort of single loop architecture occurs both in stem cells and in some cancers (see Sec. 8).
The phenotype of this network is an endlessly growing tumor generating two noncancerous
cell types (C and E) for each loop, yet there is only one cancer cell that cycles through three
cell types (A, B, and D) which are part of the cancer loop. Hence, it is very slow growing tumor
with one cancer stem cell (more on this below).
This linear cancer network results from a deleterious network transformation of the normal
network NN (Fig. 4) above. Specifically, it changed the pot1 address of cell control state D to
point to A instead of F. The cell state F is now inaccessible by this network.
Eric Werner, A general theory of cancer
26
A
C
B
D
E
Fig. 13: NLMtree: Tree view of a Linear mixed cell cancer. This is the same network as
above, but in a tree view that shows the branching structure of the network. The network on
DNA is the linearized version of this tree graph with loops. The de-differentiating loop in the
tree from D to A, makes this a never ending cell proliferating network.
5.6
Linear single teratoma network NLTs
If we define a teratoma to be any cancer that generates one or more noncancerous cell types
organized in some structure, then we get a whole range of teratomas, from simple teratoma
networks generating just a few non cancerous cell types, such as the mixed cell networks, with
no recognizable structure to complex teratoma networks that generate structure like hair, bone,
etc.
Phenotype: Under this broad definition of teratoma networks can range from the simple to
very complex. In the traditional usage, teratomas have multicellular structure in addition to
diverse cell types. We can further distinguish bounded teratomas from cancerous teratomas. A
bounded teratoma results from the activation of a developmental subnetwork that terminates
when its cells are differentiated and its possible stem cells are governed by their normal linear
control networks. A bounded teratoma network is thus a natural network in the genome that
has been inappropriately activated in time and/or space. An cancerous teratoma network is a
bounded teratoma network linked inside one or more linear, geometric or exponential cancer
network loops. For example, the case of a fetus in fetu (a multicellular system resembling a
fetus developing inside the body) may be a highly developed teratoma resulting from either a
bounded teratoma network or a cancerous teratoma network. A phenotype of repeating structures would be evidence of a cancerous teratoma network.
Note, too that in general complex teratomas may contain stem cells. Then if the stem cell loop
is linked to the cancerous backwards loop at D then we will get geometric growth where A is
in effect a meta-stem cell.
Eric Werner, A general theory of cancer
27
nx
A
B
Nx
D
b
a
Fig. 14: Network NLTs: Linear single teratoma cancer where A divides into a cell of type B
and a cell controlled by network N x that generates a structure with possibly multiple cell types
before it loops back to A.
In the network NLTs in Fig. 14 results from two steps: First, we amplify the network NL by
appending a more complex network N x into the genome. Next, we modify the network link
that led from cell type A to point to this new subnetwork while keeping the link to network for
daughter cell B the same. In linking the network NL to N x , the cell of type A will divide into
an N x -cell with the potency N x and a differentiated B-cell. We assume N x is a semi terminal
network that contains no loops except at cell type D. One of the developmental paths in N x
produces a cell of type D which dedifferentiates into cell type A again or divides into daughter
cells of type A and some other cell type F. Under the influence of subnetwork N x , the N x -cell
goes through a series of differentiations and possible divisions.
Remark 1 on NLTs Assuming synchronous divisions in an ideal, discrete, force free space
then NLTs produces k = N x (t) cells after t rounds of division. Let H(D) be the path in N x to
cell state D. H(D) is the history of network states in N x that end in state D. Let there only be
one path in N x leading to state D. Let |H(D)| be the length of the path H(D). Assume H(D) be
a longest path in N x . The N x (∣H(D)∣) = k is the total number of cells produced by N x when it
reaches cell state D.
Remark 2 on NLTs We could just count the number of cell divisions in the path H(D). Or
could just count the number of cells k produced by N x when state D is reached. Assuming D is
a longest path then k is the total number of cells produced by N x when it reaches D. We need a
notation for the number of cells produced by a network N when it reaches a cell state X. If N is
terminal, we need a notation for the total number of cells produced by N. Let N ∗ be the set of
paths in N. Let N(X)∗ or N ∗ (X) be the set of all paths up to the length of path H(X). We need
to be able to talk maximal paths in N and the length |H| of paths H in N. We need to talk of a
bounded network within an infinite looping network.
The clinical phenotype appears as follows: A cancer controlled by such a network will generally, depending on the complexity of the network N x , will be a more complex tumor with
Eric Werner, A general theory of cancer
28
multiple cell types in the N x cascade together with cells of type B. It may again be benign,
with linear non-exponential growth. Only one cell will be the cancer stem cell. All the rest
will be byproducts. Note however, the growth rate of network N x may be much slower than a
cancer with network NL. This is because the cell of type A will have to go through multiple
differentiations and possibly divisions before it returns to its state A. The salient feature of this
type of cancer will be the ratio of N x generated cells to cells of type B. Let k be the number of
cells produced by one loop of A → N x → A, then the ratio of tumor cells will be k cells of type
A → N x → A to one of type B. Thus, assuming the network N x produces n B cells, then a count
of B cells in the tumor will show how many loops the cancer has executed.
Let k be the number of cells produced by one round of N x when it reaches cell state D. Assuming the path to D is a longest path in N x then multiplying the number n of B cells by k
will give the number of cells generated by the subnetwork N x up to that point. Since the number of B cells generated by the network NLTs after n loops is just n, the ideal total number
Cells(n) of cells generated by the tumor network at loop count n is then given by the following
formula:
Cells(n) = k ∗ n + n
(5.2)
Thus, the total number of cells after n loops consists of k ∗ n cells of type N x plus n cells of
type B.
Note, all equations that give the cell count after a number of rounds of divisions are upper
bounds that indicate the maximum possible number of cells generated given ideal synchronous
development without physical constraints. Given real physics with friction, tension, cell connectivity and other such physical constraints the number of cells generated will in general be
less than the upper bound.
5.7
A linear double teratoma network NLTd:
In a linear double teratoma network a cancer cell A divides into two daughter cells each of
which is controlled by a teratoma network. Compare this to the above linear single teratoma
network NLTs. A linear double teratoma still has linear growth because only one of the teratoma subnetworks N x loops back to A.
Eric Werner, A general theory of cancer
29
nx
A
Ny
Nx
D
ny
a
Fig. 15: Network NLTd: Linear double teratoma cancer where A divides into a cell controlled by a bounded teratoma network Ny and a cell controlled by network N x that generates a
structure with possibly multiple cell types including cell type D that loops back to A.
Phenotype: A linear double teratoma tumor has all the structures of the linear single teratoma
tumor plus the cell types and structures produced by the second teratoma network Ny . Because
the cells of the two teratoma networks may interact both physically and via cell signaling,
apparently novel structures and cell types may result with the linear double teratoma that are
not generated by either of the teratoma networks singly. Recall, our more general definition
of teratoma as any network that produces multiple cell types and/or structures. Hence, double
teratomas can range from the very simple to very complex.
5.8
Multi linear cancer networks NLIk of identical cells
A linear k-identical cell cancer network NLIk links an identical cell network NIk that generates
k identical cells with a downstream linear cancer network. The simplest case is where the
number of identical linear cancer cells is k = 2:
Eric Werner, A general theory of cancer
30
b
b
A
B
C
D
b
c
Fig. 16: Network NLI2 : A multi-linear cancer of identical cells where the number of starting
identical cancer cells n = 2. The parent divides into two identical daughter cancer cells of type
B. B is controlled by a linear cancer sub-network that generates cell types B and C.
If 2 is the number of identical cells controlled by the network and n is the number of loops
synchronously executed by those cells then in this case the number of cells after n cell divisions
is:
Cells(n) = 2 + (2 × n)
(5.3)
The phenotype of such a cancer would have identical linearly growing cancers in several places
at once. They would have started simultaneously (up to some error range).
More generally, let NIk be a network of k identical divisions that generates 2k cells. Then we
can generate a cancer network NLIk if we link an upstream NIk to a simple linear cancer of
type NL (see Fig. 8):
kn
a3
a2
A1
A2
a2
ak
A3
a3
Ak = B
Ak−1
ak
C
c
Fig. 17: Network NLIk Multi linear cancer starting from 2k identical cells: This network
generates k identical divisions to produce 2k identical daughter cancer cells of type Ak = B
from one founder cell A1 . The B cells are controlled by a linear cancer sub-network that after n
rounds of division, jointly produce 2k × n cells of type C while retaining a constant 2k of linear
cancer cells of type B. Starting form one founder cell, the total number of cells after n rounds
of synchronous division is: Cells(n, k) = 2k + 2k × (n − k + 1) = 2k × (n − k + 2) if n > k and
Cells(n, k) = 2n otherwise.
Eric Werner, A general theory of cancer
31
The below is function describing the ideal rate of growth after n synchronous divisions where
k is the number of identical daughter cell divisions nodes in the network in Fig. 17.
Cells(n, k) = {
6
2n
if n <= k
k
k
k
2 + 2 × (n − k + 1) = 2 × (n − k + 2) for n > k
Exponential cancers
There are cancers that grow exponentially in that they increase as an exponential function of
the number of cells at any given time. For example, if the number of cells is k at time t of
which c are cancer cells, then an exponential cancer will contain at least cn cells after n units
of time. Now the rate of growth of an exponential process cn is vastly greater than that of the
polynomial nc or a linear process n ∗ c as will see.
For example, a cancer that increases by two cells at each time unit will have 2*30 or 60 cells
after 30 time units. Hence, after thirty time units (steps or rounds of cancerous development)
we would have a tumor with only 60 cells. While a cancer that is exponential would double
itself with each time unit or step in the cytogenic network. Hence after 30 time units we would
have a tumor with over 230 cells which is over a billion cells. Now since it is relative to the
length the time unit, if the time unit is one month then after two and a half years our linear
cancer is still very small being only 60 cells in size, while our exponential cancer patient has
developed a tumor with over one billion cells. If the cancer is fast growing with a cycle of one
day, then after one month a patient with the linear cancer will not notice it. However, a patient
with an exponential cancer will have a sizable tumor of some billion cells or if the cancer is
fluid, it will have spread very quickly. An exponential tumor that doubles with every time unit,
just one extra day the cancer will a double. So that after thirty one days our patient will have a
tumor of two billion cells.
All this assumes that there are no physical constraints on tumor growth. In natural systems
tremendous pressure is built up in solid tumors that prevent the realization of exponential
growth. Hence, the formulas describe ideal growth without physical constraints. They give
an upper bound on the rate of cancer growth.
Eric Werner, A general theory of cancer
6.1
32
Exponential aggressive growth cancer network NX
a1
A
B
C
a2
Fig. 18: Network NX: Exponential cancer where A divides into daughter cells that self loop
and differentiate to A.
The network NX controls cell division such that a cell of type A divides into two cells both of
type A. The Ideal Total number of cells after n rounds of cell division:
Cells(n) = 2n
(6.1)
Phenotype This network will result in a tumor whose growth rate is exponential doubling with
every division. All the cells will be of the same type. Depending on the connective properties
of the cells the tumor will spread quickly if the cells are not easily attached to one another or to
other cells types. If the cells are strongly bound to other cells of type A, then we have a tumor
that forms a large but quickly growing mass of similar cells.
Eric Werner, A general theory of cancer
6.2
33
Exponential mixed cell cancer network NXMa
a1
b
A
d
B
C
c
D
E
e
a2
Fig. 19: Network NXMa: Exponential mixed cell where a limited number of different cell
types in possible repeated patterns are formed.
Etiology: The network results from a network transformation of a linear network NLT (??)
that switches a pot pointer. If this is a genomic cancer network, it is caused by an addressing
transformation that in this case switches the pot2 address of the pitcher at area D from E
to A. The cell types C and E are non-proliferative while A, B and D are part of the cancer
cascade.
Phenotype: This is a mixed cell network since it generates several cell types. The tumor
has a limited number cell types A, B, C, D and E. It is highly aggressive, with exponential
cell proliferation. The ideal cell proliferation rate results in 2n cells after only n rounds of
synchronous cell division.
6.2.1
Ideal treatment that would stop exponential cancers
Ideally one can cure this cancer if one is able to reverse the network transformation that led to
the pathological network. If it is genomic in origin then an addressing inverse transformation
back to the original, presumably normal, developmental pot address would make this a linear
developmental network.
There are two cases depending on if the original linear network was a normal stem cell or a
linear cancer:
1. Normal linear stem cell network: A method that applies the inverse transformation to all
cancer cells will completely stop the exponential cancer. However, then we are still left
with an abnormal number (2n ) of stem cells.
Eric Werner, A general theory of cancer
34
2. Linear cancer network: A second reverse transformation would completely stop cell
proliferation resulting in an excess of 2n normal cells.
The above method would be especially valuable for inherited cancer networks, as these could
in principle be reversed to normal developmental networks.
Another method, is simply to transform the pathological address in that is either in the pitcher
or the catcher of the pot to a nonfunctional address so that the pathological network link is
eliminated in the cancer cells. The result would be 2n cells that no longer have the cancer network. However, if they contain a still open developmental network responsive to cell signaling,
they could continue to grow to they putative multicellular end state, even if this is no longer
normal growth in their new context.
Alternatively, if one had methods to recognize the cells with the pathological network, then
drugs might be developed that specifically target cells with the particular cancer network inducing apoptosis (cell death) or otherwise destroying those cells while leaving cells with normal
networks untouched.
6.3
Exponential single teratoma cancer: Network NXTs
NXTs modifies NX resulting in a network for an exponential growth cancer with distinct cell
types. Cell A divides into cell N x and A. The subnetwork N x *A1 for cell type N x goes through
a series of cell differentiations and possible divisions before looping back to cell type A. The
dynamic phenotype of this cancer results in exponential growth even if slower than that of
NX.
nx
A
B
Nx
D
a2
a1
Fig. 20: Network NXTs: Exponential single teratoma cancer where A divides into self and
network Nx that loops back to A.
We first observe linear growth of cell types A and N x . Then the N x cells form subsystems
of cells of various cell types. These N x subsystems then generate new cancer stem cells of
type A. This leads to gradual but certain exponential growth of the whole tumor and then to
Eric Werner, A general theory of cancer
35
exponential growth of each N x subsystem. Clinically, then we have a tumor of diverse cell
types with possible distinct subsystems with possible morphological structure, yet ultimately
while initially slower are still exponential in their growth. Note, until the N x *A network loops
back to A, the cancer appears to be a linearly growing cancer of type NL. Later each of the
subsystems induced by N x *A will appear to grow linearly, until sub-pockets of N x *A cells are
produced.
6.4
Exponential double teratoma cancer network NXTd
NXTd is a cancer network that modifies NXTs by also making the other A-loop or branch also
a subnetwork that first differentiates and possibly divides before looping back to A. Thus, A
divides into daughter cells governed by networks Nx and Ny, respectively. NXTd adds the
network Ny to the previous network NXTs which contained only Nx (see NXTs.) The NXTd
network is a new multicellular subsystem of various cells types and morphology. Like its
partner Nx, Ny ultimately has one or more of its paths loop back to the A cell type.
a1
nx
A
Ny
C
Nx
D
ny
a2
Fig. 21: Network NXTd: Exponential double teratoma cancer where A activates networks
Nx, Ny, both loop back to A.
Phenotype: Clinically, the phenotype of this cancer starts as a seemingly linearly growing
group of variously differentiated cells with perhaps two distinct identifiable developing morphologies. Then at different points Ny ∗ C and N x ∗ D will loop back to the A state, causing
gradual but exponential growth of the tumor. The growth may be slower than exponential
single teratoma network NXTs, but it will nevertheless be exponential.
Diagnostically, the key to recognizing these kinds of cancers is the appearance of repeating
subsystems with possibly two distinct growth rates and developmental morphological properties. The cancer stem cells of type A will be distributed throughout the tumor. And if the
cancer is fluid or non-connective then the cancer can spread to other cell types that overtly have
the same phenotype but are non-cancerous. Thus the healthy from non healthy cells will be
difficult to detect phenomenologically.
Eric Werner, A general theory of cancer
6.5
36
Hyper cancer networks NXH
Hyper cancers are exponential cancers with possibly mixed cell types because several identical
divisions that lead to possibly distinct differentiation states. The growth rate varies with the
number of identical loops and differentiation complexity of the intermediate states.
(a) A 4-node hyper cancer network
(b) Many seemingly noncancerous cell types in the overt
phenotype of a tumor
(c) Yet, all cells are cancerous
Fig. 22: Three views of a Hyper Cancer. In (a) we have a 4 node hyper cancer network view.
The overt phenotype of cells differs from their control state. The overt phenotype shown in
phenotype view (b) hides the fact that all cells in this tumor are cancerous. This is evident in
the control state view of the tumor shown in figure (c).
6.6
Simple Hyper Cancer NXH2
a
b
A
B
C
b
a
Fig. 23: Network NXH2 Two node Hyper Cancer: A simple exponential cancer of identical
cells. The parent A divides into two identical daughter cells B such that each daughter differentiate back to the parent control state A. Note, this has two control states with the potential of
producing two cell types with different differentiation states.
Eric Werner, A general theory of cancer
6.7
37
Identical cells with one loop back is still exponential but delayed
a
b1
A
B
C
D
b2
c
Fig. 24: Network NXIm Multi-Exponential Cancer: A simple multi-exponential cancer of
identical cells. The parent A divides into two identical daughter cells B such that one daughter
differentiate back to the parent control state A and the other differentiates to a cell of type C.
The growth is exponential but delayed and hence slower per execution of the cancer loops. The
phenotype of this type of cancer is an alternation of build up of cell types A or B while steadily
building up of cell type C. Cell type C may differentiate and grow into a complex structure.
These structures would be repeated in the resulting tissue each developing out of cell of type
C if it is linked to a further network.
6.8
Invariant transformation of exponential networks
In terms of growth this cancer is equivalent (up to a delay because of the dedifferentiation
jump) in terms of ideal growth rate to the simple exponential network above. Indeed, one can
add arbitrarily many identical cell divisions to the network and get an growth rate equivalent
networks. They are all growth rate equivalent to the simplest exponential cancer of simple cell
fission growth where daughter cells have the same control state as the parent. Even though
they are growth rate equivalent all these networks may have a different overt phenotype, since
each distinct control state as the cells divide may lead to different overt characteristics of the
cell.
There is a distinction between the overt doubling of cells at each ideal synchronous cell division
and the number of cells generated after one cancer cycle is completed. A hyper-cancer will
produce more cells after a cancer loop cycle, but its growth at each step generates the same
number of cells not matter how many identical cell divisions are in the loop. In this latter
sense, the cancer networks are equivalent and, thereby, invariant under the transformation.
They are not equivalent in terms of phenotype or the number of cells generated over the whole
cancer loop cycle.
Eric Werner, A general theory of cancer
6.9
38
Exponential cancer network with only one loop NXo
a
b1
A
B
C
D
b2
Fig. 25: NXo Exponential cancer with only one loop: The parent divides into identical
daughter cells that differentiate back to A.
Interestingly, one can have an exponential cancer with only one self activating loop. This can
happen when a cell A differentiates into two identical daughter cells and they then differentiate
back into their parent cell A, as illustrated in this example network:
6.10
n-Node Hyper Cancer Network NXHn
a1
a3
a2
A1
A2
a2
an
An − 1
A3
a3
An
an
a1
Fig. 26: Network NXHn n-Hyper cancer of degree n: This hyper cancer network generates
n identical divisions after one loop of the network to produce 2n identical daughter from one
founder cell. At node An the network generates one more identical cell division as it loops
c
back to its stating point. The network generates 2n cells from one founder cell after c cycles
of the network loop which is equal to 2 x cells after x rounds of synchronous division where
x = n ∗ c.
Eric Werner, A general theory of cancer
39
c
Cells(n, c) = 2n = 2 x where x = n ∗ c
(6.2)
Here c is used for whole cancer loop cycles, n is the number of nodes in the network, and x is
used for executing all cells by one network control time step.
6.11
An almost equivalent Hyper Cancer Network with a jump
The following network is slower but almost equivalent to the above hyper network. It is slightly
slower in growth due to the dedifferentiation step.
a1
a3
a2
A1
A2
a2
an
A3
a3
An − 1
An
an
Fig. 27: Network NXHjn n-Hyper cancer: This hyper cancer network generates n identical
divisions after one cycle of the loop in n+1 time steps to produce 2n identical daughter from one
founder cell. The dashed arrow represents a jump that does not initiate cell division. Hence, at
node An the network does a jump to the starting point without undergoing cell division. This
has the effect of starting the round identical cell divisions again producing an endless series of
c
identical cell divisions. The network generates 2n cells after c cycles of the cancer loop from
one founder cell. In n time steps, it produces 2n−c cells after n − c rounds of synchronous cell
division.
6.12
Exponential hyper teratoma cancer NXHTs
An exponential hyper cancer network is linked to a teratoma subnetwork.
Eric Werner, A general theory of cancer
40
a2
nx
A
Nx
C
D
nx
a1
Fig. 28: Network NXHT: Exponential hyper teratoma cancer network where identical
daughter teratoma networks loop back to A.
NXHT can be viewed as a variation of network NXTd is where subnetwork N x = Ny . Thus, A
divides into cells of type N x each controlled by looping network N x ∗ A where the ’*’ indicates
a link from N x to the network that controls cell type A. Then both will differentiate and possibly divide into a subsystem of some morphology with different cell types. Ultimately, they
loop back to cell type A. Again we have exponential growth. Fig. 28 is more uniform than
Fig. 21 because it will not have the complexity of having two distinct multicellular subsystems
governed by two distinct subnetworks. The growth rate of the subsystem will depend on the
time path through the loop A N x1 N x2 . . . N xn C A takes to complete one
round.
6.13
Growth rates for Hyper Cancer Networks
Let NXHn be network with a subnetwork Na is linked to Nb. Let Na be a passive network that
causes no cell divisions but only differentiates to Nb. If Nb produces, say 8 identical cells after
three rounds of division and then Nb loops back to Na, then for each execution of one loop the
network N8 = Na ⊗ Nb will produce 8 times as many cells as the previous loop. For example,
after 5 loops it will have produced 85 or 32,768 cells. After 10 loops, we would have over a
billion cells (810 = 1, 073, 741, 824 cells). While Nb is by itself harmless, when mutated to
activate Na, it becomes a very explosive cancer. If we add just five more rounds, we have 815
or over 35 trillion cells.
Note, this network N8 is similar to NXHj8 Fig. 27 and is equivalent in terms of ideal per loop,
growth rate to having a network of type NX Fig. 18 . However, the latter network requires 30
10
loops instead of 10 loops to generate 810 cells (since 23 = 230 ). Each loop in networks of type
NX doubles the number of cells at each round. And, it would require 45 loops of cell division
to generate over 35 trillion cells while NXHn can do this in 15 loops. Of course, there is a
sleight of hand here since the rounds of division are hidden in the network NXHn .
Eric Werner, A general theory of cancer
41
Our theory predicts that given a subsystem that produces a mass of cells of the same type
(with possible multiple cell types during NIn Fig. 7) by means of several rounds of division,
if that system becomes cancerous in just one of those cells, by looping back to the start of
the subnetwork controlling its development (to give something like Fig. 27) , then that system
will have the clinical phenotype of a highly explosive or hyper cancer. Indeed, given there are
enough resources and the cell physics permits it, then at least kn cells are produced after n loops
where k is the number of identical cells produced by the natural healthy system.
6.14
Multi-exponential cancers NiXn
An added loop transforms a multi-linear cancer into a multi-exponential cancer where a set of
two or more identical cells in equivalent control states all proliferate exponentially.
a1
A0
a2
A1
a1
an
A2
An − 1
a2
an
An = B
an
an
Fig. 29: Network NiXn Multi exponential cancer starting from 2n identical cells: This
network generates n identical divisions to produce 2n identical daughter cancer cells of type
An = B from one founder cell A0 . The B cells are controlled by an exponential cancer subnetwork jointly producing 2n ∗ 2k cells of type B after k further rounds of division. Starting
form one founder cell, the total number of cells after r = n + k rounds of synchronous division
is:
Cells(n, k) = 2n+k
(6.3)
While this is equivalent in terms of ideal growth rate to the simple one node exponential cancer
NX above, the tumor may have a different phenotype since each of the control states Ai in
NiXkn are distinct and may initiate the production different overt cell properties. Furthermore
the growth rates of the cells C(Ai ) controlled by Ai may differ from some C(A j ) controlled by
A j ≠ Ai . Hence, we would get variations in phenotype and growth rates even though ideally
with synchronous division the growth rates of NX and NiXkn exponentially equivalent since
Σn+k (NX) = Σn+k (NiXn ) = 2n+k after n + k rounds of synchronous cell divisions.
An instance where k = 1 to give 2 identical cells that continue to grow exponentially. This
instance is equivalent to NX since all cell types generated are identical in their control states.
Eric Werner, A general theory of cancer
42
The only difference is that the founder cell C(A) may differ in phenotype from cells of type
C(B) whereas all the cells generated by NX are identical to the founder cell.
b1
A
B
C
D
b2
Fig. 30: Network NiX1 Multi-Exponential Cancer: The simplest instance of the network
NiXn with n = 1. The parent divides into two identical daughter cells that each exponentially
generate cells. Cells(n + 1) = 21+k . Σ1+k (NiX1 ) = 21+k after 1 + k rounds of synchronous cell
divisions.
In this case the number of cells after n cell divisions is:
Cells(n) = c + (c × 2n )
(6.4)
where c is the number of identical cells controlled by the network and n is the number of loops
synchronously executed by those cells. The phenotype of such a cancer would have identical
exponentially growing cancers in n places at once. They would have started simultaneously
(up to some error range).
7
Geometric Cancer Networks
Having two or more proliferative loops does not necessarily result in an exponential cancer.
However, they can transform into exponential cancers. For example, using the WHO-grading
scheme, astrocytoma grad I, which is slow growing and benign, appears to be a linear cancer
of type NG, astrocytoma grade II -> grade III. Grade III is relatively fast growing compared to
grade II, but can be present for several years and then suddenly change to grade IV Glioblastoma Multiforme (GBM). GBM are fast growing and spread quickly. .
Eric Werner, A general theory of cancer
7.1
43
1st-Order Geometric Cancer Networks G1
a
A
B
C
b
Fig. 31: Network NG1 : A 1st-order geometric cancer with one loop: The cell of type A is
like a 1st-order stem cell; it produces terminal cells of some differentiated type B (which we
also will refer to as G0 cells). The difference between stem cells and cancer stem cells has to
do with their functionality in the system as a whole.
First order geometric cancer networks G1 contain one loop and are the same as the linear cancer
networks NL that we studied in Fig. 8. They contain one loop and generate cells linearly, one
at a time. The generated cells are terminal cells that do not proliferate. We will call terminal
cells produced by geometric cancer networks G0 cells. They are related to 1st order stem cell
networks (see below Fig. 34)
7.2
2nd-Order Geometric Cancer Networks G2
a
b
A
B
b
C
D
c
Fig. 32: Network NG2 : A 2nd-order geometric cancer with two loops: Both A and B are
linear cancer cells whose joint proliferation potential is given by Equation 7.10. The cell of
type B is a stem cell; it produces cells of type C. The cell A is a meta-stem cell that produces
stem cells of type B. The difference between stem cells and cancer stem cells has to do with
their functionality in the system as a whole.
Eric Werner, A general theory of cancer
7.3
44
Meta-stem cells
Geometric networks of the above type suggest the concept of meta-stem cells (see subsection 8.2). A linear or first order stem cell is a cell that produces other cells (that are not stem
cells) by means of a linear developmental network. Meta-stem cells are stem cells that produce
stem cells. Thus a 2nd-order geometric network with two loops contains a meta-stem cell A
that produces linear stem cells of type B. A 3rd-order three loop geometric network contains
a meta-meta-stem cell A that produces meta-stem cells B that produce stem cells C that produce cells of type D. And, so on. Whether a geometric meta-stem cell network is a cancer
meta-stem cell network will depend on the properties of the cells and how they function in the
organism. Any stem cell or meta-stem cell network can by mutations become an exponential
cancer (see below). A transformation of linear stem cell network into a meta-stem cell network
will make the stem cell proliferate according to Equation 7.4 below and be potentially harmful.
Indeed, even the metastatic behavior of cancer stem cells will be seen to be directly related to
the properties of their geometric, meta-networks (subsection 15.7).
7.4
Geometric cancer networks NGk with k loops
α1
α2
A1
A2
α2
αk
Ak
D
d
Fig. 33: Network NGk An k-th Order Geometric Cancer: A geometric cancer with k loops:
Each of the cell types Ai , i = 1 . . . k are linear cancer cells whose joint proliferation potential is
give by Equation 7.18. Each cell of type Ai , i < k, is a higher order meta-stem cell; it produces
stem cells of type Ai+1 . Thus, cell type Ak−1 is a meta-stem cell that produces stem cells of
type Ak , etc. The cell type Ak is a stem cell that produces regular non-stem cells of type D.
The difference between stem cells and cancer stem cells has to do with their functionality in
the system as a whole.
7.5
Mathematical properties of geometric cancer networks
A cancer network of the above type with one or more loops in a linear connected sequence leads
has interesting mathematical properties. It turns out that such networks are directly related to
Pascal’s Triangle, binomial coefficients and geometric numbers. Under the appropriate interpretation, one loop simple produces a line. Two loops produce a triangle. Three loops produce
Eric Werner, A general theory of cancer
45
an equilateral pyramid (tetrahedron), four loops a pentalope (a four dimensional pyramid), etc.
In general, an network of k linearly connected loops will produce an k-dimensional pyramid.
Or viewed numerically, after n synchronous rounds of ideal division, a cancer network with k
linearly connected single loops produce the sum of the first k binomial coefficients at level n of
Pascal’s triangle.
Recall that by definition:
(
n!
n
)=
k
k!(n − k)!
(7.1)
Then the number of cells produced by a single cell that is controlled by an cancer network with
k linearly connected loops, after undergoing n rounds of synchronous cell division division is
given by the following formula (for n > 0):
k
Cells(n, k) = ∑ (
i=0
k
n!
n
)=∑
i
i=0 i!(n − i)!
(7.2)
The above formula shows the direct relationship between such cancer networks and the binomial coefficients of Pascal’s triangle, i.e., the coefficients the binomial theorem.
Given the following standard definitions:
1. Linear number:
n
) when n ≥ 1
1
(7.3)
n2 + n n(n + 1)
n
=
=( )
2
2
2
(7.4)
n(n + 1)(n + 2)
n
=( )
3
6
(7.5)
n(n + 1)(n + 2)(n + 3)
n
=( )
4
24
(7.6)
Lin(n) = 1 + n = (
2. Triangular number:
T ri(n) =
3. Tetrahedral number:
T et(n) =
4. Pentalope number:
Pen(n) =
Then the number of cells, Cells(n, k), that develop after n rounds of synchronous division obey
the sum of the coefficients of Pascal’s Triangle where n is the height of the triangle and k is
horizontal coordinate corresponding to the number of single connected loops in the regulatory
network. These sums correspond to the volumes of the corresponding k-dimensional geometric
form in an ideal k-dimensional discrete space:
Eric Werner, A general theory of cancer
46
1. For one loop we get a 1-dimensional structure where its length gives the number of cells:
Cells(n, 1) = Lin(n) = 1 + n
(7.7)
2. Two loops add the triangular number to give the area of a 2-dimensional triangle:
Cells(n, 2) = Lin(n) + T ri(n − 1)
= 1+n+
(7.8)
n(n − 1)
2
(7.9)
(7.10)
3. For three loops add the tetrahedral number to give the volume of 3-dimensional pyramid:
Cells(n, 3) = Lin(n) + T ri(n − 1) + T et(n − 2)
= 1+n+
(7.11)
n(n − 1) n(n − 1)(n − 2)
+
2
6
(7.12)
(7.13)
4. For four loops add the pentalope number to give the volume of a 4-dimensional pyramid:
Cells(n, 4) = Lin(n) + T ri(n − 1) + T et(n − 2) + Pen(n − 3)
= 1+n+
(7.14)
n(n − 1) n(n − 1)(n − 2) n(n − 1)(n − 2))(n − 3)
+
+
(7.15)
2
6
24
(7.16)
5. For k loops sum the sequence of numbers through
dimensional pyramid:
n!
k!(n−k)!
to give the volume of a k-
Cells(n, k) = Lin(n) + T ri(n − 1) + T et(n − 2) + . . . + (
= 1+n+
n
)
k
n!
n(n − 1) n(n − 1)(n − 2)
+
+ ⋅⋅⋅ +
2
6
k!(n − k)!
(7.17)
(7.18)
6. So, in general we have:
⎧
1
⎪
⎪
⎪
Cells(n, k) = ⎨ k
⎪
∑i=0
⎪
⎪
⎩
if n = 0
n!
i!(n−i)!
= ∑ki=0 (
n
) otherwise
i
(7.19)
Eric Werner, A general theory of cancer
47
This view shows the direct relationship between this type of cancer and stem cell network
and the geometric numbers. What is fascinating is that these ideal networks have a generative competence with such interesting numerical properties. Their mathematical history spans
back from the binomial theorem, to Pascal’s triangle, to the Greek’s discovery of geometric
numbers.
8
Stem cell networks and cancer stem cells
Normal networks that come closest to cancer networks are stem cell networks. What distinguishes normal from cancer stem cell networks is not primarily their network architecture but
rather when and where a stem cell network is linked and activated, and what further networks
the stem cell network activates. A cancer stem cell network is either formed by mutation in
the wrong place in the global developmental network, it is activated inappropriately in the
multicellular system, or it activates inappropriate, nonfunctional networks with abnormal and
possibly pathological cellular or multicellular phenotypes relative to the overall multicellular
system.
From our theory it follows that there are at least two main kinds of stem cell networks, linear
and geometric. However, linear stem cell networks are just a special case of geometric stem
cell networks, namely, 1st order geometric stem cell networks. These are also fundamental
cancer stem cell networks. A normal linear stem cell produces a cell of a particular type that
is either terminal itself or activates a terminal network. A second order geometric stem cell
network is a meta-stem cell network that produces linear stem cells. Thus, linear stem cells
produce no further stem cells. Meta stem cells do produce stem cells.
Since a linear stem cell cannot produce further stem cells there must exist meta-stem cells that
produce more than one stem cell. These in turn are produced by yet further upstream embryonic
networks. The original embryonic fertilized egg is controlled by the global embryonic network
that is mostly terminal with the exception of its stem cells. It consists of many subnetworks
that may be multiply employed.
A linear cancer stem cell produces no additional cancer cells. Interestingly, we will see that
the order of a cancer meta-stem cell network controlling a cancer cell determines the possible
metastases producible by that cancer cell (see subsection 15.7).
A further possibility is that there exist normal exponential stem cell networks that are activated
by cell signaling, in effect a communication network linked with a cytogenic control network
to produce cells quickly on demand. However, such networks are be dangerous, leading to a
proliferative explosion if something goes wrong with the communication network.
8.1
First order stem cells
A first order stem cell produces no additional stem cells. Instead a first order stem cell A
produces cells B that is in a different control state than its parent cell A. The stem cell thus
Eric Werner, A general theory of cancer
48
produces one daughter cell that inherits the control state of its parent which is A itself and one
daughter cell B that is in a new control state. Thus stem cells have a self reflexive control
system. Note, that the cell B may still be multi-potent in that it may be controlled by a network
that generates a whole multicellular system.
First order stem cells have conditional control networks. This means that their activation depends on not just being linked into another network that activates them but that they can only
be active if the conditions Φ are satisfied.
a
Φ→A
B
b
Fig. 34: Network LSC1: 1st Order (Linear) Stem Cell Network The cell of type A is a
regular stem cell that divides conditionally. If condition Φ hold, then A produces cells of type
B. One daughter of A self loops giving a daughter cell of the same cell type A as the parent.
The other daughter cell differentiates to type B.
8.2
Second order stem cells or meta-stem cells
Unlike first order stem cells, stem cells or meta-stem cells produce other first order stem cells.
They generate stem cells. Since a first order stem cell cannot generate any further stem cells to
produce stem cells we need meta-stem cells or they must be produced by networks of the type
NIn that could produce 2n identical first order stem cells. It is possible that organisms use both
strategies.
Meta stem cells, like first order stem cells, also have conditional activation in that being linked
into another network is not sufficient for their potential activation. Their preconditions Θ must
be satisfied at each pass of the loop to execute the next loop.
Eric Werner, A general theory of cancer
m
49
a
Θ→M
Φ→A
a
B
C
b
Fig. 35: Network DSC2: Second order, meta-stem cell network: The cell M is a meta-stem
cell that given condition Θ holds, produces stem cells of type A which in turn conditionally
produces cells of type B. Thus, the cell of type A is a regular stem cell. If condition Φ hold,
then A produces cells of type B.
8.3
Higher order stem cells
Since stem cells obey the properties of geometric networks, theoretically it is possible to have
stem cell networks of arbitrary order k, such as 3rd order and 4th order meta-stem cell networks.
All such networks can have various preconditions Φi attached to the initiation of each the
contained stem cell network loops of order i. Without preconditions in an idealized setting
without physical constraints, an k−th order stem cell network would have the proliferations
properties of an k-th order geometric network (see 7.4).
8.4
Cancer susceptibility of stem cells
We will see why stem cells are especially susceptible to develop into cancer networks. First,
all that has to happen is that the first order or second order conditional antecedents are mutated to be always true, then the stem cell networks, once activated produces cells indefinitely
independent of outside signaling. Second, any of the links from the other non-proliferative
daughter cell differentiation site could loop back to a first or second order loop activation point
to produce an exponential cancer.
8.5
Normal stem cells and cancer stem cells
Stem cells have the function of continually producing new cells in tissue that is meant to regenerate. Stem cells have functional developmental control networks with preexisting loops.
This makes the progression to cancer a shorter route, because the linear, proliferating network already exists. We just need to mutate in one more loop to makes it exponential and
then with additional mutations it becomes invasive. Prostate cancer may be an instance of this
process.
Eric Werner, A general theory of cancer
50
If the stem cell proliferates conditionally, i.e., based on some contextual cellular or environmental condition, then a stem cell cancer can develop by a constant activation of the antecedent
condition necessary for the stem cell to divide. This can happen if the activating signal is constant or the receptor system for that signal is mutated to be constantly on. It may be the case
that a stem cell may already be potentially, linearly, geometrically or exponentially, proliferative, but its antecedent conditions may not be satisfied. Hence, if the conditional system
remains functional the potential stem cell proliferation whether it be linear or exponential will
not show itself. However, when the antecedent conditions for proliferation are fulfilled or the
antecedent testing system is mutated to a constant on state then the stem cell cancer will flower.
An example of this would be a stem cell whose proliferation depends on some signal such as a
hormone that, together with a signal transduction pathway, activates the stem cell proliferative
loop. Then if the signal is constantly turned on or the signal transduction mechanism is in a
constant "signal received" state, then the stem cell will proliferate either linearly, geometrically
or exponentially depending on the structure of its controlling network.
A further danger with some types of stem cells is that they may already be functionally, tissue
invasive in order to move to the appropriate site, as, for example, in wound healing. This
invasive property would then make a stem cell cancer even more dangerous.
Stem cells by themselves are not cancerous, but they have the properties of conditional, possibly invasive, linear, geometric or exponential cancers. The boundary between a cancer and
normal cell may not always be clear since in the wrong context a normal stem cell may be
cancerous.
Distinguishing linear cancer cells from stem cells based purely on the network architecture is
difficult if not impossible. Stem cells generate normal cell that are functional in the context of
the organism. Any stem cell that generates such "normal" cells inappropriately is a cancer cell.
Thus, the dividing line between linear cancers and stem cells may be indistinct. Thus, a normal
stem cell network becomes cancerous if it is activated inappropriately with respect to the overall functioning of a biological system in the organism, or if the cell differentiation networks
the stem network activates generate abnormal cells (e.g., invasive cells, or having abnormal
phenotype) or cells that generate structures inappropriately in the overall system.
9
Stochastic differentiation in stem cell networks
It has been argued that the traditional theory of linear stem cells is wrong [15]. On the traditional theory of stem cells, stem cells are immortal and constant in number. However, experimental data appears to show that for some stem cells appear to generate further stem cells. On
a deterministic theory stem cells that produce stem cells are what we have termed meta-stem
cells. However, another model would have stem cells divide and differentiate stochastically
either exponentially into two stem cells, linearly into a stem cell and a terminal cell, or into
two terminal cells. The evidence, suggests that epidermal cells divide stochastically according
to the distribution 8% double stem cells, 84% one stem cell and one terminal cell, and 8% two
terminal cells. The biological mechanism is not known.
Eric Werner, A general theory of cancer
51
l
p = 0.08
A
l1
SS
p = 0.08
t
L
TT
T
l2
t
t
p = 0.84
Fig. 36: Network TriSSC: Tripartite stochastic stem cell network that activates one of
three developmental networks. One that produces one meta-stem cell (SS) which, in turn,
produces two linear stem cells (L), one that produces one stem cell (L) which generates terminal cells (T) and one that produces a pre-terminal cell (TT) that, in turn, generates two terminal
cells (T). Globally, network generates either two terminal cells or a minimum of one and at
most two nonterminal stem cells.
For example, a simple stochastic model might have ternary stochastic differentiation where a
cell differentiates by activating, according to a tripartite probability distribution, one of three
developmental networks, namely one that produces a meta-stem cell that generates two stem
cells, one that produces one stem cell that generates terminal cells and one that produces a preterminal cell that generates two terminal cells. However, such a network would at most produce
two stem cells, or one stem cell or no stem cells. No dedifferentiation to the originating cell A
is possible under this network.
9.1
Historical background: A network that generates Till’s stochastic stem cell
model
One of the first models of stem cells was the stochastic model of hemopoietic cell proliferation
[Till et.al. [29]]. Till distinguished what he called “colony forming cells” that have the capacity
to form colonies from differentiated cells without that capacity. When a colony develops from
a single cell only a small number of these cells have “colony-forming capacity”. The rest are
differentiated cells without this capacity. This he considered a "birth-and-death" process, the
generation of colony cells being a birth process and the generation of differentiation a death
process. According to the model, a given cell with colony-forming capacity can either divide
into two colony forming cells (with probability p2 ) or differentiate into a terminal cell (with
probability p0 = 1 − p2 ) having no colony forming capacity. The case of generating mixtures of
one colony-forming cells and one differentiated cell was considered a birth process followed
by death and was handled by adjusting the probabilities.
Eric Werner, A general theory of cancer
52
Production numbers of blood cells is relatively constant under normal conditions. Under stress
or increased demand there is a rapid increase in production of blood cells. The properties of
hemopoiesis (blood cell creation) implies that the production of differentiated cells is precisely
controlled. Hence, Till argues, there must be control mechanisms. Till then considers if cell
proliferation (the numbers of cells produced) is also under precise control or lax control. He
argues for lax control and that the data suggests cell proliferation is stochastic. While a stem
cell network was not given by Till, the behavior of the Till model of colony forming cells can
be generated by following stochastic stem cell network:
p2
p0 = 1 − p2
j1
S
J1
T
j2
Fig. 37: Network TillSSC: Classic stochastic stem cell network. This network models one of
the earliest stochastic stem cell theories (Till [29]). A stem cell (S) divides to produce two cell
of type J1 that then stochastically change either into a terminal network cell T or “self renew”
by dedifferentiating to the parent stem cell S type. The cell J1 stochastically either loops back
to activate the stem cell S with probability p2 or it activates the terminal network cell T with
probability p0 = 1 − p2 . The network’s behavior approaches a deterministic exponential stem
cell network as the probability p2 approach 1. Stem cells controlled by such a network can
spontaneously stop because of the fact that for all points in all possible paths in the network
there is the possibility of reaching the terminal cell state T. The multicellular system controlled
by this network consists of two main cell types: Stem cells (S) that can proliferate and terminal
cells (T) with no proliferative potential.
Under this network (Fig. 37) a stem cell divides into identical daughter cells J1. Each J1
cell can with probability p2 dedifferentiate into a stem cell S or differentiate with probability
p0 = 1 − p2 into a terminal cell T. Hence, unlike Till’s original model, this network explicitly
handles all the possible outcomes of stem cell division. Upon stem cell division, the network
allows the generation of two stem cells (S,S), as well as the mixed case (S,T) of one stem cell
and one differentiated cell, and the case (T,T) of two differentiated cells.
Till’s model has been criticized because it predicts that stem cells eventually differentiate into
terminal cells and, thus, no longer self renew. However, this depends on the probability distribution. If p is high then exponential proliferation outruns terminal cell differentiation.
A central problem with this network is that the key to its behavior is dependent on the value
Eric Werner, A general theory of cancer
53
of the probability p2 . On the one hand, if the probability p2 of stem cell self-renewal is high,
it tends toward exponential growth of stem cells. This leads to too many stem cells versus
progenitor and terminal, specialized cells. On the other hand, if the self-renewal probability p2
is low, it eventually leads to the elimination of all stem cells. In the latter case, this network
does not represent a stem cell with unlimited proliferative potential. In this network, no dedifferentiation is possible for terminal cells. Hence, the greater the probability p0 = 1 − p2 that
the terminal network T is activated, the more likely is the permanent quiescence of the parent
stem cell.
If the probability p2 = 1 we have exponential growth of stem cells. If p2 = 0 we have differentiation to the terminal cells T. Hence, this network cannot model the developmental dynamics
of deterministic linear networks of the type NL Fig. 8. However, with the right choice of the
probability distribution, it can approximate a relatively constant production of stem cells and
terminal cells.
In their Monte Carlo simulation the probability used was p2 = 0.6 and p0 = 0.4. In our Monte
Carlo simulation, using their probability distribution, the network proliferation dynamics tends
toward a 21% proportion of stem cells after 52 generations. This appears to be much higher
than the empirical data for hemopoietic stem cells. However, these percentages depend on
how many cells are generated by the terminal cell T. If T is not a terminal cell and instead a
progenitor cell controlled by a terminal network T*, and is instead a terminal cell state that
does not divide further (as was the case in Till’s original model), then the stem cell percentage
is lower being 12% after 20 generations of cell division. We add progenitor cell capacity to T
in the network in (Fig. 39) below.
Note, in this sort of modeling approach, precise differentiation networks (architectures) are
coupled with stochastic activation networks. The topology of the network architecture sets the
boundaries for what stochastic paths are possible at all.
Eric Werner, A general theory of cancer
(a) Probability p2 = 0.52, gives 4%
stem cells
(b) When p2 = 0.7 gives 36% stem
cells
54
(c) Probability p2 = 0.9 gives 80%
stem cells
Fig. 38: Three views of stem cells under different probabilities. All three tumors are controlled by same Till network (Fig. 37 ). It shows the growth response to changes in the probability p2 . As the probability p2 is increased, more and more stem cells are generated due
to increasing dedifferentiation of the daughter cells to their parent stem cell. The stochastic
exponential network begins to dominate as p2 approaches 1.
9.2
The Till stochastic network extended to generate progenitor cells
It is very simple to extend Tills original model of hemopoietic (blood) stem cells (S) [Till
et.al. [29]] to include progenitor networks that control the proliferative capacity of stem cell
progeny. We simply link the terminal cell state to a bounded network T* that in the case below
is simply a network that results in one further cell division to produce two distinct daughter
cells T1 and T2. But, T* could be any bounded network. Thus, in principle, given the right
network T*, such progenitor cells could develop into arbitrarily complex structures.
Eric Werner, A general theory of cancer
55
p2
t1
p0 = 1 − p2
j1
S
J1
T∗
T1
T2
j2
t2
Fig. 39: Network TillSSCpro: Classic stochastic stem cell network (Fig. 37) extended to
include progenitor networks. This network extends one of the earliest stochastic stem cell
theories (Till [29]) to include progenitor cells. A stem cell (S) divides to produce two cell of
type J1 that then either divide into a terminal network cell T* or “self renew” by dedifferentiating to the parent stem cell S. The cell J1 stochastically loops back to activate the stem cell
S with probability p2 or it activates the terminal network cell T* with probability p0 = 1 − p2 .
In Till, T* is a terminal cell type T. In this example, T* is a terminal network for a progenitor
cell that generates two terminal cell types T1 and T2. As with the original model (Fig. 37), this
network’s behavior approaches a deterministic exponential stem cell network as the probability
p2 approach 1. Stem cells controlled by such a network can spontaneously stop because of the
fact that for all points in all possible paths in the network there is the possibility of reaching
the terminal cell state T*. The multicellular system controlled by this network consists of three
main cell types: Stem cells (S), progenitor cells (T*) with limited proliferative potential, and
terminal cells (T1, T2) with no proliferative potential.
As with the previous network Fig. 37, the proliferative potential of stem cells under this network Fig. 39 is dependent on the probability p2 . Since no dedifferentiation is possible in this
network for terminal cells T1 and T2. Furthermore, it is not an inherent property of the network topology that stem cells self renew. However, if the dedifferentiation probability p2 of
self renewal is high enough then stem cells do self renew and may even dominate the colonies
they generate.
If the probability p2 = 1 we have exponential growth of stem cells, with no progenitor cells T*.
If p2 = 0 we have differentiation to to the terminal progenitor network T* which in this case
ends in just two terminal cells T1 and T2. Hence, this network cannot model the developmental
dynamics of deterministic linear networks of the type NL Fig. 8. However, with the right choice
of the probability distribution, it can approximate that a relatively constant ratio of stem cells,
progenitor and terminal cells.
It follows that under the Till model of stem cells we do not need dedifferentiation of terminal
cells to maintain stem cell capacity. All we need is for the 1st order stem cells to be physically
loose, mixing in with the terminal cells in the tumor as it grows. In this case, we would see
proliferating stem cells within the tumor in the context of what appear to be only terminal cells.
Eric Werner, A general theory of cancer
56
It may appear as if terminal cells have spontaneously dedifferentiated into stem cells, but this
need not be the case since it may be the result of stem cell mixing.
What could also be happening is that we have a 2nd order geometric stem cell network that
is generating 1st order stem cells which due to developmental control in conjunction with
physics leads to a mixture of 1st order stem cells with terminal cells. Since the control state
of a cell my be hidden by the overt differentiation state of the cell, the 1st order stem cells
may be indistinguishable relative to a set of markers from terminal cells. Stochastically or
under particular conditions these 1st order stem cells may then begin to proliferate generating
terminal or progenitor cells.
9.3
A flexible stochastic network architecture with exponential and linear potential
p
1− p
j1
S
J1
J2
j2
T
1−q
q
Fig. 40: Network XLSSC: Flexible exponential-linear stochastic stem cell network. One
stem cell (S) divides to produce two cells (J1 and J2) that each stochastically activate either the
stem cell itself or a terminal cell.
In the network in Fig. 40 is very flexible. Depending on the probability distribution, the network can range between being exponential, linear, or terminal, as well as every mixture in
between. Furthermore, the network can be deterministic, stochastic or mixture of both. This
flexibility is partly the result of separating out the probability distributions for the behaviors
of the two daughter cells. It shows that the architecture of the network imposes constraints on
what kinds of developmental dynamics are in principle possible. The probability distribution
presupposes a network architecture of possible developmental paths.
The cell S is only a stem cell stochastically and not intrinsically when p < 1 and q < 1. When
p = q = 1 the network is deterministic exponential. When p = 1, q = 0 or p = 0, q = 1 the
network is deterministic linear, i.e., a deterministic 1st order geometric stem cell network. If
Eric Werner, A general theory of cancer
57
p = q = 0 the network is terminal. When p = 1 and q < 1, or when q = 1 and p < 1, then the
network is mixed deterministic linear with stochastic exponential tendencies.
If probabilities p = q then as p and q approach 1 the network approaches the behavior of a
deterministic exponential network. However, if p and q are different then this network can
simulate a linear stochastic network as well when, for example, p approaches 1 and q approaches 0, or vice versa. As the probabilities p and q decrease, the more frequently the cancer
stem cell results in a terminal tumor that does not develop further because it consists only of
cells of terminal type T. This shows that stochastic cancer stem cell networks can in some cases
go into spontaneous remission. While this network can also exhibit exponential growth even
in a stochastically linear probability distribution, because of the two backward loops, there is
a diminishing probability that it remains exponential. Thus, whether this network results in
linear or exponential proliferation depends on the probability distribution.
For the network XLSSC, if the probabilities p = 1 − q then the higher the probability of p
the more the network approximates a deterministic linear developmental network. Since, in
this case the distribution is anti-symmetric, the cell population partition of cell types consists
of an equal number of exponential stem cells and terminal cells, with the majority of cells
being linear stem cells. This corresponds to the observed distribution in epidermal basal stem
cells. If, on the other hand we have a symmetric distribution where p = q then the higher
the probability of p the more the network approximates a deterministic exponential network.
Thus the type of cancer network we have depends on the probability distributions over the
connecting stochastic links.
The network XLSSC (Fig. 40) has some similarity to the Till model (TillSSC Fig. 37), when
p = q (and value of probability p = p2 as in the original model by Till [29] model) and T.
The Till model forces that daughter cells of S are the same. In contrast, this model is more
flexible in that, depending on the probability distribution (p, q), it can model both exponential
and linear dynamics of multicellular development.
Eric Werner, A general theory of cancer
9.4
58
Linear stochastic stem cell network LSSC
p
1− p
j1
L
J1
T
b
Fig. 41: Network LSSC: A linear stochastic stem cell network. A linear stem cell (L) divides
to produce a terminal cell T and cell J1. The cell J1 stochastically loops back to activate the
stem cell T or it activates the terminal cell T. The network’s behavior approaches a deterministic
linear stem cell network as the probability p approach 1. Cancer stem cells controlled by such a
network can go into spontaneous remission because of the fact that for all points in all possible
paths in the network there is the possibility of reaching the terminal cell state T.
1− p
j1
L
J1
T
t
p
Fig. 42: Network DLSSC: Deterministic linear stem cell network with stochastic delay. A
linear stem cell (L) divides to produce a terminal cell T and cell J1. The cell J1 stochastically
loops back to activate the stem cell T with probability p. Unlike with a true linear stochastic
network like LSS which stochastically self-differentiates to L with probability p or differentiates to T, this is a deterministic linear stem cell network that always self-differentiates back
to L. The stochasticity only effects the rate at which the cell divides. The probability p only
effects the frequency of looping back to L and, thereby, changes the cycle time cell division.
Unlike LSSC, the linear cancer stem cells controlled by this network never go into spontaneous
remission since the cell endlessly self-renews. The lower the probability p the longer it takes
for the cell to cell renew and divide.
Eric Werner, A general theory of cancer
9.5
59
1st-Order Geometric Cancer Networks with open stochastic dedifferentiation
1− p
a
S
J
T
j
p
Fig. 43: Network G1SD: A 1st-order geometric cancer with stochastic dedifferentiation:
In this linear network, a cell of type S is a 1st-order stem cell. It produces progenitor cells of
type J that have a stochastic dedifferentiation potential. A cell J can either dedifferentiate to its
parent state S or differentiate into the terminal cell T. The effect of the stochasticity is seen in
the population of progenitor cells J produced by the network. The lower the probability p of
dedifferentiation, the greater the number of terminal cells in the cell population generated by
the stem cell S.
9.6
1st-Order Geometric Cancer Networks with closed stochastic dedifferentiation
In the above network Fig. 43 dedifferentiation can occur once prior to permanent differentiation
in a terminal cell state. The effect of the stochasticity can only be seen in the tumor as a
whole that is by the cancer stem cell. In the following network dedifferentiation is a constant
possibility for the life of the cell.
1− p
a
S
J
T
j
p
Fig. 44: Network G1SDc: A 1st-order geometric cancer with closed stochastic dedifferentiation
Eric Werner, A general theory of cancer
60
In this linear network Fig. 44, a cell of type S is a 1st-order stem cell. It produces progenitor
cells of type J that have a continual stochastic dedifferentiation potential. A cell J can either
dedifferentiate to its parent state S or maintains its present state J. In this closed stochastic
network, a cell J continues to have a dedifferentiation potential for the life of that cell. There
is no path to permanent terminal differentiation. This contrasts with the network Fig. 43 where
the cell only has a limited time frame to dedifferentiate prior to its permanent differentiation
into a terminal cell T. Thus, all in a population of cells generated by S have the potential to
dedifferentiate. In contrast, in the network Fig. 43 if the probability p is low, most cells in the
population generated by S will be terminal cells with no dedifferentiation potential.
9.7
Closed exponential stem cell network with stochastic delays
Depending on the probability distribution, the next stochastic network proliferative properties
can vary from deterministic terminal, linear, or exponential to stochastic terminal, linear or
exponential.
1−q
p
1− p
j1
S
J1
J2
T
j2
q
Fig. 45: Network DXSSC modified LSSC: Closed exponential stochastic stem cell network. If the probability q > 0 then this network is inherently nonterminal. All stochastic paths
lead to potentially nonterminal nodes in the network. p = q = 0 is terminal. p = q = 1 is
exponential. p = 1, q = 0 is linear. p = 0, q = 1 is exponential.
Unlike LSSC Fig. 41, if the probability q > 0 in cell state J2 then there are no true, probability
independent, terminal cells in DXSSC Fig. 45. True terminal cells have no possible stochastically available paths that lead to a reversal of differentiation to a more dedifferentiated cell
type. Instead, all cell states are stochastic with possible developmental paths that lead to the
original founder stem cell S. The cell in state J2 has a probability q of dedifferentiating to S.
The higher the probabilities q and p the more this network mirrors the behavior of a deterministic exponential network. If either p or q is low while the other high it is more similar to a
linear developmental network. Thus the probability distribution determines the similarity to
Eric Werner, A general theory of cancer
61
linear or exponential networks. Even if the probability of q is very low, as the number of J2
cell increase it becomes more and more likely that one of them will dedifferentiate to S. Thus,
the cancer pulls away from linearity with an increasing cell population size.
If probabilities q = p = 0 then the network is deterministic terminal network leading to only one
cell division. If q = 0 and p = 1 we have a deterministic linear stem cell network. However, if
the probability q = 0 and p > 0, then the network is a stochastic linear network where the lower
the probability p the greater the chance that proliferation terminates. If either p = 1 or p = 0,
and if q = 1 then the network is a deterministic exponential stem cell network. If q > 0 then no
matter how small the nonzero probability q is, proliferation increases with time as more and
more cells are created.
Remark: There is another way to escape the maze of exponential proliferation. Cell proliferation can be influenced by the number of times the self-loop at J2 is allowed to repeat. If it
is it is only allowed to repeat only a limited number of times relative then become quiescent,
then, depending on the probability q, it could result in a terminal differentiation state after some
stochastic tries. In other words, if, in parallel with the repetitions of the self-entry loop at J2,
there exists a separate simultaneous, parallel process of terminal differentiation by means of
some counting mechanism (where the cell at J2 becomes quiescent after some finite number of
counting steps), then this network could lead to the ultimate termination of cell proliferation
or at most linear proliferation. The counting mechanism could be dependent on the number of
loops executed at J2 or it could be a function of some other temporal variable.
9.8
A 1st order geometric/exponential stochastic stem cell network with stochastic progenitor cell dedifferentiation
p
1− p
s
j1
G
1−s
j3
J1
J2
B
q
j2
1−q
J3
j4
T∗
J4
r
1−r
Fig. 46: Network G1SSC: A 1st-order stochastic geometric/exponential stem cell network
with stochastic dedifferentiation.
Eric Werner, A general theory of cancer
62
This stochastic stem cell network (Fig. 46) generates progenitor cells B that generate cells J2
that can dedifferentiate to stem cells G or differentiate to a terminal network T ∗ . Depending on
the probability distribution, this network can exhibit terminal, linear-geometric, and exponential developmental dynamics. If the probability distribution is such that p = q = r = s = 1 then
this is equivalent to a deterministic 1st order geometric stem cell network. A 1st-order stem
cell (G) stochastically divides to produce either two progenitor cell B, or or two stem cells G or
one progenitor cell B and one stem cell G. and cell J1. The cells in states J1 or J2 stochastically
loop back to activate the stem cell G or to activate the progenitor cell B. The progenitor cell
B divides into a semi-terminal cell T ∗ by way of J3 or in divides into two cells of type J4. J4
either terminates with T ∗ or it dedifferentiates into a stem cell G.
Its behavior approaches a deterministic 1st-order geometric stem cell network as the probabilities p and q approach 1. However, the subnetwork activated by this 1st-order geometric
network can still stochastically dedifferentiate into one or two new stem cells G. Cancers controlled by such a network can go into spontaneous remission because of the fact that for all
points in all possible paths in the network there is the possibility of reaching a terminal cell
state. However, because there is the possibility at J1, J2 and J4 to dedifferentiate to earlier,
upstream stem cell states, the relative numbers of stem cells to terminal cells will be higher
than in networks that have fewer dedifferentiation pathways. Note, too that the dedifferentiation pathways introduce several possibilities for exponential growth. Even if dedifferentiation
probabilities are very small, any increase in these probabilities could have a significant destabilizing influence on the resulting tumor.
9.9
A broad spectrum Linear or 2nd order geometric or exponential stochastic
stem cell network
p
s
1− p
j1
G
1−s
j3
J1
j2
1−q
J2
B
q
J3
j4
J4
T
r
1−r
Fig. 47: Network LG2XSSC: A stochastic linear, 2nd order geometric or exponential stem
cell network. If the probability distribution is such that p = q = r = s = 1 then this is equivalent
to a deterministic 2nd order geometric stem cell network. Thus, as these probabilities approach
1 the behavior approaches a 2nd order geometric stem cell network.
Eric Werner, A general theory of cancer
63
In Fig. 47, a 2nd-order meta-stem cell (G) divides to produce a 1st-order stem cell B and cell
J1. The cell J1 stochastically loops back to activate the meta-stem cell G or it activates the first
order linear stem cell B. The 1st-order stem cell B divides into a semi-terminal cell C and a cell
J2. The cell J2 stochastically activates either its parent stem cell B, the semi-terminal cell C or
dedifferentiates to G. The cell C also can stochastically dedifferentiate to any of the previous
stem cell fates (G, B) or differentiate into the final terminal cell T.
The network consists of two linked 1st-order linear stochastic networks. Its behavior approaches a deterministic 2nd-order geometric stem cell network as the probabilities p and
q1 approach 1. Cancers controlled by such a network can go into spontaneous remission because of the fact that for all points in all possible paths in the network there is the possibility
of reaching a terminal cell state. However, because there is the possibility at J2 and C to dedifferentiate to earlier, upstream network states, the relative numbers of stem cells to terminal
cells will be higher than in networks that have fewer dedifferentiation pathways. Note, too that
the dedifferentiation pathways introduce several possibilities for exponential growth. Even if
their probabilities (q3 , r2 , r3 ) are very small, any increase in these probabilities could have a
significant destabilizing influence on the resulting tumor.
The deterministic possible behaviors include: (p = 0, q = 1, s = 0, r = 1) is terminal. (p =
1, q = 0) is exponential. (p = 1, q = 1, s = 0, r = 1) is linear. (p = 1, q = 1, s = 1, r = 1) is
2nd order geometric. (p = 1, q = 1, s = 1, r = 0) is a linear network linked to an exponential
network.
9.10
2nd Order Geometric stochastic stem cell network
p
q
G
1−q
1− p
j1
J1
j2
B
b
J2
C
c
Fig. 48: Network G2SSC: A 2nd order geometric stochastic stem cell network. A 2ndorder meta-stem cell (G) divides to produce a 1st-order stem cell B and cell J1.
In Fig. 48 The cell J1 stochastically loops back to activate the meta-stem cell G or it activates
the first order linear stem cell B. The 1st-order stem cell B divides into a terminal cell C and
a cell J2. The cell J2 stochastically activates either the stem cell B or the terminal cell C. The
network consists of two linked 1st-order linear stochastic networks. Its behavior approaches
Eric Werner, A general theory of cancer
64
a deterministic 2nd-order geometric stem cell network as the probabilities p and q approach
1. Cancers controlled by such a network can go into spontaneous remission because of the
fact that for all points in all possible paths in the network there is the possibility of reaching a
terminal cell state.
9.11
Geometric stochastic stem cell network with dedifferentiation
p
q1
j1
G
J1
r1
q2
1− p
j2
B
J2
C
T
c
b
q3
r2
r3
Fig. 49: Network GSSCdediff: A geometric stochastic stem cell network with dedifferentiation.
In Fig. 49 let probabilities q1 + q2 + q3 = 1 and r1 + r2 + r3 = 1. A 2nd-order meta-stem cell
(G) divides to produce a 1st-order stem cell B and cell J1. The cell J1 stochastically loops back
to activate the meta-stem cell G or it activates the first order linear stem cell B. The 1st-order
stem cell B divides into a semi-terminal cell C and a cell J2. The cell J2 stochastically activates
either its parent stem cell B, the semi-terminal cell C or dedifferentiates to G. The cell C also
can stochastically dedifferentiate to any of the previous stem cell fates (G, B) or differentiate
into the final terminal cell T. The network consists of two linked 1st-order linear stochastic
networks. Its behavior approaches a deterministic 2nd-order geometric stem cell network as
the probabilities p and q1 approach 1. Cancers controlled by such a network can go into spontaneous remission because of the fact that for all points in all possible paths in the network
there is the possibility of reaching a terminal cell state. However, because there is the possibility at J2 and C to dedifferentiate to earlier, upstream network states, the relative numbers of
stem cells to terminal cells will be higher than in networks that have fewer dedifferentiation
pathways. Note, too that the dedifferentiation pathways introduce several possibilities for exponential growth. Even if their probabilities (q3 , r2 , r3 ) are very small, any increase in these
probabilities could have a significant destabilizing influence on the resulting tumor.
Eric Werner, A general theory of cancer
9.12
65
Modeling deterministic stem cell networks with stochastic networks
For stochastic stem cell networks the probability distribution over a network topology determines their dynamic properties. If the stochastic network allows all possible paths in a set of
linear, 2nd-order geometric and exponential networks, then the dynamic behavior of each is
approximated as the corresponding embedded links are assigned probabilities that match existence or nonexistence of links in the particular network type, For example, if a link exists in a
network embedded in the stochastic network, then it assigned a probability close to 1 or 1 itself
if we want an exact match. If the link does not exist in the embedded network, it is assigned a
probability close to 0 or 0 itself for an exact match. Hence, the probability distribution distinguishes different network types that are embeddable in the topology or architecture of a given
stochastic network.
9.13
Transformations of probability distributions change stochastic stem cell
behavior
If we allow the probability distribution over a stochastic network to change with time because
of external factors (such as ultra violet radiation) then the same stochastic stem cell network
may exhibit various proliferative phenotypes, appearing alternatively as being in remission,
linear, geometric or exponential. Thus, meta-probability functions are at work here that conditionally change the probability distribution.
10
Conditional cancers
Some subnetworks are not activated except in special circumstances. For example the network
that controls wound healing is activated if there is injury to tissue. Now if a cell has a cancerous
mutation in its wound healing subnetwork, the cancer will not become evident until some or
sufficient surrounding tissue has been injured. As the wound healing subnetwork is activated
the cancer will start. Repeated injury to, or irritation of tissue has been associated with cancer.
For example, leukoplakia from chewing tobacco and esophageal carcinoma from repeated,
chronic episodes of reflux esophagitis. It has been speculated that the injury itself causes the
cancer. This may be.
However, we now have a possible alternative explanation: repeated injury makes it more likely
that a cell with a cancerous mutation in its wound healing subnetwork is found, i.e., induced
to initiate its wound healing control subnetwork. On this account the repeated injury serves
more to activate an already existing cancerous subnetwork, rather than being the cause of
the mutation in that subnetwork. Alternatively, the repeated activation of the linear stem cell
network regenerates damaged tissue becomes transformed because of error build up resulting
from overuse. The transformation may be to a geometric or exponential cancer. It may also be
a differentiation transformation to an earlier cell type resulting in a new stem cell which would
imply geometric or exponential growth with potentially invasive properties.
Eric Werner, A general theory of cancer
66
An additional aspect of cancerous wound healing subnetworks is that they involve cell dedifferentiation. If such immature cells are involved in the cancerous subnetwork then the result
may be an fast growing, possibly exponential cancer. For example, if cell differentiates to two
identical cells then a single loop can cause an exponential cancer (see Fig. 25).
11
Reactive communication networks
We distinguish reactive communication networks from interactive communication networks.
Reactive communication networks respond to input signals with some behavior. When that
behavior leads to sending signals and not just interpreting incoming signals the networks are
interactive.
Eric Werner, A general theory of cancer
11.1
67
A reactive signal based differentiation network architecture with linear and
exponential potential
σ1
σ2
j1
S
J1
J2
j2
T
σ4
σ3
Fig. 50: Network SigLXSC: Reactive signal based differentiation stem cell network with
linear and exponential potential. One stem cell (S) divides to produce two cells (J1 and J2).
The red and green squiggly arrows between two cell states indicate cell signaling, e.g., the
green arrow J1 ↝ S labeled with σ1 means that on receiving the signal σ1 a cell in state J1
differentiates or jumps to state S . On signal σ1 J1 activates the parent stem cell S. On signal
σ2 , J1 differentiates to a terminal cell T. Similarly, J2 responds to signal σ3 by dedifferentiating to state S and to σ4 by activating the terminal network T. Hence, as long as J1 receives
the dedifferentiation signal σ1 and J2 receives its dedifferentiation signal σ3 the network will
generate exponential cell proliferation. The network has linear proliferative behavior if J1 (or
J2) receives the dedifferentiation signal σ1 (σ3 ) while J2 (or J1) receives the terminal signal
σ4 ( or σ2 ). It is a terminal network if both cells J1 and J2 receive their terminal signals σ2
and σ4 , respectively. The ultimate output of this network whether it has linear, exponential or
terminal dynamics is just one terminal cell type T. If T is the start of a terminal network, it
would produce one or more terminal structures each generated by T. If the cells of T interact
then the structures T may form composite structures with different properties than T itself even
though the whole is terminal.
Eric Werner, A general theory of cancer
11.2
68
A reactive signal based geometric network architecture with 1st and 2nd
order geometric potential
s
l
σ1
S
l
M
σ2
L
T
t
Fig. 51: Network SigLSC: Reactive signal based geometric stem cell network with linear
and meta-stem cell potential. On signal σ1 one stem cell (M) divides to produce two cells (M
and L). M dedifferentiates back to the parent cell S. On signal σ2 , the linear stem cell network
L is activated. Once L is activated it will continue to produce terminal cells T. Hence, this
network is signal controlled. It will continue to produce M and hence S cells as long as it
receives the signal σ1 . S cells will differentiate to L cells on signal σ2 stopping meta-stem cell
production. The signal σ2 can thus eliminate all meta-stem cells. Hence, it may be undesirable
for systems that need to preserve their meta-stem cell line. Eliminating the signal σ2 -path from
S to L would insure that the supply of meta-stem cells can always be increased as long as there
are some meta-stem cells or S cells. The terminal cell type T may be a single terminal cell or
a terminal progenitor cell T* that generates a bounded number of further terminal cell types.
Eric Werner, A general theory of cancer
11.3
69
A pure reactive signal based geometric network architecture with 1st and
2nd order geometric potential
sL
sM
σ3
σ1
SM
M
SL
L
T
sL
σ2
t
σ4
Fig. 52: Network SigG2SC: Pure reactive signal based geometric stem cell network with
flexible geometric, linear and terminal potential. On signal σ1 , the signal dependent cell S M
differentiates to the meta-stem cell (M) which then divides to produce two cells (S M and S L ).
Both S M and S L are controlled by equivalent signal dependent networks. Their further behavior
depends on the signals they receive. The signal combination < σ1 , σ3 > produces a meta-stem
cell network with 2nd order geometric potential. The signal combination < σ1 , σ4 > and
< σ2 , σ3 > produce a 1st order stem cell network with linear potential. The signal combination
< σ2 , σ3 > causes terminal differentiation stopping all cell production. This example network
shows that different signal combinations can change the architecture of signal based networks
leading to distinct stem cell types with different generative potential. The terminal cell type T
may be a single terminal cell or a terminal progenitor cell T* that generates a bounded number
of further terminal cell types.
Eric Werner, A general theory of cancer
11.4
70
A hybrid signal stochastic network architecture with exponential and linear potential
p
1− p
j1
S
J1
J2
j2
T
σ2
σ1
Fig. 53: Network SigLXSSC: Hybrid signal stochastic stem cell network. One stem cell
(S) divides to produce two cells (J1 and Σ). J1 stochastically activates either the stem cell
itself or a terminal cell. J2 responds to signal σ1 by dedifferentiating to state S and to σ2 by
activating the terminal network T. Hence, as long as a stem cell receives the signal σ1 it will
proliferate exponentially or linearly depending on the probability p. Given the signal σ1 , if the
probability p is high the network will proliferate exponentially otherwise if the probablility p
is low, the network will proliferate linearly. Alternatively, if the network S receives the signal
σ2 and the probability p is high it will proliferate linearly, if low, it will eventually terminate
in T. If signal σ2 = σ1 where σ1 represents the absence or repression of the signal σ1 then the
default behavior of the network without the signal σ1 is to proliferate linearly if the probability
p is high or if the probability p is low, to terminate with T.
12
Communicating cancer networks in systems of cooperating social cells
A type of conditional cancer are social cancers that depend on cell signaling to be active. This
variety of cancer occurs when the genomic network interacts with receptors, signals and signal
transduction pathways. In that case we can have what might be called social cancers. For
example, a cell A will signal α to cell B and differentiates to A1, B on receiving the signal
divides into B1 and B1. B1 sends a signal β back to A and then B1 dedifferentiates to B
(its network loops back). When A1 receives β its network loops back into state A. Now the
process repeats with the A cells sending signals α to B cells. One can see that this process
is potentially exponential as long as we have sufficient cells of type A with sufficient signal
capacity to continue to activate all the developing B cells.
With conditional cancers that depend on social communication, the growth rate of the cancer
will depend on the accessibility of the signal. If the receiver of the signal requires direct contact
with the sender then even if the cytogenic cell contains a conditional exponential network,
Eric Werner, A general theory of cancer
71
since the signal may not be received by those daughter cells the exponential potential may not
be realized. Furthermore, if it is a linear network then if the passive daughter cell is interposed
between the sender and the receiving cytogenic cell then growth will stop after several divisions
when the signal no loner can reach the cytogenic cells. If later, because of physical pressure or
other conditions, the cytogenic cell is again in close enough proximity to the sender then the
cancer can start again.
12.1
Interactive signaling mono generative networks
In the figure below we emphasize the communication links. The network has one cell only signaling and the other cell signaling and conditionally dividing, however it leaves open whether
the cytogenic subnetwork is linear or exponential. The more detailed communication networks
and their properties will described in the sections below.
α
A
β
B
Fig. 54: A social signaling cancer loop where cell A is cancerous The cancer is social and
conditional in that it depends on a signal from B to divide. Since the network shows only one
loop, it leaves open the potential of the other daughter cell. Hence, it leaves open whether the
cytogenic network has linear or exponential or some other potential.
An in vivo instance of this network type is seen in bone cancer where a signal from the adjacent
tissue is required for the other to become cancerous, for example, see (Logothetis [17]).
12.2
Interactive signaling mono-linear network architecture
The following network controls two cells types A and B that communicate by cell signaling.
The cell A divides only when it receives the signal β from cell B. The cell B only sends its
signal after it has received the signal α from cell A.
Eric Werner, A general theory of cancer
72
β
α
β
A
A1
A2
a2
α
t
T
B
B1
b
a
Fig. 55: Network SigL1: An interactive signal based stem cell network with linear potential. Cell A only divides if it receives a signal β from cell B. Cell B only sends a signal if it
receives a signal from cell A. On signal β the cell A enters control state A1 . Then A1 sends a
signal α to B. Subsequently A1 switches to control state A2 whereupon it divides to produce
two cells (the dedifferentiated parent cell type A and a terminal cell T). Hence, one of the
daughter cells of A2 dedifferentiates back to the parent cell A. The cell type B does not divide
in response to the signal from A. Instead, B responds to the signal α by switching to state B1
and sending a signal β to A. After that B1 dedifferentiates back to B. The cell A will continue
to divide as long the signal loop is maintained. Notation: The blue and green squiggly arrows
between two cell states indicate cell signaling. The blue arrow A1 ↝ B labeled with the triangular α means that a cell in state A1 sends a signal α that is received by a cell in state B. The
green arrow B ↝ B1 labeled with the rectangular α means that a cell in state B on receiving the
signal α differentiates or jumps to state B1 .
12.3
Interactive signaling mono-exponential network with linear growth
The following network controls two cells types A and B that communicate by cell signaling.
The cell A divides only when it receives the signal β from cell B. The cell B only sends its
signal after it has received the signal α from cell A.
Eric Werner, A general theory of cancer
73
β
a
α
β
A
A1
A2
a2
α
B
B1
b
a
Fig. 56: Network SigX1: Signaling mono exponential cancer network. Cell A only divides
if it receives a signal β from cell B. Cell B only sends a signal if it receives a signal from cell A.
On signal β the cell A enters control state A1 . Then A1 sends a signal α to B. Subsequently A1
switches to control state A2 whereupon it divides to produce two identical daughter cells (both
of the dedifferentiated parent cell type A). Hence, both daughter cells of A2 dedifferentiate
back to the parent cell A. The cell type B does not divide in response to the signal from A.
Instead, B responds to the signal α by switching to state B1 and sending a signal β to A. After
that B1 dedifferentiates back to B. The cell A will continue to divide as long the signal loop
is maintained. Prior to any communication the network is in a waiting, passive or deadlocked
state. is an external signal such as a growth hormone to jump start the network.
12.3.1
Preconditions
For this network to be actively cancerous it requires some systemic preconditions: The existence of at least two cells of type A and B. Furthermore, the cells A and B have to be positioned
so that they can receive signals from each other. Yet even if these conditions are fulfilled, the
network may be inactive. This is results from what is called deadlock where each cell is waiting
for a signal from the other before it does anything. Thus, this network is inactive until an external source sends either the signals or β to cell A or the signal α to cell B. The external signal
might be a growth hormone, for example, that is sent at a certain period of development or
the result of environmental factors. Once one of the recognized signals is received, the network
starts endless cell proliferation given the above systemic conditions are maintained.
12.3.2
Signal differentiation determines linear or exponential growth
Interestingly, in this system exponential growth is bounded by the number of B cells. Even if
the cells of type A have the potential to proliferate exponentially, they can only divide if they
are close enough to a B cell to receive its signals. We assume only close (connected) cells
can have communication links to each other. Let mb be the number of cells of type B. Let
ca,b be the maximum number of A cells that can connect to a single cell B and, thereby, be
Eric Werner, A general theory of cancer
74
close enough to a B cell to send and receive signals. Then the maximum number of cells that
can be generated by the above network after n synchronous divisions is given by the following
formula:
Cells(n) = n × ca,b × mb
(12.1)
Therefore, exponential networks controlled by signaling communication protocols need not
exhibit their exponential potential. Indeed, in this example, the exponential network generates
only linear growth.
12.3.3
Social cell proliferation as systemic and not cell autonomous
Under this network only a whole system of cells including both types A and B have the potential
for multicellular proliferation. The individual cells alone do not have that potential since each
cell type A or B in isolation will not divide.
12.3.4
Signal identity changes can transform linear to exponential growth
A transformation in just one signal type that, for example changes the sender α to send β
instead, changes the proliferation dynamics of this exponential network to true unhindered
exponential growth. For in this case, the A cells signal each other, and possibly themselves, to
divide. Therefore, the signal differentiation is crucial in constraining this exponential network
to exhibit only linear growth.
12.4
Dual cancer signaling networks
A the interacting between two cooperative cell types, called partners, is driven by a network
where each partner has a separate role driven by a different subnetwork with signaling driving
their actions. In the figure below we emphasize the communication links. The network has
both cells signaling and conditionally dividing, however it leaves open whether the network
is linear or exponential. The more detailed communication networks and their properties will
described in the following sections.
Eric Werner, A general theory of cancer
75
α
A
β
B
Fig. 57: A dual cancer signaling network where both communicating partner cells A and B
are cancerous. The cancer is social and conditional in that each cell depends on a signal from
the other to divide. The rate and extent of growth depends on whether the properties of the
cancer subnetworks of the cell types A and B.
What we have here is a signal loop that is essential for the whole cancer network. The dual
cancer signaling network in Fig. 57 consists of two subnetworks, and A-subnetwork for cell
type A and a B-subnetwork for cell type B. Once activated, the A-subnetwork (e.g., Fig. 58,
Fig. 60 or Fig. 59) causes a signal α to be sent to cell B followed by activation of subnetwork
A2 whereupon the cell divides into two daughter cells at least one of which enters the signal
receptor state A which enables the cell to receive signals of type β. The signaling partner
subnetwork of cell type B is activated by a signal transduction cascade initiated by the receipt
of signal α. The B-subnetwork in state B1 directs the sending of a signal β to cell A. After
sending the signal, the B-subnetwork enters state B2 that directs B to divide into two daughter
cells, one or both of which loop back, to enter the receptor state B.
Clinically, we would observe depends on nature of the cytogenic subnetworks, as well as the
developing morphology of the tumor in the tissue. The cells A and B only proliferate if have
a neighbor that is a signaling partner. Thus, the growth of B cells depends on being adjacent
to A cells and vice versa. We would observe the stopping of cell proliferation if either the A
or the B cells are removed or some other agent interferes with the signaling loop. The latter
can be effected either by inhibiting the receptor for α or β, the signal α or β, or by interfering
with either theαor β signal transduction pathway. Alternatively, one could interfere with the
genome by inhibiting the areas responsible for initiation of signaling or the inhibiting the areas
responsible for cellular division. One can see there are many potential areas where one can
break the cancer loop. In actual in vivo systems one would choose that with minimal risk of
side effects.
We now investigate some of the particular types of dual cancer signaling networks.
Eric Werner, A general theory of cancer
12.5
76
A interactive signaling dual linear network architecture
β
α
A
α
t1
β
A1
A2
T1
B
t2
B1
a2
a
B2
T2
b2
b
Fig. 58: Network SigL2: A signal based dual linear network with linear potential. Two
cells A and B have equivalent signaling protocols and behavior. Cell A only divides if it
receives a signal β. Upon receiving the signal β, cell A changes to control state A1 . Prior to
dividing A1 sends a signal α to B. Then A1 changes to control state A2 whereupon it divides
to produce two cells (the dedifferentiated parent cell type A and a terminal cell type T1 ). The
communication protocol for B is similar to A. Cell B only divides if it receives a signal α. Upon
receiving the signal α cell B differentiates to control state B1 which then sends a signal β to A.
After that, cell B1 switches to control state B2 . Then cell B2 divides to produce two different
cell types (the dedifferentiated parent cell type B and the terminal cell type T2 ). The cell types
A and B will continue to divide as long the signal loop interaction protocol is sustained.
Eric Werner, A general theory of cancer
12.6
77
Interactive signaling network with exponential and linear subnetworks
β
a
α
α
β
A
A1
A2
T1
B
t2
B1
a2
a
B2
T2
b2
b
Fig. 59: Network SigXL: A signal based dual network with exponential and linear subnetworks. Two cells A and B have equivalent signaling protocols but different proliferation
behaviors. Cell A only divides if it receives a signal β. Upon receiving the signal β, cell A
changes to control state A1 and sends a signal α to B. Thereafter A1 changes to control state
A2 and divides to produce two identical daughter cells both of which are of the dedifferentiated
parent cell type A. This leads to potential exponential growth of cell type A. Similarly, cell
B only divides if it receives a signal α. Upon receiving the signal α cell B differentiates to
control state B1 which then sends a signal β to A. After that, cell B1 switches to control state
B2 . However, unlike A2 , B2 , divides producing two different cells types -the dedifferentiated
parent cell type B and the terminal cell type T2 . The cell types A and B will continue to divide
as long the signal loop interaction protocol is sustained.
12.7
A doubly exponential signaling network with linear dynamics
Interestingly, even a doubly exponential signaling network can have linear proliferative dynamics. If we transform the linear subnetwork in Fig. 59 to an exponential subnetwork or
equivalently add an exponential network to the mono exponential signaling network above in
Fig. 56 then we get a signaling dual exponential network:
Eric Werner, A general theory of cancer
78
β
a
b
α
β
A
A1
α
A2
B
B1
a2
a
B2
b2
b
Fig. 60: Network SigX2: Signaling dual exponential cancer network. Cell A only divides
if it receives a signal β from cell B. Cell B only divides if it receives a signal from cell A.
On signal β the cell A enters control state A1 . Then A1 sends a signal α to B. Subsequently
A1 switches to control state A2 whereupon it divides to produce two identical daughter cells
that are both of the dedifferentiated parent cell type A. Cell B behaves in an equivalent manner
in response to the signal α from A. On receipt of α cell B switches to state B1 and sends a
signal β to A. Subsequently B1 switches to control state B2 whereupon it divides to produce
two identical daughter cells that are both of the dedifferentiated parent cell type B. Both the
cells A and B will continue to divide as long their signal loops are maintained. Prior to any
communication the network is in a waiting, passive or deadlocked state. is an external signal
such as a growth hormone to jump start the network.
As in the case above (see network 56), two exponential subnetworks in a signaling network
do not necessarily lead to exponential growth. Instead cell proliferation is limited to those
cells of type A and B that are in sufficient contact to maintain their signaling loops. As the
cells proliferate the number of cells in communication contact may change depending on their
location in space. Hence, the context created by the multicellular development determines the
rate of growth.
n−1
Let ma be the number of cells of type A and mb be the number of cells of type B. Let Ka,b
be
the total number of cells A and B in communication contact. They are the cells that receive a
growth signal from a communication partner cell at time step n − 1. These are the cells close
enough to be in signal contact with a complementary communication partner cell so as to send
and receive signals at time step n − 1. Then the maximum number of cells generated by the
above network after n synchronous divisions is given by the following formula:
n−1
Cells(n) = 2 × Ka,b
(12.2)
If for each synchronous round of cell division, all the cells generated by step n are still in
contact at time n + 1 then this network will exhibit exponential growth. If however only a
Eric Werner, A general theory of cancer
79
constant number Ka,b of cells remain in contact after each synchronous round of division then
we have linear or polynomial growth.
12.7.1
Linear development with communicating dual exponential networks
Therefore, exponential networks controlled by signaling communication protocols need not
exhibit their exponential potential. Indeed, in this example, if the cells in communication
i
contact Ka,b
remains constant for all rounds of cell division i then the exponential network
generates only linear growth even if it is double the linear growth of the single exponential
network above.
Fig. 61: Cells generated from a signaling double exponential cancer network. The yellow
and red cells are the only cells that are dividing. The green and blue cells are signaling partners,
while potentially exponential in growth, are mostly passive for lack of a signal from each other.
Hence, while their network (SigX2 in Fig. 60) is a doubly exponential signaling network, the
overall growth is linear or at best first order geometric. The reason is the lack of communication
links to the more distant cells. In consequence, the communication between blue and green
cells generates a boundary of red and yellow proliferating cells that in turn produce more blue
and green signaling cells.
In simulations such as the typical example in (Fig. 61), the exponential networks exhibits only
linear or at most 2nd order geometric growth. The reason is that the physics of development
is such that there is no full mixture of cells rather there is a boundary or interface between
the two cell types where the communicating cells meet and interact. In the boundary region
they can grow. As cells are formed away from the proliferative communication interface while
potentially proliferative are not actually.
12.8
Self-signaling cancer networks
Cell communication networks may be self-reflexive. The cell sends a signal to itself that it
receives and interprets by activating a cytogenic subnetwork. The resulting self-induced cell
proliferation depends on the properties of the cytogenic subnetwork. In another scenario, a
receptor may also get stuck with a constant “signal received” message in which case the cell
divides without an external signal.
Eric Werner, A general theory of cancer
80
α
A
(a) A social self signaling cancer loop
(b) A self-signaling tumor generated by exponential and
linear subnetworks
Fig. 62: A self-signaling cancer cell schemata and its exponential and linear tumor phenotypes. A cell with self signaling cancer network that induces the cell to send a signal to itself
that activates cell division into at least one similar self signaling cells. The growth rate will
depend on the properties of the cytogenic subnetwork. The tumor in 63b on the right is the result of two self signaling networks. The cells in dark blue are active tumor cells controlled and
generated by of an exponential cytogenic subnetwork. The pink cells are passive, generated by
a self-signaling linear network.
The schematic network illustrated in figure 62a is a variation of signaling cancers (Fig. 54)
where instead of needing a outside signaling partner B, a cell A sends a signal α to itself.
If the cytogenic subnetwork is exponential then, on receiving the signal, A divides into two
copies of itself, A and A. These in turn send signals α to themselves and neighbors of the
same cell type A, etc. This results in an infinite loop with a corresponding exponentially
proliferating self induced cancer. The tumor in 63b shows the growth of a self-signaling
exponential network.
If the cytogenic subnetwork is linear, then we have a slow growing linear cancer, where the
cancer stem cell does not produce more cancer stem cells. Instead it produces cells of some
other cell type B. If these B cells still inherit the cancer receptor for the signal α, then they may
dedifferentiate to the A cell type and start linear proliferating.
Treatment of signaling cancers would involve agents that attack key control points in the
genome network activation signaling loop complex. Another approach would be just to kill
off fast growing cells without regard to their control. That therapy would work for signal
autonomous networks like NX networks as well.
Eric Werner, A general theory of cancer
12.8.1
81
A self-signaling network architecture with an exponential subnetwork
Clinically, when the cytogenic subnetwork is exponential, then we have a very fast growing
and exponentially proliferating cancer. From a diagnostic perspective, we would observe a
relatively uniform cell mass if the cells are cohesive, attaching to cells of their own type. Else,
if the cells are fluid we would have a quickly proliferating cancer permeating all systems of the
body.
σ1
s
A
AS R
s
σ1
(a) Network SelfSigX1
(b) A state of a tumor controlled (c) A advanced stage of a tumor
by network SelfSigX1
controlled by network SelfSigX1
growing within a bilateral cellular
system
Fig. 63: A self-signaling exponential network and its tumor phenotypes On the left fig.
(a) shows a self reflexive signal based cancer network (SelfSigX1). The cell A divides and
differentiates into two sender-receiver cells of type AS R with receptors for the signal σ1 . In the
role of sender cell AS R sends a signal σ1 to itself and other neighboring cells. On receiving a
signal σ1 , possibly from itself or from its neighbors, the cell in state AS R dedifferentiates to its
parent cell type A and the process starts again. Fig. (b) shows a stage of a tumor generated by
the network and growing independently where dividing cells are in red with communicating
cells in yellow and green. Fig. (c) shows an advanced stage of the same type of tumor growing
bilaterally in a bilateral multicellular system. It has breeched the bilateral midline. Tumor cells
are in red. Blue and yellow cells are part of the normal, noncancerous bilateral cellular context.
This network outwardly at the cellular level behaves much like the self activating network
exponential network NX (see Fig. 18). A differential diagnosis would result if the cancer was
restrained by interference with the cell signaling system since that is needed by network Fig. 54
for it to be cancerous. The distinguishing feature of the NX network would be its resilience
and resistance to any form of signaling system therapies, e.g., signaling inhibitors, receptor
deactivation, signal-transduction pathway interference.
12.8.2
Two self-signaling networks with linear versus exponential subnetworks
A linear and an exponential self-signaling network shown below makes the self-signaling versus the cytogenic subnetworks more apparent. The network on the left in the figure below is
Eric Werner, A general theory of cancer
82
grows linearly in response to its own signal while the network on the right has exponential
potential and once started is equivalent to the network in Fig. 63 above.
b
β
A
β
A1
a
α
l
T
B
α
B1
b
Fig. 64: Network SelfSigSC: One linear and one exponential self-signaling subnetwork
architectures. There are two unconnected subnetworks. In the first subnetwork on the left,
the cell type A sends a signal β to itself. On receiving the signal β it changes to control state
A1 . It then divides into a dedifferentiated cell of the parent type A and a cell of type T. This
subnetwork exhibits linear growth. In the second subnetwork, the cell type B sends a signal
α to itself. On receiving the signal α it changes to control state B1 . It then divides into two
dedifferentiated daughter cells of the parent type A, leading to exponential growth.
13
Bilateral cancers
Some cancers are bilaterally symmetric. Examples include symmetric breast cancers where
the tumor develops in the same place in each breast. Based on computational modeling and
simulation, we hypothesize that the mutations that lead to bilateral symmetric cancers must
occur very early in development since bilateral symmetry is established very early in the first
few divisions of the embryo (Gardner[11], Werner [44]). If the mutation occurred in a cell
after the establishment of bilateral symmetry then the resulting cancer would only occur in
one of the body halves. Most likely, therefore, is that such bilateral cancerous mutations are
inherited.
Eric Werner, A general theory of cancer
(a) Top view of two bilateral tumors (red,
orange and yellow cells) in a slice of a bilaterally symmetric organism.
83
(b) Side view of two bilateral tumors (red, orange and
yellow cells) in a slice of a bilaterally symmetric organism
Fig. 65: Bilateral Cancers: The top view of a bilateral cancer is shown on the left in figure
(a). A side view of the same bilateral cancer in a slice of blue and purple context cells is shown
on the right in (b).
Bilateral cancers can be controlled by linear, exponential, geometric cancer networks or signaling cancer networks. Note, a bilateral symmetric cancer can result from just one mutation.
The very same genome controls the development of each mirror body half. Thus, the tumor
itself is not symmetric, rather the tumors (to the extent that they are not stochastic or context
dependent) are mirror images of each other and occupy mirror positions in the respective body
half of the person.
It is extremely unlikely that the very same mutation would occur in the two genomes of two
mirror cells whose position and type mirror each other in the two symmetric body halves.
Instead, the best explanation, for symmetrically located cancers that mirror each other in phenotype, is either that they are germline mutations, or that a somatic mutation occurred in the
genome of a single cell early in development prior to the establishment of bilateral symmetry
in the embryo, or that the mutation was inherited.
13.1
Apparent unilateral cancers generated by submerged bilateral networks
If tumor development contains stochastic activation points, then Monte Carlo simulations show
that in many cases only one of the two potential bilateral cancers will actually continue to grow.
The other remains passive because the cancer signaling context did not establish itself. Thus,
the rarity of bilateral cancers need not imply that bilateral cancer networks are as rare. They
may just not show themselves in normal circumstances. And, some apparently unilateral cancers may in fact be based on stochastic or context dependent bilateral cancer networks.
Eric Werner, A general theory of cancer
13.2
84
Temporally displaced bilateral tumors
It may also happen that the bilaterality of the cancer is temporally displaced and that the complementary tumor starts to grow at some future time. In this case, apparently distinct tumors
are actually caused by the very same underlying bilateral cancer network. For example, in
breast cancer, later tumor development in the second breast may be the result of a stochastic,
signal or context dependent bilateral cancer network.
14
Vacuous and implicit non growing "cancers"
Interestingly, some cancers do not grow at all either because the loop contains no cytogenic
initiation sequence or because because the network is skipped and never activated by the main
genomic and epigenetic control networks.
14.1
Vacuous non-cytogenic networks with cyclic differentiation
If we remove the crucial step of cell division we have a loop where the cell just cycles through
a series of states endlessly without any cell proliferation. Such loops can be viewed from two
perspectives: On the one hand, viewed therapeutically, they offer a possible way to convert
an explicit cancer into a relatively harmless loop. On the other hand, viewed diagnostically,
such implicit non cytogenic loops, are potentially pre-cancerous in the sense that a mutation
that inserts or activates a cell division switch into the loop, transforms the harmless loop into
an explicit cancerous loop. Furthermore, one would expect that some house keeping networks
use loops to continually repeat needed cellular processes. If a copy mutation inserts a cell
generating subnetwork into the house keeping cycle then we have a possible cancer.
d
b
A
B
C
D
c
a
Fig. 66: Network NVD: Vacuous non-cytogenic cyclic network with cyclic differentiation,
where ↪ indicates a non-proliferation jump to a different differentiation state. The network
cycles through different differentiation states without producing any daughter cells.
Eric Werner, A general theory of cancer
14.2
85
Vacuous cancer sub-networks skipped by the developmental controlling
network
Below, we have not just an empty loop but rather a potentially proliferating cycle that would be
cancerous if it were not skipped by the dominant activating cascade. A mutation in A that would
lead it to differentiate to B instead of C would then result in an exponential cancer. Such networks can potentially be activated by cell signaling or by being linked to another network that
is linked in to the dominant control network governing development and sustainability.
c
b1
A
B
C
D
b2
Fig. 67: Network NVX: A Vacuous potentially exponential cancer where the activation path
from A to C skips B
On the positive side, such networks point to a possible way of curing cancer by converting a
cancer network into a vacuous network by skipping it. This would be achieved by interfering
with any one of the steps in the cancer cycle such as changing the catching address that initiates
the cancer loop.
15
Metastases
In an orthogonal process that adds yet another dimension to cancer networks, cells can break
their normal boundaries, enter the blood stream or lymphatic system to travel to new areas
of the organism forming potential new cancerous regions called metastases. The underlying
cytogenic network controlling the cell is not changed by the cell’s change of location. However,
because the cell’s context may change, new interaction protocols may change its proliferative
behavior by changing its network activation state. Even the cell’s lack of a cellular matrix, may
change its control state, causing it to generate its own cellular context. We consider just a few
possible cases of cancer networks in metastatic scenarios.
15.1
Metastatic potential of cancer network types
Each major type of cancer network has different metastatic potential. A linear cancer is generated by only one stem cell that produces terminal cells or at most terminal progenitor cells.
Eric Werner, A general theory of cancer
86
A 2nd order geometric cancer produces endless 1st order cancer stem cells which in turn produce terminal cells. Geometric cancers of order 3 produce 2nd order geometric cancer cells
and, hence, have greater metastatic potential. Exponential cancers have the greatest metastatic
potential for distributed, invasive growth since the daughter cells produce further exponential
cancer cells. If we add signaling and conditional activation to a given cancer network then,
in addition to cell physics, the cellular environment can be both an enabler or a constraint on
metastatic potential.
15.2
Metastasis via signaling
If a cell with an inducible signaling cancer network cell manages to break its boundaries and
travel to other parts of the organism, it would not necessarily lead to a cancer developing. First,
the invasive cell A cannot endlessly divide without a cooperating communicating partner cell B
with the correct complementary signaling protocol. Second, the cell still needs an appropriate
starter signal to activate its cancer network either from a partner B or from an external source
such as a growth hormone.
(a) Three metastases in a dual exponential signaling
network
(b) Three metastases in a mono exponential signaling network in a context of signal cells
Fig. 68: Two views of metastasis by different networks. Figure (a) shows the growth response to 3 metastases controlled by dual exponential signaling subnetworks (Network SigX2
Fig. 60 . The right hand Figure (b) shows 3 metastases controlled by the same signaling mono
exponential network (Network SigX1 Fig. 56).
The green and blue cells (Fig. 68 (a) above) are communication partners in the control paths
of the network but, while they potentially differentiate to exponential yellow and red cells, are
passive unless they receive the right signal from each other. The red and yellow cells divide
exponentially but are limited to the differentiation potential of their parent cells, the blue and
yellow cells, respectively. Note the irregular growth.
The orange cells (Fig. 68(b) divide to form two blue cells. The blue cells are passive receiver
cells until they receive a signal from a yellow sender cell. Once the blue cells receive their
signal they differentiate into active dividing orange cells. Yellow sender cells and receiver
Eric Werner, A general theory of cancer
87
blue cells interact by means of a signaling protocol (Network SigX1 Fig. 56) that regulates
the production of the proliferating orange cells. The yellow cells in the center of metastases
are purely signaling cells without proliferative potential. The blue cells only differentiate to
divide in the context of the yellow signaling cells. Thus, even though they do not proliferate,
the yellow signaling cells are the drivers of this type of cancer. Here the growth is more regular
with fewer distinct differentiation control states.
Thus, Fig. 68 illustrates different growth characteristics of metastases resulting from two different cancer signaling networks. Both types of metastases in the two figures illustrates that
we can have signal induced metastasis if the appropriate signal reaches any of the invading
cancer cells. Both networks require that an interloper A cell has traveled to the a region or
context of cooperating partner cells B. However, the networks have different dynamic growth
and differentiation characteristics.
The the network (Fig. 68(b) on the left side initiates metastases when the interloper cell A or
context cells B are activated by an appropriate signal. Yet, even here the doubly exponential
network does not produce exponential growth because growth is still dependent on a signaling
loop between neighboring cells.
The second network (Fig. 68(b)) on the right side, has a more uniform phenotype than the more
complex metastases generated by the dual exponential signaling network.
15.3
Moving signal-based metastases
In any signaling cancer network where the sender cells stay constant or not increasing in number, any metastasis would move to a different region if the sender cells move. For example,
the mono exponential signaling network shown in Fig. 56, that controls the yellow signaling
cells in subfigure 68b, does not generate any new sender cells. Hence, without an active other
external network that generates further sender cells, the sender cells will not increase in number. Thus, because the yellow sender cells stay constant in number yet can activate this type
of cancer network in a receptive cellular context, we could have the phenomenon of mobile
moving cancer areas or moving metastasis that stop in one area but then start anew in another
area because the underlying sender cells have moved to new tissue.
15.4
Context dependence of signal based metastasis
Cancer cells controlled by signaling cancer networks cannot form metastases independently
without cooperating partner sending cells with a fitting communication protocol. Thus, they
require a preexisting context of complementary, passive cancer helper cells. In an unresponsive
context without communicating helper cells, a potentially metastatic cell would lay dormant
until a fitting cellular context together with an initiating signal became available.
Given the context dependence of these kinds of cancers, it is possible that cancer cells with
metastasizing potential may be spread more widely through the organism than is apparent
Eric Werner, A general theory of cancer
88
from just the presence of explicit cancerous growths. Many such potentially cancerous cells
will just remain inactive in cellular contexts that do not have the partner signaling protocol
available. However, as conditions change, such potential cancer cells may suddenly become
active.
15.5
Stochastic signal induced metastasis
Another possible avenue of signal induced metastasis is if there is stochastic signaling where
cells with low probability send an proliferation inducing signal to their neighbors. Yet, another
possibility is that these sender cells could form by an occasional stochastic dedifferentiation
transforming a passive receiver cell into an active sender cell.
15.6
Stopping signal induction of metastases can have side effects
One way to prevent metastasis from developing for this kind of cancer network is to inhibit the
inducing cell signaling. However, this may have side effects if the signaling system is used for
normal cellular coordination.
15.7
A Hierarchy of metastases formed by geometric cancer networks
Geometric cancer stem cell networks form a hierarchy of interlinked stem cell networks such
that an n-th order stem cell network generates a cell controlled by an n-1 order stem cell network. Thus, 3rd order or meta-meta-stem cell networks produce cells controlled by 2nd order,
meta-stem cells. A 2nd order or meta-stem cell generates a cell controlled by a 1st order, linear stem cell network. At the base of the hierarchy is a terminal network controlling a cell
generated by a 1st order, linear stem cell.
This hierarchy of control in higher order stem cell networks has direct consequences for the
types of metastases that can, in principle, be generated by such higher order geometric networks. In other words. the metastatic potential of a cancer stem cell is determined by the
order of its geometric control network. As we will see, this the dynamic phenotypic properties
of the metastases are a result of their place in the metastatic hierarchy. This can be used for
diagnosis and reverse inference as to the type of stem cell network controlling the metastatic
tumor.
Eric Werner, A general theory of cancer
(a) Terminal metastases from a primary
1st order geometric cancer network.
89
(b) 1st order and terminal metastases from
a 2nd order geometric cancer network.
Fig. 69: Two views of metastasis by 1st order and 2nd order geometric networks. In
Figure (a): One primary 1st order geometric, linear cancer cell (in red) generates terminal
metastatic cells (in blue). But the blue terminal cells are terminal generating no further cells.
The right hand Figure (b): A 2nd order geometric cancer network generates 1st order geometric
and terminal metastases. Against a background of identical cells (in light pink) there is one
primary 2nd order geometric cancer cell (in red) that generates metastases consisting of 1st
order geometric cancer cells (in green). These in turn, generate secondary metastatic passive
terminal cells (in blue). The blue terminal cells may metastasize, moving to other regions, but
they do not generate further cells.
A 2nd order geometric cancer cell generates 1st order linear cancer cells (see Fig. 32 and
Fig. 35). If these 1st order linear cancer cells metastasize then we have a phenotype of multiple
linear cancer growths in different regions of the body. For example, one might have multiple
slow growing cysts in different regions of the body. In this case, these cysts are metastases that
were generated from a single meta-stem cell that generates progenitor cells that generate cysts.
Other examples, might include slowly progressing cancers that move in the lymphatic system.
Here again in may, but need not, be the case that a single meta-cancer stem cell is generating
sub-cancers that have the overt phenotype of a uniform cancer progressing in some direction
across a region or through the whole body.
Eric Werner, A general theory of cancer
15.7.1
90
Secondary and tertiary metastases generated by 3rd order cancer networks
(a) 1st oder and terminal metastases
from a primary 2nd order geometric
cancer network.
(b) 2nd, 1st order, and terminal metastases from a primary 3rd order geometric cancer network.
Fig. 70: Comparing metastases by 2nd order and 3rd order geometric cancer networks.
The left hand Figure (a): A 2nd order geometric cancer network (Fig. 32) generates 1st order
geometric and terminal metastases. Against a background of identical cells (in light pink) there
is one primary 2nd order geometric cancer cell (in red) that generates metastases consisting
of 1st order geometric cancer cells (in green). These in turn, generate secondary metastatic
passive terminal cells (in blue). The blue terminal cells may metastasize, moving to other
regions, but they do not generate further cells. In the right hand Figure (b): A primary tumor
controlled by a 3rd order geometric cancer network generates 2nd order and 1st order geometric
and terminal metastases. The primary 3rd order geometric cancer network (G3) generates 2nd
order (G2) metastases. These G2 networks in turn generate 1st order (G1) metastases. The 1st
order geometric (G1) cancer cells in turn generate terminal (G0) cells that are also potentially
metastatic (invasive) but do not generate further cells.
In general, n-th order geometric cancer networks Gn (see Fig. 33) generated cells controlled
by n − 1 order geometric cancer networks Gn−1 . These Gn−1 networks, in turn, generate cells
controlled by n − 2 order cancer networks Gn−2 , etc. Therefore, third-order geometric cancer
networks G3, such as the above simulated network in Fig. 70, generate cells controlled by
second-order cancer networks G2. In turn, G2-networks generate cells controlled by 1st-order
geometric networks G1. G1-networks are linear cancer networks which generate terminal cells
G0 that do not proliferate.
Eric Werner, A general theory of cancer
91
(a) Active cells in 1st oder metastases (b) Active cells in 2nd and 1st order metastases generated by
from a primary 2nd order geometric can- a primary 3rd order geometric cancer network.
cer network.
Fig. 71: View of the active proliferating cells in G1, G2 and G3 geometric metastases in
Fig. 70. Cellular proliferation alone does not distinguish higher order from lower order geometric tumors including metastatic tumors. Only terminal cells of type G0 can be distinguished
since they do not divide. Rather the control state and downstream proliferation potential distinguishes geometric networks from each other.
15.7.2
Relating metastatic phenotype with geometric cancer networks
Hence, since any of the cells generated by a cancer network have the potential to metastasize
by moving to other regions of the organism, each primary tumor controlled by a cancer network will have a distinct metastatic phenotype. Therefore, we can use the phenotype of the
metastases to draw inferences about the nature and architecture of the network generating that
phenotype. Moreover, if we find the metastases that together form related patterns as described
for the geometric metastases above, it warrants looking for the primary tumor generating the
secondary and tertiary metastases.
15.7.3
Treatment options for metastases generated by geometric cancer networks
Treatment would focus on eliminating or transforming cells with the higher level control networks generating the less dangerous lower level networks. Treatment can be by cell death
or by network transformations that block or modify the cancer network. Different transformations are necessary for different cancer networks. Once metastases have been created by
Eric Werner, A general theory of cancer
92
a higher level network, it is no longer sufficient to block or transform the higher level initiating subnetwork. Instead, each of the cells with lower level cancer subnetworks also have
to be transformed or destroyed. If they are 1st order linear networks, they may be relatively
harmless since they only produce cells that no longer proliferate. However, this will depend
on their location and effect on the other multicellular contexts in which they reside. Because
of cell interactions such as cell signaling, stochastic reactivation of higher level cancer networks such cells, depending on their differentiation state and the cellular context, may still be
dangerous.
15.8
Stochastic reactivation of terminal cells in geometric networks
If terminal cells can reactivate a proliferative control state in a network, then the more numerous
the number of terminal cells the more likely it is that one or more of them will become an active
cancer cell. Hence, given the possibility of stochastic activation of ancestral cancer networks,
formerly passive metastatic cells can become active tumors.
15.9
Metastasizing signal-autonomous exponential cancer networks
Metastasizing cells controlled by exponential cancer networks can be curable (e.g., a signalautonomous mole pregnancy), persistent but manageable if signal dependent on particular partner cells (e.g., Fig. 68 and Fig. 61), or fatal if resistant to treatment (e.g., resistant self-signaling
exponential cancers, 62a or resistant signal autonomous exponential cancers, Fig. 18). The resistance of a cancer to drugs or other treatment may depend on other factors and molecular
pathways and networks somewhat independent and orthogonal to the cytogenic control networks. The control network theory of cancer offers possible ways to stop resistant cancers by
directly attacking the control network topology.
Eric Werner, A general theory of cancer
93
Fig. 72: Metastases of a signal-autonomous exponential cancer network. All metastases
have exponential proliferation potential. Hence, they can generate further metastases with the
same exponential potential. This different from geometric cancer networks that generate cells
of a lower proliferative potential.
16
Network competence and performance in real continuous physical systems
We distinguish network competence and ideal growth from its actual performance in real spacetime. A network’s competence is the ideal synchronous growth of cells in a discrete space with
discrete physics with no friction or other properties that would constrain growth. Network
performance involves the dynamic growth of cells in a continuous space governed by the laws
of continuous physics. Linguists will recognize the analogy to Chomsky’s distinction between
linguistic competence as given by an ideal generative grammar and a speaker’s actual performance (Chomsky [6]).
Similarly, the (ideal) competence or capacity of a generative network exhibits its essential
mathematical properties unencumbered by the vagaries physical and temporal contingency.
The performance of the network may vary depending on the conditions in a realistic spacetime cell physics-in the context of a dynamic physical environment of other growing cells.
The ideal competence of a network shows the ideal, abstract topology of its developmental
possibility space (phase space) and the boundaries or limits of its growth potential. The real
performance with real continuous physics will in general be less than the ideal competence or
capacity.
This is not to say that the ideal competence is not complex, for the social, communicative
and cooperative interactions between cells will also be part of the ideal performance of cells
controlled by a proliferative network. This means we have to distinguish between ideal performance and real performance. The measures we apply here apply to the ideal performance of
cancer control networks.
Eric Werner, A general theory of cancer
16.1
94
Ideal developmental networks in real contexts
Bodily resources and cell physics are mitigating factors that limit the growth of cancers. Hence,
our ideal cancers are like the laws of Newtonian mechanics, being limited by friction and other
circumstances. In spite of this it is important to understand the nature of cancer in an idealized
developmental environment if we are to decipher cancer networks and thereby potential cancer
targets. These and other mitigating factors will be discussed.
16.2
Cell physics
Cell physics becomes important when considering diagnosis of a cancer. A repeating structure under pressure will not necessarily look the same as the same structure grown in a noncongested environment. Furthermore, inter-cellular pressures will cause such structures not
only to distort, but to break up leading to a confused cell mass of different tissues. And, this is
something that is actually observed in many tumors.
16.3
Invasive cancers
Cancers can become invasive when the cells change their physical properties or by pressure
break out into the vascular system. If they are cancer cells then they are distributed by the
blood vessels or lymphatic system to other organs and areas of the body.
16.4
Tumor vasculature
Many interactions between cells appear to from genome autonomous cell strategies that control
cell communication and cell actions. Even though cell strategies are constructed using control
information in the genome, once created they may have a life of their own not directly controlled by that genome. Vascularization is a process that appears to be largely independent of
developmental control information in the genome. It is what is referred to as a developmentally
regulated rather than a mosaic developmental process. A regulated development process is flexible relying on cell communication and cell interaction strategies, while mosaic development is
inflexible driven without cell signaling. Vascularization appears to be regulated development
being controlled by cell to cell interactions including cell communication strategies. It results
in the formation of blood vessels to feed a growing tumor. However, even though vascularization requires particular conditions to be initiated, such as anoxia (the lack of oxygen) and even
though it may involve genome autonomous cell signaling strategies to spatially and temporally
coordinate blood vessel formation, there may still be genomic developmental control networks
involved in the process. These developmental control networks need to be activated for cells
to divide and proliferate.
Eric Werner, A general theory of cancer
16.5
95
Multiple parallel networks
Developmental networks or cenes cooperate with the cell’s interpretive executive system (IES)
to generate cell actions such as cell division and cell communication. The IES is a complex molecular system that contains lower level networks that interact with the higher level
controlling developmental networks. However, we suggest that the architecture relating these
networks is something akin to a subsumption architecture (Brooks [4]) where there is local
reactivity in the lower level IES but global control resides in higher level developmental networks. Many of these networks act in parallel with in the cell. Within the multicellular system
we have vast multi-parallel developmental networks that interact via cell signaling protocols.
The cells are then agents with local strategies implemented by the state of their developmental
network and the state of their IES.
Cancer can be caused by either a transformation in the developmental network alone, or it can
result from mutations in the IES or its subsystems. Thus if a cancer causing mutation affects
only cis catchers or trans pots of the genome it is pure developmental cancer. If the cancer
results from a mutation in the cell cycle or mutated receptor while the cenome is left intact,
it is a cene autonomous cancer. And, we can have the mixed cases where mutations in the
cenome combine with IES mutations to give mixed control cancers.
16.6
Relations between gene-based and cene-based networks
Thus, if the IES is largely implemented by gene-based molecular systems then gene-based cancers may exist that are largely cenome autonomous. Pure gene-based cancers are conceptually
different from cene-based cancers even though their dynamics would still fall under the general
architecture and topology of cancer control networks. Many cancers are a mixture of gene and
cene mutations. For example, gene mutations may increase the invasiveness of a cene based
cancer thus contributing to the cancer’s metastatic potential.
Viewed more abstractly where cell’s are agents with various strategies, then a cell has a local
strategy of behavior and it has a global strategy for multicellular development. These strategies
run in parallel and interact with in the cell. For a multicellular system of cells the gene based
local strategies for cell behavior run in parallel and cooperate with the global cene based strategies of development. Each cell may be in a different local and developmental state. Things can
get quite complex and software modeling and simulation is essential for understanding complex multicellular developing systems such as embryo genesis and cancer.
Both the gene based and the cene based strategies are forms of control. Hence, the dividing line
between cenes and genes can be fuzzy. But overall developmental networks (cenes) separate
out from local gene based networks.
Eric Werner, A general theory of cancer
16.7
96
The addressing architecture and control
There is a further complication in that catchers and pitchers such as pots may be implemented
by more than one addressing architecture. Thus we may have an addressing architecture based
on additive but unordered protein transcription factors and we may have more a more precise
ordered addressing architecture based on RNA, DNA or RNA in held by proteins such as
Argonaut. Cenes may be implemented by either or any of these addressing systems. The local
control within IES and interactions between the IES and the cenome may involve both protein
and RNA addressing systems (Werner [45]).
17
Network classification of cancers
Using the properties of the above networks types, we can classify cancers according to their
phenotype as predicted by their cancer network architecture or by the network itself. A phenotypic classification can then be used in reverse for diagnosis where the cancer network
is inferred from the tumor phenotype. Furthermore, using the network-phenotype mapping,
treatment strategies can then be based on a better knowledge of the underlying cancer network.
We have attempted to provide a theory of all possible cancers giving a kind of periodic table
of cancers. What is striking is the variability of possible cancer networks. We have a multidimensional space of cancer networks. At the core are the developmental subnetworks that
control cell division. These can be mixed and matched with reactivity to context, cell signaling,
stochasticity, developmental position (via network linking within the global developmental
network that controls the temporal and spatial ontogeny of the organism), and the physics and
chemistry of cells.
The dimensions of the multidimensional space of cancer networks include: Cytogenic capacity
or proliferative potential (linear, exponential, geometric), reactive/nonreactive, cell communication, stochasticity, bilaterality, early-mid-late developmental stages depending on where the
network is located (including degree of differentiation, complexity of network-MCO, ranging
from undifferentiated, totipotent, pluripotent, differentiated, terminal, teratoma, dermoid cyst,
fetus in fetu, invasiveness (metastasis, metastatic potential, interactivity with normal tissue,
e.g., teratocarcinoma, invasive metastatic teratoma), metabolic control networks, the cell cycle
network, cell physics and chemistry.
17.1
A higher dimensional table of cancer types
The following table is a summary of all possible dimensions of cancer types and stages. A
cancer can have any or all of the properties of the dimensions listed in the vertical column of
the table.
Eric Werner, A general theory of cancer
97
Together these dimensions of cancer give rise to a large higher-dimensional space, a combinatorial space or manifold of possible cancer network types and their possible implementations.
Each possible cancer has its own phenotype reflecting the cooperative interaction of its dimensions in its formation. The ontogenic phase of a cancer refers to where in the normal developmental network the cancer is situated. It distinguishes, for example, embryonic, childhood,
adolescent, adult and late life cancers. The proliferation rate depends on the extent of the cancer loops, the degree or complexity of differentiation of the cell types, and the morphological
complexity of the developing multicellular structure (e.g., teratomas, fetus in fetu).
The space of possible cancers is roughly divided into three sections: The network topology section describes the range of possible topologies of cancer networks. The network entropy section
combines features that make the developmental path the network takes more or less uncertain.
Stochasticity associates probabilities with network links making activation of downstream networks uncertain. The cell signaling subsection is makes the outcome of a network dependent
on what signals the cell sends and receives. Hence, depending on the multicellular context, the
path taken in the network can vary. At the same time cell signaling is used to coordinate the
actions of cells, making the outcome more robust. The conditional subsection is to distinguish
cell signal dependent conditionality and entropy, from external environmental conditions such
as temperature that can effect what path is taken in a developmental network.
Eric Werner, A general theory of cancer
98
Space of Possible Cancer Networks and their Phenotypic Properties
Dimensions
Effect on Phenotype
Linear L
Slow growth -one founder cancer stem cell, the rest terminal cells
Geometric
Dynamics depend on number of loops -meta-stem cells with varying
potential
Exponential
Very fast growth -all cancer cells have equal proliferative potential
Determistic
Development is not stochastic with no network entropy and, hence,
no developmental entropy resulting from the network topology
Stochastic
Development is probabilistic with developmental entropy
Signal Autonomous
Development is independent of cell signaling with no uncertainty or
diversity generated by signals
Signal Dependent
Development is dependent on cell signaling -alternate developmental
paths possible
Conditional
Development is conditional on some property external or internal to
the cell.
Net Locality
Locations where the cancer net links into global net - Cell differentiation states, complexity and maturity of structure depends on network
location via links to and from the global network
Net Extent
The length of the range a network covers with its cancer loops - Number of cell types and complexity cell of structure depends on network
range and location in the global network
Vacuous Nets
Unconnected networks not linked to the active global network controlling development - The network phenotype is hidden showing
having no effect on development. But, has the potential to be activated by cis or trans mutations
Topological
Degree of complexity of net topology - The complexity of the network determines the possible complexity of the multicellular dynamics and phenotype
Node Complexity
Node complexity affects states - The complexity of the nodes has
effects on multicellular phenotype orthogonal to network topology
Physics
Cell physical properties orthogonal to
cenes
The connective physical properties of cells and their matrix - Physics
plays an essential role on the phenotypic effect, the space-time development, generated by a developmental network
Cell Properties / IES
Cell strategies orthogonal to cenes
Vascularization strategies, invasiveness, acidity, connectedness, flexibility, strength, IES = interpretive-executive system
Network Topology
Network Entropy
Inter-Network Links
Net Complexity
Table 2: Dimensions of the space of all possible cancers
Eric Werner, A general theory of cancer
99
The Inter-Network Links section places a network within the cenome (the global developmental control network of the genome). If the cancer network is located early in the cenome and its
extent is short we get an immature cell mass that may proliferate more quickly than one that occurs later where more complex cell types and structures are developed. Furthermore, the extent
of the network determines how many steps are involved in a cancer loop path. The longer the
path and the more complex the cell phenotype, the slower the execution of the cancer network.
Vacuous networks are included for the sake of theoretical completeness. Though there may
exist such, perhaps archaic networks that may be activated under stressful conditions.
The Physics section is yet another semi-orthogonal category. Cell physics makes a substantial
contribution to real as opposed to ideal development. There is an interaction between the
activated genes used to structure the cell and the cancer network that in part coordinates the
activation of those structural and functional genes.
The Cell Properties/IES section is another semi-orthogonal dimension: The IES, the
interpretive-executive system, that interprets and executes the information in the developmental control network is somewhat autonomous from the network itself. The very same network
can give very different results depending on the interpretation and execution.
Included in cell properties are its strategies that are somewhat autonomous from its developmental control networks, such as vascularization, invasiveness and other gene based properties
of the cell. Recall, we distinguish the cell differentiation genetic pathways that don’t result in
multicellular proliferation from from developmental control networks that do result in multicellular genesis.
17.2
Request for comments
Representing this multi-dimensional space of possible cancers graphically is not easy. There
are many ways to do it. Each has its advantages and disadvantages. Any suggestions from you,
the reader, are welcome.
One possible and provisional way of making these dimensions more formal is the following
notation.
Notation: Xn = Exponential hyper cancer of degree n, Gn = Geometric cancer of order n,
L = Linear cancer, E = Early, M = Mid, L = Late, F = Fast, ME = Medium, S = Slow,
U = Undifferentiated, SD = Semi-differentiated, TT = Terminal, MSC = Meta Stem Cell,
SC = Stem Cell, PR = Progenitor cell, I = Invasive, V = Vascularization, DD = Degree of
Differentiation, Y = Yes, N = No.
Dimensional ranges: Conditional = Signaling = Stochastic = Bilateral = Vacuous = [Yes, No],
Teratoma = [Multiple cell types, Dermoid cyst, fetus in fetu], Ontogenic phase = [E, M, L],
Proliferation rate = [S, ME, F], Degree of Differentiation = [Immature, Mid, Late, Terminal],
Cell types = [U, SD, TT , MSC, SC, PR], Stage = [I, II, III, IV].
Eric Werner, A general theory of cancer
18
100
Network theory and the underlying cause of cancer
Many networks can lead to cancer. One can go on to construct arbitrarily many varieties of
cancer networks based on this basic theory. The key is that cancer may occur when a given
genomic network is short-circuited by a backwards loop that activates an earlier site in the
network. This can then lead to the dedifferentiation of the cell to this earlier state. Then the
cell starts developing again according to the network at this earlier activation state until one of
the resulting developmental paths inevitably reach the loop point and the process repeats itself,
endlessly.
18.1
Network theory and gene mutations
What causes cancer? Perhaps, with cancer, as with developmental biology, the focus on genes
(protein coding sequences) has prevented a deeper understanding of the mechanisms underlying cancer. Based on our theory, we can now give a more detailed answer than merely saying it
is caused by a mutation of an oncogene. In fact, such an explanation is circular since an oncogene is by definition a cancer causing gene. On the other hand, if we say that cancer is caused
by one or more mutations, it does not explain why some mutations of genes cause cancer and
most mutations do not cause cancer. We suspect that for many, if not most, cancers there are
no genetic mutations that cause it. Even for those mutations that are caused directly by genetic
mutations, our theory provides an explanation not only why a mutation can cause cancer but
precisely under what conditions and how it causes the cancer.
18.2
How mutations lead to cancer
Mutations that lead to cancer transform the developmental control network of a cell. Developmental networks are based on an addressing system that allows precise control of activation
of control states in the cell. The transformation are the result of a change in the addresses
such that looping occurs in the control network. When these loops are nonfunctional and lead
to endless cell proliferation, cancer is the result. We saw that adding further loops to a linear cancer network can convert it to a geometric or exponential cancer network. Hence, more
mutations can lead to more severe cancer phenotypes.
18.3
Viruses and cancer
Because our theory of developmental evolution postulates that viruses may have be one factor
in the build up of the genome and that these repeats were subsequently appropriated as basic
control areas linked together into networks, it becomes evident that viruses could change the
addressing architecture of genomes and in particular generating specific control networks that
are possibly cancerous.
Eric Werner, A general theory of cancer
18.4
101
Aneuploidy and network theory
Some argue that too many or too few chromosomes in the nucleus (aneuploidy) is a primary
cause of cancer. It is argued that the probability of cancerous mutations is orders of magnitude
too small to explain the high incidence of cancer in human and animal populations. Instead
aneuploidy (chromosomal aberrations in number and structure) is a much more probable phenomenon. Furthermore, aneuploidy, being associated with a larger than normal cell nucleus,
is a common phenotype of aggressive cancers. The cell then becomes unstable and the normal
cell cycle goes awry.
One problem with this hypothesis is that even if true, it provides no explanation as to the
regulatory mechanisms that control and generate the cancerous tissue. Instability by itself does
not explain the distinct phenotypes seen in different types of cancer. The regulatory control
system still needs to be understood.
Some exponential cancers are not associated with aneuploidy. Other exponential cancers do
satisfy the hypothesis. For example, in mole pregnancies (hydatidiform mole) which is, in the
form of choriocarcinoma, an exponentially proliferating cancer, the female haploid genome is
missing in the fertilized egg.
We have two seemingly contradictory takes on the aneuploidy hypothesis:
First, aneuploidy may the be result of cancer and not its cause. If the cancerous control system
and its associated network issues commands too quickly for the cell to execute then chromosomes may be duplicated without the cell physically dividing. The result is an excess of
chromosomes or aneuploidy. In other words, aneuploidy could result if the control system initiates cell division when the cell is not large enough (does not have the parts or resources) to
divide. When the command to divide is given, chromosomes are duplicated, yet no physical
cell division occurs because the cell is not ready. The control system assumes physical cell
division has occurred and initiates post-divisional activation of the genome. However, since
the chromosomes are now partially or totally duplicated, we may get dual activation of both
daughter chromosome sets. This may lead to differentiation ambiguities with corresponding
phenotypic mixtures. As a result, we may get mixed phenotypes in the cell that normally would
be distributed between the potential but nonexistent daughter cells that were supposed to have
been generated by the non-occurring division.
Thus timing problems can make a mess of the cellular structure and also the resulting control
system. However, there is an alternative explanation for the generation of aneuploidy in a cell.
It can also occur because a mutation (sequence, deletion, or shift) causes the cell division command sequence to be skipped, jumped over or not executed by the control system. The result
of such a mutation is the same as with the timing problem above. We again get duplication of
chromosomes without the normally subsequent physical cellular division.
Whether aneuploidy by itself leads to cancer is independent of the above two possible explanations of the generation of aneuploidy. The first explanation stated that cancer leads to
aneuploidy when the control network is too fast for the cell to keep up with its directives.
Eric Werner, A general theory of cancer
102
Then it may partially or fully achieve the first phase of the cycle in copying the chromosomes ,
however it may not achieve the second phase where it actually duplicates and divides into two
daughter cells. This however, need not cause any additional cancer if no additional loops in
a cell proliferating are induced or activated. We just get more chromosomes not necessarily
more cells. In the case of plants chromosome duplications do not usually lead to cancer.
The second cause of aneuploidy is when a mutation functionally skips the step needed to initiate and complete physical cell division. Since it executed the step to duplicate the chromosome,
the result is aneuploidy. However, again, by itself this need not lead to cancer. Without further regulatory transformations that result in a regulatory and proliferating loop or activation
thereof, such mutations are not the cause of cancer.
18.5
Telomeres, aging and network theory
Cell division has been associated with reduction in the length of telomeres (the ends of eukaryotic chromosomes). Once the telomeres of the chromosomes are exhausted, the cell stops
dividing or cell death occurs. In some cancers telomeres are not reduced and it is postulated
that this leads to endless cellular division. This fits perfectly into our theory of cancer control.
Telomeres molecularly implement what is known as conditional loop. The conditional loop is
executed as long as some property (exit-property) holds. In this case, the cell is able to divide
as long as the telomeres are of sufficient length.
In many programming languages there is a directive that is called a for-loop, the process is
executed a set number of times and then stops (e.g., for value 1 to n = 10 do something).
In the normal case, telomeres give the cell the capacity to exhibit the behavior of a for-loop.
However, if some mechanism prevents the exit-property from ever being true then the loop
goes on forever (e.g., in effect making n infinite in our above for-loop example).
In the case of telomeres, the counting-property that permits the cell to exit the proliferating loop
after a certain number of loops is the finite number of telomeres, plus the reduction condition
that at each cell cycle the length of the telomeres is reduced by one and the exit or stopping
condition for cell division when the length goes below some value. Since, in some cancer cells
the telomeres are not reduced or added again after division, the exit property is never fulfilled
because the reduction condition is never realized. Hence we have an endless proliferating loop
which results in cancer.
Note, however that while telomeres may be one form of control over cell division by themselves
they do not necessarily cause cancer. If a genome in a cell does not contain and does not
participate in a cancer network via cell communication, then even if the telomeres are intact
and of sufficient length or numbers, then that cell may not be cancerous. Thus the existence
of telomeres is not a sufficient condition to generate cancer. At most their existence may be a
necessary condition for the execution of a cancer network. Therefore, for a cell to be cancerous
it needs not only to preserve its telomeres, but most importantly it needs a cancer network that
is active and that the cell executes.
Eric Werner, A general theory of cancer
18.5.1
103
Aging, Cancer Networks and Regenerative Medicine
Related to this is the problem of aging. If we want to prevent aging we must prevent cells
death. Hence, we need to regenerate cells endlessly. However, that may lead to endless loops
which may be cancerous. To prevent aging we have to create networks that are regenerative
without being cancerous. Therefore, regenerative medicine presupposes a deep understanding
of cancer networks and their relation to stem cell networks.
19
Curing cancer
The cure for a particular cancer then lies in either somehow blocking the causative loops,
stopping the cell cycle, promoting apotosis (cell death) of the cells controlled by the cancer
network, reversing the mutation by an inverse network transformation that converts the cancer
network into its original natural noncancerous network state.
Different theories of cancer offer different methods for the treatment and cure of cancer. Previous attempts to explain cancer have failed to understand the inherent network properties of cancer. As a result, possible treatment methodologies based on such theories are practically blind.
Therefore, such theories can offer only treatment methods that have inherent, often debilitating,
side effects. They don’t eliminate the root cause and instead either kill too many noncancerous
cells or not enough cancer cells or both. In particular, drugs based on gene centered theories
of cancer can lead to both overkill and under kill because the gene that supposedly causes or
prevents a cancer has functions in normal tissue. Therefore, activating or deactivating that gene
can have significant consequences for normal multicellular developmental processes (e.g., tissue damage, developmental defects, other cancers). Surgery and radiation are brute force and
cannot treat metastases or otherwise inaccessible tumors. Chemical agents that attack the cell
cycle attack all cells that use the cell cycle (e.g., hair loss, intestinal damage, possible new
cancers).
1. Blocking any point in the cancer loop
2. In the case of cell signaling cancer networks, blocking cell signaling either by blocking
receptors, blocking signal generation, interfering with the signal transduction pathway
that leads to activation of the cancer subnetwork
3. Stopping the cell cycle
4. Using genetic properties or markers of cancer cells to kill those cells that express those
properties
5. Forcing apotosis (cell death) of the cells controlled by the cancer network
6. Reversing the cancerous network mutations by a series of inverse network transformations that convert the cancer network into the original natural noncancerous network
Eric Werner, A general theory of cancer
20
104
Comparison of Traditional Theories with our Control Network Theory
20.1
The cell cycle and cancer
The standard explanation of cancer is that there is uncontrolled growth with a failure of apoptosis (cell death). Here uncontrolled growth means that control of the cell cycle is somehow
evaded. Prior to division, cells go through several stages with checkpoints before each stage.
This is known as the cell cycle. Cell signaling is assumed to somehow control progression
through the cell cycle by way of influencing these checkpoints (Tyson [31, 34, 32]). For example, a signal to progress to cell division may be constantly switched on, leading to endless
cell division. Or, a tumor suppressor gene may be may be deactivated by a signal or mutated
leading to uncontrolled growth.
The problem with these accounts is that they are at best partial explanations of the initiation
or control of some particular instances of cancer. They do not explain the dynamics of cancer
development and phenotypes. They do not explain different growth rates in cancer nor the
fact that cancers can generate multiple cell types and go through diverse stages of differentiation. For example, it does not explain how teratomas differ from mole pregnancies or other
differences in the ultrastructure of cancer temporal and spatial phenotypes.
20.1.1
Higher level developmental networks control the cell cycle
The fundamental problem is that the cell cycle view of cancer fails to see that cancer is a
developmental process similar to embryonic development. In particular, proliferative cancer
networks with their loops are not the same as the cell cycle. Instead, the cell cycle is controlled
by developmental control networks. In other words the cell cycle molecular system is a lower
level, cellular control loop that is a sub-network meta-controlled by higher level, developmental
epigenetic and genomic control networks. These higher level developmental networks control
the cell cycle by activating or inhibiting the cell cycle at various checkpoints.
20.2
Problems with the traditional gene-centered theory of cancer
Recall our comparison of the gene-centered theory of cancer with our network control theory
of cancer in Table 1.
First, there is no systematic causal relationship between phenotype and genetic mutations. The
reason is because the network controlling the activation of genes is left out of the picture.
Second, a deeper problem is that the theoretical and explanatory level is too low, omitting the
essential factor that causes cancer. For example, a broken transistor in may cause the computer
to not work properly. However, this does not explain why it malfunctions the way it does.
Being told a transistor is broken does not in itself have predictive power; we cannot predict
what the future the behavior of the computer will be. A predictive explanation requires that
Eric Werner, A general theory of cancer
105
we understand the role of the transistor in the overall architecture of the computer, its control
network. Similarly, knowing a gene or non-coding area DNA is mutated will have limited
predictive and explanatory value, unless we understand the role of that gene or area in the
overall control architecture of the cell. Phenotypic effect depends on the role of the broken unit
in the overall network.
20.2.1
Knudson’s two hit model
The idea of Knudson [16] is that there may be a predisposition to cancer if one of two allelic
genes is non-functional because of a preexisting mutation at birth. This thereby makes the
cancer more likely because only one more mutation in the other complementary allele/gene is
required to make the gene nonfunctional. If the gene is a tumor suppressor genes then a cancer
may develop if both alleles of the gene are knocked out by mutations.
We can see there are several assumptions in Knudson’s theory of cancer.
1. It assumes that genes, and not networks, are the root cause of cancer.
2. These genes cause cancer if they are not functional because of mutations.
3. Hence, cancer is the result of the lack of suppression of a process of cell proliferation.
4. Consequently, all cancers are caused by tumor suppressor genes.
Knudson’s model might be used to try to explain the difference in growth rates of cancers. If
we link the degree of suppression with the number of genes suppressing the cancer, then we
might link the growth rate with the number of suppressed genes. If only one gene is suppressing
growth, the rate of growth may still be inhibited. If both alleles of a tumor suppressor genes are
disabled then the growth rate might be twice as fast than if only one gene is disabled. Hence, if
only one tumor suppressor gene is mutated we might get only partial suppression and hence a
slow growing tumor. If both genes are suppressed we can have a faster growing tumor.
However, this would not explain why we get exponential growth, since the suppression of the
second gene would, presumably, without other assumptions in the theory, only double the rate
of a linear cancer making it twice as fast but not exponential. Thus, Knudson’s theory cannot
explain why some cancers exhibit linear growth and others are exponential, let alone geometric.
All cancers would be either linear, linear by a factor of two, with no explanation of why some
cancers are inherently exponential or geometric.
This failure of explanatory power is a result of the inherent limitations of atomistic nature of
Knudson’s model. It cannot not have the explanatory capacity of the network model we are
proposing because there is not enough ontological structure in Knudson’s theory.
Cancer, on Knudson’s model, is seen not as an active developmental process but rather as a
result of an insufficiency, a lack of a suppressing component. As a consequence, Knudson’s
model cannot account for the differences in phenotype of cancers. His theory can only be used
Eric Werner, A general theory of cancer
106
to explain the existence or nonexistence of a cancer that is suppressed by genes, but it cannot
explain the developmental dynamics and phenomenological nature of that cancer.
20.2.2
Repressor genes and cancer
The network model we are proposing explains the underlying reasons why Knudson’s two hit
model works for some cancers. If a gene is suppressing a cancer network that would otherwise
be activated then one suppressor gene is sufficient to prevent the cancer from growing. If
both alleles of a suppressor gene are disabled by mutation then the network is activated and
generates the cancer. Different suppressor genes suppress different networks and hence we
have different genes associated with different cancers. Yet, note, that the actual formation and
dynamics of the particular cancer is the result of the network that is suppressed and not the
suppressor gene itself.
20.3
How does our theory differ from the classical account of cancer?
1. Dynamic and causal properties We are able to describe the actual dynamic properties
of cancers and how their networks influence these dynamics. The classical account only
describes the etiology of cancer, namely that cancer is caused by mutations in oncogenes. How oncogenes actually cause cancer is often left unsaid with some exceptions.
And those exceptions involve loops which confirm our theory. In fact, the accounts are
consistent with each other.
2. Targets -safer and unsafe A safe target is a cancer target that does not have major
developmental or other side effects. We see that the targets of cancers are points in a
network. However, we have to be careful, since preventing a cancer may also interfere
with a functional network. Our theory gives an account as to which points in a network
may be less subject to side effects if they are interfered with by some therapy. For example, if a cancer occurs in childhood then it may involve networks that are responsible
for the further development of the child. In such a case, the target to block the cancer
has to be chosen carefully, limited to those that have the least negative developmental
consequences.
3. Typology of cancer We have an entirely new view of cancers and their typology. We
classify cancers by their control network architectures and not by individual genes or
phenotypes. However, we also see that phenotype can help us in determining the kind of
cancer network that may underlie the phenomenologically and molecular properties of
the tumor.
4. Cancer is a form of multicellular development Cancer is just a type of multicellular
process. The looping conditions cause a normal developmental process to potentially
result in pathology. We view multicellular development as a multiagent process where
cells are active complex agents that have the capacity to perform various actions such
Eric Werner, A general theory of cancer
107
as cell division, cell communication and cooperation. A multicellular organism is a vast
distributed multiagent system where many of the agents are in different control states.
Control networks are executable by the cell. As we have seen, when control networks
have a certain architecture they can lead to cancer. We saw in the case of stem cells that
the dividing line between cancers and normal development may depend on additional
differentiation control networks that determine strategic properties such as motility and
invasiveness of the cell. A stem cell network that activates an non-normal cytogenic or
differentiation network is usually viewed as being cancerous.
5. General theory with broad applicability We have said nothing of the biochemical details of the loops. We have left open how the control network architecture is implemented
in the cell and its genome. It may be purely genetic (protein based), it may be a combination of protein and RNA, or it may be pure RNA control with protein helper agents.
That is something for experimentalists to resolve. However, the overall cellular-genomic
control architecture of development and its pathological sibling, cancer, is at a level of
abstraction that permits any of these scenarios. In spite of its generality and abstract conceptual basis, it has predictive and prescriptive power. It is a paradigm shift that will help
lead future research efforts to find the actual basis or biochemical implementation for the
network control architecture and topologies responsible for different types of cancers.
6. Abstraction has broader explanatory and predictive power Thus, we are operating
at a higher level of ontology and information that, while dependent on the lower levels
for its realization, exhibits regularities and laws that are difficult to deduce and predict
using only the ontological and conceptual machinery available the lower levels of the
ontological and informational hierarchies we see in living systems. And, this is at the
basis of the utility of investigating such levels of network control. They help us see patterns and crucial points of control in the cellular system that will help us in the discovery
of drug targets as well as the avoidance of the dangerous repercussions of ignorance of
such developmental networks.
7. A new research paradigm As a research paradigm it shows that we will have to pay
much closer attention to the control architecture of the cell and the control information
in the genome. Genome semantics is one way of describing the meaning of a genome in
the context of the cell as agent that interprets that control information and executes it in
the process of development. If it turns out, as we suspect from other theoretical considerations, that RNA plays a crucial role in combinatorial selection of genomic commands
to the cell, then our entire concept of the cell will change. Genes will be subservient
helper agents in executing the commands in the form of non-coding regions in DNA
and/or RNA generated from those non-coding regions.
20.4
The explanatory power of our theory and diagnostics
The types of possible cancer on our view can vary as greatly as the complexity of the genome
permits. Given an genomic network one could in principle calculate the combinatorial possi-
Eric Werner, A general theory of cancer
108
bilities or mutations or transformations which would result in cancerous loops.
20.5
Cancer growth rates
The network theory of cancer explains why some cancers are slow growing and others are very
fast.
It distinguishes between linearly, geometric, and exponentially growing cancers. Normally, an
exponential cancer will soon outpace a geometric or linear cancer. A geometric cancer will
grow more quickly than a linear one. Therefore, the growth rate may serve as a diagnostic
criterion to distinguish linear, geometric and exponential cancers. However, there may be
considerable variation in growth rates within these broad categories for several reasons:
1. Complexity of differentiation Prior to each cell division, the more complex the cell,
the more time each cell cycle takes. And, therefore, the slower the rate of growth of the
cancer be it linear, geometric, or exponential. In other words, before division can take
place, the cellular differentiation pathway may take more time if a cell differentiates from
one cell type to another such as from a simple to a more complex cell type. Furthermore,
more complex cells require more steps and energy to grow and divide.
2. Network complexity The rate of growth of a cancer also depends on the length and
complexity of the cancer pathway from the initiation point to the point that the network
loops back on itself. The longer the cancer pathway and the more complex the actions
initiated by that pathway, the longer it takes for the cancer cells to loop back to their
starting point. For nonlinear cancers there will be more than one cancer pathway. If for
example, two loops in an exponential cancer each generate a complex structure before
they loop back to their initial starting state, then that cancer may execute very slowly.
However, the cancer cells will still double at the completion of both loops. Hence, the
growth, though slow in each differentiation step, will be exponential over the loop count.
Eventually, it will outpace the fastest growing linear and geometric cancers.
3. Physics and contextual constraints Cell physics and the multicellular context of the
cancer will significantly influence its real versus its ideal growth rate.
4. Cell signaling constraints For social cancers depending on cell signaling, the growth
rate depends on the availability of communication system in the social context of the cell
in the multicellular system. If the appropriate communication is not there then either the
cancer fails to grow if activation of the network requires signaling or the cancer begins
to grow if signaling is required to repress the cancer network and thereby preventing its
growth. We saw an example (Fig. 61) of a doubly exponential cancer network that only
grows linearly because of limited available cell signaling between potentially exponential cancer cells.
Summing up, whether cancers are linear, geometric or exponential, their growth rates will vary.
Our theory predicts that cancers with slow growth rates will normally exhibit complex differentiation states, multiple cell types, and even distinct repeating morphologies whether they be
Eric Werner, A general theory of cancer
109
ultra-structure or larger tissue or organ like complexes. On the other hand, less differentiated
cancers that do not differentiate to more complex mature cell types, will grow faster because
the network to develop those cells will be shorter and hence the loop will execute more quickly.
However, the growth rate of a long cancer loop, in terms of production of cells over time, may
be same as a short cancer loop if the differentiation steps in both pathways are similar in complexity. In general, though, cancer networks with short control loops and simple differentiation
steps will be the fastest growing cancers. However, an exponential cancer will eventually outpace geometric and linear cancers if the patient lives long enough.
21
What cancer is not
Our theory differs in fundamental ways from previous theories and models of cancer. However,
it should be said at the outset that despite their limitations, previous models of cancer have
intrinsic value. For reviews and views of major cancer models see [5, 2, 1, 3, 21, 22]. What
we are trying to do is to is construct a general unified theory of cancer. The particular models
of cancer could then be instantiations, such as special cases of cancer, or be integrated with the
general theory by giving more detailed accounts of molecular implementation, mathematical
analysis, and physical realization and performance.
21.1
Physics-based models: Cancer is not just physics
A great deal of insight can be gained by viewing cells as purely physical entities subject to
the laws of physics. Cell behavior is modeled by a set of physical laws. The properties of
the cell can then be adjusted by changing the parameters of those laws and/or by changing the
boundary conditions under which the cell actions takes place. Parameters of the cell model can
include rate of cell division, cell spatial dimensions, various cell structural parameters such as
spring constants that may be used to model cell connections. More and more sophisticated cell
models have been developed including models that can give the cell a wide variety of shapes
and flexibility ([21, 20]). It is then hypothesized that cell physical parameters correspond to
genes or sets of genes. Changing a physical parameter is then equated with mutations in a
gene or set of genes. A simulated multicellular developing tumor is generated and compared
with a natural tumor. If there is a match then inferences are drawn that the model is confirmed
and that particular parameters that generate particular phenotypes in tumors are likely related
to mutations in certain genes. Such a methodology, of relating model parameters to genes
and then seeing transforms of model parameters as mutations of related genes, was applied
successfully to the study of bacteria [19].
Applying this methodology to multicellular development has intrinsic challenges. The physics
of a skyscraper is different than the architecture of that skyscraper. The physics does not
determine the architecture, nor does the architecture determine the physical laws. Trying to
model multicellular systems with just physics and with no developmental control networks is
like trying to build a skyscraper with no architectural plan. The difference is that developmental
Eric Werner, A general theory of cancer
110
control networks are not descriptions of an organism or system of cells. Rather, developmental
control networks are executable networks or strategies that when interpreted by the cell lead
to cellular actions. The combination of all those strategic networks in different states in a vast
multicellular system leads to development in the context of physics. While physical conditions
enforce constraints, strategic developmental control networks distributed across the cells in
a multicellular system, cooperate to generate ever more complex multicellular systems and
ultimately form an organisml. Putting physics in the lead puts the cart before the horse.
21.1.1
Physics-based models fail to explain the ontogeny of morphology
Many present models of tumor growth focus on cell physics ignoring the underlying developmental control networks. While the physical interaction between cells and the cell strategies
that cells use to respond to and control their interactions with other cells are an essential component in understanding cancer and development of organisms in general, cell physics misses
the root cause of cancer. Some attempt to explain pattern formation in development and cancer
purely as the result of the physical interaction between cells [22, 27] (see below). Such theories fail to explain the ontogeny of morphology and fine structure of dynamically developing
multicellular organisms. Since cancer is a developmental process such theories fail to explain
the ontogeny of structure and function in cancers as well.
21.1.2
Physics-based models fail to explain complexity in development
There are fundamental problems with all such approaches when they are applied to cancer or
embryonic development in general. One problem is the complexity of multicellular life and
the information required to generate complex living organisms from a single cell. Purely local
strategies of interaction between cells in a developing multicellular system (whether implemented by simple rules in discrete space-time or by a set of differential equations in continuous
space-time), do not contain sufficient control information to generate complex organisms [37].
It is one reason why reaction diffusion models of the kind proposed by Turing [30] fail to be
adequate models of development.
21.2
Rate-based models: Cancer is not just about rates of cell division
Physics-based models are usually conjoined with rate-based models of cancer. Limitations
of rate-based models lie in the assumptions they make about the nature of cell proliferation.
Almost all previous mathematical and computational models of cancer assume that cancer is
caused by differences in the rates of cell proliferation due to uncontrolled cell growth. While
partly true, this is not the cause of cancer, but an effect of the nature of the developmental
networks controlling the cancerous cells. Hence, previous models are limited to descriptions
of rates of cell proliferation over time using either differential equations or other mathematical
formalisms. Most such models limit the description of cancer cell proliferation to adjusting rate
Eric Werner, A general theory of cancer
111
of division parameters (also called cell cycling rates). These rates of cell proliferation are then
either signal induced or by some other deterministic or stochastic condition. Cell division either
terminates in a differentiated cell or cells proliferate by doubling, growing exponentially at a
certain rate. For example, the rate may be a function of another parameter that represents the
strength of a diffused signal where the greater signal strength the faster cells proliferate.
These models have some similarity with Till’s original stochastic model (Fig. 37) [29]. By
balancing rates of cell proliferation with rates of terminal cell differentiation or death, have
successfully described the growth of spheroid tumors where cells at the center of the tumor die
due to lack of oxygen.
However, like Till’s model, these models implicitly presuppose a particular type of developmental control network. Recall that the Till model has inherent limitations based on the underlying developmental cancer network. It cannot simulate most of the other cancer networks. So
too, rate-based models fail to give an account of all the other types of cancer that do not fit into
this scheme. At best they can attempt to describe other cancers by adjusting the values of the
cell cycling rate and or the probability distribution. As we saw stochastic networks can be very
flexible in the kinds of developmental network topologies that they imitate. Unless we consider
the actual architecture of the underlying network, it is questionable how much understanding
we get from adjusting probabilities or rate constants.
Purely rate-based, whether deterministic or stochastic, mathematical models not only fail to
grasp the essential unifying features common to all cancers, but also they fail to explain why
particular cancers are they way they are. For example, rate-based theories also fail to explain
the hierarchical nature of metastases or why there are teratomas or bilateral cancers. While
they may describe the rate of growth of some tumors they fail to explain how and why the
tumor grows at all.
In sum, rate-based models of cancer fail to provide an explanation of why the tumor grows
differently from another tumor. All they can do is to adjust rate constants or probabilities to
get different rates of growth. They cannot explain why some cancers grow very slowly or why
mutations can transform them into suddenly growing exponentially, except by saying the rates
of cell proliferation have changed. But, saying that is just circular reasoning: The rate of the
tumor growth has changed because the rate of cell proliferation has changed.
21.3
Systemic models: Cancer is not the result of self-organization based on
physical emergent properties of cells
Soto and Sonnenschein claim that cells cannot be controlled by a program, a strict recipe or set
of instructions:
. . . it is unlikely that each cell can follow and execute each of the encoded rules
independently of the crowded environment present in a tissue or organ. Since cells
can only sense their local environment, the emergence of tissues can only be driven
by rules governed by coordinated interactions with the local environment of each
Eric Werner, A general theory of cancer
112
cell. This leads to the conclusion that the dynamic process of tissue formation
must mainly be governed by self-organization.[22]
While multicellular development certainly depends on the local social, communicative and
physical context of each each cell, the conclusion that tissue formation is mainly controlled by
self organization is not warranted. We have shown that cellular developmental control networks
can be conditional, reactive, and interactive. Developmental control networks contain complex
executable strategies. Furthermore, each cell can be in a different execution state that activates
a distinct developmental subnetwork (cene) in the global developmental network (cenome).
Hence, each cell can have its own unique net-based strategy that interacts with the strategies of
other cells -all in the social, communicative and physical context of neighboring cells.
This is not to say that the physical and social context in which a cell resides is not important.
The physical and social context of other cells influences the how the cell’s developmental
network interacts and how the network strategy is physically interpreted and realized. But, selforganization based on physical properties alone without the complex control information in
developmental networks is not sufficient to generate particular, complex multicellular structure
or morphology. Physical constraints are pleiotropic that, like genes, have a general effect
leading to redundant structures.
It is further argued by Sonnenschein and Soto [27, 22] that as multicellular systems develop
the physical system changes resulting in the emergence of new properties that then influence
the development itself. We agree. However, the mistake that is made is to conclude that such
emergent properties are sufficient for the particular ontogenic development of a species.
21.4
Stochastic models: Stochasticity is not a sufficient explanation of cancer
dynamics
We saw that stochasticity in developmental control networks can add great flexibility to the
network. The probability distribution determines many of the proliferative properties of the
network. Thus, a general stochastic network like that in Fig. 40 can emulate many deterministic developmental networks that are topologically distinct. Moreover, we saw (e.g., Fig. 42 and
Fig. 45) that the rate of cancer growth of various cancers can be adjusted by varying probabilities of stochastic delay loops.
That being said, all stochastic models make implicit assumptions about underlying developmental networks. Any probability space is based on a foundational possibility space. The
topology of the developmental network associated with stochastic probabilities provides this
space of possible dynamic paths that a cell can take. Therefore, stochasticity is not a sufficient condition for cancer development. Neither is it a necessary condition since cancerous
developmental stem cell networks may be deterministic.
Eric Werner, A general theory of cancer
21.4.1
113
Stochasticity, competence and performance in developmental networks
Interestingly, a cancer network can be deterministic, but the performance of the network in
real continuous physics may be stochastic. This is another reason to distinguish between network competence and network performance. The network acts in the context of a cell with its
interpretive-executive system (IES), in the social context of other cells, and in the context of
cell and environmental physics. Each of theses context may by itself be stochastic without the
network being stochastic. Hence, we may have stochastic performance with a network that has
a deterministic competence.
21.5
Gene-based models: Genes alone are not the cause of cancer
Previous so called network theories of cancer focus on interactions between gene products as
the cause of cancer.
21.5.1
Gene-centered view of development and cancer is a delusion
Low level protein networks based on genes are a different sort of network than higher level developmental control networks. While ultimately both kinds of networks are made of molecular
components, it is their role, place and function in the organizational, informational and ontological hierarchy of the cell and the multicellular system that distinguishes these networks. The
interpretive-executive system (IES) of the cell that gives pragmatic meaning to developmental
networks, in part consists of lower level genetically based protein and RNA networks. The
point is that, with regard to the multicellular development of complex form and function, be it
normal or pathological, it is the higher level control networks (cenes) which play the dominant
role. The gene-centered view of development is the source of the mistake that cancer is caused
by genes [43, 44].
21.5.2
Cancer genes: Cancer networks, not cancer genes, cause cancer
So called cancer genes are pleiotropic (having multiple functions and different effects throughout the development of an organism) and as such cannot, in general, be the cause of a particular
cancer phenotype. While genes are certainly involved in cancer, the gene-based theories of cancer miss the essential nature of cancer. Such gene based theories fail to see that developmental
networks are higher level meta-control networks that control lower level molecular networks
such as the cell cycle.
Eric Werner, A general theory of cancer
21.5.3
114
Cell cycle models: An apparently uncontrolled cell cycle is side-effect and not the
original cause of cancer
The view that an uncontrolled cell cycle is the cause of cancer, fits well with the widely held
assumption that cancer is uncontrolled cell proliferation. By focussing on too low a level of
organization they miss all together the overarching control networks responsible for not only
the development of embryos but also cancer. The cell cycle is controlled in cancer just as much
as it is controlled in normal development. The cell cycle does its job in cancer just as it does in
normal development. It is just activated more frequently.
21.6
Agent-based models: Agents controlled the rules of local interaction are
not sufficient to explain cancer
Agent-based systems view cells as agents that interact with other cells [35, 36]. However, the
complexity of a result is directly related to the complexity of the rules of interaction [37]. If
the rules are purely local, their effect is pervasive leading to redundancies in structure. This
can account for many systemic properties, but agents governed by simple rules of interaction alone, cannot account for structural and morphological complexity in development [41].
For that more information is necessary. That is why we need all the control information in
genomes [40]. While, in principle, we can make a multiagent phenotype as complex as we
want by making the control information in the form of rules more and more complex, doing
such amounts to constructing developmental control networks of the kind we have introduced
in this work.
21.7
Hybrid models: Cancer is not explained by hybrid rate-based, genetic
models
Attempts at hybrid theories of cancer composed of rate-based theories of cancer combined with
gene-based theories and cell physics still fail as theories of cancer because they fail to provide
an explanation for both development of both normal and cancerous multicellular development.
The underlying problem is that multicellular development is not adequately explained by genetic and rate-based theories. Genes are pleiotropic in their causal roles and cannot account for
individual differences in development or cancer [44].
The basic point is: Because previous models of cancer fail to include the topology of developmental networks, they cannot offer a way forward to curing cancer.
21.8
Evolutionary models: Cancer is not just an evolutionary process
Neither the development of organisms nor the development of cancer can be explained simply
as an evolutionary process. While cancer cells do evolve through a process of mutation, what
Eric Werner, A general theory of cancer
115
gives a particular cancer its particular dynamics and morphology is the result of its controlling
developmental networks and their transformations. Our theory explains why particular mutations or transformations of developmental networks change the dynamics and morphology of
the developing cancer. In addition, genetic mutations can and do change the physics of cellcell interactions that make it possible for tumor cells to break free and form metastases, for
example. However, while evolution of the control networks in concert with genetic evolution
do result in cancer, the actual behavior of tumors is explained by the resulting higher level
developmental control networks together with genetic and epigenetic changes that modify cell
behavior and cell physics. Cells can evolve and adapt to the new multicellular contexts that
result from cancer. In concert with developmental control networks such cellular changes or
adaptions can make a cancer more or less severe. What is essential to understand that cancers
evolve on at least two levels: The level of cell interaction strategies (both physics and signaling)
by changes in genetics and at the level of developmental control networks.
21.9
Development as complex social physical system
The cell is a semi-autonomous agent reacting locally to its social and physical context but
globally directed by control information in its developmental networks. The cell reacts locally
to its environment, both internal and external. A cancer cell has local strategies by which it
reacts to stresses such as a lack of oxygen, to pressure, to stretching, to strains, to its neighbors,
and to cell signaling. However, in developing multicellular systems, each cell’s local strategies
is subsumed by higher level, global control strategies that are in the form of developmental
control networks Werner [40].
It is similar to a man going on a long hike through the country. Locally he avoids obstacles and
may even take detours if the environment is convoluted, but globally he keeps his direction.
His actual path then is the result not just of his global strategy but his local reactive strategies
as well. The combination of theses strategies together with the strategy of his environment
leads to a particular path. In social systems these local and global strategies interact with other
agents plus the environment making things even more complex (Werner [35]).
Development is complex communicative, social process. Each cell has local reactivity to impinging circumstance but globally control is the result of developmental control information
that is interpreted and executed by that cell. The more complex and structured the adult organism is the more global control information is required [37]. Moreover, a multicellular system
is a social system of cells that communicate by signals, pushing and pulling and touch (being
aware of their neighbors). Thus a developing organism involves a complex time and space
dependent interaction of local and global strategies of the cellular agents. At any given point
in time and space, a cell finds itself in a possibly new social and physical context that includes
other cells and environmental circumstances. It reacts both locally and globally to this context
by means of its local and global control networks.
The attempts to reduce global control to purely local control would have our man getting lost
in the woods doing a random walk making the probability of his reaching his goal impossibly
Eric Werner, A general theory of cancer
116
small. So it is with the morphogenesis of complex organisms. Previous models of cancer have
focussed on local cell strategies and made attempts to deduce global properties of tumors from
local rules or equations of interaction. This research is essential to understanding the local
reactive strategies of cancer cells, but it needs to be supplemented and integrated with global
developmental strategies in the form of global developmental control network theory that we
have provided here. To get to foundation of cancer and cure it we need to understand global
developmental networks and how they interact with local cell strategies and cellular social and
physical contexts.
22
Conclusion
We have provided a theory that unifies many different types of cancers and their phenotypes by
describing the developmental control networks that underly all cancers.
22.1
Main Results
1. A new paradigm for understanding cancer. At the core the of our theory of cancer
is a control network based theory of how organisms develop from a single cell. It it
is the basis for all multicellular life. Cancer is then a special, pathological instance of
this more general theory of development. Our theory is in contrast with the dominant
gene centered theories of development and cancer. Cancer, its etiology, dynamics, and
development is not explained by just the mutations of genes.
2. Curing cancer. Our theory shows that, in principle, we can stop any cancer. However,
to apply the theory we need new discoveries that relate the theory to its molecular implementation. In concert, we need to develop new technologies for recognizing and then
transforming cancer networks into harmless normal developmental networks. In spite of
these challenges, the theory offers significant new insights into the nature of cancer and,
more generally, the development of tissue, organs and organisms.
3. The control networks underlying stem cells. We generalized the stem cell networks
beyond linear networks, to meta-stem cell networks, and ever higher level geometric
networks. We showed the strong relationship between normal stem cell networks and
cancer stem cell networks.
4. New categories of stem cells and cancer stem cells. Stem cells can be understood
by way of their control networks, namely the geometric networks. We showed that
geometric networks growth dynamics obey formulas based on the classical mathematics
of the coefficients of Pascal’s Triangle.
5. Exponential and hyper-exponential cancer networks. These networks give us a theory of the extremely fast growing cancers.
6. A novel theory of teratoma tumors. Tumors with different tissue, cell types and even
structures such as teeth and hair are understood with the control network theory. We
Eric Werner, A general theory of cancer
117
showed how the properties of the teratoma depends on which locations the cancer network is linked to the overall developmental control network of the organism. This allows
the us to distinguish and explain immature teratoma tumors from mature ones. It also
implies new strategies for treatment. Our theory further explains a very rare and unique
type of tumor, the fetus in fetu tumors where a mass of tissue resembling an embryo
growths inside person’s body.
7. Rare cases of cancer explained. Our theory also explains the properties of the rare
cases of mole pregnancies that result is a very fast growing cancer.
8. Cancer network theory explains bilaterally symmetric tumors. These occur in rare
forms of breast cancer. The properties of the particular tumor will depend on the network’s locality and its type of cancer network topology.
9. The theory links cancer and normal development. More generally, our theory implies
the properties of a cancer will depend on linking relationships of the cancer network to
the global developmental control network of the organism.
10. Stochastic cancer networks. Stochastic control networks of development involve probabilities. We showed that stochastic networks can be used to model and approximate
all the major non-stochastic, deterministic cancer control networks. At the same time, it
allowed us to model some classical cancer stem cell behavior.
11. Cancer signaling networks. We showed that cell-cell communicative interactions can,
under certain conditions, ameliorate the growth potential of even exponential cancer
networks. Yet we also developed models of cancer networks for cancers requiring cell
communication with noncancerous cells. Such is the case in some bone cancers. Here
treatment would require stopping the communication, while in the former case stopping
cell signaling might actually exacerbate the exponential cancer.
12. A hierarchy of metastases related to a cancer network hierarchy We related the
properties of cancer metastases to the properties of cancer network of their founding
cancer stem cell. We showed that the dynamics and phenotypes of different types of
possible metastases are a direct consequence of their cancer networks. Interestingly,
certain types of metastases are less harmful than others. This theory also indicates the
best way to stop metastatic growth by eliminating or transforming the most dangerous
founder cells.
13. New research paradigm. Any new scientific paradigm should open up previously inconceivable questions and problems. A theory that does this is all the more significant
if it shows the directions and methods of how to answer those questions and solve those
problems. The network theory of development and cancer opens up a vast new field of
cancer research and gives fundamental insights into how cancer can be cured.
Eric Werner, A general theory of cancer
22.2
118
Open Questions and New Research Problems
1. How are cancer networks and, more generally, developmental control networks implemented at the multicellular level, the cellular level, the genome level and the molecular
level?
2. We need to integrate our theory of cancer and developmental control networks with the
extensive research done on the mathematical modeling on the local physics of cancer
cells in chemical, signaling and multicellular contexts. This will involve integrating
genetics and genomics, etc. with cenomics (the study of cenes and the cenome -the
global control networks in the genome).
3. We need to link the higher level developmental control networks with lower level molecular networks such as the cell cycle and other intracellular signaling networks.
4. Will we be able to recognize cancer cells just by looking at the control networks implicit
in their genomes?
5. If so, then it calls for new methods to manipulate and transform cancer networks into
harmless and normal developmental control networks.
6. Our cancer theory shows how, in principle, one can stop any cancer in any given cell.
How can our theory be used to develop new research strategies for the development of
medicines for curing cancer?
22.3
Beyond Cancer Networks
Looking beyond cancer networks, our network theory of development has further utility.
As we saw, the problem of aging and rejuvenation are intimately related to stem cell networks
and cancer networks. The control network theory of cancer will help researchers determine the
boundary that separates a possible fountain of youth, promised by tissue and organ regeneration, and cancer.
Our theory points to new strategies and methods for stem cell research and regenerative
medicine. Once the network transformations necessary for tissue and organ generations are
understood and made safe by avoiding cancerous network transformations, then we may indeed have a new era of rejuvenating medicine. This is done today by means of surgical operations or exogenous transplantation. In the future it could be by direct, controlled regrowth of
needed cells, tissues and organs using computationally designed, synthetically bio-engineered
developmental control networks.
Our theory of developmental networks make a shift of focus from the single cell to multicellular
systems. It implies that a new era of multicellular biotechnology will be possible. Multicellular biotechnology will combine computational with synthetic systems biology to design and
construct single cells and their genomes that when activated generate designed multicellular
systems to perform various functions.
Eric Werner, A general theory of cancer
119
Finally, more fundamentally, our theory of development has deep scientific consequences for
the theory of evolution. It predicts the existence of an as yet hidden developmental control code
in genomes, a possibly universal code of life. More this in subsequent publications.
23
Materials and methods
All the stem cell and cancer networks were modeled and simulated with Cellnomica’s Software Suite ( http://www.cellnomica.com ). The normal and cancerous developmental control
networks were constructed with Cellnomica’s graphical Network Model Builder. The multicellular images of tumors were snapshots taken at temporal instances of networks being executed
in parallel by cells in a dynamic, simulated multicellular system taking place in continuous
space-time with continuous physics. Simulations of networks in action in developing multicellular systems used Cellnomica’s Developmental Simulation System. All the multicellular
systems represented in the images in this work were grown from a single cell. The tumors resulted from mutational transformations of previously designed, non-cancerous developmental
networks, again using Cellnomica’s software.
24
Request for comments
Author contact for comments on this paper: ([email protected] Please put the word
Cancer in the subject line)
This paper is meant to motivate researchers, doctors, clinicians, those in the pharmaceutical
and health industries to view cancer from a different perspective -to break the hold of the genecentered paradigm. This is consistent with incorporating the tremendous knowledge gained
by means of the gene-centered paradigm. The network control theory is consistent with genetic discoveries. We only claim that gene-centered paradigms is not sufficient to understand
multicellular development and in particular cancer. Hence, we welcome all researchers to participate in a new dialogue on the nature of cancer and how we can cooperate to find a cure for
cancer.
The author welcomes comments. Please contact the author especially if:
• You have particular examples of cancers that fit or do not seem to fit this theory
• You have stem cell examples that seem to be higher order stem cells.
• You have examples of metastases that seem to fit the metastatic hierarchy hypothesis in
this paper.
• You have a publication that is not in the references and is relevant. There are thousands
of publications on cancer. Please forgive any particularly relevant omissions.
• You have ideas for extending the theory and/or linking it with more detailed molecular
theories.
Eric Werner, A general theory of cancer
120
• You see errors of any kind in the paper.
• Any other helpful comments are welcome.
For reasons of space only a limited number of references are given. Let me know if your article
is especially relevant and should be included.
References
[1] A. R. A. Anderson and M. A. J. Chaplain. Continuous and discrete mathematical models of tumourinduced angiogenesis. Bull. Math. Biol., 60:857–
899, 1998.
[2] A. R. A. Anderson and V. Quaranta. Integrative mathematical oncology. Nat Rev Cancer,
8(3):227–234, Mar. 2008.
[3] M. Bizzarri, A. Cucina, F. Conti, and
F. D’Anselmi. Beyond the oncogene paradigm:
Understanding complexity in cancerogenesis.
Acta Biotheoretica, 56:173–196, 2008.
10.1007/s10441-008-9047-8.
[4] R. Brooks. A robust layered control system for a
mobile robot. IEEE Journal of Robotics and Automation, vol RA-2(No. 1), 1986.
[5] H. Byrne and D. Drasdo. Individual-based and
continuum models of growing cell populations: a
comparison. J. Math. Biol., 58:657–687, 2009.
[6] N. Chomsky. Aspects of the Theory of Syntax.
Cambridge, MA: MIT Press, 1965.
[7] J. Dick, R. H. Begent, and T. Meyer. Sarcoidosis
and testicular cancer: A case series and literature
review. Urologic oncology, 28(4):350–4, 2010.
[8] J. E. Dick. Stem cell concepts renew cancer research. Blood, 112(13):4793–807, 2008.
[9] J. E. Dick. Looking ahead in cancer stem cell research. Nature biotechnology, 27(1):44–6, 2009.
[10] R. L. Gardner. Polarity in early mammalian development. Current opinion in genetics and development, 9(4):417–21, 1999.
[11] R. L. Gardner. Specification of embryonic axes
begins before cleavage in normal mouse development. Development, 128(6):839–47, 2001.
[12] R. L. Gardner. The axis of polarity of the mouse
blastocyst is specified before blastulation and independently of the zona pellucida. Human reproduction, 22(3):798–806, 2007.
[13] R. L. Gardner and T. J. Davies. An investigation
of the origin and significance of bilateral symmetry of the pronuclear zygote in the mouse. Human
reproduction, 21(2):492–502, 2006.
[14] D. Hanahan and R. Weinberg. The hallmarks of
cancer,. Cell, 100:57–70, 2000.
[15] P. H. Jones, B. D. Simons, and F. M. Watt. Sic
transit gloria: farewell to the epidermal transit amplifying cell? Cell stem cell, 1(4):371–81, 2007.
[16] A. G. Knudson. Mutation and cancer: statistical study of retinoblastoma. Proc. Natl Acad. Sci.
USA, 68:820–823, 1971.
[17] C. J. Logothetis and S. H. Lin. Osteoblasts in
prostate cancer metastasis to bone. Nat Rev Cancer, 5(1):21–8, 2005.
[18] M. Nakada, D. Kita, T. Watanabe, Y. Hayashi,
L. Teng, I. V. Pyko, and J.-I. Hamada. Aberrant signaling pathways in glioma. Cancers,
3(3):3242–3278, 2011.
[19] B. Palsson. Systems Biology: Properties of Reconstructed Networks. Cambridge University Press,
2006.
[20] V. Quaranta, K. A. Rejniak, and A. R. Anderson. Invasion emerges from cancer cell adaptation to competitive microenvironments: quantitative predictions from multiscale mathematical
models. Semin. Cancer Biol., 18:338–348, 2008.
[21] K. A. Rejniak and A. R. A. Anderson. Hybrid models of tumor growth. Wiley Interdisciplinary Reviews: Systems Biology and Medicine,
3(1):115–125, 2011.
[22] K. Saetzler, C. Sonnenschein, and A. Soto. Systems biology beyond networks: Generating order
from disorder through self-organization. Seminars
in Cancer Biology, 21(3):165 – 174, 2011.
[23] M. Shackleton. Melanoma stem cells–are there
devils in the detail? Pigment cell & melanoma
research, 23(5):693–4, 2010.
[24] M. Shackleton. Moving targets that drive cancer progression. The New England journal of
medicine, 363(9):885–6, 2010.
[25] M. Shackleton. Normal stem cells and cancer stem
cells: similar and different. Seminars in cancer biology, 20(2):85–92, 2010.
[26] M. Shackleton and E. Quintana. Progress in understanding melanoma propagation. Molecular oncology, 4(5):451–7, 2010.
Eric Werner, A general theory of cancer
121
[27] A. Sonnenschein, C. & Soto. The Society of Cells,
Cancer and control of cell proliferation. Bios Scientific Publishers, 1999.
Aspects of Reasoning About Knowledge: Proceedings of the Second Conference, pages 129–143,
1988.
[28] P. J. Tesar, J. G. Chenoweth, F. A. Brook, T. J.
Davies, E. P. Evans, D. L. Mack, R. L. Gardner,
and R. D. McKay. New cell lines from mouse epiblast share defining features with human embryonic stem cells. Nature, 448(7150):196–9, 2007.
[36] E. Werner. Cooperating agents: a unified theory of
communication and social structure, pages 3–36.
Morgan Kaufmann Publishers Inc., San Francisco,
CA, USA, 1989.
[29] J. E. Till, E. A. McCulloch, and L. Siminovitch. A
stochastic model of stem cell proliferation, based
on the growth of spleen colony-forming cells. Proceedings of the National Academy of Sciences of
the United States of America, 51:29–36, 1964.
[30] A. M. Turing. The chemical basis of morphogenesis. 1953. Bull Math Biol, 52(1-2):153–97; discussion 119–52, 1990.
[31] J. J. Tyson. Modeling the cell division cycle: cdc2
and cyclin interactions. Proc Natl Acad Sci U S A,
88(16):7328–32, 1991.
[32] J. J. Tyson, W. T. Baumann, C. Chen, A. Verdugo, I. Tavassoly, Y. Wang, L. M. Weiner, and
R. Clarke. Dynamic modelling of oestrogen signalling and cell fate in breast cancer cells. Nat Rev
Cancer, 11(7):523–532, July 2011.
[33] J. J. Tyson, K. C. Chen, and B. Novak. Sniffers,
buzzers, toggles and blinkers: dynamics of regulatory and signaling pathways in the cell. Curr Opin
Cell Biol, 15(2):221–31, 2003.
[34] J. J. Tyson, A. Csikasz-Nagy, and B. Novak. The
dynamics of cell cycle regulation. Bioessays,
24(12):1095–109, 2002.
[35] E. Werner. Toward a theory of communication and
cooperation for multiagent planning. Theoretical
[37] E. Werner. What ants cannot do. In J. Perram and
J. Mueller, editors, Distributed Software Agents
and Applications. Springer Verlag, 1996.
[38] E. Werner. In silico multicellular systems biology
and minimal genomes. DDT, 8(24):1121–1127,
2003.
[39] E. Werner. Genome semantics, in silico multicellular systems and the central dogma. FEBS Letters, 579(7):1779–1782, 2005.
[40] E. Werner. How central is the genome? Science,
317(5839):753–754, 2007.
[41] E. Werner. Evolutionary embryos.
460(7251):35, 2009.
Nature.,
[42] E. Werner. Meaning in a quantum universe. Science, 329(5992):629–630, 2010.
[43] E. Werner. The cenome: On the global architecture of genomes. arXiv (forthcoming), 2011.
[44] E. Werner. Developmental delusions. arXiv (forthcoming), 2011. A critiqud of gene-centered developmental biology.
[45] E. Werner. On programs and genomes. ArXiv eprints, Nov. 2011. arXiv:1110.5265v1 [q-bio.OT]
(pdf at: http://arxiv.org/abs/1110.5265).
| 5cs.CE
|
1
Distributed Multi-User Wireless Charging
Power Allocation
arXiv:1801.06835v1 [cs.IT] 21 Jan 2018
Yanju Gu
Abstract
Wireless power charging enables portable devices to be permanently unplugged. Due to its low transmission power and low transmission efficiency, it requires much longer time slot to charge users
compared with that for data transmission in wireless communication networks. Besides, each user’s
demand urgency needs to be taken into consideration for power allocation. Therefore, new algorithms
are essential for wireless power allocation in multi-user wireless charging networks. In this paper, this
problem is formulated as a static noncooperative game. It is shown that there exists a unique Nash
equilibrium, which is the static state of the wireless power charging network. A distributed power
allocation algorithm is proposed to compute the Nash equilibrium of the game. The main result of the
paper consists of rigorous analysis of the distributed algorithm for power allocation. The algorithm is
shown to converge to Nash equilibrium of the game with exponentially convergence rate for arbitrary
initial value with synchronous scheduling. Moreover, the distributed algorithm is also convergence
guaranteed with asynchronous scheduling under communication delay and packet drops. Numerical
simulations prove the correctness of the analysis and demonstrate the fast convergence of the algorithm
and the robustness to synchronous scheduling.
I. I NTRODUCTION
Wireless communication aims at providing communication freedom by bringing network to
anywhere at anytime. However, wired charging remains the main way to feed a device battery and
the battery endurance shortage has become the major bottleneck that hinders the development of
ubiquitous wireless communication systems. The cables supporting power charging has become
the only cables that would require removal.
Wireless power charging is being developed to support wireless communication systems. It is
expected that wireless power charging will play a role in future wireless networks, including the
upcoming Internet of Things systems, wireless sensor networks, RFID networks and small cell
networks. It has the potential of powering devices larger than low-power sensors and tags, e.g.,
January 23, 2018
DRAFT
2
wearable computing devices and smartphones. Many leading smartphone manufacturers, such as
Apple, Samsung and Huawei, have begun to release new-generation devices featured with built-in
wireless charging capability. Driven by the practical needs, wireless power charging has gained
momentum from theoretics. Existing works on wireless charging communication networks can be
categorized in two directions. The first one focuses on exclusive wireless charging, i.e., wireless
power transfer and information transmission are separated. The second direction is the research
that wireless charging and information transmission are coupled to achieve some tradeoff. In each
direction, wireless power charging has been conducted in various contexts, e.g., point-to-point
channels [1], [2], relay channels [3], [4], full duplex channels [5], [6].
In general, a wireless power charging system consists of a “transmitter” device connected
to a source of power such as mains power lines, which converts the power to a time-varying
electromagnetic field, and one or more “receiver” devices which receive the power and convert
it back to electric power. From the perspective of transmission range, wireless power charging
can be divided into near-field and far-field power charging. The far-field power charging that
utilizes diffused RF/microwave as a medium to carry radiant energy has a long history and has
recently received attention of researchers from communications society. It is shown that radio
frequency signal ambient in the air can be harvested for data transmission, and thus interference
signal plays a positive role from the wireless power charging perspective [3].
On the other hand, for the near-field wireless power charging, traditionally, transmission range
of centimeters has been realized based on the principle of electromagnetic inductive coupling.
For instance, in the Qi standard, the distance between a receiver, e.g., a mobile phone, and its
power transmitter should not exceed 4 cm [7]. More recently, to boost the efficiency of power
charging, magnetic resonant coupling (MRC) has been extensively studied which has higher
power charging efficiency as well as longer operation range up to a couple of meters. MRC
effectively avoids power leakage to non-resonant externalities and thus ensures safety to the
neighboring environment. In [8], an experiment system with multiple transmitters is successfully
delivered which can simultaneously charge a cell phone at distances of about half meter, and
work independently how the phone is oriented in user’s pocket and in [9] the authors further
prototyped charging, simultaneously, multiple devices including smart phones and tablets.
No matter for which kinds of wireless power charging technology, the number of power
transmitter is expected to be small and each transmitter needs to serve multiple users at the
same time. Thus how to allocate the charging power to different users needs to be studied
January 23, 2018
DRAFT
3
and is still missing. Different from the multi-user communication, multi-user wireless power
charging has its unique feature and raise new challenges. First, wireless power charging, due to
low transmission power and low transmission efficiency, takes much longer time to full charge
users compared with data transmission in wireless communication networks. Besides, the battery
capacities for each user may vary and lead to different levels of aspiration for power charging.
Therefore, with limited power at the transmitter, multi-user charging power allocation is an
important issue to be solved with consideration of different users expected amount and duration.
In this paper, we model the competition of limited charging power among users via game
theory. In general, the priority of power allocation should give to a user with low state of charge
and tight charging deadline. In order to be able to ensure at which state the multi-user wireless
charing networks will effectively operate, we provide rigourous analysis of the Nash equilibrium.
The contributions can be summarized as follows
•
We propose a game framework for wireless power allocation to multiple users with different
charging needs and deadlines. Users compete for the limited power by independently submitting a unit price bid to the power transmitter, and proportional sharing of power follows
for the power allocation.
•
We analyze this wireless power allocation problem using rigorous game-theoretic analysis.
We show the existence and uniqueness of the Nash equilibrium for wireless power allocation.
Besides, we show for arbitrary initial values the convergence to the Nash equilibrium with
exponential rate.
•
In many practical scenarios, the inter-user information exchange is asynchronous since random data packet losses may occur, and different nodes may update at different frequencies.
These impacts are considered in computing the Nash equilibrium, which represent a general
framework including the synchronous updating as a special case.
•
The power allocation algorithm takes the advantage of distributed computation [10]–[12].
The users’ charging capacity and charging deadline is not required to be shared while
computing the Nash equilibrium.
The rest of this paper is organized as follows. Section II gives the problem modeling and
formulation. Section III rigourously shows the existence and uniqueness of the Nash equilibrium,
proposes a distributed algorithm to reach the Nash equilibrium with exponentially convergence
rate, which is also robust to communication delay and packet loss. Section IV shows the
January 23, 2018
DRAFT
4
simulation results of distributed multi-user wireless charging power allocation. Finally, Section
V concludes the paper.
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
We consider a wireless charging server with total transition power P , which serves the charging
services of M users in the its effective charging area. V = {1, . . . , M } denotes the set of users.
Similar to the single-input-multiple-output communication systems, the transmitter’s power is
wirelessly transferred to multiple users. Let hi be the wireless power charging efficiency to
user i, with value varying from 0 to 1, which is defined as the received power at the receiver
divided by the input power for charing i at the charging server. Note that with different types of
wireless power transmission schemes, hi has different physical meaning. More specifically, in
the MRC based wireless power charging system, hi depends on the mutual inductance between
electromagnetic coils of the transmitter and ith receiver [8]. While in the RF based wireless power
transmission schemes, hi is the product of the transmit efficiency, the free-space propagation
efficiency and the receive efficiency [4]. In the state-of-the-art, in order to improve the endto-end power charging efficiency, researchers efforts were focused on enhancing the wireless
power transmission efficiency via designing the non-radiated magnetic field or RF signal array
to reduce the side lobes of the beam pattern while keep its main lobe keeps spillover losses to
a minimum [9]. In the RF based wireless power transmission schemed, attaining higher receive
efficiency was also attempted through the design of high-performance rectifying antennas [13]
(i.e. rectennas), which convert the incident RF power back to DC.
In contrast to the high data transmission rate achieved in wireless communication networks,
to transmit certain amount of power for wireless charging takes much longer time due to the
limited of wireless transmission power and low wireless charging effiency. Therefore, each user
i has an expected charging period [0, Di ], after Di user i has to leave the effective charging
area, and Ci is the amount of energy that is needed to fulfill i’s power consumption, such as
communication, requirement at current time. Each time, user i submits a unit price bid xi to
compete for the amount of power transmitted from the charging server, and proportional sharing
mechanism [14] applies, i.e., P is allocated to each user in proportion to their unit price bids.
Hence, the received power at user i is
xi P
yi = P
j∈V
January 23, 2018
xj
hi .
(1)
DRAFT
5
Note that a reasonable unit price bid should be non-negative i.e.,
Si = {xi > 0}.
(2)
The user’s satisfaction to the wireless charging service relies on whether the charged device can
fulfil its requirement. Hence, the satisfaction metric of each user can be defined as
si =
yi
.
Ci /Di
(3)
Note that Ci /Di means the received amount of power that is able to full charge user i, which
is the expected minimum charging rate for user i. Thus, si is the ratio of the allocated power
charging rate yi with the minimum expected charging rate. The physical meaning is that if si = 1,
user i is charged with Ci within exactly charging period Di ; if si > 1, it denotes that user i will
full fill the charge of Ci power within the charging time Di ; and if si < 1, the charging power
rate xi can not satisfy user i’s power charging requirement within [0, Di ] charging period.
In order to obtain energy from the charging server, user i needs to pay for the service.
Therefore, the utility function that user i tries to maximize can be formulated as
P xi
Ui (xi , xV\i ) = si − λxi P
j∈V
xj
where xi P P xi xj denotes the payment required to obtain
j∈V
,
i ∈ V,
P P xi
j∈V
xj
(4)
amount of power, and λ > 0
implies how to balance between the satisfaction of charging power rate and the payment. The
utility function Ui (xi , xV\i ) is a measure of user i’s preference expressed by the amount of
satisfaction he or she receives minus the amount of he/she pays Note that if the urgency is
not take into consideration, i.e., λ = +∞, users would bid as low as possible to reduce their
payments. Substituting (1) and (3) into (4), the utility function is
Ui (xi , xV\i ) =
P x2
D i P hi x i
P
− λP i ,
Ci j∈V xj
j∈V xj
i ∈ V.
(5)
The multi-user wireless charging problems can then be formalized as a non-cooperative game
with the following triplet structure:
G = hV, {Si }i∈V , {Ui (xi , xV\i )}i∈V i,
(6)
where V denotes the game players; Si being the range of xi represents the game strategy, which
is a subset of positive real numbers; and {Ui (xi , xV\i )}i∈V is the utility function for player i. The
January 23, 2018
DRAFT
6
concept of a best response is central for problem formulation and analysis with game theory.
The formal definition is given as follows.
Definition 1 The best response function Fi (xV\i ) of user i is the best strategy for user i given
the power allocation of the others, which is then the solution to the following maximization
problem:
Fi (xV\i ) = arg max Ui (xi , xV\i ).
(7)
xi
Note that for each i, the maximum is taken over xi for a fixed xV\i , [x1 , . . . , xi−1 , xi+1 , . . . , xM ]T .
We show in Appendix A that the best response function for user i in analytically form is
X
X 1
X
Fi (xV\i ) = (
xj ) 2 + K i
xj 2 −
xj ,
j∈V\i
where Ki =
Di hi
.
λCi
(8)
j∈V\i
j∈V\i
The best response function (8) reflects the fact that all users take independently
their best available actions in order to pursue their own individual objectives as expressed by the
P
particular choice of xi . Since j∈V\i xj > 0, the best response function (8) can be reformulated
as
Ki
Fi (xV\i ) = Ki 1 + P
j∈V\i
1/2
xj
+1
−1
,
(9)
and it is evident that when some user j 6= i increases, Fi (xV\i ) will decrease. This observation
indicates that what is best for one link depends in general upon actions of other links due to
the total energy constraint at the charging server. Thus, an equilibrium for the whole system is
reached when every user is unilaterally optimum, i.e., when, given the current strategies of the
others, any change in his own strategy would result in a in terms of Fi (xV\i ). This equilibrium
constitutes the celebrated notion of Nash equilibrium. We define the joint best response function
as F , {F1 (xV\1 ), . . . , FM (xV\M )} and S = S1 × . . . , ×SM , which are a Cartesian product of
Fi (xV\i ) and S1 , . . . , SM , respectively. Then the corresponding Nash equilibrium for the multiuser wireless charging power allocation game is formally defined as follows.
Definition 2 A pure strategy profile x∗ = [x∗1 , . . . , x∗M ] ∈ S is a Nash equilibrium of the game
G in (6) if x∗ = F(x∗ ).
In general, the game G in (6) may admit multiple equilibria. In the next section, we will prove
the uniqueness of the Nash equilibrium and we analyze the property of how to converge to such
January 23, 2018
DRAFT
7
an equilibrium in a fully distributed way [].
III. NASH E QUILIBRIUM AND D ISTRIBUTED A LGORITHM A NALYSIS
A. Nash Equilibrium Analysis
In this subsection, we will show that there is always a unique Nash equilibrium for the
game G in (6) is always exists. Note that the standard interference functions introduced by
Yates [15] have been very influential on the Nash equilibrium analysis and design of distributed
power control laws for wireless communication networks. Quite a number of works [16]–[18]
has been motivated by [15] which conduct the Nash equilibrium analysis via showing that
the corresponding best response function is a standard function. While the standard function
framework is powerful, the framework does not shown the existence of fixed-points and there is
no guarantees on the rate of convergence of the iterates. Note that, recently, [19]–[21] shows that
the distributed belief propagation algorithm equals maximizing local utility function to find a
Nash equilibrium and the convergence rate is analyzed. Further, the convergence rate of a variant
of standard belief propagation is also analyzed in [22]. However, these conclusions cannot be
used in our problem due to the fact that the type of utility functions are different. Motivated by
[15] and [19]–[22], we show some unique properties of the best response function in (8) first
and then conduct the Nash equilibria and distributed computation algorithm analysis.
Property 1 The updating operator F(·) satisfies the following properties:
P 1.1: For arbitrary x, y ∈ S, x ≤ y implies F(x) ≤ F(y).
P 1.2 For arbitrary x ∈ S and α < 1, F(·) satisfies F(αx) > αF(x) and α−1 F(x) > F(α−1 x).
Proof 1 The proof is shown in Appendix B.
Note that Nash equilibrium does not always exist. If the set Si for all i ∈ V was a closed,
bounded, and convex subset of the one dimensional Euclidean space, and Ui (xV\i ) is a strictly
concave function in S, then the noncooperative game (6) has a Nash equilibrium [23, p. 173].
This theorem is often used to check the existenceness of Nash equilibirum in literature [24]–[26].
However, Si is not a compact set as shown in (2), and traditional analysis method fails. Next
the existence of Nash equilibria is studied.
Theorem 1 The non-cooperative wireless charging game in (6) admits at least one Nash equilibrium, i.e., x∗ = F(x∗ ).
January 23, 2018
DRAFT
8
Proof 2 Since F(·) is a continous function and considering P 1.1, F(xn ) must be upper bounded
(n−1)
by F(+∞1) if limxV\i →∞ Fi (+∞1) exists. By substituting xV\i
= +∞1 into F(·) in (9), we
have
ui =
Ki
lim
xV\i →∞
(1 + Ki P
1
(n−1)
j∈V\i xj
=
21
+1
1
Ki .
2
(10)
By defining u = 21 K = 12 [K1 , . . . Ki , . . . KM ]T . we have F(x) < u for all x > 0. Then we
choose arbitrary x(0) > 21 K, we have x(1) = F(x(0) ) < u. By repeatly applying F(·) on both
sides of the inequality we have x(n) < x(n−1) . Therefore x{n} is a monotonically decreasing
sequence lower bounded by 0. According to monotonic theorem, it must converge. Thus there
exsit a fixed point x∗ which satifies x∗ = F(x∗ ). Thus the non-cooperative power game in (6)
admits at least one Nash equilibrium, i.e., x∗i = Fi (x∗V\i ) for all i ∈ V.
It is possible that a game has multiple Nash equilibirums. Next, we show the uniqueness of
the Nash equilibrium in (6).
Theorem 2 The Nash equilibrium for the non-cooperative wireless charging game in (6) is
unique.
Proof 3 The uniqueness property of the converged x(n) , denoted as x∗ is proved by contradiction. Suppose there are two distinct fixed points x∗ and x̃∗ . Without loss of generality, there exist
i ∈ V such that x̃∗i < x∗i . It is clear that, there must exist 0 < αi < 1 such that x̃∗i = αi x∗i . Let
α denote the the smallest αi for all i ∈ V, we have
αx∗ ≤ x̃∗ ,
α < 1.
(11)
Applying F(·) on (11), and considering P 1.1, we have
F(αx∗ ) ≤ F(x̃∗ ) = x̃∗ ,
(12)
where the equality follows the definition of fixed point of x∗ . Besides, as x∗j > 0 and 0 < α < 1,
following P 1.2, we have F(αx∗ ) > αx∗ . According to (12), we obtain
x̃∗ > αx∗ .
January 23, 2018
(13)
DRAFT
9
Now we can conclude that (11) and (13) is a contradiction, and therefore x∗ and x̃∗ are the
same fixed point. Therefore, x(n) converges to a unique fixed positive value.
B. Distributed Algorithm and Convergence Analysis
In the previous subsection, the existence and uniqueness of the Nash equilibrium have been
established by Theorem 1 and Theorem 2, respectively. In order to compute the unique Nash
equilibrium, the charging server could gather all the users’ information together to perform
the computation. However, the parameter Ki , which depends on user i’s charging deadline and
capacity, is the privacy information of user i, and individual user may not be willing to transmit
these information to the charging server. In order to reach the Nash equilibrium of the game
without sharing individual privacy information, distributed iterative computation procedure is
preferred. In this section, we first give the synchronous updating algorithm and then analytically
prove that for arbitrary initial value of the distributed computation algorithm, the distance between
initial value and the Nash equilibrium decreases exponentially. We will also study how to choose
the initial value to make the distributed algorithm converges faster.
We introduce some preliminary definitions to provide a formal description of the distributed
computation algorithm. We assume, without any loss of generality, that the set of times at which
(n)
one or more users update their strategies is the discrete set T = {0, 1, 2, . . .}. Let xi
with
n ∈ T denote the unit price bid of user i at the nth iteration. Hence, according to (8), the unit
price bid of user i is given by
(n)
xi
(n−1)
= Fi (xV\i )
X (n−1)
X (n−1) 1
X (n−1) 2
2
−
xj
= (
xj
) + Ki
xj
j∈V\i
(14)
j∈V\i
j∈V\i
(n)
It is noteworthy that user i only needs to compute xi
(n−1)
with the updated xj
where j 6= i,
without user j sharing the privacy parameter Kj . We summarize the distributed iterative algorithm
(0)
as follows. The algorithm is started by setting user i’s bid as xi . Each node i computes its
(n−1)
updated message according to (14) at independent time n ∈ T with its available xV\i . Then
the joint distributed computation function is
(n−1)
(n−1)
x(n) = F(x(n−1) ) , {F1 (xV\1 ), . . . , FM (xV\M )}.
January 23, 2018
(15)
DRAFT
10
Mathematically speaking, analyzing the Nash equilibrium of (6) equals to analyze the corresponding fixed point problem of (15) [27]. It is shown in [19], [20] that if an iterative function
has some specific properties, the convergence is guaranteed and the convergence rate can be
identified. We next show the convergence property of F(·) follow the proof procedure in [19],
[20].
Theorem 3 x(n) converges to the unique Nash equilibrium x∗ of the non-cooperative wireless
charging game in (6) for any initial bids x(0) ∈ S.
Proof 4 As shown in Theorem 2, the Nash equilibrium is unique. Then the posiibility of inital
bid x(0) can be categorized into three groups: 1) the bid for each user is larger than its Nash
euilibrium, i.e., x∗ < x(0) . 2) the bid for each user is smaller than its Nash euilibrium, i.e.,
x(0) < x∗ . 3) some of the bid for each user is larger than its Nash euilibrium, and some of the
bid for each user is smaller than its Nash euilibrium.
1) x∗ < x(0) : There exist α ≥ 1 which satisfies x(0) ≤ αx∗ . Following the monotinic property, we
have x∗ = F(x∗ ) < F(x(0) ) < F(αx∗ ) ≤ αF(x∗ ) = αx∗ . Let sequence x̃(0) = αx∗ and x̃(1) =
F(x̃(0) ) Then we define a new sequence {x̃(t) }. As x̃(1) = F(x̃(0) ) = F(αx∗ ) ≤ αF(x∗ ) = x̃(0) .
Then, x̃(1) < x̃(0) implies that x̃(t) is a monotinocally decreasing sequence. Thus x̃(t) converges.
Since x∗ is the unique fixed point, x̃(t) also converges to x∗ . As x∗ < x(0) < x̃(0) , F(x∗ ) <
F(x(0) ) < F(x̃(0) ). Thus x(0) with x(0) > x∗ , must converges to x∗ .
2) 0 < x(0) < x∗ : There exist 1 > β > 0 which satisfies x̃(0) = βx(∗) < x(0) . x̃(1) = F(x̃(0) ) =
F(βx∗ ) ≥ βF(x∗ ) = βx∗ = x̃(0) . Thus x(0) converges to x(∗) with 0 < x(0) < x∗ .
3) indefinite: There exist xu and xl such that 0 < xl < x(0) < xu and 0 < xl < x∗ < xu .
Follwing case 1) lim`→∞ F (`) (xu ) = x∗ ; and Follwing case 2) lim`→∞ F (`) (xu ) = x∗ ; Thus,
lim`→∞ F (`) (xu ) ≤ lim`→∞ F (`) (x(0) ) ≤ lim`→∞ F (`) (xu ).
Another fundamental question is how fast the convergence would be. Define a set
C ={x(n) |xU ≥ x(n) ≥ x∗ + 1}
(16)
∪ {x
(n)
∗
(n)
|x − I ≥ x
≥ xL },
with > 0 being an arbitrary small value. Next, we give the definition of the part metric first, and
then show that the distance between {x(n) }n=1,... and the unique Nash equilibrium x∗ decreases
doubly exponentially fast with respect to part metric in C.
January 23, 2018
DRAFT
11
For a proof of the fact that d(x, y) is a metric and for other properties of the part metric we
refer to [28]. Then following the proof in [19], we can show the convergence rate is doubly
exponential.
Theorem 4 With the initial bid x(0) set to be arbitrary value in S, the distance between x(n)
and x∗ decreases doubly exponentially fast with n increases.
The physical intuition of Theorem 4 is that the sequence {x(n) }l=1,... converges at a geometric
rate before x(n) enters x∗ ’s neighborhood, which can be chose arbitrarily small. Next we show
how to make sure the monotonic convergence.
Theorem 5 x(n) converges monotonically for any initial value x(0) > 0 if the function F(·)
satisfies {x ≥ F(x) ∪ F(x) ≥ x} =
6 ∅.
Proof 5 On one hand, if x ≥ F(x) holds for some x, we can establish that x(0) ≥ F(x(0) ) or
equivalently, x(0) ≥ x(1) . By induction, this relationship can be extended to x(0) ≥ x(1) . . . ≥
x(n) . Besides, if F(x) ≥ x holds, it implies . . . ≥ x(n) ≥ . . . ≥ x(1) ≥ x(0) . Therefore,
{x|x ≥ F(x) ∪ F(x) ≥ x} 6= ∅ implies x(0) . . . x(n) must be a monotonic sequence. Since
x(n) > 0, for all n ≥ 0 and considering P 1.1, we have F(x) must within the range of F(0)
and F(+∞1). As shown in (10), we have F(+∞1) = u. On the other hand, substituting
(n−1)
xV\i
= 0 into (14), we obtain F(0) = 0. Hence, the bounded monotonic x(n) converges.
Note that to guarantee {x ≥ F(x) ∪ F(x) ≥ x} =
6 ∅ we need to choose x(0) such that x(0) ≥
x(1) or x(1) ≥ x(0) . Also, if x(0) ≥ x(1) or x(1) ≥ x(0) , we have {x ≥ F(x) ∪ F(x) ≥ x} =
6 ∅.
Thereby, Theorem 1 is equivalent to x(n) converges if x(0) ≥ x(1) or x(1) ≥ x(0) with x(0) > 0.
Hence, to guarantee the convergence in a monotonic fashion, we can simply set x(0) > [u, xU )
if u ≥ xU . Next, we will show how to choose the initial value x(0) to make x(n) converges
faster in a monotonic fashion.
Theorem 6 x(n) is convergence guaranteed with arbitrary initial value x(0) within the set
[u, xU ), and with different x(0) in this range, x(n) converges to the same value. With x(0) = u,
the iteration (14) converges faster than the same iteration with any other x(0) ∈ (u, +∞).
January 23, 2018
DRAFT
12
Proof 6 We denote by x(n) the vector sequence of (14) with initial value x(0) = u and by y (n)
the vector sequence with y (0) ∈ (u, +∞). We shall prove that
x(n) < y (n)
and
||x(n) − x∗ || < ||y (n) − y ∗ ||
(17)
for n = 0, 1, . . . ,. We prove the above results by induction. At the start, we have x(0) = u < y (0) .
Assume that x(n) < y (n) . According to P 1.1, we obtain x∗ ≤ x(n+1) = F(x(n) ) < y (n+1) =
F(y (n+1) ). Thus, (17) is proved and with x(0) = u, then the iteration (14) converges faster than
the same iteration with any other x(0) ∈ [u, +∞).
(n)
We have analyzed the convergence property of synchronous updating xi
for all i ∈ V. The
synchronous updating can be used in centralized computation, where all the information including
Ki for all i ∈ V are gathered in one computation centre. However, from privacy preserving
perspective, users are averse from sharing their charging capacity and charging deadline. The
distributed power allocation algorithm takes the advantage of privacy preserving. In distribueed
(n)
algorithm, the updating of xi
(n−1)
xi
at each iteration only occurs after every user receiving updated
from all its neighbors. However, in practice networks, due to communication packet drops
or processing delay the convergence speed of the distributed algorithm would be slow. Next, a
totally asynchronous bid updating is analyzed.
C. Asynchronous Power Allocation
With the totally asynchronous updating scheme, all the users compute their individual best
response function (8) in a totally asynchronous way. More specifically, some users are allowed
to update (8) more frequently than others with outdated information about the updated bid from
(n)
(s)
the others. For example, when node i computes xi , it may only have xj computed by other
node j with s ≤ n − 1. In order to capture these asynchronous properties of message exchanges,
totally asynchronous updating has been adopted in wireless communication networks [?], [29],
[30]. Next we introduce the totally asynchronous scheduling definition and formulate the totally
asynchronous distributed best response updating.
Definition 3 Totally Asynchronous Scheduling: Let the bid available to node i from user j at
τ (n−1)
time n are xji
, with τi (n − 1) satisfying 0 ≤ τi (n − 1) ≤ n and lim τi (n) = ∞ for all
n→∞
i = 1, ..., M .
January 23, 2018
DRAFT
13
According to Definition 3, the asynchronous updating version of (14) can be expressed as
(n)
xi
τ (n−1)
i
= Fi (xV\i
)
X
X τ (n−1) 1
τ (n−1) 2
2
= (
xj i
) + Ki
xj i
j∈V\i
−
X
j∈V\i
τ (n−1)
xj i
.
(18)
j∈V\i
(n)
Note that xi
is the outgoing information from user i, the corresponding avaiable bid used for
τ (n−1)
computation at user j is denoted by xi j
τ (n−1)
i
. and xV\i
τ (n−1)
= [x1i
τ (n−1)
i
, . . . , xi−1
τ (n−1)
i
, xi+1
τ (n−1) T
, . . . , xMi
The synchronous version of (14) is a special case of asynchronous updating in (14) without
considering the possible information loss due to communications and different user may have
different updating frequencies. Suppose user i only update and the time instance Ti , then the
outgoing information is
(n)
xi
F (xτi (n−1) ), if n ∈ T
i
i
V\i
=
x(n) ,
otherwise.
i
(19)
Following the same arguments in the proof of Theorem 1, we have the following convergence
properties for the asynchronous updating.
Corollary 1 With the asynchronous updating as in (19), x(n) is convergence guaranteed for any
initial value x(0) ∈ S.
IV. P ERFORMANCE E VALUATION
This section provides simulation results on the tests of the developed synchronous and asynchronous distributed multi-user wireless charging power allocation algorithms proposed in Section III. The related wireless charging parameters are generated according to a wireless charging
prototype in [9] which charges users iPhone 4s based on MRC method. More specifically, for
different users, the received power
Ci
Di
varies within R = [1, 3]W The corresponding power
transmission efficiency hi , according to the experiment results is varies within H = [11%, 19%].
The input power P is set to be 20W, which is the mean of standardized wireless charging
product including Duracell Powermat, Energizer Qi, LUXA2 and WiTricity WiT-500. Note that,
with charging methods based on different wireless power charging principle, such as MRC, RF
energy transfer and inductive coupling [31], the charging efficiency hi is different. Besides, each
user may require different amount of energy Ci to fulfill their devices and have different expected
January 23, 2018
DRAFT
] .
14
charging time Di . Thus there may be different methods to set these simulation parameters.
However, the proposed distributed multi-user charging power allocation algorithm and Nash
equilibrium analysis is adapt to different charging methods and users. For the asynchronous
(n)
updating, the probability of user i successfully pass xi
to user j is pi,j , with pi,j ≤ 1. Therefore,
pi,j emulate the asynchronous network due to packet loss or communicating delay. The weighting
parameter λ is set to be 1 without specification.
Fig. 1 shows the convergence behavior of the proposed algorithms for computing the Nash
equilibrium x∗i . Two users compete for the wireless power charging is considered. Ci /Di for i =
1, 2 are randomly picked up from R. It is assumed there is no packet loss nor any communication
delay, thus p1,2 = p2,1 = 1. We start the distributed synchronous updating algorithm with different
(n)
values. It shows in Fig. 1 that each user’s xi
converges to the unique Nash equilibrium within
4 iterations. Moreover, Fig. 2 shows the convergence property for asynchronous scheduling with
the same parameter set up but a successfully communication probability p1,2 = p2,1 = 0.8. It
can be seen that: The distributed approach converges still very rapidly and after convergence the
corresponding Nash equilibrium is unique.
Fig. 3 shows the relationship between iteration number upon convergence versus the network
size for both synchronous and asynchronous scheduling (pi,j = 0.8). It is shown that for both
cases, the the iteration number increases slightly when the network size increases. The slow
increase of iteration number is due to the exponential convergence rate as shown in Theorem
4. Thus for different network sizes, the iteration number is quite small and almost the same.
Fig. 3 also shows with pi,j = 0.8, the relationship between iteration number upon convergence
versus the network size under asynchronous updating. It can be seen that even with 20% chance
of packet loss between each pair of users, the asynchronous algorithm can still reach the Nash
equilibrium quickly. Moreover, the computational complexity for the distributed power allocation
(n)
algorithm is quite low since for each iteration, the computation for the unit price bid xi
in
(14) only involves serval scalar multiplications.
In Figure 4, we show the Nash equilibrium with respect to the number of users M . For each
case all the parameters are set to be the same for a fair comparison. For each user, Ci /Di is set
to be 1.2. It is demonstrated that the Nash equilibrium for each user x∗i monotonically increases
as M increases. This is due to the completion among the users gradually become more severe
such that the bid offered by each user at the equilibrium also increases.
Next we demonstrate the social welfare behavior of wireless charging user. The social welfare
January 23, 2018
DRAFT
15
0.4
User 1
User 2
Unit price bid
0.3
0.2
0.1
0
0
2
4
6
8
10
Number of iteration (Synchronous Scheduling)
Fig. 1. Convergence performance of the proposed algorithm for different initial values with synchronous scheduling.
0.4
0.3
Unit price bid
User 1
User 2
0.2
0.1
0
0
2
4
6
8
10
Number of iteration (Asynchronous Scheduling )
Fig. 2. Convergence performance of the proposed algorithm for different initial values with asynchronous scheduling.
of the users is defined as
∗
U(x ) =
M
X
Ui (x∗i , x∗V\i )
i=1
=
M
X
P (x∗i )2
Di P hi x∗i
P
P
.
−
λ
∗
∗
C
x
x
i
j
j
j∈V
j∈V
i=1
For each fixed user number, Ci /Di are uniformly generated from R and hi is uniformly generated
from D. Each point in the figure is an average of 1000 times simulation. First, Figure 5 shows
that the social welfare of the users monotonically decreases for the reason of the increasing
competition among the users, which implies the increasing social welfare of the charging server.
Second, if as the portion of urgency in the utility function decreases the utility function is
increases.
January 23, 2018
DRAFT
16
Iteration number
10
8
6
4
Asynchronous Scheduling
Synchronous Scheduling
2
0
10
20
30
40
Number of users
50
60
Fig. 3. The iteration numbers upon convergence versus the network size for both synchronous and asynchronous scheduling.
0.4
Unit price bid
0.3
0.2
User number 6, 10, 12
and 14
0.1
0
0
1
2
3
4
Number of iteration
5
6
Fig. 4. Convergence to the Nash equilibrium solution x∗i with respect to different number of users. (ratio = 0.7, asynchronous
0.2 success)
The efficiency of a noncooperative game is defined as the ratio of the highest value of the
social welfare to the worse Nash equilibrium of the game. By assuming users are cooperative,
the highest value of the social welfare is computed by
xP = [xP1 , . . . xPM ] = arg min
x
M
X
Ui (xi , xV\i ).
(20)
i=1
Besides, since we have shown in Theorem 2 that there is a unique Nash equilibrium in G, the
price of anarchy of the game G is
PM
i=1
PoA = PM
Ui (xPi , xPV\i )
∗
∗
i=1 Ui (xi , xV\i )
.
(21)
Figure 6 shows that the price of anarchy increases slowly with the number of users which implies
January 23, 2018
DRAFT
17
5
Users social welfare
4.8
λ=0.9
λ=1
λ=1.1
4.6
4.4
4.2
4
3.8
6
8
10
Number of users
12
14
Fig. 5. Users’ Social welfare U(x∗ ) with respect to users number for different weighting λ. (ratio = 0.7, synchronous 0.2
success)
1.98
Price of anarchy
1.96
1.94
1.92
1.9
1.88
1.86
10
20
30
40
Number of users
50
60
Fig. 6. Price of anarchy with respect to the number of users.
the social welfare gap between cooperative optimal and noncooperative optimal increases slowly.
V. C ONCLUSION
In this paper, a fully distributed multi-user wireless charging power allocation scheme has been
proposed. It has taken into consideration of the impact of charging time on the power allocating
problem. The proposed distributed algorithm only involves limited information exchanges among
users and avoids sharing privacy information as in the centralized algorithm. We have presented
a rigorous analysis for the existence and unique of the Nash equilibrium for both synchronous
and asynchronous updating. Moreover, we have proved that the distributed algorithm converges
with a exponential rate. Simulations have verified the theatrical analysis and shown the efficiency
of the power allocation algorithm.
January 23, 2018
DRAFT
18
A PPENDIX A
By taking the second derivative of Ui (xi , xV\i ) in (20) with respect to xi , we obtain
P
Di P hi j∈V\i xj
∂ 2 Ui (xi , xV\i )
P
=−2
∂x2i
Ci ( j∈V xj )3
P
( j∈V\i xj )2
P
− 2λP
.
Ci ( j∈V xj )3
(22)
As xj > 0 for all j ∈ V, it is clear that ∂ 2 Ui (xi , xV\i )/∂x2i < 0. Therefore, the utility function
Ui (xi , xV\i ) is strictly concave, and the best response function, according to Definition 1, can
be obtained by setting the first derivative of Ui (xi , xV\i ) to be zero, i.e.,
P
xj
∂Ui (xi , xV\i ) Di P hi
Pj∈V\i 2
=
∂xi
Ci ( j∈V xj )
P
x2i + 2xi j∈V\i xj
P
− λP
( j∈V xj )2
(23)
= 0.
After some algebraic manipulations, the best response function for user i can be readily given
by
X 1
X
X
Fi (xV\i ) = (
xj ) 2 + K i
xj 2 −
xj ,
j∈V\i
where Ki =
j∈V\i
(24)
j∈V\i
Di hi
.
λCi
A PPENDIX B
By taking the first-order derivative of Fi (xV\i ) in (8) with respect to xi , we obtain
P
1
∂Fi (xV\i )
j∈V\i xj + 2 Ki
=P
1 − 1.
P
∂xi
( j∈V\i xj )2 + Ki j∈V\i xj 2
(25)
After performing some algebraic manipulations, the numerator of (25) can be reformulated as
X
j∈V\i
= [(
1
xj + K i
2
X
j∈V\i
January 23, 2018
xj )2 + Ki
X
j∈V\i
1
xj + ( Ki )2 ]1/2 .
2
(26)
DRAFT
19
Therefore,
X
j∈V\i
and hence, it is clear that
X
X
1
1
xj + Ki > [(
xj )2 + Ki
xj ]1/2 ,
2
2
j∈V\i
∂Fi (xV\i )
∂xi
(27)
j∈V\i
> 0. Thus, Fi (·) or equivalently, F(·) satisfies that x ≥ y
implies F(x) ≥ F(y).
Next, we will show P 1.2. For arbitrary x ∈ S and α < 1, it can be shown that
X
X
X
1
αxj
(
αxj )2 + Ki
αxj 2 −
j∈V\i
j∈V\i
j∈V\i
(28)
X
X
X
1
> (
αxj )2 + Ki
α 2 xj 2 −
αxj .
j∈V\i
j∈V\i
j∈V\i
Thus, by the definition of Fi (·), (28) implies Fi (αxV\i ) > αFi (xV\i ) = αxV\i . Thus we have
F(αx) > αF(x). With similar argument, we can also show that α−1 F(x) > F(α−1 x).
R EFERENCES
[1] L. Liu, R. Zhang, and K. C. Chua, “Wireless information transfer with opportunistic energy harvesting,” IEEE Trans.
Wireless Commun., vol. 12, no. 1, pp. 288–300, 2013.
[2] S. Luo, R. Zhang, and T. J. Lim, “Optimal save-then-transmit protocol for energy harvesting wireless transmitters,” IEEE
Trans. Wireless Commun., vol. 12, no. 3, pp. 1196–1207, 2013.
[3] Y. Gu and S. Aissa, “Interference aided energy harvesting in decode-and-forward relaying systems,” in Communications
(ICC), 2014 IEEE International Conference on, 2014, pp. 5378–5382.
[4] ——, “Rf-based energy harvesting in decode-and-forward relaying systems: Ergodic and outage capacities,” IEEE Trans.
Wireless Commun., vol. 14, no. 11, pp. 6425–6434, 2015.
[5] Z. D. S. Hu and Q. Ni, “Beamforming optimization in energy harvesting cooperative full-duplex networks with self-energy
recycling protocol,” IET Communications, vol. 12, no. 3, pp. 1196–1207, 2015.
[6] M. Mohammadi, B. Chalise, H. Suraweera, C. Zhong, G. Zheng, and I. Krikidis, “Throughput analysis and optimization
of wireless-powered multiple antenna full-duplex relay systems,” IEEE Trans. Commun., vol. PP, no. 99, pp. 1–1, 2016.
[7] “Qi interface specification, system description: Wireless power transfer,” ser. ver. 1.1.2, 2012. [Online]. Available:
http://www.wirelesspowerconsortium.com/downloads/wireless-power-specification-part-1.html
[8] J. Jadidian and D. Katabi, “Magnetic mimo: How to charge your phone in your pocket,” in Proceedings of the 20th Annual
International Conference on Mobile Computing and Networking (MobiCom ’14), 2014, pp. 495–506.
[9] L. Shi, Z. Kabelac, D. Katabi, and D. Perreault, “Wireless power hotspot that charges all of your devices,” in Proceedings
of the 21st Annual International Conference on Mobile Computing and Networking MobiCom ’13, 2015, pp. 2–13.
[10] J. Du, S. Ma, Y.-C. Wu, and H. V. Poor, “Distributed hybrid power state estimation under PMU sampling phase errors,”
IEEE Transactions on Signal Processing, vol. 62, no. 16, pp. 4052–4063, 2014.
[11] J. Du and Y.-C. Wu, “Network-wide distributed carrier frequency offsets estimation and compensation via belief
propagation,” IEEE Transactions on Signal Processing, vol. 61, no. 23, pp. 5868–5877, 2013.
January 23, 2018
DRAFT
20
[12] J. Du, S. Kar, and J. M. Moura, “Distributed convergence verification for Gaussian belief propagation,” Asilomar Conference
on Signals, Systems and Computers (ASILOMAR), arXiv preprint arXiv:1711.09888, 2017.
[13] A. Massa, G. Oliveri, F. Viani, and P. Rocca, “Array designs for long-distance wireless power transmission: State-of-the-art
and innovative solutions,” Proceedings of the IEEE, vol. 101, no. 6, pp. 1464–1481, 2013.
[14] N. Nisan, T. Roughgarden, E. Tardos, and V. V. Vazirani, Algorithmic Game Theory, 2ed.
Cambridge University Press,
2007.
[15] R. Yates, “A framework for uplink power control in cellular radio systems,” IEEE J. Sel. Areas Commun., vol. 13, no. 7,
1995.
[16] S. Bu, F. R. Yu, Y. Cai, and X. P. Liu, “When the smart grid meets energy-efficient communications: Green wireless
cellular networks powered by the smart grid,” IEEE Trans. Wireless Commun., vol. 11, no. 8, 2012.
[17] H. Chen, Y. Li, Y. Jiang, Y. Ma, and B. Vucetic, “Distributed power splitting for swipt in relay interference channels using
game theory,” IEEE Trans. Wireless Commun., vol. 14, no. 1, pp. 410–420, 2015.
[18] C. K. Ho, D. Yuan, L. Lei, and S. Sun, “Power and load coupling in cellular networks for energy optimization,” IEEE
Trans. Wireless Commun., vol. 14, no. 1, pp. 509–519, 2015.
[19] J. Du, S. Ma, Y.-C. Wu, S. Kar, and J. M. Moura, “Convergence analysis of distributed inference with vector-valued
Gaussian belief propagation,” arXiv preprint arXiv:1611.02010, 2016.
[20] ——, “Convergence analysis of the information matrix in Gaussian belief propagation,” IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP), arXiv preprint arXiv:1704.03969, 2017.
[21] ——, “Convergence analysis of belief propagation for pairwise linear Gaussian models,” IEEE Global Conference on
Signal and Information Processing (GlobalSIP), arXiv preprint arXiv:1706.04074, 2017.
[22] J. Du, X. Liu, and L. Rao, “Proactive doppler shift compensation in vehicular cyber-physical systems,” arXiv preprint
arXiv:1710.00778, 2017.
[23] T. Basar and G. J. Olsder, Dynamic Noncooperative Game Theory, 2ed.
Philadelphia, PA, USA: SIAM, 1999.
[24] S. Kucera, S. Aissa, K. Yamamoto, and S. Yoshida, “Asynchronous distributed power and rate control in ad hoc networks:
a game-theoretic approach,” IEEE Transactions on Wireless Communications, vol. 7, no. 7, pp. 2536–2548, 2008.
[25] S. Lasaulce, M. Debbah, and E. Altman, “Methodologies for analyzing equilibria in wireless games,” IEEE Signal
Processing Magazine, vol. 26, no. 5, pp. 41–52, 2009.
[26] H. Chen, Y. Li, Y. Jiang, Y. Ma, and B. Vucetic, “Distributed power splitting for swipt in relay interference channels using
game theory,” IEEE Transactions on Wireless Communications, vol. 14, no. 1, pp. 410–420, 2015.
[27] J. F. Nash, “Equilibrium points in n-points games,” Proc. Nat. Acad. Sci., vol. 36, no. 1, pp. 48–49, 1950.
[28] M. Hirsch and H. Smith., Monotone dynamical systems. In Handbook of Differential Equations: Ordinary Differential
Equations, volume 2, chapter 4, pages 239-357. Elsevier, 2006.
[29] J. Du and Y.-C. Wu, “Distributed clock skew and offset estimation in wireless sensor networks: Asynchronous algorithm
and convergence analysis,” IEEE Trans. Wireless Commun., vol. 12, no. 11, 2013.
[30] ——, “Fully distributed clock skew and offset estimation in wireless sensor networks,” in IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP), 2013, pp. 4499–4503.
[31] C. Kim, D. Seo, J. You, J. Park, and B. H. Cho, “Array designs for long-distance wireless power transmission: State-ofthe-art and innovative solutions,” IEEE Trans. Ind. Electron., vol. 48, no. 6, 2001.
January 23, 2018
DRAFT
| 7cs.IT
|
arXiv:1512.04647v3 [math.AC] 29 Dec 2015
Krull dimensions of rings of holomorphic functions
Michael Kapovich
Abstract. We prove that the Krull dimension of the ring of holomorphic functions of a connected complex manifold is at least the cardinality of continuum
iff it is > 0.
Let R be a commutative ring. Recall that the Krull dimension dim(R) of R is
the supremum of cardinalities lengths of chains of distinct proper prime ideals in
R. Our main result is:
Theorem 1. Let M be a connected complex manifold and H(M ) be the ring
of holomorphic functions on M . Then the Krull dimension of H(M ) either equals
0 (iff H(M ) = C) or is infinite, iff M admits a nonconstant holomorphic function
M → C. More precisely, unless H(M ) = C, dim H(M ) ≥ c, i.e., the ring H(M )
contains a chain of distinct prime ideals whose length has cardinality of continuum.
Our proof of this theorem mostly follows the lines of the proof by Sasane [S],
who proved that for each nonempty domain M ⊂ C the Krull dimension of H(M )
is infinite (he did not prove that dim H(M ) ≥ c).
Remark 2. We note that Henricksen [H] was the first to prove that the Krull
dimension of the ring of entire functions on C has cardinality at least continuum.
In our proof we will use the Axiom of Choice in two ways: (a) to establish existence of certain maximal ideals and (b) to get existence of a nonprincipal ultrafilter
ω on N and, hence of the ordered field ∗ R of nonstandard real (or, surreal) numbers.
The field ∗ R contains ∗ N, the nonstandard natural (or surnatural) numbers.
Q
The field ∗ R is a certain quotient of the countable direct product k∈N R; we
will denote the equivalence class (in ∗ R) of a sequence (xk ) in R by [xk ]. Accordingly,
∗
N consists of equivalence classes [nk ] of sequences of natural numbers. Roughly
speaking, we will use ∗ N and certain order relation on it to compare rates of growth
of sequences of natural numbers.
Definition 3. A commutative unital ring R is ample if there exists a sequence
of valuations νk on R such that for each β ∈ ∗ N, there a = aβ ∈ R with the property
(1)
[νk (a)] = β.
1991 Mathematics Subject Classification. Primary 32A10, 16P70.
Key words and phrases. Krull dimension; ring of holomorphic functions.
The author was supported in part by the NSF Grant DMS-12-05312 and by the Korea
Institute for Advanced Study (KIAS).
1
2
MICHAEL KAPOVICH
The main technical result of this paper is:
Theorem 4. For each ample ring R, dim(R) ≥ c. In particular, R has infinite
Krull dimension.
This theorem and its proof are inspired by Theorem 2.2 of [S], although some
parts of the proof resemble the ones of [H].
We will verify, furthermore, that whenever M is a connected complex manifold
which has a nonconstant holomorphic function, the ring H(M ) is ample. This,
combined with Theorem 4, will immediately imply Theorem 1.
Remark 5. 1. We refer the reader to Section 5.3 of [Cla] for further discussion
of algebraic properties of rings of holomorphic functions.
2. Theorem 1 shows that for every Stein manifold M (of positive dimension),
the ring H(M ) has infinite Krull dimension. In particular, this applies to any
noncompact connected Riemann surfaces (since every such surface is Stein, [BS]).
3. Noncompact connected complex manifolds M of dimension > 1 can have
H(M ) = C; for instance, take M to be the complement to a finite subset in a
compact connected complex manifold (of dimension > 1).
Acknowledgements. This note grew out of the mathoverflow question,
http://mathoverflow.net/questions/94537,
and I am grateful to Georges Elencwajg for asking the question. I am also grateful
to Pete Clark for pointing at several errors in earlier versions of the paper (most
importantly, pointing out that Lemma 14 is needed for the proof, which forced me
to use ultrafilters) and providing references.
1. Surreal numbers
We refer the reader to [Go] for a detailed treatment of surreal numbers, below
is a brief introduction. A nonprincipal ultrafilter on N can be regarded as a finitelyadditive probability measure on N which vanishes on each finite subset and takes
the value 0 or 1 on each subset of N. The existence of nonprincipal ultrafilters
(the ultrafilter lemma) follows from the Axiom of Choice. Subsets of full measure
are called ω-large. Using ω one defines the following equivalence relation on the
product
Y
R.
k∈R
Two sequences (xk ) and (yk ) are equivalent if xk = yk for an ω-all k, i.e. the set
{k : xk = yk }
is ω-large. The quotient by this equivalence relation, denoted
Y
∗
R=
R/ω,
k∈N
is the set of surreal numbers. Let [xk ] be the equivalence class of the sequence (xk ).
The binary operations on sequences of real numbers project to binary operations on ∗ R making ∗ R a field. The total order ≤ on ∗ R is defined by [xk ] ≤ [yk ] iff
xk ≤ yk for an ω-all k ∈ N. With this order, ∗ R becomes an ordered field.
The set of real numbers embeds into ∗ R as the set of equivalence classes of
constant sequences; the image of a real number x under this embedding is still
denoted x. We set ∗ R+ := {α ∈ ∗ R : α > 0}.
KRULL DIMENSIONS OF RINGS OF HOLOMORPHIC FUNCTIONS
The projection of
Y
k∈N
N⊂
Y
3
R
k∈N
to ∗ R is denoted ∗ N, this is the set of surnatural numbers. We define a further
equivalence relation ∼u on ∗ R by:
α ∼u β
if there exist positive real numbers a, b such that
aα ≤ β ≤ bα.
The equivalence class (α) of α ∈ R (for this equivalence relation) is a multiplicative
analogue of the galaxy gal(α) of α, see [Go]:
∗
Definition 6. The galaxy gal(α) of a surreal number α ∈ ∗ R is the union
[
[α − n, α + n] ⊂ ∗ R.
n∈N
In other words, β ∈ gal(α) iff there exist a real number a such that α−a ≤ β ≤ α+a.
The next lemma is immediate:
Lemma 7. For α ∈ ∗ R+ , the equivalence class (α) of α equals exp(gal(log(α))).
We let u R denote the quotient ∗ R/ ∼u and u N the projection of ∗ N to u R.
Define the total order ≫ on u R by
(β) ≫ (α)
if for every real number c, cα < β. By abusing the notation, we will simply say
that β ≫ α, with α, β ∈ ∗ R.
For the reader who prefers to think in terms of sequences of (positive) real
numbers, the relation (β) ≫ (α) is an analogue of the relation
(an ) = o((bn )),
n → ∞.
Remark 8. The equivalence relation ∼u and the order ≫ are similar to the
ones used by Henricksen in [H].
Proposition 9. The set u N has the cardinality of continuum.
Proof. Note first, that ∗ R has cardinality of continuum, hence, the cardinality of
N is at most c. The proof of the proposition then reduces to two lemmata.
u
Lemma 10. The set gal(∗ R+ ) of galaxies {gal(α) : α ∈ ∗ R+ } has the cardinality
of continuum.
Proof. For each α = [ak ] ∈ ∗ R+ , the galaxy gal(α) contains the surnatural number
⌈α⌉ = [bk ], where bk = ⌈ak ⌉. For each surnatural number β ∈ ∗ N, and natural
number n ∈ N, the intersection
[β − n, β + n] ∩ ∗ N
is finite, equal {β − n, ..., β + n}. Therefore, gal(β) ∩ ∗ N = {β} + Z. It follows that
the map
∗
N → gal(∗ R+ ), β 7→ gal(β)
is a bijection modulo Z. Lastly, the set of surnatural numbers ∗ N has the cardinality
of continuum.
4
MICHAEL KAPOVICH
Lemma 11. The map λ : ∗ N → gal(∗ R+ ), λ : β 7→ gal(log(n)), is surjective.
Proof. For each α ∈ ∗ R+ let β = ⌈exp(α)⌉ ∈ ∗ N. Since log(x + 1) − log(x) ≤ 1 for
x ≥ 1, we have that
log(β) ∈ gal(α).
Now, we can finish the proof of the proposition. The map λ : ∗ N → gal(∗ R+ )
descends to a map µ : u N → gal(∗ R+ ). According to Lemma 11, the map µ is
surjective. By Lemma 10 the set gal(∗ R+ ) has the cardinality of continuum.
We will prove Theorem 4 in the next section by showing that for each ample
ring R, the ordered set (u N, ≫) embeds into the poset of prime ideals in R reversing
the order:
(β) ≫ (α) ⇒ Pβ ( Pα
for certain prime ideals Pγ ⊂ R determines by (γ) ∈ u N. Proposition 9 will then
imply that the Krull dimension of R is at least c.
2. Krull dimension of ample rings
Recall that a valuation on a unital ring R is a map ν : R → R+ ∪ {∞} such that:
1. ν(a + b) ≥ min(a, b),
2. ν(ab) = ν(a) + ν(b).
3. ν(a) = ∞ ⇐⇒ a = 0.
4. ν(1) = 0.
For the following lemma, see Theorem 10.2.6 in [Coh] (see also Proposition 4.8 of
[Cla] or Theorem 1 in [K]).
Lemma 12. Let I be an ideal in a commutative ring A and M ⊂ A\I be a subset
closed under multiplication. Then there exists an ideal J ⊂ A containing I and
disjoint from M , so that J is maximal with respect to this property. Furthermore,
J is a prime ideal in A.
Let R be an ample ring and νk the corresponding sequence of valuations on R.
For each β ∈ ∗ N we define
Iβ := {a ∈ R| [νk (a)] ≫ [β]} ⊂ R.
Lemma 13. Each Iα is an ideal in R.
Proof. We will check that Iα is additive since it is clearly closed under multiplication
by elements of R. Take p′ , p′′ ∈ Iα ,
[νk (p′ )] ≫ α, [νk (p′′ )] ≫ α.
By the definition of a valuation,
nk := νk (p′ + p′′ ) ≥ min(νk (p′ ), νk (p′′ )),
for each k ∈ N. For m ∈ N, define the ω-large sets
A′ = {k : νk (p′ ) ≥ mα},
A′′ = {k : νk (p′′ ) ≥ mα}.
Therefore, their intersection A = A′ ∩ A′′ is ω-large as well, which implies that
∀m ∈ N, [nk ] ≥ mα ⇒ [nk ] ≫ α.
KRULL DIMENSIONS OF RINGS OF HOLOMORPHIC FUNCTIONS
5
Then for each γ ≫ β, the element aγ as in Definition 3, belongs to Iβ . It follows
that Iβ 6= 0 for every β. Define the subsets
Mβ := {a ∈ R|∃n ∈ N, [νk (a)] ≤ nβ} ⊂ R;
each Mβ is closed under the multiplication. It is immediate that whenever α ≤ β,
we have the inclusions
Iβ ⊂ Iα , Mα ⊂ Mβ .
It is also clear that Iβ ∩ Mβ = ∅. At the same time, for each β ≫ α,
aβ ∈ Iα ∩ Mβ .
For each α we let Jα denote the set of ideals P ⊂ R such that
Iα ⊂ P, P ∩ Mα = ∅.
By Lemma 12, every maximal element P ∈ Jα is a prime ideal.
Lemma 14. Every Jα contains unique maximal element, which we will denote
Pα in what follows.
Proof. Suppose that P ′ , P ′′ are two maximal elements of Jα . We define the ideal
P = P ′ + P ′′ . Clearly, P contains Iα . To prove that P is disjoint from Mα , take
p′ ∈ P ′ , p′′ ∈ P ′′ , since p′ ∈
/ Mα , p′′ ∈
/ Mα . Then the same proof as in Lemma 13
′
′′
shows that [νk (p + p )] ≫ α which means that p′ + p′′ ∈
/ Mα . Thus, P ∈ Jα and,
in view of maximality of P ′ , P ′′ , we obtain
P ′ = P = P ′′ .
For each β ≫ α we define the ideal Qαβ := Iα + Pβ .
Lemma 15. Qαβ ∩ Mα = ∅.
Proof. The proof is similar to the one of the previous lemma. Let q = c + p,
c ∈ Iα , p ∈ Pβ . Since p ∈
/ Mβ , p ∈
/ Mα as well. Therefore,
[νk (p)] ≫ α.
Since c ∈ Iα ,
[νk (c)] ≫ α.
Hence,
[νk (c + p)] ≫ α
as well. Thus, q ∈
/ Mα .
Corollary 16. Qαβ ∈ Jα . In particular, Qα ⊂ Pα .
Proof. It suffices to note that Iα ⊂ Qαβ according to the definition of Qαβ .
Lemma 17. The inequality β ≫ α implies Pβ ⊂ Pα and this inclusion is proper.
Proof. By the definition of Qαβ and Corollary 16, we have the inclusions
Pβ ⊂ Qα ⊂ Pα .
We now claim that Pβ 6= Qαβ = Iα + Pβ . Recall that aα ∈ Iα ⊂ Qαβ and aα ∈ Mβ ,
while Mβ ∩ Pβ = ∅. Thus, aα ∈ Qαβ \ Pβ .
6
MICHAEL KAPOVICH
According to Proposition 9, the set ∗ N of surnatural numbers contains a subset
S of cardinality continuum such that for all α < β in S, we have β ≫ α. The map
α 7→ Pα
sends each α ∈ S to a prime ideal in R; α < β implies that Pβ ( Pα .
We conclude that the ring R contains the (descending) chain of distinct prime
ideals Pα , α ∈ S; the length of this chain has the cardinality of continuum. In
particular, dim(R) ≥ c. Theorem 4 follows.
3. Ampleness of rings of holomorphic functions
We will need the following classical result, see e.g. [Con, Ch. VII, Theorem 5.15]:
Theorem 18. Let D ⊂ C be a domain, and let ck ∈ D be a sequence which
does not accumulate anywhere in D and let mk be a sequence of natural numbers.
Then there exists a holomorphic function g in D which has zeroes only at the points
ck and such that mk is the order of zero of g at ck , k ∈ N.
Corollary 19. If M is a connected complex manifold which admits a nonconstant holomorphic function h : M → C, then the ring H(M ) is ample.
Proof. We let D denote the image of h. Pick a sequence ck ∈ D which converges to
a point in Ĉ \ D and which consists of regular values of h. (Here Ĉ is the Riemann
sphere.) For each ck the preimage Ck := h−1 (ck ) is a complex submanifold in M ;
in each Ck pick a point bk . Define valuations
νk : H(M ) → Z+ ∪ {∞}
by νk (f ) := ordbk (f ), the total order of f at bk , cf. [Gu, Chapter C, Definition 1].
Now, given β ∈ ∗ N, β = [mk ], we let g = gβ denote a holomorphic function on
D as in Theorem 18. Define a = aβ := g ◦ h ∈ H(M ). Then νk (a) = mk , which
implies that the ring H(M ) is ample.
Ampleness of H(M ) together with Theorem 4 imply Theorem 1.
References
[BS]
[Cla]
[Coh]
[Con]
[Go]
[Gu]
[H]
[K]
[S]
H. Behnke, K. Stein, Entwicklung analytischer Funktionen auf Riemannschen Flächen,
Math. Ann. Vol. 120 (1949) p. 430–461.
P. Clark, “Commutative Algebra”. Preprint.
P. M. Cohn, “Basic Algebra: Groups, Rings and Fields”, Springer Verlag, 2004.
J. B. Conway, “Functions of One Complex Variable” I, 2nd edition. Springer Verlag,
New York (1978).
R. Goldblatt, “Lectures on the hyperreals,” Graduate Texts in Mathematics, Vol. 188,
Springer-Verlag, 1998.
R. Gunning, “Introduction to Holomorphic Functions of Several Variables,” Volume 1,
Wadsworth & Brooks/Cole, 1990.
M. Henricksen, On the prime ideals of the ring of entire functions, Pacific J. Math. Vol.
3 (1953) p. 711–720.
I. Kaplansky, “Commutative Rings”, Allyn and Bacon, Inc., Boston, Mass. 1970.
A. Sasane, On the Krull dimension of rings of transfer functions, Acta Appl. Math. Vol.
103 (2008) p. 161–168.
Department of Mathematics, University of California, Davis, CA 95616
E-mail address: [email protected]
| 0math.AC
|
1
Interpolation and Extrapolation of Toeplitz Matrices
via Optimal Mass Transport
arXiv:1711.03890v1 [eess.SP] 10 Nov 2017
Filip Elvander, Student Member, IEEE, Andreas Jakobsson, Senior Member, IEEE,
and Johan Karlsson, Member, IEEE
Abstract—In this work, we propose a novel method for quantifying distances between Toeplitz structured covariance matrices.
By exploiting the spectral representation of Toeplitz matrices,
the proposed distance measure is defined based on an optimal
mass transport problem in the spectral domain. This may then
be interpreted in the covariance domain, suggesting a natural
way of interpolating and extrapolating Toeplitz matrices, such
that the positive semi-definiteness and the Toeplitz structure of
these matrices are preserved. The proposed distance measure is
also shown to be contractive with respect to both additive and
multiplicative noise, and thereby allows for a quantification of
the decreased distance between signals when these are corrupted
by noise. Finally, we illustrate how this approach can be used
for several applications in signal processing. In particular, we
consider interpolation and extrapolation of Toeplitz matrices,
as well as clustering problems and tracking of slowly varying
stochastic processes.
Keywords Covariance interpolation, Optimal mass transport, Toeplitz matrices, Spectral estimation
I. I NTRODUCTION
Statistical modeling is a key methodology for estimation
and identification and is used throughout the signal processing
field. An intrinsic component of such models is covariance
estimates, which are extensively used in application areas such
as spectral estimation, radar, and sonar [1], [2], wireless channel estimation, medical imaging, and identification of systems
and network structures [3], [4]. Although being a classical
subject (see, e.g., [5]), covariance estimation has recently
received considerable attention. Such contributions include
works on finding robust covariance estimates with respect to
outliers, as well as methods suitable for handling different
distribution assumptions, including families of non-Gaussian
distributions [6]–[10]. Another important class of problems is
covariance estimation with an inherent geometry that gives
rise to a structured covariance matrix. Such structures could
arise from stationarity assumptions of the underlying object
[11]–[15] or be due to assumptions in, e.g., the underlying
network structures in graphical models [16], [17]. In this work,
we focus on Toeplitz structures which naturally arise when
modeling stationary signals and processes.
This work was supported in part by the Swedish Research Council.
Parts of the material herein has been submitted for possible publication
to the IEEE International Conference on Acoustics, Speech and Signal
Processing, Calgary, Canada, April 2018.
F. Elvander and A. Jakobsson are with the Center for
Mathematical Sciences, Lund University, SE-22100 Lund, Sweden
(email: [email protected], [email protected]). J. Karlsson is
with the Department of Mathematics, KTH Royal Institute of Technology,
Stockholm, Sweden (email: [email protected]).
Although many methods rely on stationarity for modeling
signals, such assumptions are typically not valid over longer
time horizons. Therefore, tools for interpolation and morphing
of covariance matrices are important for modeling and fusing
information. Several such tools for interpolating covariances
have recently been proposed in the literature, for example
methods based on g-convexity [9], optimal mass transport
[18], and information geometry [19]. An alternative approach
for such interpolation is to relax, or ”lift”, the covariances
and instead consider interpolation between the lifted objects.
For example, in [20] (see also [21]), interpolation between
covariance matrices is induced from the optimal mass transport
geodesics between the Gaussian density functions with the
corresponding covariances.
The topic of optimal mass transport (see, e.g., [22], [23])
was originally introduced in order to address the problem
of, in a cost efficient way, supplying construction sites with
building material and has been used in many contexts, such
as, e.g., economics and resources allocation. Lately, it has also
gained interest in application fields such as image processing
[24], [25] signal processing [26]–[28], computer vision and
machine learning [29]–[33]. In this work, we will utilize
optimal mass transport to model changes in the covariance
structure of stochastic processes, or signals. To this end, we
propose a new lifting approach, where the lifting is made from
the covariance domain to the frequency domain, using the fact
that any positive semi-definite Toeplitz matrix has a spectral
representation. We combine this approach with the frequency
domain metric based on optimal mass transport, proposed in
[27], in order to define pairwise distances between Toeplitz
matrices. This is done by considering the minimum distance,
in the optimal mass transport sense, between the sets of power
spectra consistent with each of the Toeplitz matrices. The
proposed distance measure is shown to be contractive with
respect to additive and multiplicative noise, i.e., it reflects the
increased difficulty of discriminating between two stochastic
processes if these are corrupted by two realizations of a noise
process. Also, we show that the proposed distance measure
gives rise to a natural way of interpolating and extrapolating
Toeplitz matrices. The interpolation method preserves the
Toeplitz structure, the positive semi-definiteness, as well as
the zeroth moment of the interpolating/extrapolating matrices.
The proposed optimal mass transport problem is in its
original form an infinite-dimensional problem. As an alternative to finding solutions using approximations based on
discretizations of the underlying space, we show that certain
formulations of the problem allows for approximations by a
2
semi-definite program using a sum-of-squares representation.
Also, we illustrate how the method can be used for interpolation, extrapolation, tracking, and clustering.
This paper is organized as follows. In Section II, we provide
a brief background on the moment problem, i.e., determining
the power spectrum from a partial covariance sequence or
finite covariance matrix, as well as introduce the problem of
optimal mass transport. Section III introduces the proposed
distance notion for positive semi-definite Toeplitz covariance
matrices. Here, the dual problem is derived, and properties of
the proposed distance notion are described. In Section IV, we
describe applications of the proposed distance notion, such as
induced interpolation, extrapolation, tracking, and clustering.
Section V formulates a sum-of-squares relaxation of the dual
problem. Section VI provides numerical illustrations of the
proposed distance notion, as well as the described applications.
Finally, Section VII concludes upon the work.
Notation
Let Mn denote the set of Hermitian n × n matrices and let
(·)T denote the transpose, (·)H the Hermitian transpose, and
(·) the complex conjugate. Let T = (−π, π] and let Cperio (T)
denote the set of continuous and 2π-periodic functions on T.
The set of linear bounded functionals on Cperio (T), or equivalently, the dual space of Cperio (T), is the set of generalized
integrable functions on the set T, here denoted by M(T).
Thus, M(T) includes, e.g., functions containing singular parts
such as Dirac delta functions [34].1 Further, we let M+ (T)
denote the subset of such functions that are non-negative. We
use hΦ, f i to denote the application of the functional Φ on f ,
e.g.,
Z
hΦ, f i =
f (θ)Φ(θ)dθ
T
if f ∈ Cperio (T) and Φ ∈ M(T). For Hilbert spaces, h·, ·i is
the standard inner product, e.g., when X and Y are vectors
or matrices, then hX, Yi = trace(XYH ). We denote matrices
by boldface upper-case letters, such as X, whereas vectors are
denoted by boldface lower-case letters, such as x.
II. BACKGROUND
A. Stochastic processes and spectral representations
We will in this work consider complex-valued discrete time
stochastic processes, or signals, y(t) for t ∈ Z. These will be
assumed to be zero mean and wide sense stationary (WSS),
i.e., E(y(t)) = 0 for all t ∈ Z, and the covariance
rk , E(y(t)y(t − k))
being independent of t. Here, E(·) denotes the expectation
operator. The frequency content of the process y(t) may then
be represented by the power spectrum, Φ, i.e., the non-negative
function on T whose Fourier coefficients coincide with the
covariances:
Z π
1
rk =
Φ(θ)e−ikθ dθ
(1)
2π −π
1
Strictly speaking M(T) is the set of signed bounded measures on T [35].
for k ∈ Z (see, e.g., [36, Chapter 2]). Typically in spectral
estimation, one considers the inverse problem of recovering
the power spectrum Φ from a given set of covariances rk ,
for k ∈ Z, with |k| ≤ n − 1. The condition for any such
reconstruction to be valid is that Φ should be consistent with
the covariance sequence {rk }|k|≤n−1 , i.e., (1) should hold for
|k| ≤ n − 1. The corresponding n × n covariance matrix is
r0
r1
r2
..
.
R=
rn−1
r−1
r0
r1
..
.
r−2
r−1
r0
..
.
···
···
···
..
.
rn−2
rn−3
···
r−n+1
r−n+2
r−n+3
..
.
(2)
r0
which is a Hermitian Toeplitz matrix, since y(t) is WSS. Thus,
expressed in the form of matrices, a spectrum is consistent with
an observed partial covariance sequence, or, equivalently, a finite covariance matrix, if Γ(Φ) = R, where Γ : M(T) → Mn
is the linear operator
Z
1
a(θ)Φ(θ)a(θ)H dθ
(3)
Γ(Φ) ,
2π T
and
a(θ) ,
1
eiθ
···
ei(n−1)θ
T √
/ n
(4)
is the Fourier vector. Note that Γ(Φ) is a Toeplitz matrix,
since a(θ)a(θ)H is Toeplitz for any θ. It may be noted that
for any positive semi-definite Toeplitz matrix, R, there always
exists at least one consistent power spectrum; in fact, if R
is positive definite, there is an infinite family of consistent
power spectra [37]. It may be noted that for singular Toeplitz
covariance matrices, the spectral representation is unique. This
fact has recently been successfully utilized in atomic norm
minimization problems for grid-less compressed sensing of
sinusoidal signals [38]. In this work, we are mainly interested
in the non-singular case, where several power spectra are
consistent with given covariance matrices. In Section III, we
will utilize such spectral representations in order to define a
notion of distance between pairs of Toeplitz matrices. This
distance will be defined in terms of the minimum optimal mass
transport cost between the sets of power spectra consistent
with the matrices.
B. Optimal mass transport
The Monge-Kantorovich problem of optimal mass transport
is the problem of finding an optimal transport plan between
two given mass distributions [22], [23]. The cost of moving a
unit mass is defined on the underlying space, and the optimal
transport plan is defined as the plan minimizing the total
cost. The resulting minimal cost, associated with the optimal
transport plan, can then be used as a measure of similarity,
or distance, between the two mass distributions. The idea of
utilizing the optimal mass transport cost as a distance measure
has been used, e.g., for defining metrics on the space of power
spectra [27], whereas the optimal transport plan has been
used for tracking stochastic processes with smoothly varying
spectral content and for spectral morphing for speech signals
3
[28]. Recently, the interpretation of the optimal transport plan
as providing an optimal association between elements in two
mass distributions has been used as a means of clustering in
fundamental frequency estimation algorithms [39].
As in [27], we consider the following distance between two
spectra Φ0 and Φ1 :
Z
S(Φ0 , Φ1 ) ,
min 2
c(θ, ϕ)M (θ, ϕ)dθdϕ (5a)
M ∈M+ (T )
T2
Z
subject to Φ0 (θ) =
M (θ, ϕ)dϕ (5b)
ZT
Φ1 (ϕ) =
M (θ, ϕ)dθ (5c)
T
where T2 = T × T denotes the 2-D frequency space. Here,
the cost function, c(θ, ϕ), details the cost of moving one
unit of mass between the frequencies θ and ϕ. The transport
plan, M (θ, ϕ), specifies the amount of mass moved from
frequency θ to frequency ϕ. The objective in (5a) is the total
cost associated with the transport plan M and the constraints
(5b) and (5c) ensure that M is a valid transport plan from
Φ0 to Φ1 , i.e., the integration marginals of M coincide with
the spectra Φ0 and Φ1 . It may be noted that, due to these
marginal constraints, the distance measure S is only defined
for spectra of the same mass, or total power. However, S
may be generalized in order to allow for mass differences by
including a cost for adding and subtracting mass by postulating
that the spectra Φ0 and Φ1 are perturbations of functions
Ψ0 and Ψ1 that have equal mass. As in [27], this may be
formulated as
1
X
Sκ (Φ0 , Φ1 ) =
min
S(Ψ0 , Ψ1 ) + κ
kΦj − Ψj k1
Ψj ∈M+ (T)
Then, Wκ (Φ0 , Φ1 ) is contractive with respect to the additive
and normalized multiplicative noise, i.e., with respect to
• Φ 7→ Φ + Φa
• Φ 7→ Φ ∗ Φm
R
for any Φa , Φm ∈ M+ (T), with T Φm (θ)dθ = 1.
As we shall see, it is possible to construct notions of distance on the space of positive semi-definite Toeplitz matrices
that have properties similar to those stated in Theorem 2.
III. A NOTION OF DISTANCE FOR T OEPLITZ MATRICES
As noted above, any positive semi-definite Toeplitz matrix
R has at least one spectral representation, i.e., there exists at
least one spectrum Φ that is consistent with it. Thus, we define
the distance, T , between two positive semi-definite Toeplitz
matrices, R0 and R1 , as the minimum transportation cost, as
measured by S, between spectra consistent with the respective
matrices, i.e.,
T (R0 , R1 ) ,
min
S(Φ0 , Φ1 )
Φ0 ,Φ1 ∈M+ (T)
(8)
Γ(Φj ) = Rj , j = 0, 1
subject to
Considering the definition of S in (5), this can equivalently be
formulated as the convex optimization problem
Z
T (R0 , R1 ) =
min 2
c(θ, ϕ)M (θ, ϕ)dθdϕ
M ∈M+ (T ) T2
Z
subject to Γ
M (θ, ϕ)dϕ = R0 (9)
ZT
Γ
M (θ, ϕ)dθ = R1
j=0
T
(6)
where κ > 0 is a used-defined parameter detailing the cost
of adding or subtracting mass. Also, if the cost function
in the optimal mass transport problem in (5) is chosen as
c(θ, φ) = d(θ, φ)p , p ≥ 1 for any metric d(θ, φ) on T,
then W (Φ0 , Φ1 ) = S(Φ0 , Φ1 )1/p is the so-called Wasserstein
metric on M+ (T). Similarly, for Sκ , the following theorem
holds.
Theorem 1 ([27]). Let p ≥ 1, κ > 0, and let the cost function
be c(θ, φ) = |θ − φ|p . Then,
Wκ (Φ0 , Φ1 ) = Sκ (Φ0 , Φ1 )1/p
(7)
is a metric on M+ (T).
Consider a situation where we need to discriminate between
two signals on the basis of their statistics or of their power
spectra. In such cases, additive or multiplicative noise typically
impede our ability to differentiate between the two. This was
considered in [27] and it was shown that the transportation
distance respects this property in the sense that corrupting
two signals with additive and (normalized) multiplicative
noise decreases their transportation distance. Specifically, the
following theorem holds.
Theorem 2 ([27]). Let p ≥ 1, κ > 0, and let the cost function
be c(θ, φ) = |θ − φ|p , and let Wκ (Φ0 , Φ1 ) be defined by (7)
Note that the formulation in (9) is only defined for covariance matrices with the same zeroth covariances, or, equivalently, covariance matrices whose consistent spectra have the
same mass. However, in order to allow for mass differences,
T (R0 , R1 ) can be generalized in analog with (6) as
Tκ (R0 , R1 ) ,
min
Φj ∈M+ (T)
subject to
Sκ (Φ0 , Φ1 )
(10)
Γ(Φj ) = Rj for j = 0, 1
or, equivalently,
Z
Tκ (R0 , R1 ) =
min
M ∈M+ (T2 )
Ψ0 ,Ψ1 ∈M+ (T)
c(θ, ϕ)M (θ, ϕ)dθdϕ
T2
Z
M (θ, ϕ)dϕ − Ψ0
+κ
(11)
1
T
Z
M (θ, ϕ)dθ − Ψ1
+κ
T
1
subject to Γ(Ψj ) = Rj for j = 0, 1
Typically, the cost function c would be selected to be symmetric in its arguments, in which case Tκ would be symmetric
as well, i.e., Tκ (R0 , R1 ) = Tκ (R1 , R0 ) for any Toeplitz
matrices R0 , R1 ∈ M+ . Although many possible choices of
such cost functions exist, we will in most examples presented
2
herein consider c(θ, ϕ) = eiθ − eiϕ , i.e., the cost function
quantifies distances as the square of the distance between the
4
corresponding points on the unit circle. As we show in Section
V, this particular choice of cost function allows for a sum-ofsquares relaxation of the dual formulation of (11). This dual
formulation is presented next.
where
Ωκc = Λ0 , Λ1 ∈ Mn |
Γ∗ (Λ0 )(θ) + Γ∗ (Λ1 )(ϕ) ≤ c(θ, ϕ),
Γ∗ (Λ0 )(θ) ≤ κ, Γ∗ (Λ1 )(ϕ) ≤ κ for all θ, ϕ ∈ T
A. The dual formulation
In order to study properties of the distance notion Tκ ,
we consider the dual formulation of (11), where we assume
that the cost function, c, is non-negative on T2 . In order to
address this problem, we first note that the adjoint operator2
Γ∗ : Mn → Cperio (T) of the operator Γ is
Γ∗ (R)(θ) = a(θ)H Ra(θ)
where we have used that the cost function c is non-negative.
This yields the dual problem.
Proposition 1. For any cost function c(θ, ϕ), such that
c(θ, ϕ) ≥ 0, for all θ, ϕ ∈ T, the dual problem of (11) is
maximize
(Λ0 ,Λ1 )∈Ωκ
c
hΛ0 , R0 i + hΛ1 , R1 i
(12)
where
since
Ωκc = Λ0 , Λ1 ∈ Mn |
Z
Γ∗ (Λ0 )(θ) + Γ∗ (Λ1 )(ϕ) ≤ c(θ, ϕ),
hΓ(Φ), Ri = h a(θ)Φ(θ)a(θ)H dθ, Ri
ZT
=
Φ(θ)a(θ)H Ra(θ)dθ = hΦ, Γ∗ (R)i
Γ∗ (Λ0 )(θ) ≤ κ, Γ∗ (Λ1 )(ϕ) ≤ κ ∀θ, ϕ ∈ T
T
where, in the first line, the inner product is the one associated
with Mn , and the second line is the bilinear form with
arguments Γ∗ (R) ∈ Cperio (T) and Φ ∈ M+ (T). With this
result, we can derive an expression of the dual problem by
considering the Lagrangian relaxation of (11). The Lagrangian
is given by
Z
Lκ (M, Ψ0 , Ψ1 , Λ0 , Λ1 ) =
c(θ, ϕ)M (θ, ϕ)dθdϕ
Since the primal and dual problems are convex and the set
of feasible points Ωκc has non-empty interior for any κ ≥ 0,
Slater’s condition (see, e.g., [40]) gives that strong duality
holds and hence the duality gap between (11) and (12) is zero.
Also, note the strong resemblance in form compared to the
dual of (6), which is given by [27]
Z
Z
maximize
λ0 (θ)Φ0 (θ)dθ + λ1 (ϕ)Φ1 (ϕ)dϕ
λ0 ,λ1 ∈Cperio (T)
subject to
T2
+ hR0 −Γ(Ψ0 ), Λ0 i + hR1 −Γ(Ψ1 ), Λ1 i
Z
Z
M (θ, ϕ)dθ − Ψ1
M (θ, ϕ)dϕ − Ψ0 + κ
+κ
T
T
Z 1
= hΛ0 , R0 i + hΛ1 , R1 i +
c(θ, ϕ)M (θ, ϕ)dθdϕ
λ1 (ϕ) ≤ κ ∀ϕ ∈ T
1
Similarly, for the case when the zeroth covariances of R0 and
R1 are required to be equal, the dual problem is
T2
1
T
maximize
(Λ0 ,Λ1 )∈Ωc
1
T
where 1{·} is the indicator function. Using this, and considering the infimum with respect to M , we arrive at
(
hΛ0 , R0 i+hΛ1 , R1 i if (Λ0 , Λ1 ) ∈ Ωκc
inf 2 Lκ =
−∞
otherwise
M ∈M+ (T )
2 Strictly
speaking, this is the pre-adjoint operator of Γ.
hΛ0 , R0 i + hΛ1 , R1 i
(13)
where
Note that the Lagrange multiplier matrices Λ0 and Λ1 may be
taken as Hermitian matrices, as R0 − Γ (Ψ0 ) and R1 − Γ (Ψ1 )
are Hermitian, and thus all inner products are real. Considering
the infimum of Lκ with respect to Ψ0 and Ψ1 , it may be noted
that this is only finite if Γ∗ (Λ0 )(θ) ≤ κ and Γ∗ (Λ1 )(ϕ) ≤ κ,
for all θ, ϕ ∈ T. If this is satisfied, the Lagrangian is, for any
fixed non-negative M , minimized by Ψ0 and Ψ1 given by
Z
Ψ0 (θ) = 1{Γ∗ (Λ0 )(θ)∈[−κ,κ]} M (θ, ϕ)dϕ
ZT
Ψ1 (ϕ) = 1{Γ∗ (Λ1 )(ϕ)∈[−κ,κ]} M (θ, ϕ)dθ
Ψ0 ,Ψ1 ∈M+ (T)
T
λ0 (θ) ≤ κ ∀θ ∈ T
− hΨ0 , Γ∗ (Λ0 )i − hΨ1 , Γ∗ (Λ1 )i
Z
Z
M (θ, ϕ)dϕ − Ψ0 + κ
M (θ, ϕ)dθ − Ψ1
+κ
T
T
λ0 (θ)+λ1 (ϕ) ≤ c(θ, ϕ) ∀θ, ϕ ∈ T
Ωc = Λ0 , Λ1 ∈ Mn |
Γ∗ (Λ0 )(θ) + Γ∗ (Λ1 )(ϕ) ≤ c(θ, ϕ) ∀θ, ϕ ∈ T
This is the dual problem of (8), where adding and subtracting
mass is not allowed in the transport problem.
B. Properties of distance notion Tκ
For the distance notion Tκ in (10), the following proposition
holds.
Proposition 2. Let the cost function c be non-negative and
such that c(θ, θ) = 0, for all θ ∈ T. Then, the distance measure
Tκ defined in (11) is contractive with respect to additive noise.
If c is also shift-invariant, i.e., c(θ − φ, ϕ − φ) = c(θ, ϕ), for
all θ, ϕ, φ ∈ T, then Tκ is also contractive with respect to
multiplicative noise whose zeroth covariance is smaller than
or equal to unity.
The statement of this proposition is that when two stochastic
processes become contaminated by noise, the distance notion
Tκ decreases and hence the processes become harder to
5
distinguish. Intuitively, this is a desirable property of Tκ , as
additive or multiplicative noise should indeed impede ones
ability to discriminate between two processes. Proposition 2
may be proven by utilizing results in [27]. However, in the
interest of making the exposition self-contained, we will here
provide a direct proof based on the dual formulation in (12).
Here, we prove the first part of the proposition, i.e., the case of
additive noise. For the case of multiplicative noise, the proof
is deferred to Appendix A.
Proof. Consider two processes with covariance matrices R0
and R1 and assume that they are both additively corrupted
by an independent noise process with covariance Rw . This
results in processes with covariances R00 = R0 + Rw and
R01 = R1 + Rw , respectively. From the dual formulation in
(12), the distance between these covariance matrices is given
by
Tκ (R00 , R01 ) =
max
hΛ0 , R0 i + hΛ1 , R1 i
κ
(Λ0 ,Λ1 )∈Ωc
+ hΛ0 +Λ1 , Rw i
Let Φw be any spectrum consistent with the noise covariance
Rw , i.e., with Γ(Φw ) = Rw . We then have that
hΛ0 + Λ1 , Rw i = hΛ0 + Λ1 , Γ(Φw )i
= hΓ∗ (Λ0 ) + Γ∗ (Λ1 ), Φw i
Z
=
(Γ∗ (Λ0 )(θ) + Γ∗ (Λ1 )(θ)) Φw (θ)dθ
T
As (Λ0 , Λ1 ) ∈ Ωκc , it holds that
Γ∗ (Λ0 )(θ) + Γ∗ (Λ1 )(θ) ≤ c(θ, θ) = 0 for all θ ∈ T
and since Φw ≥ 0, we get
hΛ0 + Λ1 , Rw i ≤ 0
Hence, it follows that
Tκ (R00 , R01 ) ≤
max
hΛ0 , R0 i+hΛ1 , R1 i = Tκ (R0 , R1 )
(Λ0 ,Λ1 )∈Ωc
It may be noted that the assumptions made in Proposition 2 regarding the cost function c are quite mild, allowing
for a large class of potential cost functions. In particular,
2
c(θ, ϕ) = eiθ − eiϕ satisfies the conditions of Proposition 2
for both additive and multiplicative noise.
IV. I NTERPOLATION , EXTRAPOLATION , AND TRACKING
The formulation in (8) does not only define a notion of distance between two Toeplitz covariance matrices, R0 and R1 ;
it also provides a means of forming interpolating matrices, i.e.,
defining intermediate covariance matrices Rτ , for τ ∈ (0, 1).
In order to define the interpolating matrices, we again utilize
the spectral representation of positive semi-definite Toeplitz
matrices. To this end, we note that, given an optimal transport
plan, M , found as the functional minimizing R(8), one may
define
R spectra intermediate to the marginals T M (θ, ϕ)dϕ
and T M (θ, ϕ)dθ by linearly shifting the frequency locations
of the spectral mass, as dictated by M . That is, any mass
transferred from φ to φ + ϕ is defined to, at τ ∈ [0, 1],
be located at frequency φ + τ ϕ. Using this, the intermediate
spectrum is given by
Z
ΦM
(θ)
,
δθ ({φ + τ ϕ}mod T )M (φ, φ + ϕ)dφdϕ
τ
ZT2
=
M (θ − τ ϕ, θ + (1 − τ )ϕ)dϕ
(14)
T
Here, δθ is the Dirac delta function localized at θ, i.e.,
δθ (φ) , δ(φ − θ) and the integrands are extended periodically
with period 2π outside the domain of integration. Also, we
denote by {x}mod T the value in T that is congruent with x
modulo 2π. Based on this definition of the interpolating spectrum ΦM
τ , the corresponding interpolating covariance matrix,
Rτ is defined as the unique Toeplitz matrix consistent with
this spectrum, i.e.,
Rτ , Γ(ΦM
(15)
τ )
Z
Z
= a(θ)
M (θ − τ ϕ, θ + (1 − τ )ϕ)dϕ a(θ)H dθ
T
ZT
=
a({φ+τ ϕ}mod T )M (φ, φ+ϕ)a({φ+τ ϕ}mod T )H dφdϕ
T2
for τ ∈ [0, 1]. To simplify the following exposition, let
Iτ (M ) , Γ(ΦM
τ ) denote the linear operator in (15) that maps
a transport plan to an interpolating covariance matrix, i.e.,
Rτ = Iτ (M ). It may be noted from (15) that Iτ (M ) is welldefined also for τ 6= [0, 1], i.e., the formulation allows also
for extrapolation. The following proposition follows directly
from the definition in (15).
Proposition 3. For any τ ∈ R, the following basic properties
hold for Rτ = Iτ (M ):
a) If R0 and R1 have the same zeroth covariance, then it
is also the zeroth covariance of Rτ .
b) The matrix Rτ is a Toeplitz matrix.
c) The matrix Rτ is positive semi-definite.
Due to these properties, the proposed method offers a way
of interpolating the covariances of, e.g., slowly varying time
series as the interpolant Rτ allows for modeling linear changes
in the spectrum of the process.
Remark 1. The interpolation approach generalizes trivially
to the formulation in (11) between the covariances Γ(ΦM
0 )
and Γ(ΦM
)
under
the
assumption
that
Ψ
and
Ψ
are
0
1
1
M
perturbations of ΦM
0 and Φ1 , respectively. In order to define
an interpolation and extrapolation procedure from R0 to R1
where there is a cost κ for adding and subtracting mass, one
may, along with the interpolation path Iτ (M ), linearly add
the part corresponding to the added and subtracted mass, i.e.,
M
Rτ = Iτ (M ) + (1 − τ )Γ(Ψ0 − ΦM
0 ) + τ Γ(Ψ1 − Φ1 ) (16)
Note that in this scenario, positive semi-definiteness cannot be
guaranteed for the extrapolation case.
A. Comparison with other methods
The properties in Proposition 3 distinguish the proposed
interpolant Rτ from other proposed matrix geodesics. As an
6
and the geodesic in [20], [21], which builds on optimal mass
transport of Gaussian distributions and can be expressed as
H
1/2
1/2
1/2
1/2
R̆τ = (1 − τ )R0 +τ R1 U (1 − τ )R0 +τ R1 U
(18)
where
−1/2
U = R1
−1/2
R0
1/2
1/2
R0 R1 R0
1/2
Although both of these geodesics preserve positive semidefiniteness, they neither preserve the Toeplitz structure nor
the zeroth covariance. As noted above, the three properties
in Proposition 3 hold for any τ ∈ R for the proposed
approach, and thus directly allows for extrapolating using (15).
In contrast, it may be noted that for the linear combination
there are no guarantees that the resulting matrix is
Rconv
τ
positive semi-definite if τ ∈
/ [0, 1]. Also, note that the alternative geodesics, R̃τ and R̆τ , do not naturally generalize to
extrapolation.
B. Tracking of slowly varying processes
The proposed interpolant Rτ = Iτ (M ) may also be readily
used for tracking slowly varying stochastic process. As noted
above, Iτ (M ) allows for the modeling of slow, i.e., locally
linear, shifts in the location of spectral power. Building on
this property, we can extend the optimal transport problem
in (8), in order to fit a covariance path Rτ to a sequence
of J covariance matrix estimates, R̂τj , for j = 1, . . . , J. As
Rτ is unambiguously determined from a transport plan M
via Iτ (M ), this tracking problem may be formulated as the
convex optimization problem
Z
c(θ, ϕ)M (θ, ϕ)dθdϕ+λ
minimize
M ∈M+ (T2 )
T2
J
X
j=1
2
Iτj (M )− R̂τj
F
(19)
where λ > 0 is a user-specified regularization parameter.
As may be noted from (19), the optimal transport plan M
is here determined as the one that minimizes not only the
transport cost, but also takes into account the deviations of
the interpolant Rτj = Iτj (M ) from the available covariance
matrix estimates R̂τj . The behavior of this construction is
illustrated in the numerical section.
1.2
r 1, = 0
r 1, = 1
1
r 1,
unit circle
imaginary part
example, consider the basic method of forming interpolants
using convex combinations of R0 and R1 , i.e., Rconv
=
τ
τ R0 + (1 − τ )R1 , for τ ∈ [0, 1]. This preserves the Toeplitz
structure, as well as the zeroth covariance and the positive
semi-definiteness. However, from a spectral representation
point of view, the convex combination gives rise to fade-in
fade-out effects, i.e., only spectral modes directly related to
R0 and R1 can be represented, and there can be no shift
in the location of these modes (see also Example VI-B and
Figure 3). Other more sophisticated options include, e.g., the
geodesic with respect to g-convexity [9]
τ
1/2
−1/2
−1/2
1/2
R̃τ = R0
R0 R1 R0
R0
(17)
0.8
0.6
0.4
0.2
0
-1
-0.5
0
0.5
1
real part
Fig. 1. Path of the off-diagonal element of the covariance matrix Rτ for the
case n = 2, when choosing R0 and R1 according to (26).
C. Clustering: Barycenter computation
As a further example, we will see that the barycenter
with respect to the distance notion Tκ may be formulated as
a convex optimization problem. This might be desirable in
clustering or classification applications, where one is interested
in either identifying classes of signals or processes based
on their covariance matrices or associate a given covariance
matrix with such a signal class. Considering the case of clustering, assume that L covariance matrices, R` , ` = 1, . . . , L,
are available. Then, we may define their barycenter via Tκ
according to
Rbary = arg min
R∈Mn
+
L
X
Tκ (R, R` )
(20)
`=1
i.e., as the covariance matrix that minimizes the sum of Tκ
for the set of covariance matrices R` . Explicitly, Rbary solves
the convex optimization problem
L Z
X
c(θ, ϕ)M` (θ, ϕ)dθdϕ
minimize
n
R∈M+
M` ∈M+ (T2 )
Ψ` ∈M+ (T)
Φ` ∈M+ (T)
`=1
+κ
+κ
T2
Z
L
X
`=1
L
X
`=1
subject to
M` (θ, φ)dθ − Φ`
1
T
(21)
Z
M` (θ, φ)dφ − Ψ`
1
T
Γ(Φ` ) = R
Γ(Ψ` ) = R` , ` = 1, . . . , L
This formulation allows for using, e.g., K-means clustering
(see, e.g., [41]) in order to identify classes of covariance
matrices, as well as classify a given covariance matrix according to these classes. Classification of a covariance matrix
(j)
R according to classes defined by a set of barycenters Rbary ,
j = 1, . . . , J, may then be formulated as
(j)
arg min Tκ (R, Rbary )
j∈{1,...,J}
(22)
7
2
0.4
1.5
0.2
2
0.6
corr
corr
0.6
0.4
1.5
0.2
1
0
1
0
0.5
-100
0.5
-100
-50
0
0
50
100
(degrees)
In Section VI-D, we present a simple illustration of this
potential application, considering unsupervised clustering of
phonemes.
V. S UM - OF - SQUARES RELAXATION
In order to solve the dual problem in (12) in practice, it has
to be implemented as a finite dimensional problem, e.g., by
gridding the space T2 and thereby approximating the set Ωκc
using a finite number of constraints. However, for the special
2
case of c(θ, ϕ) = eiθ − eiϕ , this can also be done using
a sum-of-squares (SOS) relaxation (see, e.g., [42]) of (12).
For simplicity of notation, we present only the case without
κ (κ = ∞), i.e., the dual of (8) as formulated in (13). To this
end, identify
z = eiθ , w = eiϕ
[Λ1 ]k,` w`−k
Thus, the set of constraints defining the feasible set Ωc is given
by Γ∗ (Λ0 )(θ) + Γ∗ (Λ1 )(ϕ) ≤ c(θ, ϕ), for all θ, ϕ ∈ T, or,
equivalently,
[Λ0 ]k,` z `−k −
k,`=1
n
X
[Λ1 ]k,` w`−k ≥ 0
Note that in the two-dimensional trigonometric polynomial
(23), the coefficient for z −k1 w−k2 is equal to
2 − diag (Λ0 + Λ1 ) 1
T
−diag (Λ0 , k1 ) 1
T
−diag (Λ1 , k2 ) 1
−1
−1
0
where Zn\0 = {k ∈ Z | |k| ≤ n, k 6= 0}. Here, diag (X, k)
denotes the column vector containing the elements on the kth
super-diagonal of the matrix X, if k > 0, and the elements
on the kth sub-diagonal, if k < 0, with 1 denoting a column
vector of ones of appropriate dimension.
In order to formulate a computationally feasible optimization problem, we remove the non-negativity constraint for the
two-dimensional polynomial in (23) and instead impose that
it should have a sum-of-squares representation [42], [43]. In
particular, we impose that the polynomial in (23) should be of
the form
P (z, w) = z−1
T
Qz
(24)
and ⊗ is the Kronecker product. Note that any polynomial on
this form is non-negative by definition, and, furthermore, for
any non-negative polynomal P ∗ , there is a sequence of polynomials, Pm , on the form (24) such that kPm − P ∗ k∞ → 0
as m → ∞.
Next, note that the coefficients of P are associated to the
elements of Q in (24) according to
k,`=1
(23)
T
Fig. 3. Interpolated spatial spectrum estimated as a(θ)H Rconv
a(θ), where
τ
is the linear combination of R0 and R1 .
Rconv
τ
T
T
z = 1 w · · · wm−1 ⊗ 1 z · · · z m−1
T
T
z−1 = 1 w−1 · · · w−m+1 ⊗ 1 z −1 · · · z −m+1
k=1 `=1
n
X
100
2
c(θ, ϕ) = 2 − zw−1 − z −1 w
n X
n
X
Γ∗ (Λ0 )(θ) =
[Λ0 ]k,` z `−k
2−zw−1 −z −1 w−
0
50
where Q ∈ Mm is positive semi-definite,
which allows us to write
Γ∗ (Λ1 )(ϕ) =
0
(degrees)
Fig. 2. Interpolated spatial spectrum estimated as a(θ)H Rτ a(θ), where Rτ
is obtained by solving (9).
k=1 `=1
n X
n
X
-50
for k1 = k2 = 0
for k1 ∈ Zn\0 and k2 = 0
for k1 = 0 and k2 ∈ Zn\0
for k1 = 1 and k2 = −1
for k1 = −1 and k2 = 1
otherwise
pk1 ,k2 = trace (Tk1 ,k2 Q) ,
for − m + 1 ≤ k1 , k2 ≤ m − 1
where
Tk1 ,k2 = Tk2 ⊗ Tk1
and Tk is the matrix with ones on the kth diagonal and zeros
elsewhere. Putting these facts together by requiring that the
polynomial (23) can be written as a sum-of-square (24), we
8
2
1.5
1
2
1
corr
corr
2
1.5
0.5
1
1
0
0
0.5
-100
-50
0.5
-100
0
50
(degrees)
-50
0
100
0
50
(degrees)
0
100
Fig. 4. Interpolated spatial spectrum estimated as a(θ)H R̃τ a(θ), where R̃τ
is obtained from (17).
Fig. 5. Interpolated spatial spectrum estimated as a(θ)H R̆τ a(θ), where R̆τ
is obtained from (18).
approximate (13) by the semi-definite program (SDP)
matrices of dimension n = 2. Consider covariance matrices
of the form
5π
1 r
1
rei 6
(26)
R0 =
, R1 =
5π
r 1
1
re−i 6
maximizen
Λ0 ,Λ1 ∈M
Q∈M
hΛ0 , R0 i + hΛ1 , R1 i
(25)
m2
subject to
Q0
T
trace (Q) = 2 − diag (Λ0 + Λ1 ) 1
trace (T−1,1 Q) = −1
trace (T1,−1 Q) = −1
T
trace (Tk1 ,0 Q) = −diag (Λ0 , k1 ) 1 , k1 ∈ Zn\0
T
trace (T0,k2 Q) = −diag (Λ1 , k2 ) 1 , k2 ∈ Zn\0
trace (Tk1 ,k2 Q) = 0 , if k1 k2 = 1
or 1 < |k1 k2 | ≤ m2
It is worth noting that the P defined by the optimal Q in
(24) is non-negative (actually, strictly positive) on the unit
torus. Thus, for the solution of the SDP problem in (25), there
will be a corresponding feasible solution of (13). Therefore,
any solution to (25) will give a lower bound for the optimal
objective value of (13). However, any non-negative polynomial
may be arbitrarily well approximated by P of the form (24) by
a suitable choice the degree m, and thus the maximal objective
value of (25) will converge to the maximal objective value
of (13) as m grows. By comparison, directly discretizing T,
thereby approximating Ωc using a finite number of constraints,
yields an upper bound for the maximal objective value of (13).
In this case, the maximal objective value will also converge
to that of (13), this time from above, as the spacing of the
discretization of the grid becomes finer.
VI. N UMERICAL EXAMPLES
In this section, we present some numerical examples illustrating different aspects and application areas of the proposed
distance notion Tκ , as well as the interpolant Rτ = Iτ (M ).
where r ∈ [−1, 1]. Thus, considering interpolating paths Rτ ,
these will be on the form
1
r1,τ
Rτ =
(27)
r̄1,τ
1
Figure 1 displays the real and imaginary part of r1,τ for
τ ∈ [0, 1] when varying the magnitude r of the off-diagonal
element of R0 and R1 between 0 and 1. As can be seen,
the trajectories of r1,τ approximately correspond to convex
5π
combinations (1 − τ )r + τ rei 6 when r is close to zero,
whereas they are considerably curved for r closer to 1. It
may be noted that for the singular case, i.e., r = 1, the
trajectory of r1,τ coincides with the unit circle. Thus, we see
that for covariance matrices that have consistent spectra that
are essentially flat, the interpolant Rτ will be approximately
equal to the convex combination (1−τ )R0 +τ R1 . At the other
extreme, the interpolant corresponding to covariance matrices
R0 and R1 that have consistent spectra that are close to being
singular will also have almost singular consistent spectra.
B. Interpolation and extrapolation for DOA
Next, we illustrate the proposed methods ability to produce
interpolants Rτ that are consistent with locally linear changes
in the frequency location of spectral power. This is illustrated using a direction-of-arrival (DOA) estimation problem.
Consider a uniform linear array (ULA) with 15 sensors with
half-wavelength sensor spacing and a scenario where two
covariance matrices
2
1X
(0)
(0)
R0 =
a(θ` )aH (θ` ) + σ 2 I
2
`=1
A. Trajectory example
In order to illustrate the behavior of the proposed interpolation method, we consider a simple scenario with covariance
3
R1 =
1
1X
(1)
(1)
(1)
(1)
a(θ1 )aH (θ1 ) +
a(θ` )aH (θ` ) + σ 2 I
2
4
`=2
9
1
1
14
0.75
6
0.75
5
0.5
12
0.25
10
0.5
4
0.25
0
8
0
-0.25
6
-0.25
-0.5
4
-0.5
-0.75
2
-0.75
3
2
-1
0
0.2
0.4
0.6
-1
0.8
0
Fig. 6. Spectrum estimated as aH (θ)R̂a(θ), where R̂ is estimated as the
sample covariance matrix based on 100 samples in each window.
(0)
(1)
(0)
are available3 . Here θ1 = θ1 = −50◦ , θ2 = 30◦ ,
(1)
(1)
θ2 = 20◦ , and θ3 = 40◦ , and σ 2 = 0.05. Such a scenario
(0)
may be interpreted as a target at θ2 splitting up into two
(1)
(1)
targets at θ2 and θ3 as time progresses, whereas the target
(0)
at θ1 stays put. We use the proposed method in (9), as R0
and R1 have the same zeroth covariance, in order to find
the optimal transport map M . Then, using (15), we compute
covariance matrices Rτ , for τ ∈ [0, 2], i.e., we both interpolate
on τ ∈ [0, 1] and extrapolate on τ ∈ (1, 2]. This is then
based on convex
compared to the basic interpolant Rconv
τ
and linear combinations of R0 and R1 , as well as the more
sophisticated covariance matrix geodesics R̃τ and R̆τ , as
defined in (17) and (18), respectively. For these four cases, we
then estimate the corresponding inter- and extrapolated spectra
using the correlogram, i.e., as
Φcorr (X, θ) = a(θ)H Xa(θ)
1
(28)
where X is substituted for the four different covariance interpolants. The results for the proposed interpolant Rτ are shown
in Figure 2. As can be seen, Rτ indeed models a scenario
where one of the targets has a constant location, whereas
the second target splits upp into two smaller targets. Note
also that Rτ implies that the smaller target continue linearly
with respect to the look-angle, θ, also for the extrapolation
case, i.e., for τ > 1. In contrast, the convex (linear for
τ > 1) combination Rconv , as shown in Figure 3, display
undesirable behavior; clear fade-in fade-out effects are visible,
and non-negativity is violated as τ → 2 due to Rconv
becoming indefinite. Similar objections may be raised against
the geodesics R̃τ and R̆τ , shown in Figures 4 and 5, both
displaying fade-in fade-out effects and thereby fail to model
any displacement of the targets. Also, note that the total power
of the signal varies greatly as τ goes from 0 to 2, especially
for R̆τ .
3 Note that θ in this example denotes spatial frequency. For simplicity, we
retain the notation a(θ) also for this case.
0.2
0.4
0.6
0.8
1
Fig. 7. Spectrum estimated as aH (θ)Rτ a(θ), where Rτ is obtained by
solving (19), fitted to a sequence of five covariance estimates.
C. Tracking of an AR-process
Next, we illustrate the approach in (19) for the tracking
of signals with slowly varying spectral content. To this end,
consider a complex autoregressive (AR) process with one complex, time-varying pole. The pole is placed at a constant radius
of 0.9, and moves from the frequency 0.4π to 0.6π. Spectral
estimates based directly on covariance matrix estimates R̂
are shown in Figure 6. These covariance matrix estimates are
obtained as the outer product estimate, based on 100 samples
each, where the overlap between each estimate is 80 samples.
As can be seen, the spectral estimates are very noisy and
vary greatly in power. Using five of these covariance matrix
estimates R̂, evenly spaced throughout the signal, we solve
(19) in order to obtain an estimated covariance path, Rτ . The
resulting spectra, estimated using (28), are shown in Figure 7.
As can be seen, the path resulting from the proposed method
allows for a smooth tracking of the shift in spectral content.
D. K-means clustering
As a simple illustration of how to utilize the barycenter
formulation in Section IV-C, we consider the application of
unsupervised clustering of phonemes. Specifically, we consider
7 utterances; 3 utterances of the phoneme /ae/, 2 utterances of
/oy/, and 2 utterances of /n/. For each utterance, we estimate
a covariance matrix of dimension n = 10 and then run a
K-means algorithm that alternates between classifying each
estimated covariance matrix according to (22) and computing
new barycenters according to (20). We initiate the algorithm
by choosing initial barycenters as convex combinations of the
available covariance matrix estimates, and demand separation
into three clusters. The algorithm is then run until convergence,
i.e., until the classification has stabilized. The results are
shown in Table I. As can be seen, the K-means algorithm
utilizing (20) and (22) is indeed able to correctly identify
three clusters corresponding to the three different phonemes.
Table II presents the corresponding distance matrix, i.e., the
matrix detailing the distance Tκ between each covariance
10
1
1
0.5
0.5
0
corr
corr
1
1
0.5
0.5
0
-100
-100
-50
-50
0
50
0
0
50
100
(degrees)
0
100
(degrees)
Fig. 8. Interpolated spatial spectrum estimated as a(θ)H Rτ a(θ), where Rτ
is obtained by solving (9), for the case of one moving source and one static
2
interferer, using with the cost function c(θ, ϕ) = eiθ − eiϕ .
Fig. 9. Interpolated spatial spectrum estimated as a(θ)H Rτ a(θ), where Rτ
is obtained by solving (9), for the case of one moving source and one static
interferer, using with the modified cost function in (29).
matrix estimate and each barycenter. As can be seen, the
clusters are quite well-separated.
and thus the proposition holds also for this choice of c(θ, ϕ).
The resulting estimated spectra obtained from the interpolant
Rτ resulting from solving (9) using this modified cost function
is shown in Figure 9. As can be seen, the source and the
interferer are now well separated throughout the interpolated
path.
E. Fixed cost for moving mass
In some scenarios, there might be a relatively large noise
component present in both R0 and R1 , where the noise power
is localized in frequency and this localization is the same
in both R0 and R1 . This might be the case in, e.g., DOA
estimation scenarios where a source of interest is moving in
the presence of a stationary interferer. Then, if the source for
example moves past the location of the interferer, the optimal
transport problem in (9) may result in a power association such
that the source and the interferer are mixed together. Such a
scenario is illustrated in Figure 8, showing estimated spectra
obtained from the interpolant Rτ , as given by (9) with the cost
2
function c(θ, ϕ) = eiθ − eiϕ . Here, the source is moving
from look-angle θ = 30◦ to θ = −20◦ , whereas there is a fixed
interferer located at θ = 0◦ having a third of the power of the
source. In order to avoid this type of problem and promoting
transport plans that avoid transporting stationary masses, the
cost function may be modified according to
(
1 + |eiθ − eiϕ |2 if θ 6= ϕ
c(θ, ϕ) =
(29)
0
if θ = ϕ
Thus, the cost function is here formulated such that there is a
fixed, baseline cost of moving any mass. It may be noted that
this cost function satisfies the assumptions of Proposition 2,
Utterance
True phoneme
Class
1
/ae/
1
2
/ae/
1
3
/ae/
1
4
/oy/
2
5
/oy/
2
6
/n/
3
7
/n/
3
VII. C ONCLUSIONS AND FURTHER DIRECTIONS
In this work, we have proposed a notion of distance for
positive semi-definite Toeplitz matrices. By considering spectral representations of such matrices, the proposed measure
is based on distances, in an optimal mass transport sense, between families of spectra consistent with the Toeplitz matrices.
We have shown that the proposed distance measure, under
some mild assumptions, is contractive with respect to additive
and multiplicative noise. The proposed measure may be used
to, for example, define inter- and extrapolation of Toeplitz
matrices, being of interest in applications such as tracking of
slowly varying signals.
A future direction of this methodology is to generalize
the distance measure for structured matrices corresponding
to input-to-state covariances in the THREE framework for
spectral analysis [12]. This generalizes ideas from beamspace
processing, enabling the user to improve resolution and robustness in power spectral estimation over selected frequency
bands [44], [45]. This will be the subject of further research.
Utterance
Class 1
Class 2
Class 3
1
0.001
0.168
0.447
2
0.017
0.143
0.426
3
0.061
0.098
0.389
4
0.121
0.028
0.319
5
0.258
0.123
0.205
6
0.384
0.264
0.037
7
0.510
0.383
0.087
TABLE I
C LUSTERING OF 7 UTTERANCES INTO THREE CLUSTERS USING A
K- MEANS ALGORITHM UTILIZING THE BARYCENTER FORMULATION IN
S ECTION IV-C. T HE THIRD ROW INDICATES THE IDENTIFIED CLASSES AS
GIVEN BY THE ALGORITHM .
TABLE II
D ISTANCE , AS MEASURED BY Tκ , BETWEEN EACH UTTERANCE AND EACH
BARYCENTER FOR THE THREE IDENTIFIED CLUSTERS .
11
Fourier transform, that (Φw ∗ δθ ) (ϕ) is a spectrum consistent
with the covariance matrix Rw Rθ , i.e.,
A PPENDIX
A. Proof of Proposition 2 for multiplicative noise
Here, we present the proof for Tκ being contractive with
respect to multiplicative noise.
Proof. Let the noise covariance matrix be Rw , implying
that the covariances of the contaminated processes are
R00 = R0 Rw and R01 = R1 Rw , respectively. Let Φw be
any spectrum consistent with Rw , i.e., Γ(Φw ) = Rw . Also,
let the zeroth covariance of the noise process be smaller than
or equal to 1, so that
Z
Φw (θ)dθ ≤ 1
hΛ0 , Γ (Φw ∗ δθ )i = hΛ0 , Rw
Rθ i
and therefore
Γ∗ (Λ0 ) ∗ Φ̆w (θ) = hΛ0 , Rw
Rθ i
= hΛ0
= hΛ̃0 , Rθ i
= hΛ̃0 , a(θ)a(θ)H i
= a(θ)H Λ̃0 a(θ)
= Γ∗ (Λ̃0 )(θ)
T
We have
Tκ (R00 , R01 )
=
=
=
max
hΛ0 , R0 Rw i+hΛ1 , R1 Rw i
max
hΛ0 Rw , R0 i+hΛ1 Rw , R1 i
(Λ0 ,Λ1 )∈Ωκ
c
(Λ0 ,Λ1 )∈Ωκ
c
max
(Λ0 ,Λ1 )∈Ωκ
c
Λ̃0 ,Λ̃1 ∈Mn
Λ̃0 = Λ0
subject to
where the third equality follows from the definition of Λ̃0
and the last from the definition of Γ∗ (·). Using the same
κ
reasoning
for Λ1 , we thus have that (Λ0 , Λ1 ) ∈ Ωc implies
κ
Λ̃0 , Λ̃1 ∈ Ωc . Combined, this yields
Tκ (R00 , R01 ) ≤
hΛ̃0 , R0 i + hΛ̃1 , R1 i
Rw , Rθ i
max hΛ̃0 , R0 i + hΛ̃1 , R1 i = Tκ (R0 , R1 )
(Λ̃0 ,Λ̃1 )∈Ωκc
Rw
Λ̃1 = Λ1
Rw
In order to show that (Λ0 , Λ1 ) ∈ Ωκc implies Λ̃0 , Λ̃1 ∈ Ωκc ,
we note the following. Assume that (Λ0 , Λ1 ) ∈ Ωκc . Then,
Γ∗ (Λ0 )(θ − φ)+Γ∗ (Λ1 )(ϕ − φ) ≤ c(θ − φ, ϕ − φ) = c(θ, ϕ)
(30)
for all φ, θ, ϕ ∈ T, where the last equality follows from the
assumption that c is shift-invariant. Define
Φ̆w (φ) = Φw (−φ)
R
which satisfies Φ̆w (φ)dφ ≤ 1 and Φ̆w (φ) ≥ 0, ∀φ ∈ T.
Multiplying both sides of (30) with Φ̆w (φ) and integrating
with respect to φ then yields
Γ∗ (Λ0 ) ∗ Φ̆w (θ) + Γ∗ (Λ1 ) ∗ Φ̆w (ϕ) ≤ c(θ, ϕ)
for all θ, ϕ ∈ T. Similarly,
Γ∗ (Λ0 ) ∗ Φ̆w (θ) ≤ κ, Γ∗ (Λ1 ) ∗ Φ̆w (ϕ) ≤ κ
(31)
for all θ, ϕ ∈ T. Further, for any θ ∈ T, we have
Z
∗
Γ (Λ0 ) ∗ Φ̆w (θ) = a(ϕ)H Λ0 a(ϕ)Φ̆w (θ − ϕ)dϕ
T
Z
= trace Λ0 a(ϕ)a(ϕ)H Φ̆w (θ − ϕ)dϕ
T
Z
= hΛ0 , a(ϕ)a(ϕ)H Φw (ϕ − θ)dϕi
ZT
= hΛ0 , a(ϕ)a(ϕ)H (Φw ∗ δθ ) (ϕ)dϕi
T
= hΛ0 , Γ (Φw ∗ δθ )i
where the last equality uses the definition of the operator Γ(·).
As δθ (ϕ) is a spectrum consistent with the rank-one covariance
matrix Rθ = a(θ)a(θ)H , we have, by the properties of the
R EFERENCES
[1] J. P. Burg, “Maximum Entropy Spectral Analysis,” in Proceedings of the
37th Annual International SEG Meeting, Oklahoma City, OK., 1967.
[2] P. Stoica and R. Moses, Spectral Analysis of Signals, Prentice Hall,
Upper Saddle River, N.J., 2005.
[3] A. P. Dempster, “Covariance Selection,” Biometrics, vol. 28, no. 1, pp.
157–175, 1972.
[4] V. Chandrasekaran, P. A. Parrilo, and A. S. Willsky, “Latent variable
graphical model selection via convex optimization,” The Annals of
Statistics, vol. 40, no. 4, pp. 1935–1967, 2012.
[5] A. K. Kshirsagar, Multivariate Analysis, Dekker, 1972.
[6] A. M. Zoubir, V. Koivunen, Y. Chakhchoukh, and M. Muma, “Robust
Estimation in Signal Processing: A Tutorial-Style Treatment of Fundamental Concepts,” IEEE Signal Process. Mag., vol. 29, no. 4, pp. 61–80,
July 2012.
[7] E. Ollila, D. E. Tyler, V. Koivunen, and H. V. Poor, “Complex Elliptically
Symmetric Distributions: Survey, New Results and Applications,” IEEE
Trans. Signal Process., vol. 60, no. 11, pp. 5597–5625, Nov 2012.
[8] E. Ollila and D. E. Tyler, “Regularized m -Estimators of Scatter Matrix,”
IEEE Trans. Signal Process., vol. 62, no. 22, pp. 6059–6070, Nov 2014.
[9] A. Wiesel, “Geodesic Convexity and Covariance Estimation,” IEEE
Trans. Signal Process., vol. 60, no. 23, pp. 6182–6189, Dec 2012.
[10] Y. Sun, P. Babu, and D. P. Palomar, “Robust Estimation of Structured
Covariance Matrix for Heavy-Tailed Elliptical Distributions,” IEEE
Trans. Signal Process., vol. 64, no. 14, pp. 3576–3590, July 2016.
[11] J. P. Burg, D. G. Luenberger, and D. L. Wenger, “Estimation of
Structured Covariance Matrices,” Proc. IEEE, vol. 70, no. 9, pp. 963–
974, 1982.
[12] C. L. Byrnes, T. T. Georgiou, and A. Lindquist, “A new approach to
spectral estimation: a tunable high-resolution spectral estimator,” IEEE
Trans. Signal Process., vol. 48, no. 11, pp. 3189–3205, Nov 2000.
[13] H. Li, P. Stoica, and J. Li, “Computationally efficient maximum
likelihood estimation of structured covariance matrices,” IEEE Trans.
Signal, vol. 47, no. 5, pp. 1314–1323, May 1999.
[14] T. T. Georgiou, “Structured covariances and related approximation questions,” in Directions in Mathematical Systems Theory and Optimization.
2003, pp. 135–140, Springer.
[15] J. Karlsson and P. Enqvist, “Input-to-state covariances for spectral
analysis: The biased estimate,” in Proc. of Int. Symp. Mathematical
Theory of Networks and Systems, 2012.
[16] E. Avventi, A. G. Lindquist, and B. Wahlberg, “ARMA Identification
of Graphical Models,” IEEE Trans. Autom. Control, vol. 58, no. 5, pp.
1167–1178, May 2013.
[17] M. Zorzi and R. Sepulchre, “AR identification of latent-variable
graphical models,” IEEE Trans. Autom. Control, vol. 61, no. 9, pp.
2327–2340, Oct 2016.
12
[18] K. Yamamoto, Y. Chen, L. Ning, T. T. Georgiou, and A. Tannenbaum,
“Regularization and Interpolation of Positive Matrices,” arXiv preprint
arXiv:1611.07945, 2016.
[19] F. Barbaresco, “Information geometry of covariance matrix: CartanSiegel homogeneous bounded domains, Mostow/Berger fibration and
Frechet media,” in Matrix Information Geometry. 2013, pp. 199–255,
Springer.
[20] M. Knott and C. S. Smith, “On the optimal mapping of distributions,”
Journal of Optimization Theory and Applications, vol. 43, no. 1, pp.
39–49, 1984.
[21] L. Ning, X. Jiang, and T. T. Georgiou, “On the Geometry of Covariance
Matrices,” IEEE Signal Process. L, vol. 20, no. 8, pp. 787–790, August
2013.
[22] C. Villani, Optimal transport: old and new, Springer Science & Business
Media, 2008.
[23] S. Kolouri, S. R. Park, M. Thorpe, D. Slepcev, and G. K. Rohde,
“Optimal Mass Transport: Signal processing and machine-learning applications,” IEEE Signal Process. Mag., vol. 34, no. 4, pp. 43–59, July
2017.
[24] S. Haker, L. Zhu, A. Tannenbaum, and S. Angenent, “Optimal Mass
Transport for Registration and Warping,” International Journal of
Computer Vision, vol. 60, no. 3, pp. 225–240, 2004.
[25] F. De Goes, D. Cohen-Steiner, P. Alliez, and M. Desbrun, “An Optimal
Transport Approach to Robust Reconstruction and Simplification of 2D
Shapes,” Computer Graphics Forum, vol. 30, no. 5, pp. 1593–1602,
2011.
[26] B. Engquist and B. D. Froese, “Application of the Wasserstein metric
to seismic signals,” Communications in Mathematical Sciences, vol. 12,
no. 5, 2014.
[27] T. T. Georgiou, J. Karlsson, and M. S. Takyar, “Metrics for power
spectra: an axiomatic approach,” IEEE Trans. Signal, vol. 57, no. 3, pp.
859–867, March 2009.
[28] X. Jiang, Z. Q. Luo, and T. T. Georgiou, “Geometric Methods for
Spectral Analysis,” IEEE Trans. Signal Process., vol. 60, no. 3, pp.
1064–1074, March 2012.
[29] Y. Rubner, C. Tomasi, and L. Guibas, “The Earth Mover’s Distance as a
Metric for Image Retrieval,” International Journal of Computer Vision,
vol. 40, no. 2, pp. 99–121, 2000.
[30] H. Ling and K. Okada, “An Efficient Earth Mover’s Distance Algorithm
for Robust Histogram Comparison,” IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 29, no. 5, pp. 840–853, 2007.
[31] J. Rabin, G. Peyré, J. Delon, and M. Bernot, “Wasserstein Barycenter
and its Application to Texture Mixing,” in International Conference
on Scale Space and Variational Methods in Computer Vision, 2011, pp.
435–446.
[32] M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein GAN,” arXiv
preprint arXiv:1701.07875, 2017.
[33] J. Adler, A. Ringh, O. Öktem, and J. Karlsson, “Learning to solve inverse
problems using Wasserstein loss,” arXiv preprint arXiv:1710.10898,
2017.
[34] D. G. Luenberger, Optimization by Vector Space Methods, John Wiley
and Sons, New York, 1969.
[35] A. Friedman, Foundations of Modern Analysis, Courier Corporation,
1970.
[36] D. H. Johnson and D. E. Dudgeon, Array Signal Processing: Concepts
and Techniques, Prentice Hall, Englewood Cliffs, N.J., 1993.
[37] U. Grenander and G. Szegö, Toeplitz Forms and Their Applications,
University of California Press, Los Angeles, 1958.
[38] B. N. Bhaskar, G. Tang, and B. Recht, “Atomic Norm Denoising with
Applications to Line Spectral Estimation,” IEEE Trans. Signal Process.,
vol. 61, no. 23, pp. 5987 – 5999, July 2013.
[39] F. Elvander, S. I. Adalbjörnsson, J. Karlsson, and A. Jakobsson, “Using
Optimal Transport for Estimating Inharmonic Pitch Signals,” in 42nd
IEEE Int. Conf. on Acoustics, Speech, and Signal Processing, New
Orleans, LA, USA, March 5-9 2017, pp. 331–335.
[40] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge
University Press, Cambridge, UK, 2004.
[41] J. R. Deller, J. H. L. Hansen, and J. G. Proakis, Discrete-Time Processing
of Speech Signals, Wiley-IEEE Press, 1999.
[42] B. Dumitrescu, “Trigonometric Polynomials Positive on Frequency
Domains and Applications to 2-D FIR Filter Design,” IEEE Trans.
Signal Process., vol. 54, no. 11, pp. 4282–4292, Nov 2006.
[43] B. Dumitrescu, “Optimization of two-dimensional iir filters with
nonseparable and separable denominator,” IEEE Trans. Signal Process.,
vol. 53, no. 5, pp. 1768–1777, May 2005.
[44] A. N. Amini and T. T. Georgiou, “Tunable Line Spectral Estimators
Based on State-Covariance Subspace Analysis,” IEEE Trans. Signal
Process., vol. 54, no. 7, pp. 2662–2671, July 2006.
[45] J. Karlsson and T. T. Georgiou, “Uncertainty Bounds for Spectral
Estimation,” IEEE Trans. Autom. Control, vol. 58, no. 7, pp. 1659–
1673, July 2013.
| 10math.ST
|
ResumeVis: A Visual Analytics System to Discover
Semantic Information in Semi-structured Resume Data
Chen Zhang
Hao Wang
Yingcai Wu
Institute of Software, Chinese
Institute of Software, Chinese
State Key Lab of CAD & CG, Zhejiang
Academy of Sciences, Beijing, China Academy of Sciences, Beijing, China
University, Hangzhou, China
008613811513670
008615001098360
008613738047313
[email protected]
[email protected]
ABSTRACT
Massive public resume data emerging on the WWW indicates
individual-related characteristics in terms of profile and career
experiences. Resume Analysis (RA) provides opportunities for
many applications, such as talent seeking and evaluation. Existing
RA studies based on statistical analyzing have primarily focused on
talent recruitment by identifying explicit attributes. However, they
failed to discover the implicit semantic information, i.e., individual
career progress patterns and social-relations, which are vital to
comprehensive understanding of career development. Besides, how
to visualize them for better human cognition is also challenging. To
tackle these issues, we propose a visual analytics system ResumeVis
to mine and visualize resume data. Firstly, a text-mining based
approach is presented to extract semantic information. Then, a set
of visualizations are devised to represent the semantic information
in multiple perspectives. By interactive exploration on ResumeVis
performed by domain experts, the following tasks can be
accomplished: to trace individual career evolving trajectory; to
mine latent social-relations among individuals; and to hold the full
picture of massive resumes’ collective mobility. Case studies with
over 2500 online officer resumes demonstrate the effectiveness of
our system. We provide a demonstration video1.
CCS Concepts
• Information systems➝Web mining
computing➝Visual analytics.
• Human-centered
Keywords
Visual analytics; resume analysis; semantic information mining;
text visualization
1. INTRODUCTION
Massive publicly accessible resumes have emerged on the
Internet. They outline people’s attributes, experiences and
qualifications previously and currently [1]. Therefore, Resume
Analysis (RA) is an important means for personnel assessment. For
1
https://1drv.ms/v/s!AstTHdadq8_-gz6Tku2v10TwhGzi
[email protected]
example, with resumes one can assess people’s current
qualifications or predict their behaviors in the future by analyzing
past experiences [2]. In general, RA has many applications, such as
talent seeking, recruitment and job-performance evaluating in
company’s Human Resources department, and cadre assessment
and appointment in government’s think-tank department.
Besides, resumes can also reveal implicit people’s career
advancement process and latent social relations. For example, a
government-officer or scholar’s resume shows how his or her
title/rank advanced from lower level through higher level. The
implicit social-relations can be detected by identifying intersections
of career experiences, e.g., co-worker relationship. We believe that
such semantic information hidden in semi-structured resume text is
vital for in-depth RA tasks, such as semantic searching, talent
recommendation, career prediction and social network analysis.
However, existing RA studies neglected to discover such
information. Instead, they were statistical analysing-based and
basically focused on identifying explicit attribute information from
structured resumes to solve problems such as talent recruitment.
Besides, how to design the appropriate visualizations for enhancing
human cognition in order to perceive career patterns and mobility
from a huge amount of resumes is also a challenge.
To address these challenges, we firstly define three semantic
analytical tasks in multiple perspectives:
How to trace individual career evolving trajectory and uncover
common growth pattern? (Individual perspective)
How to mine latent social-relations among individuals?
(Interpersonal perspective)
How to hold the full picture of massive resumes for visualizing
collective mobility? (Collective perspective)
Then, we propose a visual analytics prototype called ResumeVis
to mine and visualize resume data. Specifically, a text-mining
based approach is presented to discover semantic information, such
as personal career paths and interpersonal social-relations, from
semi-structured resume text. Besides, three visualization modules
are devised to represent the mined semantic information:
career trajectory chart. A timeline-styled chart which
represents the career progress “trajectory” of an individual.
interpersonal relationship graph. An ego-network based graph
which reveals relations between individuals.
organization-individual mobility map. A map which employs
“quadrant diagram” metaphor to indicate people mobility.
Based on that, ResumeVis provides an integrated visual analytics
environment, where users can gain knowledge and insights through
visual interaction and in-depth exploration. Finally, the previously
defined three semantic analytical tasks can be accomplished.
To the best of our knowledge, visual analytics on semantic
information in resume data has not been researched previously.
Based on ResumeVis, the analysis of resume can be extended into
the analysis of people, which may trigger the application of
ResumeVis in other fields such as public opinion monitoring.
{Paragraph #1} Jim; male; ethnic Han; born on August 2nd, 1975; come from
Changsha city, Hunan province. Work on January, 1990;
Joined the Chinese Communist Party on December, 1991.
Currently is the governor of Hunan Province.
{Paragraph #2} Year 1989 ~ 1992: Appointed as the secretary of Party Branch
of Health Bureau of Ningxiang county, Hunan province.
…… //Paragraphs #3 to #6 are omitted.
{Paragraph #7} Year 2010-Up to now: Appointed as the governor of Hunan
Province.
Fig. 1. An example of officer resume.
The rest of this paper is organized as follows: Section 2 describes
the input data, design requirements and an overview of our system.
Section 3 proposes a semantic information extraction approach.
Section 4 presents the visual design of our system. Case studies are
performed in Section 5. Section 6 introduces related work. The
conclusion is summarized in Section 7.
2. SYSTEM OVERVIEW
2.1 Description of Input Data
Electronic resume can be classified into two forms: structured
(sr) and semi-structured (smr). sr is commonly table-based and
exists in database. The data format of sr is well-organized and
usually fixed, so sr can be easily stored and managed. However, it
is hard to acquire sufficient sr for high-level semantic analysis
because the available sr on the Internet is limited and most of sr as
internal resources are private. On the other hand, smr is a type of
text document and usually exists on the Internet. The data format
of smr is diverse and inconsistent because of text forms (txt, word,
pdf, etc.), so it is non-trivial to uniformly manage and analyze smr.
However, smr as publicly accessible resources can be obtained
easily, which provide great opportunities for semantic based
intelligence analysis. The examples of intelligence analysis are
semantic search, person classification, person profile, etc [8].
We use smr crawled on the Internet as input. smr can be
instantiated to government-officer’s resume, enterprise-stuff
resume, executive resume, scholar resume or pop-star resume.
Specifically, we use government-officer’s resume in our system.
However, the analysis approaches and visualizations are general; it
is easy to make an extension to address other types of smr.
A resume usually contains two sections: basic information (BIsection) and experience (E-section) [35]. BI-section covers a
person’s name, gender, age, date of birth, etc. E-section
summarizes a person’s education and/or career experiences.
Fig. 1 shows an example, which translates a Chinese resume.
Due to the difference between Chinese and English, we add
semicolons so that the pauses in a sentence can be easily recognized.
In Fig. 1, the first 3 sentences constitute BI-section. Sentence #2
gives time stamps of two milestones, and sentence #3 outlines the
current career title. The rest sentences, i.e., paragraphs #2 to #7,
constitute E-section. Each paragraph describes an experience
record containing time span, title, organization and location.
Another case is that all experience records are within one paragraph.
We omit the education experiences which have the similar structure
with career experiences, and we only focus on the analysis of career
experiences in this paper. However, the approach is general and can
be extended into the analysis of education experiences.
2.2 Design Requirements
Our system is under a three-year International Innovation Team
project founded by K. C. Wong Education Foundation. This project
aims at public opinion surveillance from online media. Five domain
experts are incorporated to supervise this project, including a vice
head of an institute who is in charge of this project, two government
administrators whose duties are cadre management and training, an
intelligence analyst, and a HR expert in a company. Regular
research discussion meetings and unscheduled idea exchange
meetings with these experts have held since the year 2014. During
these communications, the experts have clarified their requirements
and evaluated the prototypes developed for the project.
2.2.1 Requirements in mining phase
R1 Semantic elements extraction. Firstly, to extract careerrelated information mentioned in Section 2.1. Secondly, to quantify
the career rank of the resume’s title at each career stage. Taking
officer title as an example: President, vice President, Governor,
vice Governor, Mayor, vice Mayor, County Head, vice County
Head and civilian are quantified as rank 8 through rank 0.
R2 Pattern discovery. Firstly, to mine personal career progress
patterns. A domain expert claims that an effective feature for
pattern discovery may be the resume’s progress rate at each career
stage. Secondly, to discover interpersonal social patterns based on
the social relations among resumes. The social relations may be
explicit where the intersection between individuals’ career
experiences exists. The relations may also be implicit that
individuals’ career progress paths are similar. The interpersonal
social patterns can be further used to construct a social network,
upon which numerous network analysis tasks can be performed.
R3 Resume validation. Given an unknown resume with no
person name, incomplete, outdated or even wrong information, to
identify who this resume may belong to and validate the incorrect
information in this resume. Applications of this requirement could
be updating the outdated resumes, or revising the inaccurate
resumes produced intentionally or by accident.
2.2.2 Requirements in visualization phase
R4 Representing the resume from overview to details. When
dealing with numerous resumes, users need a multi-level
visualization so as to investigate career paths from overview to
details and perceive career progress patterns comprehensively.
R5 Facilitating visual comparisons between different
resumes. The visualization should facilitate the comparison
between the resumes with various progress patterns and social
patterns through metaphor representation, where the career path of
a resume and the social relations between resumes can be
intuitively portrayed (Figs. 5a & 5b).
R6 Visualizing collective mobility from massive resumes. The
visualization should summarize the collective mobility across
temporal dimension and organization dimension so that users can
quickly hold the full picture of numerous resumes and gain insights
easily. For example, the visualization should reveal the transfer
process of people between various communities, such as
government, grass roots, state-owned enterprises and non-profit
organizations (Right part of Fig. 2).
R7 Flexible data labeling to collect user feedback. The pattern
discovery model in R2 depends on predefined categories and
labeled data. In order to incorporate human knowledge and improve
model performance, the visualization and corresponding
interactions should be designed to support smoothly data labeling
and model tunning (Fig. 6).
R8 Easily browsing of raw data. The raw data such as text in
each resume is the strongest support for validating the semantic
elements and patterns discovered by the system. The visualization
should enable users to explore the raw data easily (Fig. 6).
2.3 System Workflow
Fig. 2 illustrates the system workflow. It has three phases:
The mining phase will be described in Section 3. It loads input
data, extracts semantic elements and mines pattern information.
The visualization phase and interaction phase will be described
in Section 4. The former contains three visualization modules
Visualization Phase
Mining Phase
Resume Text
*Ascending
*Steady
*Recessionary
Pre-processing
Organization-Individual Mobility Map
Career Trajectory Chart
2013.07
Animation
1960 1970 1980 1990 2000 2010 2020
Government
National Chief #8
National Deputy #7
Chief #6
Provincial
Deputy #5
Chief #4
Municipal
Deputy #3
Chief #2
Prefectural
Basic Info.
Non-profit Org.
Parsing
Resume base
(.xml)
County Head
Deputy #1
Experience Info.
Grass-roots #0
Rank Quantification
Career Trajectory
Interpersonal Relationship Graph
Time
Relations between A and B (Colleague):
During 2001.1.1~2004.3.1 at Party committee of
Hebei province
Semantic Mining
Grass-roots
Career Patterns
Mobility Info. (Appoint and Dismiss)
Feedback & Tuning
Social-relations
Perceiving &
Gaining Insights
I
D
Person A
J
G
E
K
F
2013.07; Person A appointed as the vice mayor of Changsha city,
Hunan province;
2013.07; Person B appointed as the secretary of Party Branch of Health
Bureau of Changsha city, Hunan province;
2012.12; Person C appointed as the secretary of Party Committee of
Beijing city;
2012.12; Person C dismissed as the mayor of Beijing city;
Interaction Phase
Fig. 2. Overview of our system workflow.
supporting visual exploration and analytics of mined semantic
information. The latter forms a loop between mining and
visualization, where users are involved for perceiving, gaining
insights and generating feedback, such as tuning parameters.
3. SEMANTIC EXTRACTION
In left part of Fig. 2, the approach contains four components: preprocessing, parsing, rank quantification and semantic mining.
3.1 Parsing
We propose a method to parse semi-structured resume text using
Chinese Word Segmentation and NER techniques (R1). Fig. 3
shows an example of extraction result based on the resume in Fig.
1. In Fig. 3, BI-section and E-section are extracted and saved into
structured resume base as .xml format (Fig. 3).
BI-section includes some tags such as /Name, /Gender and
/Nation. We use an information extraction technique [14] to extract
these tags.
E-section contains a list of experience records in a chronological
order. Each record consists of some tags: /DateBegin, /DateEnd,
/Location, /Org, and /Title. In the case that all experience records
are within one paragraph of the original resume, we use /DateBegin
and /DateEnd as separators to divide one paragraph into several
paragraphs. We use the similar technique [14] to extract tags
/DateBegin, /DateEnd, /Location, /Org and /Title. Specifically, we
choose Chinese characters “year”, “month” and “day” as the
keywords to extract the time range of each experience record. We
choose Chinese characters “province”, “city”, “county”, “town”, etc.
to extract the location of each experience record. We choose
Chinese characters “division”, “bureau”, “department”, “institute”,
“academy”, etc. to extract the organizations of each experience
record. We use Chinese characters “head”, “secretary”, “committee
member”, etc. to extract the titles of each experience record. These
keywords are collected from a large number of Chinese government
websites.
In Fig. 3, we use the extracted semantic elements to form a
resume base. The resume base is designed as .xml format with the
tree structure of the experience section. On base of structured
resume base, the subsequent data mining and visualization
techniques can be easily performed. Due to space limitation we
only show one experience record in Fig. 3 while the rest are omitted.
Finally, the semi-structured resume text has been converted into
the structured resume base.
3.2 Rank Quantification
We quantify the career rank of each record, (<rank> in Fig. 3).
We take administrative levels mentioned in Chinese Civil-Servants
Law [36, 37] as the reference in which quantification rules are
extracted. For example, the mayors of the ordinary cities should be
quantified as rank 4 (R1). However, there are many exceptions. For
example, the mayors of Beijing, Shanghai, Tianjin and Chongqing
should be quantified as rank 6, because these four cities are
municipalities which directly under the central government and
have higher rank than ordinary cities.
Eventually, we extract a career trajectory sequence based on
these rules and exceptions. Table 1 shows an example of the career
trajectory sequence extracted from the resume base in Fig. 3.
Please note each section in the career trajectory sequence,
including rank value, can be modified by users in UI shown in Fig.
6. This is essential when the information is incorrect due to the
complexity of Chinese administration and numerous special cases.
3.3 Semantic Mining
After obtained the resume base with quantified career rank, we
perform various mining methods to discover semantic patterns (R2).
<?xml version="1.0" encoding="UTF-8"?>
<resume_base version="1.0">
<resume category="1"> //ID of this resume
<basic_info_section>
<name>Jim</name> <nation>Han</nation> <birth_place>Changsha city,
Hunan province</birth_place>
<age>39</age> <gender>Male</gender>
<date category="birth"> <year>1975</year> <month>8</month>
<day>2</day> </date>
<date category="party"> <year>1991</year> <month>12</month>
<day>1</day> </date>
<date category="work"> <year>1990</year> <month>1</month>
<day>1</day> </date>
</basic_info_section >
<experience_section>
<experience_record>
<date category="begin"> <year>1989</year> <month>1</month>
<day>1</day> </date>
<date category="end"> <year>1992</year> <month>1</month>
<day>1</day> </date>
<location> <province>Hunan</province> <city>Changsha</city> </location>
<organization_array>
<organization>
<organization_name>Party Branch of Health Bureau</name >
<title_array>
<title> <title_name>Secretary</post_name> <rank>0</rank> </title>
</title_array>
</organization>
</organization_array>
</experience_record>
//The rest of experience records are omitted
</experience_section>
</resume>
</resume_base >
Fig. 3. The resume base built based on raw resume in Fig. 1.
Individual trajectory
Average trajectory
(a) ascending
(c) recessionary
Rank
(b) stead
Let x {r1 , r2 ,.., rn } as a resume’s feature set; C { y1 , y 2 , y3 } as
the pattern set, where y1 , y 2 and y 3 separately denote ascending,
steady, and recessionary.
Firstly, we calculate the conditional probabilities by using (2).
Time
Time
Time
P ( ri | y j ) g ( ri , y j ,
Fig. 4. Three career progress patterns.
Table 1. The career trajectory sequence
/DateBegin
/DateEnd
/Location
1989.1.1
1992.1.1
Changsha,
Hunan
1992.1.1
1995.1.1
idem
1995.1.1
1998.1.1
idem
1998.1.1
2002.1.1
idem
2002.1.1
2010.1.1
idem
2010.1.1
2015.12.11
idem
/Org
Party
Branch of
Health
Bureau
County
Party
Committee
Municipal
Party
Committee
Municipal
Party
Committee
Provincial
Party
Committee
Provincial
Party
Committee
/Rank
Secretary
0
County
head
2
Vice
mayor
3
Mayor
4
Vice
governor
5
Governor
6
P( y j | x)
e
2
(2)
2
yj
yj
j
P( x | y j ) P( y j )
P( x)
P( y j )
n
n
P(r | y ) P( y ) P(r | y )
P( x)
i
j
j
i 1
i
j
i 1
3.3.2 Mining interpersonal social-relations (R2)
We employ Naïve Bayes [18] to discover individual patterns.
The resumes are classified into various categories based on their
experiences information; each category is considered as a career
progress pattern. The unlabeled resume can be automatically
classified by the model trained on labeled resumes.
1. Definition of career progress patterns
Taking the mean of all resumes’ trajectories as baseline, the
domain experts defined three types of career progress patterns to
describe resumes with various career trajectories: ascending,
steady and recessionary. In Fig. 4, ascending represents the resume
that the rank growth rate (red line) is observably larger than the
average rate (black dot line) across the whole time dimension;
steady represents the resume that the rank growth rate is roughly
equal to the average rate across the whole time dimension;
recessionary represents the resume that the rank growth rate is
observably smaller than the average rate across the whole time
dimension. Specifically, the rank of recessionary can even decrease
in real cases. The trajectory lines in Fig. 4 depict the rank growth
rate along the time axis, which are calculated based on the time span
of each rank value in Table 1.
Please note the three patterns are specified by the domain experts
based on their knowledge and observation. There are other more
complicated patterns, such as the combination of the three patterns
in different career stages. Here we choose three basic patterns to
demonstrate our system’s functionality. Also, the labeling process,
in which each resume in the training set is assigned as one pattern,
is conducted by the domain experts.
2. Selection of career progress features
We take the time span on each rank value as the feature to describe
the career progress pattern. It can be formalized as (1).
(1)
r1 = t1 T , r2 = t2 T ,..., rn = tn T
where n 9 , which denotes the number of rank value. T
2
(3)
Thirdly, we classify x as yk , where k argmax j P( y j | x) .
3.3.1 Discovering individual career patterns
n
)
( ri y j ) 2
where y and y separately denote the mean and standard
deviation of ri on the pattern y j . Because ri is real number, we
assume that ri follows a Gaussian distribution.
Secondly, for each unlabeled resume x , we calculate the
posterior probabilities of each pattern y j by using (3).
j
/Title
yj
1
t ,
i 1 i
which denotes the time span across whole career experiences. For
example, the feature of the sequence in Table 1 is formalized as
{0.11, 0, 0.11, 0.11, 0.15, 0.3, 0.22, 0, 0}, where T 27 .
3. Training classifier using labeled resumes
1. Explicit relation mining
Let M {M 1 , M 2 ,.., M n } as the resume base containing n
resumes. Firstly, we construct a basket dataset where the
organization is considered as the “basket”. By scanning M , a list
of organizations O {O1 , O2 ,.., Om } is obtained. For each
organization Oi , we add a line into the basket dataset; the elements
in this line are the resumes which belong to, or belonged to Oi .
Secondly, we employ Apriori [24] to mine frequent resume set.
Thirdly, we propose a resume matching algorithm to measure the
matching degree between each resume pair in the frequent resume
set. The basic principle of the algorithm is that more experience
intersection produce higher matching degree; the matching degree
represents the relation value between the resume pair. The value
range of matching degree is from 0 (totally mismatching) through
1 (exactly matching, which means two resumes are the same).
Given an unknown resume Mu with no person name, incomplete
or even incorrect information, we can identify a resume M c from
the resume base M so that the matching degree between Mu and
M c reaches a peak value. It indicates that Mu most potentially
belongs to Mc . Besides, the possible incomplete or wrong
information of Mu can be validated based on the experience
intersection between Mu and M c (R3). Please refer to resume
validation view in Fig. 7 for details.
2. Implicit relation mining
We calculate the cosine similarity between resumes to measure
the implicit relations. The features used are the career progress
features mentioned in Section 3.3.1. Higher similarity indicates that
the career trajectories of resumes are more similar.
Given a resume M , top K resumes can be extracted from the
resume base M based on explicit relation and implicit relation
mining. These similar resumes will be visualized in interpersonal
relationship graph in Fig. 5(b).
4. VISULIZATION DESIGN
To satisfy the requirements in visualization phase (R4~R8), we
provide our system with three visualization modules.
4.1 Career Trajectory Chart
We present a timeline-based chart (Fig. 5a) to visually encode
the career progress process of an individual. This chart as a variant
of classical line chart converts the abstract individual career
evolving information into an intuitive representation. Specifically,
the temporal variation of individual’s rank, such as official title, and
the promotion/demotion timestamp are visualized. In addition, this
chart can assist users in understanding various career progress
patterns intuitively.
(d)
(a)
(b)
(c)
Fig. 5. Main window of ResumeVis. (a) career trajectory chart.
(b) interpersonal relationship graph. (c) statistical histogram. (d)
object manage panel.
some vertical lines. Mouse hovering over a rectangle reveals
detailed experience information via a floating tooltip window. To
enhance dynamic evolving effect when rendering the trajectory, the
animated transition is adopted so that the rectangles emerge one
after another in a chronological order. Besides, the animated
transition also occurs to preserve mental model when switching
between mode age and mode year. Legends on this chart shows the
names and the career progress patterns of corresponding resumes.
Eventually, a “ladder” representation is constructed. To avoid
overlapping among different trajectories, we add a tiny y-axis offset
on trajectories for easily distinguishing them.
In Fig. 5c, we employ statistical histogram to get an overview of
total resumes’ trajectories. Besides, we can compare the statistics
of trajectories between different resumes where the average value
of all resumes (denoted as gray bars) is depicted as a baseline.
In Fig. 6, we use “Resume Text” view to display the raw resume
as a baseline for easily browsing of raw data (R4).
Based on that, users can drag the icons of individuals on object
manage panel (Fig. 5d) into the chart (Fig. 5a) to perform visual
comparison, where different resumes are classified with various
colors (R5).
Users can click the icon on object manage panel to popup a
window to depict the basic information of the resume, as shown in
Fig. 6. This window contains three sections: Basic Info., Resume
Text and Career Experiences. In “Basic Info.”, users can review
and modify the basic information, such as uploading ID photo or
editing name. In “Resume Text”, users can review raw resume text.
In “Career Experiences” section, user can modify the semantic
elements of career experiences (R8). For example, rank values are
indicated as stars; users can adjust the number of stars by clicking
the stars when rank quantification is incorrect. After the rank values
are modified, career trajectory chart with new rank values is
updated smoothly.
Users can label the pattern type by clicking “Career pattern”
combo box at the bottom of Fig. 6. Users can also modify the
current pattern type predicted by the classifier in Section 3.3.2 (R7).
This label/annotation function not only facilitates expressing the
insight of users but supports the classification learning process.
4.2 Interpersonal Relationship Graph
Fig. 6. Basic information and setting window.
(a)
(b)
Fig. 7. Resume validation view. (a) The collapsed view; (b)
The expanded view.
In Fig. 5a, x-axis encodes the temporal dimension. There are two
modes, “age” and “year”, which are toggled alternately by pressing
the corresponding buttons. In addition, y-axis encodes the
dimension “rank”, i.e., quantified rank value of the career trajectory
sequence. The ordinal scale is from rank 0 to rank 8. In this paper,
these rank values are mapped into levels of the administrative titles
from Civilian (0) through President (8).
Then, the resume’s career trajectory can be visualized as a set of
color-filled horizontal rectangles, each of which denotes an
experience record. These rectangles are assembled head-to-tail by
We present an ego-network based spiral graph, in Fig. 5b, to
reveal interpersonal relations between a focus person, denoted as a
central node, and other related people, denoted as surrounding
nodes. As a complement of career trajectory chart, this graph can
enable users to understand the “context” information of the focus
person.
When users right-click the icon of a resume M , the graph is
constructed in which M is visualized as the central point and M k
are shown as the K surrounding nodes. The node is visualized as
a rounded rectangle with the name of the individual on it. Please
note M k are calculated in Section 3.3.2. The distance between the
central node and a surrounding node encodes the relation value
between the focus person and a related person. That is, shorted edge
means closed relations.
The parameter K and relation type can be set via tool buttons at
the bottom of the graph. The mouse hovering over a surrounding
node shows detailed relation information. When the explicit
relations are chosen, the relation value is shown as in Fig. 9a;
otherwise, the experience intersection is shown as in Fig. 9b. Users
can drag the surrounding node to adjust the position of the node
along the reference circle for optimized layout. When dragging a
node, an auxiliary red circle emerges as a baseline to facilitate the
comparison of relation value among other nodes (see Fig. 5b).
When users load an unknown resume M u with no person name,
incomplete, outdated or even wrong information, we propose
resume validation view shown in Fig. 7 to identify who M u may
belong to and validate the incorrect information in M u (R3). In
Fig. 7a, the most similar person of M u is calculated as Jim and the
similarity degree, i.e., matching degree in Section 3.3.2, between
M u and Jim is 86%, which means 86% of Jim’s career experience
are matched with M u . Users may click “Show Details” button to
investigate the detailed information in the expanded view. In Fig.
7b, other similar people as candidates are listed in a descending
order according to the similarity degree. After clicking an icon in
the list which denotes a candidate M s is selected, M u is shown
in “Test Resume” section and M s is shown in “Standard Resume”
section. The mismatch between two resumes’ text is highlighted as
red bold font.
4.3
Organization-Individual Mobility Map
We propose a map for gaining insights on collective mobility
through visualizing massive resumes. This map employs the
metaphor “quadrant diagram” (right part of Fig. 2) to represent the
collective mobility across temporal and organization dimension
(R6). Here organizations are classified as four categories by
considering the suggestions and comments of several domain
experts. They are government, grass roots, state-owned enterprises
(or companies) and non-profit organizations. Resumes belonging to
same type of organization form a community. Therefore, resumes
are divided as five communities: four basic communities and one
compound community where individuals belong to multiple types
of organizations.
In right part of Fig. 2, each node denotes an individual, of which
the size encodes the rank value. The map is divided as five regions
to encode five communities: four sectors denote four basic
communities and the circle region in the center denotes the
compound community. Four radials as the boundaries of four basic
communities encode the dimension “time”, which increases from
center to around. Therefore, the location of the node encodes the
type of organization to which the individual belongs at a specific
time. As time evolves, the size and the location of the node vary.
The transition process is visualized with the form of animation so
as to reveal the collective mobility. A force-directed layout method
[38] is applied to avoid node overlapping while keeping the relative
positions of the nodes as much as possible. The detailed mobility
information, with two forms of appointment and dismissing, is
shown in “Mobility Info.” window (bottom-right corner of Fig. 2).
Users can zoom-in, zoom-out and drag the map for exploration
using various resolution. Users can select a node to investigate the
detailed information regarding the individual and the
organization(s) he/she belongs to (Figs. 10b and 10c). Users can
also select a community to gain the information regarding the
individuals belong to the selected community (Fig. 10d). When
selecting a community, the nodes in this community are highlighted
and the other nodes are hyalinized so that users can focus on
interesting area. When users select a node from the compound
community, several auxiliary lines emerge to indicate which
communities this node belongs to (Fig. 10e).
In top-right corner of Fig. 2, users can press “Animation” button
to play the transition process over time for perceiving the collective
mobility during the whole time span. Besides, users can slide the
slider so as to discover the status at a specific time stamp.
5.
CASE STUDY
We evaluated the system through three case studies. In each case
study some real-world tasks were designed by three of our domain
experts. The other two experts were then required to use the system
for solving the tasks. Before the evaluation the experts spent two
hours using the system so that they were familiar with the system.
Feedbacks from the experts were received after the evaluation. The
dataset contains 2640 online resumes collected from websites of
Chinese governments at various levels.
5.1
Individual Career Progress Analysis
T1: Summarizing the general rule of career development.
T2: Comparing various career paths of different resumes.
T3: Searching target officer based on semantic information.
These tasks are vital for government to select and promote
potential officer. It is believed that selecting proper officer for
further promotion is non-trivial because many aspects of the
individual need to be considered, such as the career growth rate at
various stage. However, there is no quantitative standard for
assessing these aspects. Currently, this task is performed manually
based on human knowledge and experience. When facing massive
resumes, evaluating each resume merely by reading resume text is
rather tedious and time-consuming.
In Fig. 8, after experts imported the resumes into our system, the
system provided statistical histogram to indicate the statistical
pattern of career progress. Fig. 8a shows the statistical results of
total resumes, i.e., the average value of all resumes on each rank’s
time span. Figs. 8b, 8c and 8d separately denote the statistical
results of the individuals named “Jack”, “Tom” and “Karl”, whose
resumes were automatically classified as ascending, steady and
recessionary, respectively. From this view experts can gain the
following insights (T1):
1. There are two “bottleneck” period for Chinese officers during
their early career. The first is rank#0. They usually spend 13 years
at grass-roots organizations before they get the first promotion to
rank#1, such as vice county head. The second is rank#5. They
usually spend 10 years at rank#5, such as vice governor, before they
get the promotion to rank#6, such as governor.
(a)
(c)
(b)
(d)
(e)
Fig. 8. The statistical histogram views (a, b, c and d) and career
trajectory chart (e) of individuals’ career trajectories.
cluster X
cluster Y
(a)
(b)
Intersection
experience
(c)
(d)
Fig. 9. Interpersonal relationship graphs (a: implicit relation;
b: explicit relation) and corresponding career trajectory charts
(c: time mode “age”; d: time mode “year”).
2. Officers with career progress pattern ascending, such as Jack,
get faster promotion at the two “bottleneck” period than the average.
Officers with pattern steady, such as Tom, have approximate
promotion speed with the average at their early career. Officers
with pattern recessionary, such as Karl, have slower promotion
speed than the average at their early career.
Fig. 8e shows career trajectory chart with time mode “age”
visualizing the career progress trajectories of Jack, Tom and Karl.
From this chart experts easily perceived visual insights mentioned
above by comparing career trajectories with various patterns from
individual perspective (T2).
To perform T3, experts quickly skimmed the resumes with
career progress pattern ascending (about 200 resumes) by
visualizing them in career trajectory chart. Among these resumes
experts then filtered 39 resumes as candidates of potential officers
which have “sharp” trajectories. At last, experts selected 10
resumes as potential officers by investigating and examining the
detailed semantic information of each resume shown in Fig. 6. The
whole process took nearly 2 hours. In contrast, without our system
experts may spend several days to read over 2600 resumes for
selecting target resumes.
5.2
Interpersonal Social-relations Analysis
T4: Finding the similar or relevant resumes (individuals) based
on the focus resume (individual).
T5: Validating unknown/ unfaithful resumes.
Please note T4 is important for government to discover the
interpersonal relationship and even social network in Chinese
government. This task may facilitate anti-corruption.
When experts investigated interpersonal relationship graph of
the focus individual named “Tony” shown in Figs. 9a and 9b, they
may perceive the following phenomena (T4).
1. Through interaction with UI, e.g., switching the Relation Type
and changing the value of K, experts observed that surrounding
points can be classified as two clusters: clusters X and Y. By
investigating the detailed information regarding the relations
between the central point and clusters, experts realized that the
duration of career intersection between the central point and X is
significantly larger than that of Y. Besides, the intersection type, i.e.,
the organization type, of X is “Company” while the intersection
type of Y is “Non-profit Org”.
Fig. 10. (a) organization-individual mobility map. A selected
node (b) and its information (c). (d) The information of the
selected community (Government). (e) A selected node
belonging to multiple communities. (f) and (g) are maps created
by Excel based on the statistics of resumes.
2. The individual Lifeng has close relations with Tony.
Specifically, they have similar career progress trajectories in Figs.
9c and 9d. Another evidence in Fig. 9a shows the implicit relations,
i.e., the similarity, between Lifeng and Tony reaches a peak value.
In addition, the explicit relations in Fig. 9b shows that they were
colleagues during year 2001. Based on that, experts can infer that
the promotion of Lifeng may benefit from the intersection
experiences with Tony (blue mark in Fig. 12(d)). Experts can even
predict Lifeng will get further promotion soon based on such
trajectories similarity and relations.
Experts can interact with the UI in Fig. 7 for performing T5. We
collected 50 resumes from the officer base of Xinhua Net [3]. Two
participants (not experts) were requested to delete 10%~30% of text
content (including “name” section) from first 25 resumes for
producing “unknown resumes”, and modified 15% of text content
from second 25 resumes for producing “unfaithful resumes”. Such
operations were performed by using our system as shown in Fig. 6.
Based on these produced resumes, experts successfully identified
21 unknown resumes and 19 unfaithful resumes, as well as their
missing or wrong information being corrected.
5.3
Collective Mobility Analysis
T6: Discovering career evolving patterns across temporal
dimension.
T7: Discovering career transfer patterns across organization
dimension.
Solving T6 and T7 enables decision makers to be aware of career
situation of officers in a macroscopic perspective and thus make
reasonable strategies accordingly, such as guiding and controlling
the process of officer training, appointing and dismissing. Through
animation and interaction with organization-individual mobility
map in Fig. 10a, experts can gain two valuable insights.
First, there is a “booming” of officer emergence during
2009/11/01 and 2010/01/01 (The dense area of government section
in Fig 10a). Fig. 10f proves this “booming” phenomenon, which
shows the emergence trend during the whole time range. From this
phenomenon domain experts inferred that the reason behind may
be the fourth plenary session of the 17th central committee of the
communist party of China held in Sep. 2009 (T6).
Second, the number of resumes in government significantly
outweigh the number of resumes in other organizations.
Specifically, the number of grass-roots’ resumes is the least. Fig.
10g can prove that, which shows the proportion among various
organizations. This phenomenon demonstrates the fact that
currently being a civil servant in the government is the best career
choice for Chinese job seekers (T7).
Besides, experts can conduct in-depth exploration by ample
interactive tools. For example, when experts select a node (Fig.
10b), the information regarding this node is shown in Fig. 10c.
Experts can also select a community so that the nodes in this
community are highlighted and the information of this community
is shown in Fig. 10d. Fig. 10e indicates that experts can check the
nodes belonging to multiple communities.
5.4
User Feedback
We collected the feedback from the two experts. In Table 2, the
experts were required to rank a score on functionality and usability
of our system for each task. The scores could be 5 (excellent),
4(good), 3 (average), 2 (not good) or 1 (poor).
Then, the experts gave their comments about user experience.
Both experts were very impressed by the system, particularly by the
revealed semantic information which can well support their
decision making. They suggested that representing resumes in
multiple perspectives is “comprehensive”, and “valuable for human
cognition and perceiving insights”. In particular, the first expert
highlighted the novelty of career trajectory chart for outlining
career path. He said “Never have I seen such a visualization that
can intuitively represent the career progress trajectory from the
resume data”. He was very impressed by investigating the resume
data from overview to details. He believed that the metaphor
employed in the system can definitely improve the efficiency of RA.
He also mentioned that “the idea of representing social-relations
between different resumes as a spiral graph is excellent for
understanding social patterns among individuals”. The second
expert particularly liked organization-individual mobility map
because “the semantic information regarding collective behaviours
of resumes can be easily perceived in such a compact view”. In
addition, he believed that resume validation is valuable and
practical for real-world RA task; indeed, “there is no similar
function existed in current RA tools”.
Lastly, the experts pointed out that our system still has the
following limitations which should be improved in the future.
1. The information on spatial dimension has not considered yet.
Intuitively, the spatial career trajectory and spatial mobility on
geographical map are also important.
2. The ego-spiral graph should be extended into a social network
for better visual comprehension, because the relations among the
surrounding nodes are also important.
3. Case study qualitatively evaluates the functionality of
visualization phase of our system. However, the mining phase
should be further validated via some quantitative experiments.
6. RELATED WORK
6.1 Resume Analysis
The prototype of resume analysis is personal experience analysis
born during the World War II [4]. In 1996, resumes of scientific
researchers were used to evaluate researchers’ research ability and
to discover technology transfer process [5]. Gaughan and Robin [6]
compared career trends of scientists between different countries
using resumes of 800 physicists and bio-scientists from France and
the USA. Sabatier et al. [7] used resumes of 583 French bioscientists to research what kind of impact career mobility had on
various career progress patterns.
DaXtra [8] is a CV analysis software suite for recruitment
management. DaXtra contains resume parsing, semantic searching
and intelligent matching components which facilitate recruitment
processes. Visualize.me [9] employs the infographic representation
to turn a structured resume into several informative charts.
6.2 NLP and Data Mining
Much research has been conducted on NLP-related sub-tasks:
Tokenization, Part-of-speech (POS) Tagging, Named Entity
Recognition (NER), Sentiment Analysis, Word Segmentation etc.
[10, 11, 12, 13, 14]. There are plenty of NLP toolkits, such as
Stanford CoreNLP [10], UIMA [15], and GATE [16].
In data mining field, Classification and Association Rules
Mining are related to our system since our system classifies people
into various categories based on their experiences, and mines
relations between people. Popular classification techniques include
Decision Trees [17], Naïve Bayes [18] and Support Vector
Machine [19]. Many algorithms for generating association rules
were presented, such as Apriori [20], Eclat [21] and FP-growth [22]
6.3 Text Visualization
Multiple visualizations have been developed to represent
information in text data [23, 24, 25, 26, 27, 28].
Concrete text visualization. Word cloud has been used to
provide the content overview of corpus [29]. Cui et al. [23]
proposed a hybrid tag cloud visualization which incorporated a
trend chart to illustrate temporal content evolutions in a set of
documents. Wattenberg and Viégas [30] proposed the Word Tree
to represent information-retrieval results in a tree form in which
each node denoted a keyword. Ham et al. [31] presented the Phrase
Net to display a graph whose nodes are phrases and whose edges
indicate that two phrases are linked by a specific relation.
Abstract text visualization. Havre et al. [32] proposed
ThemeRiver to visualize thematic variations over time using a
river-like metaphor. Luo et al. [33] used a similar metaphor to
presents events based on event-based text analysis. Cui et al. [34]
presented TextFlow for analyzing topical evolution patterns in
which critical events were detected as flags of topic
birth/disappearance and topic merging/splitting. Wu et al. [24]
proposed OpinionFlow to visualize the diffusion of public opinions
on social media by combining a Sankey graph with a density map.
7.
CONCLUSION
This paper presents a visual analytics prototype based on resume
data, aiming to help users understand the semantic patterns hidden
in the resumes. First, a text-mining based approach is presented to
extract the semantic information. Then, three visualization modules
are designed to represent the semantic information in multiple
perspectives. Through in-depth interaction and exploration
performed by domain experts, our system are proved to be effective.
Further work can focus on the extension of the visualization on
geographical dimension by considering the temporal-spatial
mobility information. We also need to design better visual
encodings for representing social relations and coordinated
interaction between various views. Lastly, we consider to make a
formal evaluation of our system, especially on mining performance.
8. ACKNOWLEDGMENTS
This work is supported by National Basic Research Program of
China (2013CB329305), Natural Science Foundation of China
(61303164, 61402447, 61502466) and K. C. Wong Education
Foundation.
Table 2. Scoring results for each task
Task#
Score (expert#1 / #2)
1
4/3
2
4/5
3
4/4
4
5/5
5
5/4
6
5/5
7
5/4
9. REFERENCES
[1] Curriculum vitae.
https://en.wikipedia.org/wiki/Curriculum_vitae.
[2] Stokes, G. S. Introduction to special issue: the next one
hundred years of biodata. Human Resource Management
Review, 9(2): 111-116, 1999.
[3] Officer base of Xinhua Net. http://www.news.cn/rwk/.
[4] J. P. Guilford. https://en.wikipedia.org/wiki/J._P._Guilford.
[5] Research Value Mapping Program.
http://archive.cspo.org/rvm/.
[6] Gaughan, M., Robin, S. National science training policy and
early scientific careers in France and the United States.
Research Policy, (33): 569-581, 2004.
[7] Sabatier, M., Carrere, M., Mangematin, V. Profiles of
Academic Activities and Careers: DoesGender matter? An
analysis based on French life scientists’ CVs. Journal of
Technology Transfer, (31):311-324, 2006.
[8] DaXtra. https://www.daxtra.com/.
[9] Visualize.me. http://vizualize.me/.
[10] Manning, C. D., Surdeanu, M., Bauer, J., Finkel, J. R.,
Bethard, S., McClosky, D. The Stanford CoreNLP Natural
Language Processing Toolkit. In Proceedings of 52nd
Annual Meeting of the Association for Computational
Linguistics: System Demonstrations, 55-60, 2014.
[11] Tseng, H., Chang, P., Andrew, G., Jurafsky, D., Manning, C.
A Conditional Random Field Word Segmenter. In Fourth
SIGHAN Workshop on Chinese Language Processing, 2005.
[12] Toutanova, K., Klein, D., Manning, C. D., Singer, Y.
Feature-Rich Part-of-Speech Tagging with a Cyclic
Dependency Network. In Proceedings of HLT-NAACL. 173180, 2003.
[13] Finkel, J. R., Grenager, T., Manning, C. Incorporating Nonlocal Information into Information Extraction Systems by
Gibbs Sampling. In Proceedings of the 43nd Annual Meeting
of the Association for Computational Linguistics, 363-370,
2005.
[14] Soderland, S. Learning information extraction rules for
semistructured and free text, Machine Learning, 34(1-3):
233–272, 1999.
[15] Ferrucci, D., Lally, A. UIMA: an architectural approach to
unstructured information processing in the corporate research
environment. Natural Language Engineering, 10(3-4): 327–
348, 2004.
[16] Cunningham, H., Maynard, D., Bontcheva, K., Tablan, V.
GATE: an architecture for development of robust HLT
applications. In Proceedings of ACL, 168-175, 2002.
[17] Safavian S R, Landgrebe D. A survey of decision tree
classifier methodology. IEEE Transactions on Systems, Man,
and Cybernetics, 21(3): 660-674, 1991.
[18] Rish, I. An empirical study of the naive Bayes classifier. In
IJCAI 2001 workshop on empirical methods in artificial
intelligence. IBM New York, 3(22): 41-46, 2001.
[19] Suykens, J. A., Vandewalle, J. Least squares support vector
machine classifiers. Neural Processing Letters, 9(3): 293300, 1999.
[20] Rakesh, A., Ramakrishnan, S. Fast algorithms for mining
association rules in large databases, In Proceedings of the
20th International Conference on Very Large Data Bases,
487-499, 1994.
[21] Zaki, M. J. Scalable algorithms for association mining. IEEE
Transactions on Knowledge and Data Engineering, 12 (3):
372–390, 2000.
[22] Han, J., Pei, J., Yin, Y. Mining frequent patterns without
candidate generation. In ACM Sigmod Record, 29(2): 1-12,
2000.
[23] Cui, W., Wu, Y., Liu, S., Wei, F., Zhou, M. X., Qu, H.
Context preserving dynamic word cloud visualization. In
Pacific Visualization Symposium, 121-128, 2010.
[24] Wu, Y., Liu, S., Yan, K., Liu, M., Wu, F. OpinionFlow:
Visual analysis of opinion diffusion on social media. IEEE
Transactions on Visualization and Computer Graphics,
20(12): 1763-1772, 2014.
[25] Liu, S., Wu, Y., Wei, E., Liu, M., Liu, Y. Storyflow:
Tracking the evolution of stories. IEEE Transactions on
Visualization and Computer Graphics, 19(12): 2436-2445,
2013.
[26] Wu, Y., Wei, F., Liu, S., Au, N., Cui, W., Zhou, H., Qu, H.
OpinionSeer: interactive visualization of hotel customer
feedback. IEEE Transactions on Visualization and Computer
Graphics, 16(6): 1109-1118, 2010.
[27] Sun, G., Wu, Y., Liu, S., Peng, T. Q., Zhu, J. J., Liang, R.
EvoRiver: Visual analysis of topic coopetition on social
media. IEEE Transactions on Visualization and Computer
Graphics, 20(12): 1753-1762, 2014.
[28] Xu, P., Wu, Y., Wei, E., Peng, T. Q., Liu, S., Zhu, J. J., Qu,
H. Visual analysis of topic competition on social media.
IEEE Transactions on Visualization and Computer Graphics,
19(12): 2012-2021, 2013.
[29] Seifert, C., Kump, B., Kienreich, W., Granitzer, G.,
Granitzer, M. On the beauty and usability of tag clouds. In
International Conference Information Visualisation, 17–25,
2008.
[30] Wattenberg, M., Viégas, F. B. The word tree, an interactive
visual concordance. IEEE Transactions on Visualization and
Computer Graphics, 14(6): 1221-1228, 2008.
[31] Van Ham, F., Wattenberg, M., Viégas, F. B. Mapping text
with phrase nets. IEEE Transactions on Visualization and
Computer Graphics, 15(6): 1169-1176, 2009.
[32] Havre, S., Hetzler, E., Whitney, P., Nowell, L. Themeriver:
Visualizing thematic changes in large document collections.
IEEE Transactions on Visualization and Computer Graphics,
8(1): 9-20, 2002.
[33] Luo, D., Yang, J., Krstajic, M., Ribarsky, W., Keim, D.
Eventriver: Visually exploring text collections with temporal
references. IEEE Transactions on Visualization and
Computer Graphics, 18(1):93-105, 2012.
[34] Cui, W., Liu, S., Tan, L., Shi, C., Song, Y., Gao, Z. J., Tong,
X., Qu, H. Textflow: Towards better understanding of
evolving topics in text. IEEE Transactions on Visualization
and Computer Graphics, 17(12): 2412-2421, 2011.
[35] Clifton, T. C., Mumford, M. D., Baughman, W. A.
Background data and autobiographical memory: Effects of
item types and task characteristics. International Journal of
Selection and Assessment, 7(1): 57-71, 1999.
[36] Administrative Level in China.
http://baike.baidu.com/view/1672894. htm.
[37] Chinese Civil-Servants Law.
http://baike.baidu.com/view/2361565.htm.
[38] Fruchterman, T. M., Reingold, E. M. Graph drawing by
force-directed placement. Software: Practice and experience,
21(11): 1129-1164, 1991.
| 2cs.AI
|
Learning the LMP-Load Coupling From Data:
A Support Vector Machine Based Approach
Xinbo Geng∗1 and Le Xie†1
arXiv:1603.07276v1 [cs.SY] 23 Mar 2016
1
Department of Electrical and Computer Engineering, Texas A&M University
March 24, 2016
Abstract
This paper investigates the fundamental coupling between loads and locational marginal
prices (LMPs) in security-constrained economic dispatch (SCED). Theoretical analysis based
on multi-parametric programming theory points out the unique one-to-one mapping between
load and LMP vectors. Such one-to-one mapping is depicted by the concept of system pattern
region (SPR) and identifying SPRs is the key to understanding the LMP-load coupling. Built
upon the characteristics of SPRs, the SPR identification problem is modeled as a classification
problem from a market participant’s viewpoint, and a Support Vector Machine based datadriven approach is proposed. It is shown that even without the knowledge of system topology
and parameters, the SPRs can be estimated by learning from historical load and price data.
Visualization and illustration of the proposed data-driven approach are performed on a 3-bus
system as well as the IEEE 118-bus system.
1
Introduction
A fundamental issue with electricity market operation is to understand the impact of operating
conditions (e.g. load levels at each bus) on the locational marginal prices (LMPs). This paper
examines this key issue of the relationship between nodal load levels and LMPs. This issue is
further compounded by the increasing levels of demand response and variable resources in the grid.
In the power systems literature, reference [1] is among the pioneering works that uses perturbation techniques to compute the sensitivities of the dual variables in SCED (e.g. LMPs) with respect
to parameters (e.g. the nodal load levels). This sensitivity calculation method is widely used in
subsequent researches. However, this approach is valid only for small changes and the marginal
generator stays the same. Reference [2] observed the “step changes” of LMPs with respect to increasing system load level and discovered that new binding constraints (transmission or generation)
are the reason of the “step changes”. This is followed by further analysis on identifying the critical
load levels (CLLs) that trigger such step changes of LMPs [3], [4], [5]. This line of work assumes
that the system load change is distributed to each bus proportional to the base case load, which,
in many instances, do not necessarily represent the real-world situations. Reference [6] analyzed
this problem using quadratic-linear programming (QLP) and the concepts of system patterns and
system pattern regions (SPRs) were first introduced. The SPRs depict the relationship between
∗
†
[email protected]
[email protected]
1
loads and LMPs in the whole load space, which is not confined in a small neighborhood of an
operating point or constrained by a specific load distribution pattern. This paper is inspired by [6]
but focuses on the case of piecewise linear generation costs, instead of the quadratic cost case in [6].
The reason that we study the piecewise linear cost case is that piecewise linear cost curves are often
quite representative of the market practice in the real world. In addition, some new theoretical
results based on piecewise linear cost curves are derived, and are generalizable towards quadratic
cost cases.
Characterizing the SPRs would provide important insights to both system operators and market
participants. Reference [7] advances the theory of SPR from system operator’s perspective where
the knowledge of system topology and parameters is available. For market participants, such
knowledge is not necessarily available. Our previous work [8] examines the issue from market
participant’s viewpoint and applies the geometric features of SPRs to identify them.
This paper significantly advances our previous work by (1) completing the theoretical characterization of SPRs as a function of nodal load levels; (2) proposing a computational algorithm
to identify SPRs using historical data; (3) introducing the posterior probabilities of SPRs with
the presence of uncertain system parameters such as transmission limits; and (4) extending the
algorithm to consider practical factors such as partial load information and loss component of
LMPs.
The rest of the paper is organized as follows. Section 2 provides the analysis of LMP-load
coupling in SCED problem from the viewpoint of MLP theory, with an illustrative example. Section
3 illustrates the changes of SPRs given changes of system parameters such as transmission limits.
Based on the theoretical analysis, a data-driven algorithm for market participants to identify SPRs
is described in Section 4. Section 5 illustrates the performance of the algorithm on the IEEE 118bus system. Section explores the impact of nodal load information, and Section 7 provides critical
assessment of the proposed method. Concluding remarks and future works are presented in Section
8.
2
2.1
Theoretical Analysis
Notations
The notations of this paper are summarized below: mathematical symbols in hollowed-out shapes
(e.g. R) represent spaces and symbols in Calligra font (e.g. Sπ ) stand for sets. The superscript
“∗ ” indicates the variable is optimal, “ ˆ ” denotes estimated values (e.g. λ̂). Variables with “ ¯ ”
are expectations or average values (e.g. λ̄). “| ” denotes the transpose of a vector or matrix (e.g.
1|n ). The subscript “i ” represents the ith element of the vector (e.g. PGi ), and the superscript “(i) ”
(i)
represents the ith element in a set (e.g. PD ). The vector of n × 1 ones, matrix of m × n zeros and
the n × n identity matrix are denoted by 1n and 0m×n and In respectively.
2
2.2
Security Constrained Economic Dispatch
In real-time energy market operations, the LMPs are the results from the security-constrained
economic dispatch (SCED), which is formulated as follows:
min
(k)
PG
s.t.
nb
X
i=1
nb
X
(k)
ci (PGi )
(k)
PGi =
i=1
−F
(1a)
nb
X
(k)
PDj
: λ1
(1b)
: µ+ , µ−
(1c)
: η+, η−
(1d)
j=1
+
(k)
(k)
≤ H(PG − PD ) ≤ F +
(k)
PG− ≤ PG ≤ PG+
(k)
(k)
where PG is the generation vector at time k, and PD is the load vector at time k. We assume
there are both generation and load at each bus. Let nb denote the number of buses and nl denote
(k)
(k)
the number of transmission lines, then PG , PD ∈ Rnb . H ∈ Rnl ×nb is the shift factor matrix.
This formulation considers each snapshot independently, therefore it is called static SCED in
(k)
(k)
this paper. For simplicity, we write PG and PD as PG and PD when discussing the static SCED.
The objective of SCED is to minimize the total generation cost and satisfy the transmission and
generation capacity constraints while keeping the real-time balance between supply and demand.
(k)
The generation cost function ci (PGi ) of generator i is increasing and convex, and it is usually
regarded as a quadratic function or approximated by a piecewise linear function. To better reflect
the current practice in electricity markets, this paper studies the SCED problem with piecewise
linear
functions. And for the consideration of simplicity, the simplest form, i.e.
Pnb generator bidding
| P is being considered in this paper.
)
=
c
c
(P
G
i=1 i Gi
A fundamental concept in electricity markets is the Locational Marginal Price. The LMP λi at
bus i is defined as the change of total system cost if the demand at node i is increased by 1 unit [9].
According to [10], the LMP vector λ can be calculated by the following equation:
λ = λ1 1nb + H | (µ+ − µ− )
(2)
For better understanding, we start with the simplest case of static SCED. More elaborated
SCED formulations are in Section 5.4. Since the line losses are not explicitly modeled the SCED
formulation, the LMPs in this paper do not contain the loss components. Further discussions on
the loss component are in Section 7.4.
2.3
SCED Analysis via MLP
In real-world market operations, the parameters associated with the SCED above are typically
time-varying. Therefore, it is essential to understand the effects of parameters on the optimality of
the problem. Multi-parametric Programming (MP) problem aims at exploring the characteristics of
an optimization problem which depending on a vector of parameters [11]. Multi-parametric Linear
Programming (MLP) theory, which is the foundation of this paper, pays special attention to Linear
Programming (LP) problems.
In this paper, we would like to understand the impact of parameters (i.e., load levels, line
capacities, etc) on the outcome of SCED (namely, the prices). We pose the problem in view of
MLP, and analyze the theoretical properties.
3
In the reality, the LMP vector depends upon a number of factors, including: (1) the loads in
the system; (2) line flow limits; (3) ramp constraints; (4) generation offer prices; (5) topology of
the system; (6) unit commitment results. We first focus on the relationship between loads and
LMPs assuming the other five factors remain unchanged; then Section 3 takes the line flow limits
and ramp constraints into account; the influence of generation offer prices is explored in Section
7.3. Future work will investigate the impacts of unit commitment results and the system topology
changes on the prices.
Consider the static SCED in the standard MLP form1 :
Primal: min{c| PG : APG + s = b + W PD , s ≥ 0}
|
|
Dual: max{−(b + W PD ) y : A y = −c, y ≥ 0}
(3a)
(3b)
where:
|
0
1nb
1|nb
−1|nb
0
−1|nb
+
F
H
H
A=
−H , b = −F + , W = −H
+
P
0n ×n
In
b
b
b
G
−PG−
0nb ×nb
−Inb
(4)
The load vector PD is the vector of parameters θ, and the load space D is the parameter space Θ.
Since not every PD in the load space leads to a feasible SCED problem, D ∈ D denotes the set of
all feasible vectors of loads. [13] shows that D is a convex polyhedron in D.
Definition 1 (Optimal Partition/System Pattern). For a load vector PD ∈ D, we could find a
finite optimal solution PG∗ and s∗ . Let J = {1, 2, · · · , nc } denote the index set of constraints where
nc = 2 + 2nl + 2ng for Eqn. (3). The optimal partition π = (B, N ) of the set J is defined as follows:
B(PD ) := {i : s∗i = 0 for PD ∈ D}
N (PD ) := {j :
s∗j
(5a)
> 0 for PD ∈ D}
(5b)
B(PD ) := {i : yi∗ > 0 for PD ∈ D}
(6a)
Or in the dual form:
N (PD ) := {j :
yj∗
= 0 for PD ∈ D}
(6b)
Obviously, B ∩ N = ∅ and B ∪ N = J . The optimal partition π = (B, N ) divides the index set into
two parts: binding constraints B and non-binding constraints N . In SCED, the optimal partition
represents the status of the system (e.g. congested lines, marginal generators), and is called system
pattern.
Definition 2 (Critical Region/System Pattern Region). The concept critical region refers to the set
of vectors of parameters which lead to the same optimal partition (system pattern) π = (Bπ , Nπ ):
Sπ := {PD ∈ D : B(PD ) = Bπ }
(7)
For the consideration of consistency, the critical region is called system pattern region (SPR) in
this paper.
1
In other references (e.g. [12, 13]), the primal form of the MLP problem is different. For the consideration of
convenience of analyzing SCED problem, we follow the formulations in [11]. Those two forms are interchangeable.
4
According to the definitions, each SPR is one-to-one mapped to a system pattern, the SPRs are
therefore disjoint and the union of all the SPRs is the feasible set of vectors of loads: ∪i Sπi = D.
All the SPRs together represent a specific partition of the load space. The features of SPRs, which
directly inherit from critical regions in MLP theory, are summarized as follows:
Theorem 1. The load space could be decomposed into many SPRs. Each SPR is a convex polytope.
The relative interiors of SPRs are disjoint convex sets and each corresponds to a unique system
pattern [6]. There exists a separating hyperplane between any two SPRs [8].
Lemma 1 (Complementary Slackness). According to complementary slackness:
AB PG∗ = (b + W PD )B
(8a)
AN PG∗
A|B yB
< (b + W PD )N
(8b)
= −c, yB > 0
(8c)
yN = 0
(8d)
where the (·)B is the sub-matrix or the sub-vector whose row indices are in set B, same meaning
applies for (·)N .
Remark 1. The supply-demand balance equality constraint in is rewritten as two inequalities in
Eqn. (3). These two inequalities will always be binding and appear in the binding constraint set
B at the same time. One of them is redundant and therefore eliminated from the set B. In the
remaining part of the paper, set B denotes the set after elimination.
Remark 2. If the problem is not degenerate, the cardinality of binding constraint set B is equal
to the number of decision variables (i.e. number of generators ng ) 2 . The matrix AB is invertible
and PG∗ is uniquely determined by A−1
B (b + W PD )B .
Remark 3. SCED problems with different generation costs will have different SPRs. For a system
pattern π = (B, N ), its SPR would remain the same as long as the generation cost vector c satisfies
Eqn. (8c).
Lemma 2. Within each SPR, the vector of LMPs is unique [7][8].
The proof of this lemma follows Eqn. (8c) (dual form of system pattern definition). Since the
system pattern π is unique within an SPR Sπ , therefore the solution y ∗ is unique for any PD ∈ Sπ
And the vector of LMPs can be calculated using Eqn. (2). This lemma also illustrates that the
LMP vectors are discrete by nature in the case of linear costs.
Theorem 2. If the SCED problem is not degenerate, then different SPRs have different LMP
vectors.
The proof of Theorem 2 turns out to be non-trivial, and is described as follows. If two SPRs
have the same LMP: λ(i) = λ(j) , their energy components are the same because of the entrywise equality, then Eqn. (2) suggests that the congestion components should also be the same:
H | (µ(i) − µ(j) ) = 0. Given the fact that the null space of H | is always non-empty3 , a critical
question arises: “is it possible that µ(i) − µ(j) belongs to the null space of H | ?” Or equivalently,
“is it possible that different congestion patterns have the same LMP vector?” We show that the
answer is “no”. A complete proof of the theorem is provided in Appendix C.
2
This is consistent with the statement that the number of marginal generators equals to the number of congested
lines plus one.
3
dim(N (H | )) = nl − nb + 1 ≥ 0. The equality holds if and only if the topology of the system is a tree, where
nl = nb − 1.
5
2.4
An Illustrative Example
The 3-bus system in Fig. 1 serves as an illustrative example in this paper. It was first analyzed using
the Multi-Parametric Toolbox 3.0 (MPT 3.0) [14], results are shown in Fig. 2(a). A Monte-Carlo
simulation is conducted, with load vectors colored according to their LMPs. The theoretical results
are verified by the Monte-Carlo simulation results. Notice that PD2 and PD3 could be negative.
This is for the consideration of renewable resources in the system, which are typically considered
as negative loads.
Figure 1: 3-bus System
(a) Theoretical Results Using MPT 3.0
(b) Monte-Carlo Simulation
Figure 2: SPRs of the 3-bus System (Static SCED)
3
SPRs with Varying Parameters
Section 2.3 shows construction properties of the load space with fixed parameters of the system
(e.g. transmission constraints). However, these parameters might be time-varying due to reasons
like dynamic ratings or active ramping constraints. This subsection reveals more features of SPRs
with respect to varying factors in the system.
Lemma 3 (Analytical Form of SPRs). Let IB · (b + W PD ) represent the sub-vector (b + W PD )B ,
where IB is the sub-matrix of the identity matrix whose row indices are in set B. Then the analytical
6
form of the SPRs could be solved from Eqn. (8a) and Eqn. (8b) as follows:
(IN A · (IB A)−1 IB − IN )(b + W PD ) < 0
(9)
We can calculate the analytical expressions of the SPRs using Eqn. (9). An illustrative example
with complete details is provided in Appendix D.
Remark 4. Eqn. (9) could be written as:
(IN A(IB A)−1 IB − IN ) · W PD < (IN − IN A(IB A)−1 IB )b
(10)
This indicates the shape of the SPR Sπ only depends on two factors: (1) the corresponding system
pattern π = (B, N ); (2) matrices A and W , namely the shift factor matrix H according to Eqn.
(4). Small changes of vector b only parallel-shift the SPRs’ boundaries.
3.1
Dynamic Line Rating
Dynamic line rating (DLR), contrary to the static line rating (SLR), refers to the technology that
optimizes the transmission capacity based on the real-time conditions such as ambient temperature
and wind speed [15]. It is considered to be more adaptive in maximizing the line potential while
keeping the secure grid operation.
From dispatch point of view, DLR can be represented by the changes of transmission limits F +
in Eqn. (1c). It changes the vector b in Eqn. (4) and thus translate the boundaries of SPRs.
The 3-bus system in Fig. 1 with different transmission limits is analyzed via MPT 3.0. Compared with the standard transmission limits [60; 60; 80], when we increase the limits by 10% (Fig.
3(a)), SPR #3 expands but SPR #1, #2 and #4 shrink; when we decrease the limits by 10% (Fig.
3(b)), SPR #3 shrinks but SPR #1, #2 and #4 expand. This verifies the claim that dynamic line
ratings only shift the boundaries without altering the shapes of SPRs. The implication of having
DLR is that SPRs in Fig. 4 are overlapping instead of completely separable in Fig. 2(b). Details
of the Monte-Carlo simulation are provided in Section 5.3.1.
(a) Line Limits: (66, 66, 88)
(b) Line Limits: (54, 54, 72)
Figure 3: SPRs of the 3-bus System (Static SCED with DLRs)
7
Figure 4: Monte-Carlo Simulation (Static SCED with DLRs)
3.2
Ramping Constraints
The analysis of SPRs can also be generalized to the dispatch models that include inter-temporal
constraints such as ramping:
PGk−1 − R− ∆t ≤ PGk ≤ PGk−1 + R+ ∆t
(11)
In Eqn.(11), R+ and R− represent the ramp up and down constraints of generators.
Adding ramp constraints to the static SCED problem is equivalent with replacing the generation
capacity constraints Eqn. (1d) with:
max{PG− , PGk−1 − R− ∆t} ≤ PGk ≤ min{PG+ , PGk−1 + R+ ∆t}
(12)
When the ramp capacity is not binding, i.e. PG− > PGk−1 − R− ∆t and PG+ < PGk−1 + R+ ∆t, the
SCED problem is the same as the case where no ramp constraints are considered. The SPRs would
be exactly the same as in Fig. 2(a) and 2(b). However, active ramp constraints change the actual
generation constraints, and therefore change the parameter b in Eqn.(4). This leads to parallel shift
of the boundaries of SPRs. The impacts of ramping constraints on SPRs is similar with the case
of dynamic line ratings.
The 3-bus system, again, is analyzed via both MPT 3.0 and Monte-Carlo simulation. Fig.
5(a) and 5(b) demonstrate the cases where ramp constraints are active. SPRs look similar with
parallel changes on the boundaries. When analyzing the load and LMP data, we will again see the
overlapping SPRs (Fig. 6).
4
A Data-driven Approach to Identifying SPRs
The SPRs depict the fundamental coupling between loads and LMP vectors. Massive historical
data could help market participants estimate SPRs, understand the load-LMP coupling and then
forecast LMPs. This section proposes a data-driven method to identify SPRs, which is a significant improvement of the basic method in [8] by considering varying system parameters and the
probabilistic nature of system parameters.
8
(a) Previous Generation: (30; 30)
(b) Previous Generation: (100; 100)
Figure 5: SPRs of the 3-bus System (SCED with Ramp constraints)
Figure 6: Monte-Carlo Simulation (SCED with Ramp constraints)
9
4.1
4.1.1
The SPR Identification Problem
SPR Identification as a Classification Problem
A classifier is an algorithm to give a label y to each feature vector x. The feature vectors sharing the
same labels belong to the same class. The objective of the classification problem is to find the best
classifier which could classify each feature vector accurately. For the parametric classifiers, there is
always a training set, i.e. a group of feature vectors whose labels are known. There are two steps in
a classification problem: training and classifying. Training usually means solving an optimization
problem over the training set to find the best parameters of the classifier. And classifying is to
classify a new feature vector with the trained classifier.
According to Section 2.3, the load vectors in an SPR share many common features (e.g. vectors
of LMPs). Theorem 2 proved that the LMP vectors are distinct for different SPRs. Therefore,
one SPR can be regarded as a class and the LMP vector is the label of each class. Theorem 1
proves the existence of the separating hyperplanes. Since each separating hyperplane labels two
SPRs with different sides, it turned out that the separating hyperplanes are classifiers and the key
of identifying SPRs is to find optimal hyperplanes, which is exactly the objective of Support Vector
Machine (SVM).
4.1.2
SPR Identification with SVM
Suppose there is a set of labeled load vectors for training and those load vectors belong to only two
distinct SPRs (labels y (i) ∈ {1, −1}). Then the SPR identification problem with a binary SVM
classifier (separable case) is stated below:
min
w,b
s.t
1 |
w w
2
(13a)
(i)
y (i) (w| PD − b) ≥ 1, y (i) ∈ {−1, 1}
(13b)
The word “binary” here specifies only two classes (i.e. SPRs) are being considered. Eqn. (13b)
is feasible only when the two SPRs are not overlapping and there exists at least one hyperplane
thoroughly separating them (separable case). For any load vector PD in the load space, w| PD −b = 0
represents the separating hyperplane where w is the norm vector to the hyperplane. Two lines
satisfying w| PD − b = ±1 separate all the training data and formulate an area with no points
inside. This empty area is called margin. The width of the margin is 2/||w||, which is the distance
between those two lines. The optimal solution refers to the separating hyperplane which maximizes
the width of the margin 2/||w||, therefore the objective of the binary SVM classifier is to minimize
the norm of vector w.
Due to the existence of multiple SPRs, multi-class classifiers are needed. Since Theorem 1
guarantees the existence of separating hyperplanes between every pair of SPRs, the “one-vs-one”
multi-class SVM classifier is incorporated in the data-driven approach to identifying SPRs. Detailed
procedures are summarized in Section 4.2.
4.2
4.2.1
A Data-driven Approach
SPR Identification with Varying System Parameters
When the system parameters are varying (e.g. dynamic line ratings), two SPRs may overlap with
each other. The SPR identification problem is no longer a separable case as in Section 4.1.2. The
SVM classifier needs to incorporate soft margins to allow some tolerance of classification error.
10
Figure 7: SPR Identification Problem with SVM (Separable Case)
The slack variable si is added to Eqn. (13a) and penalties of violation C
objective function. Large C indicates low extent of tolerance.
min
w,b,s
s.t
P
i si
are added in the
X
1 |
w w+C
s(i)
2
(14a)
i
(i)
y (i) (w| PD − b) ≥ 1 − s(i)
s
(i)
≥ 0, y
(i)
(14b)
∈ {−1, 1}
Figure 8: SPR Identification Problem with SVM (Non-Separable Case)
4.2.2
Fitting Posterior Probabilities
The posterior probability is the probability that the hypothesis is true given relevant data or observations. In the classification problem, the posterior probability can be stated as: P(class|input).
Estimating the posterior probability is very helpful in practical problems [16]. When identifying
SPRs, knowing the posterior probability P(y = i|PD and y ∈ {1, 2, · · · , n}) is not only about knowing the classification result y = i (PD belongs to SPR#i), but also understanding the confidence or
possible risk. The market participants could accordingly adjust their bidding strategy and reduce
possible loss.
Although the posterior probabilities are desired, the standard SVM algorithm provides an uncalibrated value which is not a probability as output [16]. Modifications are needed to calculate
11
the binary posterior probabilities P(y = i|PD and y ∈ {i, j}). Common practice is to add a link
function to the binary SVM classifier and train the data to fit the link function. Some typical
link functions include sigmoid functions [16] and Gaussian approximations [17]. In this paper, the
sigmoid link function is selected due to its general better performance than other choices [16].
In general, there are more than two SPRs. What we really want to know is the multi-class
posterior probabilities P(y = i|PD and y ∈ {1, 2, · · · , n}). For short, we will use P(y = i|PD ) to
represent multi-class posterior probabilities. [17] proposed a well-accepted algorithm to calculate
multi-class posterior probabilities from pairwise binary posterior probabilities. This algorithm is
incorporated in our approach and briefly summarized in Appendix B.
4.2.3
The Data-driven Approach
There are three steps in the proposed data-driven approach (Fig. 9):
Figure 9: The Data-driven Approach
Training Suppose there are n different SPRs in the training data set. Each time two SPRs are
selected, trained and we get a binary SVM classifier. This pairwise training procedure is repeated
Cn2 = n(n−1)/2 times and we collect n(n−1)/2 binary classifiers, namely the n(n−1)/2 separating
hyperplanes between any two out of n SPRs.
Classifying/Predicting Given load forecast PD , we could use the “max-vote-wins” algorithm
to get the classification results: each binary classifier provides a classification result (vote) for the
load forecast PD , the SPR which collects the most votes will be the final classification result. The
∗
load forecast PD is therefore pinpointed to an SPR. The LMP forecast: λ̂(PD ) = λ(i ) where i∗ is
the index of the SPR winning the most votes. This step is independent of the data post-processing
procedure.
Data Post-processing Calculate posterior probabilities P(y = i|PD ) for i = 1, 2, · · · , n by
applying Platt’s algorithm and then Hastie and Tibshirani’s algorithm4 . It is worth noting that
the proposed approach is generalizable to many other scenarios with overlapping SPRs in the data,
possible extensions are discussed in Section 7.1.
5
Case Study
In this section, we illustrate the proposed data-driven approaches on two systems.
4
Details of these two algorithms are summarized in Appendix A and B.
12
5.1
Performance Metrics
We first introduce the performance metrics.
5.1.1
5-fold Cross Validation
To evaluate the performance of the model to an independent data set and avoid overfitting, the
k-fold cross validation technique is being used. In k-fold cross-validation, the overall data set is
randomly and evenly partitioned into k subsets. Every time a subset is chosen as validation data
set, and the remaining k − 1 subsets are used for training. This cross-validation process is repeated
k times (k folds), and each subset serves as the validation data set once. The 5-fold cross validation
is being used in this paper.
5.1.2
Classification Accuracy
Classification accuracy is the most common criteria to evaluate the performance of classifiers.
The classification accuracy α is the ratio of the correctly classified points in the validation data set.
When incorporating 5-fold cross validation, the classification accuracy of each fold (α1 , α2 , · · · , α5 ) is
calculated first,P
then the overall performance of the method is evaluated by the average classification
accuracy ᾱ = ( 5i=1 αi )/5.
5.1.3
LMP Forecast Accuracy
The proposed approach forecasts the LMP at every bus. The performance of LMP forecast at bus
i is evaluated by the nodal LMP forecast accuracy βi , which is the average forecast accuracy of all
the validation data points (j = 1, 2, · · · , nv )
βi =
nv
1 X
|λ̂i [j] − λi [j]|
nv
λi [j]
(15)
j=1
The overall LMP forecast accuracy β evaluates the performance of LMP forecast for the whole
system. It is the average of all the nodal LMP forecast accuracy βi (i = 1, 2, · · · , nb ):
β=
nb
1 X
βi
nb
(16)
i=1
5.2
Static SCED with Static Line Ratings
This section explores the simplest case: static SCED with SLRs. Since [8] discusses the 3-bus
system as well as the IEEE 24-bus system, we only examine the data-driven approach on an 118bus system. The same dataset generated in this section is used in Section 6.1 as well.
System Configuration Most of the system settings follow the IEEE 118-bus, 54-unit, 24-hour
system in [18] but with the following changes: (1) the lower bounds of generations are set to zero,
but the upper bounds of generators remain the same as in [18]; (2) generation costs are linear.
Details of the parameters are summarized in [19].
13
Load [18] also provides an hourly system load profile and a bus load distribution profile. With
linear interpolation, the hourly system load profile is modified to be 5-min based. To account for
the variability of loads, we assume the load at each bus follows normal distribution N (µ, σ). The
expectation µ of each nodal load is calculated from the system load profile and bus load distribution
profile, the standard deviation σ is set to be 10% of the expectation. 1440 (5 days, 5-min based)
load vectors are generated, then Matpower [20] solves these 1440 SCED problems and records 1440
LMP vectors. These 1440 load vectors and LMP vectors are the training and validation data.
Simulation Results Results are summarized in Table 4. The classification accuracy is around
67% but the LMP forecast is satisfying. When the classification result of a load vector is correct,
the LMP forecast is correct for every bus, i.e. β = 100%. It is worth noting that even if the
classification fails, the overall LMP forecast still has accuracy about 90%. This is because the
classification errors happen between one SPR and it neighbors. LMPs of adjacent SPRs are similar
due to the fact that only one active constraint is different (Lemma 7). Therefore, the LMP forecast
result is much more accurate than classification.
Table 1: Results of the 118-bus System (Static SCED with SLR)
Fold
1
2
3
4
5
avg
5.3
5.3.1
Classification
64.24%
67.36%
64.93%
71.18%
65.63%
66.67%
LMP Forecast
96.82%
96.71%
96.95%
97.34%
96.84%
96.93%
Static SCED with Dynamic Line Ratings
3-bus System
We start with an illustrative 3-bus system example. This succinct example provides key insights
and visualization of the proposed method.
Data The parameters of the 3-bus system are presented in Fig. 1. The data set is generated
using Matpower with the following assumptions: (1) the load vector is evenly distributed in the
load space; and (2) the transmission limits F is time-varying: for simplicity, we utilize the following
model to calculate the real-time transmission limits F :
F = (1 + ξ)F0
(17)
F0 is the “standard” transmission limits and F0 = [60; 60; 80]. It is the same as the case of static
line ratings. ξ ∼ N (0, 0.1) represents the major factor (e.g. ambient temperature or wind speed)
that impacts the transmission capacities. All the data generated is visualized in Fig. 4.
Simulation Results Table 2 summarizes the classification and LMP forecast accuracies. The
accuracies are around 95% because of the overlapping SPRs.
14
Table 2: Results of 3-bus System (5-fold Validation)
Fold
1
2
3
4
5
avg
Classification
93.967%
93.236%
94.150%
95.612%
94.150%
94.23%
LMP Forecast
96.218%
96.054%
95.767%
96.700%
96.405%
96.23%
Posterior Probabilities The posterior probabilities are visualized. The posterior probabilities
of an SPR compose a surface (Fig. 10(b) and Fig. 10(b)). When putting all the 5 surfaces of
5 SPRs together (shown in Fig. 11), the five surfaces intersect with each other and formulate
some “mountains” and “valleys”. The “mountains” correspond to the inner parts of SPRs, where
the overlapping of SPRs is almost impossible to happen. And the “valleys” always locate at the
boundaries among SPRs.
(a) SPR#3: LMP = (50,50,50)
(b) SPR#4: LMP = (20,50,35)
Figure 10: Posterior Probabilities of Two SPRs
5.3.2
118-bus System
A more comprehensive case study is conducted on the 118-bus system to evaluate the performance
and computational burden of the data-driven approach on a complex system with realistic settings.
System Configuration The only difference from the system configuration in Section 5.2 is about
transmission limits. To consider DLR, we use the same model as Eqn. (17). F0 is the same as the
transmission limits in [18] and ξ ∼ N (0, 0.1).
Performance The algorithm is implemented using the Statistics and Machine Learning Toolbox
of Matlab. Table 3 summarizes the computation time of each step in the data-driven approach
on a PC with Intel i7-2600 8-core [email protected] and 16GB RAM memory. There are 181 SPRs
2 = 16290 SVM classifiers are trained in 58.72 seconds. On
found in 1152 points for training, C181
average, one SVM classifier is trained within 0.004 seconds. This is because most of the SPRs are
15
Figure 11: Posterior Probability Surfaces
completely separable, these cases will be solved in an extremely short time. Those adjacent SPRs
are overlapping and are the major source of the computational burden.
Table 3: Average Computation Time (in seconds)
Steps
(a) training
(b) predicting (288 points)
(c) data post-processing
5.4
Computation Time (s)
58.73
26.8504
701.22
Case Studies with Ramp Constraints
Settings The parameters of the 118-bus system are the same as in Section 5.2. And the ramp
capacities of generators follows the simplified assumption below: each generator could ramp up
(down) to its generation limits in 15 min. For example, a generator with G+ = 200MW and
G− = 125MW, its ramp capacity is: R+ = R− = (200 − 125)/15 = 5MW/min. This setting is
called R0 in Table 5. Due to the temporal coupling of SCED with ramp constraints, a daily load
curve is necessary. The settings of loads are the same as in Section 5.2. 1440 SCED problems are
solved consecutively with Matpower, and 1440 load vectors and LMP vectors are recorded.
Simulation Results The classification and LMP forecast accuracies are summarized in Table. 5.
With the ramp settings above, the classification and LMP forecast are satisfying. However, different
ramp settings would change the results dramatically. As shown in Table. 5: when generators ramp
up/down 2 times faster (R/R0 = 2), the ramp constraints would rarely be active, then it is the same
as static SCED; when generators ramp up/down 2 times slower (R/R0 = 0.5), the actual generation
upper/lower bounds are determined by previous dispatch results and ramp constraints. Generation
16
Table 4: Results of the 118-bus System (Dynamic Line Rating)
Fold
1
2
3
4
5
avg
Classification
61.11%
59.38%
60.76%
51.39%
55.90%
57.71%
LMP Forecast
95.11%
94.53%
95.24%
93.34%
94.22%
94.49%
limits become time-varying and the SPRs are overlapping. This explains the unsatisfying results
when the system is lack of ramp capacities. Furthermore, varying SPRs could also explain the price
spikes during ramping up hours in the morning and ramping down hours in the early evening.
Table 5: Results on 118-bus System with Different Ramp Settings
R/R0
LMP Forecast
6
0.5
44.57%
1
85.10%
2
96.33%
The Impact of Nodal Load Information
We would like to point out that one possible contribution of this paper is to consider the LMP
changes due to nodal load variations. This section dedicates to a detailed discussion about the
impact of nodal load information on the understanding of LMP changes. We first demonstrate the
benefits of having nodal load information in Section 6.1; then Section 6.2 illustrates the effects of
incomplete load information and the attempts to solve the issue.
To concentrate on the effects that incomplete load information brings, we make the following
assumptions: (1) transmission limits are constant, no dynamic line ratings are being considered;
(2) ramp constraints are not taken into account.
6.1
On Nodal Load Levels
Previous literature such as [2] studied the impact of system load levels on the LMPs. An important
concept “critical load level” is defined as the system load level where the step changes of LMPs
happen. Many LMP forecast methods were proposed based on identifying CLLs. But the definition
of CLL assumes that the nodal load levels of all the buses change proportionally. This assumption
constrains the load vectors in the load space to be on a straight line, and the CLLs are indeed the
intersection points of the straight line with the boundaries of SPRs.
We would like to point out that one possible contribution of this paper is to consider the LMP
changes due to nodal-level load variations. Contrary to CLL-based methods, which solve a onedimensional problem, the proposed SVM-based method could explore all the dimensions of the load
space and is indeed a generalization of the CLL-based method.
Consider the SPR identification problem with only one feature vector: the total demand of the
system. Fig. 12 illustrates the problem formulation. Since only the total demand PD = PD1 + PD2
is available, the load vectors in the original SPRs are projected to the axis of total demand. Because
this is a one-dimension problem, the SVM classifier degenerates to the case that there is only one
decision variable b, the direction of the separating hyperplane w is represented by the positivity
17
of b. The objective becomes finding the optimal value b which has the least overlapping points of
different LMPs.
X
min
s(i)
(18a)
b,s
s.t
i
(i)
y (i) (PD − b) ≥ 1 − s(i)
(i)
s
≥ 0, y
(i)
(18b)
∈ {−1, 1}
Figure 12: Identifying Critical Load Levels
We compare this CLL-based method and SVM-based method on the 3-bus system and 118-bus
system. Results are demonstrated in Table. 7, 6 and Fig. 13. The performance of both methods
are close for the nodal LMP forecast of many buses, but the CLL-based method failed to provide
correct forecast of some specific buses (e.g. bus 23 in Fig 13), while the SVM-based method provides
much better results. The SVM-based method is also better on forecasting high prices.
Table 6: Comparison of CLL and SVM (118-bus system)
LMP Forecast
Overall
Price > 45 $/MWh
Worst Forecast (Bus No.)
6.2
CLL
94.82%
88.86%
73.92% (23)
SVM
95.95%
96.32%
88.17% (23)
Incomplete Load Information
In practice, LSEs or other market participants may not have the complete information about load
levels at all buses. We investigate the performance of the algorithm by assuming LSEs have access
to only: (1) the total system-level load; and (2) the nodal load levels in its own area.
To better illustrate the problem formulation. We add a load PD1 at bus 1 to the 3-bus system
in Fig.15 . Modified system is shown in Fig.22. Assume there are three LSEs in the system. LSE
#i at P
bus i has access to the following information: (1) load at bus i: PDi ; and (2) system-level
load: 3i=1 PDi .
5
If there are still two loads in the system, knowing system-level load PD2 +PD3 and PD2 is equivalent with knowing
PD2 and PD3 .
18
Figure 13: Nodal LMP Forecast Accuracy
Figure 14: 3-bus system with 3 loads
19
With incomplete load information, the SPR identification problem becomes more difficult. For
example, LSE 2 observes two SPRs which almost completely overlap with each other (blue and red
in Fig. (15(a))). Since the one-to-one mapping of the LMP vectors and SPRs is not effected by
the incomplete load information, this is still a classification problem. The data-driven approach
can still be applied but the feature vectors are the system load and a subset of nodal load levels,
instead of load levels at every bus in Section 4.
(a) LSE 2
(b) LSE 3
Figure 15: LSEs
Simulation results are summarize in Table.7. The results indicate that classification accuracy
goes down to around around 50% while the LMP forecast accuracy is still satisfactory. This could
be explained by the following observations: (1) Fig. 15(a) and 15(b) are obtained by projecting
the 3D SPRs to a lower dimension space. Since the projection is a linear transformation, although
the SPRs are overlapping, their boundaries remain linear; (2) the LSEs may care more about their
own LMPs. For example, Fig. 15(a) could be re-colored by the LMPs at bus 2 (Fig. 16(a)). Since
there are only two possibilities of LMPs at bus 2 (20 and 50), there are only two colored regions in
Fig. 16(a). Even with relatively low accuracy of the overall classification, the forecast of LMPs at
bus 2 is still accurate.
Table 7: Results 3-bus System
LSE
1
2
3
CLL
LMP@Bus 1
86.08%
70.69%
87.91%
69.48%
LMP@Bus 2
97.45%
96.13%
98.65%
97.15%
LMP@Bus 3
88.53%
89.31%
93.53%
89.24%
Overall
90.69%
85.38%
93.53%
85.29%
When forecasting a subset of nodal LMPs becomes the major concern, it might be more computationally efficient to formulate the problem in a way as Fig. 16(a) shows. The number of classes
decreases significantly and so does the computational burden. But the new colored regions might
be the union of SPRs. Though the colored regions in Fig. 16(a) are convex, the union of convex
sets are usually non-convex. Because of this, the SVM with linear kernel may not be the best
choice. Choosing the best classifier would depend upon the feature of the regions, and will be part
of the future work.
20
(a) PD2 and system load
(b) PD3 and system load
Figure 16: LMP at bus 2
Similar with the case of DLRs or ramp constraints, overlapping SPRs implies uncertainties
and the posterior probabilities are necessary. The posterior probabilities for LSE #2 and #3 are
visualized, respectively.
(a) LSE 2
(b) LSE 3
Figure 17: Posterior Probability Surfaces
Because of the relatively small resistances of transmission lines, the loss components of LMPs
are usually small compared with the other two components. Geometrically speaking, each LMP
vector is a point in the LMP space and the LMPs of the same SPR form a cluster. The center of the
cluster contains the energy, congestion component and the average loss component, the deviations
from the center represent varying loss components due to different line flows. We could run a
clustering algorithm (e.g. K-means) on the LMP data in order to find out the centers of those
clusters. Then the LMP vectors of the same cluster are regarded as the LMPs of the same SPR. By
doing so, the SPR identification problem is modeled as a classification problem. The LMP forecast
is the forecast of energy components, congestion components and the average loss components.
21
7
Discussions
7.1
On Posterior Probabilities
When dealing with uncertainties, it is natural to analyze the data in a probabilistic manner. The
calculation of posterior probabilities is essential and provides the quantification of possible risks.
We only propose the method to calculate posterior probabilities in this paper, but quantification
of the posterior probabilities could yield many interesting applications. For example, LSEs could
consider demand response mechanisms to partially change the load vector and thus shift from high
price SPRs. Market participants could also estimate the price volatilities due to renewables in a
system. Further discussions on how to utilize the posterior probabilities for specific applications
are our future work.
7.2
On the Computational Cost
The theoretical analysis reveals that the load space could be partitioned into many SPRs. This
overall structure of the load space could help solve the SCED problem and shift part of the online
computational burden to offline [21]. The total number of SPRs could help evaluate computational
burden to some extent.
With MPT 3.0, the exact number of SPRs of some IEEE benchmark systems are calculated.
Though the total number of SPRs is finite6 , it grows extremely fast with the scale of the system.
However, with the Monte-Carlo simulation, we found much less SPRs than the theoretical results.
[6] points out that because of the regular patterns of loads, only some subsets of the complete
theoretical load space could be achievable thus helpful in practice. Therefore, only a small subset
of the SPRs is meaningful to be analyzed. This suggests the great potential of reducing the
computational burden. The proposed approach is also parallel computation-friendly, which could
be very useful when dealing with large-scale simulations.
Table 8: Number of SPRs of Some Benchmark Systems
System Info
3 Bus System (Fig.1)
IEEE 6 Bus System
IEEE 9 Bus System
IEEE 14 Bus System
IEEE 24 Bus System
IEEE 118 Bus System
7.3
MPT 3.0
5
20
15
1470
∼ 106
-
Simulation (8640 points)
4
7
7
50
445
971
On Generation Offer Prices
The marginal costs of generators are fluctuating due to many factors such as oil prices. This leads
to the changes of generation offer prices c in the SCED formulation. Intuitively, the SPRs would
change with respect to large offer price variations. Eqn. 8c in Lemma 1 quantifies the variation of
offer prices: for a system pattern π = (B, N ), the corresponding SPR Sπ would remain the same
as long as the generation cost vector c satisfies Eqn. (8c).
An illustrative example is provided below. A diesel turbine is added at bus 3 in Fig. 1, the new
3-bus 3-generator 2-load system is shown in Fig. 18. Suppose the offer price of the diesel turbine
6
n −1
A loose upper bound is 2ng −1 × Cngg+nl .
22
is varying due to the fluctuations of oil prices. Fig. 19(a) shows the SPRs when the offer price of
the new generator is 65; when the offer price increases from 65 to 100, three SPRs are different
while the others remain the same7 . This shows that the SPRs have some extent of robustness to
the varying generation offer prices.
Figure 18: The 3-bus 3-generator 2-load System
(a) c = (20, 50, 65)
(b) c = (20, 50, 100)
Figure 19: System Pattern Regions with Different Generation Offer Prices
7.4
LMPs with Loss Components
Since the line losses are not explicitly modeled in the SCED formulation, all the theoretical analysis
is conducted on the “lossless”LMP vectors. The LMP forecast discussed above is the forecast of
the energy components and congestion components. In the reality, the proposed method could
be applied directly on the markets not considering line losses (e.g. ERCOT), and the electricity
markets providing the energy component, congestion component and loss component separately
(e.g. MISO). There are many possible methods to forecast the loss components, but that is a
separate story.
7
More specifically, we can calculate the condition from Eqn. (8c): if the offer price of the new generator satisfies
c3 < 2c2 − c1 = 80, then the SPRs in Fig. 19(a) would remain the same
23
There are Economic Dispatch models with line losses explicitly modeled (e.g. [22]), similar
analysis using MLP theory could be conducted but it is beyond the scope of this paper.
8
Conclusions
In this paper, we examine the fundamental coupling between nodal load levels and LMPs in realtime SCED. It is shown that the load space can be partitioned into convex system pattern regions,
which are one-to-one mapped with distinct LMP vectors. Based on the theoretical results, we
propose a data-driven learning algorithm for market participants to identify SPRs. Identifying SPRs
is modeled as a classification problem, and the proposed data-driven approach is built upon a “onevs-one” multi-class SVM classifier. The proposed algorithm is shown to be capable of estimating
SPRs solely from historical data without knowing confidential system information such as network
topology and bidding curves. The approach is shown to be extensible towards considering dynamic
line ratings, line losses and partial load information. Simulation results based on the IEEE 118bus system demonstrates that the proposed algorithm is effective in understanding the past and
predicting the future.
This paper is a first step towards developing theoretically rigorous and computationally feasible
algorithms to analyzing the market prices as a result of varying loading levels. Future work should
investigate: (1) the system pattern regions with different unit commitment results and system
topologies; (2) the impacts of multi-interval temporal constraints on the system pattern regions.
Another important avenue of research is to develop efficient learning algorithm to process a large
amount of historical data in near real-time market operations.
Appendices
A
Platt’s Algorithm
Platt’s algorithm [16] is used to calculate the posterior probabilities of a binary SVM classifier.
According to [16], training data is fitted to a sigmoid function by minimizing the negative log
likelihood function.
X
min
−
tk log(rk ) + (1 − tk ) log(1 − rk )
(19a)
A,B
where
k
(k)
rk = P(y = i|PD and y ∈ {i, j})
(k)
= (1 + eAPD
+B −1
)
(k)
(19b)
(k)
PD is the kth load vector in the training set, and the parameter tk associated with PD is calculated
by Eqn. (20), where N+ (N− ) is the number of positive (negative) examples.
(
N+ +1
N+ +2 if yk = +1
tk =
(20)
1
N− +2 if yk = −1
B
Hastie & Tibshirani’s Algorithm
Hastie & Tibshirani’s Algorithm [17] is widely accepted to calculate the multi-class posterior probabilities.
24
The basic idea behind Hastie & Tibshirani’s Algorithm is pretty straightforward: with the
multi-class posterior probabilities (i.e. pˆi = P(y = i|PD and i ∈ {1, 2, · · · , n})), we could estimate
the binary posterior probabilities:
pˆi
µˆij =
(21)
pˆi + pˆj
If pˆi s are correctly estimated, then the estimation of binary posterior probabilities µˆij should be
identical to the calculated binary posterior probabilities rij s from Platt’s algorithm. Therefore the
objective of Hastie & Tibshirani’s Algorithm is to minimize the Kullback-Leibler distance between
µˆij and rij . Details are summarized below:
The first step of Hastie & Tibshirani’s Algorithm is to calculate the binary posterior probabilities
rij based on Platt’s algorithm.
The second step of Hastie & Tibshirani’s Algorithm is to run the following algorithm until
convergence:
1. Start with the initial guess for the pˆi and µˆij = pˆi /(pˆi + pˆj ).
2. Repeat this (i = 1, 2, · · · , n, 1, 2, · · · ) until convergence:
P
j6=i nij rij
pˆi ← pˆi P
(22)
j6=i nij µij
P
Then renormalize pˆi ← pˆi / nj=1 pˆj and recompute µˆij = pˆi /(pˆi + pˆj ).
P
3. If p̂/ pˆi converges to the same p̂, then the algorithm stops, the vector p̂ will be the estimated
multi-class posterior probabilities.
C
Proof of Theorem 2
This section provides the details of the proof of the theorem “different system pattern regions
(SPRs) have different LMP vectors”.
C.1
Basics of the SCED
Consider the SCED problem in the form of Eqn. (1). Its Lagrangian L : Rnb × R × Rnl × Rnl ×
Rng × Rng → R is:
L(PG , λ1 , µ+ , µ− , η+ , η− )
= c| PG + λ1 (1| PG − 1| PD )
+ µ|+ (HPG − HPD − F+ ) − µ|− (HPG − HPD − F− )
|
|
+ η+
(PG − G+ ) − η−
(PG − G− )
(23)
According to KKT conditions, µ+ , µ− , η+ , η− ≥ 0 and
∇|PG L = 0 ⇒ c + λ1 1 + H | µ + η = 0
(24)
where µ = µ+ − µ− and η = η+ − η− . The LMP vector λ can be calculated:
λ = −∇|PD L = λ1 1 + H | µ
This is consistent with [10].
25
(25)
C.2
C.2.1
Preparation
On General Mathematics
The lemma and definition below lie the foundation of the proof of Theorem 2.
Lemma 4 (Convex Piecewise Linear Functions With Parallel Segments). Assume the piecewise
linear function f : Rn → R is composed of m linear functions fk = c|k x where k = 1, 2, · · · , m. Let
D = domf , Dk = dom fk 8 , and assume D and Di are closed convex sets 9 . If f is convex and has
two parallel segments: fi and fj (i 6= j) with ∇| fi = ∇| fj , then fi and fj have to be on the same
hyperplane. Namely:
1. ∀xi ∈ relint Di , ∀xj ∈ relint Dj , f (xj ) = f (xi ) + ∇| f (xi ) · (xj − xi ).
2. ∀xi ∈ relint Di , ∀xj ∈ relint Dj , if the convex combination of xi and xj belongs to Dk instead
of Di or Dj (i 6= j 6= k), then ∇| fk = ∇| fi = ∇| fj and f (xk ) = f (xi ) + ∇| f (xi ) · (xk − xi ).
Lemma 4 could be easily proved by applying the first-order conditions and the definition of
convex functions.
Definition 3 (Adjacent Sets). Given two closed set Di and Dj , and dim(Di ) = dim(Dj ) = d ≥ 2.
We say Di and Dj are adjacent if Di ∩ Dj 6= ∅ and dim(Di ∩ Dj ) = d − 1.
C.2.2
On the Features of SPRs
According to the literatures on Multi-parametric Linear Programming theory:
Lemma 5. If the problem is not degenerate, then the partition of the load space is unique, and Sπ
is an open polyhedron of the same dimension as D [23]. This indicates the dimensions of all the
SPRs are the same.
Lemma 6. The optimal value function f ∗ (PD ) = c| PG∗ (PD ) is convex and piecewise affine over D,
and affine in each SPR. The optimal solution PG∗ within an SPR is an affine function of the load
vector PD [24].
The 3-bus 3-generator 2-load system in Fig. 18 is analyzed via MPT 3.0. The optimal value
function and primal solutions are demonstrated in Fig. 20. This verifies Lemma 6.
Lemma 7 (System Patterns of Adjacent SPRs). Given two system pattern regions (SPRs) Si and
Sj and their system patterns πi = (Bi , Ni ) and πj = (Bj , Nj ). If Si and Sj are adjacent10 , then Bi
and Bj only differ in one entry.
Proof. Lemma 7 is a direct conclusion from Definition 3. If Si and Sj are adjacent but Bi and Bj
differ in k ≥ 2 entries. Then Si ∩ Sj is depicted by k linear constraints. If the constraints are linear
independent11 , then dimSi ∩ Sj = dimSi − k < dimSi − 1. This is contradictory with the definition
of adjacent sets, which requires dimSi ∩ Sj = dimSi − 1.
8
It is obvious that D = ∪k Dk .
The word “closed” indicates for i 6= j: Di ∩ Dj 6= ∅, but relint Di ∩ relint Dj = ∅
10
See Definition 3.
11
If they are linear dependent, we can always eliminate the redundant constraints, which will not make any
difference.
9
26
(a) Optimal Value Function
(b) Primal Solution
Figure 20: Piecewise Affine Optimal Value Function and Primal Solution
C.3
Sketch of the Proof
The optimal value function of SCED problem f ∗ is convex and piecewise affine over D and affine
in each SPR (Lemma 6). According to the definition of LMPs, the LMP vector is the norm vector
of the affine segment of f ∗ . If two different SPRs have the same LMP, then their norm vectors are
the same, thus f ∗ has two parallel linear segments. Lemma 4 claims the existence of two adjacent
SPRs 12 with the same LMP vectors. The system patterns of those two adjacent SPRs, according
to Lemma 7, differ in only one entry. There are only two possibilities: (1) those two SPRs have
one different binding generation constraints, but the binding transmission constraints are the same;
(2) those two SPRs have one different binding transmission constraints, but the binding generation
constraints are the same.
C.4
Proof of Theorem 2
A complete proof is provided in this section.
Assume there exist two SPRs (i, j) which have the same LMP vector λ(i) = λ(j) . Notice that
this equality λ(i) = λ(j) is true for each entry. The LMPs of the slack bus (i.e. bus #1) are the same:
(i)
(j)
λ1 = λ1 . In other words, the energy components of the LMP vectors are the same. According to
H | µ = λ − λ1 1, the congestion components are also the same:
H | µ(i) = H | µ(j)
(26)
η (i) = −c − λ(i) = −c − λ(j) = η (j)
(27)
According to Eqn. (24):
η (i) = η (j) means the marginal generators (which are ON) of the two SPRs (i, j) are also exactly
the same.
Also, according to the analysis in Appendix C.3, there are two adjacent SPRs (i and k 13 ) with
the same LMP vectors. And the system pattern of adjacent SPRs only differ in one entry. This
12
13
May not be exactly same as previous two.
k 6= j is possible.
27
indicates that there is only one different binding constraints: either one different congested line or
one different marginal generator.
We will discuss these two cases separately:
C.4.1
Case 1: same generation shadow prices represent different system patterns
For SPR i and k, similar arguments would show that λ(i) = λ(k) indicates η (i) = η (k) . If SPR i
and k are different, then the same vector η = η (i) = η (k) represents two different sets of marginal
generators. This is possible only when there are some “equivalent” generators. For those “equivalent” generators, if we increase the output of one and decrease the other one by the same amount,
the total generation cost remains the same and optimal. That means both generation outputs are
optimal, the SCED problem has infinite optimal solutions. This is the case that SCED is degenerate
and the Lagrange multipliers cannot be uniquely determined.
C.4.2
Case 2: same congestion component represent different congested lines
We will show this case is not possible.
There is one different congested line between SPR i and SPR j. Since we can label the congested
lines with any non-repetitive numbers, assume line 1 is congested in SPR i but not congested in
SPR j. Similarly, line 2 is congested in SPR j but not congested in SPR i. And the index set of
all the other lines congested in both SPR i and SPRj is denoted by C.
In our previous settings, the matrix AB , which relates with binding constraints, has the structure:
supply-demand balance: 1|nb
(28)
AB = shift factor matrix related with congested lines
matrix related with generation constraints
We rearrange the structure of matrix AB as follows14 :
supply-demand balance: 1|nb
matrix about generation constraints: G
AB =
shift factor matrix of commonly congested lines: HC
shift factor matrix of uniquely congested lines: HU
For SPR i and SPR k:
ABi
|
|
1nb
1nb
E
G
G E
=
HC = HC , ABk = HC HC
h1
h2
h1
h2
(29)
(30)
where [1|nb ; G] is common for both SPR i and SPR k, we use matrix E = [1|nb ; G] to represent it.
HC is the shift factor matrix related with lines congested in both SPR i and SPR k. h1 is the
row of line 1 in the shift factor matrix H, h2 is the row of line 2 in the shift factor matrix H.
h1 , h2 ∈ R1×nb are row vectors, where nb is the number of buses.
−1
The structure of A−1
Bi and ABk is as follows:
14
h
i
(i)
(i)
(i)
B1 B2 β3
h
i
= B1(k) B2(k) β3(k)
A−1
=
Bi
(31)
A−1
Bk
(32)
This step will not make any difference to the theoretical results, but will significantly simplify the notations.
28
(k)
(i)
where β3 , β3 ∈ Rnb ×1 .
Since ABk A−1
Bk = I:
(k)
(k)
(k)
EB1
EB2
Eβ3
I 0 0
= HC B1(k) HC B2(k) HC β3(k) = 0 I 0
(k)
(k)
(k)
0 0 1
h2 B1
h2 B2
h2 β3
ABk · A−1
Bk
(33)
An interesting observation:
−1 |
| |
(A|Bk )−1 A|Bi = (A−1
Bk ) ABi = (ABi ABk )
|
(k)
(k)
(k) |
I
0
0
EB1
EB2
Eβ3
I
0
= HC B1(k) HC B2(k) HC β3(k) = 0
(k)
(k)
(k)
(k)
(k)
(k)
h1 B1
h1 B2
h1 β3
h1 B1
h1 B2
h1 β3
(34)
Multiply yBi on both sides:
(A|Bk )−1 A|Bi × yBi = ((A|Bk )−1 A|Bi ) × yBi
(k) |
(k)
I 0 (B1 )| h|1
α
α
(B1 )| h1
(k) | | (i)
(i)
= 0 I (B2 ) h1 µC = µC + µ1 (B2(k) )| h|1
(k)
(k)
µ1
0
0 0
h1 β3
h1 β3
(35)
Also:
(A|Bk )−1 A|Bi × yBi = (A|Bk )−1 (A|Bi × yBi ) = (A|Bk )−1 × (−c) = yBk
Therefore:
(k) |
α
α
(B1 )| h1
(k)
(i)
µ = µ + µ1 (B (k) )| h|
1
2
C
C
(k)
µ2
0
h1 β3
α
= µC(k)
µ2
(36)
(37)
We get the following equations:
(k)
(i)
µC − µC
(k)
= µ1 (B2 )| h|1
(38)
(k)
µ1 h1 β3
(39)
µ2 =
From the assumption λ(i) = λ(k) we get H | (µ(i) − µ(k) ) = 0. Since the shadow prices of the
non-congested lines are zero:
(i)
(k)
0 = H | (µ(i) − µ(k) ) = HC| (µC − µC ) + µ1 h|1 − µ2 h|2
(40)
Using Eqn.(38):
(k)
0 = −µ1 HC| (B2 )| h|1 + µ1 h|1 − µ2 h|2
From (ABk
)−1 A
Bk
(41)
= I:
(k)
(k)
(k)
(42)
E | (B1 )| h|1 = 0
(43)
B1 E + B2 HC + β3 h2 = I
Using Eqn. (42), (39) and µ1 > 0, Eqn. (41) becomes:
(k)
29
(k)
Since EB1
(k)
= I, (B1 )| = (EE | )−1 E:
(k)
0 = EE | (B1 )| h|1 = Eh|1 = 0
(44)
Given the structure of matrix E, we get the following equation:
h1 1nb = 0
(45)
This is not possible given the feature of the shift factor matrix H.
Therefore it is not possible that different congestion patterns have the same LMP vector.
D
More Details About the SPRs of the 3-bus System in Fig. 18
This section provide complete details about the SPRs of the 3-bus 3-generator 2-load system in
Fig. 18. 10 SPRs are visualized in Fig. 21, and their corresponding system pattern, analytical
form and LMP vectors are summarized in Table 9.
To better illustrate the concept system pattern in Table 9, we provide the detailed formulation of
the SCED problem of the 3-bus system below. According to definition 1, system pattern partitions
all the constraints into two sets: binding constraints B and non-binding constraints N . In Table 9,
we only use the binding constraints to represent the system pattern15 , and each binding constraint
is represented by its index. Also it is worth noticing that the supply-demand balance constraint is
rewritten to be two inequality constraints, therefore it has indices 1 and 216
minimize
PG1 ,PG2 ,PG3
subject to
20PG1 + 50PG2 + 100PG3
PG1 , PG2 + PG3 = PD2 + PD2
: 1, 2
2
1
− (PG2 − PD2 ) − (PG3 − PD3 ) ≤ 60
:3
3
3
2
1
:4
− (PG2 − PD2 ) − (PG3 − PD3 ) ≤ 60
3
3
1
1
(PG2 − PD2 ) − (PG3 − PD3 ) ≤ 80
:5
3
3
1
2
(PG2 − PD2 ) + (PG3 − PD3 ) ≤ 60
:6
3
3
1
2
(PG2 − PD2 ) + (PG3 − PD3 ) ≤ 60
:7
3
3
1
1
− (PG2 − PD2 ) + (PG3 − PD3 ) ≤ 80
:8
3
3
PG1 ≤ PG+1
:9
PG2 ≤ PG+2
: 10
PG+3
: 11
≤ PG 1
: 12
≤ PG 2
: 13
≤ PG 3
: 14
PG 3 ≤
PG−1
PG−2
PG−3
15
16
Since the non-binding constraints are just the complement of the index set {1, 2, · · · , 14}.
Since this constraint is always binding, therefore 1 and 2 are actually equivalent. One of them is redundant.
30
Figure 21: SPRs of the 3-bus System in Fig. 18
Table 9: Details of the SPRs in Fig. 21
SPR No.
System Pattern
Analytical Form of The SPRs
LMPs
0.8944
0.4472
80.4984
0.4472
0.8944
80.4984
−0.7071 0.7071 PD2
169.7056
0.7071 −0.7071 PD ≤ 169.7056
3
0.7071
70.7107
0.7071
−0.7071 −0.7071
0
20
20
20
1
1
2
13
14
2
1
2
4
14
3
1
2
9
14
4
1
2
3
14
0
1.0000
140.0000
0
−1.0000
PD2 ≤ −80.0000
0.4472
147.5805
0.8944 PD3
−0.4472 −0.8944
−80.4984
20
50
80
0
−1.0000
−20.0000
0
1.0000
PD2 ≤ 80.0000
0.7071
0.7071
PD3
176.7767
−70.7107
−0.7071 −0.7071
50
50
50
−100.0000
1.0000
0
0.7071 −0.7071 PD2
−169.7056
−0.7071 0.7071 PD3 ≤ 205.0610
−1.0000
0
120.0000
31
20
50
35
E
−100.0000
0
1.0000
275.7716
0.7071 −0.7071 PD2
≤
120.0000
0
−1.0000 PD3
−169.7056
−0.7071 0.7071
5
1
2
8
14
6
1
2
5
13
7
1
2
4
5
8
1
2
4
10
−1.0000
0
−50.0000
−147.5805
−0.4472 −0.8944 PD2
≤
1.0000
PD3
170.0000
0
0.4472
0.8944
192.3018
9
1
2
9
10
10
1
2
3
10
−100.0000
1.0000
0
−169.7056
0.7071 −0.7071 PD2
−0.7071 0.7071 PD3 ≤ 205.0610
120.0000
−1.0000
0
−140.0000
0 −1
1
0
PD2 ≤ 50.0000
0
190.0000
PD3
1
100.0000
−1 0
1.0000
0
230.0000
−1.0000
PD2
−170.0000
0
≤
−0.7071 −0.7071 PD3
−176.7767
212.1320
0.7071
0.7071
1.0000
0
290.0000
−0.8944 −0.4472 PD2
−214.6625
≤
PD3
−1.0000
−230.0000
0
0.8944
0.4472
237.0232
20
50
−10
20
−60
100
20
50
100
20
60
100
100
100
100
20
180
100
System Pattern Regions in 3D Space
For better illustration, we only visualized the 2-dimension SPRs in previous sections. But the SPRs
are usually polyhedrons in high-dimension space. The visualization of the 3-dimension SPRs could
help readers get more intuition on the high-dimension SPR/polyhedron. One load is added to the
3-bus system in Fig. 1, and the new 3-bus 2-generator 3-load system is shown in Fig. 22. Since
there are 3 loads in the system, the SPRs locate in the 3D space. Similar with the 2D case, the
SPRs are polyhedrons and there exists a separating hyperplane between any two SPRs.
32
Figure 22: 3-bus 2-generator 3-load System
Figure 23: 3D System Pattern Regions
33
Figure 24: Pairwise Visualization of System Pattern Regions
34
References
[1] A. Conejo and E. Castillo, “Locational marginal price sensitivities,” IEEE Trans. Power Syst.,
vol. 20, no. 4, pp. 2026–2033, 2005.
[2] F. Li, “Continuous locational marginal pricing (CLMP),” IEEE Trans. Power Syst., vol. 22,
no. 4, pp. 1638–1646, 2007.
[3] F. Li and R. Bo, “Congestion and price prediction under load variation,” IEEE Trans. Power
Syst., vol. 24, no. 2, pp. 911–922, 2009.
[4] R. Bo and F. Li, “Probabilistic LMP forecasting considering load uncertainty,” IEEE Trans.
Power Syst., vol. 24, no. 3, pp. 1279–1289, 2009.
[5] ——, “Efficient Estimation of Critical Load Levels Using Variable Substitution Method,” IEEE
Trans. Power Syst., vol. 26, no. 4, pp. 2472–2482, nov 2011.
[6] Q. Zhou, L. Tesfatsion, and C. Liu, “Short-term congestion forecasting in wholesale power
markets,” IEEE Trans. Power Syst., vol. 26, no. 4, pp. 2185–2196, 2011.
[7] Y. Ji, R. Thomas, and L. Tong, “Probabilistic Forecast of Real-Time LMP via Multiparametric
Programming,” in Proc. the 48th Annual Hawaii International Conference on System Sciences,
2015.
[8] X. Geng and L. Xie, “A Data-driven Approach to Identifying System Pattern Regions in
Market Operations,” in IEEE Power and Energy Society General Meeting, 2015.
[9] D. Kirschen and G. Strbac, Fundamentals of Power System Economics. John Wiley & Sons,
Ltd., 2005.
[10] F. Wu, P. Varaiya, P. Spiller, and S. Oren, “Folk theorems on transmission access: Proofs and
counterexamples,” Journal of Regulatory Economics, vol. 23, pp. 5–23, 1996.
[11] F. Borrelli, A. Bemporad, and M. Morari, “Geometric algorithm for multiparametric linear
programming,” Journal of Optimization Theory and Applications, vol. 118, no. 3, pp. 515–540,
2003.
[12] I. Adler and R. D. C. Monteiro, “A geometric view of parametric linear programming,” Algorithmica, vol. 8, no. 1-6, pp. 161–176, dec 1992.
[13] T. Gal and J. Nedoma, “Multiparametric linear programming,” Management Science, vol. 18,
no. 7, pp. 406–422, 1972.
[14] M. Herceg, M. Kvasnica, C. Jones, and M. Morari, “Multi-Parametric Toolbox 3.0,” in Proceedings of the European Control Conference, 2013, pp. 502–510.
[15] D. A. Douglass and A.-A. Edris, “Real-time monitoring and dynamic thermal rating of power
transmission circuits,” Power Delivery, IEEE Transactions on, vol. 11, no. 3, pp. 1407–1418,
1996.
[16] J. Platt, “Probabilistic outputs for support vector machines and comparisons to regularized
likelihood methods,” Advances in large margin classifiers, vol. 10, pp. 61–74, 1999.
35
[17] T. Hastie, R. Tibshirani, and Others, “Classification by pairwise coupling,” The annals of
statistics, vol. 26, no. 2, pp. 451–471, 1998.
[18] I. I. of Technology, “IEEE 118-bus, 54-unit, 24-hour system.” [Online]. Available:
http://motor.ece.iit.edu/data/JEAS{ }IEEE118.doc
[19] X. Geng, “Understand LMP-Load Coupling from A Market Participant Perspective: Theory,
Examples and An SVM-based Data-driven Approach,” Master Thesis, Texas A&M University,
College Station, 2015.
[20] R. D. Zimmerman, C. E. Murillo-Sánchez, and R. J. Thomas, “MATPOWER: Steady-state
operations, planning, and analysis tools for power systems research and education,” IEEE
Trans. Power Syst., vol. 26, no. 1, pp. 12–19, 2011.
[21] Y. Ji, L. Tong, and R. J. Thomas, “Probabilistic Forecast of Real-Time LMP and Network
Congestion,” arXiv preprint arXiv:1503.06171, 2015.
[22] E. Litvinov, T. Zheng, G. Rosenwald, and P. Shamsollahi, “Marginal loss modeling in LMP
calculation,” IEEE Trans. Power Syst., vol. 19, no. 2, pp. 880–888, 2004.
[23] C. Filippi, “On the geometry of optimal partition sets in multiparametric linear programming,”
1997.
[24] T. Gal, Postoptimal analyses, parametric programming and related topics: degeneracy, multicriteria decision making, redundancy. Walter de Gruyter, 1994.
36
| 3cs.SY
|
Multi-Sensor Data Pattern Recognition for Multi-Target Localization: A
Machine Learning Approach
arXiv:1703.00084v1 [cs.SY] 28 Feb 2017
Kasthurirengan Suresh, Samuel Silva, Johnathan Votion, and Yongcan Cao
Abstract— Data-target pairing is an important step towards
multi-target localization for the intelligent operation of unmanned systems. Target localization plays a crucial role in numerous applications, such as search, and rescue missions, traffic
management and surveillance. The objective of this paper is to
present an innovative target location learning approach, where
numerous machine learning approaches, including K-means
clustering and supported vector machines (SVM), are used
to learn the data pattern across a list of spatially distributed
sensors. To enable the accurate data association from different
sensors for accurate target localization, appropriate data preprocessing is essential, which is then followed by the application
of different machine learning algorithms to appropriately group
data from different sensors for the accurate localization of multiple targets. Through simulation examples, the performance of
these machine learning algorithms is quantified and compared.
Index Terms— Sensor fusion, Target localization, Machine
learning, Pattern recognition
I. I NTRODUCTION
Conducting surveillance missions using sensor networks is
essential for many civilian and military applications, such as
disaster response [1], border patrol [2], force protection [3],
[4], combat missions [5], and traffic management [6]. One
main task in these missions is to collect data regarding
the operational environment and then obtain intelligence
information from the data. Because the sensors used to
collect data are often spatially distributed, extracting data
patterns becomes critical to obtain accurate knowledge about
the underlying activities.
The existing work on identifying data patterns from spatially distributed sensors is focused on developing probabilistic reasoning techniques without recognizing the specific data
association or data patterns. Existing approaches for multitarget state estimation can be characterized by two features:
a data-to-target assignment algorithm, and an algorithm for
single target state estimation under pre-existing data-to-target
associations. With unknown data association, probabilistic
data association (PDA) [7] and multiple hypothesis tracking
(MHT) [8] are two common approaches where dense measurements are available. In the study of traffic patterns, the
existing research is focused on estimating traffic density and
smart routes [6] without analyzing the data pattern to obtain
better knowledge of traffic information.
The main limitation of the existing research is the lack
of work that addresses data patterns in spatially distributed
sensors, which is crucial in obtaining accurate modeling for
The authors are with the Department of Electrical and Computer Engineering, The University of Texas, San Antonio, TX 78249.
Corresponding Author: Yongcan Cao ([email protected])
the multi-target localization problem. For example, if two
observations/measurements are obtained for a single target,
the estimated location of the target can be better calculated if
such a data correlation can be identified before estimation. If
the correlation is not identified, the target’s location cannot
be estimated accurately. Although, an incorrect estimation
of total target numbers is expected. Existing techniques do
not take into consideration the critical data patterns in multitarget localization. Our proposed method leverages machine
learning tools to uncover the data patterns in order to provide
accurate multi-target localization and state estimation.
In this paper, we consider the multi-target data pattern
recognition problem where a number of targets move along a
known road network with numerous sensors placed at unique
random known locations. The target’s data is recorded as it
passes by the sensor’s location. It is assumed that the sensors
are unable to identify individual targets. The objective of this
paper is to develop algorithms that can effectively extract and
associate correlated data to the same target. More precisely,
we seek to classify all data obtained from spatially distributed
sensors into datasets and associate each of these sets to one
target. To achieve this objective, we first generate datasets
for targets with different motions. Then a pre-processing
algorithm is developed that allows the original dataset to be
translated into another form that yields classifiable datasets.
Numerous machine learning algorithms, including K-means
clustering and support vector machines (SVM), as well as
their variations, are used to classify the translated dataset.
These algorithms are evaluated via numerous simulation
studies that suggest some well-behaved algorithms.
The remainder of this paper is organized as follows. In
Section II, two standard clustering algorithms: K-means and
support vector machine (SVM) are reviewed. Section III
gives the problem formulation. Section IV describes the proposed algorithms that classify datasets for different targets.
Section V provides numerous simulations that illustrate the
performance of the proposed algorithms. Finally, Section VI
provides a brief summary and discusses some potential future
work.
II. P RELIMINARIES
An important form of learning from observations is to
conduct classification of datasets associated with these observations. Traditional techniques generally arrange objects
or data based on their numerical measure and consequently
their similarity/disposition on a feature graph [9]. Among the
several techniques in machine learning, K-means and SVM
are two popular classification algorithms that will be used in
the context of the paper.
A. K-means clustering
K-means is one of the most popular data processing
algorithms. It is a simple unsupervised learning algorithm
used to solve the clustering problem. The K-means method
seeks to minimize the average squared distance between
points in the same cluster [10].
Let X = {x1 , ..., xn } be a set of n points in a ddimensional Euclidean space. Let k denote the number of
clusters. The Euclidean distance between two points xi and
xj is defined by kxi − xj k, where k·k is the 2-norm. For
a subset Y ⊆ X and a point x, the distance between them
is defined as d(x, Y ) = miny∈Y kx − yk [11]. The centroid
for any subset Y ⊆ X is denoted as Ω and is defined as
1 X
y.
(1)
Ω(Y ) =
|Y |
y∈Y
A cost value is assigned to each centroid and is related to the
distance between the position of the centroid and each point
in X. Let C = {c1 , ..., ck } be a set of k centroid positions.
The cost of Y with respect to C is defined as
X
X
2
ΦY (C) =
d2 (y, C) =
min ky − ci k . (2)
y∈Y
y∈Y
i=1,...,k
The K-means algorithm seeks to find a set C of k centers
to minimize ΦX (C). The label/cluster to which each data
sample belongs is defined as the one to which the distance
between ci and X is smaller than to any cj , j 6= i.
The optimal solution to the k-means problem uses Lloyd’s
iteration [12]. To achieve the placement of the centroids that
lead to a minimum cost ΦY (C), the centroids’ position is
updated at each iteration. At the first loop, the k centroids are
placed alongside with the data points at random positions.
Given the position of these centroids, for each data point
xi we find the nearest centroid Ω. After going through the
entire dataset, a cluster Ck = {x1 , ...xu } is formed containing
all xi which its closer cluster is Ck . Having all clusters
being formed, the position of the centroid is then recalculated
according to (1). For the next iteration, the current centroid
position is used to form new Ck respecting the fact that
xi ∈ Ck if and only if, its distance d(xi , Ck ) ≤ d(xi , Cu ) ∀u.
Such an iterations continues until no significant change is
observed on the centroid positions. Given the simplicity of
the algorithm, the convergence is expected to be achieved in
a few iterations.
It has been shown in [10] that with proper definition of
initial set C the accuracy and convergence can be drastically
improved. Section IV will describe this idea further.
B. Support Vector Machines (SVM)
Support Vector Machines is a supervised learning method
used commonly for classification and regression. It consists
of minimizing the empirical classification error while maximizing the geometric margin in classification. This approach
differs from most of the commonly used machine learning
algorithms because it not only aims at simply minimizing the
empirical classification error, but also increasing the dimensional feature space to optimize the classification function
[13].
The SVM implements an idea where the input vectors are
mapped into a high-dimensional feature space through an a
priori nonlinear mapping, identified later as ‘kernel’. In this
high-dimensional space a linear surface can be chosen in
order to ensure generalization of the network [14].
Given a dataset D with N samples, consisting of elements
d
th
in the pattern (xj , yj )N
sample and
j=1 , x ∈ R is the j
yj ∈ (0, 1) is the corresponding class label. Given these
two classes, the objective is to design a hyperplane able to
divide them with the largest margin between classes possible
[15]. Nonlinear SVM’s are often used when a data is not
linearly separable in the input space Rd . In this case the
nonlinear SVM maps the feature vector x ∈ Rd into a high
(up to infinity) dimensional Euclidean space, H, by using
the nonlinear function Φ : Rd 7→ H. For this nonlinear
relation, the decision boundary for a 2 class problem, takes
the following form as
w · Φ(x) + b = 0.
(3)
And we obtain H(w, b, ξ) from the optimization described
as
X
1
2
kwk + C
ξi .
(4)
min
w,b,ξ 2
subject to yi (w · Φ(xi + b) + ξi ≥ 1 and ξi ≥ 0
f or i = 1, ..., N.
Notice that C is a variable that compensates for the size
of w and the sum of ξ in order to avoid over fitting. For
numerical computation purposes, the dual form of (4) is used.
This dual form is defined in (5)-(8) as
N
N
X
1 X
αi ,
min
yi yj αi αj K(xi , xj ) −
α
2 i,j=1
i=1
subject to
N
X
yi αi = 0 and 0 ≤ αi ≤ C,
i=1
f or i = 1, .., N
(5)
where
K(xi , xj ) = Φ(xi ) · Φ(xj )
w=
f (x) =
N
X
i=1
N
X
(6)
αi Φ(xi )
(7)
yi αi K(xi , xj ) + b.
(8)
i=1
Eq. (8) is used as the decision function, defining which
label should be applied to a specific test sample. If f (x) ≤
0 the prediction is labeled +1, otherwise the prediction is
labeled 0. In this case, the binary choice of labels can be
chosen to be anything that ones need to classify. Because
this paper deals with several labels, an approach of “onevs-all”, derived from the standard SVM, will be shown in
section IV.
III. P ROBLEM F ORMULATION
In the context of this paper, we consider a 1-dimensional
road, having a length denoted as D ∈ R>0 . Let S =
{S1 , · · · , SNS } be a set of NS ∈ Z>0 sensors placed along
the road at the locations R = {R1 , · · · , RNS }, where each
element in R is unique and valued in the range (0, D).
As a target passes the sensor, the sensor collects the
target’s information. This information includes the velocity
of the target (denoted as v) and a timestamp representing
when the target passes the sensor (denoted as t). The information collected about the target is disassociated with the
target, meaning that the target for which the information was
received cannot be directly identified using the information.
Let A = {A1 , · · · , ANA } represent a set of NA ∈ Z>0
targets. It is assumed that each target passes each sensor exactly one time. The sensors store their collected information
in the matrices V ∈ RNA ×NS and T ∈ RNA ×NS . The mth
velocity and timestamp measurements obtained by the nth
sensor is recorded as Vn,m ∈ V and Tn,m ∈ T , respectively
(where n ∈ {1, · · · , NS } and m ∈ {1, · · · , NA }). Let
the information X ⊂ R1 be the collection of velocity
data V and timestamp information T , organized such that
each element Xnm ∈ X is a set of data conaining the
values Vn,m and Tn,m . More precisely, the set of all observations from the sensor network is represented as X =
NA
{X11 , X12 , ..., X1NA , ..., XN
}.
S
precisely, given a set of observations from all sensors as
NA
X = {X11 , X12 , ..., X1NA , ..., XN
}, 1 our goal is to classify
S
the data X in the desired form of Xideal as in (9). For the
simplicity of representation, we assume that no false alarm or
missing detection will occur although the proposed methods
can be augmented appropriately to deal with false alarm and
missing detection.
To evaluate the performance of the proposed method,
labeled data is needed to obtain the percentage of true
association between targets and their measurements. Let the
information Y be the collection of information that includes
the true association of targets to their measurements. The
information Y is structured similarly to X except that each
element Ynm ∈ Y also includes the value Ai , which describes
the target identification that the measurement corresponds to.
Algorithm 1 is the pseudo code that describes how multiple datasets of Y are generated. The data sets of Y are
generated in iterations. Let NK be the total number of iterations and let K be the current iteration step. In each iteration
step, there is a sub-iterative process. The sub-iterative process
generates and records the trajectory for a single target. The
trajectory is generated by first defining the initial time and
velocity values for the target (line 5 and 6 of Algorithm 1).
These values are selected from a uniform random variable,
where (vmin , vmax ) and (tmin , tmax ) indicate, respectively,
the upper and lower bound for initial velocity and time. A
unique trajectory is generated for the profile Vi (t) by the
function g(·, ·), which is not fully described here (line 7 of
Algorithm 1). Then, according to this trajectory the function
h(·, ·, ·) is used to insert measurements into the appropriate
elements of YK (line 8 of Algorithm 1). The velocity and
time measurements associated with target Ai are inserted into
the element location of YK according to the sensor positions
R. After inserting all measurements, the information YK is
saved (line 10 of Algorithm 1).
Algorithm 1 Data generation
1:
2:
3:
Fig. 1: Sensor distribution and road configuration
2
Let Xideal ⊂ R be the desired outcome of the form given
by
Xideal
1
X1
..
=
.
NA
X1
···
..
.
···
1
XN
S
..
. ,
NA
XNS
(9)
where each row of Xideal represents the measurements of all
sensors regarding the same target. For example, the ith row
of Xideal is the dataset associated with target Ai observed
by different sensors at different time instances. Because the
targets’ velocities fluctuate as they move along the road
network, the sequence of targets observed by one sensor
may not be the same as that observed by another one,
leading a necessary data pattern recognition problem. More
4:
5:
6:
7:
8:
9:
10:
11:
Output: Generate multiple Y from NS sensors on a road
with length D
R = {R1 , R2 , ..., Rm };
for K = 1 : NK do
for i = 1 : NA do
voi =∼ U [(vmin , vmax )];
toi =∼ U [(tmin , tmax )];
Vi (t) = g(voi , toi );
YK ← h(Vi (t), Ai , R);
end for
Save YK ;
end for
IV. A LGORITHM
To solve the aforementioned problem, the objective of
this section is to derive a set of machine learning algo1 In this case, X m means the mth sample collected by the nth sensor,
n
(m does not identify a specific target).
Θ : (X ⊂ R1 ) 7→ (Xideal ⊂ R2 ),
80
Sensor 1
Sensor 2
Sensor 3
Sensor 4
70
Sensor 5
Sensor 6
Sensor 7
Sensor 8
Sensor 9
Sensor10
60
Velocity
rithms that classify the information X , such that each targetmeasurement pair can be identified. In addition, we are
interested in conducting a comparison of the performance of
these algorithms for different datasets. Using the appropriate
machine learning techniques, it is expected that the random
data X can be re-organized in order to achieve the mapping
described as
50
40
(10)
30
where each row of Xideal represents a set of all measurements associated to the mth target. The performance of the
proposed machine learning algorithms will be evaluated by
the associated accuracy levels.
20
0
10
20
30
Time
40
50
Fig. 2: Dataset with a large velocity variance
A. Data Pre-processing
F : (Si ∈ S) 7→ (S1 ∈ S)
60
50
Velocity
Figures 2, 3, 4 show the case where overlapping between
measurements in X occur without any preprocessing of the
data for different velocity variances. In fact, the overlapping
of dataset becomes more significant if more targets are
involved. By using the raw data X , extracting data patterns
using machine learning algorithms, such as K-means and
SVM, is difficult. Moreover, if there exists overlapping of
data points from different targets, mis-classification of these
points into the same category is well expected. To obtain
more accurate data patterns, an appropriate pre-processing
of X is essential.
The method for pre-processing the data is to project the
measurements in X and estimate what would be that target’s
measurement obtained by sensor S1 . A notation for such a
mapping is given by
T 0 n,m = Tn,m −
(Rn − R1 )
.
Vn,m
Eq. 13 describes the backward extrapolation of the time
from the nth sensor to sensor S1 . For example, the 4th
measurement obtained by sensor S2 includes V2,4 = 20 m/s
and T2,4 = 50 s. Let the distance between sensor S1 and
S2 (evaluated as R2 − R1 ) be 100 units. The estimated time
60
30
0
10
20
30
70
80
90
Fig. 3: Dataset with a medium velocity variance
50
Sensor 1
Sensor 2
Sensor 3
Sensor 4
45
(13)
40
50
Time
Sensor 8
Sensor 9
Sensor10
40
10
∀ (0 ≤ n ≤ NS ) and ∀ (0 ≤ m ≤ NA ).
In this approach the sensor S1 (the first sensor) is taken
as a reference. All other sensor readings are used to generate
an estimate for what the reading at sensor S1 would be. For
each sensor Sn , we have a velocity (Vn,m ), time (Tn,m ) and
position (Rn ). We use the velocity and position of the sensor
to calculate a time value that represents when the target had
passed sensor S1 . The new information is represented as
m
the set of data X 0 n consisting of velocity (Vn,m ) and time
0
(T n,m ).
The definition of this mapping is given by
Sensor 5
Sensor 6
Sensor 7
20
(11)
40
Velocity
The data obtained after pre-processing is represented as
01
01
X 1 · · · X NS
..
..
..
.
(12)
X0 =
.
.
.
0 NA
N
X1
· · · X 0 NA
S
Sensor 1
Sensor 2
Sensor 3
Sensor 4
Sensor 5
Sensor 6
Sensor 7
Sensor 8
Sensor 9
Sensor10
35
30
25
20
15
0
10
20
30
Time
40
50
60
70
Fig. 4: Dataset with a small velocity variance
associated with the measurement X24 is then calculated as
T 0 n,m = 45 s. Conducting the same pre-processing strategy
to the information X results in the converted time estimation
of all measurements with respect to sensor S1 . Figures 5, 6,
7 show the plot of the datasets used in Figures 2, 3, and 4
respectively after applying the proposed data pre-processing
method. As can be observed from Figures 5, 6, 7 that the
processed data yields some patterns that can be potentially
recognized by machine learning algorithms, such as K-means
and SVM.
90
Target 1
Target 2
Target 3
Target 4
80
Velocity
70
Target 5
Target 6
Target 7
clusters. Each cluster here represents a target. A slightly
modified version of K-means algorithm, called K-means++
algorithm [10], is used to provide better performance as
shown in the simulation examples in Section V. For the Kmeans++ algorithm, a center k1 is chosen randomly as the
first step. Then the next cluster center kj+1 is chosen, with
Xnm ∈ X 0 , according to the probability given by,
Target 8
Target 9
Target 10
60
50
40
Ds (Xnm )2
m 2
m ∈X Ds (Xn )
Xn
P
30
20
35
where Ds (Xnm ) denotes the shortest distance from a data
point Xnm to the closest center kj−1 already chosen.
An alternative algorithm chosen to solve this problem is
SVM. The SVM algorithm discussed in section II-B is an
ideal classifier which involves classification between two
categories. In order to develop a multi-class SVM classifier a
“one-vs-all’ classifier is used. There are NS different binary
classifiers built such that for any target Ai , all the points in
class Ai are fed as positive samples, and all points not in
Ai are fed as negative samples. For targets Ai with classifier
functions fi , the output classifier function f (x) for a new
data x can be obtained via the following operation:
30
f (x) = arg max fi (x).
0
5
10
15
Time
20
25
30
Fig. 5: Dataset with a large velocity variance after
pre-processing
50
Target 1
Target 2
Target 3
Target 4
45
Target 5
Target 6
Target 7
Target 8
Target 9
Target 10
Velocity
40
i
25
20
(15)
C. Recovering the original dataset
5
10
15
20
25
30
Time
35
40
45
50
Fig. 6: Dataset with a medium velocity variance after
pre-processing
Target 1
Target 2
Target 3
Mis-classification
60
Target 5
Target 6
Target 7
Target 8
Target 9
Target 10
50
Velocity
(14)
In the previous section, the measurements were classified
for each target based on the pre-processed data. Predicting
and classifying the pre-processed data is an intermediate
solution to the problem and does not directly give any
meaning to our original data classification. When the dataset
is pre-processed and projected backwards with reference
to sensor S1 , the result obtained is given by (12), where
m
X 0 n consists of information Vn,m , t0 n,m and corresponds to
sensor Sn located at Rn . Therefore, the original time tji can
be obtained via the following operation:
tn,m = t0 n,m +
40
30
20
0
5
10
15
Time
20
25
30
35
Fig. 7: Dataset with a small velocity variance after
pre-processing
Rn − R1
.
Vn,m
With the original time tn,m , velocity Vn,m and position of
the sensor Rn , the final classified set X 00 can be obtained.
Each unique row in X 00 consists of measurements associated
to a single target. In particular, X 00 can be written in
the desired form given in (9). Algorithm 2 is the pseudo
code that describes the proposed learning algorithm that
includes the data preprocessing, classification, and recovering
as described in the previous part of the section.
V. S IMULATION
B. Classification Algorithms
Two different algorithms are used for data classification.
A K-means clustering algorithm is used as the primary
approach because the described problem fits the classic
example of an unsupervised learning problem. When an
unlabeled data is clustered with K means, it identifies the
clusters and assigns each of the data points to one of the
In this section, we provide some simulation examples
to demonstrate the effectiveness of the proposed machine
learning algorithms and compare their performances. To
show the benefits of pre-processing, the proposed algorithms
are applied to both the raw data X and the preprocessed
data X 0 . In addition, the accuracy levels of these algorithms
are formally compared. To fully investigate the performance
of the proposed algorithms, we select a variety of target
Algorithm 2 Learning algorithm
Input: X
X11
.
2: Output: Obtain X 00 = ..
X1n
90
X21
..
.
X2n
...
..
.
...
while not at the end
the1 file do
of
1
V1 t1 P1
..
..
.. ;
4:
X[m × n] ← .
.
.
n
n
V1 t1 Pm
5: end while
6: while not at the end of X[m × n] do
1
;
7:
tpji = tji − PiV−P
j
Xm
..
.
70
n
Xm
3:
j
Cluster 5
Cluster 6
Cluster 7
Cluster 8
Cluster 9
Cluster 10
60
50
40
30
20
0
10
20
Time
30
40
50
i
X 0 ← Vij , t0 i , Pi ;
end while
Create a model for K-means++ / one-versus-all SVM
Run / Train the model
while not at the end of X 0 do
j
1
;
13:
tji = t0 i + PiV−P
j
8:
9:
10:
11:
12:
Cluster 1
Cluster 2
Cluster 3
Cluster 4
80
1
Velocity
1:
Fig. 8: K-means++ classification of unprocessed data in
Fig. 2.
i
14:
15:
16:
X 00 ← Vij , tji , Pi ;
end while
return X 00
motions, i.e., the velocity profile of targets falls into different
ranges. The parameters used in the simulation examples are
given in Tables I, II, and III. Figures 2, 3 and 4 show the
plot of the dataset obtained via the simulation parameters in
Tables I, II and III, respectively.
TABLE I: Parameters with a large velocity variance
NA
10
D
1000
Si
every 100(i − 1)
Vi
[30, 80]
Ti
[1, 30]s
TABLE II: Parameters with a medium velocity variance
NA
10
D
1000
Si
every 100(i − 1)
Vi
[20, 50]
Ti
[1, 30]s
TABLE III: Parameters with a small velocity variance
NA
10
D
1000
Si
every 100(i − 1)
Vi
[20, 40]
Fig. 9: K-means++ classification of pre-processed data in
Fig. 5.
Ti
[1, 20]s
Figure 8 shows the classification of unprocessed data
generated using the parameters in Table I. The classification
of the unprocessed data is obtained using the K-means++
clustering algorithm. It can be observed from Figures 2 and 8
that the accuracy is very low because of the significant overlapping of raw data in the absence of data pre-processing.
For the data generated using the parameters shown in Tables
II and III, same worse performance is obtained using Kmeans++ clustering algorithm. Therefore, pre-processing of
the raw datasets is necessary to obtain more accurate data
patterns.
Figures 9, 10 and 11 shows the result of the classification obtained from the K-means++ clustering algorithm on
the pre-processed datasets, generated using the parameters
shown in Tables I, II and III, respectively. It can be observed
from Figures 9 and 10 that the accuracy of the classification
obtained from K-means++ clustering algorithm is very high.
Most data points are classified accurately with very few misclassifications. This high accuracy can be attributed to the
wide velocity and time distribution of the two datasets as
shown in Figures 5 and 7. The accuracy of the K-means++
algorithm on the dataset obtained using Table III is poor
and is shown in Figure 10. The poor accuracy is due to the
narrow distribution of velocity and time in the dataset. This
results in more overlapping of data points even after the preprocessing is performed on the dataset.
To overcome the challenge of classifying datasets with
narrow velocity and time distribution, SVM classifier is
used. The multi-class “one-vs-all’ Quadratic SVM classifier
is used for classifying all the three sets of data, where the
classifier functions fi (·) in (15) are quadratic functions. Our
observation is that the classification accuracy for datasets
generated using parameters shown in Tables I and II is 100%.
For the dataset generated using parameters shown in Table
55
Cluster 1
Cluster 2
Cluster 3
Cluster 4
50
45
Cluster 5
Cluster 6
Cluster 7
100
Cluster 8
Cluster 9
Cluster 10
80
35
Accuracy
Velocity
40
30
25
20
60
40
20
15
10
0
5
10
15
20
Time
25
30
35
40
Fig. 10: K-means++ classification of pre-processed data in
Fig. 6.
45
Target 1
Target 2
Target 3
Target 4
40
Velocity
35
Target 5
Target 6
Target 7
Target 8
Target 9
Target 10
30
25
20
15
10
0
5
10
15
Time
20
25
30
35
Fig. 11: K-means++ classification of pre-processed data in
Fig. 7.
III, the accuracy obtained is 90% and the classification is
shown in Figure 12. The increased accuracy shows that the
benefit of Quadratic SVM (QSVM) in providing a significant
performance increase on narrowly distributed datasets when
compared to the K-means++ algorithm.
55
Target 1
Target 2
Target 3
Mis-classification
50
45
Target 5
Target 6
Target 7
Target 8
Target 9
Target 10
Velocity
40
35
30
25
20
0
LSVM
QSVM
Kmns++
Technique
Kmns
Fig. 13: Accuracy for processed dataset. LSVM: Linear
SVM; QSVM: Quadratic SVM; Kmns++: K-means++
clustering; Kmns: K-means clustering
algorithms. Figure 13 shows the accuracy levels using Kmeans, K-means++, Quadratic SVM (QSVM), and Linear
SVM (LSVM). For K-means algorithm, the accuracy is based
on the average obtained using different datasets. For SVM,
the accuracy is obtained using a cross validation factor of
10 and then averaged for different datasets. Our observations
include: (1) K-means++ provides better performance than the
standard K-means; (2) SVM provides similar performance
as K-means++ in most case; and (3) QSVM provides most
robust performance even when the velocity distributions and
initial time instants for all targets are in a small range.
VI. C ONCLUSION AND F UTURE W ORK
Identifying correct data-target pairing is essential for the
situational awareness that the intelligent operation of unmanned systems relies on. This paper studied data pattern
recognition for multi-target localization from a list of spatially distributed sensors. In contrast to most existing methods that do not take into consideration the data correlation,
we proposed to analyze the data correlation of unlabeled
data from different sensors. A three-step approach, i.e., preprocessing, classification, recovering, was proposed, where
numerous machine learning algorithms, including both Kmeans and SVM, were used to provide reliable classification
for a variety of target motions. In addition, simulation studies
were provided to show that the proposed method offers
a highly accurate solution for recognizing data patterns.
Our future work will focus on studying more general road
network as well as the consideration of false alarms and
missing detections.
15
R EFERENCES
10
[1] S. M. George, W. Zhou, H. Chenji, M. Won, Y. O. Lee, A. Pazarloglou,
R. Stoleru, and P. Barooah, “Distressnet: a wireless ad hoc and sensor
network architecture for situation management in disaster response,”
IEEE Communications Magazine, vol. 48, no. 3, pp. 128–136, 2010.
[2] E. Onur, C. Ersoy, H. Delic, and L. Akarun, “Surveillance wireless sensor networks: Deployment quality analysis,” IEEE Network, vol. 21,
no. 6, pp. 48–53, 2007.
[3] N. Ahmed, D. Casbeer, Y. Cao, and D. Kingston, “Bayesian Hidden
Markov models for UAV-enabled target localization on road networks
with soft-hard data,” in SPIE Defense+Security Conference, 2015.
0
5
10
15
Time
20
25
30
35
Fig. 12: Quadratic SVM (QSVM) classification of
pre-processed data in Fig. 7.
To compare the performance of different machine learning
algorithms, we also plot the accuracy levels of various
[4] D. Casbeer, K. Meier, and Y. Cao, “Estimating the state of an intruder
with a UAV and unattended ground sensors,” in AIAA Infotech@
Aerospace Conference, 2013.
[5] T. Bokareva, W. Hu, S. Kanhere, B. Ristic, N. Gordon, T. Bessell,
M. Rutten, and S. Jha, “Wireless sensor networks for battlefield
surveillance,” in Land Warfare Conference, 2006.
[6] W. Min and L. Wynter, “Real-time road traffic prediction with spatiotemporal correlations,” Transportation Research Part C: Emerging
Technologies, vol. 19, no. 4, pp. 606–616, 2011.
[7] Y. Bar-Shalom, F. Daum, and J. Huang, “The probabilistic data
association filter: Estimation in the presence of measurement origin
uncertainty,” IEEE Transactions on Control Systems Magazine, vol. 29,
no. 12, pp. 82–100, 2009.
[8] S. S. Blackman, “Multiple hypothesis tracking for multiple target
tracking,” IEEE Aerospace and Electronic Systems Magazine, vol. 19,
no. 1, pp. 309–332, 2004.
[9] R. S. Michalski and R. E. Stepp, “Learning from observation: Conceptual clustering,” in Machine learning. Springer, 1983, pp. 331–363.
[10] D. Arthur and S. Vassilvitskii, “K-means++: The advantages of
careful seeding,” in Proceedings of the eighteenth annual ACM-SIAM
symposium on Discrete algorithms. Society for Industrial and Applied
Mathematics, 2007, pp. 1027–1035.
[11] B. Bahmani, B. Moseley, A. Vattani, R. Kumar, and S. Vassilvitskii,
“Scalable K-means++,” Proceedings of the VLDB Endowment, vol. 5,
no. 7, pp. 622–633, 2012.
[12] R. Ostrovsky, Y. Rabani, L. J. Schulman, and C. Swamy, “The
effectiveness of lloyd-type methods for the K-means problem,” in
Foundations of Computer Science, 2006. FOCS’06. 47th Annual IEEE
Symposium on. IEEE, 2006, pp. 165–176.
[13] M. V. Yeo, X. Li, K. Shen, and E. P. Wilder-Smith, “Can SVM be
used for automatic EEG detection of drowsiness during car driving?”
Safety Science, vol. 47, no. 1, pp. 115–124, 2009.
[14] C. Cortes and V. Vapnik, “Support-vector networks,” Machine learning, vol. 20, no. 3, pp. 273–297, 1995.
[15] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern classification. John
Wiley & Sons, 2012.
| 3cs.SY
|
Infinitely generated Hecke algebras with infinite presentation
Corina Ciobotaru∗
arXiv:1603.04599v1 [math.RT] 15 Mar 2016
March 16, 2016
Abstract
For a locally compact group G and a compact subgroup K, the corresponding
Hecke algebra consists of all continuous compactly supported complex functions on
G that are K–bi-invariant. There are many examples of totally disconnected locally
compact groups whose Hecke algebras with respect to the maximal compact subgroups are not commutative. One of those is the universal group U (F )+ , when F
is primitive but not 2–transitive. For this class of groups we prove that the Hecke
algebra with respect to the maximal compact subgroup K is infinitely generated
and infinitely presented. This may be relevant for constructing irreducible unitary
representations of U (F )+ whose subspace of K–fixed vectors has dimension at least
two, or answering the question whether U (F )+ is a type I group or not. On the contrary, when F is 2–transitive that Hecke algebra of U (F )+ is commutative, finitely
generated admitting only one generator.
1
Introduction
The Hecke algebras are very useful tools to study the representation theory of locally
compact groups. For example, in the particular case of a semi-simple algebraic group
G over a non-Archimedean local field there are two important Hecke algebras that can
be associated with: the Hecke algebra of G with respect to the good maximal compact
subgroup, called the spherical Hecke algebra of G, and the Hecke algebra of G with respect to the Iwahori subgroup, which is a smaller compact subgroup. The latter algebra
is called the Iwahori–Hecke algebra of G and plays a very important role in the representation theory of algebraic groups, especially in the Kazhdan–Lusztig theory, being an
intense and rich field of research. The former one is used to study the spherical unitary
dual of semi-simple and analogous groups. That Hecke algebra is moreover commutative and finitely generated with respect to the convolution product. The representation
theory of both algebras are intimately related to the representation theory of G.
In this article we restrict our attention to Hecke algebras associated with specific totally disconnected locally compact groups and their maximal compact subgroups. First,
let us recall the general definition of a Hecke algebra and some more specific motivation.
∗
[email protected]
1
Definition 1.1. Let G be a locally compact group and K ≤ G be a compact subgroup.
We denote by Cc (G, K) the space of continuous, compactly supported complex-valued
functions φ : G → C that are K-bi-invariant, i.e., functions that satisfy the equality
φ(kgk′ ) = φ(g) for every g ∈ G and all k, k ′ ∈ K. We view the C-vector space Cc (G, K)
as an algebra whose multiplication is given by the convolution product
Z
φ(xg)ψ(g−1 )dµ(g)
φ ∗ ψ : x 7→
G
where µ is the left Haar measure on G. Moreover, Cc (G, K) is called the Hecke algebra
corresponding to K ≤ G and (G, K) forms a Gelfand pair if the convolution algebra
Cc (G, K) is commutative.
For a locally compact group G it is well known that there is a natural one-to-one
correspondence between (irreducible) unitary representations of G and (irreducible) nondegenerate ∗–representations of L1 (G) (see for example Dixmier [Dix77, Chapter 13]).
When restricted to a Hecke algebra of G we have the following situation. To any (irreducible) unitary representation (π, H) of G admitting a non-zero K–invariant vector
there is associated a canonical (irreducible) non-degenerate ∗–representation of the Hecke
algebra Cc (G, K) to End(HK ), where HK is the space of K–invariant vectors with respect
to (π, H). In general the converse is not true, namely: the category of non-degenerate
∗–representation of the Hecke algebra Cc (G, K) is not necessarily equivalent to the category of unitary representations of G generated by K–fixed vectors (see for example the
PhD thesis of Hall [Hal99]).
The importance of Gelfand pairs in the theory of unitary representations of locally
compact groups is given by the following well known two results. For the corresponding
definitions one can consult van Dijk [vD09].
Proposition 1.2 (See Proposition 6.3.1 in [vD09]). Let G be a locally compact group
and K a compact subgroup of it. The pair (G, K) is Gelfand if and only if for every
irreducible unitary representation π of G on a Hilbert space H the dimension of the
subspace of K–fixed vectors is at most one.
Corollary 1.3 (See Corollary 6.3.3 in [vD09]). Let (G, K) be a Gelfand pair. The
positive-definite spherical functions on G correspond one-to-one to the equivalence classes
of irreducible unitary representations of G having a one-dimensional K-fixed vector
space.
All known examples of connected and totally disconnected non-compact locally compact groups that admit Gelfand pairs are:
1. all semi-simple non-compact real Lie groups, with finite center, together with their
maximal compact subgroups
2. locally compact groups G that act continuously, properly, by type-preserving automorphisms and strongly transitively on a locally finite thick Euclidean building
∆ together with the stabilizer in G of a special vertex of ∆.
2
In the latter case, by the main theorem of Caprace–Ciobotaru [CC15] those are all
Gelfand pairs that can appear. Recall that the second family of groups includes all
semi-simple algebraic groups over non-Archimedean local fields and closed subgroups of
Aut(T ), that act 2–transitively on the boundary of T , where T is a bi-regular tree with
valence at least 3 at every vertex. Using the polar decomposition of those groups (see for
example [Cio15, Remark 4.6, Lemma 4.7]) and the Lemma of Bernstein [Ber74], we have
that the commutative Hecke algebra Cc (G, K), where K is the stabilizer in G of a special
vertex of its corresponding Bruhat–Tits building, is finitely generated. Moreover, the
main result of Bernstein [Ber74] states that all reductive p-adic Lie groups are groups of
type I (see Definition 5.4.2 in [Dix77]). For the case of closed non-compact subgroups of
Aut(Td ) that act 2–transitively on the boundary of Td , the type I property is still open
in this generality (see [Dem04], [Cio]).
Apart from the case of Gelfand pairs, the structure of Hecke algebras with respect
to maximal compact subgroups is in general much less studied, even if those algebras
can provide very useful information about the unitary representations of that locally
compact groups. In this article we propose to study the structure of non-commutative
Hecke algebras that are associated with a particular family of totally disconnected locally
compact groups and their maximal compact subgroups. This family of groups is given
by the universal groups U (F ) introduced by Burger–Mozes in [BM00, Section 3]. In his
PhD thesis [Ama03], Amann studies these groups from the point of view of their unitary
representations.
First, let us recall the basic definitions.
Definition 1.4. Denote by T the d-regular tree, with d ≥ 3, and by Aut(T ) its
full group of automorphisms, endowed with the compact-open topology. Let ι : E(T ) →
{1, ..., d} be a function from the set E(T ) of unoriented edges of the tree T such that
its restriction to the star E(x) of every vertex x ∈ T is in bijection with {1, ..., d}. A
function ι with those properties is called a legal coloring of the tree T .
Definition 1.5. Let F be a subgroup of permutations of the set {1, ..., d} and let ι be
a legal coloring of T . We define the universal group, with respect to F and ι, to be
U (F ) := {g ∈ Aut(T ) | ι ◦ g ◦ (ι|E(x) )−1 ∈ F, for every x ∈ T }.
By U (F )+ we denote the subgroup generated by the edge-stabilizing elements of
U (F ). Moreover, Proposition 52 of [Ama03] tells us that the groups U (F ), U (F )+ are
independent of the legal coloring ι of T .
From the definition we deduce that U (F ) and U (F )+ are closed subgroups of Aut(T ).
When F is the full permutation group Sym({1, ..., d}) we have that U (F ) = Aut(T ). If
F = id then U (F ) is the group of all automorphisms preserving the coloring ι of T .
Another key property which is used in the sequel is the following.
Definition 1.6 (See [Tit70]). Let T be a locally finite tree and let G ≤ Aut(T ) be a
closed subgroup. We say that G has Tits’ independence property if for every edge
3
e of T we have the equality Ge = GT1 GT2 , where Ti are the two infinite half sub-trees of
T emanating from the edge e and GTi is the pointwise stabilizer of the half-tree Ti .
Following [BM00, Ama03] we know that U (F ) and U (F )+ have Tits’ independence
property.
To avoid heavy notation, for the rest of the article we refer to the following convention.
Convention 1.7. Fix d ≥ 3. We denote by distT the usual distance on T . Let F be a
primitive subgroup of Sym{1, · · · , d}. Consider fixed a coloring ι of T , a vertex x of T
and an edge e in the star of x. For simplicity set G := U (F )+ and K := U (F )+
x , which
is the stabilizer in G of the vertex x. Moreover, for a finite subset A ⊂ T we denote by
GA the pointwise stabilizer in G of the set A. For ξ ∈ ∂ T , Gξ denotes the stabilizer in
G of the ideal point ξ. Let S(x, r) := {y ∈ T | distT (x, y) = r}, where r ∈ N∗ . Let T x,e
be the half-tree of T that emanates from the vertex x and that contains the edge e. Set
Vx,r := S(x, r) ∩ T x,e , for every r ∈ N∗ . We have that |Vx,r | = (d − 1)r−1 . For every two
points y, z ∈ T ∪∂ T , we denote by [y, z] the unique geodesic between y and z in T ∪∂ T .
For a hyperbolic element γ in G let |γ| := minx∈T {distT (x, γ(x))}, which is called the
translation length of γ. Set Min(γ) := {x ∈ T | distT (x, γ(x)) = |γ|}.
By [BM00, Ama03] the group G act 2–transitively on the boundary ∂ T if and only
if F is 2–transitive. From this fact together with the main theorem in [CC15], applied
to our case of d-regular trees, we obtain that (G, K) is a Gelfand pair if and only if F
is 2–transitive. Therefore, when F is primitive but not 2–transitive the Hecke algebra
Cc (G, K) is not commutative. In particular, by Proposition 1.2 the group G admits an
irreducible unitary representation whose subspace of K–fixed vectors has dimension at
least two. Moreover, the theory of unitary representations of G is not at all developed
when F is primitive but not 2–transitive. It is therefore natural to study the structure of
the Hecke algebra Cc (G, K) when F is primitive but not 2–transitive and its irreducible
non-degenerate ∗–representations. Regarding the structure of this algebra we obtain the
following theorem, which is the main result of this article.
Theorem 1.8. Let F be primitive. If F is 2-transitive then the Hecke algebra Cc (G, K)
is finitely generated admitting only one generator. If F is not 2–transitive, then the
Hecke algebra Cc (G, K) is infinitely generated with an infinite presentation.
Understanding irreducible non-degenerate ∗–representations of Cc (G, K) when F is
primitive but not 2–transitive is left for a further study. It would be also interesting to
see if Theorem 1.8 could be used to answer the open question whether G is not a type I
group when F is primitive but not 2–transitive and to construct an explicit irreducible
unitary representation of G whose subspace of K–fixed vectors has dimension at least
two.
The article is structured as follows. In Section 2 we prove combinatorial formulas
that are essential for the proof of our main theorem. In Section 3 we study the structure
of the Hecke algebra Cc (G, K) when F is primitive. This is used in Section 4 to prove
Theorem 1.8.
4
Acknowledgement
We would like to thank Anders Karlsson for his encouragement to continue working on
this article and his useful comments and Laurent Bartholdi, Ofir David, Iain Gordon,
Alex Lubotzky and Agata Smoktunowicz for further discussions regarding finitely and
infinitely generated algebras. This article was written when the author was a post-doc
at the University of Geneva and University of Münster. We would like to thank those
institutions for their hospitality and good conditions of working.
2
Some combinatorial formulas
This section is meant to prove combinatorial formulas relating multinomial and binomial
coefficients, as shown by Proposition 2.4 below. These formulas are proved using Newton’s General Binomial Theorem. As a consequence we obtain Lemma 2.5 below, which
is one of the key ingredients involved in the main results of this article.
Definition 2.1. Let f, f ′ : N∗ × N∗ → N∗ be defined by
f (r, k) := kr (k − 1)r−1 and f ′ (r, k) := (k − 1)r−1 ,
for every (r, k) ∈ N∗ × N∗ , where N∗ := N \ {0} and with the convention that 00 = 1.
n−1
P
For n ∈ N, set Sum(n) := {(k1 , · · · , kn−1 ) | ki ∈ N and n =
ki · i}.
i=1
For l ≥ 2 and (k1 , · · · , kl ) ∈ Nl the corresponding multinomial coefficient is
defined to be
(k1 + · · · + kl )!
k1 + · · · + kl
,
:=
k1 , · · · , kl
k1 !k2 ! · · · kl !
with the convention that 0! = 1. The multinomial coefficient counts the total number
of different words of length k1 + · · · + kl formed with l distinct letters, where ki is the
multiplicity of the ith letter. For l = 2 we obtain the binomial coefficient.
For every r ∈ N with r ≥ 2 let
P1,r (X) := X r−2 + · · · + X + 1
where X is the variable of the polynomial. Notice that P1,r (1) = r − 1. Moreover
r−2
r−2
X
X
(X j − 1) = (X − 1) (X j−1 + · · · + X + 1) .
P1,r (X) − (r − 1) =
j=1
j=1
For every r ∈ N with r ≥ 3 define
r−2
P1,r (X) − (r − 1) X j−1
P2,r (X) :=
(X
+ · · · + X + 1).
=
X −1
j=1
5
More generally, for every i ∈ N with i ≥ 2 and every r ∈ N with r ≥ i + 1 define
Pi,r (X) :=
Pi−1,r (X) − Pi−1,r (1)
.
X −1
Lemma 2.2. For every r ∈ N with r ≥ 3 and every i ∈ {2, · · · , r − 1} we have that
Pi,r (X) =
r−1
X
Pi−1,j (X).
(1)
j=i
Proof. Given r ≥ 3, we prove the lemma by induction on i. First we verify it for i = 2.
Indeed, we have
r−2
P2,r (X) =
P1,r (X) − P1,r (1)
X r−2 + X r−3 + · · · + X + 1 − (r − 1) X X j − 1
=
=
X −1
X −1
X −1
j=1
=
r−1
X
(X j−2 + · · · + 1) =
r−1
X
P1,j (X).
j=2
j=2
Suppose that the formula (1) is true for i and we want to prove it for i + 1. We have
r−1
r−1
X
Pi,r (X) − Pi,r (1) X Pi−1,j (X) − Pi−1,j (1)
Pi,j (X),
=
=
Pi+1,r (X) =
X −1
X −1
j=i
j=i+1
as for j = i we have that Pi−1,i (X) = Pi−1,i (1) because the degree of the polynomial
Pi−1,i (X) is zero.
We also need the following easy lemma.
Lemma 2.3. For every j ∈ N with j ≥ 2 and every i ∈ {1, · · · , j − 1} we have that
(j − 1)!
j−1
.
(2)
Pi,j (1) =
=
i
i!(j − i − 1)!
j−1
Proof. First notice that P1,j (1) = j − 1 =
=
.
1
Given j ≥ 2, one can easily verify formula (2) by induction on i and by using formula
(1) for X = 1 and the known equality
(j−1)!
1!(j−2)!
n
X
l=1
l(l + 1) · · · (l + k) =
n(n + 1) · · · (n + k + 1)
.
k+2
6
Proposition 2.4. For every r ∈ N with r ≥ 2 and every i ∈ {1, · · · , r − 1} we have that
X
r−1
r−i
Pi,r (1) =
=
.
(3)
i
k1 , · · · , kr−1
(k1 ,··· ,kr−1 )∈Sum(r)
k1 +···+kr−1 =r−i
Proof. The key ingredient in order to prove the proposition is the following equality,
known as Newton’s General Binomial Theorem. Let n ∈ N∗ . Formally we have
∞
X
!n
n
1
= (1 − X)−n
(1 + X + X + · · · ) =
=
X
1−X
k=0
∞
∞
X
X
−n
n+k−1
k
k
=
(−1)
X =
Xk,
k
k
2
n
k
k=0
where by definition
−n
k
:=
(4)
k=0
(−n)(−n−1)···(−n−k+1)
,
k!
thus
(−1)k
−n
k
=
n(n+1)···(n+k−1)
.
k!
i
Take n := r− i. Then for
k = i the coefficient of X provided by formula (4) is
r−i+i−1
r−1
=
.
i
i
Now compute the coefficient of X i with respect to the product (1 + X +
X 2 + · · · )r−i .
P
r−i
.
By the definition of the multinomial coefficient this equals
k
,
·
1 · · , ki+1
k2 +2·k3 +···+i·ki+1 =i
k1 +k2 ···+ki+1 =r−i
Notice the following equality of sets:
{(k1 , k2 , · · · , ki+1 ) | k2 + 2 · k3 + · · · + i · ki+1 = i and k1 + k2 · · · + ki+1 = r − i}
=
{(k1 , k2 , · · · , ki+1 ) | (k1 , k2 · · · , ki+1 , 0, · · · , 0) ∈ Sum(r) and k1 + k2 · · · + ki+1 = r − i}
=
{(k1 , k2 , · · · , kr−1 ) | (k1 , k2 · · · , kr−1 ) ∈ Sum(r) and k1 + k2 · · · + kr−1 = r − i},
as in the last set, ki+2 = · · · = kr−1 are always zero. Therefore
X
X
r−i
r−i
=
.
k1 , · · · , ki+1
k1 , · · · , kr−1
k2 +2k3 +···iki+1 =i
k1 +k2 ···+ki+1 =r−i
We conclude that indeed
(k1 ,··· ,kr−1 )∈Sum(r)
k1 +···+kr−1 =r−i
P
(k1 ,··· ,kr−1 )∈Sum(r)
k1 +···+kr−1 =r−i
7
r−i
k1 , · · · , kr−1
=
r−1
i
.
Lemma 2.5. Let r, k ∈ N with r, k ≥ 1. We have that
X
k1 + · · · + kr−1
2r−1
k
−
f (1, k)k1 f (2, k)k2 · · · f (r−1, k)kr−1 = f (r, k)
k1 , · · · , kr−1
(k1 ,··· ,kr−1 )∈Sum(r)
Proof. By Proposition 2.4, we obtain that
X
k1 + · · · + kr−1
f (1, k)k1 f (2, k)k2 · · · f (r − 1, k)kr−1
M := k2r−1 −
k1 , · · · , kr−1
(k1 ,··· ,kr−1 )∈Sum (r)
X
k1 + · · · + kr−1
2r−1
kr (k − 1)k2 +2k3 +···+(r−2)kr−1
=k
−
k1 , · · · , kr−1
(k1 ,··· ,kr−1 )∈Sum(r)
X
k1 + · · · + kr−1
= kr kr−1 −
(k − 1)k2 +2k3 +···+(r−2)kr−1
k1 , · · · , kr−1
(k1 ,··· ,kr−1 )∈Sum (r)
r−1
X
X
r−1
i
(k − 1)r−i
= kr
k
−
1
−
k1 , · · · , kr−1
i=2
=k
r
k
r−1
−1−
r−1
X
(k1 ,··· ,kr−1 )∈Sum (r)
k1 +···+kr−1 =i
r−i
Pr−i,r (1)(k − 1)
i=2
= kr
(k − 1)(kr−2 + · · · + k + 1) −
r−1
X
!
Pr−i,r (1)(k − 1)r−i
i=2
r
= k (k − 1) P1,r (k) − P1,r (1) −
= kr (k − 1)2
r−2
X
r−i−1
Pr−i,r (1)(k − 1)
i=2
r−3
X
P2,r (k) − P2,r (1) −
= ···
!
r−1−(r−3)
X
= kr (k − 1)r−3 Pr−3,r (k) − Pr−3,r (1) −
= kr (k − 1)r−3
!
Pr−i,r (1)(k − 1)r−i−2
i=2
!
(k − 1)Pr−2,r (k) −
2
X
i=2
Pr−i,r (1)(k − 1)r−i−(r−3)
Pr−i,r (1)(k − 1)r−i−(r−3)
i=2
r
r−3
r
r−2
= k (k − 1)
= k (k − 1)
!
((k − 1)Pr−2,r (k) − Pr−2,r (1)(k − 1))
r
r−1
(k − 1)Pr−1,r (k) = k (k − 1)
= kr (k − 1)r−1 = f (r, k).
8
r
r−1
Pr−1,r (1) = k (k − 1)
r−1
r−1
3
Computation of the Hecke algebra
Every function of the Hecke algebra Cc (G, K) is a finite linear combination of functions
of the form 1KgK with g ∈ G. To better understand the Hecke algebra Cc (G, K) one
should be able to evaluate the convolution product of a finite number of functions 1KgK
with g ∈ G. The structure of the group G allows us to characterize its K–double cosets
and to perform computations of the desired convolution products.
3.1
K–double cosets of G
The goal of this subsection is to describe and count K–double cosets of G. Along the
way we also give a description of some special hyperbolic elements of G that play an
important role in that study.
When F is primitive but not 2–transitive, the group G still has some of the properties
of closed non-compact subgroups of Aut(T ) that act 2–transitively on the boundary ∂ T .
Remark 3.1. As F is primitive, given an edge e′ of T at odd distance from e, one
can construct (using the definition of G) a hyperbolic element in G translating e to e′ .
Moreover, every hyperbolic element in G has even translation length.
Lemma 3.2. (KA+ K decomposition) Let F be primitive. Then G admits a KA+ K
decomposition, where
A+ := {γ ∈ G | γ is hyperbolic, e ⊂ Min(γ), γ translates the edge e inside T x,e } ∪ {id}.
Proof. Let g ∈ G. If g(x) = x, then g ∈ K. Suppose that g(x) 6= x. Consider the
geodesic segment [x, g(x)] in T and denote by e1 the edge of the star of x that belongs
to [x, g(x)]. Notice that [x, g(x)] has even length and that there exists k ∈ K such that
k(e1 ) = e; therefore, kg(x) ∈ T x,e . By Remark 3.1 there is a hyperbolic element γ ∈ G
of translation length equal to the length of [x, g(x)] that translates the edge e inside T x,e
and such that γ(x) = kg(x); thus γ −1 kg ∈ K. Notice that the KA+ K decomposition of
an element g ∈ G is not unique.
Remark 3.3. By Lemma 3.2 one notices that every K–double coset of G is of the form
KγK for some γ ∈ A+ . When F is 2–transitive we have that A+ =< a >+ , where a
is a hyperbolic element of G, with |a| = 2 and x ∈ Min(a). In this case we obtain the
well-known polar decomposition of G and the K–double cosets of G are just {Kan K}n≥0 .
In order to describe the hyperbolic elements in A+ we need the following easy but
important remarks.
Remark 3.4. Let γ, γ ′ ∈ A+ be two hyperbolic elements. We claim that γγ ′ is again
a hyperbolic element of A+ and that |γ| + |γ ′ | = |γγ ′ |. Indeed, the edge e endowed
with the orientation pointing towards the boundary ∂ T x,e is sent by any η ∈ A+ to an
edge of T x,e with induced orientation that points towards the boundary ∂ T x,e . This
observation proves the claim.
9
Remark 3.5. For i ∈ {1, · · · , d} let Fi be the stabilizer in F of i. As F is transitive
on the set {1, · · · , d} the number of orbits of Fi acting on the set {1, · · · , d} \ {i} is
independent of the choice of the color i. Therefore, we denote by k the total number of
all such orbits. In addition, independently of the choice of a color i ∈ {1, · · · , d} we are
allowed to denote by nj , where j ∈ {1, · · · , k}, the number of elements of each Fi –orbit
in {1, · · · , d} \ {i}. Notice that n1 + · · · + nk = d − 1. If F is 2–transitive k = 1. If F is
not 2–transitive we have that k ≥ 2.
Let us now describe the elements in A+ of translation length 2. By Remarks 3.1
and 3.5 it is easy to see that there are exactly k distinct K–double cosets KγK, with
γ ∈ A+ and |γ| = 2. Let KγK be such a K–double coset and let γ1 , γ2 ∈ A+ ∩ KγK.
Then |γi | = 2, for every i ∈ {1, 2}. We notice that there are two cases: either there is
k ∈ Ge such that kγ1 (e) = γ2 (e), or kγ1 (e) 6= γ2 (e) for every k ∈ K. Moreover, the
latter mentioned case occurs if and only if F is not 2–transitive.
Definition 3.6. Let F be primitive and let γ1 , γ2 ∈ A+ be such that |γ1 | = |γ2 | = 2. We
say that (γ1 , γ2 ) is a K–primitive pair if there exists k ∈ K such that kγ1 (e) = γ2 (e).
If this is the case, notice that k ∈ Ge and Kγ1 K = Kγ2 K. In particular, for every
k′ ∈ G[x,γ2 (e)] k G[x,γ1 (e)] we have that k′ γ1 (e) = γ2 (e).
Lemma 3.7. Let F be primitive. Then every K–double coset KγK, with γ ∈ A+ of
translation length 2, admits exactly k K–primitive pairs up to K–left-right multiplication. In particular there are exactly k2 K–primitive pairs of A+ , up to K–left-right
multiplication.
Proof. The proof follows from Remark 3.5.
For a further use and following Lemma 3.7 let us make the following notation.
Notation 3.8. For every γ ∈ A+ , with |γ| = 2, we denote by {γ j }j∈{1,··· ,k} ⊂ A+ the
set of different elements of translation length 2 such that for every η ∈ KγK ∩ A+ there
exists a unique j ∈ {1, · · · , k} with the property that (η, γ j ) is a K–primitive pair.
Lemma 3.9. Let F be primitive. Let γ, γ ′ ∈ A+ be such that there exists k ∈ K with
kγ(e) = γ ′ (e) and let (γ1 , γ1′ ) be a K–primitive pair. Respectively, let γ2 , γ2′ ∈ A+ with
|γ2 | = |γ2′ | and such that Kγ2 K = Kγ2′ K. Then there exists k′ ∈ K with the property
that k′ γγ1 (e) = γ ′ γ1′ (e), respectively, k′ γγ2 (x) = γ ′ γ2′ (x).
Proof. We have to find k′ ∈ K such that k′ sends the geodesic segment [x, γγ1 (e)]
into the geodesic segment [x, γ ′ γ1′ (e)]. Notice that by hypothesis, it already exists k ∈
K that sends the geodesic segment [x, γ(e)] ⊂ [x, γγ1 (e)] into the geodesic segment
[x, γ ′ (e)] ⊂ [x, γ ′ γ1′ (e)]. We just need to find k′ ∈ G[x,γ ′ (e)] that sends [γ ′ (e), kγγ1 (e)]
into [γ ′ (e), γ ′ γ1′ (e)]. Suppose the contrary that there is no such k′ ∈ G[x,γ ′ (e)] . Using the
definition of G (or the fact that G has Tits’ independence property, see [Ama03]), we
have that by fixing the edge γ ′ (e), the segment [γ ′ (e), γ ′ γ1′ (e)] is not contained in the
Gγ ′ (e) –orbit of the segment [γ ′ (e), kγγ1 (e)]. Apply now (γ ′ )−1 . The latter assumption
10
becomes that by fixing the edge e, the segment [e, γ1′ (e)] is not contained in the Ge –orbit
of the segment [e, (γ ′ )−1 kγγ1 (e)]. But (γ ′ )−1 kγ ∈ Ge < K and recall that (γ1 , γ1′ ) is a
K–primitive pair. We obtained thus a contradiction and the conclusion follows.
If we just consider that Kγ2 K = Kγ2′ K, one proceeds in the same way as above by
replacing the edge e with the vertex x where is needed.
The importance of K–primitive pairs is expressed by the next proposition where two
K–double cosets of G are compared. Its proof relies on the fact that every element in
A+ can be written as a product of elements of A+ that are of translation length 2; the
aim of describing the elements of A+ being thus achieved.
Proposition 3.10. Let F be primitive and let γ ∈ A+ . Then γ = γ1 γ2 · · · γt , where γi ∈
A+ is such that |γi | = 2, for every i ∈ {1, · · · , t}. Moreover, for any γ1 , · · · , γt , γ1′ , · · · , γt′ ∈
A+ , with |γi | = |γi′ | = 2 for every i ∈ {1, · · · , t}, the following equivalence is true:
Kγ1 γ2 · · · γt K = Kγ1′ γ2′ · · · γt′ K if and only if (γi , γi′ ) are K–primitive for every i ∈
{1, · · · , t − 1} and Kγt K = Kγt′ K.
Proof. Let us prove the first part of the proposition. Notice that if |γ| = 2 the conclusion
follows. Assume that |γ| > 2 and denote the vertices of the geodesic [x, γ(x)] ⊂ T x,e by
x = x0 , x1 , x2 , · · · , x2t . Then by Remark 3.1 there exists γ1 ∈ A+ with |γ1 | = 2 such that
γ1 (e) = [x2 , x3 ]. Notice that γ1−1 γ is still an element of A+ , with |γ1−1 γ| = |γ|−2. Indeed,
this is because the orientation of the edge γ1−1 γ(e) induced from the orientation of e that
points towards the boundary of the half-tree T x,e is preserved. Apply now the above
procedure to γ1−1 γ. There exists thus γ2 ∈ A+ with |γ2 | = 2 and such that γ2−1 γ1−1 γ is
in A+ with |γ2−1 γ1−1 γ| = |γ| − 4. By induction we obtain the elements γ1 , γ2 , · · · , γt with
the property that γ = γ1 γ2 · · · γt . The conclusion follows.
Let us prove the second part of the proposition.
Assume that Kγ1 γ2 · · · γt K = Kγ1′ γ2′ · · · γt′ K. This means that there exist k, k′ ∈ K
such that kγ1 γ2 · · · γt k′ = γ1′ γ2′ · · · γt′ ; this is equivalent to saying that there is k ∈ K
such that the geodesic segment [x, γ1 γ2 · · · γt (x)] is sent by k into the geodesic segment
[x, γ1′ γ2′ · · · γt′ (x)]. From here we deduce that in fact kγ1 (e) = γ1′ (e) ⊂ [x, γ1′ γ2′ · · · γt′ (x)].
Therefore, (γ1′ )−1 kγ1 =: k1 ∈ Ge < K and k1 γ2 · · · γt k′ = γ2′ · · · γt′ . We apply again
the above procedure and by induction we obtain that (γi , γi′ ) are K–primitive for every
i ∈ {1, · · · , t − 1} and Kγt K = Kγt′ K. The implication follows.
Assume now that (γi , γi′ ) are K–primitive for every i ∈ {1, · · · , t − 1} and Kγt K =
′
Kγt K. We need to find k ∈ K such that k([x, γ1 γ2 · · · γt (x)]) = [x, γ1′ γ2′ · · · γt′ (x)]. We
prove this by induction on i ∈ {1, · · · , t} by applying Lemma 3.9. Notice that for i = 1
we just use that (γ1 , γ1′ ) is a K–primitive pair. Suppose that the induction step is true
for i and we want to prove it for i + 1. This means that there is ki ∈ K that sends
the geodesic segment [x, γ1 γ2 · · · γi (e)] into the geodesic segment [x, γ1′ γ2′ · · · γi′ (e)] and
we want to find ki+1 ∈ K that sends the geodesic segment [x, γ1 γ2 · · · γi γi+1 (e)] into the
geodesic segment [x, γ1′ γ2′ · · · γi′ γi+1 (e)]. This follows by applying Lemma 3.9, even for
the last induction step when i + 1 = t. The conclusion is thus proved.
11
The following lemma counts the K–double cosets of G. Together with Proposition 3.10 the goal of this subsection is achieved.
Lemma 3.11. Let F be primitive. Using Convention 1.7, for every r ≥ 1 the total
number of orbits of the group Ge acting on Vx,2r is exactly k2r−1 .
For any two distinct Ge –orbits [x1 ], [x2 ] of the 2r–level Vx,2r , there exist hyperbolic
elements γ1 , γ2 ∈ A+ such that γi (x) = xi , for i ∈ {1, 2}. In addition, for any such
choice of γ1 , γ2 having the above properties we have that Kγ1 K 6= Kγ2 K.
Therefore, for every r ≥ 1 there are exactly k2r−1 disjoint K–double cosets KγK,
where γ ∈ A+ and |γ| = 2r.
Proof. We prove the first assertion by induction on r. For r = 1, this is Remark 3.5.
Suppose now that the induction hypothesis is true for r ≥ 1 and we want to prove it
for r + 1. Fix any vertex y of the 2r–level Vx,2r and denote by [z, y] the corresponding
last edge of the geodesic segment [x, y]. We claim that there are exactly k2 orbits of
G[z,y] acting on Vz,3 . Indeed, this follows from the fact that (d − 1)2 = (n1 + · · · + nk )2 =
Pk
2
i,j=1 ni nj , where we have k terms as the pairs (i, j) and (j, i) give disjoint orbits of
the group G[z,y] acting on Vz,3 . Applying then Tits’ independence property of G for the
subgroup G[z,y] , we obtain that in fact there are k2 orbits of G[x,y] acting on Vz,3 . From
the above claim the induction step follows easily.
The first part of the second assertion of the lemma is Remark 3.1. Notice that for a
given Ge –orbit [x1 ], the hyperbolic element γ1 does not depend, up to Ge –conjugation,
on the chosen representative x1 .
It remains to prove that for any two such hyperbolic elements γ1 , γ2 we have that
Kγ1 K 6= Kγ2 K. Indeed, consider two disjoint Ge –orbits [x1 ], [x2 ] of the 2r–level Vx,2r
and two hyperbolic elements γ1 , γ2 ∈ G such that γi (x) = xi , for i ∈ {1, 2}. By contraposition, suppose that (Kγ1 K) ∩ (Kγ2 K) 6= ∅. Then γ1 = k1 γ2 k2 , for some k1 , k2 ∈ K.
As γi (x) = xi and x1 , x2 ∈ Vx,2r , we have that x1 = k1 (x2 ) and k1 ∈ Ge . This is a
contradiction with [x1 ] 6= [x2 ] as Ge –orbits. The conclusion follows.
3.2
Convolution products of K–double cosets
Recall that our main goal is to prove that the Hecke algebra Cc (G, K) is infinitely
generated when F is primitive but not 2–transitive. Therefore, we need to understand
the elements f ∈ Cc (G, K) and convolution products of those. As the support of any
function f ∈ Cc (G, K) is compact and K–bi-invariant, using Lemma 3.2 the support
of f is covered with a finite number of compact-open subsets of the form KηK, with
η ∈ A+ . We obtain that
X
f=
ai 1Kηi K ,
(5)
i∈I
where I is a finite set, ai ∈ C and ηi ∈ A+ for every i ∈ I. This implies that the set
of all functions 1KγK , with γ ∈ A+ , forms an infinite (countable) base for the C–vector
space Cc (G, K) endowed with the addition of functions.
12
To study convolution products of elements in Cc (G, K) it is enough to compute
convolution products of the form 1Kγ1 K ∗ 1Kγ2 K ∗ · · · ∗ 1Kγn K , where γi ∈ A+ for
every i ∈ {1, · · · , n}. Firstly, given two elements γ1 , γ2 ∈ A+ not necessarily of the
same translation length, we want to evaluate the K–double cosets appearing in the
decomposition (5) of the convolution product 1Kγ1 K ∗ 1Kγ2 K .
Recall that by definition
Z
1Kγ1 K ∗ 1Kγ2 K (x) := 1Kγ1 K (xg)1Kγ2 K (g−1 )dµ(g).
G
In order to determine the support of 1Kγ1 K ∗ 1Kγ2 K , we need to have that xg ∈ Kγ1 K
and g−1 ∈ Kγ2 K. This gives that x ∈ Kγ1 Kγ2 K, implying that the support of the
function 1Kγ1 K ∗ 1Kγ2 K is contained in Kγ1 Kγ2 K. It is easy to check that in fact the
support of 1Kγ1 K ∗ 1Kγ2 K equals Kγ1 Kγ2 K.
Definition 3.12. Let X be a compact K–bi-invariant subset of G. Notice that X is
covered with a finite number of (open) K–double cosets KηK, where η ∈ A+ . We say
that KηK is a maximal K–double coset of X, where η ∈ A+ , if the translation length
|η| is maximal among all K–double cosets that
P appear in the above decomposition of X.
ai 1Kηi K as in equality (5). We say that ai
Let γ1 , γ2 ∈ A+ . Let 1Kγ1 K ∗ 1Kγ2 K =
i∈I
is a maximal coefficient of 1Kγ1 K ∗ 1Kγ2 K if the corresponding K–double coset Kηi K
of the support of 1Kγ1 K ∗ 1Kγ2K is maximal.
Proposition 3.13. Let F be primitive and let γ1 , γ2 ∈ A+ . By Proposition 3.10, let
γ1 = γ1′ γ2′ · · · γt′1 and γ2 = γ1′′ γ2′′ · · · γt′′2 , where γi′ , γj′′ ∈ A+ are such that |γi′ | = |γj′′ | = 2,
for every i ∈ {1, · · · , t1 } and every j ∈ {1, · · · , t2 }.
Then the maximal K–double cosets that appear in Kγ1 Kγ2 K are of the form
Kγ1′ γ2′ · · · γt′1 −1 γ γ1′′ γ2′′ · · · γt′′2 K
where γ can be any element of A+ ∩ Kγt′1 K. In particular, we obtain exactly k such
maximal K–double cosets.
Proof. Let (ξi− , ξi+ ) be the translation axis of γi , for i ∈ {1, 2}. As Kγ1 γ2 K is evidently
a maximal K–double coset of Kγ1 Kγ2 K, it remains to find all other K–double cosets
KηK such that η ∈ A+ and |η| = |γ1 | + |γ2 |.
To compute the maximal K–double cosets that appear in the decomposition of
Kγ1 Kγ2 K, it is enough to study elements of the form γ1 kγ2 for k ∈ K with the property
that k[x, ξ2+ ] ∩ [ξ1− , ξ1+ ] ⊂ [x, ξ1+ ]. Indeed, if k[x, ξ2+ ] ∩ [ξ1− , ξ1+ ] was not just equal to
x and not a subset of (x, ξ1+ ], one would see that Kγ1 kγ2 K is not a maximal K–double
coset of Kγ1 Kγ2 K.
Consider first the case when k ∈ Ge . For such k we claim that Kγ1 kγ2 K = Kγ1 γ2 K.
Indeed, we apply Lemma 3.9, as we need to find k1 ∈ Ge such that k1 γ1 kγ2 (x) = γ1 γ2 (x).
The claim follows. In addition, notice that such k1 fixes point-wise the geodesic segment
[x, γ1 (e)].
13
Consider now the case when k ∈ K such that k(e) ∈
/ (γ1− , γ1+ ). Notice that for
every edge in the star of x, which is not on the bi-infinite geodesic (γ1− , γ1+ ), there
exists some k ∈ K sending e to it. For such k ∈ K we need to decompose the element
γ1 kγ2 , which is in A+ , using elements γ ∈ A+ with |γ| = 2. By the hypothesis on k
we have that |γ1 kγ2 | = |γ1 | + |γ2 | = t1 + t2 . There exist thus γ ′ , γ ′′ ∈ A+ such that
γ1 kγ2 = γ ′ γ ′′ , with |γ ′ | = |γ1 | and |γ ′′ | = |γ2 |. Because of the choice of k, we have that
γ1 (x) = γ ′ (x); thus (γ ′ )−1 γ1 ∈ K and γ ′ (e) belongs to the star of γ1 (x), being different
from γ1 (e) and the edge belonging to the geodesic segment [x, γ1 (x)]. Therefore, the first
t1 − 1 terms of the decomposition of γ1 kγ2 are γ1′ γ2′ · · · γt′1 −1 and the t1 –term appearing
in the decomposition of γ1 kγ2 is an element γ ∈ A+ ∩ Kγt′1 K. Notice that this t1 –term
depends strictly on the element k and up to K-left-right multiplication, every element
of A+ ∩ Kγt′1 K appears as a t1 –term of γ1 kγ2 . From the equality γ1 kγ2 = γ ′ γ ′′ and the
above properties we obtain that Kγ2 K = Kγ ′′ K. Following Lemma 3.9, we conclude
that there exists k′ ∈ K such that k′ γ ′ γ ′′ (x) = γ ′ γ2 (x).
Combining the two cases studied above and using Proposition 3.10, we obtain that
Kγ1 kγ2 K = Kγ1′ γ2′ · · · γt′1 −1 γγ1′′ γ2′′ · · · γt′′2 K, where γ ∈ A+ ∩ Kγt′ K. Moreover, up to
K–primitivity, there are k such γ elements in A+ ∩ Kγt′ K and the last part of the
proposition follows.
The next proposition computes the maximal coefficients appearing in 1Kγ1 K ∗ 1Kγ2 K .
Proposition 3.14. Let F be primitive and let γ1 , γ2 ∈ A+ . By Proposition 3.10, let
γ1 = γ1′ γ2′ · · · γt′1 and γ2 = γ1′′ γ2′′ · · · γt′′2 , where γi′ , γj′′ ∈ A+ are such that |γi′ | = |γj′′ | = 2,
for every i ∈ {1, · · · , t1 } and every j ∈ {1, · · · , t2 }.
Then every maximal coefficient appearing in 1Kγ1 K ∗ 1Kγ2 K equals µ(K), where µ is
the left Haar measure on G. In particular, by normalizing µ(K) = 1, we have that all
maximal coefficients equal one.
P
ai 1Kηi K , where I is a finite
Proof. As in equality (5) we have that 1Kγ1 K ∗ 1Kγ2 K =
i∈I
set, ai ∈ C and ηi ∈ A+ for every i ∈ I. Let Kηi K be a maximal K–double coset and
let h ∈ Kηi K. By Proposition 3.13
ηi = γ1′ γ2′ · · · γt′1 −1 γ γ1′′ γ2′′ · · · γt′′2 = γ1 (γt′1 )−1 γγ2
where γ is an element of A+ ∩ Kγt′1 K. In particular, using Proposition 3.10, we can
suppose, without loss of generality, that γ(x) = γt′1 (x).
We want to compute 1Kγ1 K ∗ 1Kγ2 K (h) which equals ai . As all functions involved
are K–bi-invariant, we can suppose that h = ηi . It remains to evaluate
Z
1Kγ1 K (ηi g)1Kγ2 K (g−1 )dµ(g).
G
This reduces to find all g ∈ Kγ2−1 K such that ηi g ∈ Kγ1 K; this is equivalent to evaluate
14
the intersection ηi−1 Kγ1 K ∩ Kγ2−1 K. We would have that
1Kγ1 K ∗ 1Kγ2 K (ηi ) = ai = µ(ηi−1 Kγ1 K ∩ Kγ2−1 K)
= µ(γ2−1 (γ1 (γt′1 )−1 γ)−1 Kγ1 K ∩ Kγ2−1 K)
= µ(γ −1 (γ1′ γ2′ · · · γt′1 −1 )−1 Kγ1 K ∩ γ2 Kγ2−1 K).
Notice the following. For every g ∈ γ2 Kγ2−1 K, we have
distT (g(x), γ2 (x)) = distT (γ2 k1 γ2−1 k2 (x), γ2 (x)) = |γ2 |.
For g ∈ γ −1 (γ1′ γ2′ · · · γt′1 −1 )−1 Kγ1 K we have
distT (g(x), γ −1 (γ1′ γ2′ · · · γt′1 −1 )−1 (x)) = distT (k3 γ1 (x), x) = |γ1 |.
As γ2 (x) ∈ T x,e and γ −1 (γ1′ γ2′ · · · γt′1 −1 )−1 (x) ∈ T \ T x,e , and because |γ −1 (γ1′ γ2′ · · · γt′1 −1 )−1 | =
|γ1−1 | we obtain that
B(γ −1 (γ1′ γ2′ · · · γt′1 −1 )−1 (x), |γ1 |) ∩ B(γ2 (x), |γ2 |) = {x}
where B(y, r) ⊂ T denotes the closed ball centered at the vertex y and of radius r. This
implies that for every g ∈ γ −1 (γ1′ γ2′ · · · γt′1 −1 )−1 Kγ1 K ∩ γ2 Kγ2−1 K we necessarily have
that g(x) = x; therefore, γ −1 (γ1′ γ2′ · · · γt′1 −1 )−1 Kγ1 K ∩ γ2 Kγ2−1 K ⊂ K. In fact, we claim
that
γ −1 (γ1′ γ2′ · · · γt′1 −1 )−1 Kγ1 K ∩ γ2 Kγ2−1 K = K.
Indeed, it is immediate that K ⊂ γ2 Kγ2−1 K. It remains to show that K ⊂ γ −1 γt′1 γ1−1 Kγ1 K.
As we have supposed that γ(x) = γt′1 (x), we have that γ −1 γt′1 ∈ K, so γ −1 γt′1 γ1−1 γ1 K =
K. This proves the claim and we obtain that
ai = µ(γ −1 (γ1′ γ2′ · · · γt′1 −1 )−1 Kγ1 K ∩ γ2 Kγ2−1 K) = µ(K).
The conclusion follows.
4
The Hecke algebra is infinitely generated
The goal of this section is to prove that the Hecke algebra Cc (G, K) endowed with the
convolution product is infinitely generated.
By contraposition suppose that Cc (G, K) would be a finitely generated algebra with
respect to the convolution product. Let S = {f1 , · · · , fm } be a finite set of generators
for Cc (G, K). Then every f ∈ Cc (G, K) would be written as a finite linear combination
of {f1 , · · · , fm } and finite
By equation (5) we decompose
P i convolution products of those.
every fi ∈ S as fi = nj=1
aij 1Kγij K , where γij ∈ A+ and aij ∈ C. One remarks that
the set S ′ := {1Kγij K }ij ⊂ Cc (G, K) also finitely generates Cc (G, K). In addition, by
letting N := maxγij (|γij |) < ∞ the set S ′′ := {1KγK | γ ∈ A+ with |γ| ≤ N } finitely
15
generates Cc (G, K) too. Therefore, if Cc (G, K) was finitely generated, we could assume,
without loss of generality, that there exists a set S ′′ = {1KγK | γ ∈ A+ with |γ| ≤ N }
that finitely generates Cc (G, K).
It remains to investigate the subsets {γ1 , · · · , γn } ⊂ A+ such that {1Kγ1 K , · · · , 1Kγn K }
would finitely generate Cc (G, K) with respect to the convolution product. For this we
need to solve in Cc (G, K) systems of equations of convolution products and to introduce
a notion of linear independence for such systems of equations.
4.1
Sub-bases and weakly linearly independent equations of degree n
In Section 3.2 we have computed the convolution product of two functions 1Kγ1 K , 1Kγ2 K ∈
Cc (G, K), where γ1 , γ2 ∈ A+ . We have obtained the equation
X
X
ai 1Kηi K +
µ(K)1Kηj K ,
(6)
1Kγ1 K ∗ 1Kγ2 K =
|ηi |<|γ1 |+|γ2 |, i∈I
|ηj |=|γ1 |+|γ2 |, j∈J
where I is a finite set, ai ∈ C, |J| = k, ηi , ηj ∈ A+ and ηj , with j ∈ J, |J| = k, are
as in Proposition 3.13.
From now on we consider that the left Haar measure µ of G is normalized with
µ(K) = 1.
More generally, by considering γ1 , γ2 , · · · , γm ∈ A+ we obtain a more general equation
1Kγ1 K ∗ 1Kγ2 K ∗ · · · ∗ 1Kγm K =
ai 1Kηi K +
X
|ηi |<|γ1 |+···+|γm |, i∈I
where I is a finite set, ai ∈ C, |J| =
Proposition 3.13.
km−1 ,
ηi , ηj ∈
X
1Kηj K ,
|ηj |=|γ1 |+···+|γm |, j∈J
A+
(7)
and ηj are accordingly as in
Definition 4.1. Let f ∈ Cc (G, K). Recall that by equality (5)
X
f=
ai 1Kηi K ,
i∈I
where I is a finite set, ai ∈ C and ηi ∈ A+ for every i ∈ I. We say that (the equation)
f is of degree n if for every i ∈ I, |ηi | ≤ n and there exists i ∈ I such that |ηi | = n. In
particular, equation (7) is of degree n if |γ1 | + · · · + |γm | = n.
Definition 4.2. Let I be finite. We say that {Ei }i∈I is a system of equations of
degree n if:
1. for every i ∈ I, Ei := 1Kγi1 K ∗ 1Kγi2 K ∗ · · · ∗ 1Kγim K for some γij ∈ A+ , where
i
mi is finite and j ∈ {1, · · · , mi }
2. Ei is of degree n for every i ∈ I.
16
Notation 4.3. To simplify the notation, from now on we reserve the letter E to represent
a convolution product of the form 1Kγ1 K ∗ 1Kγ2 K ∗ · · · ∗ 1Kγm K , where γi ∈ A+ for every
i ∈ {1, · · · , m}.
Definition 4.4. A system of equations {E1 , · · · , Er } of degree n is weakly linearly
dependent if there exist b1 , · · · , br ∈ C not all zero such that b1 E1 + · · · + br Er is of
degree strictly less than n. We say that a system of equations {E1 , · · · , Er } of degree n
is weakly linearly independent if it is not weakly linearly dependent.
Definition 4.5. We say that the set {1Kγi K }i∈I , where I is finite and {γi }i∈I ⊂ A+ ,
forms a sub-base of Cc (G, K) of degree ≤ n if the following conditions are satisfied:
1. |γi | ≤ n, for every i ∈ I
2. every function 1KηK , with η ∈ A+ and 0 < |η| ≤ n, can be written as a sum
of a finite linear combination of {1Kγi K }i∈I and a finite linear combination of
convolution products of {1Kγi K }i∈I having degree ≤ n
3. |I| is minimal with the above properties.
Remark 4.6. Let F be primitive. It is immediate that a sub-base of Cc (G, K) of degree
2 exists and it is unique. Its cardinality is k.
Before considering the general case, let us warm up proving the following lemma.
Lemma 4.7. Let F be primitive. Then Cc (G, K) admits a sub-base of degree ≤ 4. Its
cardinality is k + k2 (k − 1). A sub-base of Cc (G, K) of degree ≤ 4 is not unique.
Proof. For the proof recall Definition 2.1 and Notation 3.8. Let also {1Kγj K }j∈{1,··· ,k}
be the sub-base of Cc (G, K) degree 2.
Let η1 , η2 ∈ {γ1j1 , · · · , γkjk }j1 ,··· ,jk ∈{1,··· ,k} . The unique way to obtain functions of the
from 1Kηi1 η K , with i1 ∈ {1, · · · , k}, as maximal K–double coset using the sub-base of
1 2
degree 2 is by convoluting 1Kη1 K and 1Kη2 K :
1Kη1 K ∗ 1Kη2 K =
ai 1Kηi K +
X
k
X
1Kη1i η2 K .
(8)
i=1
|ηi |<|γ1 |+|γ2 |, i∈I
Following Proposition 3.10, in equation (8) appears all functions
1Kηi1 η2 K , with i1 ∈
1
{1, · · · , k}. It is the only one where they can appear.
In order to find a sub-base of Cc (G, K) of degree ≤ 4 we have to choose from
equation (8) f ′ (2, k) = k − 1 different elements of A+ of translation length 4. This
number is independent of the choices made for the elements {γiji }kj,i=1 . By doing this,
one of the terms of degree 4 appearing in equation (8) will be written in terms of
the chosen elements. Without loss of generality, we can suppose that we have chosen
{1Kηi η2 K }i∈{2,··· ,k} to form a sub-base of degree ≤ 4. Moreover, by the same reasoning
1
and making a choice, we choose the following set {1Kγ j γ k | i, l ∈ {1, · · · , k} and j ∈
i
17
l
{2, · · · , k}} ∪ {1Kγj K }j∈{1,··· ,k} . It is immediate to see that this set is a sub-base of
Cc (G, K) of degree ≤ 4. Indeed, for example the function 1Kη11 η2 K can be written using
equation (8) in the chosen sub-base of degree ≤ 4. No fewer elements than above can be
chosen to form such a sub-base. The cardinality of the chosen sub-base of degree ≤ 4 is
f (2, k) + f (1, k) = k2 (k − 1) + k; this number is independent of the choices made. Notice
that the sub-base {1Kγ j γ k | i, l ∈ {1, · · · , k} and j ∈ {2, · · · , k}} ∪ {1Kγj K }j∈{1,··· ,k} is
i l
not unique.
Proposition 4.8. Let F be primitive and let {1Kγj K }j∈{1,··· ,k} be the sub-base of Cc (G, K)
degree 2. For every r ≥ 3 there exists a sub-base of Cc (G, K) of degree ≤ 2r. Its cardinality is f (1, k) + f (2, k) + · · · + f (r, k) and it is uniquely determined given the sub-base
of degree ≤ 2(r − 1). In particular, if k = 1 then f (1, k) + f (2, k) + · · · + f (r, k) = 1, for
every r ≥ 1.
Proof. We prove the proposition by induction on r. First recall Definition 2.1 and
Notation 3.8. By Lemma 4.7 we take {1Kγ j γ k | i, l ∈ {1, · · · , k} and j ∈ {2, · · · , k}} ∪
i l
{1Kγj K }j∈{1,··· ,k} as a sub-base of Cc (G, K) of degree ≤ 4.
Fix a sequence {ηi }i≥1 ⊂ {γ1j1 , · · · , γkjk }j1 ,··· ,jk∈{1,··· ,k} .
Let r = 3. Using the chosen sub-base of Cc (G, K) of degree ≤ 4, we form all equations
E of degree 2r = 6 where functions of the form 1Kηi1 ηi2 η K , with i1 , i2 ∈ {1, · · · , k},
1
2
3
can appear (the total number of these functions is k2 ). By a combinatorial argument,
Lemmas 2.5, 4.7 and equality (7) the total number of these equations is exactly
X
k1 + · · · + kr−1
f ′ (1, k)k1 f ′ (2, k)k2 · · · f ′ (r−1, k)kr−1 = k2 −f ′ (3, k).
k1 , · · · , kr−1
(k1 ,··· ,kr−1 )∈Sum (r)
We claim that all the above k2 − f ′ (3, k) equations form a system of weakly linearly
independent equations of degree 6. Indeed, notice that all functions 1Kηi1 ηi2 η K , with
1
2
3
i1 , i2 ∈ {1, · · · , k}, appear at least once in the above system of equations. In particular,
these functions appear from the convolution product 1Kη1 K ∗ 1Kη2 K ∗ 1Kη3 K . Moreover,
the function 1Kη11 η21 η3 K appears only once, hence, only from the convolution product
1Kη1 K ∗ 1Kη2 K ∗ 1Kη3 K . This is because the sub-base of degree ≤ 4 is chosen to be
{1Kγ j γ K | i, l ∈ {1, · · · , k} and j ∈ {2, · · · , k}}∪{1Kγj K }j∈{1,··· ,k} . Moreover, functions
i l
of the form 1Kη1i η21 η3 K , 1Kη11 η2i η3 K , for every i ∈ {2, · · · , k} appear twice: once from the
convolution product 1Kη1 K ∗ 1Kη2 K ∗ 1Kη3 K and once, respectively, from the convolution
product 1Kηj η2 K ∗ 1Kη3 K , 1Kη1 K ∗ 1Kηj η3 K , where j ∈ {2, · · · , k}. These remarks imply
1
2
that the system of equations we are interested in are indeed weakly linearly independent.
This proves our claim.
By the theory of linear algebra, in order to form a sub-base of degree ≤ 6, given the
sub-base of degree ≤ 4, we have to choose (from the above system of equations of degree
6) f ′ (3, k) characteristic functions corresponding to K–double cosets of degree 6. This
argument is independent and valid for every of the k3 systems of equations.
18
By the proof of the above claim and above facts, to obtain a sub-base of degree ≤ 6
one uniquely can add the set {1Kγ i1 γ i2 γ K }, where i1 , i2 ∈ {2, · · · , k} and j1 , j2 , j3 ∈
j1 j2 j3
{1, · · · , k}. This set together with the sub-base of Cc (G, K) of degree ≤ 4 is minimal
and its cardinality is indeed f (1, k) + f (2, k) + f (3, k). In particular, all the functions of
the form 1Kγ 1 γ i2 γ K , 1Kγ i1 γ 1 γ K , with i1 , i2 , j1 , j2 , j3 ∈ {1, · · · , k}, are written using
j1 j2 j3
j1 j2 j3
the chosen sub-base of degree ≤ 6; they do not appear as elements of that sub-base.
Let us now suppose that the conclusion of the proposition is true for all ≤ r and that
the sub-base of Cc (G, K) of degree ≤ 2r does not contain any function of the form
1Kγ i1 γ i2 ···γ il−1 γ
j1 j2
jl−1 jl K
where l ≤ r and at least one of {i1 , · · · , il−1 } is 1. We have to prove this is also true
for r + 1. Indeed, using only the sub-base of degree ≤ 2r, constructed in the previous
induction steps, we form all equations E of degree 2(r + 1) where functions of the form
1Kηi1 ηi2 ···ηir η K , with i1 , i2 , · · · , ir ∈ {1, · · · , k}, can appear (the total number of these
1
2
r
r+1
functions is kr ). By a combinatorial argument, Lemmas 2.5, 4.7 and equality (7) the
total number of these equations is exactly
X
k1 + · · · + kr
f ′ (1, k)k1 f ′ (2, k)k2 · · · f ′ (r, k)kr = kr − f ′ (r + 1, k).
k1 , · · · , kr
(k1 ,··· ,kr )∈Sum (r+1)
We claim that all the above kr −f ′ (r+1, k) equations form a system of weakly linearly
independent equations of degree 2(r + 1). Indeed, every function 1Kηi1 ηi2 ···ηir η K with
1
2
r
r+1
i1 , · · · , ir ∈ {1, · · · , k} appears at least once in the above system of equations, specifically,
from the convolution product 1Kη1 K ∗ 1Kη2 K ∗ · · · ∗ 1Kηr+1 K . In addition, the function
1Kη11 η21 ···ηr1 ηr+1 K appears exactly once in that system of equations. Moreover, every equation of the above system determines in a unique way a function 1Kηi1 ηi2 ···ηir η K where
r r+1
1 2
the number of appearances of 1 among the coefficients i1 , · · · , ir is maximal. This function uniquely depends on the convolution product involved in that equation. In addition,
there exists at least one different equation of that system of equations such that the latter mention function appears, but it is not anymore ‘maximal’. The only exception is
the function 1Kη11 η21 ···ηr1 ηr+1 K .
If the above system of equations (E1 , · · · , En ), where n = kr − f ′ (r + 1, k), was not
weakly linearly independent there would exist coefficients c1 , · · · , cn ∈ C, not all zero,
such that c1 E1 + · · · + cn En is of degree strictly less then 2(r + 1). By the remarks made
above this would imply that we have reduced all functions 1Kηi1 ηi2 ···ηir η K where at
1
2
r
r+1
least one of {i1 , · · · , il } is 1. This cannot be possible as every equation determines in
a unique way a ‘maximal’ such function. Therefore, the weakly linearly independence
follows. This argument is independent and valid for every of the other kr+1 systems of
equations.
It remains to prove that there is only one choice in order to complete the sub-base
of degree ≤ 2r to a sub-base of degree ≤ 2(r + 1) and that the chosen sub-base of degree
19
2(r+1) is indeed 1Kηi1 ηi2 ···ηir η
1
2
r
r+1 K
, for i1 , · · · , ir ∈ {2, · · · , k}. This follows from the fact
that every function of the form 1Kηi1 ηi2 ···ηir η K , where i1 , · · · , ir ∈ {2, · · · , k} appears
r r+1
1 2
more than every other function 1Kηi1 ηi2 ···ηir η K , where at least one of the coefficients
r r+1
1 2
i1 , · · · , ir is 1. This concludes the induction step and also the proof of the proposition.
4.2
The proof of the main theorem
Theorem 4.9. Let F be primitive. If k = 1 then the Hecke algebra Cc (G, K) is finitely
generated admitting only one generator. If k > 1 then the Hecke algebra Cc (G, K) is
infinitely generated with an infinite presentation.
Proof. Let k = 1. Then F is 2–transitive. By Remark 3.3 we have that A+ =< a >+
where a is a hyperbolic element of G, with |a| = 2 and x ∈ Min(a). The K–double cosets
of G are just {Kan K}n≥0 . It is then easy to see that the function 1KaK alone generates
the Hecke algebra Cc (G, K) (one can also apply the general result of Proposition 4.8).
Consider now the case k > 1 and suppose that Cc (G, K) is a finitely generated
algebra with respect to the convolution product. Arguing as in the introduction to
Section 4 we can assume, without loss of generality, that Cc (G, K) is finitely generated
by the set S ′′ = {1KγK | γ ∈ A+ with |γ| ≤ N }, for some N ∈ N∗ . By Proposition 4.8,
there exists a sub-base of Cc (G, K) of degree ≤ N and by our assumption this sub-base
of Cc (G, K) of degree ≤ N finitely generates Cc (G, K). This is in contradiction with
Proposition 4.8 as the sub-base of Cc (G, K) of degree ≤ N + 1 strictly contains the one
of degree ≤ N . Therefore, Cc (G, K) is infinitely generated. Its infinitely presentation is
the one coming from the convolution products of the generators.
References
[Ama03] Olivier Amann, Group of tree-automorphisms and their unitary representations, ETH Zürich,
2003. PhD thesis. ↑3, 4, 10
[Ber74] I.N. Bernstein, All reductive p-adic groups are of type I, Functional Anal. Appl. 8 (1974), 91–93.
English translation. ↑3
[BM00] Marc Burger and Shahar Mozes, Groups acting on trees: from local to global structure, Inst.
Hautes Études Sci. Publ. Math. 92 (2000), 113–150 (2001). ↑3, 4
[CC15] P-E. Caprace and C. Ciobotaru, Gelfand pairs and strong transitivity for Euclidean buildings,
Ergodic Theory and Dynamical Systems 35 (2015), 1056–1078. arXiv:1304.6210. ↑3, 4
[Cio15] C. Ciobotaru, A unified proof of the Howe–Moore property, Journal of Lie Theory 25 (2015),
no. 1, 065–089. arXiv:1403.0223. ↑3
[Cio]
, A note on type I groups acting on d–regular trees. arXiv:1506.02950. ↑3
[Dem04] Selçuk Demir, Some finiteness results in the representation theory of isometry groups of regular
trees, Geom. Dedicata 105 (2004), 189–207. ↑3
[Dix77] Jacques Dixmier, C ∗ -algebras, North-Holland Publishing Co., Amsterdam-New York-Oxford,
1977. Translated from the French by Francis Jellett; North-Holland Mathematical Library, Vol.
15. MR0458185 ↑2, 3
20
[FTN91] Alessandro Figà-Talamanca and Claudio Nebbia, Harmonic analysis and representation theory
for groups acting on homogeneous trees, London Mathematical Society Lecture Note Series,
vol. 162, Cambridge University Press, Cambridge, 1991. ↑
[Hal99] Rachel Hall, Hecke C ∗ -algebras, Ph.D thesis, Penn. State University, 1999. ↑2
[Tit70] Jacques Tits, Sur le groupe des automorphismes d’un arbre, Essays on topology and related
topics (Mémoires dédiés à Georges de Rham), Springer, New York, 1970, pp. 188–211 (French).
↑3
[vD09] G. van Dijk, Introduction to Harmonic Analysis and Generalized Gelfand Pairs, Walter de
Gruyter Studies in Mathematics 36, 2009. ↑2
21
| 4math.GR
|
1
SER for Optimal Combining in Rician Fading
with Co-channel Interference
arXiv:1708.06336v1 [eess.SP] 14 Aug 2017
Muralikrishnan Srinivasan
S. Kalyani
Department of Electrical Engineering,
Indian Institute of Technology, Madras,
Chennai, India 600036.
{ee14d206,skalyani}@ee.iitm.ac.in
Abstract
Approximate Symbol error rate (SER) expressions are derived for receive diversity system employing optimum combining, when both the desired and the interfering signals are subjected to Rician
fading, for the case of a) equal power uncorrelated interferers b) unequal power interferers c) interferer
correlation. The derived expressions are applicable for an arbitrary number of receive antennas and
interferers and for any QAM constellation. Furthermore, we derive a simple closed form expression
for SER in the interference-limited regime, for the special case of Rayleigh faded interferers. A close
match is observed between the SER result obtained through the derived analytical expression and the
one obtained from Monte-Carlo simulations, for a variety of QAM constellations.
Index Terms
Optimum combining, Rician fading, SER, QAM, Wishart matrices, Hypergeometric functions
I. I NTRODUCTION
Various diversity combining schemes have been proposed to exploit space diversity offered by
adaptive antenna arrays. Optimum combining (OC) proposed in [1] is a scheme that maximizes
signal to interference plus noise ratio (SINR). Performance of OC receivers has been extensively
studied for various cases, such as a) when both the desired and interfering signals are subjected to
Rayleigh fading [2]–[5], b) when the number of interferers is more than the number of receiving
antennas [6], c) when dual-antenna diversity reception is used with multiple interferers in M-PSK
September 18, 2017
DRAFT
2
[7], d) for arbitrary number of receiving antennas and equal power interferers [8], [9], e) when
user signal is Rayleigh faded and interferers are correlated [10], [11]. Outage probability of OC,
in the presence of Rayleigh faded interferers, has also been sufficiently studied [12]. Performance
of OC in co-operative relay systems like amplify and forward (AF) and decode and forward (DF)
relays systems, in the presence of co-channel interference, has also been recently well studied
[13]–[16].
Many practical scenarios exist, where both the desired and interfering signals may have line-ofsight (LOS) paths. Such scenarios include indoor propagation, micro-cellular channels, satellite
channels, inter-vehicular communications, etc. Symbol error probability expressions (SEP) for
OC have been derived, when either the desired signals or the interfering signals undergo Rician
fading, while the other undergoes Rayleigh fading for arbitrary number of antennas and interferers
[2].
Existing 4G and emerging 5G systems are both interference-limited. Hence, receiver techniques
like OC, which mitigate interference, will play a key role in the performance analysis of these
systems. In this context, a characterization of OC receivers which takes into account practical
scenarios such as unequal interference power and correlation among interferers, is important.
Though OC has been well studied in literature, the focus has typically been on equal power
uncorrelated interferers. We believe ours is the first work to derive SER expressions for OC
considering a) mixture of Rayleigh and Rician faded interferers, b) unequal power interferers
(which occurs typically in most wireless systems) and c) correlated interferers (which occurs due
to correlated channel fading, shadowing and from spatial distribution of transmitters [17]–[19]).
The SER expressions are derived by determining expressions for the moment generating
function (MGF) of the SINR η. The derived SER expressions are functions of a double infinite
series. However, for evaluation, we truncate all infinite series to finite series with arbitrarily small
truncation error. The series terms are functions of Tricomi hypergeometric functions, which has
been used extensively in analyzing throughput and rate of MISO systems over various fading
channels [20]–[22]. We also derive expression for the moments of the SINR η. Our results are
compared with corresponding Monte-Carlo simulations and a close match is observed.
All our SER expressions and all other SER expressions in [2], [10] involve an explicit
evaluation of determinant. Hence, a simple approximation, which avoids determinant evaluation,
is also derived when the interferers are subjected to Rayleigh fading and this approximation holds
when the noise power σ 2 ≈ 0. We exploit the shifted factorial determinant and Vandermonde
September 18, 2017
DRAFT
3
determinant properties in [23] to derive this approximation.
II. S YSTEM M ODEL
Let NR denote the number of receive antennas, NI denote the number of interferers, c
denote the NR × 1 channel from the transmitter to the user, ci denote the NR × 1 channel
from the ith interferer to the user, x denote the desired user symbol belonging to unit energy
QAM constellation and xi denote the ith interferer symbol also belonging a unit energy QAM
constellation. The received vector is given by
y = cx +
NI
X
ci xi + n,
(1)
i=1
where n is the NR × 1 additive white complex Gaussian noise vector, with power of σ 2 per
dimension i.e., n ∼ CN (0, σ 2 INR ). The interferer channels are modeled as i.i.d Rician i.e.,
√
i
, b′ = κi1+1 , κi the ratio of the power of line of sight
ci ∼ CN ( a′ m′i , b′ INR ), where a′ = κiκ+1
component to the scattering component of the interferer signals and m′i is an NR × 1 arbitrary
vector with elements of unit magnitude. The user channel is also assumed to be i.i.d Rician i.e.,
√
s
, b = κs1+1 . Note that, the Rician parameter κs , is the
c ∼ CN ( am, bINR ), where a = κsκ+1
ratio of the power of line of sight component to that of the scattering component and m is an
NR × 1 mean vector with elements of unit magnitude and uniform phase. Let EI′′ = NI × EI′
denote the total energy of the interfering signals, where EI′ is the mean energy of each of the
interfering signals. The covariance matrix of the interference term plus the noise term is given
by
H
R = EI′ C′ C′ + σ 2 I = EI CCH + σ 2 I,
(2)
√
where C′ = [c1 , ..., cNI ], C′ ∼ CN ( a′ M′ , b′ INR ⊗ INI ). M′ is an arbitrary deterministic matrix
obtained by stacking m′i s, such that, M′ = [m′1 , m′2 , m′3 , ...., m′NI ] and tr(M′ M′ H ) = NR NI .
√ √
√
Here, C ∼ CN (M, INR ⊗ INI ), EI = EI′ × b′ and M = a′ / b′ M′ = κi M′ . The received
SINR for the OC is given by [8]
η = ED cH R−1 c,
(3)
where ED is the mean energy of the user signal. Using the standard assumption that the
contribution of the interference and the noise at the output of optimal combiner, for a fixed
September 18, 2017
DRAFT
4
η, can be well-approximated to be Gaussian as in [24] and [25] and references therein, the
probability of symbol error for an M-ary square QAM constellation is given by [26],
p
p
Pe ≈ k1 Q( k2 η) − k3 Q( k2 η)2 ,
(4)
k2
where k1 = 4 1 − √1M , k2 = M3−1 , k3 = 41 and the Q-function is given by Q(x) =
R ∞ −u2 /2
1
e
du. The assumption that the contribution of the interference and the noise at the
2π x
output of the optimum combiner, for a fixed η, is Gaussian, is valid even when the number of
interferers NI is small [25] and such a system model assumption is made in a number of papers
[8], [27], [28] to derive the SER expression. Using the approximation Q(x) ≈
one can write Pe as [29],
Pe ≈
where
a1 = k121 ,
a2 = k41 ,
3
a3 = −k
,
144
3
a4 = −k
,
16
5
X
2 2
1 − 12 x2
e
+ 14 e− 3 x ,
12
al e−bl η ,
(5)
l=1
3
a5 = −k
,
24
b1 = k22 , b2 = 2k32 , b3 =k2 , b4 = 4k32 and b5 = 7k62 . The
average SER obtained by averaging Pe over all channel realizations is,
5
5
X
X
al Eη [e−bl η ]
al e−bl η ] =
SER ≈ Eη [Pe ] = Eη [
l=1
l=1
=
5
X
al Mη (s)|s=−bl ,
(6)
l=1
where Mη (s) is the MGF of η and Eη denotes expectation over SINR η. Finding an expression
for SER in (6) requires determining an expression for the MGF of SINR η. This is derived in
the following section.
III. SER
EXPRESSIONS FOR EQUAL POWER UNCORRELATED INTERFERERS
A general expression for the MGF Mη (s) of SINR η, can now be obtained from Theorem 1 of
[2]. We now further simplify for the specific case of Rician distribution. Let n2 = max(NR , NI )
and n1 = min(NR , NI ).
Mη (s) = (−1)NR (σ 2 /EI )(NR −n1 ) EΛR
"
n1
Y
σ 2 /EI + λi
i=1
(NR −n1 )
λi
!
#
det(J)
,
Vn1 (ΛR )
(7)
where Vn1 (ΛR ) is the determinant of the Vandermonde matrix formed by eigen values of noncentral Wishart matrix CCH . J is an n1 × n1 matrix with elements,
h1 (s, λi ) − PNR −n1 ht (s, 0)λt−1 , j = 1,
i
t=1
Ji,j =
λNR −j ,
j = 2, ..., n1 ,
i
September 18, 2017
DRAFT
5
and
ht (s, x) =
aNR s
1 F1 (t; NR ; xEI /ED +σ2 /ED −bs )
,
(bsED /EI − σ 2 /EI − x)t
(8)
where 1 F1 (.) is the confluent hypergeometric function [30] with the series expansion of 1 F1 (.)
P
(a)k z k
given by 1 F1 (a; b; z) = ∞
z=0 (b)k k! .
SER is derived in [2] for two cases: a) Rician signal with Rayleigh interferers b) Rayleigh
signal with Rician interferers. For the Rician-Rayleigh case, (7) is used along with the eigen
value distribution of central Wishart matrix to arrive at a closed form expression for the MGF.
In the later case, the fact that user signal c exhibits Rayleigh fading and hence invariant under
unitary transformation is exploited to derive a closed form expression for the MGF.
To the best of our knowledge, there is no open literature that proves that Rician distribution
is invariant under unitary transformation. Therefore, for the case of Rician signals with Rician
interferers, we propose to evaluate the expectation in (7), by using the eigen value distribution of
non-central Wishart matrix and subsequently simplify it by using properties of hyper-geometric
functions.
The joint pdf of ordered eigen values (λ1 > λ2 > ... > λn1 ) of non-central Wishart matrix
is given by [31],
f (λ1 , ..., λs ) = c1 |Υ|
n1
n1
Y
Y
(λi − λj )
λnk 2 −n1 e−λk ,
i<j
(9)
k=1
where Υ is a n1 × n1 matrix whose (i, j)th entry ∀i = 1, ..., n1 is given by,
0 F1 (n2 − n1 + 1; wj λi ), j = 1, ..., L,
Υi,j =
λn1 −j (n2 −n1 )! ,
j = L + 1, ..., n1 ,
i
(n2 −j)!
and
e−tr(Ω) ((n2 − n1 )!)−n1
c1 = Qn1 −L
.
QL
n1 −L QL
i=1 (n1 − L − i)!
i=1 wi
i<j (wi − wj )
Note that wi s are the ordered L non-zero eigen-values of the non-centrality matrix, Ω = MH M
P
zk
and the series expansion of hypergeometric function 0 F1 (.) is given by 0 F1 (b; z) = ∞
k=0 (b)k k! .
September 18, 2017
DRAFT
6
Substituting (9) in (7),
Mη (s) = c1 (−1)NR (σ 2 /EI )(NR −n1 )
∞
Z
0
×
"
n1
Y
σ 2 /EI + λi
i=1
(NR −n1 )
λi
n1
n1
Y
Y
(λi − λj )
λnk 2 −n1 e−λk dλ1 ...dλn1
i<j
!
#
det(J)
|Υ|
Vn1 (ΛR )
k=1
= (−1)NR (σ 2 /EI )(NR −n1 )
Z
0
∞
c1
n1
Y
i=1
!
(σ 2 /EI + λi ) × e−λi |J||Υ|dλ1...dλn1 .
(10)
From Theorem 2 in Appendix of [32], it can be observed that, for two arbitrary m × m matrices
φ(y) and ψ(y) with ij th element φi (yj ) and ψi (yj ) and arbitrary function g(.), where y =
[y1 , y2 , ...ym ]T , the following identity holds:
Z b
Z
Z
m
Y
g(yk )dy1dy2 ...dym = det
{φi (y)ψj (y)g(y)dy}i,j=1,..,m .
...
|φ(y)||ψ(y)|
b≥yi ≥a
a
k=1
Using the above property to simplify MGF in (10), we get,
Mη (s) = c|N|,
where
Ni,j =
(11)
e−tr(Ω) ((n2 − n1 )!)−n1 (−1)NR (σ 2 /EI )(NR −n1 )
,
c = Qn1 −L
QL
n1 −L QL
w
(w
−
w
)
(n
−
L
−
i)!
1
i
j
i=1 i
i=1
i<j
R ∞ σ2
PNR −n1
−x n2 −NR
(
+
x)e
x
F
(n
−
n
+
1;
w
x)[h
(s,
x)
−
ht (s, 0)xt−1 ]dx,
0
1
2
1
i
1
t=1
0
EI
j = 1, i = 1, ..., L,
R
∞ ( σ2 + x)e−x xn2 −NR xn1 −i (n2 −n1 )! [h1 (s, x) − PNR −n1 ht (s, 0)xt−1 ]dx,
t=1
0
EI
(n2 −i)!
j = 1, i = L + 1, ..., n1 ,
R ∞ σ2
( EI + x)e−x xn2 −NR 0 F1 (n2 − n1 + 1; wi x)xNR −j dx, j = 2, ..., n1 i = 1, ..., L,
0
R
∞ ( σ2 + x)e−x xn2 −NR xn1 −i (n2 −n1 )! xNR −j dx, j = 2, ..., n1 i = L + 1, ..., n1 .
0
EI
(n2 −i)!
Further simplification of Ni,j is given in Appendix A. The final expression for the entries Ni,j
is given in (12) and this can be substituted in (11) to obtain the final MGF of SINR and the
SER can then be obtained using (6). In (12) p = σ 2 /EI , q = n2 − n1 + 1, u = aNR sED /EI ,
v = bsED /EI − σ 2 /EI . Note that the expression for Ni,j , for L = n1 or L = 0, gets significantly
simplified.
a) L = n1 : If the number of eigen values is L = n1 corresponding to all the interferers being
Rician faded, we get Mη (s) = c|NL=n1 |, where c =
September 18, 2017
e−tr(Ω) ((n2 −n1 )!)−n1
Qn1
(−1)NR (σ 2 /EI )(NR −n1 )
i<j (wi −wj )
DRAFT
7
and the value of |NL=n1 | is evaluated using (12) at L = n1 .
b) L = 0: If the number of non-zero eigen values is zero i.e., L = 0 corresponding to all the
interferers being Rayleigh faded, we get Mη (s) = c|NL=0 |, where
1
Qn1
(−1)NR (σ 2 /EI )(NR −n1 )
i=1 (n1 − i)!
i=1 (n2 − i)!
c = Q n1
and the value of |NL=0 | is evaluated using (13). Also, when we do a Laplace expansion of the
determinant |NL=0| given in (13) along the first column and substitute ζt (k) =
σ2
Γ(n1
EI
+ n2 −
NR + t − k + 1) + Γ(n1 + n2 − NR + t − k + 2), we observe that the expression for Mη (s) for
the Rayleigh interferers case is the same as the one obtained in [2, Eq. 13].
A. SER for interference limited scenario
Recently, there has been a lot of interest in characterizing the performance of cellular networks/
wireless system in an interference limited scenario. Throughput and rate have been studied
in [33]–[37] and references therein, assuming noise can be neglected (i.e., σ 2 = 0), in an
interference limited scenario. Motivated by these works, we now derive SER for the interference
limited scenario and show that the SER expression can be substantially simplified in the case of
Rayleigh interferers. Note that an interference-limited scenario is possible only for NI > NR .
For NR > NI and σ 2 = 0, the receive antennas can cancel every interfering signal.
When σ 2 is neglected, Nij can be approximated as in (14) and can be substituted in (11)
to obtain the final MGF of SINR. The SER can then be obtained using (6). If the number of
non-zero eigen values L = 0 as is the case for Rayleigh fading, then
Mη (s) = c|Nσ2 =0,L=0 |,
where
((n2 − n1 )!)−n1
(−1)NR (σ 2 /EI )(NR −n1 ) ,
c = Qn1
i=1 (n1 − i)!
(15)
(16)
and the value of |Nσ2 =0,L=0 | is evaluated using (14) at L = 0. Note that in c, we do not neglect
σ 2 . The expression can be further simplified as shown in Appendix C to obtain,
Qn1
n1
(−1)NR +1 (σ 2 /EI )(NR −n1 ) n2 ! X
j=1 (j − 1)!
i+1
Qn1
(−1) A(i)
Mη (s) =
,
(n2 − n1 )! i=1 (n1 − i)! i=1
(n1 − i)!(i − 1)!
where A(i) is given in Appendix C. This can be further substituted in (6), to obtain the expression
for SER as,
SER ≈
5
X
September 18, 2017
l=1
Qn1
n1
(−1)NR +1 (σ 2 /EI )(NR −n1 ) n2 ! X
j=1 (j − 1)!
i+1
Qn1
al
(−1) A(i)
|s=−bl .
(n2 − n1 )! i=1 (n1 − i)! i=1
(n1 − i)!(i − 1)!
(17)
DRAFT
8
Ni,j
"
k
P
PT1 ul
w
T2
i
k=0 qk k!
l=0 (NR )l − Γ(k + n2 − NR + 2)U(l + 1, l − k − n2 + NR , −v)
#
−pΓ(k + n2 − NR + 1)U(l + 1, −k − n2 + NR + l + 1, −v)
PNR −n1 1 F1 (t;NR ;u/v)
× pΓ(t + n2 − NR ) 1 F1 (t + n2 − NR ; q; wi )
−
t=1
vt
+ Γ(t + n2 − NR + 1)1 F1 (t + n2 − NR + 1; q; wi ) ,
j = 1, i = 1, ..., L,
"
(n2 −n1 )! PT1
ul
l=0 (NR )l − Γ(n2 − NR + n1 − i + 2)
(n2 −i)!
U(l + 1, l − n2 + NR − n1 + i, −v)
=
−pΓ(n2 − NR + n1 − i + 1)U(l + 1, −n2 + NR − n1 + i + l + 1, −v)
PNR −n1 1 F1 (t;NR ;u/v)
pΓ(t + n2 + n1 − NR − i)
t=1
vt
−
#
+ Γ(t + n2 + n1 − NR − i + 1) ,
j = 1, i = L + 1, ..., n1 ,
p 1 F1 (n2 − j + 1; q; wi )Γ(n2 − j + 1) + Γ(n2 − j + 2)1 F1 (n2 − j + 2; q; wi),
j = 2, ..., n1 i = 1, ..., L.
(n2 −n1 )!
[pΓ(n2 + n1 − i − j + 1) + Γ(n2 + n1 − i − j + 2)],
(n2 −i)!
j = 2, ..., n1 i = L + 1, ..., n1 .
Ni,j |L=0
(12)
"
PT1 ul
l=0 (NR )l − Γ(n2 − NR + n1 − i + 2)U(l + 1, l − n2 + NR − n1 + i, −v)
−pΓ(n2 − NR + n1 − i + 1)U(l + 1, −n2 + NR − n1 + i + l + 1, −v)
PNR −n1 1 F1 (t;NR ;u/v)
= −
pΓ(t + n2 + n1 − NR − i)
t=1
vt
#
+Γ(t + n2 + n1 − NR − i + 1) ,
j = 1, i = 1, ..., n1 ,
pΓ(n + n − i − j + 1) + Γ(n + n − i − j + 2), j = 2, ..., n , i = 1, ..., n .
2
1
2
1
1
1
(13)
September 18, 2017
DRAFT
9
Ni,j |σ2 =0
"
k
P
PT1 (aNR sED /EI )l
wi
T2
− Γ(k + n2 − NR + 2)
k=0 (n2 −n1 +1)k k!
l=0
(NR )l
#
U(l + 1, l − k − n2 + NR , −bsED /EI )
PNR −n1 1 F1 (t;NR ;aNR sED /(−bsED )
Γ(t + n2 − NR + 1)
−
t=1
(bsED /EI )t
× 1 F1 (t + n2 − NR + 1; n2 − n1 + 1; wi ),
j = 1, i = 1, ..., L,
"
(n2 −n1 )! PT1 (aNR sED /EI )l − Γ(n − N + n − i + 2)
2
R
1
l=0
(n2 −i)!
(NR )l
=
U(l + 1, l − n2 + NR − n1 + i, −bsED /EI )
PNR −n1 1 F1 (t;NR ;aNR sED /(−bsED )
−
t=1
(bsED /EI )t
#
× Γ(t + n2 + n1 − NR − i + 1) , j = 1, i = L + 1, ..., n1 ,
Γ(n2 − j + 2)1 F1 (n2 − j + 2; n2 − n1 + 1; wi ),
j = 2, ..., n1 i = 1, ..., L.
(n2 −n1 )! [Γ(n + n − i − j + 2)],
j = 2, ..., n1 , i = L + 1, ..., n1 .
2
1
(n2 −i)!
(14)
For an interference limited scenario (NI > NR and σ 2 = 0), by substituting n1 = NR and
n2 = NI , the SER becomes,
SER ≈
5
X
l=1
QNR
NR
X
(−1)NR +1 NI !
j=1 (j − 1)!
i+1
al
|s=−bl .
(−1) A(i)
QNR
(NR − i)!(i − 1)!
(NI − NR )! i=1 (NR − i)! i=1
(18)
Ours is the first work to obtain SER expression in an interference limited scenario, for Rayleigh
faded interferers in a closed form. All existing work, so far, require an explicit evaluation of the
determinant. Further, the expression derived also gives an approximation of SER, for NI > NR
for very low noise values σ 2 ≈ 0. Note that the dependence of c term on σ 2 is not present for
NI > NR . On the other hand, the σ 2 term exists in c term for NR > NI . We derive in the
succeeding subsection, an SER approximation for NR > NI , by substituting σ 2 = 0 only in the
determinant.
SER approximation for NR > NI : If we substitute n1 = NI and n2 = NR in (11) and ignore
the σ 2 term inside the determinant, we also obtain an approximation for the SER as,
QNI
NI
5
X
(−1)NR +1 (σ 2 /EI )(NR −NI ) NR ! X
j=1 (j − 1)!
i+1
(−1) A(i)
|s=−bl .
SER ≈
al
QNI
(N
−
i)!(i
−
1)!
(N
−
i)!
(N
−
N
)!
I
I
R
I
i=1
i=1
l=1
September 18, 2017
(19)
DRAFT
10
Note that SER expressions obtained for Rayleigh interferers in [2] involve not only an explicit
evaluation of determinants but also numerical integration, while results here require neither. This
approximation works very well when σ 2 is actually small or when interferer powers are large
compared to σ 2 .
B. Moments of Rician-Rician SINR
Similar to [2], we can obtain the moments of the SINR η. The lth moment of SINR for Rician
faded user and Rician faded interferers is given by,
µRic−Ric
l
where c =
n1
X
dl
Ric
(−1)k |Yk |dl = αlRic µRay−Ric
= l Mη (s)|s=0 = αl c
l
ds
k=1
e−tr(Ω) ((n2 −n1 )!)−n1
Qn1
(−1)NR (σ 2 /EI )(NR −n1 ) ,
i<j (wi −wj )
th
st
by omitting the k
row and 1
dl is given by (51), Yk is the matrix formed
column of the matrix Yi,j =
σ2
F (n2
EI 1 1
1; wk )Γ(n2 −j +1)+Γ(n2 −j +2)1 F1 (n2 −j +2; n2 −n1 +1; wi ) and αlRic
− j + 1; n2 − n1 +
R /b)k
P
.
= bl lk=0 kl (aN
(NR )k
The derivation is given in Appendix D. From the above relation it is clear that the remarks 1-5
in [2] hold true, irrespective of whether the interferers undergo Rician or Rayleigh fading.
IV. SER
EXPRESSIONS FOR CORRELATED INTERFERERS AND UNEQUAL POWER
INTERFERERS
In the previous section, SER expressions are derived for the case of equal power uncorrelated
interferers. But, in practice, the interferers can have different power and/or can be correlated.
In a practical cellular systems, there can be one or more of the following: a) Receiver side
correlation, b) Interferer correlation, c) Unequal power interferers.
The general non-central Wishart matrix W is written as W = C′ C′ H , where C′ ∼ CN (M, Σ⊗
Ψ). Here, the NR × NR matrix Σ denotes the receive correlation and the NI × NI matrix Ψ
denotes the transmit correlation or interferer correlation in our case.
Suppose, we consider only receive side correlation and assume that the interferer correlation is
not present, i.e., Ψ is an identity matrix. This reduces to the non-central Wishart matrix denoted
by W ∼ CW(NI , Σ, Σ−1MMH ). This case, where Ψ is assumed to be an identity matrix, is
widely discussed in literature. The eigen-value distribution of this case, i.e., a non-central Wishart
matrix with a covariance matrix Σ which is not an identity matrix, is analyzed in [38] in terms of
zonal polynomials. However, using this eigen-value distribution to obtain the MGF expression of
September 18, 2017
DRAFT
11
η and hence derive the SER expressions becomes mathematically intractable. Hence, considering
receive correlation is beyond the scope of this work.
On the other hand, the cases of Ψ being a diagonal matrix, i.e., unequal power interferers
or Ψ being a full matrix, i.e., correlated interferers, have barely received attention in statistic
literature. There do not even exist matrix variate and eigen-value distribution results for this case.
However, we do provide results for this case by considering the problem as two sub-problems
a) for NR ≥ NI exact results are provided, b) for NR < NI approximate results are provided.
In short, in this section we derive SER expressions for C′ ∼ CN (M, INR ⊗ Ψ).
1
Note W = C′ C′ H can be decomposed into W = CΨCH , such that C ∼ CN (MΨ− 2 , INR ⊗
INI ) [39]. Let us first consider the case of correlated interferers. The covariance matrix of the
interference term plus the noise term is given by
R = CΨCH + σ 2 I.
(20)
The received SINR for the OC is given by [8],
η = ED cH R−1 c,
(21)
where ED is the mean energy of the user signal. We will consider this problem as two cases:
a) NR ≥ NI , b) NR < NI .
A. NR ≥ NI
Like in the case of equal power uncorrelated interferers, we consider the following general
expression for the MGF Mη (s) of SINR η, from [2],
#
!
" n
1
2
Y
det(J)
σ
+
λ
i
,
Mη (s) = (−1)NR (σ 2 )(NR −n1 ) EΛR
(NR −n1 )
Vn1 (ΛR )
i=1 λi
(22)
where Vn1 (ΛR ) is the determinant of the Vandermonde matrix formed by eigen values of noncentral Wishart matrix CΨCH . J is an n1 × n1 matrix with elements,
h1 (s, λi ) − PNR −n1 ht (s, 0)λt−1 , j = 1,
i
t=1
Ji,j =
λNR −j ,
j = 2, ..., n1 ,
i
and
ht (s, x) =
September 18, 2017
aNR s
1 F1 (t; NR ; x/ED +σ2 /ED −bs )
.
(bsED − σ 2 − x)t
(23)
DRAFT
12
Recall that, in the case of equal power uncorrelated interferers, we used the above MGF
expression of η and simplified the expression using the eigen value distribution of the noncentral Wishart matrix. But for the case of correlated interferers, there exists no matrix variate
distribution forumla in the open literature and deriving one requires an integration over the Steifel
manifold [38]. Also, there exists no eigen value distribution for this case. Hence, initially, we
consider the case of Rayleigh-faded interferers, i.e., M = 0.
1) Rayleigh faded correlated interferers: We exploit the property that W = CΨCH has the
1
1
same non-zero eigen values as that of Ψ 2 CH CΨ 2 , where CH C is also a Wishart matrix. The
1
1
eigen value distribution of Ψ 2 CH CΨ 2 , is given by [40],
NI
NI
Y
Y
R −NI
f (λ1 , ..., λs ) = c1 |Υ| (λi − λj )
λN
,
k
i<j
(24)
k=1
where Υ is a NI × NI matrix whose (i, j)th entry ∀i, j = 1, ..., NI is given by,
λ
− ri
Υi,j = e
and
1
c1 = (−1) 2 NI (NI −1) QNI
1
i<j ( ri
j
|Ψ|−NR
.
Q I
− r1j ) N
k=1 (NR − k)!
Note that ri s are the ordered NI distinct non-zero eigen-values of Ψ. Further simplification of
the MGF using (24) is given in Appendix E. The MGF after simplification becomes,
Mη (s) = c|N|,
1
c = (−1)NR (σ 2 )(NR −NI ) (−1) 2 NI (NI −1) QNI
1
i<j ( ri
(25)
|Ψ|−NR
,
Q I
− r1j ) N
k=1 (NR − k)!
and N is given by
P∞ (aNR sED )l 2
(σ |bsED − σ 2 |−l U(1, 1 − l, ri |bsED − σ 2 |)
l=0
(NR )l
+|bsED − σ 2 |−l+1U(2, 2 − l, ri |bsED − σ 2 |))
Ni,j =
aNR s
)
PNR −NI 1 F1 (t;NR ; σ2 /E
D −bs
(σ 2 rit Γ(t) + rit+1 Γ(t + 1)], j = 1, i = 1, ..., NI ,
−
2
t
t=1
(bsED −σ )
σ 2 r NR −j+1 Γ(NR − j + 1) + r NR −j+2 Γ(NR − j + 2), j = 2, ..., NI i = 1, ..., NI .
i
i
(26)
where ri , 1 ≤ i ≤ NI are the eigen values of Ψ. We can truncate the converging infinite series
for j = 1 at a finite value, with an arbitrarily small truncation error. The convergence proof is
similar to the one given in Appendix A. Recall that the the SER can then be obtained substituting
(25) in (6).
September 18, 2017
DRAFT
13
Note that this is an exact MGF expression and is novel for the case of Rayleigh faded
correlated interferers with Rician faded users. Earlier works like [2], considers only equal power
uncorrelated interferers, while recent works like [10] consider only Rayleigh faded user. An
approximation which works for σ 2 ≈ 0 is also derived in Appendix E. The expression is as
follows:
Mη (s) ≈ c
where,
NI
X
i=1
(−1)i+1 A(i)ri−NR +NI −2 |V i (r)|,
1
c = (−1)NR (σ 2 )(NR −NI ) (−1) 2 NI (NI −1) QNI
1
i<j ( ri
A(i) =
∞
X
(aNR sED )l
l=0
−
(NR )l
NX
R −NI
t=1
(27)
|Ψ|−NR
,
Q I
− r1j ) N
k=1 (NR − k)!
(|bsED |−l+1 U(2, 2 − l, ri |bsED |))
aNR s
1 F1 (t; NR ; −bs ) t+1
ri Γ(t
(bsED )t
+ 1),
and V i (r) denotes the Vandermonde matrix formed from all elements of r = (r1 , r2 , .., rNI )
except the ith element. This can be further substituted in (6), to obtain the approximate expression
for SER as,
SER ≈
5
X
l=1
1
al (−1)NR (σ 2 )(NR −NI ) (−1) 2 NI (NI −1) QNI
×
1
i<j ( ri
NI
X
i=1
(−1)i+1 A(i)s=−bl ri−NR +NI −2 |V i (r)|.
|Ψ|−NR
Q I
− r1j ) N
k=1 (NR − k)!
(28)
2) Rician faded correlated interferers: For the case of Rician faded interferers, the above
approach is not possible, even for NR ≥ NI . This is because, we have to use the zonal-polynomial
based eigen value distribution from [38] to simplify the MGF, which is not mathematically
tractable. Nevertheless, we arrive at a mathematically tractable solution, wherein we propose
to approximate the non-central Wishart matrix by a central Wishart matrix through moment
matching and then use the derived expression given in (25).
Expected value of any matrix of the form AH A, if A ∼ CN (M, INI ⊗ INR ), with NR degrees
1
1
of freedom, is given by E[AH A] = NR + MH M. This implies that for W = Ψ 2 CH CΨ 2 ,
the first moment is given by, E[W] = NR Ψ + Ψ1/2 MH MΨ1/2 . The first moment of any
September 18, 2017
DRAFT
14
central Wishart matrix W1 ∼ CW(NR , Φ) with the same degree of freedom NR , is given by
E[W1 ] = NR Φ. When the first moments of the W and W1 are equated, we obtain
NR Φ = NR Ψ + Ψ1/2 MH MΨ1/2
1 1/2 H
Ψ M MΨ1/2 .
NR
We have now obtained a central Wishart approximation of the non-central Wishart matrix. Hence,
Φ=Ψ+
W can be approximated as a central Wishart CW(NR , Ψ +
1
Ψ1/2 MH MΨ1/2 ).
NR
A similar
approximation is performed in [41]. Now that we have a central Wishart matrix, the expressions
derived for the case of Rayleigh faded interferers holds, but with the matrix Ψ in (25) replaced
by Ψ +
1
Ψ1/2 MH MΨ1/2 .
NR
The simplified MGF expression is now given by
Mη (s) ≈ c|N|,
|Ψ + N1R Ψ1/2 MH MΨ1/2 |−NR
1
NR
2 (NR −NI )
N
(N
−1)
I
I
c = (−1) (σ )
(−1) 2
QNI 1
QNI
1
i<j ( ri − rj )
k=1 (NR − k)!
(29)
and N is given by
P∞ (aNR sED )l 2
(σ |bsED − σ 2 |−l U(1, 1 − l, ri |bsED − σ 2 |)
l=0
(NR )l
+|bsED − σ 2 |−l+1U(2, 2 − l, ri |bsED − σ 2 |))
Ni,j =
aNR s
)
PNR −NI 1 F1 (t;NR ; σ2 /E
D −bs
(σ 2 rit Γ(t) + rit+1 Γ(t + 1)], j = 1, i = 1, ..., NI ,
− t=1
2 )t
(bsE
−σ
D
σ 2 r NR −j+1 Γ(NR − j + 1) + r NR −j+2 Γ(NR − j + 2), j = 2, ..., NI i = 1, ..., NI ,
i
i
(30)
where ri 1 ≤ i ≤ NI are the eigen values of Ψ +
1
Ψ1/2 MH MΨ1/2 .
NR
The SER approximation
given by (28), can also be used, since we are anyway approximating non-central Wishart matrix
by central Wishart matrix. But the tightness of the SER approximation given by (28), now also
depends on how good is the non-central Wishart to central Wishart matrix for the specific case.
3) Unequal power interferers: All the above analysis holds for a general Ψ. For the case of
unequal power interferers, Ψ is just a diagonal matrix, with the interferer powers occupying the
diagonal. Hence, the MGF expressions (25) and (29) can be used for unequal power Rayleighfaded and Rician-faded interferers respectively.
B. NI > NR
1) Rayleigh faded correlated interferers: For the case of Rayleigh-faded correlated interferers,
for NI > NR , the covariance matrix of the interference term plus the noise term is given by
R = CΨCH + σ 2 I.
September 18, 2017
(31)
DRAFT
15
Here, C ∼ CN (0, INR ⊗ INI ). From [42], the distribution of W = CΨCH is same as that of
PNI
i=1 λi Wi , where λi are the eigen values of Ψ and Wi ∼ CW(1, INR ). Though this method
works for NR ≥ NI , we use the accurate analysis given in the previous subsection for calculating
SER.
From [43], sum of central Wishart matrices can be approximated by another central
" Wishart
#
matrix. In our case, from [43], W ≈ S
PNI
i=1
ps
λi
, where S ∼ CW(ps , INR ) and ps =
PNI
( i=1
λi )2
PNI 2
i=1 λi
rounded to the nearest integer. Note that, this has reduced to a case of a Wishart matrix with
an identity covariance matrix. Hence, the MGF expression derived for the case of equal power
Rayleigh interferers, i.e., expressions corresponding to L = 0 given in Section III, can now
be used. Also, the determinant simplification that has been derived in the case of equal power
Rayleigh faded interferers holds for this case.
2) Rician faded interferers for NI > NR : In case, of correlated or unequal power Rician
faded interferers for NI > NR , it is mathematically intractable to give an MGF expression
and hence derive an SER expression. Nevertheless, the existing SER expressions in Section III
derived for the case of equal power uncorrelated interferers can be used as an upper bound. If
we consider all the interferers to have the same power as that of the maximum-power interferer,
our expression gives an upper bound for the actual SER, i.e., our expressions give the worst
case SER. Similarly, the expressions for uncorrelated case gives the worst case SER, i.e., a
good upper bound on for the actual SER of correlated interferers. This is because, correlated
interferers cause partial interference alignment [10] and hence the receive antennas can cancel
the interferers better, leading to lower SER when compared to the uncorrelated case.
V. N UMERICAL R ESULTS
The derived SER expressions are first verified using Monte-Carlo simulations, for both NR >
NI and NR < NI , for the case of equal power uncorrelated interferers. The total interference
power is denoted as EI′′ , from which average interference power per interferer is obtained as
EI′ = EI′′ /NI . The mean energy of the received signal, ED , is taken to be unity without loss of
generality. Signal to interference noise ratio (SIR) is given by ED /EI′′ . The determinant of N
matrix whose entries are given by (12) is determined with the infinite summation truncated to
T1 = T2 = 100. For T1 = T2 = 100, the truncation error has negligible magnitude, as evinced by
our simulations. By substituting the determinant of N matrix in (11), the MGF is evaluated for
September 18, 2017
DRAFT
16
s = bl , ∀l = 1 to 5. These MGF values are substituted in (6) and theoretical SER is calculated
for values of signal to noise ratio (SNR) in the range 5 dB to 25 dB.
For the Monte-Carlo simulation, the deterministic matrix M′ is first obtained with unit magnitude and uniform phase, satisfying the condition tr(M′ M′H ) = NR NI and the M matrix is
√
then obtained from M′ using the relation M = κi M′ . This matrix is fixed during a set of
simulations. For each Monte-Carlo simulation, the random matrix C ∼ CN (M, INR ) is generated
and the random covariance matrix R of the interference terms plus the noise term is calculated
using (2). The i.i.d Rician user channel c is then generated and using (3), SINR η can be
calculated. Finally, the simulated SER is obtained by substituting the value of SINR η in (4)
and evaluating it. The number of Monte-Carlo simulations for each SNR value is of the order
of 107 . This is repeated for various combinations of parameters NR , NI , EI′′ , κs , and κi and
numerical results obtained are compared with the simulation for each case.
For the case of equal power and uncorrelated interferers, we see that there is a close match
between the theoretical and simulated SER as seen from Fig. 1 and Fig. 1b. The exponential
approximation Q(x) ≈
2 2
1 − 12 x2
e
+ 41 e− 3 x ,
12
from [29] provides a very tight upper bound for values
of x > 0.5 and the bound becomes tighter as x increases. Since, for NI > NR , the average SINR
is much lower when compared to the case NR > NI , we can observe a small mismatch between
the theoretical and the simulated SER in Fig. 1 (b). Also, the SER approximation 1 plot in Fig.
1, calculated by using the Nij matrix whose entries are given by (14), is tight beyond 15 dB. For
Rayleigh interferers, the SER approximation computed using (19), match with the simulation
results at high SNR, for NR > NI as seen from Fig. 2(a). We can also see from Fig. 2(a) that,
when the interference power dominates the noise power, as is the case when EI′′ = −2 dB, the
SER approximation is tight even at 10 dB SNR. For NI > NR , the high SNR approximation
computed using (18) is compared with Monte-Carlo simulations in Fig. 2 (b) for an SNR of
20 dB and a series of SIR values. An excellent match is observed in an interference limited
scenario.
Similar Monte-Carlo simulations are performed for the case of correlated and/or unequal power
interferers. The difference is that, random covariance matrix R of the interference terms plus the
noise term is now calculated using (20). Also note that the determinant of N matrix whose entries
are given by (26) and (30), for Rayleigh and Rician faded interferers respectively, is determined
with the infinite summation truncated to T1 = 400. From Fig. 3 (a), we can observe that for
unequal power Rayleigh faded interferers, SER computed by means of (25) and (26) matches
September 18, 2017
DRAFT
17
10
10
0
10 0
Simulation
Theoretical
SER approximation 1
-1
16 QAM
-3
10
-4
10
-5
10
-1
16 QAM
SER
SER
10 -2
10
Simulation
Theoretical
SER approximation 1
4 QAM
4 QAM
10
-2
10
-3
10 -6
10
-7
10
-8
5
10
15
20
25
5
SNR in dB
10
15
20
25
SNR in dB
(a)
(b)
NR = 8, NI = 6, EI′′ = −1dB, κs = 3, κI = 6
NR = 2, NI = 3, EI′′ = −8dB, κs = 2, κI = 3
Fig. 1: SER for Rician interferers
the simulated values. We can also see that the approximation for σ 2 ≈ 0 computed using (28)
gives good match with the simulated results for higher values of SNR. For the case of correlated
interferers, we consider exponential correlation between interferers [10], i.e., Ψ(i, j) = ρ|i−j| and
0 ≤ ρ ≤ 1, where Ψ is the interferer covariance matrix. A good match between the theoretical
and simulation results, is observed for the case of correlated Rayleigh faded interferers, in Fig.
3 (b).
In Fig. 4 (a), we studied the case of mix of Rayleigh and Rician faded unequal power
interferers. The theoretical SER computed by means of (29) and (30) match the simulation
results. Also, the SER approximation given by (28) gives a good match to the simulated SER
for high SNR values. For the case of only Rician faded interferers, the simulated value matches
the theoretical value for some cases as seen in Fig.4(b) and doesn’t perform very well for
some cases as seen from Fig.5 (a). This is so because, the expressions (29) and (30) used for
Rician faded interferers are obtained using moment matching approximation. For Rayleigh faded
interferers with unequal power and NI > NR , the approximation in Section IV.B, gives a fairly
good match to the theoretical values as seen from Fig.5 (b), as long as the interferer powers do
not vary widely.
VI. C ONCLUSIONS
In this paper, we considered receive diversity systems in the presence of multiple equal power
correlated interferers. Approximate SER expressions have been derived for optimum combining
September 18, 2017
DRAFT
18
10 0
10 0
Simulation
Theoretical
Simulation
Theoretical
SER approximation
10
16 QAM
-2
16 QAM, E ''=-6 dB
10 -1
I
10
-4
4 QAM,
4 QAM, E ''=-2dB
SER
SER
I
10
-2
10 -6
10
10
10 -3
-8
4 QAM, E ''=-6 dB
I
-10
5
10
15
20
10
25
-4
0
2
4
SNR in dB
6
8
10
12
SIR in dB
(a) NR = 6, NI = 4, κs = 3
(b) Interference dominated scenario
NR = 2, NI = 4, κs = 4
Fig. 2: SER for Rayleigh interferers
10 0
10
-1
10
-2
10 1
Simulation
Theoretical
SER approximation
10
-1
10
-2
16 QAM
SER
SER
16 QAM
10 -3
10 -3
4 QAM
10
Simulation
Theoretical
SER approximation
10 0
4 QAM
-4
10 -4
10 -5
10 -6
10
5
10
15
20
25
SNR in dB
-5
10 -6
5
10
15
20
25
SNR in dB
(a) unequal power Rayleigh interferers κs =
(b) correlated Rayleigh interferers
2, NR = 4, NI = 3, EI′′ = −5, −8, −10dB
κs = 3, NR = 4, NI = 2, EI′′ = −1dB, ρ = 0.6
Fig. 3: SER for correlated/unequal power Rayleigh interferers
system for the Rician-Rician case, i.e., when the desired and interfering signals is subjected
to Rician fading and interferers can be a mix of Rayleigh and Rician faded signals. SER is
also derived for an interference limited scenario and the expressions obtained are significantly
simpler than existing expressions. Both, interferer correlation and unequal power interferers
are also considered. The Monte-Carlo simulation closely match the derived results. We believe
extending this analysis to take into account receiver side correlation may be an interesting future
work.
September 18, 2017
DRAFT
19
10 0
10
10 0
Simulation
Theoretical
SER approximation
-1
10
-1
10
-2
10
-3
16 QAM
16 QAM
-2
SER
SER
10
Simulation
Theoretical
SER approximation
10 -3
4 QAM
10 -4
10
-4
4 QAM
10 -5
10
-6
5
10
15
20
25
10
-5
10
-6
10
-7
5
10
SNR in dB
15
20
25
SNR in dB
(a) mixture of two Rician and two Rayleigh
(b) unequal power Rician interferers κs =
interferers κs = 2, κi = 5, NR = 5, NI =
3, κi = 2, NR = 4, NI = 2, EI′′ = −3, −10dB
4, EI′′ = −1, −1, 5, −2, −2.5dB
Fig. 4: SER for correlated/unequal power Rayleigh interferers
10 0
10 0
Simulation
Theoretical
10 -1
Simulation
Theoretical
high SNR approximation
16 QAM
16 QAM
10 -1
-2
SER
SER
10
10 -3
10 -2
4 QAM
10
4 QAM
-4
10 -3
10 -5
10 -6
5
10
15
20
10 -4
25
5
10
SNR in dB
15
20
25
SNR in dB
(a) unequal power Rician interferers
(b) unequal power Rayleigh interferers
κs = 5, κi = 4, NR = 5, NI = 4, EI′′ =
κs = 2, NR = 3, NI = 5, EI′′ =
−2, −5, −1, −9dB
−6, −9, −10, −10, −10dB
Fig. 5: SER for unequal power interferers
A PPENDIX A
S IMPLIFICATION
OF
Ni,j
We will simplify Nij substantially and obtain a series expansion by exploiting various properties of special functions. We first substitute the value of h(t, x) from (8) for Nij j = 1 and
September 18, 2017
DRAFT
20
then use the following identities [44], for p < q and Re(s) > 0,
Z ∞
e−x xs−1 p Fq (a1 , .., ap ; b1 , .., bq ; ax)dx
=
Γ(s) p+1Fq (s, a1 , .., ap ; b1 , .., bq ; a), (32)
0
and
Z
∞
e−x xs−1 dx = Γ(s),
(33)
0
to solve the integrals in Ni,j entries for j = 2, .., n1 in (11). For j = 1, i = 1, ...L and
i = L + 1, ..., n1 , the integrals to be solved are of the form,
Z ∞
u
p + x −x z
e x (0 F1 (q, wi x))(1 F1 (1; NR ;
))dx,
I=
v−x
x−v
0
(34)
where p = σ 2 /EI , q = n2 − n1 + 1, u = aNR sED /EI , v = bsED /EI − σ 2 /EI and z is a positive
integer greater than zero. To obtain a solution for I, we substitute the series expansion for 0 F1
and 1 F1 , and interchange summations and integration. The integral to be solved becomes,
X
Z ∞
∞
∞
X
ul (1)l
wik
p + x −x xk+z
e
dx .
(35)
I=
l
(q)
(N
v
−
x
(x
−
v)
k k!
R )l l! 0
l=0
k=0
The justification for the interchange of summations and integration is provided in Appendix B.
R ∞ p −x xk+z
R ∞ x −x xk+z
Let A1 = 0 v−x
e (x−v)l dx and A2 = 0 v−x
e (x−v)l dx. The Tricomi function or confluent
Hyper-geometric function of the second kind is given by [44],
Z ∞
1
U(α, γ, z) =
e−zt tα−1 (1 + t)γ−α−1 dt [Re(α)
Γ(α) 0
>
0, Re(z)
>
0]. (36)
In our case −v > 0 and k + z > 0. Hence using the above identity, A1 and A2 can be simplified
as,
A1 = −pΓ(k + z + 1)(−v)l−k−z U(k + z + 1, k + z + 1 − l, −v),
(37)
A2 = −Γ(k + z + 2)(−v)l−k−z−1U(k + z + 2, k + z + 2 − l, −v).
(38)
Further, using the functional identity U(a, b, z) = z 1−b U(a − b + 1, 2 − b, z) from [45], (37) and
(38) are simplified and substituted back in (35) to obtain,
" ∞
∞
X ul (1)l
X
wik
− Γ(k + z + 2)U(l + 1, l − k − z, −v)
I=
(q)k k! l=0 (NR )l l!
k=0
− pΓ(k + z + 1)U(l + 1, −k − z + l + 1, −v)
#
(39)
To the best of our knowledge, there are no known identities available in open literature that give
a closed form expression for the above double infinite summation.
September 18, 2017
DRAFT
21
Convergence of the infinite summations
To prove"the convergence of the above infinite summation,#first consider the summation I1 =
P∞ wik P∞ |u|l (1)l
l=0 (NR )l l! Γ(k + z + 2)U(l + 1, l − k − z, −v) . From Theorem 3 in [46] we get
k=0 (q)k k!
the identity U(a, b, x) < x−a for x > 0, a > 0 and a − b + 1 > 0. In our case, we can see that
a = l + 1 > 0 and a − b + 1 = k + z + 2 > 0 and x = −v > 0. Therefore,
∞
∞
X
wik X |u|l (1)l
I1 <
Γ(k + z + 2)(−v)l+1
(q)
(N
k k!
R )l l!
k=0
l=0
∞
∞
X
wik (z + 2)k X |uv|l (1)l
= (−v)Γ(z + 2)
(q)k k! l=0 (NR )l l!
k=0
= (−v)Γ(z + 2) 1 F1 (z + 2, q, wi ) 1 F1 (1, NR , |uv|)
The last equality is obtained from the series expansion definition of 1 F1 Hypergeometric function
[45]. A similar argument can be used to prove the absolute convergence of the other infinite
summation. Hence, I is convergent, which implies thatwe can truncate the double summation
P 1 PT2 ul wik
to T1 and T2 values such that I − Tl=0
k=0 (NR )l (q)k − (k + 1)U(l + 1, l − k, −v) − pU(l +
1, −k + l + 1, −v)
≤ ǫ for any ǫ > 0. Hence we approximate I by
IT1 ,T2
" T
T2
1
X
X
ul
wik
− Γ(k + z + 2)U(l + 1, l − k − z, −v)
=
(q)
(N
k k!
R )l
l=0
k=0
− pΓ(k + z + 1)U(l + 1, −k − z + l + 1, −v)
#
(40)
with arbitrarily low approximation error. Hence, the simplified Ni,j entry is given by (12). The
Ni,j entries for L = 0 is given in (13).
A PPENDIX B
I NTERCHANGE
OF I NTEGRATION AND
S UMMATION
From (34) we have,
I=
Z
0
∞
u
p + x −x z
e x (0 F1 (q, wi x))(1 F1 (1; NR ;
))dx.
v−x
x−v
Substituting the series expansion of 0 F1 term and using the property 1 F1 (a; b; z) = ez 1 F1 (b −
a; b; −z) [44], we obtain,
Z ∞X
∞
u
−u
wik p + x −x k+z
e x ke x−v 1 F1 (NR − 1; NR ;
) dx.
I=
(q)
x−v
k k! v − x
0
k=0
September 18, 2017
DRAFT
22
We see that v < 0 and u < 0 and hence v = −|v| and the argument in confluent hypergeometric
−u
x−v
is positive. By application of a special case of Tonelli’s theorem [47], according
RP
PR
fn (x)dx =
to which, if fn (x) ≥ 0, ∀n, x, then
fn (x)dx, we can interchange the
term
summation and integration in I since all the terms inside the integration are positive. After
interchanging, again using the property 1 F1 (a; b; z) = e−z 1 F1 (b − a; b; −z) [44], we get,
Z ∞
∞
X
wik
p + x −x k+z
u
I=−
e x
)dx.
1 F1 (1; NR ;
(q)k k! 0 |v| + x
x−v
k=0
Writing 1 F1 in terms of its series expansion, we get,
Z ∞ X
∞
∞
X
ul (1)l p + x −x xk+z
wik
e
dx .
I=
l
(q)
(N
(x
−
v)
k k!
R )l l! v − x
0
l=0
k=0
Since u < 0, ul alternates between positive and negative values. Therefore, Tonelli’s theorem
cannot be applied as in the previous interchange. So we use Lebesgue dominated convergence theorem to interchange the inner summation and integration. By Lebesgue dominated convergence
P
theorem [48], if a sequence of measurable functions fl are such that |fl (x)| ≤ dl (x) ∀l, x, dl (x)
RP
P
PR
PR
gl (x)dx =
Jl converges ∀x, then
fl (x)dx.
converges ∀x and
fl (x)dx =
ul (1)l p+x −x xk+z
p+x
e (x−v)l . | (v−x)
| < 1 because v = −p − y, where y > 0. Therefore,
(NR )l l! v−x
P
l
k+z
|u| (1)l −x x
e
an converges
|fl | < dl = (N
l . Applying ratio test, according to which, the series
)
l!
(x−v)
R l
P
dl+1
|u|
sup al+1
if lim
dl converges. Let,
l→∞ | al | < 1, we get, liml→∞ | dl | = liml→∞ NR +l = 0. Therefore
Assume fl =
Jl =
Z
∞
0
|u|l (1)l −x xk+z
e
dx
(NR )l l!
(x − v)l
(41)
Using (36) , equation (41) can be simplified as,
Jl =
|u|l (1)l
Γ(k + z + 1)U(l, l − k − z, −v).
(NR )l l!
Since U(a, b − k, x) is monotonically decreasing with k [20],
|Jl | < Hl =
|u|l (1)l
Γ(k + z + 1)U(l, l + 1, −v).
(NR )l l!
Using the identity U(a, a + 1, z) = z −a ,
|Jl | < Hl =
|u|l (1)l
Γ(k + +z1)(−v)−l .
(NR )l l!
P
|u|
|
=
lim
=
0.
Hence
by
ratio
test,
Hl converges.
It can be seen that, liml→∞ | HHl+1
l→∞
(N
+l)(−v)
R
l
P
P
Since |Jl | is upper bounded by Hl and
Hl converges,
Jl converges. In our case, since
P
PR
P
|fl | ≤ dl ∀l and
dl converges and
dl (x)dx =
Jl converges, the integration and the
inner summation can be interchanged.
September 18, 2017
DRAFT
23
A PPENDIX C
S ER
APPROXIMATION FOR
R AYLEIGH
INTERFERERS
Consider the expression to be simplified, Mη (s) = c|Nσ2 =0,L=0 |, where
((n2 − n1 )!)−n1
(−1)NR (σ 2 /EI )(NR −n1 )
c = Qn1
(n
−
i)!
1
i=1
and Nσ2 =0,L=0 is from (14) for L = 0. First the common terms inside each column or row
of the determinant is taken out of the determinant and canceled with the existing terms in the
constant c. All columns from j = 2, ..., n1 are flipped and all rows from i = 1, .., n1 are flipped.
Γ(n2 − n1 + i + j − 1) is then removed from each row, to obtain Ñ. Now, Mη (s) = c|Ñ|, where
Qn1
(n2 − n1 + i)!
Qn1
c = Qn1 i=1
(−1)NR +1 (σ 2 /EI )(NR −n1 ) ,
(42)
(n
−
i)!
(n
−
i)!
1
2
i=1
i=1
A(i)
j = 1, i = 1, ..., n1 ,
Ñi,j = 1,
(43)
j = 2 i = 1, ..., n1 .
Qj−2 (n − n + i + k), j = 3, ..., n , 1 ≤ i ≤ n
2
1
1
1
k=1
where
" T
1
X
1
(aNR sED /EI )l
A(i) =
− Γ(n2 − NR + i + 1)
(n2 − n1 + i)! l=0
(NR )l
NX
R −n1
1 F1 (t; NR ; aNR sED /(−bsED )
U(l + 1, l − n2 + NR − i + 1, −bsED /EI ) −
(bsED /EI )t
t=1
#
× Γ(t + n2 − NR + i)
From [23], shifted factorials are defined by,
1
; n = 0,
(z)s;n =
z(z + s)....(z + (n − 1)s) ; n = 1, 2, ...
A special case of this is the Pochhammer’s symbols, when s = 1.
1
; n = 0,
(z)n = (z)1;n =
z(z + 1)....(z + (n − 1)) ; n = 1, 2, ...
(44)
(45)
In our case, in the Ñ matrix, we have such Pochhammer’s symbols in all columns except in the
first. From [23, Lemma.1], we have the relation that determinant of a matrix with ij th element
for 0 ≤ i, j ≤ n − 1, being a shifted factorial (zj )s;i is given by |(zj )s;i | = ∆n (z), where
September 18, 2017
DRAFT
24
∆n (z) =
Q
0≤i<j≤n−1 (zj
− zi ). Evaluating |Ñ| by Laplace expansion along the first column and
using the above relation from [23], we get
|Ñ| =
n1
X
i=1
(−1)i+1 A(i)∆in1 −1 (z),
(46)
where z = [n2 −n1 +1+1, n2 −n1 +1+2, ...., n2 −n1 +1+n1] and ∆in1 −1 (z) is the Vandermonde
determinant formed by all elements of the vector z except the ith element. Any Vandermonde
determinant remains unchanged if from each element of the matrix, one subtracts the same
Q
Q
constant, i.e., ∆n (z + c) = 0≤k<j≤n−1((zj + c) − (zk + c)) = 0≤k<j≤n−1(zj − zk ). Hence, the
constant n2 − n1 + 1 can be subtracted from each element of the vector z. Hence,
|Ñ| =
n1
X
i=1
(−1)i+1 A(i)∆in1 −1 (z),
(47)
where z = [1, ..., n1 ]. The Vandermonde determinant ∆n (z), whose nodes are given by first
Q
n1 integers, i.e., z = [1, ..., n1 ], is given by ∆n (z) = 1≤k<j≤n1 (j − k). For simplifying this
expression, we expand the double product as follows:
n1
Y
∆n (z) = (n1 − 1)!(n1 − 2)!...(1)! =
(j − 1)!.
(48)
j=1
However, we actually want to evaluate ∆in1 −1 (z) and not ∆n (z). Note that the Vandermonde
determinant ∆in1 −1 (z) in which the ith element is missing, is given by,
Y
(j − k)
∆in1 −1 (z) =
1≤k<j≤n1 ;k,j6=i
Note that the above expression is difficult to evaluate. Hence to obtain a simplified expression
we multiply and divide the expression for ∆in1 −1 (z) by the terms that are present in ∆n1 (z), but
are missing in ∆in1 −1 (z). We thus obtain,
∆in1 −1 (z) =
Q
− k)
(i − 1)!(n1 − i)!.
1≤k<j≤n1 ; (j
Substituting (48) in the above expression, we obtain, ∆in1 −1 (z) in terms of ∆n1 (z) as,
Qn1
∆
(z)
n
j=1 (j − 1)!
1
∆in1 −1 (z) =
=
.
(i − 1)!(n1 − i)!
(i − 1)!(n1 − i)!
NR +1
2
(NR −n1 )
/EI )
n2 !
Hence the final expression becomes Mη (s) = c|Ñ| where, c = (−1)(n −n(σ)! Q
and
n1
(n
−i)!
2
1
1
i=1
Qn1
n1
X
j=1 (j − 1)!
i+1
.
(49)
|Ñ| =
(−1) A(i)
(n1 − i)!(i − 1)!
i=1
September 18, 2017
DRAFT
25
A PPENDIX D
M OMENTS
OF
SINR
For the case of L = n1 we will derive the lth moment. The mgf equation for this case can be
written as
Mη (s) = c
n1
X
k=1
R∞
2
( Eσ I
0
(−1)k+1 ρk |Yk |,
(50)
aNR s
)
xEI /ED +σ 2 /ED −bs
2
(bsED /EI −σ /EI −x)
1 F1 (1;NR ;
−x n2 −NR
+ x)e x
dx
where ρk =
0 F1 (n2 − n1 + 1; wk x)
PNR −n1 1 F1 (t;NR ;aNR sED /(σ2 −bsED )
2
−
× Eσ I Γ(t + n2 − NR ) 1 F1 (t + n2 − NR ; n2 − n1 + 1; wk )
t=1
(bsED /EI −σ2 /EI )t
2
+ Γ(t + n2 − NR + 1)1 F1 (t + n2 − NR + 1; n2 − n1 + 1; wi ) , Yi,j = Eσ I 1 F1 (n2 − j + 1; n2 −
n1 + 1; wk )Γ(n2 − j + 1) + Γ(n2 − j + 2)1 F1 (n2 − j + 2; n2 − n1 + 1; wi ) and Yk is the matrix
Y with k th row and first column removed. The lth moment is given by µl =
need to evaluate
dl
Mη (s)|s=0.
dsl
We
dl
ρ .
dsl k
We use the relations in [2] to evaluate the differential and obtain,
Z
ED l ∞ σ 2
dl
Ric
)
(
+ x)−l e−x xn2 −NR 0 F1 (n2 − n1 + 1; wk x)dx
ρk |s=0 = −αl l!(
dsl
EI
E
I
0
2
NX
R −n1
σ
−
Γ(t + n2 − NR ) 1 F1 (t + n2 − NR ; n2 − n1 + 1; wk )
αlRic
E
I
t=1
EI ED
+ Γ(t + n2 − NR + 1)1 F1 (t + n2 − NR + 1; n2 − n1 + 1; wi ) (t)l (− 2 )t ( 2 )l
σ
σ
R
P
k
∞ 2
R /b)
. The integral 0 ( Eσ I +x)−l e−x xn2 −NR 0 F1 (n2 −n1 +1; wk x)dx
where αlRic = bl lk=0 kl (aN
(NR )k
can only be solved by expanding the hypergeometric function and exchanging the integration
and summation to obtain,
dl
ρ |
dsl k s=0
= −αlRic dl where
∞
σ2
ED l X (wk )n (n2 − n1 + 1)n
)
Γ(n2 − NR + n + 1)U(l, l − n2 + NR − n, )
dl = l!(
EI n=0
n!
EI
NX
R −n1
σ2
+
Γ(t + n2 − NR ) 1 F1 (t + n2 − NR ; n2 − n1 + 1; wk )
EI
t=1
EI ED
+ Γ(t + n2 − NR + 1)1 F1 (t + n2 − NR + 1; n2 − n1 + 1; wi ) (t)l (− 2 )t ( 2 )l
(51)
σ
σ
September 18, 2017
DRAFT
26
A PPENDIX E
C ORRELATED
AND / OR
U NEQUAL
POWER I NTERFERERS
Using (24) to simplify (22), we obtain the MGF of η as
#
!
Z ∞" Y
n1
2
det(J)
σ + λi
|Υ|
Mη (s) = c1 (−1)NR (σ 2 )(NR −NI )
(NR −NI )
VNI (ΛR )
0
i=1 λi
×
NI
Y
i<j
(λi − λj )
= (−1)NR (σ 2 )(NR −NI )
Z
NI
Y
R −NI
λN
dλ1 ...dλNI
k
k=1
∞
0
c1
n1
Y
(σ 2 + λi )|J||Υ|dλ1 ...dλNI .
(52)
i=1
From Theorem 2 in Appendix of [32], it can be observed that, for two arbitrary m × m matrices
φ(y) and ψ(y) with ij th element φi (yj ) and ψi (yj ) and arbitrary function g(.), where y =
[y1 , y2 , ...ym ]T , the following identity holds:
Z b
Z
Z
m
Y
g(yk )dy1dy2 ...dym = det
{φi (y)ψj (y)g(y)dy}i,j=1,..,m .
...
|φ(y)||ψ(y)|
b≥yi ≥a
a
k=1
Using the above property to simplify MGF in (52), we get,
Mη (s) = c|N|,
where
1
c = (−1)NR (σ 2 )(NR −NI ) (−1) 2 NI (NI −1) QNI
1
i<j ( ri
and
Ni,j =
(53)
|Ψ|−NR
,
Q I
− r1j ) N
k=1 (NR − k)!
R
x
∞ (σ 2 + x)e− ri [h1 (s, x) − PNR −n1 ht (s, 0)xt−1 ]dx,
0
R
x
∞ (σ 2 + x)e− ri xNR −j dx,
0
t=1
j = 1, i = 1, ..., NI ,
j = 2, ..., NI i = 1, ..., NI .
R∞
Solving the integrals using the identity 0 e−px xs−1 dx = p−s Γ(s) from [44], we obtain
aNR s
R∞ 2
− rx 1 F1 (1;NR ; x/ED +σ 2 /ED −bs )
i
dx
(σ
+
x)e
(bsED −σ2 −x)
0
aNR s
PNR −NI 1 F1 (t;NR ; σ2 /ED −bs ) 2 t
Ni,j =
−
(σ ri Γ(t) + rit+1 Γ(t + 1)], j = 1, i = 1, ..., NI ,
t=1
(bsED −σ2 )t
σ 2 riNR −j+1 Γ(NR − j + 1) + riNR −j+2 Γ(NR − j + 2), j = 2, ..., NI , i = 1, ..., NI .
We can solve the integral for j = 1, by expanding the 1 F1 hypergeometric series and interchanging the integration and summation. The approach followed in Appendix A can be followed here.
After making simplifications similar to the ones made for (35), we will obtain the final N matrix
as (26).
September 18, 2017
DRAFT
27
Approximation for σ 2 ≈ 0
The determinant evaluation of |N| can be significantly simplified for σ 2 ≈ 0. We first substitute
σ 2 = 0 in Ni,j , to obtain
A(i), j = 1, i = 1, ..., NI ,
Ni,j =
r NR −j+2Γ(NR − j + 2),
i
where A(i) =
1)). By taking
j = 2, ..., NI , i = 1, ..., NI .
aN s
R )
P R −NI 1 F1 (t;NR ; −bs
(aNr s)l
−l+1
U(2, 2 − l, ri |bsED |)) − N
(rit+1 Γ(t +
t=1
l=0 (NR )l (|bsED |
(bsED )t
riNr −NI +2 and common gamma terms outside the determinant term we obtain,
P∞
Mη (s) ≈ c|N|,
1
c = (−1)NR (σ 2 )(NR −NI ) (−1) 2 NI (NI −1) QNI
1
i<j ( ri
×
NI
Y
riNR −NI +2
j=2
i=1
Ni,j =
NI
Y
Γ(NR − j + 2),
A(i)r −NR +NI −2 ,
i
r NI −j ,
i
(54)
|Ψ|−NR
Q I
− r1j ) N
k=1 (NR − k)!
j = 1, i = 1, ..., NI ,
j = 2, ..., NI , i = 1, ..., NI .
Expanding along the first column, we obtain an approximation for the MGF for σ 2 = 0 as,
Mη (s) ≈ c
i
NI
X
i=1
(−1)i+1 A(i)ri−NR +NI −2 |V i (r)|,
(55)
where V (r) denotes the Vandermonde matrix formed from all elements of r = (r1 , r2 , .., rNI )
except the ith element. Note that, we do not substitute σ 2 ≈ 0 in the c term but only in the |N|
term, to obtain the approximation.
R EFERENCES
[1] J. Winters, “Optimum Combining in Digital Mobile Radio with Cochannel Interference,” IEEE Journal on Selected Areas
in Communications, vol. 2, no. 4, pp. 528–539, July 1984.
[2] M. McKay, A. Zanella, I. Collings, and M. Chiani, “Error Probability and SINR Analysis of Optimum Combining in
Rician Fading,” IEEE Transactions on Communications, vol. 57, no. 3, pp. 676–687, March 2009.
[3] O. B. S. Ali, C. Cardinal, and F. Gagnon, “Performance of Optimum Combining in a Poisson Field of Interferers and
Rayleigh Fading Channels,” IEEE Transactions on Wireless Communications, vol. 9, no. 8, pp. 2461–2467, August 2010.
[4] P. D. Rahimzadeh and N. C. Beaulieu, “Limits to Performance of Optimum Combining with Dense Multiple Correlated
Antennas,” IEEE Transactions on Communications, vol. 58, no. 7, pp. 2014–2022, July 2010.
September 18, 2017
DRAFT
28
[5] K. Chung, “An Analytical Expression for Performance of Optimum Combining with Multiple Correlated CCIs and Two
Antennas,” IEEE Communications Letters, vol. 16, no. 4, pp. 458–461, April 2012.
[6] A. Shah and A. Haimovich, “Performance Analysis of Optimum Combining in Wireless Communications with Rayleigh
Fading and Cochannel Interference,” IEEE Transactions on Communications, vol. 46, no. 4, pp. 473–479, Apr 1998.
[7] J. S. Kwak and J. H. Lee, “Performance Analysis of Optimum Combining for Dual-Antenna Diversity with Multiple
Interferers in a Rayleigh Fading Channel,” IEEE Communications Letters, vol. 6, no. 12, pp. 541–543, Dec 2002.
[8] M. Chiani, M. Win, A. Zanella, R. Mallik, and J. Winters, “Bounds and Approximations for Optimum Combining of
Signals in the Presence of Multiple Cochannel Interferers and Thermal Noise,” IEEE Transactions on Communications,
vol. 51, no. 2, pp. 296–307, Feb 2003.
[9] R. Mallik, M. Win, M. Chiani, and A. Zanella, “Bit-Error Probability for Optimum Combining of Binary Signals in the
Presence of Interference and Noise,” IEEE Transactions on Wireless Communications, vol. 3, no. 2, pp. 395–407, March
2004.
[10] S. Menon and S. Kalyani, “SER for Optimal Combining in the Presence of Multiple Correlated Co-channel Interferers,”
IEEE Communications Letters, vol. PP, no. 99, pp. 1–1, 2015.
[11] D. Yue, X. Wang, and F. Xu, “Performance Analysis for Optimum Combining of Rayleigh Fading Signals with Correlated
Rayleigh Interferers and Noise,” IEEE Signal Processing Letters, vol. 13, no. 5, pp. 269–272, May 2006.
[12] Q. Zhang and X. Cui, “Outage Probability for Optimum Combining of Arbitrarily Faded Signals in the Presence of
Correlated Rayleigh Interferers,” IEEE Transactions on Vehicular Technology, vol. 53, no. 4, pp. 1043–1051, July 2004.
[13] N. Suraweera and N. C. Beaulieu, “Optimum Combining for Cooperative Relaying in a Poisson Field of Interferers,” IEEE
Transactions on Communications, vol. 63, no. 9, pp. 3132–3142, Sept 2015.
[14] ——, “Optimum Combining in Dual-Hop AF Relaying for Maximum Spectral Efficiency in the Presence of Co-Channel
Interference,” IEEE Transactions on Communications, vol. 63, no. 6, pp. 2071–2080, June 2015.
[15] ——, “Optimum Combining With Joint Relay and Antenna Selection for Multiple-Antenna Relays in the Presence of
Co-Channel Interference,” IEEE Communications Letters, vol. 18, no. 8, pp. 1459–1462, Aug 2014.
[16] A. Afana, S. Ikki, T. M. N. Ngatched, and O. A. Dobre, “Performance analysis of cooperative networks with optimum
combining and Co-channel interference,” in 2015 IEEE International Conference on Communication Workshop (ICCW),
June 2015, pp. 949–954.
[17] Y. Zhong, W. Zhang, and M. Haenggi, “Managing Interference Correlation Through Random Medium Access,” IEEE
Transactions on Wireless Communications, vol. 13, no. 2, pp. 928–941, February 2014.
[18] S. Kumar and S. Kalyani, “Impact of Correlation between Interferers on Coverage Probability and rate in Cellular Systems,”
arXiv preprint arXiv:1707.08802, 2017.
[19] ——, “Impact of Correlated Interferers on Coverage and Rate of FFR and SFR Schemes,” IEEE Transactions on Vehicular
Technology, vol. 65, no. 1, pp. 434–440, Jan 2016.
[20] J. Zhang, Z. Tan, H. Wang, Q. Huang, and L. Hanzo, “The Effective Throughput of MISO Systems Over κ-µ Fading
Channels,” IEEE Transactions on Vehicular Technology, vol. 63, no. 2, pp. 943–947, Feb 2014.
[21] M. You, H. Sun, J. Jiang, and J. Zhang, “Effective Rate Analysis in Weibull Fading Channels,” IEEE Wireless
Communications Letters, vol. 5, no. 4, pp. 340–343, Aug 2016.
[22] J. Zhang, L. Dai, Z. Wang, D. W. K. Ng, and W. H. Gerstacker, “Effective Rate Analysis of MISO Systems over α-µ
Fading Channels,” in 2015 IEEE Global Communications Conference (GLOBECOM), Dec 2015, pp. 1–6.
[23] J.-M. Normand, “Calculation of some determinants using the s -shifted factorial,” Journal of Physics A: Mathematical
and General, vol. 37, no. 22, p. 5737, 2004. [Online]. Available: http://stacks.iop.org/0305-4470/37/i=22/a=003
September 18, 2017
DRAFT
29
[24] H. V. Poor and S. Verdu, “Probability of error in MMSE multiuser detection,” IEEE Transactions on Information Theory,
vol. 43, no. 3, pp. 858–871, May 1997.
[25] A. Giorgetti and M. Chiani, “Influence of Fading on the Gaussian Approximation for BPSK and QPSK with Asynchronous
Cochannel Interference,” Wireless Communications, IEEE Transactions on, vol. 4, no. 2, pp. 384–389, March 2005.
[26] J. Proakis and M. Salehi, Fundamentals of Communication Systems.
Pearson Education, 2007.
[27] M. Chiani, M. Win, A. Zanella, and J. Winters, “Exact Symbol Error Probability for Optimum Combining in the Presence
of Multiple Co-channel Interferers and Thermal Noise,” in Global Telecommunications Conference, 2001. GLOBECOM
’01. IEEE, vol. 2, 2001, pp. 1182–1186 vol.2.
[28] ——, “A Simple and Asymptotically Tight Upper Bound on the Symbol Error Probability of Adaptive Antennas with
Optimum Combining,” in Communications, 2002. ICC 2002. IEEE International Conference on, vol. 3, 2002, pp. 1878–
1881 vol.3.
[29] M. Chiani, D. Dardari, and M. K. Simon, “New Exponential Bounds and Approximations for the Computation of Error
Probability in Fading Channels,” IEEE Transactions on Wireless Communications, vol. 2, no. 4, pp. 840–845, July 2003.
[30] H. Exton, Multiple hypergeometric Functions and Applications, ser. Ellis Horwood series in mathematics and its
applications.
E. Horwood, 1976.
[31] S. Jin, M. McKay, X. Gao, and I. Collings, “MIMO Multichannel Beamforming: SER and Outage Using New Eigenvalue
Distributions of Complex Noncentral Wishart Matrices,” IEEE Transactions on Communications, vol. 56, no. 3, pp. 424–
434, March 2008.
[32] M. Chiani, M. Win, and A. Zanella, “On the Capacity of Spatially Correlated MIMO Rayleigh-Fading Channels,” IEEE
Transactions on Information Theory, vol. 49, no. 10, pp. 2363–2371, Oct 2003.
[33] J. F. Paris, “Outage Probability in η-µ/ η-µ and κ-µ/ η-µ Interference-Limited Scenarios,” IEEE Transactions on
Communications, vol. 61, no. 1, pp. 335–343, January 2013.
[34] S. Kumar and S. Kalyani, “Coverage Probability and Rate for κ -µ/η-µ Fading Channels in Interference-Limited Scenarios,”
IEEE Transactions on Wireless Communications, vol. 14, no. 11, pp. 6082–6096, Nov 2015.
[35] S. Kumar, G. Chandrasekaran, and S. Kalyani, “Analysis of Outage Probability and Capacity for κ- µ/η- µ Faded Channel,”
IEEE Communications Letters, vol. 19, no. 2, pp. 211–214, Feb 2015.
[36] G. Zhou, W. Xu, and G. Bauch, “A practical scheme to achieve sum capacity for strong interference-limited scenarios,”
in 2015 IEEE International Conference on Communications (ICC), June 2015, pp. 2797–2802.
[37] C. Chayawan and V. A. Aalo, “On the outage probability of optimum combining and maximal ratio combining schemes
in an interference-limited Rice fading channel,” IEEE Transactions on Communications, vol. 50, no. 4, pp. 532–535, Apr
2002.
[38] T. Ratnarajah, R. Vaillancourt, and M. Alvo, “Complex random matrices and Rician channel capacity,” Problems of
Information Transmission, vol. 41, no. 1, pp. 1–22, 2005. [Online]. Available: http://dx.doi.org/10.1007/s11122-005-0006-6
[39] M. R. McKay and I. B. Collings, “General Capacity Bounds for Spatially Correlated Rician MIMO Channels,” IEEE
Transactions on Information Theory, vol. 51, no. 9, pp. 3121–3145, Sept 2005.
[40] R. Couillet and M. Debbah, Random Matrix Methods for Wireless Communications.
Cambridge University Press, 2011.
[Online]. Available: https://books.google.co.in/books?id= j7pT9HjKAUC
[41] R. Xu, Z. Zhong, and J. M. Chen, “Approximation to the Capacity of Rician Fading MIMO Channels,” in VTC Spring
2009 - IEEE 69th Vehicular Technology Conference, April 2009, pp. 1–5.
[42] M. Ohlson and T. Koski, “On the Distribution of Matrix Quadratic Forms,” Communications in Statistics-Theory and
Methods, vol. 41, no. 18, pp. 3403–315.
September 18, 2017
DRAFT
30
[43] S. Kumar, G. F. Pivaro, G. Fraidenraich, and C. F. Dias, “On the exact and approximate eigenvalue distribution for sum
of Wishart matrices,” arXiv preprint arXiv:1504.00222, 2015.
[44] A. Jeffrey and D. Zwillinger, Table of Integrals, Series, and Products, ser. Table of Integrals, Series, and Products Series.
Elsevier Science, 2007.
[45] M. Abramowitz and I. Stegun, Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables,
ser. Applied mathematics series. U.S. Government Printing Office, 1972.
[46] J. A. C.M. Joshi, “Some inequalities for the gauss and Kummer Hypergeometric Functions ,” Indian Journal of Pure and
Applied Mathematics, vol. 22, no. 8, pp. 637–644.
[47] T. Tao, An Introduction to Measure Theory, ser. Graduate studies in mathematics. American Mathematical Society, 2011.
[48] T. Apostol, Mathematical Analysis, ser. Addison-Wesley series in mathematics.
Addison-Wesley Publishing Company,
1974.
September 18, 2017
DRAFT
| 7cs.IT
|
Right Amenable Left Group Sets and the
Tarski-Følner Theorem
arXiv:1603.06460v2 [math.GR] 24 Mar 2016
Simon Wacker
Karlsruhe Institute of Technology
[email protected]
http://www.kit.edu
Abstract We introduce right amenability, right Følner nets, and right
paradoxical decompositions for left homogeneous spaces and prove the
Tarski-Følner theorem for left homogeneous spaces with finite stabilisers.
It states that right amenability, the existence of right Følner nets, and
the non-existence of right paradoxical decompositions are equivalent.
Keywords: amenability, group actions, Tarski-Følner theorem
The notion of amenability for groups was introduced by John von Neumann
in 1929, see the paper ‘Zur allgemeinen Theorie des Maßes’[4]. It generalises the
notion of finiteness. A group G is left or right amenable if there is a finitely
additive probability measure on P(G) that is invariant under left and right
multiplication respectively. Groups are left amenable if and only if they are
right amenable. A group is amenable if it is left or right amenable.
The definitions of left and right amenability generalise to left and right group
sets respectively. A left group set (M, G, ⊲) is left amenable if there is a finitely
additive probability measure on P(M ) that is invariant under ⊲. There is in
general no natural action on the right that is to a left group action what right
multiplication is to left group multiplication. Therefore, for a left group set there
is no natural notion of right amenability.
A transitive left group action ⊲ of G on M induces, for each element m0 ∈ M
and each family {gm0 ,m }m∈M of elements in G such that, for each point m ∈ M ,
we have gm0 ,m ⊲ m0 = m, a right quotient set semi-action P of G/G0 on M with
−1
⊲ m, where G0 is the stabiliser of
defect G0 given by m P gG0 = gm0 ,m ggm
0 ,m
m0 under ⊲. Each of these right semi-actions is to the left group action what
right multiplication is to left group multiplication. They occur in the definition
of global transition functions of cellular automata over left homogeneous spaces
as defined in [5]. A coordinate system is a choice of m0 and {gm0 ,m }m∈M .
A left homogeneous space is right amenable if there is a coordinate system
such that there is a finitely additive probability measure on P(M ) that is semiinvariant under P. For example finite left homogeneous spaces, abelian groups,
and finitely right generated left homogeneous spaces of sub-exponential growth
are right amenable, in particular, quotients of finitely generated groups of subexponential growth by finite subgroups acted on by left multiplication.
2
Right Amenable Left Group Sets and the Tarski-Følner Theorem
A net of non-empty and finite subsets of M is a right Følner net if, broadly
speaking, these subsets are asymptotically invariant under P. A finite subset
E of G/G0 and two partitions {Ae }e∈E and {Be }e∈E of M constitute a right
paradoxical decomposition if the map _ P e is injective on Ae and Be , and the
family {(Ae P e) ∪· (Be P e)}e∈E is a partition of M . The Tarski-Følner theorem
states that right amenability, the existence of right Følner nets, and the nonexistence of right paradoxical decompositions are equivalent.
The Tarski alternative theorem and the theorem of Følner, which constitute the Tarski-Følner theorem, are famous theorems by Alfred Tarski and
Erling Følner from 1938 and 1955, see the papers ‘Algebraische Fassung des
Maßproblems’[3] and ‘On groups with full Banach mean value’[2]. This paper is
greatly inspired by the monograph ‘Cellular Automata and Groups’[1] by Tullio
Ceccherini-Silberstein and Michel Coornaert.
For a right amenable left homogeneous space with finite stabilisers we may
choose a right Følner net. Using this net we show in [6] that the Garden of Eden
theorem holds for such spaces. It states that a cellular automaton with finite set
of states and finite neighbourhood over such a space is surjective if and only if
it is pre-injective.
In Sect. 1 we introduce finitely additive probability measures and means, and
kind of right semi-actions on them. In Sect. 2 we introduce right amenability. In
Sect. 3 we introduce right Følner nets. In Sect. 4 we introduce right paradoxical
decompositions. In Sect. 5 we prove the Tarski alternative theorem and the
theorem of Følner. And in Sect. 6 we show under which assumptions left implies
right amenability and give two examples of right amenable left homogeneous
spaces.
Preliminary Notions. A left group set is a triple (M, G, ⊲), where M is a set,
G is a group, and ⊲ is a map from G × M to M , called left group action of G
on M , such that G → Sym(M ), g 7→ [g ⊲ _], is a group homomorphism. The
action ⊲ is transitive if M is non-empty and for each m ∈ M the map _ ⊲ m
is surjective; and free if for each m ∈ M the map _ ⊲ m is injective. For each
m ∈ M , the set G ⊲ m is the orbit of m, the set Gm = (_ ⊲ m)−1 (m) is the
stabiliser of m, and, for each m′ ∈ M , the set Gm,m′ = (_ ⊲ m)−1 (m′ ) is the
transporter of m to m′ .
A left homogeneous space is a left group set M = (M, G, ⊲) such that ⊲ is
transitive. A coordinate system for M is a tuple K = (m0 , {gm0 ,m }m∈M ), where
m0 ∈ M and, for each m ∈ M , we have gm0 ,m ⊲ m0 = m. The stabiliser Gm0 is
denoted by G0 . The tuple R = (M, K) is a cell space. The set {gG0 | g ∈ G}
of left cosets of G0 in G is denoted by G/G0 . The map P : M × G/G0 → M ,
(m, gG0 ) 7→ gm0 ,m g ⊲ m0 is a right semi-action of G/G0 on M with defect G0 ,
which means that
∀ m ∈ M : m P G0 = m,
∀ m ∈ M ∀ g ∈ G ∃ g0 ∈ G0 : ∀ g′ ∈ G/G0 : m P g · g′ = (m P gG0 ) P g0 · g′ .
Right Amenable Left Group Sets and the Tarski-Følner Theorem
3
It is transitive, which means that the set M is non-empty and for each m ∈ M
the map m P _ is surjective; and free, which means that for each m ∈ M the
map m P _ is injective; and semi-commutes with ⊲, which means that
∀ m ∈ M ∀ g ∈ G ∃ g0 ∈ G0 : ∀ g′ ∈ G/G0 : (g ⊲ m) P g′ = g ⊲ (m P g0 · g′ ).
For each A ⊆ M , let 1A : M → {0, 1} be the indicator function of A.
1
Finitely Additive Probability Measures and Means
In this section, let R = ((M, G, ⊲), (m0 , {gm0 ,m }m∈M )) be a cell space.
Definition 1. Let µ : P(M ) → [0, 1] be a map. It is called
1. normalised if and only if µ(M ) = 1;
2. finitely additive if and only if,
∀ A ⊆ M ∀ B ⊆ M : A ∩ B = ∅ =⇒ µ(A ∪ B) = µ(A) + µ(B) ;
3. finitely additive probability measure on M if and only if it is normalised and
finitely additive.
The set of all finitely additive probability measures on M is denoted by PM(M ).
Definition 2. The group G acts on [0, 1]P(M) on the left by
P(M)
→ [0, 1]P(M) ,
: G × [0, 1]
(g, ϕ) 7→ [A 7→ ϕ(g −1 ⊲ A)],
such that G PM(M ) ⊆ PM(M ).
Definition 3. The quotient set G/G0 kind of semi-acts on [0, 1]P(M) on the
right by
P(M)
× G/G0 → [0, 1]P(M) ,
: [0, 1]
(ϕ, g) 7→ [A 7→ ϕ(A P g)].
Definition 4. Let ϕ be an element of [0, 1]P(M) . It is called -semi-invariant if
and only if, for each element g ∈ G/G0 and each subset A of M such that the
map _ P g is injective on A, we have (ϕ g)(A) = ϕ(A).
Remark 1. Let R be the cell space ((G, G, ·), (eG , {g}g∈G )). Then, G0 = {eG }
and P = ·. Hence, : (ϕ, g) 7→ [A 7→ ϕ(A·g)]. Except for g not being inverted, this
is the right group action of G on PM(M ) as defined in [1, Sect. 4.3, Paragraph 4].
Moreover, for each element g ∈ G, the map _ P g is injective. Hence, being semi-invariant is the same as being right-invariant as defined in [1, Sect. 4.4,
Paragraph 2].
4
Right Amenable Left Group Sets and the Tarski-Følner Theorem
Definition 5. The vector space of bounded real-valued functions on M with
pointwise addition and scalar multiplication is denoted by ℓ∞ (M ), the supremum
norm on ℓ∞ (M ) is denoted by k_k∞ , the topological dual space of ℓ∞ (M ) is denoted by ℓ∞ (M )∗ , the pointwise partial order on ℓ∞ (M ) is denoted by ≤, and
the constant function [m 7→ 0] is denoted by 0.
Definition 6. Let ν : ℓ∞ (M ) → R be a map. It is called
1. normalised if and only if ν(1M ) = 1;
2. non-negativity preserving if and only if
∀ f ∈ ℓ∞ (M ) : (f ≥ 0 =⇒ ν(f ) ≥ 0);
3. mean on M if and only if it is linear, normalised, and non-negativity preserving.
The set of all means on M is denoted by M(M ).
Definition 7. Let Ψ be a map from ℓ∞ (M ) to ℓ∞ (M ). It is called non-negativity
preserving if and only if
∀ f ∈ ℓ∞ (M ) : (f ≥ 0 =⇒ Ψ (f ) ≥ 0).
Lemma 1. Let G0 be finite, let A be a finite subset of M , and let g be an element
of G/G0 . Then, |(_ P g)−1 (A)| ≤ |G0 | · |A|.
Proof. Let a ∈ A such that (_ P g)−1 (a) 6= ∅. There are m and m′ ∈ M such
that Gm0 ,m = g and m′ P g = a. For each m′′ ∈ M , we have m′′ P g = gm0 ,m′′ ⊲ m
and hence
m′′ P g = a ⇐⇒ m′′ P g = m′ P g
−1
⇐⇒ gm
′ gm0 ,m′′ ⊲ m = m
0 ,m
−1
⇐⇒ gm
′ gm0 ,m′′ ∈ Gm
0 ,m
⇐⇒ gm0 ,m′′ ∈ gm0 ,m′ Gm .
Moreover, for each m′′ and each m′′′ ∈ M with m′′ 6= m′′′ , we have gm0 ,m′′ 6=
gm0 ,m′′′ . Thus,
|(_ P g)−1 (a)| = |{m′′ ∈ M | m′′ P g = a}|
= |{m′′ ∈ M | gm0 ,m′′ ∈ gm0 ,m′ Gm }|
≤ |gm0 ,m′ Gm |
= |Gm |
= |G0 |.
S
Therefore, because (_ P g)−1 (A) = a∈A (_ P g)−1 (a), we have |(_ P g)−1 (A)| ≤
|G0 | · |A|.
⊓
⊔
Right Amenable Left Group Sets and the Tarski-Følner Theorem
5
Definition 8. The group G acts on ℓ∞ (M ) on the left by
: G × ℓ∞ (M ) → ℓ∞ (M ),
(g, f ) 7→ [m 7→ f (g −1 ⊲ m)].
Lemma 2. Let G0 be finite. The quotient set G/G0 kind of semi-acts on ℓ∞ (M )
on the right by
: ℓ∞ (M ) × G/G0 → ℓ∞ (M ),
(f, g) 7→ [m 7→
X
f (m′ )],
m′ ∈(_Pg)−1 (m)
such that, for each tuple (f, g) ∈ ℓ∞ (M )×G/G0 , we have kf gk∞ ≤ |G0 |·kf k∞ .
Proof. Let g ∈ G/G0 . Furthermore, let f ∈ ℓ∞ (M ). Moreover, let m ∈ M .
Because G0 is finite, according to Lemma 1, we have |(_ P g)−1 (m)| ≤ |G0 | < ∞.
Hence, the sum in the definition of is finite. Furthermore,
X
|f (m′ )|
|(f g)(m)| ≤
m′ ∈(_Pg)−1 (m)
≤
X
m′ ∈(_Pg)−1 (m)
1 · kf k∞
= |(_ P g)−1 (m)| · kf k∞
≤ |G0 | · kf k∞ .
Therefore, f
g ∈ ℓ∞ (M ), kf
gk∞ ≤ |G0 | · kf k∞ , and
is well-defined.
⊓
⊔
Remark 2. In the situation of Remark 1, we have : (f, g) 7→ [m →
7
f (m ·
g −1 )]. Hence, is the right group action of G on RG as defined in [1, Sect. 4.3,
Paragraph 5].
Lemma 3. Let G0 be finite and let g be an element of G/G0 . The map _ g is
linear, continuous, and non-negativity preserving.
Proof. Linearity follows from linearity of summation, continuity follows from
linearity and k_ gk∞ ≤ |G0 | · k_k∞ , and non-negativity preservation follows
from non-negativity preservation of summation.
⊓
⊔
Lemma 4 ([1, Proposition 4.1.7]). Let ν be a mean on M . Then, ν ∈
ℓ∞ (M )∗ and kνkℓ∞ (M)∗ = 1. In particular, ν is continuous.
⊓
⊔
Definition 9. The group G acts on ℓ∞ (M )∗ on the left by
∞
∗
∞
∗
: G × ℓ (M ) → ℓ (M ) ,
(g, ψ) 7→ [f 7→ ψ(g −1
such that G M(M ) ⊆ M(M ).
f )],
6
Right Amenable Left Group Sets and the Tarski-Følner Theorem
Definition 10. Let G0 be finite. The quotient set G/G0 kind of semi-acts on
ℓ∞ (M )∗ on the right by
∞
∗
∞
∗
: ℓ (M ) × G/G0 → ℓ (M ) ,
(ψ, g) 7→ [f 7→ ψ(f
g)].
Proof. Let ψ ∈ ℓ∞ (M )∗ and let g ∈ G/G0 . Then, ψ g = ψ ◦ (_
ψ and _ g are linear and continuous, so is ψ g.
g). Because
⊓
⊔
Definition 11. Let G0 be finite and let ψ be an element of ℓ∞ (M )∗ . It is called
-invariant if and only if, for each element g ∈ G/G0 and each function f ∈
ℓ∞ (M ), we have (ψ g)(f ) = ψ(f ).
Remark 3. In the situation of Remark 2, we have : (ψ, g) 7→ [f 7→ ψ(f g].
Except for g not being inverted, this is the right group action of G on ℓ∞ (G)∗
as defined in [1, Sect. 4.3, Paragraph 6]. Hence, being -invariant is the same as
being right-invariant as defined in [1, Sect. 4.4, Paragraph 3].
Theorem 1 ([1, Theorem 4.1.8]). The map
Φ : M(M ) → PM(M ),
ν 7→ [A 7→ ν(1A )],
is bijective.
⊓
⊔
Theorem 2 ([1, Theorem 4.2.1]). The set M(M ) is a convex and compact
subset of ℓ∞ (M )∗ equipped with the weak-∗ topology.
⊓
⊔
2
Right Amenability
In Definition 13 we introduce the notion of right amenability using finitely additive probability measures. And in Theorem 3 we characterise right amenability
of cell spaces with finite stabilisers using means.
Definition 12. Let (M, G, ⊲) be a left group set. It is called left amenable if
and only if there is a -invariant finitely additive probability measure on M .
Definition 13. Let M = (M, G, ⊲) be a left homogeneous space. It is called
right amenable if and only if there is a coordinate system K = (m0 , {gm0 ,m }m∈M )
for M such that there is a -semi-invariant finitely additive probability measure
on M , in which case the cell space R = (M, K) is called right amenable.
Remark 4. In the situation of Remark 1, being right amenable is the same as
being amenable as defined in [1, Definition 4.4.5].
In the remainder of this section, let R = ((M, G, ⊲), (m0 , {gm0 ,m }m∈M )) be
a cell space such that the stabiliser G0 of m0 under ⊲ is finite.
Right Amenable Left Group Sets and the Tarski-Følner Theorem
7
Lemma 5. Let g be an element of G/G0 and let A be a subset of M such that
the map _ P g is injective on A. Then, 1APg = 1A g.
Proof. For each m ∈ M , because _ P g is injective on A,
(
1, if m ∈ A P g,
1APg (m) =
0, otherwise,
= |{m′ ∈ A | m′ P g = m}|
X
=
1A (m′ )
m′ ∈(_Pg)−1 (m)
= (1A
In conclusion, 1APg = 1A
g)(m).
⊓
⊔
g.
Lemma 6 ([1, Lemma 4.1.9]). The vector space
E(M ) = {f : M → R | f (M ) is finite} (= span{1A | A ⊆ M })
is dense in the Banach space (ℓ∞ (M ), k_k∞ ).
⊓
⊔
Lemma 7. Let ψ be an element of ℓ∞ (M )∗ such that, for each element g ∈ G/
G0 and each subset A of M such that the map _ P g is injective on A, we have
(ψ g)(1A ) = ψ(1A ). The map ψ is -invariant.
Proof. Let g ∈ G/G0 .
First, let A ⊆ M . Moreover, let m ∈ M . According to Lemma 1, we have
km = |(_ P g)−1 (m)| ≤ |G0 |. Hence, there are pairwise distinct mm,1 , mm,2 ,
. . . , mm,km ∈ M such that (_ P g)−1 (m) = {mm,1 , mm,2 , . . . , mm,km }. For each
i ∈ {1, 2, . . . , |G0 |}, put
Ai = {mm,i | m ∈ M, km ≥ i} ∩ A.
Because, for each m ∈ M and each m′ ∈ M such that m 6= m′ , we have
(_ P g)−1 (m) ∩ (_ P g)−1 (m′ ) = ∅, the sets A1 , A2 , . . . , A|G0 | are pairwise
disjoint and the map _ P g is injective on each of these sets. Moreover, because
S|G0 |
P|G0 |
S
P g)−1 (m) = M , we have
m∈M (_
i=1 Ai = A. Therefore, 1A =
i=1 1Ai .
Thus, because ψ g and ψ are linear,
|G0 |
|G0 |
|G0 |
X
X
X
(ψ g)(1A ) = (ψ g)
(ψ g)(1Ai ) =
ψ(1Ai ) = ψ(1A ).
1 Ai =
i=1
i=1
i=1
Therefore, ψ g = ψ on the set of indicator functions. Thus, because the indicator
functions span E(M ), and ψ g and ψ are linear, ψ g = ψ on E(M ). Hence,
because E(M ) is dense in ℓ∞ (M ), and ψ g and ψ are continuous, ψ g = ψ on
ℓ∞ (M ). In conclusion, ψ is -invariant.
⊓
⊔
8
Right Amenable Left Group Sets and the Tarski-Følner Theorem
Theorem 3. The cell space R is right amenable if and only if there is a invariant mean on M .
Proof. Let Φ be the map in Theorem 1.
First, let R be right amenable. Then, there is -semi-invariant finitely additive probability measure µ on M . Put ν = Φ−1 (µ). Then, for each g ∈ G/G0 and
each A ⊆ M such that _ P g is injective on A, according to Lemma 5,
(ν g)(1A ) = ν(1A
g) = ν(1APg ) = µ(A P g) = µ(A) = ν(1A ).
Thus, according to Lemma 7, the mean ν is -invariant.
Secondly, let there be a -invariant mean ν on M . Put µ = Φ(ν). Then, for
each g ∈ G/G0 and each A ⊆ M such that _ P g is injective on A, according to
Lemma 5,
(µ g)(A) = µ(A P g) = ν(1APg ) = ν(1A
g) = ν(1A ) = µ(A).
⊓
⊔
Hence, µ is -semi-invariant.
3
Right Følner Nets
In this section, let R = ((M, G, ⊲), (m0 , {gm0 ,m }m∈M )) be a cell space.
Definition 14. Let {Fi }i∈I be a net in {F ⊆ M | F 6= ∅, F finite} indexed by
(I, ≤). It is called right Følner net in R indexed by (I, ≤) if and only if
∀ g ∈ G/G0 : lim
i∈I
|Fi r (_ P g)−1 (Fi )|
= 0.
|Fi |
Remark 5. In the situation of Remark 1, for each element g ∈ G and each index
i ∈ I, we have (_ P g)−1 (Fi ) = Fi · g −1 . Hence, right Følner nets in R are exactly
right Følner nets for G as defined in [1, First paragraph after Definition 4.7.2].
Lemma 8. Let V be a set, let W be a set, and let Ψ be a map from V × W to
R. There is a net {vi }i∈I in V indexed by (I, ≤) such that
∀ w ∈ W : lim Ψ (vi , w) = 0,
i∈I
(1)
if and only if, for each finite subset Q of W and each positive real number
ε ∈ R>0 , there is an element v ∈ V such that
∀ q ∈ Q : Ψ (v, q) < ε.
(2)
Proof. First, let there be a net {vi }i∈I in V indexed by (I, ≤) such that (1)
holds. Furthermore, let Q ⊆ W be finite and let ε ∈ R>0 . Because (1) holds, for
each q ∈ Q, there is an iq ∈ I such that,
∀ i ∈ I : (i ≥ iq =⇒ Ψ (vi , q) < ε).
Right Amenable Left Group Sets and the Tarski-Følner Theorem
9
Because (I, ≤) is a directed set and Q is finite, there is an i ∈ I such that, for
each q ∈ Q, we have i ≥ iq . Put v = vi . Then, (2) holds.
Secondly, for each finite Q ⊆ W and each ε ∈ R>0 , let there be a v ∈ V such
that (2) holds. Furthermore, let
I = {Q ⊆ W | Q is finite} × R>0
and let ≤ be the preorder on I given by
∀(Q, ε) ∈ I ∀(Q′ , ε′ ) ∈ I : (Q, ε) ≤ (Q′ , ε′ ) ⇐⇒ Q ⊆ Q′ ∧ ε ≥ ε′ .
For each (Q, ε) ∈ I and each (Q′ , ε′ ) ∈ I, the element (Q ∪ Q′ , min(ε, ε′ )) of I is
an upper bound of (Q, ε) and of (Q′ , ε′ ). Hence, (I, ≤) is a directed set.
By precondition, for each i = (Q, ε) ∈ I, there is a vi ∈ V such that
∀ q ∈ Q : Ψ (vi , q) < ε.
Let w ∈ W and let ε0 ∈ R>0 . Put i0 = ({w}, ε0 ). For each i = (Q, ε) ∈ I with
i ≥ i0 , we have w ∈ Q and ε ≤ ε0 . Hence,
∀ i ∈ I : (i ≥ i0 =⇒ Ψ (vi , w) < ε0 ).
Therefore, {vi }i∈I is a net in V indexed by (I, ≤) such that (1) holds.
⊓
⊔
Lemma 9. There is a right Følner net in R if and only if, for each finite subset
E of G/G0 and each positive real number ε ∈ R>0 , there is a non-empty and
finite subset F of M such that
∀e ∈ E :
|F r (_ P e)−1 (F )|
< ε.
|F |
Proof. This is a direct consequence of Lemma 8 with
Ψ : {F ⊆ M | F 6= ∅, F finite} × G/G0 → R,
(F, g) 7→
|F r (_ P g)−1 (F )|
.
|F |
⊓
⊔
Lemma 10. Let m be an element of M , and let g be an element of G/G0 . There
is an element g ∈ g such that
∀ g′ ∈ G/G0 : (m P g) P g′ = m P g · g′ .
Proof. There is a g ∈ G such that gG0 = g. Moreover, because P is a semi-action
with defect G0 , there is a g0 ∈ G0 such that
∀ g′ ∈ G/G0 : (m P gG0 ) P g′ = m P g · (g0−1 · g′ ).
Because g · (g0−1 · g′ ) = gg0−1 · g′ and gg0−1 ∈ g, the statement holds.
⊓
⊔
10
Right Amenable Left Group Sets and the Tarski-Følner Theorem
Lemma 11. Let A and A′ be two subsets of M , and let g and g′ be two elements
of G/G0 . Then, for each element m ∈ (_ P g)−1 (A) r (_ P g′ )−1 (A′ ),
mPg∈
[
A r (_ P g −1 · g′ )−1 (A′ ),
g∈g
′
mPg ∈
[
(_ P (g ′ )−1 · g)−1 (A) r A′ .
g′ ∈g′
Proof (Lemma 11). Let m ∈ (_ P g)−1 (A) r (_ P g′ )−1 (A′ ). Then, m P g ∈ A
and m P g′ ∈
/ A′ . According to Lemma 10, there is a g ∈ g and a g ′ ∈ g′ such
that (m P g) P g −1 · g′ = m P g′ ∈
/ A′ and (m P g′ ) P (g ′ )−1 · g = m P g ∈ A.
−1
′
−1
′
Hence, m P g ∈
/ (_ P g · g ) (A ) and m P g′ ∈ (_ P (g ′ )−1 · g)−1 (A). Therefore,
Pg −1 ·g′ )−1 (A′ ) and m Pg′ ∈ (_P (g ′ )−1 ·g)−1 (A)rA′ . In conclusion,
mPg ∈ Ar(_
S
S
m P g ∈ g∈g Ar(_ P g −1 ·g′ )−1 (A′ ) and m P g′ ∈ g′ ∈g′ (_ P (g ′ )−1 ·g)−1 (A)rA′ .
⊓
⊔
Lemma 12. Let G0 be finite, let F and F ′ be two finite subsets of M , and let
g and g′ be two elements of G/G0 . Then,
2
|F r (_ P g −1 · g′ )−1 (F ′ )|,
|G0 | · max
g∈g
|(_ P g)−1 (F ) r (_ P g′ )−1 (F ′ )| ≤
2
|(_ P (g ′ )−1 · g)−1 (F ) r F ′ |.
|G0 | · max
′
′
g ∈g
Proof. Put A = (_ P g)−1 (F ) r (_ P g′ )−1 (F ′ ). For each g ∈ g, put Bg =
F r (_ P g −1 · g′ )−1 (F ′ ). For each g ′ ∈ g′ , put Bg′ ′ = (_ P (g ′ )−1 · g)−1 (F ) r F ′ .
According to Lemma 11, the restrictions (_Pg)↾A→Sg∈g Bg and (_Pg′ )↾A→Sg′ ∈g′ B ′ ′
g
are well-defined. Moreover, for each m ∈ M , according to Lemma 1, we have
|(_ P g)−1 (m)| ≤ |G0 | and |(_ P g′ )−1 (m)| ≤ |G0 |. Therefore, because |g| = |G0 |,
|A| ≤ |G0 | · |
[
Bg | ≤ |G0 | ·
g∈g
X
2
|Bg | ≤ |G0 | · max|Bg |
g∈g
g∈g
and analogously
2
|A| ≤ |G0 | · max
|Bg′ ′ |.
′
′
g ∈g
⊓
⊔
Lemma 13. Let G0 be finite and let {Fi }i∈I be a net in {F ⊆ M | F 6=
∅, F finite} indexed by (I, ≤). The net {Fi }i∈I is a right Følner net in R if
and only if
|(_ P g)−1 (Fi ) r Fi |
∀ g ∈ G/G0 : lim
= 0.
(3)
i∈I
|Fi |
Proof. Let g ∈ G/G0 . Furthermore, let i ∈ I. Because Fi = (_ P G0 )−1 (Fi ),
according to Lemma 12,
2
|(_ P g)−1 (Fi ) r Fi | ≤ |G0 | · max|Fi r (_ P g −1 G0 )−1 (Fi )|
g∈g
Right Amenable Left Group Sets and the Tarski-Følner Theorem
and
11
|Fi r (_ P g)−1 (Fi )| ≤ |G0 |2 · max|(_ P g −1 G0 )−1 (Fi ) r Fi |.
g∈g
Moreover, |g| = |G0 | < ∞. Therefore, if {Fi }i∈I is a right Følner net in R, then
lim
|(_ P g)−1 (Fi ) r Fi |
= 0;
|Fi |
lim
|Fi r (_ P g)−1 (Fi )|
= 0.
|Fi |
i∈I
and, if (3) holds, then
i∈I
In conclusion, {Fi }i∈I is a right Følner net in R if and only if (3) holds.
Lemma 14. Let G0 be finite. There is a right Følner net in R if and only if,
for each finite subset E of G/G0 and each positive real number ε ∈ R>0 , there
is a non-empty and finite subset F of M such that
∀e ∈ E :
|(_ P e)−1 (F ) r F |
< ε.
|F |
Proof. This is a direct consequence of Lemma 13 and Lemma 8 with
Ψ : {F ⊆ M | F 6= ∅, F finite} × G/G0 → R,
(F, g) 7→
4
|(_ P g)−1 (F ) r F |
.
|F |
⊓
⊔
Right Paradoxical Decompositions
In this section, let R = ((M, G, ⊲), (m0 , {gm0 ,m }m∈M )) be a cell space.
Definition 15. Let A and A′ be two sets. The set A ∪ A′ is denoted by A ∪· A′
if and only if the sets A and A′ are disjoint.
Definition 16. Let E be a finite subset of G/G0 , and let {Ae }e∈E and {Be }e∈E
be two families of subsets of M indexed by E such that, for each index e ∈ E,
the map _ P e is injective on Ae and on Be , and
!
!
[
[
[
[
M = · Ae = · Be =
· Ae P e ∪·
· Be P e .
e∈E
e∈E
e∈E
e∈E
The triple (N, {Ae }e∈E , {Be }e∈E ) is called right paradoxical decomposition of
R.
Remark 6. In the situation of Remark 1, for each element g ∈ G, the map _ P g
is injective. Hence, right paradoxical decompositions of R are the same as right
paradoxical decompositions of G as defined in [1, Definition 4.8.1].
12
Right Amenable Left Group Sets and the Tarski-Følner Theorem
Lemma 15. Let G0 be finite and let (N, {Ae }e∈E , {Be }e∈E ) be a right paradoxical decomposition of R. Then,
X
X
X
X
(1Be e).
1M =
1 Ae =
1Be =
(1Ae e) +
e∈E
e∈E
e∈E
e∈E
Proof. This is a direct consequence of Definition 16 and Lemma 5.
5
Tarski’s and Følner’s Theorem
In this section, let R = ((M, G, ⊲), (m0 , {gm0 ,m }m∈M )) be a cell space such that
the stabiliser G0 of m0 under ⊲ is finite.
Lemma 16. Let g be an element of G/G0 . The map _ g is continuous, where
ℓ∞ (M )∗ is equipped with the weak-∗ topology.
Proof. For each f ∈ ℓ∞ (M ), let evf : ℓ∞ (M )∗ → R, ψ 7→ ψ(f ). Furthermore, let
f ∈ ℓ∞ (M ). Then, for each ψ ∈ ℓ∞ (M )∗ ,
(evf ◦(_ g))(ψ) = evf (ψ g) = (ψ g)(f ) = ψ(f
Thus, evf ◦(_ g) = evf g . Hence, because evf
Therefore, the map _ g is continuous.
g
g) = evf
g (ψ).
is continuous, so is evf ◦(_ g).
⊓
⊔
Lemma 17. Let {νi }i∈I be a net in M(M ) such that, for each element g ∈ G/
G0 , the net {(νi g) − νi }i∈I converges to 0 in ℓ∞ (M )∗ equipped with the weak-∗
topology. The cell space R is right amenable.
Proof. Let g ∈ G/G0 . According to Theorem 2, the set M(M ) is compact in
ℓ∞ (M )∗ equipped with the weak-∗ topology. Hence, there is a subnet {νij }j∈J
of {νi }i∈I that converges to a ν ∈ M(M ). Because, according to Lemma 16, the
map _ g is continuous, the net {(νij g) − νij }j∈J converges to (ν g) − ν
in ℓ∞ (M )∗ . Because it is a subnet of {(νi g) − νi }i∈I , it also converges to 0 in
ℓ∞ (M )∗ . Because the space ℓ∞ (M )∗ is Hausdorff, we have (ν g) − ν = 0 and
hence ν g = ν. Altogether, ν is a -invariant mean. In conclusion, according
to Theorem 3, the cell space R is right amenable.
⊓
⊔
Lemma 18. Let m be an element of M , and let E and E ′ be two subsets of
G/G0 . There is a subset E ′′ of G/G0 such that (m P E) P E ′ = m P E ′′ ; if
G0 ∈ E ∩ E ′ , then G0 ∈ E ′′ ; if E and E ′ are finite, then |E ′′ | ≤ |E| · |E ′ |; and
if G0 · E ′ ⊆ E ′ , then E ′′ = {g · e′ | e ∈ E, e′ ∈ E ′ , g ∈ e}.
Proof. For each e ∈ E, according to Lemma 10, there is a ge ∈ e such that
∀ g ∈ G/G0 : (m P e) P g = m P ge · g.
Put E ′′ = {ge · e′ | e ∈ E, e′ ∈ E ′ }. Then, (m P E) P E ′ = m P E ′′ . Moreover, if
G0 ∈ E∩E ′ , then G0 = gG0 ·G0 ∈ E ′′ ; if E and E ′ are finite, then |E ′′ | ≤ |E|·|E ′ |;
and if G0 · E ′ ⊆ E ′ , then E ′′ is as stated.
⊓
⊔
Right Amenable Left Group Sets and the Tarski-Følner Theorem
13
Main Theorem 4. Let R = ((M, G, ⊲), (m0 , {gm0 ,m }m∈M )) be a cell space
such that the stabiliser G0 of m0 under ⊲ is finite. The following statements
are equivalent:
1. The cell space R is not right amenable;
2. There is no right Følner net in R;
3. There is a finite subset E of G/G0 such that G0 ∈ E and, for each finite
subset F of M , we have |F P E| ≥ 2|F |;
4. There is a 2-to-1 surjective map φ : M → M and there is a finite subset E
of G/G0 such that
∀ m ∈ M ∃ e ∈ E : φ(m) P e = m;
5. There is a right paradoxical decomposition of R.
Proof. 1 implies 2. Let there be a right Følner net {Fi }i∈I in R. Furthermore,
let i ∈ I. Put
νi : ℓ∞ (M ) → R,
1 X
f (m).
f 7→
|Fi |
m∈Fi
Then, νi ∈ M(M ). Moreover, let g ∈ G/G0 and let f ∈ ℓ∞ (M ). Then,
(νi g)(f ) = νi (f g)
1 X
=
(f g)(m)
|Fi |
m∈Fi
X
1 X
=
|Fi |
−1
′
m∈Fi m ∈(_Pg)
1
=
|Fi |
X
f (m′ )
(m)
f (m).
m∈(_Pg)−1 (Fi )
Hence,
1
(νi g − νi )(f ) =
|Fi |
X
m∈(_Pg)−1 (Fi )rFi
f (m) −
X
m∈Fi r(_Pg)−1 (Fi )
Therefore,
f (m).
1
|f (m)|
|f (m)| +
|Fi |
−1
−1
P
P
m∈(_ g) (Fi )rFi
m∈Fi r(_ g) (Fi )
|(_ P g)−1 (Fi ) r Fi | |Fi r (_ P g)−1 (Fi )|
· kf k∞ .
+
≤
|Fi |
|Fi |
|(νi g − νi )(f )| ≤
X
X
14
Right Amenable Left Group Sets and the Tarski-Følner Theorem
According to Definition 14 and Lemma 13, the nets {|(_Pg)−1 (Fi )rFi |/|Fi |}i∈I
and {|Fi r (_ P g)−1 (Fi )|/|Fi |}i∈I converge to 0. Hence, so does {|(νi g −
νi )(f )|}i∈I . Thus, the net {νi g− νi }i∈I converges to 0 in ℓ∞ (M )∗ equipped
with the weak-∗ topology. Hence, according to Lemma 17, the cell space R is
right amenable. In conclusion, by contraposition, if R is not right amenable,
then there is no right Følner net in R.
2 implies 3. Let there be no right Følner net in R. According to Lemma 9,
there is a finite E1 ⊆ G/G0 and an ε ∈ R>0 such that, for each non-empty
and finite F ⊆ M , there is an eF ∈ E1 such that
|F r (_ P eF )−1 (F )|
≥ ε.
|F |
Put E2 = {G0 } ∪ E1 .
Let F ⊆ M be non-empty and finite. Then, F ⊆ F ∪ (F
Thus,
|F
P
E2 | − |F | = |(F
= |(F
≥ |(F
P
P
P E2 |
− |F | ≥
E2 .
E2 ) r F |
E1 ) r F |
P eF ) r F |.
P
eF ) r F )| ≤
|(_ P eF )−1 ((F P eF ) r F )|
.
|G0 |
Therefore, because F r (_ P eF )−1 (F ) ⊆ (_ P eF )−1 ((F
|F
P
P
Moreover, according to Lemma 1, we have |(_ P eF )−1 ((F
|G0 | · |(F P eF ) r F |. Hence,
|F
E1 ) = F
P E2 | − |F |
P eF ) r F ),
|F r (_ P eF )−1 (F )|
|G0 |
ε
≥
|F |.
|G0 |
≥
Put ξ = 1 + ε/|G0 |. Then, |F P E2 | ≥ ξ|F |. Because ε does not depend on
F , neither does ξ. Therefore, for each non-empty and finite F ⊆ M , we have
|F P E2 | ≥ ξ|F |.
Let F ⊆ M be non-empty and finite. Because ξ > 1, there is an n ∈ N such
that ξ n ≥ 2. Hence,
|(((F
|
P
E2 ) P · · · ) P E2 ) P E2 | ≥ ξ|((F
{z
}
P
E2 ) P · · · ) P E2 |
n times
≥ ···
≥ ξ n |F |
≥ 2|F |.
Moreover, according to Lemma 18, there is an E ⊆ G/G0 such that E is
finite, G0 ∈ E, and F P E = (((F P E2 ) P · · · ) P E2 ) P E2 . In conclusion,
|F P E| ≥ 2|F |.
Right Amenable Left Group Sets and the Tarski-Følner Theorem
M
15
M
Nr (F ) = F
F
P
E
F
_Pe
m
Nl (F ′ )
=
{m ∈ M | (m P E) ∩ F =
6 ∅}
F′
F′
_ P e′
m′
Figure 1: Schematic representation of the set-up of the proof of Theorem 4, Item 3
implies Item 4: Each region enclosed by one of the two rectangles is M ; the
regions enclosed by the smaller circles with solid borders are subsets F and F ′
of M respectively; the regions enclosed by the circles with dashed borders are
m P E andSm′ P E respectively; the region enclosed by the circle with dotted
border is e′ ∈E (_ P e′ )−1 (m′ ); the regions enclosed by the larger circles are
Nr (F ) and Nl (F ′ ) respectively.
3 implies 4 (see Fig. 1). Let there be a finite E ⊆ G/G0 such that, for each
finite F ⊆ M , we have |F P E| ≥ 2|F |. Furthermore, let G be the bipartite
graph
(M, M, {(m, m′ ) ∈ M × M | ∃ e ∈ E : m P e = m′ }).
Moreover, let F ⊆ M be finite. The right neighbourhood of F in G is
Nr (F ) = {m′ ∈ M | ∃ e ∈ E : F
Pe
∋ m′ } = F
P
E
and the left neighbourhood of F in G is
Nl (F ) = {m ∈ M | ∃ e ∈ E : m P e ∈ F } =
[
(_ P e)−1 (F ).
e∈E
By precondition |Nr (F )| = |F P E| ≥ 2|F |. Moreover, because G0 ∈ E,
we have F = (_ P G0 )−1 (F ) ⊆ Nl (F ) and hence |Nl (F )| ≥ |F | ≥ 2−1 |F |.
Therefore, according to the Hall harem theorem, there is a perfect (1, 2)matching for G. In conclusion, there is a 2-to-1 surjective map φ : M → M
16
Right Amenable Left Group Sets and the Tarski-Følner Theorem
M
M
_Pe
ψ
ψ(Ae ) = Ae P e
ψ(M )
φ
m
Ae
Be′
φ
ψ ′ (Be′ ) = Be′
P e′
ψ′
_ P e′
ψ ′ (M )
Figure 2: Schematic representation of the set-up of the proof of Theorem 4, Item 4
implies Item 5: Each region enclosed by one of the three columns, two with
solid and one with dashed border, is M ; the dot in the right column, called m,
is an element of M , and the two dots in the left column are its preimages under
φ, which are its images under ψ and ψ ′ ; there are elements e and e′ of E such
that m P e = ψ(m) and m P e′ = ψ ′ (m), in other words, m ∈ Ae and m ∈ Be′ ;
as depicted in the right columns with solid and dashed borders, the families
{Ae }e∈E and {Be′ }e′ ∈E are partitions of M ; as depicted in the left column, the
set {ψ(M ), ψ ′ (M )} is a partition of M , the family {ψ(Ae )}e∈E = {Ae P e}e∈E
is a partition of ψ(M ), and the family {ψ ′ (Be′ )}e′ ∈E = {Be′ P e′ }e′ ∈E is a
partition of ψ ′ (M ).
such that, for each m ∈ M , we have (φ(m), m) ∈ E, that is, there is an
e ∈ E such that φ(m) P e = m.
4 implies 5 (see Fig. 2). Let there be a 2-to-1 surjective map φ : M → M
and a finite subset E of G/G0 such that
∀ m ∈ M ∃ e ∈ E : φ(m) P e = m.
By the axiom of choice, there are two injective maps ψ and ψ ′ : M → M
such that, for each m ∈ M , we have φ−1 (m) = {ψ(m), ψ ′ (m)}. For each
e ∈ E, let
Ae = {m ∈ M | m P e = ψ(m)}
and Be = {m ∈ M | m P e = ψ ′ (m)}.
Right Amenable Left Group Sets and the Tarski-Følner Theorem
17
Let m ∈ M . There is an e ∈ E such that φ(ψ(m)) P e = ψ(m). Because
φ(ψ(m)) = m, we have m ∈ Ae . And, because P is free, for each e′ ∈ E r{e},
we have m P e′ 6= m P e = ψ(m) and thus m ∈
/ Ae′ . Therefore,
[
[
M = · Ae and analogously M = · Be .
e∈E
e∈E
′
Moreover, ψ(Ae ) = Ae P e and ψ (Be ) = Be P e. Hence, because M =
ψ(M ) ∪· ψ ′ (M ), and ψ and ψ ′ are injective,
!
!
!
!
[
[
[
[
′
M=
· ψ(Ae ) ∪·
· ψ (Be ) =
· Ae P e ∪·
· Be P e .
e∈E
e∈E
e∈E
e∈E
′
Furthermore, because ψ and ψ are injective, for each e ∈ E, the maps
(_ P e)↾Ae = ψ↾Ae and (_ P e)↾Be = ψ ′ ↾Be are injective. In conclusion,
(N, {Ae }e∈E , {Be }e∈E ) is a right paradoxical decomposition of R.
5 implies 1. Let there be a right paradoxical decomposition (N, {Ae }e∈E , {Be }e∈E )
of R. According to Lemma 15,
X
X
X
X
1M =
1 Ae =
1Be =
(1Ae e) +
(1Be e).
e∈E
e∈E
e∈E
e∈E
Suppose that R is right amenable. Then, according to Theorem 3, there is
a -invariant mean ν on M . Because ν is linear and normalised,
1 = ν(1M )
X
=
ν(1Ae
e) +
e∈E
=
X
X
e∈E
ν(1Be
e)
e∈E
(ν e)(1Ae ) +
X
(ν e)(1Be )
e∈E
e∈E
=
X
ν(1Ae ) +
X
ν(1Be )
e∈E
= ν(1M ) + ν(1M )
=1+1
= 2,
which contradicts that 1 6= 2. In conclusion, R is not right amenable.
⊓
⊔
Corollary 1 (Tarski alternative theorem; Alfred Tarski, 1938). Let M
be a left homogeneous space with finite stabilisers. It is right amenable if and only
if there is a coordinate system K for M such that there is no right paradoxical
decomposition of (M, K).
⊓
⊔
Corollary 2 (Theorem of Følner; Erling Følner, 1955). Let M be a left
homogeneous space with finite stabilisers. It is right amenable if and only if there
is a coordinate system K for M such that there is a right Følner net in (M, K).
⊓
⊔
Remark 7. In the situation of Remark 1, Corollaries 1 and 2 constitute [1, Theorem 4.9.1].
18
6
Right Amenable Left Group Sets and the Tarski-Følner Theorem
From Left to Right Amenability
Lemma 19. Let R = ((M, G, ⊲), (m0 , {gm0 ,m }m∈M )) be a cell space and let H
be a subgroup of G such that, for each element g ∈ G/G0 , there is an element
h ∈ H such that the maps _ P g and h ⊲ _ are inverse to each other. If (M, H,
⊲↾H×M ) is left amenable, then R is right amenable.
Proof. Let µ ∈ PM(M ). Furthermore, let g ∈ G/G0 . There is an h ∈ H such
that _ P g and h ⊲ _ are inverse to each other. Moreover, let A ⊆ M . Because
_ P g = (h ⊲ _)−1 = h−1 ⊲ _, we have A P g = h−1 ⊲ A. Therefore,
(µ g)(A) = µ(A P g)
= µ(h−1 ⊲ A)
= (h µ)(A).
Thus, µ g = h µ. Hence, if µ is ↾H×[0,1]P(M ) -invariant, then µ is -semiinvariant. In conclusion, if (M, H, ⊲↾H×M ) is left amenable, then R is right
amenable.
⊓
⊔
Lemma 20. Let R = ((M, G, ⊲), (m0 , {gm0 ,m }m∈M )) be a cell space and let H
be a subgroup of G such that G = G0 H, for each element g ∈ G/G0 , the map
_ P g is injective,
∀ h ∈ H : _ P hG0 = h ⊲ _,
and
∀ h ∈ H ∀ g ∈ G/G0 : (_ P hG0 ) P g = _ P h · g.
If (M, H, ⊲↾H×M ) is left amenable, then R is right amenable.
Proof. Let gG0 ∈ G/G0 . Because g −1 ∈ G = G0 H, there is a g0 ∈ G0 and there
is an h ∈ H such that g −1 = g0 h. Thus, h = g0−1 g −1 ∈ H. Hence, for each
m ∈ M,
((_ P gG0 ) ◦ (h ⊲ _))(m) = (h ⊲ m) P gG0
= (m P hG0 ) P gG0
= m P hgG0
= m P g0−1 g −1 gG0
= m P G0
= m.
Therefore, h⊲ _ is right inverse to _ P gG0 . Hence, _ P gG0 is surjective and thus,
because it is injective by precondition, bijective. Therefore, _ P gG0 and h⊲_ are
inverse to each other. In conclusion, according to Lemma 19, if (M, H, ⊲↾H×M )
is left amenable, then R is right amenable.
⊓
⊔
Definition 17. Let G be a group. The set
Z(G) = {z ∈ G | ∀ g ∈ G : zg = gz}
is called centre of G.
Right Amenable Left Group Sets and the Tarski-Følner Theorem
Lemma 21. Let G be a group. The centre of G is a subgroup of G.
19
⊓
⊔
Lemma 22. Let R = ((M, G, ⊲), (m0 , {gm0 ,m }m∈M )) be a cell space and let H
be a subgroup of G such that G is equal to G0 H, ⊲↾H×M is free, and {gm0 ,m }m∈M
is included in Z(H). If (M, H, ⊲↾H×M ) is left amenable, then R is right amenable.
Proof. Let g ∈ G. For each m ∈ M ,
m P gG0 = gm0 ,m g ⊲ m0 = gm0 ,m ⊲ (g ⊲ m0 ).
Let m ∈ M . For each m′ ∈ M , because ⊲↾Z(H)×M is free and gm0 ,m , gm0 ,m′ ∈
Z(H),
m′ P gG0 = m P gG0 ⇐⇒ gm0 ,m′ = gm0 ,m
⇐⇒ m′ = m.
Therefore, _ P gG0 is injective.
Let m ∈ M and let h ∈ H. Because gm0 ,m ∈ Z(G),
m P hG0 = gm0 ,m h ⊲ m0
= hgm0 ,m ⊲ m0
= h ⊲ m.
Put m′ = m P hG0 . Then,
gm0 ,m h ⊲ m0 = hgm0 ,m ⊲ m0
=h⊲m
= m′ .
Hence, because gm0 ,m′ ⊲ m0 = m′ also and ⊲↾H×M is free, gm0 ,m′ = gm0 ,m h.
Therefore,
(m P hG0 ) P gG0 = m′ P gG0
= gm0 ,m′ g ⊲ m0
= gm0 ,m hg ⊲ m0
= m P hgG0 .
In conclusion, according to Lemma 20, if (M, H, ⊲↾H×M ) is left amenable, then
R is right amenable.
⊓
⊔
Example 1. Let M = K be a field, let
G = {f : M → M, x 7→ ax + b | a, b ∈ M, a 6= 0}
be the group of affine functions with composition as group multiplication, and
let
H = {f : M → M, x 7→ x + b | b ∈ M }
20
Right Amenable Left Group Sets and the Tarski-Følner Theorem
be the group of translations also with composition as group multiplication. The
group H is an abelian subgroup of G, which in turn is a non-abelian subgroup
of the symmetry group of M . Moreover, according to [1, Example 4.6.2 and
Theorem 4.6.3], the group G is left amenable and hence, according to [1, Proposition 4.5.1], so is its subgroup H. Furthermore, the group G acts transitively
on M by function application by ⊲ and so does H by ⊲↾H×M , even freely so. Because the groups G and H are left amenable, so are the left group sets (M, G, ⊲)
and (M, H, ⊲↾H×M ). The stabiliser of m0 = 0 is the group of dilations
G0 = {f : M → M, x 7→ ax | a ∈ M r {0}}.
We have G = G0 H. For each m ∈ M , let
gm0 ,m : M → M,
x 7→ x + m,
be the translation by m. Then, {gm0 ,m }m∈M is included in Z(H) = H. Hence,
according to Lemma 22, the cell space R = ((M, G, ⊲), (m0 , {gm0 ,m }m∈M )) is
right amenable.
Lemma 23. Let H and N be two groups, let φ : H → Aut(N ) be a group homomorphism, let G be the Cartesian product H × N , and let
· : G × G → G,
((h, n), (h′ , n′ )) 7→ (hh′ , nφ(h)(n′ )).
The tuple (G, ·) is a group, called semi-direct product of H and N with respect
to φ, and denoted by H ⋉φ N .
⊓
⊔
Lemma 24. Let G be a semi-direct product of H and N with respect to φ. The
neutral element of G is (eH , eN ) and, for each element (h, n) ∈ G, the inverse
of (h, n) is (h−1 , φ(h−1 )(n)).
⊓
⊔
Definition 18. Let (M, G, ⊲) be a left homogeneous space. It is called principal
if and only if the action ⊲ is free.
Lemma 25. Let (M, H, ⊲H ) be a principal left homogeneous space. Furthermore, let G0 be a group, let φ : G0 → Aut(H) be a group homomorphism, let
m0 be an element of M , for each element m ∈ M , let hm0 ,m be the unique
element of H such that hm0 ,m ⊲ m0 = m, and let
⊲G0 : G0 × M → M,
(g0 , m) 7→ φ(g0 )(hm0 ,m ) ⊲H m0 .
Moreover, let G be the semi-direct product of G0 and H with respect to φ, and
let
⊲ : G × M → M,
Right Amenable Left Group Sets and the Tarski-Følner Theorem
21
((g0 , h), m) 7→ h ⊲H (g0 ⊲G0 m).
The triple (M, G0 , ⊲G0 ) is a left group set and the group G0 is the stabiliser of m0
under ⊲G0 . Furthermore, the tuple R = ((M, G, ⊲), (m0 , {(eG0 , hm0 ,m )}m∈M )) is
a cell space and the group G0 × {eH } is the stabiliser of m0 under ⊲. Moreover,
under the identification of G0 with G0 × {eH } and of H with {eG0 } × H, the left
group sets (M, G0 , ⊲G0 ) and (M, H, ⊲H ) are left group subsets of (M, G, ⊲).
Proof. Because φ(eG0 ) = idAut(H) , for each m ∈ M ,
eG0 ⊲G0 m = φ(eG0 )(hm0 ,m ) ⊲H m0
= hm0 ,m ⊲H m0
= m.
Let g0 and g0′ ∈ G0 , and let m ∈ M . Because ⊲H is free and hm0 ,φ(g0′ )(hm0 ,m )⊲H m0 ⊲H
m0 = φ(g0′ )(hm0 ,m )⊲H m0 , we have hm0 ,φ(g0′ )(hm0 ,m )⊲H m0 = φ(g0′ )(hm0 ,m ). Therefore,
g0 g0′ ⊲G0 m = φ(g0 g0′ )(hm0 ,m ) ⊲H m0
= (φ(g0 ) ◦ φ(g0′ ))(hm0 ,m ) ⊲H m0
= φ(g0 )(φ(g0′ )(hm0 ,m )) ⊲H m0
= φ(g0 )(hm0 ,φ(g0′ )(hm0 ,m )⊲H m0 ) ⊲H m0
= g0 ⊲G0 (φ(g0′ )(hm0 ,m ) ⊲H m0 )
= g0 ⊲G0 (g0′ ⊲G0 m).
In conclusion, (M, G0 , ⊲G0 ) is a left group set.
Because hm0 ,m0 = eH , for each g0 ∈ G0 ,
g0 ⊲G0 m0 = φ(g0 )(eH ) ⊲H m0
= e H ⊲ H m0
= m0 .
In conclusion, G0 is the stabiliser of m0 under ⊲G0 .
For each m ∈ M ,
(eG0 , eH ) ⊲ m = eH ⊲H (eG0 ⊲G0 m) = m.
Let g0 ∈ G0 , let h ∈ H, and let m ∈ M . Because hhm0 ,m ⊲H m0 = h ⊲H m, we
have hhm0 ,m = hm0 ,h⊲H m . Hence,
φ(g0 )(h) ⊲H (g0 ⊲G0 m) = φ(g0 )(h) ⊲H (φ(g0 )(hm0 ,m ) ⊲H m0 )
= φ(g0 )(h)φ(g0 )(hm0 ,m ) ⊲H m0
= φ(g0 )(hhm0 ,m ) ⊲H m0
= φ(g0 )(hm0 ,h⊲H m ) ⊲H m0
= g0 ⊲G0 (h ⊲H m).
22
Right Amenable Left Group Sets and the Tarski-Følner Theorem
Therefore, for each g0 ∈ G0 , each g0′ ∈ G0 , each h ∈ H, each h′ ∈ H, and each
m ∈ M,
(g0 , h)(g0′ , h′ ) ⊲ m = (g0 g0′ , hφ(g0 )(h′ )) ⊲ m
= hφ(g0 )(h′ ) ⊲H (g0 g0′ ⊲G0 m)
= h ⊲H φ(g0 )(h′ ) ⊲H g0 ⊲G0 (g0′ ⊲G0 m)
= h ⊲H g0 ⊲G0 h′ ⊲H (g0′ ⊲G0 m)
= (g0 , h) ⊲ h′ ⊲H (g0′ ⊲G0 m)
= (g0 , h) ⊲ (g0′ , h′ ) ⊲ m .
In conclusion, (M, G, ⊲) is a left group action.
Because ⊲H is transitive and, for each h ∈ H and each m ∈ M , we have
(eG0 , h) ⊲ m = h ⊲ m, the left group action ⊲ is transitive and hence M = (M,
G, ⊲) is a left homogeneous space. Moreover, because, for each m ∈ M ,
(eG0 , hm0 ,m ) ⊲ m0 = hm0 ,m ⊲H (eG0 ⊲G0 m0 )
= hm0 ,m ⊲H m0
= m,
the tuple K = (m0 , {(eG0 , hm0 ,m )}m∈M ) is a coordinate system for M. Therefore, R = (M, K) is a cell space.
Because G0 is the stabiliser of m0 under ⊲G0 , for each (g0 , h) ∈ G, we have
(g0 , h) ⊲ m0 = h ⊲H (g0 ⊲ m0 ) = h ⊲ m0 . Because ⊲H is free, G0 × {eH } is the
stabiliser of m0 under ⊲.
Under the identification of G0 with G0 × {eH } and of H with {eG0 } × H, we
⊓
⊔
have ⊲↾G0 ×M = ⊲G0 and ⊲↾H×M = ⊲H .
Corollary 3. In the situation of Lemma 25, let H be abelian. The cell space R
is right amenable.
Proof. According to [1, Theorem 4.6.1], because H is abelian, it is left amenable.
Therefore, (M, H, ⊲H ) is left amenable. Identify G0 with G0 × {eH } and identify
H with {eG0 }×H. Then, H is a subgroup of G, and G = G0 H, and ⊲↾H×M = ⊲H
is free, and, for each m ∈ M , we have (eG0 , hm0 ,m ) ∈ H = Z(H). Hence,
according to Lemma 22, the cell space R is right amenable.
⊓
⊔
Example 2. Let d be a positive integer; let E be the d-dimensional Euclidean
group, that is, the symmetry group of the d-dimensional Euclidean space, in
other words, the isometries of Rd with respect to the Euclidean metric with
function composition; let T be the d-dimensional translation group; and let O be
the d-dimensional orthogonal group. The group T is abelian, a normal subgroup
of E, and isomorphic to Rd with addition; the group O is isomorphic to the
quotient E/T and to the (d × d)-dimensional orthogonal matrices with matrix
multiplication; the group E is isomorphic to the semi-direct product O ⋉ι T ,
Right Amenable Left Group Sets and the Tarski-Følner Theorem
23
where ι : O → Aut(Rd ) is the inclusion map. The groups T , O, and E act on Rd
on the left by function application, denoted by ⊲T , ⊲O , and ⊲, respectively; under
the identification of T with Rd by t 7→ [v 7→ v + t], of O with the orthogonal
matrices of Rd×d by A 7→ [v 7→ Av], and of E with O ⋉ι T by (A, t) 7→ [v 7→
Av + t], we have
⊲ T : T × Rd → Rd ,
(t, v) 7→ v + t,
and
⊲ O : O × Rd → Rd ,
(A, v) 7→ Av,
and
⊲ : E × Rd → Rd ,
((A, t), v) 7→ Av + t,
and
ι : O → Aut(Rd ),
A 7→ [v 7→ Av].
Hence, for each vector v ∈ Rd , we have v ⊲T 0 = v, therefore, ⊲O = [(A, v) 7→
ι(A)(v) ⊲T 0], and thus ⊲ = [((A, t), v) 7→ t ⊲T (A ⊲O v)]. Moreover, because
the group (T, ◦) ∼
= (Rd , +) is abelian, according to [1, Theorem 4.6.1], it is
left amenable and so is (Rd , Rd , +) ∼
= (Rd , T, ⊲). In conclusion, according to
d
Corollary 3, the cell space ((R , E, ⊲), (0, {−v}v∈Rd )) is right amenable.
References
1. Ceccherini-Silberstein, Tullio and Michel Coornaert. Cellular Automata and Groups.
In: Springer Monographs in Mathematics. Springer-Verlag, 2010.
2. Følner, Erling. On groups with full Banach mean value. In: Mathematica Scandinavica 3 (1955), pages 245-254.
3. Tarski, Alfred. Algebraische Fassung des Maßproblems. In: Fundamenta Mathematicae 31.1 (1938), pages 47-66.
4. von Neumann, John. Zur allgemeinen Theorie des Maßes. In: Fundamenta Mathematicae 13.1 (1929), pages 73–111.
5. Simon Wacker. Cellular Automata on Group Sets and the Uniform Curtis-HedlundLyndon Theorem. In: Cellular Automata and Discrete Complex Systems (2016), to
be published. arXiv:1603.07271 [math.GR].
6. Simon Wacker. The Garden of Eden Theorem for Cellular Automata on Group Sets.
arXiv:1603.07272 [math.GR].
24
Right Amenable Left Group Sets and the Tarski-Følner Theorem
In Appendix A we present the basic theory of topologies. In Appendix B we
present the basic theory of nets. In Appendix C we introduce initial and product
topologies. In Appendix D we introduce the notion of compactness for topological
spaces. In Appendix E we introduce topological dual spaces of topological spaces.
In Appendix F we state Hall’s marriage and harem theorems. And in Appendix G
we state Zorn’s lemma.
A
Topologies
The theory of topologies and nets as presented here may be found in more detail
in Appendix A in the monograph ‘Cellular Automata and Groups’[1].
Definition 19. Let X be a set and let T be a set of subsets of X. The set T is
called topology on X if, and only if
1. {∅, X} is a subset of T ,
S
2. for each family {Oi }i∈I of elements in T , the union i∈I Oi is an element
of T ,
T
3. for each finite family {Oi }i∈I of elements in T , the intersection i∈I Oi is
an element of T .
Definition 20. Let X be a set, and let T and T ′ be two topologies on X. The
topology T is called
1. coarser than T ′ if, and only if T ⊆ T ′ ;
2. finer than T ′ if, and only if T ⊇ T ′ .
Definition 21. Let X be a set and let T be a topology on X. The tuple (X, T )
is called topological space, each subset O of X with O ∈ T is called open in X,
each subset A of X with X r A ∈ T is called closed in X, and each subset U of
X that is both open and closed is called clopen in X.
The set X is said to be equipped with T if, and only if it shall be implicitly
clear that T is the topology on X being considered. The set X is called topological
space if, and only if it is implicitly clear what topology on X is being considered.
Example 3. Let X be a set. The set P(X) is the finest topology on X. Itself as
well as the topological space (X, P(X)) are called discrete.
Definition 22. Let (X, T ) be a topological space, let x be a point of X, and let
N be a subset of X. The set N is called neighbourhood of x if, and only if there
is an open subset O of X such that x ∈ O and O ⊆ N .
Definition 23. Let (X, T ) be a topological space and let x be a point of X. The
set of all open neighbourhoods of x is denoted by Tx .
Right Amenable Left Group Sets and the Tarski-Følner Theorem
B
25
Nets
Definition 24. Let I be a set and let ≤ be a binary relation on I. The relation
≤ is called preorder on I and the tuple (I, ≤) is called preordered set if, and
only if the relation ≤ is reflexive and transitive.
Definition 25. Let ≤ be a preorder on I. It is called directed and the preordered
set (I, ≤) is called directed set if, and only if
∀ i ∈ I ∀ i′ ∈ I ∃ i′′ ∈ I : i ≤ i′′ ∧ i′ ≤ i′′ .
Definition 26. Let ≤ be a preorder on I, let J be a subset of I, and let i be an
element of I. The element i is called upper bound of J in (I, ≤) if, and only if
∀ i′ ∈ J : i′ ≤ i.
Definition 27. Let M be a set, let I be a set, and let f : I → M be a map. The
map f is called family of elements in M indexed by I and denoted by {mi }i∈I ,
where, for each index i ∈ I, mi = f (i).
Definition 28. Let I be a set, let ≤ be a binary relation on I, and let {mi }i∈I
be a family of elements in M indexed by I. The family {mi }i∈I is called net in
M indexed by (I, ≤) if, and only if the tuple (I, ≤) is a directed set.
Definition 29. Let {mi }i∈I and {m′j }j∈J be two nets in M . The net {m′j }j∈J
is called subnet of {mi }i∈I if, and only if there is a map f : J → I such that
{m′j }j∈J = {mf (j) }j∈J and
∀ i ∈ I ∃ j ∈ J : ∀ j ′ ∈ J : (j ′ ≥ j =⇒ f (j ′ ) ≥ i).
Definition 30. Let (X, T ) be a topological space, let {xi }i∈I be a net in X
indexed by (I, ≤), and let x be a point of X. The net {xi }i∈I is said to converge
to x and x is called limit point of {xi }i∈I if, and only if
∀ O ∈ Tx ∃ i0 ∈ I : ∀ i ∈ I : (i ≥ i0 =⇒ xi ∈ O).
Definition 31. Let (X, T ) be a topological space and let {xi }i∈I be a net in X
indexed by (I, ≤). The net {xi }i∈I is called convergent if, and only if there is a
point x ∈ X such that it converges to x.
Remark 8. Let {mi }i∈I be a net that converges to x. Each subnet {m′j }j∈J of
{mi }i∈I converges to x.
Lemma 26. Let (X, T ) be a topological space, let Y be a subset of X, and let
x be an element of X. Then, x ∈ Y if, and only if there is a net {yi }i∈I in Y
that converges to x.
Proof. See Proposition A.2.1 in ‘Cellular Automata and Groups’[1].
⊓
⊔
26
Right Amenable Left Group Sets and the Tarski-Følner Theorem
Lemma 27. Let (X, T ) be a topological space. It is Hausdorff if, and only if
each convergent net in X has exactly one limit point.
Proof. See Proposition A.2.2 in ‘Cellular Automata and Groups’[1].
⊓
⊔
Definition 32. Let (X, T ) be a Hausdorff topological space, let {xi }i∈I be a
convergent net in X indexed by (I, ≤), and let x be the limit point of {xi }i∈I .
The point x is denoted by limi∈I xi and we write xi → x.
i∈I
Definition 33. Let (X, T ) be a topological space, let {xi }i∈I be a net in X
indexed by (I, ≤), and let x be an element of X. The point x is called cluster
point of {xi }i∈I if, and only if
∀ O ∈ Tx ∀ i ∈ I ∃ i′ ∈ I : (i′ ≥ i ∧ xi′ ∈ O).
Lemma 28. Let (X, T ) be a topological space, let {xi }i∈I be a net in X indexed
by (I, ≤), and let x be an element of X. The point x is a cluster point of {xi }i∈I
if, and only if there is a subnet of {xi }i∈I that converges to x.
Proof. See Proposition A.2.3 in ‘Cellular Automata and Groups’[1].
⊓
⊔
Lemma 29. Let (X, T ) and (X ′ , T ′ ) be two topological spaces, let f be a continuous map from X to X ′ , let {xi }i∈I be a net in X, and let x be an element
of X.
1. If x is a limit point of {xi }i∈I , then f (x) is a limit point of {f (xi )}i∈I .
2. If x is a cluster point of {xi }i∈I , then f (x) is a cluster point of {f (xi )}i∈I .
Proof. Confer the last paragraph of Sect. A.2 in ‘Cellular Automata and Groups’[1].
⊓
⊔
Definition 34. Let R = R ∪ {−∞, +∞} be the affinely extended real numbers
and let {ri }i∈I be a net in R indexed by (I, ≤).
1. The limit of the net {inf i′ ≥i ri′ }i∈I is called limit inferior of {ri }i∈I and
denoted by lim inf i∈I ri .
2. The limit of the net {supi′ ≥i ri′ }i∈I is called limit superior of {ri }i∈I and
denoted by lim supi∈I ri .
C
Initial and Product Topologies
Definition 35. Let X be a set, let I be a set, and, for each index i ∈ I, let
(Yi , Ti ) be a topological space and let fi be a map from X to Yi . The coarsest
topology on X such that, for each index i ∈ I, the map fi is continuous, is called
initial with respect to {fi }i∈I .
Lemma 30. Let (X, T ) be a topological space, where T is the initial topology
with respect to {fi : X → Yi }i∈I , let (Z, S) be a topological space, and let g be a
map from Z to X. The map g is continuous if, and only if, for each index i ∈ I,
the map fi ◦ g is continuous.
⊓
⊔
Right Amenable Left Group Sets and the Tarski-Følner Theorem
27
Lemma 31. Let (X, T ) be a topological space, where T is the initial topology
with respect to {fi : X → Yi }i∈I , let {xi′ }i′ ∈I ′ be a net in X, and let x be a point
in X. The point x is a limit point or cluster point of {xi′ }i′ ∈I ′ if, and only if,
for each index i ∈ I, the point fi (x) is a limit point or cluster point, respectively,
of {fi (xi′ )}i′ ∈I ′ .
Proof. Confer the last paragraph of Sect. A.3 in ‘Cellular Automata and Groups’[1].
⊓
⊔
Definition
36. Let {(Xi , Ti )}i∈I be a family of topological spaces, let X be the
Q
set i∈I Xi , and, for each index i ∈ I, let πi be the projection of X onto Xi .
The initial topology on X with respect to {πi }i∈I is called product.
Q
Remark 9. The product topology on X has for a base the sets i∈I Oi , where, for
each index i ∈ I, the set Oi is an open subset of Xi , and the set {i ∈ I | Oi 6= Xi }
is finite.
Definition 37. Q
Let {(Xi , Ti )}i∈I be a family of discrete topological spaces and
let X be the set i∈I Xi . The product topology on X is called prodiscrete.
Lemma
32. Let {(Xi , Ti )}i∈I be a family of Hausdorff topological spaces. The
Q
set i∈I Xi , equipped with the product topology, is Hausdorff.
Proof. See Proposition A.4.1 in ‘Cellular Automata and Groups’[1].
⊓
⊔
Definition 38. Let X be a topological space. It is called totally disconnected if,
and only if, for each non-empty and connected subset A of X, we have |A| = 1.
Lemma 33. LetQ{(Xi , Ti )}i∈I be a family of totally disconnected topological
spaces. The set i∈I Xi , equipped with the product topology, is totally disconnected.
Proof. See Proposition A.4.2 in ‘Cellular Automata and Groups’[1].
⊓
⊔
Lemma 34. Let {(Xi , Ti )}i∈I be a family of topological
Q spaces and, for each
index
i
∈
I,
let
A
be
a
closed
subset
of
X
.
The
set
i
i
i∈I Ai is a closed subset
Q
of i∈I Xi , equipped with the product topology.
Proof. See Proposition A.4.3 in ‘Cellular Automata and Groups’[1].
D
⊓
⊔
Compactness
Definition 39. Let (X, T ) be a topological space and let {Oi }i∈I be a family
of elements of T . The family {Oi }i∈I is called open cover of X if, and only if,
S
i∈I Oi = X.
Definition 40. Let (X, T ) be a topological space. It is called compact if, and
only if, for each open cover {Oi }i∈I of X, there is a finite subset J of I such
that {Oj }j∈J is an open cover of X.
28
Right Amenable Left Group Sets and the Tarski-Følner Theorem
Lemma 35. Let (X, T ) be a topological space. It is compact if, and only if, for
each
S subsets of X such that, for each finite subset J of
T family {Ai }i∈I of closed
I, j∈J Aj 6= ∅, we have i∈I Ai 6= ∅.
Proof. Confer first paragraph of Sect. A.5 in ‘Cellular Automata and Groups’[1].
⊓
⊔
Theorem 5. Let (X, T ) be a topological space. The following statements are
equivalent:
1. The space (X, T ) is compact;
2. Each net in X has a cluster point with respect to T ;
3. Each net in X has a convergent subnet with respect to T .
Proof. See Theorem A.5.1 in ‘Cellular Automata and Groups’[1].
⊓
⊔
Theorem 6 (Andrey Nikolayevich Tikhonov,
1935). Let {(Xi , Ti )}i∈I be
Q
a family of compact topological spaces. The set i∈I Xi , equipped with the product
topology, is compact.
Proof. See Theorem A.5.2 in ‘Cellular Automata and Groups’[1].
⊓
⊔
Corollary 4. Let {(Xi , Ti )}i∈I be a family of finite topological spaces. The set
Q
i∈I Xi , equipped with the product topology, is compact.
Proof. See paragraph before Corollary A.5.3 in ‘Cellular Automata and Groups’[1].
⊓
⊔
E
Dual Spaces
The theory of dual spaces as presented here may be found in more detail in
Appendix F in ‘Cellular Automata and Groups’[1].
In this section, let (X, k_k) be a normed R-vector space.
Definition 41. The vector space
X ∗ = {ψ : X → R linear | ψ is continuous}
is called topological dual space of X.
Definition 42. 1. The norm
k_kX ∗ : X ∗ → R,
ψ 7→
|ψ(x)|
,
x∈Xr{0} kxk
sup
is called operator norm on X ∗ .
2. The topology on X ∗ induced by k_kX ∗ is called strong topology on X ∗ .
Right Amenable Left Group Sets and the Tarski-Følner Theorem
29
Definition 43. Let x be an element of X. The map
evx : X ∗ → R,
ψ 7→ ψ(x),
is called evaluation map at x.
Definition 44. The initial topology on X ∗ with respect to {evx }x∈X is called
weak-∗ topology on X ∗ .
Lemma 36. Let {ψi }i∈I be a net in X ∗ , let ψ be an element of X ∗ , and let
X ∗ be equipped with the weak-∗ topology. The net {ψi }i∈I converges to ψ if, and
only if, for each element x of X, the net {ψi (x)}i∈I converges to ψ(x).
⊓
⊔
Lemma 37. The weak-∗ topology on X ∗ is coarser than the strong topology on
X ∗.
⊓
⊔
Corollary 5. Let {ψi }i∈I be a net in X ∗ that converges to ψ with respect to
the strong topology on X ∗ . The net {ψi }i∈I converges to ψ with respect to the
weak-∗ topology on X ∗ .
⊓
⊔
Definition 45. Let Y be a subset of X. The set Y is called convex if, and only
if,
∀(y, y ′ ) ∈ Y × Y ∀ t ∈ [0, 1] : ty + (1 − t)y ′ ∈ Y.
Definition 46. The topological vector space X is called locally convex if, and
only if, the origin has a neighbourhood base of convex sets.
Lemma 38. Let X ∗ be equipped with the weak-∗ topology and let ψ be an element
of X ∗ . A neighbourhood base of ψ is given by the sets
B(ψ, F, ε) = {ψ ′ ∈ X ∗ | ∀ x ∈ F : |ψ(x) − ψ ′ (x)| < ε},
for F ⊆ X finite and ε ∈ R>0 .⊓
⊔
Corollary 6. Let X ∗ be equipped with the weak-∗ topology. The space X ∗ is
locally convex.
⊓
⊔
Lemma 39. The space X ∗ , equipped with the weak-∗ topology, is Hausdorff.
Proof. Confer last paragraph of Sect. F.2 in ‘Cellular Automata and Groups’[1].
⊓
⊔
Theorem 7 (Stefan Banach, 1932; Leonidas Alaoglu, 1940). Let X ∗ be
equipped with the weak-∗ topology. The unit ball {ψ ∈ X ∗ | kψkX ∗ ≤ 1}, equipped
with the subspace topology, is compact.
Proof. See Theorem F.3.1 in ‘Cellular Automata and Groups’[1].
⊓
⊔
30
F
Right Amenable Left Group Sets and the Tarski-Følner Theorem
Hall’s Theorems
The theory concerning Hall’s theorems as presented here may be found in more
detail in Appendix H in the monograph ‘Cellular Automata and Groups’[1].
Definition 47. Let X and Y be two sets, and let E be a subset of X × Y .
The triple (X, Y, E) is called bipartite graph, each element x of X is called left
vertex, each element y of Y is called right vertex, and each element e of E is
called edge.
Definition 48. Let (X, Y, E) and (X ′ , Y ′ , E ′ ) be two bipartite graphs. The graph
(X, Y, E) is called bipartite subgraph of (X ′ , Y ′ , E ′ ) if, and only if, X ⊆ X ′ ,
Y ⊆ Y ′ , and E ⊆ E ′ .
Definition 49. Let (X, Y, E) be a bipartite graph, and let (x, y) and (x′ , y ′ ) be
two elements of E. The edges (x, y) and (x′ , y ′ ) are called adjacent if, and only
if, x = x′ or y = y ′ .
Definition 50. Let (X, Y, E) be a bipartite graph.
1. Let x be an element of X. The set
Nr (x) = {y ∈ Y | (x, y) ∈ E}
is called right neighbourhood of x.
S
2. Let A be a subset of X. The set a∈A Nr (a) is called right neighbourhood
of A and is denoted by Nr (A).
3. Let y be an element of Y . The set
Nl (y) = {x ∈ X | (x, y) ∈ E}
is called left neighbourhood of y.S
4. Let B be a subset of Y . The set b∈B Nl (b) is called left neighbourhood of
B and is denoted by Nl (B).
Definition 51. Let (X, Y, E) be a bipartite graph. It is called
1. finite if, and only if, the sets X and Y are finite.
2. locally finite if, and only if, for each element x of X, the set Nr (x) is finite,
and for each element y of Y , the set Nl (y) is finite.
Remark 10. Let (X, Y, E) be a locally finite bipartite graph. Then, for each finite
subset A of X, the set Nr (A) is finite; and, for each finite subset B of Y , the set
Nl (B) is finite.
Definition 52. Let (X, Y, E) be a bipartite graph and let M be a subset of E.
The set M is called matching if, and only if, for each element (e, e′ ) of M × M
with e 6= e′ , the edges e and e′ are non-adjacent.
Right Amenable Left Group Sets and the Tarski-Følner Theorem
31
Definition 53. Let (X, Y, E) be a bipartite graph and let M be a matching. The
matching M is called
1. left-perfect if, and only if,
∀ x ∈ X ∃ y ∈ Y : (x, y) ∈ M ;
2. right-perfect if, and only if,
∀ y ∈ Y ∃ x ∈ X : (x, y) ∈ M ;
3. perfect if, and only if, it is left-perfect and right-perfect.
Definition 54. Let (X, Y, E) be a locally finite bipartite graph. Is is said to
satisfy the
1. left Hall condition if, and only if,
∀ A ⊆ X finite : |Nr (A)| ≥ |A|;
2. right Hall condition if, and only if,
∀ B ⊆ Y finite : |Nl (B)| ≥ |B|;
3. Hall marriage conditions if, and only if, it satisfies the left and right Hall
conditions.
Theorem 8. Let (X, Y, E) be a locally finite bipartite graph. It satisfies the left
or right Hall condition if, and only if, there is a left- or right-perfect matching,
respectively.
Proof. See Theorem H.3.2 in ‘Cellular Automata and Groups’[1].
⊓
⊔
Theorem 9. Let (X, Y, E) be a bipartite graph such, that there is a left-perfect
matching and there is a right-perfect matching. There is a perfect matching.
Proof. See Theorem H.3.4 in ‘Cellular Automata and Groups’[1].
⊓
⊔
Corollary 7 (Georg Ferdinand Ludwig Philipp Cantor, Friedrich Wilhelm Karl Ernst Schröder, Felix Bernstein). Let X and Y be two sets
such, that there is an injective map f from X to Y and there is an injective map
g from Y to X. There is a bijective map from X to Y .
Proof. See Theorem H.3.5 in ‘Cellular Automata and Groups’[1].
⊓
⊔
Theorem 10 (Hall’s marriage theorem, Philip Hall, 1935). Let (X, Y, E)
be a locally finite bipartite graph. It satisfies the Hall marriage conditions if, and
only if, there is a perfect matching.
Proof. See Theorem H.3.6 in ‘Cellular Automata and Groups’[1].
⊓
⊔
32
Right Amenable Left Group Sets and the Tarski-Følner Theorem
Definition 55. Let X and Y be two sets, and let f be a surjective map from X
to Y . The map f is called k-to-1 if, and only if,
∀ y ∈ Y : |f −1 (y)| = k.
Definition 56. Let (X, Y, E) be a bipartite graph, let k be a positive integer,
and let M be a subset of E. The set M is called perfect (1, k)-matching if, and
only if,
∀ x ∈ X : |{y ∈ Y | (x, y) ∈ E}| = k
and
∀ y ∈ Y : |{x ∈ X | (x, y) ∈ E}| = 1.
Remark 11. The set M is a perfect (1, k)-matching if, and only if, there is a
k-to-1 surjective map ψ : Y → X such, that {(ψ(y), y) | y ∈ Y } = M .
Remark 12. The set M is a perfect (1, 1)-matching if, and only if, it is a perfect
matching.
Definition 57. Let (X, Y, E) be a locally finite bipartite graph and let k be a
positive integer. The graph (X, Y, E) is said to satisfy the Hall k-harem conditions if, and only if, for each finite subset A of X, we have |Nr (A)| ≥ k|A|, and
for each finite subset B of Y , we have |Nl (B)| ≥ k −1 |B|.
Theorem 11 (Hall’s harem theorem, Philip Hall). Let (X, Y, E) be a locally finite bipartite graph and let k be a positive integer. The graph (X, Y, E)
satisfies the Hall k-harem conditions if, and only if, there is a perfect (1, k)matching.
Proof. See Theorem H.4.2 in ‘Cellular Automata and Groups’[1].
G
⊓
⊔
Zorn’s Lemma
Definition 58. Let ≤ be a preorder on I. It is called partial order on I and the
preordered set (I, ≤) is called partially ordered set if, and only if the relation ≤
is antisymmetric.
Definition 59. Let ≤ be a partial order on I. It is called total order on I and
the partially ordered set (I, ≤) is called totally ordered set if, and only if the
relation ≤ is total.
Definition 60. Let ≤ be a preorder on I and let i be an element of I. The
element i is called maximal in (I, ≤) if, and only if
∀ i′ ∈ I : (i′ ≥ i =⇒ i′ ≤ i).
Definition 61. Let ≤ be a preorder on I and let J be a subset of I. The set J
is called chain in (I, ≤) if, and only if the restriction of ≤ to J is a total order
on J.
Lemma 40 (Zorn’s Lemma). Let (I, ≤) be a preordered set such that each
chain in I has an upper bound. Then, I has a maximal element.
⊓
⊔
| 4math.GR
|
A NOTE ON LOGARITHMIC SPACE STREAM ALGORITHMS FOR
MATCHINGS IN LOW ARBORICITY GRAPHS
arXiv:1612.02531v3 [cs.DS] 14 Aug 2017
ANDREW MCGREGOR AND SOFYA VOROTNIKOVA
1. Introduction
We present a data stream algorithm for estimating the size of the maximum matching of a low
arboricity graph. Recall that a graph has arboricity α if its edges can be partitioned into at most α
forests and that a planar graph has arboricity α = 3. Estimating the size of the maximum matching
in such graphs has been a focus of recent data stream research [1–3, 5, 7]. See also [6] for a survey
of the general area of graph algorithms in the stream model.
A surprising result on this problem was recently proved by Cormode et al. [3]. They designed
an ingenious algorithm that returned a (22.5α + 6)(1 + ǫ) approximation using a single pass over
the edges of the graph (ordered arbitrarily) and O(ǫ−2 α · log n · log1+ǫ n) space1. We improve the
approximation factor to (α + 2)(1 + ǫ) via a tighter analysis and show that, with a modification of
their algorithm, the space required can be reduced to O(ǫ−2 log n).
2. Results
Let match(G) be the maximum size of a matching in a graph G and let Eα be the set of edges
uv where the number of edges incident to u or v that appear in the stream after uv are both at
most α.
2.1. A Better Approximation Factor. We first show a bound for match(G) in terms of |Eα |.
Cormode et al. proved a similar but looser bound.
Theorem 1. match(G) ≤ |Eα | ≤ (α + 2) match(G).
Proof. We first prove the left inequality. To do this define ye = 1/(α + 1) if e is in Eα and 0
otherwise. Note that ye is a fractional matching with maximum weight 1/(α + 1) and hence2
X
1
α+2
|Eα |
=
ye ≤ 1 +
match(G) .
match(G) =
α+1
α+1
α+1
e
It remains to prove the right inequality. Define H to be the set of vertices with degree α + 1 or
greater. We refer to these as the heavy vertices. For u ∈ H, let Bu be the set of the last α + 1
edges incident to u that arrive in the stream.
Supported by NSF CAREER Award CCF-0953754 and CCF-1320719 and a Google Faculty Research Award.
1Here, and throughout, space is specified in words and we assume that an edge or a counter (between 0 and α)
can be stored in one word of space.
2It can be shown as a corollary of Edmonds Matching Polytope Theorem [4] that any fractional matching in which
all edge weights are bounded by ǫ is at most a factor 1+ǫ larger than the maximum integral matching. See [7, Theorem
5] for details.
1
Say an edge uv is good if uv ∈ Bu ∩ Bv and wasted if uv ∈ Bu ⊕ Bv , i.e., the symmetric difference.
Then |Eα | is exactly the number of good edges. Define
w = number
x = number
y = number
z = number
of
of
of
of
good edges with exactly no end points in H ,
good edges with exactly one end point in H ,
good edges with two end points in H ,
wasted edges with two end points in H ,
and note that |Eα | = w + x + y.
We know x + 2y + z = (α + 1)|H| because Bu contains exactly α + 1 edges if u ∈ H. Furthermore,
z + y ≤ α|H| because the graph has arboricity α. Therefore
x + y ≥ (α + 1)|H| − α|H| = |H| .
Let EL be the set of edges with no endpoints in H. Since every edge in EL is good, w = |EL |.
Hence, |Eα | ≥ |H| + |EL | ≥ match(G) where the last inequality follows because at most one edge
incident to each heavy vertex can appear in a matching.
Let Gt be the graph defined by the stream prefix of length t and let Eαt be the set of good edges
with respect to this prefix, i.e., all edges uv from Gt where the number of edges incident to u or v
that appear after uv in the prefix are both at most α. By applying the theorem to Gt , and noting
that E ∗ ≥ |Eα | and match(Gt ) ≤ match(G), we deduce the following corollary:
Corollary 2. Let E ∗ = maxt |Eαt |. Then match(G) ≤ E ∗ ≤ (α + 2) match(G).
2.2. A (Slightly) Better Algorithm. See Figure 1 for an algorithm that approximates E ∗ to
a (1 + ǫ)-factor in the insert-only graph stream model. The algorithm is a modification of the
algorithm for estimating |Eα | designed by Cormode et al. [3]. The basic idea is to independently
sample edges from Eαt with probability that is high enough to obtain an accurate approximation
of |Eαt | and yet low enough to use a small amount of space. For every sampled edge e = uv, the
algorithm stores the edge itself and two counters cue and cve for degrees of its endpoints in the rest
of the stream. If we detect that a sampled edge is not in Eαt , i.e., either of the associated counters
exceed α, it is deleted.
Cormode et al. ran multiple instances of this basic algorithm corresponding to sampling probabilities 1, (1 + ǫ)−1 , (1 + ǫ)−2 , . . . in parallel; terminated any instance that used too much space; and
returned an estimate based on one of the remaining instantiations. Instead, we start sampling with
probability 1 and put a cap on the number of edges stored by the algorithm. Whenever the capacity
is reached, the algorithm halves the sampling probability and deletes every edge currently stored
with probability 1/2. This modification saves a factor of O(ǫ−1 log n) in the space use and update
time of the algorithm. We save a further O(α) factor in the analysis by using the algorithm to
estimate E ∗ rather than |Eα |. The proof of correctness is similar to that for the original algorithm.
Theorem 3. With high probability, Algorithm 1 outputs a (1 + ǫ) approximation of E ∗ .
Proof. Let k be such that 2k−1 τ ≤ E ∗ < 2k τ where τ = 20ǫ−2 log n. First suppose we toss O(log n)
coins for each edge in Eαt and say that an edge e is sampled at level i if at least the first i − 1 coin
tosses at heads. Hence, the probability that an edge is sampled at level i is pi = 1/2i and that the
probability an edge is sampled at level i conditioned on being sampled at level i − 1 is 1/2. Let sti
be the number of edges sampled. It follows from the Chernoff bound that for i ≤ k,
2 ∗
2
2 ∗
t
ǫ E pk
ǫ τ
ǫ E pi
t
∗
≤ exp −
≤ exp −
= 1/ poly(n) .
P |si − pi |Eα || ≥ ǫpi E ≤ exp −
4
4
8
By the union bound, with high probability, sti /pi = |Eαt | ± ǫE ∗ for all 0 ≤ i ≤ k, 1 ≤ t ≤ n.
2
Algorithm 1: Approximating E ∗
(1) Initialize S ← ∅, p = 1, max = 0
(2) For each edge e = uv in the stream:
(a) With probability p add e to S and initialize counters cue ← 0 and cve ← 0
(b) For each edge e′ ∈ S, if e′ shares endpoint w with e:
• Increment cw
e′
>
α,
remove
e′ from S and corresponding counters
• If cw
′
e
−2
(c) If |S| > 30ǫ log n:
• p ← p/2
• Remove each edge in S and corresponding counters with probability 1/2
(d) max ← max(max, |S|/p)
(3) Return max
Figure 1. Approximating E ∗ Algorithm.
The algorithm initially maintains the edges in Eαt sampled at level i = 0. If the number of these
edges exceeds the threshold, we subsample these to construct the set of edges sampled at level
i = 1. If this set of edges also exceeds the threshold, we again subsample these to construct the
set of edges at level i = 2 and so on. If i never exceeds k, then the above calculation implies that
the output is (1 ± ǫ)E ∗ . But if stk is bounded above by (1 + ǫ)E ∗ /2k < (1 + ǫ)τ for all t with high
probability, then i never exceeds k.
It is immediate that the algorithm uses O(ǫ−2 log n) space since this is the maximum number of
edges stored at any one time. By Corollary 2, E ∗ is an (α + 2) approximation of match(G) and
hence we have proved the following theorem.
Theorem 4. The size of the maximum matching of a graph with arboricity α can be (α + 2)(1 + ǫ)approximated with high probability using a single pass over the edges of G given O(ǫ−2 log n) space.
Acknowledgement. In an earlier version of the proof of Theorem 3, we erroneously claimed that,
conditioned on the current sampling rate being 1/2j , all edges in Eαt had been sampled at that
rate. Thanks to Sepehr Assadi, Vladimir Braverman, Michael Dinitz, Lin Yang, and Zeyu Zhang
for raising this issue.
References
[1] Marc Bury and Chris Schwiegelshohn. Sublinear estimation of weighted matchings in dynamic data streams. In
Algorithms - ESA 2015 - 23rd Annual European Symposium, Patras, Greece, September 14-16, 2015, Proceedings,
pages 263–274, 2015.
[2] Rajesh Chitnis, Graham Cormode, Hossein Esfandiari, MohammadTaghi Hajiaghayi, Andrew McGregor, Morteza
Monemizadeh, and Sofya Vorotnikova. Kernelization via sampling with applications to finding matchings and
related problems in dynamic graph streams. In Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium
on Discrete Algorithms, SODA 2016, Arlington, VA, USA, January 10-12, 2016, pages 1326–1344, 2016.
[3] Graham Cormode, Hossein Jowhari, Morteza Monemizadeh, S. Muthukrishnan. The Sparse Awakens: Streaming
Algorithms for Matching Size Estimation in Sparse Graphs. CoRR, abs/1608.03118, 2016.
[4] Jack Edmonds. Maximum matching and a polyhedron with 0,1-vertices. Journal of Research of the National
Bureau of Standards, 69:125-130, 1965.
[5] Hossein Esfandiari, Mohammad Taghi Hajiaghayi, Vahid Liaghat, Morteza Monemizadeh, and Krzysztof Onak.
Streaming algorithms for estimating the matching size in planar graphs and beyond. In Proceedings of the TwentySixth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2015, San Diego, CA, USA, January 4-6,
2015, pages 1217–1233, 2015.
[6] Andrew McGregor. Graph stream algorithms: a survey. SIGMOD Record, 43(1):9–20, 2014.
[7] Andrew McGregor and Sofya Vorotnikova. Planar Matching in Streams Revisited. APPROX, 2016.
3
College of Computer and Information Sciences, University of Massachusetts.
4
| 8cs.DS
|
Hand Pose Estimation through Semi-Supervised and Weakly-Supervised Learning
Natalia Neverovaa,∗, Christian Wolfa , Florian Neboutb , Graham W. Taylorc
a Université
de Lyon, INSA-Lyon, CNRS, LIRIS, F-69621, France
b Awabot SAS, France
c School of Engineering, University of Guelph, Canada
arXiv:1511.06728v4 [cs.CV] 15 Sep 2017
Abstract
We propose a method for hand pose estimation based on a deep regressor trained on two different kinds of input. Raw depth data
is fused with an intermediate representation in the form of a segmentation of the hand into parts. This intermediate representation
contains important topological information and provides useful cues for reasoning about joint locations. The mapping from raw
depth to segmentation maps is learned in a semi- and weakly-supervised way from two different datasets: (i) a synthetic dataset
created through a rendering pipeline including densely labeled ground truth (pixelwise segmentations); and (ii) a dataset with
real images for which ground truth joint positions are available, but not dense segmentations. Loss for training on real images
is generated from a patch-wise restoration process, which aligns tentative segmentation maps with a large dictionary of synthetic
poses. The underlying premise is that the domain shift between synthetic and real data is smaller in the intermediate representation,
where labels carry geometric and topological meaning, than in the raw input domain. Experiments on the NYU dataset [1] show
that the proposed training method decreases error on joints over direct regression of joints from depth data by 15.7%.
Keywords: Hand pose estimation, Deep learning, Semantic segmentation
intermediate
representation
CNN
depth&map
weakly'
supervised'
learning
joint&coordinates
CNN
Figure 1: A rich intermediate representation is fused with raw input for hand
joint regression. The intermediate representation is learned in a semi/weaklysupervised setting from real and synthetic data (see Figure 3 for an illustration
of the training procedure). The input image is from the NYU dataset [1].
1. Introduction
Hand pose estimation and tracking from depth images, i.e. the
estimation of joint positions of a human hand, is a first step
for various applications: hand gesture recognition, humancomputer interfaces (moving cursors and scrolling documents),
human-object interaction in virtual reality settings and many
more. While real-time estimation of full-body pose is now
available in commercial products, at least in cooperative environments [2], the estimation of hand pose is more complex.
In situations where the user is not directly placed in front of the
∗ Corresponding
author (currently at Facebook AI Research, Paris)
Email address: [email protected] (Natalia Neverova)
Preprint submitted to Computer Vision and Image Understanding
computer, and therefore not close to the camera, the problem
is inherently difficult. In this case, typically the hand occupies
only a small portion of the image, and fingers and finger parts
are only vaguely discernible.
Existing discriminative solutions to articulated pose from
depth maps (see Section 2 for a full description of related work)
have concentrated on two different strategies: (i) early work
first contructed an intermediate representation of the body or
hand into parts, from which joints were then estimated in a second step [2, 3]; (ii) subsequent work proceeded by direct regression from depth either through heat maps or to joint coordinates
[1, 4, 5, 6, 7, 8, 9, 10].
There is clear interest in the second strategy, i.e. direct regression of joint positions, with a large body of recent work
pointing in this direction. One of the reasons is that intermediate representations are only really efficient when training is
restricted to synthetic data, where intermediate labels can be
easily obtained. On real data, precise and dense annotations
like part segmentations are difficult to come by, with the exception of finger painting datasets [11]. On the other hand, the
amount of information passed to the training algorithm is significantly higher in the case of the intermediate segmentation:
several bits per pixel of the input image vs. 2 or 3 real values
per joint and per image. In this paper we argue that this advantage is important, and we propose a new model for regression as
well as a semi-supervised and weakly-supervised training algorithm which allows us to extract this information automatically
from real data.
The novelty of our approach compared to existing work lies
in the integration of the intermediate representation as a latent
September 18, 2017
variable during training: no dense annotation on real data is
required, only joint positions are needed.
In our target configuration, pose estimation is performed
frame-by-frame without any dynamic information. A model
is learned from two training sets: (i) a (possibly small) set of
real images acquired with a consumer depth sensor. Ground
truth joint positions are assumed to be available, for instance
obtained by multi-camera motion capture systems. (ii) a second (and possibly very large) set of synthetic training images
produced from 3D models by a rendering pipeline, accompanied by dense ground truth in the form of a segmentation into
hand parts. This ground truth is easy to come by, as it is usually
automatically created by the same rendering pipeline. The main
arguments we develop are the following:
• an intermediate representation defined as a segmentation
into parts contains rich structural and topological information, since the label space itself is structured. Labels
have adjacency, topological and geometric relationships,
which can be leveraged and translated into loss for weaklysupervised training;
plains the experimental setup and provides results. Section 6
concludes.
2. Related work
Compared to the study of hand pose, a much larger body of
work has focused on full-body pose estimation. We draw influence from this literature and therefore include it in our brief
review.
Learning — The majority of recent work on pose estimation
is based on machine learning. Body part segmentation as an
intermediate representation for joint estimation from depth images was successfully used by Shotton et al. [2], where random
forests were trained to perform pixel-wise classification. This
was adapted to hand pose estimation by Keskin et al. [3], where
an additional pose clustering step was introduced. Later, Sun
et al. [14] proposed a cascaded regression framework where
positions of hand joints are predicted in a hierarchical manner
adapted to the kinematic structure of a hand. In work by Tang
et al. [12], a random forest performs different tasks at different
levels: viewpoint clustering in higher levels, part segmentation
in intermediate levels, and joint regression in lower levels. An
explicit transfer function is learned between synthetic and real
data. In follow-up work by Tang et al. [13], a latent regression forest is learned, which automatically extracts a hierarchical and topological model of a human hand from data. Instead
of pixel-wise voting, the forest is descended a single time starting from the center of mass of the hand. The traditional split
nodes in the RF are accompanied by division nodes, which trigger parallel descents of sub-trees for multiple entities (joints or
groups of joints). Li et al. [15] went further in this direction
and proposed a method where the topological model of a hand
is learned jointly with hand pose estimation. Later work uses
deep networks for body part segmentation from RGB images
[16] or depth images [17].
Recent work estimates joint positions by regression with
deep convolutional neural nets [1, 5, 4, 18, 19, 10, 20, 21].
Typically such models have been trained to produce heatmaps
encoding the joint positions as spatial Gaussians, though direct
regression to an encoding of joints has also been attempted. Recently, training with a combination of classification and regression losses for producing heatmaps has been proven particularly
effective [22].
Post-processing to enforce structural constraints is based on
graphical models [19, 4, 5, 7] or inverse kinematics [1]. In a
work by Oberweger et al. [23] a deep learning framework is
regularized by a bottleneck layer, forcing the network to model
underlying structure of joint positions. Ge at al. [24] proposed
a multi-view deep learning framework based on feature extraction from several projections of a point cloud representing a
depth image of a hand. Stacked hourglass networks [25] introduce a series of modules, each of which perform convolutions
followed by deconvolutions with skip connections. The different modules allow the method to model context, in the spirit of
auto-context models for semantic segmentation [26].
In a recent overview [27], deep learning models were shown
to perform the best among existing approaches, but still far from
• a regression of joint positions is easier and more robust if
the depth input is combined with a rich semantic representation like a segmentation into parts, provided that this
semantic segmentation is of high fidelity (see Figure 1).
We show that the additional information passed to the training
algorithm is able to improve pose regression performance. A
key component to obtaining this improvement is obtaining reliable segmentations for real data. While purely supervised training on synthetic data has proven to work well for full-body pose
estimation [2, 3], hand pose estimation is known to require real
data captured from depth sensors for training [12, 13, 1] due
to low input resolution and data quality. Manually annotating
dense segmentations of large datasets is not an option, and estimating segmentation maps from ground truth joint positions
is unreliable in the case of low quality images. We propose a
semi/weakly supervised setting in order to tackle this problem,
where this intermediate representation is learned from densely
labeled synthetic depth images as well as from real depth images associated with ground truth joint positions.
In particular, the proposed training method exploits the rich
geometrical and topological information of the intermediate
representation. During the training process, predicted segmented patches from real images are aligned with a very large
dictionary of labelled patches extracted from rendered synthetic
data. The novelty here lies in the fact that we do not match input
patches but patches taken from the intermediate representation,
which include the to-be-inferred label and its local context.
We call the proposed training method weakly supervised;
since part of the ground truth data only contains joint locations
and not the dense labels of part segmentations used during inference. The method is also semi-supervised, as part of the dataset
is fully labeled, while another part is not.
The paper is organized as follows. Section 2 discusses related work. Section 3 introduces the model and the semi/weakly
supervised learning setting. Section 4 gives details about the
deep architectures employed in the experiments. Section 5 ex2
human performance. In this context of data-driven methods,
there were a number of recent works dedicated to automating
data labeling [28] and weakly-supervised learning for sparsely
annotated videos [29].
Our work can be seen as an improvement of the discriminative stage which could be augmented with a generative counterpart of choice.
Correspondence — Pose estimation has been attempted by
solving correspondence problems in other work. In a method
by Taylor et al. [37], correspondence between depth pixels and
vertices of an articulated 3D model is learned. In [38], approximate directed Chamfer distances are used to align observed
edge images with a synthetic dataset.
Graphical models — In recent work by Chen and Yuille [5], a
graphical model is implemented with deep convolutional nets,
which jointly estimate unary terms, given evidence of joint
types and positions, and binary terms, modelling relationships
between joints. This work was later extended in [7] for modelling occluding parts by introducing an additional connectivity
prior. Tompson et al. [4] jointly learn a deep full-body part detector with a Markov Random Field which models spatial priors. Joint learning is achieved by designing the priors as convolutions and implementing inference as forward propagation
approximating a single step in a message passing algorithm.
Alternatively, Chu et al. [8] proposed a method for learning
structured representations by capturing dependencies between
body joints during training with introduced geometrical transform kernels.
Recently, model fitting and training a network on dense correspondances between the input and the template has proven to
be extremely efficient in the context of facial landmark localization [39]. Along these lines, creation of large-scale synthetic
datasets for human pose estimation [16] will facilitate adaptation of dense prediction methods to the human pose. Our work
is essentially motivated by the same idea of exploiting dense
signals as a source of universal and rich supervision.
Other work — Work on applications other than pose estimation share similarities with our method. Patchwise alignment of
segmentations in a transductive learning setting has been performed on 3D meshes [40], matching real unlabelled shape segments to shapes from a large labelled database. A large number
of candidate segments is created and the optimal segmentation
is calculated solving an integer problem. Our method can be
viewed as a kind of multi-task learning, a topic actively pursued
by the deep vision community [41, 42]. While these techniques
rely on subnetworks that share parameters, our approach does
not use weight sharing, instead, it co-ordinates subnetworks via
a patchwise restoration process and a joint error function.
Our work bears a certain resemblance to the recently proposed N 4 -Fields [43]. This method, which was proposed
for different applications, also combines deep networks and a
patchwise nearest neighbor (NN) search. However, whereas
in [43], NN-search is performed in a feature space learned
by deep networks, our method performs NN-search in a patch
space corresponding to semantic segmentation learned by deep
networks. This part of our work also bears some similarity to
the way label information is integrated in structured prediction
forests [44], although no patch alignment with a dictionary is
carried out there.
Top down methods — A different group of methods is based
on top-down processes which fit 3D models to image data. A
method by Oikonomidis et al. [30] is based on pixelwise comparison of rendered and observed depth maps. Inverse rendering of a generative model including shading and texture is used
for model fitting in [31]. Several works [32, 33] employ ICP
for hand pose reconstruction and 3D fingertip localization under spatial and temporal constraints. More recent work by Tang
et al. [9] is based on hierarchical sampling optimization, where
a sequence of predictors is aligned with kinematic structure of
a hand.
A hybrid method by Qian et al. [34] for real-time hand
tracking uses a simple hand model consisting of a number
of spheres and combines a gradient-based discriminative step
with stochastic optimization. More recent work by Sharp et
al. [11] directly exploits a hand mesh consisting of triangles
and vertices to formulate a multi-level discriminative strategy
followed by generative model-based refinement. In [35], a
person-specific and hybrid generative/discriminative model is
built for a system using five RGB cameras plus a ToF camera.
In the spirit of hybrid methods, Oberweger et al. [10] proposed
a neural model including a feedback loop based on iterative
image generation from obtained predictions followed by selfcorrection.
3. Semi-supervised and weakly-supervised learning of joint
regression
Top down down methods can be very accurate, if the geometric models are flexible enough and can closely fit the real
geometric data. In recent work, geometric models are adapted
to each instance (body or hand). In [36], for instance, the two
fitting steps (pose estimation and shape estimation) are integrated into a single joint optimization procedure. The downside of these models is their requirement for initialization. Hybrid models combine generative and discriminative steps, for
instance by performing discriminative initialization followed by
fitting of a generative model.
When prediction models are learned automatically, representation learning and full end-to-end training are often desirable and
described as the holy grail in machine learning. This approach
indeed benefits from properties, such as freedom of the training procedure to explore and find the best representation and a
lower burden on the scientist or practitioner, who is not required
to handcraft representations from knowledge of the application
and/or the input data domain.
In practice, however, depending on the difficulty of the problem, the depth of the model and the amount of available training
data, it might be suboptimal to give the model complete freedom over the intermediate representations it explores. Prob-
0 https:/www.youtube.com/watch?v=7GMiExWKM8c
3
• For a given view and pose, the part label itself carries
strong geometrical information: the label alone of a given
pixel is a very strong prior on the position of the pixel in
3d. This provides important cues for regression to the
desired joint positions. This property will also be exploited in our system to motivate patchwise searches in
label space.
VORONOI BOUNDARIES
In our setting, the intermediate representation is available during training time for the synthetic data only. For real data and
during test time, it is automatically inferred. More precisely,
our training dataset is organized into two partitions: a set of real
depth images with associated ground truth joint positions, and
a second set of synthetic depth images with associated ground
truth label images (segmentation maps in the intermediate representation). We will denote by D(j) the j th pixel in image D.
The synthetic images have been rendered from different 3D
hand models using a rendering pipeline (details will be given in
Section 5). As in [35], we also sample different pose parameters and hand shape parameters. Variations in viewpoints and
hand poses are obtained taking into account physical and physiological constraints. In contrast to other weakly-supervised or
semi-supervised methods, for instance [12], we do not suppose
that any ground truth data for the intermediate representation
is available for the real training images. Manually labelling
segmentations is extremely difficult and time consuming. Labelling a sufficiently large number of images is hardly practical.
We do, however, rely on ground truth for joint positions,
which can be obtained in several ways: In [2], external motion capture using markers is employed. In [1], training data is
acquired from multiple views from three different depth sensors
and an articulated model is fitted offline.
The functional decomposition of the method as well as the
dataflow during training and testing are outlined in Figure 3.
The goal is to regress joint positions from input depth maps,
and to this end, the proposed method leverages two different
mappings learned on two training sets. A segmentation network learns a mapping fs (·, θs ) from raw depth data to intermediate segmentation maps, parametrized by a parameter vector θs . A regression network learns a mapping fr (·, θr ) from
raw depth data combined with segmentation maps to joint positions, parametrized by a vector θr .
The training procedure uses both synthetic and real data, and
proceeds in three steps:
VORONOI BOUNDARIES
Figure 2: Complementarity of segmentation maps and key points: the space
occupied by the index finger and the ring finger is difficult to segment, as seen
by the Voronoi borders (on the left); depth values are often very similar and
close for different fingers at realistic sensor resolutions (taken from the NYU
Hand Pose dataset, on the right).
lems can either arise from overfitting or from suboptimal solutions found by the minimization procedure. Instead of falling
back to handcrafting feature representations, solutions can be
intermediate supervision, or a decomposition approach, where
an intermediate representation is temporarily imposed during a
pre-training step before full end-to-end training.
In this work we propose an intermediate representation in the
form of a segmentation map. In contrast to traditional decomposition approaches, the intermediate representation is available as additional information to the final regressor, which also
receives raw input. Our intermediate representation is a segmentation into 20 parts, illustrated in Figure 1. 19 parts correspond to finger parts, one part to the palm (see Figure 6 for exact
definitions). The background is considered to be subtracted in a
preprocessing step and is not part of the segmentation process.
This dense segmentation is complementary to groundtruth
key points and the former is very difficult to obtain from the
latter in the case of strong auto occlusions. This is illustrated
in Figure 2a: the space occupied by the index finger and the
ring finger is difficult to segment. Resorting to spatial distances
alone, traditionally done in easier settings, fails due to complex
curved shapes of fingers and the low amount of keypoints which
can be obtained with automatic methods (2 points per finger in
the NYU dataset [1]). In this context, a Voronoi diagram gives
regions which are very different from the complex finger regions. Depth values might theoretically help, but in practice
this fails since the values are often very similar and close for
different fingers at realistic sensor resolutions, as shown in Figure 2b taken from the NYU dataset.
Compared to the initial input depth image, this representation
has several important advantages:
• The part label space is characterized by strong topological properties. In contrast to other semantic segmentation problems (for instance, semantic full scene labeling),
strong neighborhood relationships can be defined on the
label space. They can be leveraged to restore noisy part
label images and to generate a loss function for training.
This property serves as a key component of creating reliable estimators and for performing transfer from synthetic
to real images.
4
1. First, the segmentation network fs is pre-trained on synthetic training data in a supervised way using dense
ground truth segmentations, resulting in a prediction
model fs (·, θs ). The parameters are learned minimizing
classical negative log-likelihood (NLL). This training step
is shown as blue data flow in Figure 3.
2. Then the prediction model for fs is fine-tuned in a subsequent step by weakly supervised training on real data,
resulting in a refined prediction model fs (·, θs0 ). This step,
shown as green data flow in Figure 3, is described in more
detail in Subsection 3.1.
3. Finally, the regression network fr is trained on real data.
It is implemented as a mapping fr (·, θr ) : (D, N ) → z
dictionary
of synthetic
patches
synthetic data:
segmentation
maps
Notations:
real data labels:
joint coordinates
patchwise
restoration
produced
segmentation
maps
produced
joint coordinates
error
segm. vs joints
segmentation
network fs
regression
network fr
backward pass
Paths:
supervised training
of fs (synthetic data)
supervised training
of fr (real data)
region
localization
real data:
depth images
synthetic data:
depth images
forward pass
weakly supervised
supervised
weakly
training of
of ffss
training
(real data)
(synthetic
data)
Figure 3: A functional overview of the method. Blue data flow corresponds to supervised training of the segmentation network fs . Green flow corresponds to
weakly supervised training of fs . Red arrows show supervised training of the regressor fr (for more detail see a supplementary video available online 2 ).
Layer
Filter size / units
Segmentation network fs
Depth input
48×48
Conv. layer 1
32×5×5
Conv. layer 2
64×5×5
Conv. layer 3
128×5×5
500
Hidden layer
Hidden layer
500
Output
20×48×48
Regression network fr
Depth input
24×24
Conv. layer c1
32×3×3
Conv. layer c211, c221
16×1×1
Conv. layer c212, c222
8×3×3
Pooling p231
Conv. layer 232
8×1×1
Conv. layer 241
16×1×1
Hidden layer fc
1200
Output
14×3
Pooling
3.1. Weakly supervised fine-tuning of the segmentation network
Supervised pre-training of the segmentation network results in
a prediction model fs (·, θs ). To address the domain shift between synthetic data and real data shot with depth sensors, the
model is fine-tuned by training on real data. Since no ground
truth segmentation maps exist for this data, we generate a loss
function for training based on two sources:
2×2
2×2
1×1
-
• sparse information in the form of ground truth joint positions, and
2×2
• a priori information on the local distribution of part labels
on human hands through a patch-wise restoration process,
which aligns noisy predictions with a large dictionary of
synthetic poses.
2×2
The weakly supervised training procedure is shown as green
data flow in Figure 3. Each real depth image is passed through
the pre-trained segmentation network fs , resulting in a segmentation map. This noisy predicted map is restored through a
restoration process fnn described further below. The quality
of the restored segmentation map is estimated by comparing it
to the set of ground truth joint positions for that image. In particular, for each joint, a corresponding part-label is identified,
and the barycenter of the corresponding pixels in the segmentation map is calculated. A rough quality measure for a segmentation map can be given as the sum (over joints) of the L2
distances between barycenters and ground truth joint positions.
The quality measure is used to determine whether the restoration process has lead to an improvement in segmentation quality, i.e. whether the barycenters of the restored map are closer
to the ground truth joint positions than the barycenters of the
original prediction. For images where this is the case, the segmentation network is updated for each pixel, minimizing NLL
using the labels of the restored map as artificial “ground truth”.
-
Table 1: Hyper-parameters chosen for the deep networks.
from a full size input depth image D and a full size segmentation map N to a joint location vector z. Parameters θr are trained classically by minimizing the L2 norm
between output joint positions and ground truth joint positions. This training step is shown as red data flow in
Figure 3.
Subsection 3.1 provides more details on step 2, the weakly supervised training procedure. The actual deep architectures employed will be described separately in Section 4.
5
c1
c211 c212
p1
c221 c222
real input
p231 c232
real
segmentation
region
selection
fc
output
c241
inception module
Figure 4: Organization of the regression network fr . All functional modules are shown in red and the produced feature maps are shown in blue. The grey areas
correspond to masked regions on the feature maps. Masked areas are shown as rectangles in the figure, but are of general shape.
(a)
(b)
(c)
(d)
(e)
(f)
(a)
(b)
(c)
(d)
(e)
(f)
Figure 5: Different segmentation results: (a) input image; (b) output of the segmentation network after supervised training; (c) after restoration; (d) output of the
(a)
(b)
(c)
(d)
(e)
(f)
segmentation network (a)
after joint training; (e) (b)
ground truth segmentation;
The image itself
(c) (f) estimated joint positions.
(d)
(e)was not part of the training
(f) set.
e 4: Different segmentation results: (a) input image; (b) output of the segmentation network after supervised training; (c) after restoration; (d) output of th
(a) (a)(e)input
(b)(b)
(c)
(d) afterThe
4: Different
segmentation
image;
output of the(f)
segmentation
network
supervised
training;
(c) after
restoration;
(d)set.
output of the
entation
network
after jointresults:
training;
ground
truth
segmentation;
estimated joint
positions.
image(e)
itself
was not
part(f)
of
the training
re 4: Different
segmentation
results:(e)(a)
input truth
image;
(b) output of(f)the
segmentation
network after
supervised
(c) after
restoration;
(d) output of t
ntation
network
after joint training;
ground
segmentation;
estimated
joint
positions.
itselftraining;
was
not part
of the
training
3.2. Patchwise
restoration
proach,
which The
aimsimage
to match
at
a patch-local
level
ratherset.
than
mentation
network
after joint training;
(e)
ground
truth segmentation;
(f)
estimated
joint network
positions.after
Thesupervised
image itself
was not(c)
part
of the
training set.
ure 4: Different
segmentation
results:
(a)
input
image;
(b)
output
of
the
segmentation
training;
after
restoration;
(d) output of
matching
(a)
(b)patches of size P(c)
(d) whole images.(e)
(f)
We proceed patchwise,
extracting
×P from
mentation
network
after
joint
training;
(e)
ground
truth
segmentation;
(f)
estimated
joint
positions.
The
image
itself
was
not
part
of
the
training
set.
(k)
given real input depth image D is aligned with this diction pixel j,
the assigned
patches
n ofimages,
all neighbors
Architectures
a large
set of synthetic
segmentation
resulting k
in are
a dic- 4. A
(k)
in a patchwise
process using
the (c)
intermediate
representapixel and
j,tionary
the assigned
patches
n
of. all
neighbors
k are
4. nary
Architectures
Different
segmentation
results:
(a)(l)
input
image;
(b)
of
the
segmentation
network
after supervised
training;
after restoration;
(d) output of
of patches
P pixel
={p
N
}.output
Inisour
experiments,
(k)
mined
of
j},n
inl∈{1
each
calculated.
en
pixel
j, the
the position
assigned
patches
of. .patch
all
neighbors
kestimated
are
4.
Architectures
(j)
tion.
For
each
pixel
j,
a
patch
q
is
extracted
from
the
segThe
structure
of
the
segmentation
network
f
is
motivated
b
ion
network
after
joint
training;
(e)
ground
truth
segmentation;
(f)
joint
positions.
The
image
itself
was
not
part
of
the
training
set.
s
we
patches
size
27×27
and
aa dictionary
of
36(.):
million
ined
theused
position
ofofpixel
j in each
patch
is calculated.
belpixel
isand
estimated
by voting,
resulting
inof
mapping
fnn
(k)
mined
and
the
position
of
pixel
j
in
each
patch
is
calculated.
en
j,
the
assigned
patches
n
all
neighbors
k
are
4.
Architectures
mentation
produced
by
the
network
f
,
and
the
nearest
neighbor
The
structure
of
the
segmentation
network
f
is
motivated
by
s
s
patches by
is extracted
the training
set (see Section
idea
of performing
efficient pixelwise
image
el is estimated
voting, from
resulting
in a mapping
fnn (.):5). As the
The
structure
of
the segmentation
network
fs is segmentatio
motivated b
(j)
ν(q
is found
by searching
the dictionary
P : image
abel
is estimated
by voting,
resulting
inposes
a patch
mapping
fnnoccur
(.): 380
mined
and
position
of pixel
j in each
is calculated.
idea) of
performing
efficient
pixelwise
segmentation
alsothe
reported
in [35],
the
range
of
which
can
in the
X
preserving
the
original
resolution
of
the
input,
and
is inspire
the
idea
of
performing
efficient
pixelwise
image
segmentatio
(j)
(k,j@k)
The structure
of (j)
the segmentation network
fs is motivated
(k)
natural
motion
is
extremely
large,
full global
matching
X
(q
)=
arg
max
I(lmaking
=aneighbors
nmapping
) fknn
(2)
abel
estimated
by
voting,
resulting
in
(.):
(j)
(j) input,
(l)
xel fj,isnnthe
assigned
patches
n
of
all
are
4.
Architectures
380
preserving
the
original
resolution
of
the
and
is
by
networks
were
proposed
object
detec
,
ν(q
) = which
arg
min
dHof
(q
), forand
(1)isinspired
(j)
(k,j@k)
XThis
l difficult.
380
preserving
original
resolution
the, pinput,
inspire
theOverFeat
ideanofthe
performing
efficient
f
(q
)
=
arg
max
I(l
=
n
)
(2)
with
pose
datasets
motivates
our
patch-based
ap(l) ∈P pixelwise image segmentati
nn
(j)
(k,j@k)
p
k2W
⇥W
by OverFeat networks which were proposed for object deteced and the position of pixel j in each patch is calculated.
fnn (q ) = arg max
I(l = n
)
(2) 380 tion
l
and localization
[38,which
39]. The
consists
ofis3 inspir
convo
by
OverFeat
networks
werenetwork
object
dete
preserving
the
original
resolution
ofproposed
the input,
structure
of
the [38,
segmentation
network
fsfor
isand
motivated
⇥W
l k2W X
(j)
(k,j@k)
6tion
and
localization
39].
The
network
consists
of (see
3 convois estimated
resulting
a mapping
fnn) (.): (2) The
k2W ⇥WinI(l
fnn (q by
) =voting,
arg max
=n
lutional
layers,
followed
by
a
fully
connected
layer
Tab
tion
and
localization
[38,
39].
The
network
consists
of
3
conv
by
OverFeat
networks
which
were
proposed
for
object
dete
the
idea layers,
of performing
efficient
pixelwise
image
segmentati
l
re n(k) =⌫(q (j) ) is the nearest
neighbor
result
for
pixel
k,
lutional
followed
by
a
fully
connected
layer
(see
Table
k2W
⇥W
1).
Max
pooling,
which
typically
follows
convolutional
layer
(k)
(j)
lutional
followed
a fully
connected
layer
(see
Tab
tion andlayers,
localization
[38,by
39].
Theofnetwork
consists
3inspir
conv
Xnneighbor
(j)
the original
the convolutional
input,
and of
is
em)ndenotes
=⌫(q
) ismthe
nearest
result
pixel
k,380
(k)
(j)
pixel
of
, I(!)=1
if !for
holds
and
0385 preserving
1).
Max in
pooling,
which resolution
typically
follows
layers,
(j)
(k,j@k)
results
downsampling
of
feature
maps,
destroying
precis
ere
n(q
=⌫(q
) is
thepatch
nearest
result
for pixel
k,
f) nn
)
=
arg
max
I(l
=
n
)
(2)
(j) neighbor
1).
Max
pooling,
which
typically
follows
convolutional
layer
layers,
followed
by awere
fullyproposed
connectedfor
layer
(seedete
Tab
denotes
pixel(j)m oflj@k
patchdenotes
n (j)
, I(!)=1
if ! holds
andj 0in
bylutional
OverFeat
networks
which
object
m)
,ere
and
position
pixel
(k) expression
ininformation.
downsampling
of feature
maps, destroying
precise
spatial
Instead,
we perform
pooling
over
2laye
⇥
denotes
pixel)mis of
n neighbor
,the
I(!)=1
if !ofholds
and 385
0385 results
nthe
=⌫(q
thepatch
nearest
result
for
pixel
k,
k2W
⇥W
results
in
downsampling
of
feature
maps,
destroying
preci
1).
Max
pooling,
which
typically
follows
convolutional
and
the
expression
j@k denotes
the position of pixel j in
tion
and
localization
[38,
39].
The
network
consists
of
3
conv
patch
centered
on
neighbor
k.
,m)
(j)
spatial
information.
Instead,
we
perform
pooling
over
2
⇥
2
, anddenotes
the expression
j@k
denotes
position
pixeland
j in0 overlapping regions produced by shifting the feature maps ob
pixel m of
patch
n ,the
I(!)=1
if !ofholds
where dH (q, p) is the Hamming distance between the two
patches q and p. The search performed in (1) can be calculated
efficiently using KD-trees.
In a naı̈ve setting, a restored label for pixel j could be obtained by choosing the label of the center of the retrieved patch
n(j) . This, however, leads to noisy restorations, which suggest
the need for spatial context. Instead of chosing the center label
of each patch only, we propose to use all of the labels in each
patch. For each input pixel, the nearest neighbor results in a
local window of size W ×W are integrated. In particular, for a
given pixel j, the assigned patches n(k) of all neighbors k are
examined and the position of pixel j in each patch is calculated.
A label is estimated by voting, resulting in a mapping fnn (.):
X
fnn (q (j) ) = arg max
I(l = n(k,j@k) )
(2)
l
The regression network, taking a depth image as a single input and producing 3 coordinates for a given joint, also incorporates the information provided by the segmentation network
during training. Structurally, it resembles an Inception module [49, 50] where the output of the first convolutional layer
after max pooling as passed through several parallel feature extractors capturing information at different levels of localization.
The organization of this module is shown in Figure 4. and the
corresponding parameters are, as before, provided in Table 1.
The output of the first convolutional layer c1 (followed by pooling p1) is aligned with the segmentation maps produced by the
segmentation network. From each feature map, for a given joint
we extract a localized region of interest filtered by the mask of a
hand part to which it belongs (or a number of hand parts which
are naturally closest to this joint). These masks are calculated
by performing morphological opening on the regions having
the corresponding class label in the segmentation map. Once
the local region is selected, the rest of the feature map area is
set to 0. The result, along with the original feature maps is then
fed to the next layer, i.e. an Inception module. The rest of the
training process is organized in such a way that the network’s
capacity is split between global structure of the whole image
and the local neighborhood, and a subset of Inception 3×3 filters is learned specifically from the local area surrounding the
point of interest.
Experiments showed that individual networks for each joint
do perform better than networks sharing parameters over joints.
We conjecture, that sharing parameters would be the optimal
choice for smaller amounts of training data. As the number
of examples increases, separating networks allows all layers to
pick up the fine nuances required for regressing each individual
joint.
Both networks have ReLU activation functions at each layer
and employ batch normalization [51]. The regression network
during test time uses the batch normalization parameters estimated on the training data, however, in the segmentation network, the batch normalization is performed across all pixels
from the same image, for both training and test samples.
k∈W ×W
where n(k) =ν(q (j) ) is the nearest neighbor result for pixel k,
n(j,m) denotes pixel m of patch n(j) , I(ω)=1 if ω holds and 0
else, and the expression j@k denotes the position of pixel j in
the patch centered on neighbor k.
This integration bears some similarity to [43], where information of nearest neighbor searches is integrated over a local
window, albeit through averaging a continuous mapping. It is
also similar to the patchwise integration performed in structured
prediction forests [44].
If real-time performance is not required, the patch alignment
process in Equation 1 can be regularized with a graphical model
including pairwise terms favoring consistent assignments, for
instance, a Potts model or a term favoring patch assignments
with consistent overlaps. Interestingly, this produces only very
small gains in performance, especially given the higher computational complexity. Moreover, the gains vanish if local integration (Equation 2) is added. More information is given in
Section 5.
4. Architectures
The structure of the segmentation network fs is motivated by
the idea of performing efficient pixelwise image segmentation
preserving the original resolution of the input, and is inspired
by OverFeat networks which were proposed for object detection and localization [45, 46]. The network consists of 3 convolutional layers, followed by a fully connected layer (see Table 1). Max pooling, which typically follows convolutional layers, results in downsampling of feature maps, destroying precise spatial information. Instead, we perform pooling over 2×2
overlapping regions produced by shifting the feature maps obtained at the previous step by a single pixel along one or another
axis. As opposed to patchwise training of pixel classification
based on its local neighborhood, such an architecture is more
computationally efficient, as it benefits from dense computations at earlier layers. Compared to recently introduced fullyconvolutional [47] and deconvolutional networks [48], which
tackle a similar problem in the context of semantic segmentation, the proposed network requires less upsampling and interpolation.
5. Experimental results
We evaluated the proposed method on the NYU Hand Pose
Dataset, which was published in [1] and is publicly available3 .
It comprises 70,000 images captured with a depth sensor in
VGA resolution accompanied by ground truth annotations of
positions of hand joints.
A video illustrating the results of our method is available
online4 . It shows the obtained pose as well as the intermediate representation during testing and during training, i.e. after
restoration.
To train the segmentation network, the synthetic training images were selected from our own dataset consisting of two subsets: (i) 170,974 synthetic training images rendered using the
3 http://cims.nyu.edu/ tompson/NYU_Hand_Pose_
˜
Dataset.htm
4 https://www.youtube.com/watch?v=7GMiExWKM8c
7
Method
— per pixel—
No restoration
NN-search — no integration
NN-search — integration w. equation (2)
CRF – Potts-like model
CRF – Hamming distance on overlapping patch area
100
100
51.03
48.76
54.55
53.10
52.45
— per class —
(+3.52)
(+2.07)
(+1.42)
39.38
39.72
46.38
43.64
42.68
(+7.00)
(+4.26)
(+3.30)
Table 2: Restoration (=segmentation) accuracy on 100 manually labelled images of the NYU dataset (=NYU-100).
90
90
80
80
70
70
60
60
50
50
40
40
30
30
20
20
10
10
00
11
22
33
44
55
66
77
88
99
10
10
Supervised
training only
only
Supervised
training
Supervised*training*only
training*only
Supervised*
11
11
12
12
13
13
14
14
15
15
16
16
17
17
18
18
19
19
20
20
Weakly*
supervised*
training
Weakly*
training
Weaklysupervised*
supervised
training
Weakly
supervised
training
Figure 6: Classification accuracy of the segmentation network fs for supervised training only (blue), and for weakly supervised training (red), where 1 . . . 20 is
100
the number100
of a segment shown in the hand legend at the right.
90
90
80
80
any training procedure. Recall that pure segmentation perforcommercial software “Poser”, including ground truth; (ii) 500
70
70
mance is of course not the goal of this work (or of the targeted
labelled synthetic images plus ground truth reserved for testing.
60
60
application), it is given as additional information only.
In our experiments, we extract hand images of normalized
50
50
metric size (taking into account depth information) and resize
Purely supervised training on the synthetic dataset gives poor
40
40
them to 48×48
pixels. The data is preprocessed by local consegmentation performance of 51.03% accuracy per pixel. We
30
30
trast normalization
with a kernel size of 9. To be able to predict
emphasize once more that training was performed on synthetic
absolute 2020
values for z-coordinates, the subtracted depth is then
images while we test on real images, thus of an unseen distri10 to the network output. For supervised training of the
10
added back
bution. This domain shift is clearly a problem, as accuracy on
00
segmentation
network
and
the
regression
network,
the
full
set
the1515synthetic
dataset
is2020very high, 90.16%. Using patchwise
10
11
12
13
14
16
17
18
19
11
22
33
44
55
66
77
88
99
10
11
12
13
14
16
17
18
19
of 170,974 training images is used.Supervised
A
third
of
this
set
is
used
restoration
of
the
predicted
real patches with a large dictionary
Supervised
training
only
training only
Supervised*training*only
training*only
Unsupervised*
training
Unsupervisedtraining
training
Supervised*
Unsupervised*
Unsupervised
training
to extract patches for the patch alignment mapping fnn , givof synthetic patches gives a performance increase of +3.5 pering a dictionary of 36M patches of size 27×27 extracted from
centage points per pixel and +7 percentage points per class.
56,991 images. Local integration as given in equation (2) was
This corroborates our intuition that the intermediate represendone using windows of size W ×W = 17×17.
tation carries important structural information. Integration of
The segmentation network was initially trained for 100
patch-labels over a local window with equation (2) is essential–
epochs with SGD, batch size 1, initial learning rate 0.1 and
pure NN-search without integration performs poorly.
learning rate decay 10−5 on the synthetic data and then fineFigure 5 provides examples of input depth maps and different
tuned for additional 10 epochs on a mixture of synthetic and
segmentations: after synthetic pre-training only, after off-line
restored real segmentations with the ratio 9:1.
restoration and a prediction after weakly-supervised finetuning.
Finally, the regression network is trained by gradient descent
We can see that in a majority of cases the restoration output is
using the Adam [52] update rule, with learning rate set 0.05 and
very close to ground truth maps. Figure 6 shows the distribution
batch size of 64.
of the error over the different hand parts. The improvement is
ConvNets were implemented using the Torch7 library [53].
consistent, and high on the important fingertips.
NN-search using KD-trees was performed using the FLANN
Restoration failure cases, corresponding to 5-10% of the real
library [54].
data, are demonstrated and explained in Figure 7. At the finetuning stage, these samples are filtered out and excluded from
training (see Section 3.1 for more detail).
5.1. Segmentation performance
We also compared local integration of equation (2) to potentially more powerful regularization methods by implementing
To evaluate the performance of the various segmentation metha CRF-like discrete energy function. The goal of the optimizaods, we manually labelled 100 images from the NYU dataset
tion problem is to regularize the patchwise restoration process
and report accuracy per pixel and per class in Table 2. These
described in Section 3.2. Instead of chosing the nearest neigh100 images were solely used for evaluation and never entered
8
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
Figure 7: Visualization of the automatic rejections made by the quality check described in section 3.1. Examples of raw noisy segmentations of real images are
shown on top and the corresponding maps after restoration are shown at bottom. These restoration failure cases may happen due to artifacts in depth maps (c), (e),
(f), increased distance between the camera and the hand (g) and particularly noisy initial predictions (a), (d), (h). As a result, the restoration process may result in
lower recall of finger segments (a)-(g) or, less often, false detections (h). Both cases can be automatically detected.
Method
Datasets used
Fully supervised training only
Semi-/weakly-supervised training
synth. segmentations
synth. segmentations + real joint positions
— per pixel—
51.03
57.18
(+6.15)
— per class —
39.38
47.20
(+7.82)
Table 3: The contribution of semi-/weakly-supervised training on segmentation accuracy.
bor in patch space for each pixel as described in equation (2),
a solution is searched which satisfies certain coherence conditions over spatial neighborhoods. To this end, we create a global
energy function E(x) defined on a 2D-lattice corresponding to
the input image to restore:
E(x) =
X
i
u(xi ) + α
X
Inference was performed through message passing using the
open-GM library [65], and the hyper-parameter α was optimized on a hold-out set. Interestingly, local patch integration
with equation (2) outperformed the combinatorial models significantly while at the same time being much faster. We conjecture that the reason is that our patchwise integration corresponds to a high-order potential (2), whereas the binary terms
in the CRF model are poorer. Calculating the global solution
of poorer model seems to be less efficient than locally solving
a high order problem. We see this as a further indication of the
strong topological information carried by the label space of the
intermediate representation.
Table 3 shows the contribution of semi/weakly-supervised
training, where sparse annotation (joint positions) are integrated
into the training process of the segmentation network fs . This
procedure achieves an improvement of +6.15 percentage points
per pixel and +7.82 percentage points per class, an essential
step in learning an efficient intermediate representation.
At first sight it might seem to be odd that the pretrained
predictor (+6.15pp w.r.t.t. baseline) is better than the off-line
restoration process (+3.52pp), on whose results it has been
trained. However, let us recall that the predictor also uses a second source of information, namely the joint locations available
on real images during the weakly-supervised fine-tuning step.
Rejecting bad segmentations during training is responsible for
the difference.
b(xi , xj )
i∼j
where i∼j indices neighbors i and j. Each pixel i is assigned a
discrete variable xi taking values between 1 and N =10, where
xi =l signifies that for pixel i the l−th nearest neighbor in patch
space is chosen. For each pixel i, a nearest neighbor search is
performed using KD-trees and a ranked list of N neighbors is
kept defining the label space for this pixel. The variable N controls the degree of approximation of the model, where N =∞
allows each pixel to be assigned every possible patch of the
synthetic dictionary.
The unary data term u(x) guides the solution towards approximations with low error. It is defined as the Hamming distance between the original patch and the synthethic patch.
We tested two different pairwise terms b(xi , xj ):
Potts-like terms — a Potts model classically favors equality of
labels of neighboring sites. In our setting, we favor equality of the center pixels of the two patches assigned to xi
and xj .
5.2. Hand joint position estimation
Patch-overlap distance — the alternative pairwise term is defined as the Hamming distance between the two synthethic
patches defined by xi and xj , in particular, the distance restricted to the overlapping area.
Table 4 illustrates the effect of incorporating segmentation information on the performance of the regression network, i.e. on
the error in joint positions — the main goal of this work. In the
9
Method
2D error, mm
3D error, mm
Tompson et al. [1]
Oberweger et al. (DeepPrior) [23]
Oberweger et al. [10]
Bouchacourt et al. (DISCO) [55]
Xu et al. (Lie-X) [56]
Zhou et al. (DeepModel) [57]
Deng et al. (Hand3D) [58]
Guo et al. (REN) [59]
Wan et al. (Crossing Nets) [60]
Fourure et al. (JTSC) [61]
Zhang et al. [62]
Madadi et al. [63]
Oberweger et al. (DeepPrior++) [64]
7.1
14.8
12.4
–
–
–
–
–
–
8.0
–
–
–
28.8
19.8
16.0
20.7
14.5
16.9
17.6
13.4
15.5
16.8
18.3
15.6
12.3
Our baseline: direct regression (a)
Our baseline: cascade direct regression (two steps) (b)
Our baseline: semi-supervised network (c)
Our method: semi/weakly supervised network (d)
13.3
12.6
12.1
11.2
17.3
16.9
16.5
14.8
Table 4: Joint position estimation error on the NYU Hand Pose dataset (some values were estimated from plots if authors did not provide numerical values). Baseline
(a) corresponds to the case of a single regression network where the segmentation results are not used. The cascaded baseline (b) is similar in architecture but features
an additional refinement step (as proposed in [23]). Method (c) is based on the full pipeline with no restoration employed during training, i.e. the segmentation
network is trained on synthetic data in supervised way and not fine-tuned on the real samples. Finally, method (d) corresponds to our proposed solution.
1
Fraction of frames within distance / %
0.9
0.8
0.7
0.6
0.5
0.4
Our method
Oberweger et al. (DeepPrior++) [63]
Ge et al. [20]
Oberweger et al. [10]
Oberweger et al. [10]
Zhou
et al. et(DeepModel)
[57]
Tompson
al. [1]
Sinha
et al. [20]
(DeepHand) [21]
DeepPrior
Oberweger
(DeepPrior)
[23]
Regressionet+al.
restored
segmentation
Tompson et al. [1]
0.3
0.2
0.1
0
0
10
20
30
40
50
60
Distance threshold / mm
70
80
90
100
Figure 8: Joint estimation accuracy for the 3D case expressed as propotion of frames where all joints are localized within a given distance threshold in mm.
bottom part of the table we compare the proposed method to
our own baselines, which are based on an ablation study of the
proposed network. In particular, we compare with direct regression without the intermediate representation, and with a version
where the intermediate representation is learned in a supervised
way only.
and its ground truth location. In comparison to a single network
regressor, the 2D mean error was improved by 15.7%. The second best model not involving segmentation (cascade regression)
was inspired by the work of [18] and more recent [23], where
the initial rough estimation of hand joints positions is then improved by zooming in.
The first part of the table provides comparison with the literature and is based on publicly available data (predictions of
uv-coordinates of joint locations) released by the authors of corresponding methods. The error is expressed as mean distance
in mm (in 2D or 3D) between predicted position of each joint
We can see that the purely supervised baseline performs
well compared to the state-of-the-art. Prior to augmenting the
method with the unsupervised pipeline, we spent additional
time on careful optimization of the baselines by tuning the architecture and the training regime. This appeared to be crucial,
10
Figure 9: Visualization of estimated hand skeletons produced by our network trained in a semi/weakly-supervised fashion (depth images are sampled uniformly
from the NYU Hand Pose dataset).
(a)
(b)
(c)
(d)
Figure 10: Visual comparison of results produced on the NYU Hand Pose dataset by different 3D pose estimation methods: (a) DeepPrior [23], (b) Oberweger et
al. [10], (c) our method, based on the network trained in a semi/weakly-supervised fashion, (d) ground truth.
11
7. Acknowledgements
in particular the choice of batch normalization [66] and Adam
optimization [67].
Figure 9 visualizes the estimated pose skeletons corresponding to 24 input depth maps randomly sampled from the test set.
Figure 10 illustrates performance of the proposed method on
challenging examples in comparison with a number of state-ofthe-art methods. Figure 8 plots quantitative performance expressed as the number of frames with all joints being localized
within a given distance threshold in 3D. In these figures, we
compare the proposed method with several recent state-of-theart approaches. For the 2D hand pose estimation method proposed in [1], we follow [23] and augment estimated x and y coordinates with depth values from the input depth maps. In those
cases when predicted locations fall on the background, we set
the corresponding z-coordinate to the median depth value of the
hand.
We should note here, that the quality of network outputs can
be further improved by optimization through inverse kinematics, as it has been done, for example, in [1]. However, the focus of this work is to explore the potential of pure learning approaches with no priors enforcing structure on the output. The
bottom part of the table contains non deep learning methods. In
a recent work [9] on optimization of hand pose estimation formulated as an inverse kinematics problem, the authors report
performance similar to [1] in terms of 2D UV-error (no error in
mm provided).
This work has been partly financed through the French grant Interabot, a project of type “Investissement’s d’Avenir / Briques
Génériques du Logiciel Embarqué”, and by the ANR project
SoLStiCe (ANR-13-BS02-0002-01), a project of the grant program “ANR blanc”. G. Taylor acknowledges the support of
NSERC, CFI, and NVIDIA.
References
References
[1] J. Tompson, M. Stein, Y. LeCun, K. Perlin, Real-time continuous pose
recovery of human hands using convolutional networks, in: ACM Transactions on Graphics, 2014.
[2] J. Shotton, A. Fitzgibbon, M. Cook, T. Sharp, M. Finocchio, R. Moore,
A. Kipman, A. Blake, Real-time human pose recognition in parts from
single depth images, in: CVPR, 2011.
[3] C. Keskin, F. Kiraç, Y. Kara, L. Akarun, Hand pose estimation and hand
shape classification using multi-layered randomized decision forests, in:
ECCV, 2014.
[4] J. Tompson, A. Jain, Y. LeCun, C. Bregler, Joint training of a convolutional network and a graphical model for human pose estimation, in:
NIPS, 2014.
[5] X. Chen, A. Yuille, Articulated pose estimation by a graphical model with
image dependent pairwise relations, in: CVPR, 2014.
[6] M. Sun, P. Kohli, J. Shotton, Conditional regression forests for human
pose estimation, in: CVPR, 2012.
[7] X. Chen, A. Yuille, Parsing occluded people by flexible compositions, in:
CVPR, 2015.
[8] X. Chu, W. Ouyang, H. Li, X. Wang, Structured feature learning for pose
estimation, in: arXiv:1603.09065v1, 2016.
[9] D. Tang, J. Taylor, K. Pushmeet, C. Keskin, T.-K. Kim, J. Shotton, Opening the black box: Hierarchical sampling optimization for estimating human hand pose, in: ICCV, 2015.
[10] M. Oberweger, P. Wohlhart, V. Lepetit, Training a feedback loop for hand
pose estimation, in: ICCV, 2015.
[11] T. Sharp, C. Keskin, D. Robertson, J. Taylor, J. Shotton, D. Kim, C. Rhemann, I. Leichter, A. Vinnikov, Y. Wei, D. Freedman, P. Kohli, E. Krupka,
A. Fitzgibbon, S. Izadi, Accurate, robust, and flexible real-time hand
tracking, in: SIGCHI, 2015.
[12] D. Tang, T. Yu, T.-K. Kim, Real-time articulated hand pose estimation
using semi-supervised transductive regression forests, in: ICCV, 2013.
[13] D. Tang, H. Chang, A. Tejani, T.-K. Kim, Latent regression forest: Structured estimation of 3d hand posture, in: CVPR, 2014.
[14] X. Sun, Y. Wei, S. Liang, X. Tang, J. Sun, Cascaded hand pose regression,
in: CVPR, 2015.
[15] P. Li, H. Ling, X. Li, C. Liao, 3d hand pose estimation using randomized
decision forest with segmentation index points, in: ICCV, 2015.
[16] G. Varol, J. Romero, X. Martin, N. Mahmood, M. Black, I. Laptev,
C. Schmid, Learning from synthetic humans, in: CVPR, 2017.
[17] M. Jiu, C. Wolf, G. Taylor, A. Baskurt, Human body part estimation from
depth images via spatially-constrained deep learning, Pattern Recognition
Letters 50 (2014) 122–129.
[18] A. Toshev, C. Szegedy, DeepPose: Human pose estimation via deep neural networks, in: CVPR, 2014.
[19] A. Jain, J. Tompson, M. Andriluka, G. Taylor, C. Bregler, Learning human pose estimation features with convolutional networks, in: ICLR,
2014.
[20] L. Ge, H. Liang, J. Yuan, D. Thalmann, 3d convolutional neural networks
for efficient and robust hand pose estimation from single depth images,
in: CVPR, 2017.
[21] A. Sinha, C. Choi, K. Ramani, Deephand: Robust hand pose estimation
by completing a matrix with deep features, in: CVPR, 2016.
[22] A. Bulat, G. Tzimiropoulos, Human pose estimation via convolutional
part heatmap regression, in: ECCV, 2016.
[23] M. Oberweger, P. Wohlhart, V. Lepetit, Hands deep in deep learning for
hand pose estimation, in: CVWW, 2015.
5.3. Computational complexity
All models have been trained and tested using GPUs, except the
patchwise restoration process which is pure CPU code and not
used at test time. Estimating the pose of a single hand takes 31
ms if the segmentation resolution is set to 24×24 pixels (which
includes the forward passes of both networks fs (12 ms) and
fr (18 ms)) and 58 ms for 48×48 segmentation outputs (40 ms
for fs , corresponding to the results reported in the experiments
section of the paper). Training of the segmentation network
requires up to 24 hours to minimize validation error, while the
regression network is trained in 20 min on a single GPU. The
results were obtained using a cluster configured with 2 x E52620 v2 Hex-core processors, 64 GB RAM and 3 x Nvidia GTX
Titan Black cards with 6GB memory per card.
6. Conclusion
We presented a method for hand pose estimation based on an
intermediate representation which is fused with raw depth input data. We showed that the additional structured information
of this representation provides important cues for joint regression which leads to lower error. Weakly supervised learning of
the mapping from depth to segmentation maps from a mixture
of densely labelled synthetic data and from sparsely labelled
real data is a key component of the proposed method. Weak supervision is dealt with by patch-wise alignment of real data to
synthetic data performed in the space of the intermediate representation, exploiting its strong geometric and topological properties.
12
[24] L. Ge, H. Liang, J. Yuan, D. Thalmann, Robust 3d hand pose estimation
in single depth images: From single-view cnn to multi-view cnns, in:
CVPR, 2016.
[25] A. Newell, K. Yang, J. Deng, Stacked hourglass networks for human pose
estimation, in: ECCV, 2016.
[26] P. Pinheiro, R. Collobert, Recurrent convolutional neural networks for
scene labeling, in: ICML, 2014.
[27] J. S. III, G. Rogez, Y. Yang, J. Shotton, D. Ramanan, Depth-based hand
pose estimation: methods, data, and challenges, in: ICCV, 2015.
[28] M. Oberweger, G. Riegler, P. Wohlhart, V. Lepetit, Efficiently creating 3d
training data for fine hand pose estimation, in: CVPR, 2016.
[29] O. Koller, H. Ney, R. Bowden, Deep hand: How to train a cnn on 1 million hand images when your data is continuous and weakly labelled, in:
CVPR, 2016.
[30] I. Oikonomidis, N. Kyriazis, A. Argyros, Efficient model-based 3D tracking of hand articulations using Kinect, in: BMVC, 2011.
[31] M. de La Gorce, D. Fleet, N. Paragios, Model-based 3d hand pose estimation from monocular video, IEEE TPAMI 33 (9) (2014) 1793–1805.
[32] H. Liang, J. Yuan, D. Thalmann, Z. Zhang, Model-based hand pose estimation via spatial-temporal hand parsing and 3D fingertip localization,
The Visual Computer 29 (2013) 837–848.
[33] A. Tagliasacchi, M. Schroder, A. Tkach, S. Bouaziz, M. Botsch, M. Pauly,
Robust articulated-icp for real-time hand tracking, in: Eurographics Symposium on Geometry Processing, 2015.
[34] C. Qian, X. Sun, Y. Wei, X. Tang, J. Sun, Realtime and Robust Hand
Tracking from Depth, in: CVPR, 2014.
[35] S. Sridhar, A. Oulasvirta, C. Theobalt, Interactive markerless articulated
hand motion tracking using RGB and depth data, in: ICCV, 2013.
[36] J. Taylor, L. Bordeaux, T. Cashman, B. Corish, C. Keskin, E. Soto,
D. Sweeney, J. Valentin, B. Luff, A. Topalian, E. Wood, S. Khamis,
P. Kohli, T. Sharp, S. Izadi, R. Banks, A. Fitzgibbon, J. Shotton, Efficient
and precise interactive hand tracking through joint, continuous optimization of pose and correspondences, in: ACM TOG - SIGGRAPH, 2016.
[37] J. Taylor, J. Shotton, T. Sharp, A. Fitzgibbon, The vitruvian manifold:
Inferring dense correspondences for one-shot human pose estimation, in:
CVPR, 2012.
[38] V. Athitsos, Z. Liu, Y. Wu, J. Yuan, Estimating 3D hand pose from a
cluttered image, in: CVPR, 2003.
[39] R. A. Guler, G. Trigeorgis, E. Antonakos, P. Snape, S. Zafeiriou, I. Kokkinos, Densereg: Fully convolutional dense shape regression in-the-wild,
in: CVPR, 2017.
[40] W. Xu, Z. Shi, M. Xu, K. Zhou, J. Wang, B. Zhou, J. Wang, Z. Yuan,
Transductive 3D shape segmentation using sparse reconstruction, in:
Transactions on Graphics, 2014.
[41] T. Devries, K. Biswaranjan, G. Taylor, Multi-task learning of facial landmarks and expression, in: 14th Canadian Conference on Computer and
Robot Vision (CRV), 2014.
[42] Z. Zhang, P. Luo, C. C. Loy, X. Tang, Facial landmark detection by deep
multi-task learning, in: ECCV, 2014.
[43] Y. Ganin, V. Lempitsky, N 4 -Fields: Neural Network Nearest Neighbor
Fields for Image Transforms, in: ACCV, 2014.
[44] P. Kontschieder, S. Bulo, M. Pelillo, H. Bischof, Structured labels in random forests for semantic labelling and object detection, in: ICCV, 2011.
[45] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, Y. LeCun,
Overfeat: Integrated recognition, localization and detection using convolutional networks, in: ICLR, 2014.
[46] N. Neverova, C. Wolf, G. Taylor, F. Nebout, Hand segmentation with
structured convolutional learning, in: ACCV, 2014.
[47] J. Long, E. Shelhamer, T. Darrell, Fully convolutional networks for semantic segmentation, in: CVPR, 2015.
[48] H. Noh, S. Hong, B. Han, Learning deconvolution network for semantic
segmentation, in: ICCV, 2015.
[49] C. Szegedy, W. Liu, J. Yangqing, P. Sermanet, S. Reed, D. Anguelov,
D. Erhan, V. Vanhoucke, A. R. et al, Going deeper with convolutions, in:
CVPR, 2015.
[50] M. Lin, Q. Chen, S. Yan, Network in network, in: ICLR, 2014.
[51] S. Ioffe, C. Szegedy, Batch normalization: Accelerating deep network
training by reducing internal covariate shift, in: ICML, 2015.
[52] D. Kingma, J. L. Ba, Adam: A method for stochastic optimization, in:
ICLR, 2015.
[53] R. Collobert, K. Kavukcuoglu, C. Farabet, Torch7: A matlab-like envi-
ronment for machine learning, in: NIPS BigLearn Workshop, 2011.
[54] M. Muja, D. Lowe, Fast approximate nearest neighbors with automatic
algorithm configuration, in: VISAPP, 2009.
[55] D. Bouchacourt, M. P. Kumar, S. Nowozin, Disco nets: Dissimilarity coefficient networks, in: NIPS, 2016.
[56] C. Xu, N. Govindarajan, Y. Zhang, L. Cheng, Liex: Depth image based
articulated object pose estimation, tracking, and action recognition on lie
groups, in: IJCV, 2016.
[57] K. Zhou, Q. Wan, W. Zhang, X. Xue, Y. Wei, Model-based deep hand
pose estimation, in: IJCAI, 2016.
[58] X. Deng, S. Yang, Y. Zhang, P. Tan, L. Chang, H. Wang, Hand3d: Hand
pose estimation using 3d neural network, in: arXiv:1704.02224, 2017.
[59] H. Guo, G. Wang, X. Chen, C. Zhang, F. Qiao, H. Yang, Region ensemble
network: Improving convolutional network for hand pose estimation, in:
ICIP, 2017.
[60] C. Wan, T. Probst, L. van Gool, L. Yo, Crossing nets: Dual generative
models with a shared latent space for hand pose estimation, in: CVPR,
2017.
[61] D. Fourure, R. Emonet, E. Fromont, D. Muselet, N. Neverova,
A. Tremeau, C. Wolf, Multi-task, multi-domain learning: application to
semantic segmentation and pose regression, in: Neurocomputing, 2017.
[62] X. Zhang, C. Xu, Y. Zhang, T. Zhu, L. Cheng, Multivariate regression
with grossly corrupted observations: A robust approach and its applications, in: arXiv:1701.02892, 2017.
[63] M. Madadi, S. Escalera, X. Baro, J. Gonzalez, End-to-end global to local
cnn learning for hand pose recovery in depth data, in: arXiv:1705.09606,
2017.
[64] M. Oberweger, V. Lepetit, Deepprior++: Improving fast and accurate 3d
hand pose estimation, in: ICCV workshop, 2017.
[65] B. Andres, T., Beier, J. Kappes, OpenGM: A c++ library for discrete
graphical models, arXiv e-prints (2012).
[66] S. Ioffe, C. Szegedy, Batch normalization: Accelerating deep network
training by reducing internal covariate shift, in: ICML, 2015.
[67] D. Kingma, J. Ba, Adam: A method for stochastic optimization, in:
ICML, 2015.
13
| 2cs.AI
|
ON TORSION IN FINITELY PRESENTED GROUPS
arXiv:1107.1489v4 [math.GR] 19 Oct 2016
MAURICE CHIODO
Abstract. We give a uniform construction that, on input of a recursive
presentation P of a group, outputs a recursive presentation of a torsionfree group, isomorphic to P whenever P is itself torsion-free. We use this
to re-obtain a known result, the existence of a universal finitely presented
torsion-free group; one into which all finitely presented torsion-free groups
embed. We apply our techniques to show that recognising embeddability of
finitely presented groups is Π02 -hard, Σ02 -hard, and lies in Σ03 . We also show
that the sets of orders of torsion elements of finitely presented groups are
precisely the Σ02 sets which are closed under taking factors.
1. Introduction
By a finite presentation ⟨X∣R⟩ of a group we mean, as usual, a finite collection
of generators X, together with a finite set R of defining relations. A recursive
(resp. countably generated recursive) presentation ⟨X∣R⟩ of a group is then a
finite (resp. countable) collection of generators X, together with a recursive
enumeration of a possibly infinite set R of defining relations. We use P to
denote the group presented by a presentation P .
The Higman embedding theorem [8] shows that every recursively presented
group embeds into a finitely presented group. Moreover, this embedding can be
made uniform; there is an algorithm that takes any recursive presentation P
and outputs a finite presentation Q and an explicit embedding φ ∶ P ↪ Q. This
embedding theorem was used by Higman to show the existence of a universal
finitely presented group; one into which all finitely presented groups embed. By
analysing Higman’s embedding theorem, we prove:
Theorem 3.10. There is a universal finitely presented torsion-free group G.
That is, G is torsion-free, and for any finitely presented group H we have that
H ↪ G if (and only if ) H is torsion-free.
Theorem 3.10 first appeared (as far as we are aware) in the appendix by Oleg
V. Belegradek of the paper [1], Theorem A.1. He gives a different proof to ours,
making use of arguments from model theory. Moreover, in [1, Remark A.2] he
2010 AMS Classification: 20F10, 03D40, 03D80.
Keywords: Higman’s embedding theorem, universal finitely presented group, embeddings,
Kleene’s arithmetical hierarchy, torsion.
The author was supported by: a University of Melbourne Overseas Research Experience
Scholarship, the Italian FIRB “Futuro in Ricerca” project RBFR10DGUA 002, and the Swiss
National Science Foundation grant FN PP00P2-144681/1.
Date: October 20, 2016.
1
2
MAURICE CHIODO
points out that theorem 3.10 can also be proved along the lines we follow in the
present paper.
Key to many of the important results in this work is the technical observation that the Higman embedding theorem can preserve the set of orders of
torsion elements; we state this as theorem 2.2. Every group G has a unique
torsion-free quotient through which all other torsion-free quotients factor (see
corollary 3.4); we call this the torsion-free universal quotient Gtf . By standard techniques in combinatorial group theory, we show in proposition 3.8 the
existence of an algorithm that takes any finite presentation P and outputs a
recursive presentation P tf of the torsion-free universal quotient of P . Theorem 3.10 then follows by combining theorem 2.2 and proposition 3.8, similar to
Higman’s original construction of a universal finitely presented group.
In [10] it was shown by Lempp that the problem of recognising torsionfreeness for finitely presented groups is Π02 -complete in Kleene’s arithmetic hierarchy (see [12], or the introduction to [10], for a description of Σ0n sets, Π0n
sets, and Kleene’s arithmetic hierarchy). Therefore the set of finitely presented
subgroups of any universal torsion-free finitely presented group is Π02 -complete,
and, in particular, not recursively enumerable. In [4] we gave another proof of
the existence of a finitely presented group whose set of finitely presented subgroups is not recursively enumerable, without the use of the results of Lempp
[10] or Oleg Belegradek [1]. Building on theorem 3.10, we show the following.
Theorem 4.5. For any recursive enumeration P1 , P2 , . . . of all finite presentations of groups, the set K = {(i, j) ∈ N2 ∣ P i ↪ P j } is Σ02 -hard, Π02 -hard, and
has a Σ03 description.
We write Tord(G) to denote the orders of non-trivial torsion elements of a
group G, and say a set A ⊆ N is factor-complete if it is closed under taking
multiplicative factors (excluding 1). Applying theorem 2.2 to an idea by Dorais
in the comments to [17], we give the following complete characterisation of sets
which can occur as Tord(G) for G a finitely (or recursively) presented group:
Theorem 5.2. For a set of natural numbers A the following are equivalent:
(1) A = Tord(G) for some finitely presented group G;
(2) A = Tord(G) for some countably generated recursively presented group G;
(3) A is a factor-complete Σ02 set.
It follows (corollary 5.5) that we can realise any Σ02 set, up to one-one equivalence, as Tord(G) for some finitely presented group G.
Acknowledgements: The author wishes to thank Jack Button, Andrew Glass,
Steffen Lempp, Vincenzo Marra, Chuck Miller and Rishi Vyas for their many
useful conversations and comments which led to the overall improvement of this
work. Thanks also go to François Dorais, Stefan Kohl, Benjamin Steinberg and
Henry Wilton for their thoughtful discussion on MathOverflow, which led to
the addition of section 5 to this work. We thank Igor Belegradek for bringing
ON TORSION IN FINITELY PRESENTED GROUPS
3
the work [1] to our attention, and to the anonymous referee for suggesting improvements. Finally, thanks must go to the late Greg Hjorth, whose suggestion
of making contact with Steffen led to the eventual writing of this work.
2. Preliminaries
2.1. Notation.
With the convention that N contains 0, we write ϕm to be the mth partial
recursive function ϕm ∶ N → N, and the domain of ϕm to be the mth partial
recursive set Wm (also known as a recursively enumerable set, abbreviated to
r.e. set). A presentation P = ⟨X∣R⟩ is said to be a countably generated recursive
presentation if X is a recursive enumeration of generators, and R a recursive
enumeration of relators. If P, Q are group presentations then we denote their
free product presentation by P ∗ Q, given by taking the disjoint union of their
generators and relators; this extends to the free product of arbitrary collections
of presentations. If X is a set, we write X ∗ for the set of finite words on
X ∪X −1 , including the empty word ∅. If φ ∶ X → Y ∗ is a set map, then we write
φ ∶ X ∗ → Y ∗ for the extension of φ to X ∗ . If g1 , . . . , gn are elements of a group G,
then we write ⟨g1 , . . . , gn ⟩G for the subgroup in G generated by these elements,
and ⟪g1 , . . . , gn ⟫G for the normal closure of these elements in G. Cantor’s
pairing function is defined by ⟨⋅, ⋅⟩ ∶ N × N → N, ⟨x, y⟩ ∶= 21 (x + y)(x + y + 1) + y,
which gives a computable bijection.
2.2. Embedding theorems.
Definition 2.1. Let G be a group. We let o(g) denote the order of a group
element g, and say g is torsion if 1 ≤ o(g) < ∞. We set
Tor(G) ∶= {g ∈ G ∣ g is torsion}
Tord(G) ∶= {n ∈ N ∣ ∃g ∈ Tor(G) with o(g) = n ≥ 2}
Thus Tord(G) is the set of orders of non-trivial torsion elements of G.
As detailed in [4, Lemma 6.9 and Theorem 6.10], the following is implicit in
Rotman’s proof [13, Theorem 12.18] of the Higman embedding theorem.
Theorem 2.2. There is a uniform algorithm that, on input of a countably
generated recursive presentation P = ⟨X∣R⟩, constructs a finite presentation
T(P ) such that P ↪ T(P ) and Tord(P ) = Tord(T(P )), along with an explicit
embedding φ ∶ P ↪ T(P ).
We will also use the following consequence to theorem 2.2.
Theorem 2.3 ([4, Lemma 6.11]). There is a uniform algorithm that, on input of any n ∈ N, constructs a finite presentation Qn such that Tord(Qn ) is
one-one equivalent to N ∖ Wn . Taking n′ with Wn′ non-recursive thus gives
that Tord(Qn′ ) is not recursively enumerable; thus the set of finitely presented
subgroups of Qn′ is not recursively enumerable.
4
MAURICE CHIODO
3. Universal finitely presented torsion-free groups
If G, H are groups with H torsion-free, a surjective homomorphism h ∶ G ↠ H
is universal if, for any torsion-free K and any homomorphism f ∶ G → K, there
is a homomorphism φ ∶ H → K such that f = φ ○ h ∶ G → K, i.e., the following
diagram commutes:
h /
H
❆❆
❆❆
❆ φ
f ❆
G❆
K
Note that if φ exists then it will be unique. Indeed, if φ′ also satisfies f = φ′ ○ h,
then φ ○ h = φ′ ○ h, and hence φ = φ′ as h is a surjection and thus is rightcancellative. Moreover, any such H is unique, up to isomorphism. Such an
H is called the universal torsion-free quotient for G, denoted Gtf . Observe
that if G is itself torsion-free, then Gtf exists and Gtf ≅ G, as the identity map
idG ∶ G → G has the universal property above.
A standard construction, showing that Gtf exists for every group G, is done
via taking the quotient of G by its torsion-free radical ρ(G), where ρ(G) is
the intersection of all normal subgroups N ⊲ G with G/N torsion-free (see
[3]). It follows immediately that G/ρ(G) has all the properties of a torsion-free
universal quotient for G.
We present here an alternative construction for Gtf which, though isomorphic
to G/ρ(G), lends itself more easily to an effective procedure for finitely (or
recursively) presented groups, as shown in proposition 3.8.
Definition 3.1. Given a group G, we inductively define Tori (G) as follows:
Tor0 (G) ∶= {e}
Tori+1 (G) ∶= ⟪ {g ∈ G ∣ g Tori (G) ∈ Tor (G/ Tori (G))} ⟫G
Tor∞ (G) ∶= ⋃ Tori (G)
i∈N
Thus, Tori (G) is the set of elements of G which are annihilated upon taking
i successive quotients of G by the normal closure of all torsion elements, and
Tor∞ (G) is the union of all these.
By construction, Tori (G) ≤ Torj (G) whenever i ≤ j. It follows immediately
that Tor∞ (G) ⊲ G. The finite presentation P ∶= ⟨x, y, z∣x2 , y 3 , xy = z 6 ⟩ defines
a group for which Tor1 (P ) ≠ Tor∞ (P ), as shown in [6, Proposition 4.1].
Lemma 3.2. If G is a group, then G/ Tor∞ (G) is torsion-free.
Proof. Suppose g Tor∞ (G) ∈ Tor (G/ Tor∞ (G)). Then g n Tor∞ (G) = e in G/ Tor∞ (G)
for some n > 1, so gn ∈ Tor∞ (G). Thus there is some i ∈ N such that
gn ∈ Tori (G), and hence g Tori (G) ∈ Tor (G/ Tori (G)). Thus g ∈ Tori+1 (G) ⊆
Tor∞ (G), and so g Tor∞ (G) = e in G/ Tor∞ (G).
Proposition 3.3. If G is a group, then ρ(G) = Tor∞ (G).
ON TORSION IN FINITELY PRESENTED GROUPS
5
Proof. Clearly ρ(G) ⊆ Tor∞ (G), by definition of ρ(G) and the fact that G/ Tor∞ (G)
is torsion-free (lemma 3.2). It remains to show that Tor∞ (G) ⊆ ρ(G). We proceed by contradiction, so assume Tor∞ (G) ⊈ ρ(G). Then there is some N ⊲ G
with G/N torsion-free, along with some minimal i such that Tori (G) ⊈ N
(clearly, i > 0, as Tor0 (G) = {e}). Then, by definition of Tori (G) and the
fact that N is normal, there exists e ≠ g ∈ Tori (G) such that g Tori−1 (G) ∈
Tor (G/ Tori−1 (G)) and g ∉ N (or else Tori (G) ⊆ N ). But then gn ∈ Tori−1 (G)
for some n > 1. Since Tori−1 (G) ⊆ N by minimality of i, we have that gN is
a (non-trivial) torsion element of G/N , contradicting torsion-freeness of G/N .
Hence Tor∞ (G) ⊆ ρ(G).
Corollary 3.4. If G is a group, then G/ Tor∞ (G) = Gtf ; the torsion-free universal quotient for G.
What follows is a standard result, which we state without proof.
Lemma 3.5. Let P = ⟨X∣R⟩ be a countably generated recursive presentation.
Then the set of words {w ∈ X ∗ ∣ w = e in P } is r.e.
Lemma 3.6. Let P = ⟨X∣R⟩ be a countably generated recursive presentation.
Then the set of words {w ∈ X ∗ ∣ w ∈ Tor(P ) in P } is r.e.
Proof. Take any recursive enumeration {w1 , w2 , . . .} of X ∗ . Using lemma 3.5,
start checking if win = e in P for each wi ∈ X ∗ and each n ∈ N (by proceeding
along finite diagonals). For each wi we come across which has a finite order,
add it to our enumeration. This procedure will enumerate all words in Tor(P ),
and only words in Tor(P ). Thus the set of words in X ∗ representing elements
in Tor(P ) is r.e.
From this, we deduce the following:
Lemma 3.7. Given a countably generated recursive presentation P = ⟨X∣R⟩,
the set Ti ∶= {w ∈ X ∗ ∣ w ∈ Tori (P ) in P } is r.e., uniformly over all i and all
such presentations P . Moreover, the union T∞ ∶= ⋃ Ti is r.e., and is precisely
the set {w ∈ X ∗ ∣ w ∈ Tor∞ (P ) in P }.
Proof. We proceed by induction. Clearly Tor1 (P ) is r.e., as it is the normal
closure of Tor(P ), which is r.e. by lemma 3.6. So assume that Tori (P ) is
r.e. for all i ≤ n. Then Torn+1 (P ) is the normal closure of Tor(P / Torn (P )),
which again is r.e. by the induction hypothesis and lemma 3.6. The rest of the
lemma then follows immediately.
Proposition 3.8. There is a uniform algorithm that, on input of a countably
generated recursive presentation P = ⟨X∣R⟩ of a group P , outputs a countably
generated recursive presentation P tf = ⟨X∣R′ ⟩ (on the same generating set X,
and with R ⊆ R′ as sets) such that P tf is the torsion-free universal quotient of
P , with associated surjection given by extending idX ∶ X → X.
Proof. By corollary 3.4, P is the group P / Tor∞ (P ). Then, with the notation
lemma 3.7, it can be seen that P tf ∶= ⟨X∣R ∪ T∞ ⟩ is a countably generated
tf
recursive presentation for P , uniformly constructed from P .
tf
6
MAURICE CHIODO
Theorem 3.9. There is a finitely presentable group G which is torsion free, and
contains an embedded copy of every countably generated recursively presentable
torsion-free group.
Proof. Take an enumeration P1 , P2 , . . . of all countably generated recursive presentations of groups, and construct the countably generated recursive presentation Q ∶= P1tf ∗P2tf ∗. . .; this is the countably infinite free product of the universal
torsion-free quotient of all countably generated recursively presentable groups
(with some repetition). As each Pitf is uniformly constructible from Pi (by
proposition 3.8), we have that our construction of Q is indeed effective, and
hence Q is a countably generated recursive presentation. Also, proposition 3.8
shows that Q is a torsion-free group, as we have successfully annihilated all the
torsion in the free product factors, and the free product of torsion-free groups is
again torsion-free. Moreover, Q contains an embedded copy of every torsion-free
countably generated recursively presentable group, as the universal torsion-free
quotient of a torsion-free group is itself. Now use theorem 2.2 to embed Q into a
finitely presentable group T(Q). By construction, ∅ = Tord(Q) = Tord(T(Q)),
so T(Q) is torsion-free. Finally, T(Q) has an embedded copy of every countably
generated recursively presentable torsion-free group, since Q did. Taking G to
be T(Q) completes the proof.
From this we immediately observe the following consequence.
Theorem 3.10. There is a universal finitely presented torsion-free group G.
That is, G is torsion-free, and for any finitely presented group H we have that
H ↪ G if (and only if ) H is torsion-free.
Note. One may ask why theorem 3.10 does not follow immediately from Higman’s embedding theorem by taking the free product of all finite presentations
of torsion-free groups, and using the fact that Higman’s theorem preserves orders of torsion elements. This cannot work, as we later shown in theorem
4.2 that the set of finite presentations of torsion-free groups is not recursively
enumerable.
Remark. Miller [11, Corollary 3.14], extending a result of Boone and Rogers
[2, Theorem 2], showed there is no universal finitely presented solvable word
problem group. It can be shown that none of the following group properties
admit a universal finitely presented group: finite, abelian, solvable, nilpotent
(simple, however, remains open).
4. Complexity of embeddings
Using the machinery described in section 2, we can encode the following
recursion theory facts into groups.
Lemma 4.1 ([12, §13.2 Theorem VIII]). The set {n ∈ N ∣ Wn = N} is Π02 complete; the set {n ∈ N ∣ ∣Wn ∣ < ∞} is Σ02 -complete.
We can thus recover the following result, first proved in [10, Main Theorem].
Theorem 4.2. The set of finite presentations of torsion-free groups is Π02 complete.
ON TORSION IN FINITELY PRESENTED GROUPS
7
Proof. Given n ∈ N, we use theorem 2.3 to construct a finite presentation Qn
such that Tord(Qn ) is one-one equivalent to N ∖ Wn . Thus Qn is torsion-free if
and only if Wn = N. From lemma 4.1, {n ∈ N ∣ Wn = N} is Π02 -complete, so the
set of torsion-free finite presentations is at least Π02 -hard. But this set has the
following Π02 description (taken from [10]):
G is torsion-free if and only if ∀w ∈ G∀n > 0(wn ≠G e or w =G e)
and hence is Π02 -complete.
Combining this with the universal torsion-free group from theorem 3.10, we
get the following immediate corollary, which extends theorem 2.3.
Corollary 4.3. There is a finitely presented group whose finitely presentable
subgroups form a Π02 -complete set.
A similar construction to the proof of theorem 2.3 (as found in [4, Lemma
6.11]) gives us the following:
Proposition 4.4. For any fixed prime p, the set of finite presentations into
which Cp embeds is Σ02 -complete.
Proof. Given n ∈ N we form the countably generated recursive presentation Pn
as follows:
Pn ∶= ⟨x0 , x1 , . . . ∣ {xpi ∣ i ∈ N} ∪ {x0 , . . . , xj ∣ j ∈ Wn }⟩
If ∣Wn ∣ < ∞ then P n ≅ Cp ∗ Cp ∗ . . .. Conversely, if ∣Wn ∣ = ∞ then P n ≅ {e}. So
Tord(P n ) = {
{p} if ∣Wn ∣ < ∞
∅ if ∣Wn ∣ = ∞
That is, Cp ↪ P n if and only if ∣Wn ∣ < ∞. Now use theorem 2.2 to construct a finite presentation T(Pn ) such that P n ↪ T(Pn ) with Tord(P n ) = Tord(T(Pn )).
Hence Cp ↪ T(Pn ) if and only if ∣Wn ∣ < ∞, so by lemma 4.1 the set of finite
presentations into which Cp embeds is Σ02 -hard. But this set has the following
straightforward Σ02 description:
Cp ↪ G if and only if ∃w ∈ G(w ≠G e and wp =G e)
and hence is Σ02 -complete.
We can now prove:
Theorem 4.5. Take an enumeration P1 , P2 , . . . of all finite presentations of
groups; Pi = ⟨Xi ∣Ri ⟩. Then the set K = {(i, j) ∈ N2 ∣ P i ↪ P j } is Σ02 -hard,
Π02 -hard, and has a Σ03 description.
Proof. Corollary 4.3 shows that K is Π02 -hard, Proposition 4.4 shows that K is
Σ02 -hard, and the following is a Σ03 description for K:
K = {(i, j) ∈ N2 ∣ (∃φ ∶ Xi → Xj∗ )(∀w ∈ Xi∗ )(φ(w) =P j e if and only if w =P i e)}
8
MAURICE CHIODO
Note that, with the aid of Cantor’s pairing function (a computable bijection
between N2 and N), we can view the set K above as being a subset of N. Hence
it makes sense to talk of K being Π02 -hard etc.
Based on theorem 4.5, we conjecture the following:
Conjecture. The set K defined above is Σ03 -complete. That is, the problem of
deciding for finite presentations Pi , Pj if P i ↪ P j is Σ03 -complete.
5. Complexity of Tord(G)
We now apply our techniques to investigate the complexity of Tord(G) for
G a finitely presented group.
Definition 5.1. Call a set A ⊆ N≥2 factor-complete if it is closed under taking
non-trivial factors. That is, n ∈ A ⇒ m ∈ A for all m > 1 with m∣n.
We give a set-theoretic description of precisely which factor-complete sets
can appear as Tord(G) for G finitely (or recursively) presented. We presented
an earlier proof of the following result in [5]; what follows is a clearer proof
pointed out to us by the anonymous referee.
Theorem 5.2. For a set of natural numbers A the following are equivalent:
(1) A = Tord(G) for some finitely presented group G;
(2) A = Tord(G) for some countably generated recursively presented group G;
(3) A is a factor-complete Σ02 set.
Proof. (2) ⇒ (1) because, by theorem 2.2, any recursively presented group can
be embedded into a finitely presented group with the same Tord.
(1) ⇒ (3). First, observe that Tord(G) is factor-complete (for any group G),
because if o(g) = mn then o(g m ) = n, for any g ∈ G. Second, Tord(G) is a Σ02
set. Indeed, if G has finite presentation ⟨X ∣ R⟩, and S is the set of words in
X ∗ which represent the trivial element in G, then
Tord(G) = {n ∣ ∃w ∈ X ∗ (n > 1 ∧ wn ∈ S ∧ ∀i(0 < i < n ⇒ wi ∉ S))}
Since S is r.e. (by lemma 3.5), it is a Σ01 subset of X ∗ , and so the result follows.
(3) ⇒ (2). As A is a Σ02 set, it has a description of the form
A = {n ∈ N ∣ ∃x∀yR(n, x, y)}
for some ternary recursive relation R on N. Let
P ∶= {(n, m) ∈ N2 ∣ (∀x ≤ m)(∃y)¬R(n, x, y)}
Clearly P is r.e. If n ∉ A then (n, m) ∈ P for all m. Conversely, if n ∈ A then
(n, m) ∈ P
⇔ m < mn ∶= min{m ∣ (∀y)R(n, m, y)}
Let I ∶= {(n, m) ∈ N2 ∣ n > 1}, and let G ∶= ⟨X∣T ⟩ where
X ∶= {anm ∣ (n, m) ∈ I},
T ∶= {annm ∣ (n, m) ∈ I} ∪ {anm ∣ (n, m) ∈ I ∩ P }
Clearly, T is r.e., and so G has countably generated recursive presentation. By
the observations above, G can be defined by the generators anm and relators
ON TORSION IN FINITELY PRESENTED GROUPS
9
annm = e, where n ∈ A and m ≥ mn . Let Kn denote the free product of countably
many cyclic groups Cn of order n. Then G is isomorphic to the free product
G ≅ ∗n∈A Kn
and therefore
Tord(G) = ⋃ Tord(Cn ) = ⋃ {k ∣ k∣n ∧ k > 1} = A;
n∈A
n∈A
the latter equality holds because A is factor-complete.
Note. Theorem 5.2 was first proved in the more restricted setting of primes
(i.e., considering sets of integers consisting only of primes) by Steinberg [14]
and Wilton [17], in response to a question asked by Kohl [9]. Moreover, in the
comments in [17], Dorais gave a sketch of an alternate proof of the version for
primes. Our original proof was a formalisation the proof by Dorais, and our
result is an extention of this to the more general setting of all factor-complete Σ02
sets. We thank Dorais, Kohl, Steinberg, and Wilton for their online discussion,
as well as their insight into key aspects of this result; our work in this section
is an extension of their ideas and results.
From the uniformity of the constructions in the proof of theorem 5.2, we
make the following observation.
Proposition 5.3. The equivalence discussed in theorem 5.2 is computable, in
the following sense:
a) Given a countably generated recursive presentation Q, we can compute from
it a finite presentation P with Tord(P ) = Tord(Q).
b) Given a finite presentation P , we can compute from it a ternary recursive
relation R on N for which Tord(P ) = {n ∈ N ∣ ∃x∀yR(n, x, y)}.
c) Given a ternary recursive relation R on N for which A ∶= {n ∈ N ∣ ∃x∀yR(n, x, y)}
is factor-complete, we can compute from it a countably generated recursive presentation Q with Tord(Q) = A.
We adopt the standard numbering of primes {pi }i∈N , ordered by size; the
following lemma is then immediate.
Lemma 5.4. Let X ⊆ N. Then the set
Xprime ∶= {pi ∣ i ∈ X}
is factor-complete and one-one equivalent to X.
Applying lemma 5.4 to theorem 5.2, we can conclude the following:
Corollary 5.5. Given any Σ02 set A, the set Aprime is one-one equivalent to
A, and can be realised as the set of orders of torsion elements of some finitely
presented group G.
6. Further work
This paper invites research into several questions. We mention some here.
10
MAURICE CHIODO
Problem 1. Given the existence of a universal torsion-free group (theorem
3.10), and the constructions of Valiev [15, 16] of explicit finite presentations of
universal finitely presented groups, one could perhaps combine these techniques
to produce an explicit finite presentation of a universal torsion-free group.
Problem 2. The positions of the following properties in the arithmetic hierarchy have not been fully determined. Techniques such as those we have covered
here may be of use in locating them.
1. Solvable: Known to have a Σ03 description.
2. Residually finite: Known to have a Π02 description.
3. Simple: Known to have a Π02 description.
4. Orderable: Known to have a Π03 description (the Ohnishi condition).
Points 1–3 are mentioned in [11, p. 20], while point 4 appears in [7, Lemma
2.2.1]. We note that it may very well be the case that some of these are neither
Π0n -complete nor Σ0n -complete, for any n.
Problem 3. Following from theorem 5.5, and the uniformity of such a realisation of a Σ02 set A as one-one equivalent to the torsion orders of a finitely presented group, one could perhaps construct an explicit finite presentation P of a
group with Tord(P ) being Σ02 -complete, by encoding the set {n ∈ N ∣ ∣Wn ∣ < ∞}
which is Σ02 -complete (lemma 4.1).
References
[1] I. Belegradek, A. Szczepański, Endomorphisms of relatively hyperbolic groups (with an
appendix by O. V. Belegradek, On universal torsion-free finitely presented groups), Internat. J. Algebra Comput. 18 no. 1, 97–110 (2008).
[2] W. W. Boone, H. Rogers Jr., On a problem of J.H.C. Whitehead and a problem of Alonzo
Church, Math. Scand. 19, 185–192 (1966).
[3] S. D. Brodsky, J. Howie, The universal torsion-free image of a group, Israel J. Math. 98,
209-228 (1997).
[4] M. Chiodo, Finding non-trivial elements and splittings in groups, J. Algebra. 331, 271–
284 (2011).
[5] M. Chiodo (mathoverflow.net/users/31415), Primes occurring as orders of elements of a
finitely presented group, http://mathoverflow.net/questions/121823 (v.2013-02-14).
[6] M. Chiodo, R. Vyas, Quotients by torsion elements, arXiv:1303.1784v1 (2013).
[7] A. M. W. Glass, Partially ordered groups, Series in algebra Vol. 7, World Sci., Singapore,
(1999).
[8] G. Higman, Subgroups of finitely presented groups, Proc. Royal Soc. London Ser. A 262,
455–475 (1961).
[9] S. Kohl (mathoverflow.net/users/28104), Primes occurring as orders of elements of a
finitely presented group, http://mathoverflow.net/questions/121178 (v.2013-02-08).
[10] S. Lempp, The computational complexity of recognising torsion-freeness of finitely presented groups, Bull. Austral. Math. Soc. 56, 273–277 (1997).
[11] C. F. Miller III, Decision problems for groups-survey and reflections. Algorithms and
classification in combinatorial group theory (Berkeley, CA, 1989), Math. Sci. Res. Inst.
Publ., 23, Springer, New York, 1–59 (1992).
[12] H. Rogers Jr, Theory of recursive functions and effective computability, MIT Press, (1987).
[13] J. Rotman, An introduction to the theory of groups, Springer-Verlag, New York, (1995).
[14] B. Steinberg (mathoverflow.net/users/15934), Primes occurring as orders of elements of
a finitely presented group, http://mathoverflow.net/questions/121267 (v.2013-02-09).
[15] M. K. Valiev, Examples of universal finitely presented groups (Russian), Dokl. Akad.
Nauk SSSR 211, 265–268 (1973). English transl., Soviet Math. 14, 987–991 (1973).
ON TORSION IN FINITELY PRESENTED GROUPS
11
[16] M. K. Valiev, Universal group with twenty-one defining relations, Discrete Math. 17 no.
2, 207–213 (1977).
[17] H. Wilton (mathoverflow.net/users/1463), Primes occurring as orders of elements of a
finitely presented group, http://mathoverflow.net/questions/121189 (v.2013-02-10).
Mathematics Department, University of Neuchâtel
Rue Emile-Argand 11, Neuchâtel, 2000, SWITZERLAND
[email protected]
| 4math.GR
|
Mobile Machine Learning Hardware at ARM:
A Systems-on-Chip (SoC) Perspective
Yuhao Zhu∗
Department of Computer Science
University of Rochester
[email protected]
Matthew Mattina
Machine learning is playing an increasingly significant role in
emerging mobile application domains such as AR/VR, ADAS, etc.
Accordingly, hardware architects have designed customized hardware for machine learning algorithms, especially neural networks,
to improve compute efficiency. However, machine learning is typically just one processing stage in complex end-to-end applications, involving multiple components in a mobile Systems-on-achip (SoC). Focusing only on ML accelerators loses bigger optimization opportunity at the system (SoC) level. This paper argues that
hardware architects should expand the optimization scope to the
entire SoC. We demonstrate one particular case-study in the domain
of continuous computer vision where camera sensor, image signal
processor (ISP), memory, and NN accelerator are synergistically
co-designed to achieve optimal system-level efficiency.
INTRODUCTION
Mobile devices are the most prevalent computing platform of the
present day, and are dominated by the ARM architecture. A large
number of emerging mobile application domains now rely heavily
on machine learning; in particular, various forms of deep neural
networks (DNNs) have been instrumental in driving progress on
problems such as computer vision and natural language processing.
On mobile platforms, DNN inference is currently typically executed
in the cloud. However, the trend is to move DNN execution from
the cloud to the mobile devices themselves. This shift is essential to
remove the communication latency and privacy issues of the cloud
offloading approach.
The increasing use of DNNs in mobile applications places significant compute requirements on the mobile System-on-chip (SoC),
which must process tens of billions of linear algebra operations per
second within a tight energy budget. In response, there has been
significant effort expended on dedicated hardware to accelerate
the computation of neural networks. This is borne out in a proliferation of designs for DNN accelerators (NNX), which typically
demonstrate high computational efficiency on the order of 0.4 –
3.8 TOPS/W on convolutional NN inference [12, 13, 15, 17]. This
is several orders of magnitude more efficient than typical mobile
CPU implementations.
Sadly, the efficiency benefits of hardware accelerators are largely
a one-time improvement, and will likely saturate, while the compute
requirement of DNNs keep increasing. Using computer vision as an
example, today’s convolutional neural network (CNN) accelerators
are not able to perform object detection (e.g., YOLO[16]) in real
time at 1080p/60fps. As the resolution, frame rate, and the need
∗ Work
done while a visiting researcher at ARM Research
Machine Learning & AI
ARM Research
[email protected]
Mobile Systems-on-a-chip (SoC)
CNN
Accelerator
DRAM
arXiv:1801.06274v2 [cs.LG] 1 Feb 2018
ABSTRACT
1
Paul Whatmough
Machine Learning & AI
ARM Research
[email protected]
CPU
Cluster
ML-enabled Application
GPU
L3 $
Accelerator
Interface (ACP)
System (SoC)
Interconnect
Sensor
Hub
Sensors
Video
Codec
Domain Library
NN (e.g.,
AndroidNN)
Image
Signal
Processor
Camera
Computer
Vision
NLP
Hardware Abstraction Layer
Camera
HAL
NNX
HAL
GPU
HAL
Device Driver
Camera
Driver
NNX
Driver
GPU
Driver
Figure 1: Mobile SoC (left) and software stack (right). Shaded
components are used in continuous vision tasks.
for stereoscopic vision grows with the emergence of AR/VR use
cases, the compute requirement will continue to increase, while the
power budget remains constant, leaving a large gap.
Therefore, we must move from a narrow focus on hardware
accelerators to begin to consider system-level optimizations for ML
on mobile. Expanding the scope beyond the DNN accelerator to
consider the whole SoC, we emphasize three areas for optimization:
• Accelerator Interfacing : hardware accelerators must be
efficiently interfaced with the rest of the SoC for full benefit.
• Software Abstractions : for cross-platform compatability,
SoC details should be abstracted with a clean API.
• System Optimizations : Co-design of algorithms and the
various hardware blocks in the system.
Section 2 will describe these optimizations further, with a case study
in Section 3. We conclude in Section 4.
2
ML ON MOBILE SYSTEMS
We have already started to see changes in mobile systems in response to the computational demands of deep learning methods.
Most notably, NNX components are now common in mobile SoCs,
e.g., the Neural Engine in the iPhoneX [2] and the HPU CNN coprocessor in the Microsoft HoloLens [7]. However, a significant
challenge still remains as to how to integrate NNX components
into the system. We identify three aspects from both hardware and
software perspectives.
Accelerator Interfacing There are two main challenges in interfacing NNX IPs to the rest of the mobile hardware system: (1)
providing an efficient path to offload an NN invocation task from the
CPU to the NNX, and (2) providing sufficient memory bandwidth
for weights and activation data. In particular, the cache/memory
interface between the accelerator and the SoC is critical, since modern DNNs typically have very large parameter sets, which demand
high memory bandwidth to feed many arithmetic units [10].
SysML’18, Feb 2018, Stanford, CA, USA
3
CASE STUDY: CONTINUOUS VISION
Computer vision (CV) tasks such as object classification, localization, and tracking are key capabilities for a number of exciting new
application domains on mobile devices, such as augmented reality
(AR) and virtual reality (VR). However, the computational cost of
modern CV CNNs far exceed the severely limited power budget of
mobile devices, especially for real time (e.g., 1080p/60fps). This is
true even with a dedicated CNN hardware accelerator [12, 13, 15].
To achieve real-time object detection with high accuracy on mobile devices, our key idea is to reduce the total number of expensive
CNN inferences through system-level optimizations. This is done
by harnessing the synergy between different hardware IPs of the
Norm. Energy
1.0
70
0.8
56
NNX
Memory
Sensor + ISP
0.6
0.4
42
28
0.2
0.0
v2
LO
YO
14
-2
EW
-4
EW
-8
EW
-1
EW
6
-3
EW
2
Ti
LO
O
Y
ny
0
Average Precision (%)
Our key insight is to leverage the L3 cache in the CPU cluster as
a bandwidth filter rather than directly interfacing the NNX with
the DRAM as some state-of-the-art NNX designs currently do [11].
This is achieved through the ARM Accelerator Coherency Port
(ACP) [3], which is available on ARM CPU clusters and allows
attached accelerators to load and store data directly to and from
the L3. In this way, the NNX can also take advantage of L3 cache
features in the CPU cluster such as prefetching,cache stashing, and
partitioning. Furthermore, ACP is also low-latency, such that the
CPUs and NNX can work together closely on data in the L3 cache.
Software Abstractions Mobile SoC hardware and ML algorithms are both evolving rapidly. Therefore, it is paramount to
present a programming interface that minimizes disruption to application developers. This would make new hardware features easy
to use, and provide compatibility across a range of mobile SoCs.
The key of such a programming interface is a clear abstraction
that allows applications to execute DNN jobs efficiently on (one of
many) hardware accelerators, or fall back to execution on a CPU or
GPU. The AndroidNN API [1] provides an example of this principle,
by abstracting common DNN kernels such as convolution, and
scheduling execution through a hardware abstraction layer (HAL).
We provide optimized implementations for DNN kernels in the
form of ARM Compute Library [4]. The library takes advantage
of recent ARM ISA enhancements that provides new instructions
for essential linear algebra operations behind DNNs, such as the
new dot-product instructions in Arm’s NEON SIMD extension [5].
Finally, we provide IP-specific drivers and HAL implementations
to support the AndroidNN API.
System Optimizations While adding specialized NNX hardware IP to the SoC improves kernel-level performance and efficiency, the DNN component is typically only one stage in a larger
end-to-end application pipeline. For instance in computer vision,
many on/off-chip components such camera sensors, Image Signal
Processors (ISP), DRAM memory, as well as the NNX have to collaborate together to deliver real time vision capabilities. The NNX IP
itself constitutes at most half of the total power/energy consumption, and there are additional opportunities in jointly optimizing
the whole system.
Once we expand our scope to the system level, we expose new optimization opportunities by exploiting functional synergies across
different IP blocks; these optimizations are not obvious when considering the NNX in isolation. We will demonstrate this principle
in the following case study.
Yuhao Zhu, Matthew Mattina, and Paul Whatmough
Figure 2: Cross-IP optimization of object detection on a mobile SoC allows over 40% reduction in energy (left) with less
than 1% accuracy loss (right).
vision subsystem. In particular, we leverage the fact that the image
signal processor (ISP) inherently calculates motion vectors (MV) for
use in its temporal denoising algorithms. Usually MVs are discarded
after de-noising, but we elect to expose them at the system level.
Instead of using CNN inference on every frame to track the movement of objects, we reuse the MVs to extrapolate the movement
of objects detected in the previous video frame, without further
CNN inferencing for the current frame. As we increase the number
of consecutively extrapolated frames (extrapolation window, or
EW), the total number of CNN inferences is reduced, leading to
performance and energy improvements.
We also leverage the ARM ACP interface [3] to use the LLC for
inter-layer data reuse (e.g., feature maps), which would otherwise
be spilled to the DRAM from the NN accelerator’s local SRAM.
A typical L3 size in mobile devices is about 2 MB [14] and ACP
provides around 20 GB/s of bandwidth, which is sufficient to capture
the reuse of most layers in today’s object detection CNNs. This
design greatly reduces DRAM and system power consumption.
Finally, we present software support that abstracts away the
hardware implementation details. As Fig. 1 shows, the high-level
CV libraries are unmodified, keeping the AndroidNN interface
unchanged. We implement specific driver and HAL modifications
that our hardware augmentation entails.
We evaluated the system-level optimizations on an in-house SoC
simulator, which we calibrated with measurements on the Jetson
TX2 development board [6]. We use commonly-used benchmarks
such as VOT [9] and OTB [8] as well as our internal datasets. Results
in Fig. 2 show that compared to state-of-the-art object detection
frameworks such as YOLO [16] that execute an entire CNN for
every frame, our system reduces the energy by over 40% with less
than 1% accuracy loss at an extrapolation window (EW) size of
two. The energy saving is greater as EW increases, while accuracy
degrades. Compared to the conventional approach of reducing the
compute intensity by down-scaling the network (e.g., TinyYOLO,
which is ∼ 5 × simpler), our system achieves higher energy savings
and higher accuracy.
4
CONCLUSION
Efficiently supporting demanding ML workloads on energy-constrained
mobile devices requires careful attention to the overall system design. We emphasized three key research priorities: accelerator interfacing, software abstractions, and cross-IP optimizations.
Mobile Machine Learning Hardware at ARM:
A Systems-on-Chip (SoC) Perspective
REFERENCES
[1] [n. d.]. Android Neural Networks API. ([n. d.]). https://developer.android.com/
ndk/guides/neuralnetworks/index.html
[2] [n. d.].
Apple’s Neural Engine Infuses the iPhone with
AI Smarts.
([n. d.]).
https://www.wired.com/story/
apples-neural-engine-infuses-the-iphone-with-ai-smarts/
[3] [n. d.]. ARM Accelerator Coherency Port. ([n. d.]). http://infocenter.arm.com/
help/index.jsp?topic=/com.arm.doc.ddi0434a/BABGHDHD.html
[4] [n. d.]. Arm Compute Library. ([n. d.]). https://github.com/ARM-software/
ComputeLibrary
[5] [n. d.]. Arm NEON Technology. ([n. d.]). https://developer.arm.com/technologies/
neon
[6] [n. d.]. Jetson TX2 Module. ([n. d.]). http://www.nvidia.com/object/
embedded-systems-dev-kits-modules.html
[7] [n. d.]. Second Version of HoloLens HPU will Incorporate AI Coprocessor for
Implementing DNNs. ([n. d.]). https://www.microsoft.com/en-us/research/blog/
second-version-hololens-hpu-will-incorporate-ai-coprocessor-implementing-dnns/
[8] [n. d.]. Visual Tracker Benchmark. ([n. d.]). http://cvlab.hanyang.ac.kr/tracker_
benchmark/datasets.html
[9] [n. d.]. VOT2014 Benchmark. ([n. d.]). http://www.votchallenge.net/vot2014/
[10] Tianshi Chen, Zidong Du, Ninghui Sun, Jia Wang, Chengyong Wu, Yunji Chen,
and Olivier Temam. 2014. DianNao: A Small-footprint High-throughput Accelerator for Ubiquitous Machine-learning. In Proc. of ASPLOS.
[11] Yu-Hsin Chen, Joel Emer, and Vivienne Sze. 2016. Eyeriss: A Spatial Architecture
for Energy-efficient Dataflow for Convolutional Neural Networks. In Proc. of
ISCA.
[12] Yu-Hsin Chen, Tushar Krishna, Joel S Emer, and Vivienne Sze. 2017. Eyeriss:
An Energy-Efficient Reconfigurable Accelerator for Deep Convolutional Neural
Networks. In Proc. of ISSCC.
[13] Giuseppe Desoli, Nitin Chawla, Thomas Boesch, Surinder-pal Singh, Elio Guidetti,
Fabio De Ambroggi, Tommaso Majo, Paolo Zambotti, Manuj Ayodhyawasi,
Harvinder Singh, et al. 2017. 14.1 A 2.9 TOPS/W Deep Convolutional Neural Network SoC in FD-SOI 28nm for Intelligent Embedded Systems. In Proc. of
ISSCC.
[14] Matthew Halpern, Yuhao Zhu, and Vijay Janapa Reddi. 2016. Mobile CPU’s Rise
to Power: Quantifying the Impact of Generational Mobile CPU Design Trends
on Performance, Energy, and User Satisfaction. In Proc. of HPCA.
[15] Bert Moons, Roel Uytterhoeven, Wim Dehaene, and Marian Verhelst. 2017. 14.5
Envision: A 0.26-to-10TOPS/W Subword-parallel Dynamic-Voltage-AccuracyFrequency-Scalable Convolutional Neural Network Processor in 28nm FDSOI. In
Proc. of ISSCC.
[16] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. 2016. You Only
Look Once: Unified, Real-Time Object Detection. In Proc. of CVPR.
[17] P. N. Whatmough, S. K. Lee, H. Lee, S. Rama, D. Brooks, and G. Y. Wei. 2017.
14.3 A 28nm SoC with a 1.2GHz 568nJ/prediction sparse deep-neural-network
engine with >0.1 timing error rate tolerance for IoT applications. In 2017 IEEE
International Solid-State Circuits Conference (ISSCC). 242–243. https://doi.org/10.
1109/ISSCC.2017.7870351
SysML’18, Feb 2018, Stanford, CA, USA
| 9cs.NE
|
Correctness and completeness of logic programs
Version of 24-04-2015
Wlodzimierz Drabent,
arXiv:1412.8739v2 [cs.LO] 17 May 2015
Institute of Computer Science, Polish Academy of Sciences, and IDA, Linköping University, Sweden
We discuss proving correctness and completeness of definite clause logic programs. We propose a method
for proving completeness, while for proving correctness we employ a method which should be well known
but is often neglected. Also, we show how to prove completeness and correctness in the presence of SLD-tree
pruning, and point out that approximate specifications simplify specifications and proofs.
We compare the proof methods to declarative diagnosis (algorithmic debugging), showing that approximate specifications eliminate a major drawback of the latter. We argue that our proof methods reflect
natural declarative thinking about programs, and that they can be used, formally or informally, in everyday programming.
CCS Concepts: •Theory of computation → Program reasoning; •Software and its engineering →
Constraint and logic languages; Software verification; Software testing and debugging;
General Terms: Theory, Verification
Additional Key Words and Phrases: logic programming, declarative programming, program completeness,
program correctness, specifications, declarative diagnosis / algorithmic debugging
1. INTRODUCTION
The purpose of this paper is to present proof methods to deal with the correctness related
issues of definite clause logic programs. The goal is to devise methods which are 1. simple
and applicable in the practice of Prolog programming, and 2. declarative (i.e. not referring
to any operational semantics), in other words depending only on the logical reading of
programs.
The notion of program correctness (in imperative and functional programming) divides in
logic programming into correctness and completeness. Correctness means that all answers
of the program are compatible with the specification, completeness – that the program
produces all the answers required by the specification. A specification may be approximate:
for such a specification some answers are allowed, but not required to be computed. We
discuss various advantages of using approximate specifications.
There exists a simple, natural, and moreover complete proof method for correctness [Clark
1979]. It should be well known but has been mostly neglected. Surprisingly little has been
done on reasoning about completeness. See Section 8 for the related work. In this paper, for
proving correctness we employ Clark’s method. For proving completeness we introduce a
simplification of the method of [Drabent and Milkowska 2005]. We introduce a new notion
of semi-completeness; semi-completeness and termination imply completeness. We discuss
a few sufficient conditions for completeness. We also propose a way of proving that completeness is preserved under pruning of SLD-trees (due to e.g. the cut of Prolog). This is
augmented by a way of proving correctness for such trees, taking into account that pruning
may remove some answers.
We would like to treat logic programming as a declarative programming paradigm, and
to reason about programs declaratively, i.e. independently from their operational semantics.
The method of proving correctness is declarative, and so is the method of proving semicompleteness. However some of the sufficient conditions for completeness are not declarative,
as they, roughly speaking, refer to an operational notion of program termination. Notice that
in most practical cases termination has to be established anyway. So such a way of proving
completeness may be a reasonable compromise between declarative and non-declarative
reasoning.
This paper provides small examples; for a more substantial application of the presented
proof methods see [Drabent 2012a]. That paper presents a construction of a practical Prolog
2
24-04-2015
program, the SAT solver of [Howe and King 2012]. Starting from a formal specification, a
logic program is constructed hand in hand with its correctness and completeness proofs.
Then control is added (delays, and pruning SLD-trees) to obtain the final Prolog program.
The added control preserves correctness, while the approach presented here makes it possible to formally prove that also completeness is preserved. This example illustrates also
the usefulness of approximate specifications. The consecutively constructed versions of the
program are not equivalent – their main predicates define different relations. So it is not
a case of a program transformation which preserves the semantics of a program. What is
preserved is correctness and completeness w.r.t. the same approximate specification.
Closely related to proving correctness and completeness of programs is program diagnosis, i.e. locating errors in programs. We show a correspondence between the presented proof
methods and declarative diagnosis (called also algorithmic debugging). We show how an important drawback of declarative diagnosis can be easily overcome by employing approximate
specifications.
Preliminaries. In this paper we consider definite clause programs (i.e. logic programs
without negation). For (standard) notation and definitions the reader is referred to [Apt
1997]. Given a predicate symbol p, by an atom for p (or p-atom) we mean an atom whose
predicate symbol is p, and by a clause for p – a clause whose head is an atom for p. The set
of the clauses for p in the program under consideration is called procedure p.
We assume a fixed alphabet of function and predicate symbols, not restricted to the
symbols from the considered program. The only requirement is that the Herbrand universe is
nonempty. In particular, the set of function symbols may be finite or infinite. The Herbrand
universe will be denoted by HU, the Herbrand base by HB, and the sets of all terms,
respectively atoms, by T U and T B. By ground(E) we mean the set of ground instances of
a term, atom, etc E. For a program P , ground(P ) is the set of ground instances of the clauses
of P . The least Herbrand model of a program P will be denoted by MP S
. It is the least fixed
∞
point of the immediate consequence operator TP : 2HB → 2HB ; MP = i=1 TPi (∅).
By a computed (respectively correct) answer for a program P and a query Q we mean an
instance Qθ of Q where θ is a computed (correct) answer substitution [Apt 1997] for Q and
P . We usually say just answer as each computed answer is a correct one, and each correct
answer (for Q) is a computed answer (for Q or for some its instance). Thus, by soundness
and completeness of SLD-resolution, Qθ is an answer for P iff P |= Qθ.
By “declarative” (property, reasoning, . . . ) we mean referring only to logical reading of
programs, thus abstracting from any operational semantics. So Q being an answer for P
is a declarative property. Properties depending on the order of atoms in clauses will not
be considered declarative, as logical reading does not distinguish equivalent formulae, like
α ∧ β and β ∧ α.
Names of variables begin with an upper-case letter. We use the list notation of Prolog.
So [t1 , . . . , tn ] (n ≥ 0) stands for the list of elements t1 , . . . , tn . Only a term of this form is
considered a list. (Thus terms like [a, a|X], or [a, a|a], where a 6= [ ] is a constant, are not
lists). N stands for the set of natural numbers, and Z for the set of integers; f : A ,→ B states
that f is a partial function from A to B. We use a name “proof tree” instead of “implication
tree” of [Apt 1997].
Organization of the paper. The next section introduces the basic notions. Section 3 discusses approximate specifications. Sections 4 and 5 deal with proving, respectively, correctness and completeness of programs. Proving completeness and correctness in the context
of pruning of SLD-trees is discussed in Section 6. Section 7 discusses declarative diagnosis.
Section 8 is an overview of related work, Section 9 provides additional discussion of the proposed methods, and the next section concludes the paper. Two appendices include results
about completeness of the proof methods, and proofs missing in the main text.
24-04-2015
3
2. CORRECTNESS AND COMPLETENESS
This section introduces basic notions of specification, correctness and completeness, and
relates them to the answers of programs.
The purpose of a logic program is to compute a relation, or a few relations. It is convenient
to assume that the relations are on the Herbrand universe. A specification should describe
them. In most cases all the relations defined by a program should be described, an n-ary
relation for each n-argument predicate symbol. It is convenient to describe such family of
relations as a Herbrand interpretation S: a tuple (t1 , . . . , tn ) is in the relation corresponding
to a predicate symbol p iff p(t1 , . . . , tn ) ∈ S. So by a (formal) specification we mean a
Herbrand interpretation; given a specification S, each A ∈ S is called a specified atom.
In imperative programming, correctness usually means that the program results are as
specified. In logic programming, due to its non-deterministic nature, we actually have two
issues: correctness (all the results are compatible with the specification) and completeness (all the results required by the specification are produced). In other words, correctness
means that the relations defined by the program are subsets of the specified ones, and completeness means inclusion in the opposite direction. In terms of specified atoms and the
least Herbrand model MP of a program P , we define:
Definition 2.1. Let P be a program and S ⊆ HB a specification. P is correct w.r.t. S
when MP ⊆ S; it is complete w.r.t. S when MP ⊇ S.
We will sometimes skip the specification when it is clear from the context. We propose to
call a program fully correct when it is both correct and complete.
It is important to relate the answers of programs to their correctness and completeness.
The relation is not trivial as, for a program P and a query Q, the equivalence MP |= Q iff
P |= Q is not true in some cases.1
Lemma 2.2. Let P be a program, and Q a query such that
(1 ) Q contains exactly k ≥ 0 (distinct) variables, and the underlying language has (at least)
k constants not occurring in P, Q, or a non-constant function symbol not occurring
in P, Q.
Then MP |= Q iff P |= Q.
In particular, the equivalence holds for ground queries. It also holds when the set of function
symbols is infinite (and P is finite). Note that condition (1) is satisfied by any Prolog
program and query, so the equivalence holds in practice. See Appendix B for a proof of the
lemma, and [Drabent 2015b] for a slightly more general result and further discussion.
Now we are ready to relate program answers to the correctness and completeness of
programs. (Remember that a query Q is an answer for P iff P |= Q.)
Proposition 2.3. Let P be a program, and S a specification.
(1 ) P is correct w.r.t. S iff P |= Q implies S |= Q for each query Q.
(2 ) P is complete w.r.t. S iff S |= Q implies P |= Q for any ground query Q.
(3 ) If the set of function symbols of the underlying language is infinite and P is finite then
P is complete w.r.t. S iff S |= Q implies P |= Q for any query Q.
(4 ) P is complete w.r.t. S iff
S |= Q implies P |= Q for any query Q such that condition (1) of Lemma 2.2 holds.
Within this proposition, “query Q” can be replaced by “atomic query Q”.
1
For instance assume a two element alphabet of function symbols, with a unary f and a constant a. Take
P = { p(a). p(f (Y )).}. Then MP = HU . Hence MP |= p(X) but P 6|= p(X).
It is quite often believed that the equivalence always holds, See for instance erroneous Th. 2 in [Paschke
2011], or the remark on the least Herbrand models on p. 512 of [Naish 2006].
4
24-04-2015
Proof. For the first case, assume MP ⊆ S (i.e. P is correct). If P |= Q then MP |= Q,
hence S |= Q. Conversely, assume that P |= Q implies S |= Q for any atomic query Q. If
A ∈ MP then P |= A (by Lemma 2.2), hence S |= A. Thus MP ⊆ S.
For cases 2, 3, 4 it is sufficient to show the “if” of case 2 and the “only-if” of case 4. For
the “only-if” of case 4, assume S ⊆ MP (P complete) and S |= Q (for a Q as described).
Then MP |= Q and, by Lemma 2.2, P |= Q.
For the “if” of case 2 assume the right hand side of the equivalence, for atomic queries.
So for every ground atom A we have that S |= A implies P |= A, hence MP |= A. Thus
S ⊆ MP .
The following auxiliary notions will be used.
Definition 2.4. The set of p-atoms from a specification S will be called the specification
for p given by S.
A predicate (or a procedure) p in P is correct w.r.t. S when each p-atom of MP is in S,
and complete w.r.t. S when each p-atom of S is in MP .
An answer Q is correct w.r.t. S when S |= Q.
P is complete for a query Q w.r.t. S when S |= Qθ implies that Qθ is an answer for P ,
for any ground instance Qθ of Q.
Informally, complete for Q means that all the answers for Q required by the specification
are computed. Note that a program is complete w.r.t. S iff it is complete w.r.t. S for any
query iff it is complete w.r.t. S for any query A ∈ S.
There is an alternative way to define program completeness, namely by S |= Q implying
P |= Q, for every query Q [Drabent and Milkowska 2005]. It was not chosen here, as the
two notions of completeness are equivalent in the practical case of an infinite alphabet of
function symbols. Also, it seems that the chosen version leads to simpler sufficient conditions
for completeness.
3. APPROXIMATE SPECIFICATIONS
This section first shows that requiring specifications to be exact leads to difficulties and
inconveniences. Then usefulness of approximate specifications is discussed.
Notice that if a program P is both correct and complete w.r.t. S then MP = S and
the specification describes exactly the relations defined by P . Often it is difficult (and not
necessary) to specify the relations exactly. The relations defined by the program are often
not exactly the ones intended by the programmer. Such discrepancy is however not an error,
because for all intended usages of the program the answers are the same as for a program
defining the intended relations. For certain atoms it is irrelevant whether they are, or are
not, in MP . We illustrate this by some examples.
Example 3.1. The well-known program
APPEND = { app( [H|K], L, [H|M ] ) ← app( K, L, M ).
app( [ ], L, L ). }
does not define the relation of list concatenation (provided that the underlying alphabet
contains some function symbol not occurring in the program, e.g. a constant a). For instance,
APPEND |= app([ ], a, a). In other words, APPEND is not correct w.r.t.
0
SAPPEND
= { app(k, l, m) ∈ HB | k, l, m are lists, k ∗ l = m },
0
where k ∗ l stands for the concatenation of lists k, l. It is however complete w.r.t. SAPPEND
,
and correct w.r.t.
0
SAPPEND = { app(k, l, m) ∈ HB | if l or m is a list then app(k, l, m) ∈ SAPPEND
}.
Proofs are given later on in Examples 4.3, 5.8. Correctness w.r.t. SAPPEND and completeness
0
w.r.t. SAPPEND
are sufficient to show that APPEND will produce the required results when
24-04-2015
5
used to concatenate or split lists.2 Understanding the exact relation defined by the program
is unnecessary. See [Drabent and Milkowska 2005] for further discussion.
0
Note that SAPPEND is the union of SAPPEND
and the complement HB \ T of the set
T = { app(k, l, m) ∈ HB | l or m is a list }. The intuition is that it is irrelevant whether
the atoms from HB \ T are in MAPPEND .
When we are only interested in using APPEND to concatenate lists, the specification for
correctness can be weakened by replacing “if l or m is a list” by “if l is a list”. Similarly, if
only list splitting is of interest then “if m is a list” can be used.
A similar example is given in Sect. 7.2 where in an insertion sort program it is irrelevant
how elements should be inserted into an unsorted list.
Example 3.2. Programmers often do not know the exact relations defined by logic programs. For many programs in textbooks, the exact relations defined by the predicates are not
explained. For instance this concerns the programs in Chapter 3.2 of [Sterling and Shapiro
1994] defining predicates member/2, sublist/2, prefix/2, suffix/2, and append/3. Similarly to the case of APPEND described above, the relations defined by the programs are
not what we understand as the list membership relation, the sublist relation etc. The programs define certain supersets of these relations. (They are complete w.r.t. specifications
describing the relations.) Suitable specifications for correctness can be easily constructed,
similarly as for APPEND. Understanding the exact relations defined by the programs is
unnecessary, and they are not discussed in the book.
The exact relations defined by programs are quite often misunderstood. For instance, in
[Deville and Lau 1994, Ex. 15] it is claimed that the constructed program, let us call it P ,
defines the relation of list inclusion
l1 , l2 are lists,
I(included ) = included (l1 , l2 ) ∈ HB
.
every element of l1 belongs to l2
In our terms, this means that predicate included of P is correct and complete w.r.t.
I(included ). However the correctness does not hold; P contains a clause included (L1 , L2 ) ←
L1 = [ ]., thus P |= included ([ ], t) for any term t.
The examples show that it is often not necessary to know the semantics of a program
exactly, and that its approximate description is sufficient. An approximate specification
is a pair of specifications Scompl , Scorr called, respectively, specification for completeness and
specification for correctness. The intention is that the program is complete w.r.t. the former,
and correct w.r.t. the latter. We say that a program P is fully correct w.r.t. Scompl , Scorr
when Scompl ⊆ MP ⊆ Scorr . In other words, the specifications Scompl , Scorr describe,
respectively, which atoms have to be computed, and which are allowed to be computed;
2
Formally: In such usage, in a query Q = app(s, t, u) the argument t (and s) is a list, or u is a list. By
the correctness w.r.t. SAPPEND and Prop. 2.3(1), for any answer Qθ of APPEND we have SAPPEND |= Qθ.
0
0
As tθ or uθ is a list, each ground instance of Qθ is in SAPPEND
. Hence SAPPEND
|= Qθ. This means that
sθ, tθ, uθ are lists and uθ is the concatenation of sθ, tθ.
0
Conversely, let us use the fact that APPEND is complete w.r.t. SAPPEND
for any set of function symbols
that includes [ ] and [ | ]. Whenever uσ is a ground list which is the concatenation of sσ and tσ (equivalently
0
SAPPEND
|= Qσ) then, by Prop. 2.3(2), Qσ is an answer of APPEND. The same holds for a non-ground
list uσ, provided that the condition of Prop. 2.3(3) or 2.3(4) holds. In particular, when u = [u1 , . . . , un ] is
a list and s, t are distinct variables then each app([u1 , . . . , ui ], [ui+1 , . . . , un ], u) (i = 0, . . . , n) is an answer
of APPEND. In other words, the program produces each splitting of u.
When the conditions of Prop. 2.3 are not satisfied, it is possible that a program P is complete
0
0
w.r.t. SAPPEND
and SAPPEND
|= Q, but Q is not an answer of P . This happens for instance when [ ]
and [ | ] are the only function symbols, Q = app([X], [ ], [X]), and P consists of three clauses: app( [ ], L, L ).;
app( [[ ]|K], L, [[ ]|M ] ) ← app( K, L, M ).; app( [[H|T ]|K], L, [[H|T ]|M ] ) ← app( K, L, M ). (A completeness
proof for P is similar to that of Ex. 5.8).
6
24-04-2015
for the atoms from Scorr \ Scompl the semantics of the program is irrelevant. Sometimes,
by abuse of terminology, a single specification Scorr or Scompl will be called approximate,
especially when the intention is that that the specification is distinct from MP .
Usually programs are intended to be used with queries Q such that if Q is atomic then it
has no instances in Scorr \Scompl . For such queries a program fully correct w.r.t. Scompl , Scorr
behaves like a program correct and complete w.r.t. Scompl :
Proposition 3.3. Let Scompl , Scorr be an approximate specification, P be a program
complete w.r.t. Scompl and correct w.r.t. Scorr , and P 0 a program correct and complete
w.r.t. Scompl . Let Q be an atomic query such that ground(Q) ∩ (Scorr \ Scompl ) = ∅. Let Qθ
be an instance of Q such that condition (1) of Lemma 2.2 holds for P with Qθ and for P 0
with Qθ. Then Qθ is an answer for P iff Qθ is an answer for P 0 .
In particular, the equivalence holds when the set of function symbols is infinite (and P, P 0
are finite).
Proof. For any program P 00 such that P 00 and Q satisfy (1), it holds that P 00 |= Qθ iff
ground(Qθ) ⊆ MP 00 . Note that MP 0 = Scompl ⊆ MP ⊆ Scorr . We have ground(Q)∩Scorr ⊆
Scompl . Hence ground(Qθ) ∩ MP ⊆ Scompl . Thus ground(Qθ) ⊆ MP iff ground(Qθ) ⊆
Scompl . Hence P |= Qθ iff P 0 |= Qθ.
Comments. An obvious role of approximate specifications is to describe selected properties
of programs. Ex. 4.2 below shows a case where only the lengths of lists are taken into account,
not their elements. Note that such a property cannot be expressed by an exact specification.
Typed logic programming [Pfenning 1992] is outside of the scope of this paper. But it is
worth mentioning here that, for some programs, introducing types may change their semantics, so that they define exactly the intended relations. For instance in some appropriate
typed logic, the APPEND program of Ex. 3.1 (with the arguments of app typed to be lists)
would define exactly the relation of list concatenation Thus such typed APPEND would
0
be correct and complete w.r.t. SAPPEND
, and the need for an approximate specification
disappears. On the other hand, to obtain a similar result for procedure insert of Ex. 7.2 it
is necessary to employ a type of sorted integer lists, which seems to be outside of the scope
of known type systems for logic programs.
Often, while developing a program, the exact relations to be defined should not be fixed in
advance. This would mean taking some design decisions too early. For instance, constructing
0
(or with SAPPEND ) as an exact specification would lead
an append program with SAPPEND
to a program less efficient and more complicated than APPEND. See also Ex. 7.2. A larger
example is shown in [Drabent 2012a; 2012b]: a program is derived by stepwise refinement,
and the semantics of (the common predicates in) the consecutive versions of the program
differ. What does not change is an approximate specification (of the common predicates);
the correctness and completeness is preserved in all the versions.
This example suggests a generalization of the paradigm of program development by transformations which preserve program semantics [Pettorossi et al. 2010, and the references
therein]. It shows that it is useful and natural to use transformations which instead preserve correctness and completeness w.r.t. an approximate specification.
Importance of approximate specifications for declarative diagnosis is discussed in Sect. 7.
4. PROVING CORRECTNESS
To prove correctness we use the following property [Clark 1979].
Theorem 4.1 (Correctness). A sufficient condition for a program P to be correct
w.r.t. a specification S is
S |= P.
24-04-2015
7
So the sufficient condition states that, for each ground instance H ← B1 , . . . , Bn (m ≥ 0)
of a clause of the program, if B1 , . . . , Bn ∈ S then H ∈ S. Equivalently, TP (S) ⊆ S.
Proof. S |= P means that S is a Herbrand model of P , thus MP ⊆ S.
Example 4.2. Consider a program SPLIT:
s([ ], [ ], [ ]).
s([X|Xs], [X|Y s], Zs) ← s(Xs, Zs, Y s).
(1)
(2)
Let |l| denote the length of a list l. A specification
S = { s(l, l1 , l2 ) | l, l1 , l2 are lists, 0 ≤ |l1 | − |l2 | ≤ 1 }
describes how the sizes of the last two arguments of S are related. To show, by Th. 4.1, that
the program is correct w.r.t. S, consider a ground instance s([h|t], [h|t2 ], t1 ) ← s(t, t1 , t2 ) of
(2). Assume s(t, t1 , t2 ) ∈ S. Thus [h|t], [h|t2 ], t1 are lists. Let m = |t1 | − |t2 |. As m ∈ {0, 1},
we have |[h|t2 ]| − |t1 | = 1 − m ∈ {0, 1}. So the head s([h|t], [h|t2 ], t1 ) is in S. The proof for
(1) is trivial.
A stronger specification with respect to which SPLIT is correct is shown below.
Example 4.3. To show that program APPEND is correct w.r.t. SAPPEND from Ex. 3.1,
take a (H ← B) ∈ ground(APPEND) such that B = app(k, l, m) ∈ SAPPEND . So H =
app( [h|k], l, [h|m] ). If l or [h|m] is a list then k, l, m are lists and k ∗ l = m. So [h|k], l, [h|m]
are lists and [h|k] ∗ l = [h|m]. Thus H ∈ SAPPEND .
For the other clause of the program, take app([ ], m, m) ∈ ground(APPEND). If m is a
list then [ ] ∗ m = m; so app([ ], m, m) ∈ SAPPEND ).
Example 4.4. Now we show correctness of SPLIT w.r.t. a more precise specification
s([t1 , . . . , t2n ], [t1 , · · · , t2n−1 ], [t2 , · · · , t2n ]),
n ≥ 0,
SSPLIT =
,
s([t1 , . . . , t2n+1 ], [t1 , · · · , t2n+1 ], [t2 , · · · , t2n ]) t1 , . . . , t2n+1 ∈ HU
where [tk , · · · , tl ] denotes the list [tk , tk+2 , . . . , tl ], for k, l being both odd or both even. (When
k > l, [tk , · · · , tl ] = [tk , . . . , tl ] = [ ].) For (1) the proof is trivial. Consider a ground instance
H ← B of (2). Assume B ∈ SSPLIT . If B = s([t1 , . . . , t2n ], [t1 , · · · , t2n−1 ], [t2 , · · · , t2n ]) then
H = s([t0 , t1 , . . . , t2n ], [t0 , t2 , · · · , t2n ], [t1 , · · · , t2n−1 ]), for some t0 . If B = s([t1 , . . . , t2n+1 ],
[t1 , · · · , t2n+1 ], [t2 , · · · , t2n ]) then H = s([t0 , t1 , . . . , t2n+1 ], [t0 , t2 , · · · , t2n ], [t1 · · · , t2n+1 ]). In
both cases H ∈ SSPLIT . (To see this, rename each ti as t0i+1 .)
Completeness of this correctness proving method is discussed in Appendix A. For more
examples (among others involving difference lists and accumulators), and for further explanations, references and discussion see [Drabent and Milkowska 2005].
5. PROVING COMPLETENESS
We first introduce a notion of semi-completeness, and sufficient conditions under which
semi-completeness of a program implies its completeness. The next subsection presents a
sufficient condition for semi-completeness. For cases where completeness does not follow
from semi-completeness, we present a way of proving completeness directly.
5.1. Semi-completeness
Before discussing semi-completeness we present a few auxiliary notions, which have been
introduced in a context of proving program termination.
Definition 5.1. A level mapping is a function | |: HB → N assigning natural numbers
to atoms.
8
24-04-2015
A program P is recurrent w.r.t. a level mapping | | [Bezem 1993; Apt 1997] if, in every
ground instance H ← B1 , . . . , Bn ∈ ground(P ) of its clause (n ≥ 0), |H| > |Bi | for all
i = 1, . . . , n. A program is recurrent if it is recurrent w.r.t. some level mapping.
Definition 5.2. A program P is acceptable w.r.t. a specification S and a level mapping
| | if P is correct w.r.t. S, and for every H ← B1 , . . . , Bn ∈ ground(P ) we have |H| > |Bi |
whenever S |= B1 , . . . , Bi−1 . A program is acceptable if it is acceptable w.r.t. some level
mapping and some specification.
This definition is more general than the original one [Apt and Pedreschi 1993; Apt 1997]
which uses a model of P , instead of a specification for which P is correct. Both definitions
are equivalent.3
We now introduce a notion of semi-completeness. It may be seen as a step on the way to
proving completeness.
Definition 5.3. A program P is semi-complete w.r.t. a specification S if P is complete
w.r.t. S for (cf. Df. 2.4) any query Q for which there exists a finite SLD-tree.
In practice, the existence of a finite SLD-tree means that P with Q terminates under some
selection rule. For a semi-complete program, if a computation for a query Q terminates then
it has produced all the answers for Q required by the specification. Obviously, a complete
program is semi-complete. We also have:
Proposition 5.4 (Completeness). Let a program P be semi-complete w.r.t. S. The
program is complete w.r.t S if
(1 ) for each query A ∈ S there exists a finite SLD-tree, or
(2 ) the program is recurrent, or
(3 ) the program is acceptable (w.r.t. a specification S 0 possibly distinct from S).
Note that condition (1) is equivalent to each A ∈ S being an instance of a query Q with a
finite SLD-tree.
Proof. For a program P semi-complete w.r.t. S, the first part of condition (1) implies
that P is complete w.r.t. S for each query A ∈ S; hence S ⊆ MP . Condition (2), or (3)
implies that for any ground query each SLD-tree, respectively each LD-tree, is finite [Apt
1997]. This implies (1).
5.2. Proving semi-completeness
We need the following notion.
Definition 5.5. A ground atom H is covered by a clause C w.r.t. a specification S if H
is the head of a ground instance H ← B1 , . . . , Bn (n ≥ 0) of C, such that all the atoms
B1 , . . . , Bn are in S.
A ground atom H is covered by a program P w.r.t. S if it is covered w.r.t. S by some
clause C ∈ P . A specification S is covered by a program P if each H ∈ S is covered by P
w.r.t. S.
For instance, given a specification S = { p(si (0)) | i ≥ 0 }, atom p(s(0)) is covered both by
p(s(X)) ← p(X) and by p(X) ← p(s(X)). The notion of an atom covered by a clause stems
from [Shapiro 1983]. A specification S covered by P is called in [Apt 1993] a supported
Herbrand interpretation of P .
3
They define the same programs to be acceptable. More precisely, a program P is acceptable w.r.t. | |
and a specification S iff, according to the original definition, P is acceptable w.r.t. | | and a model I of
P . (The “only-if” case holds for each model I such that MP ⊆ I ⊆ S, the “if” case for each S such that
MP ⊆ S ⊆ I.)
24-04-2015
9
Now we are ready to present a sufficient condition for semi-completeness, which together
with Prop. 5.4 provides a sufficient condition for completeness.
Theorem 5.6 (Semi-completeness). If a specification S is covered by a program P
then P is semi-complete w.r.t. S.
The proof is presented in Appendix B. Note that the sufficient condition of Th. 5.6 is
equivalent to S ⊆ TP (S); the latter implies S ⊆ gfp(TP ). It is also equivalent to S being a
model of ONLY-IF(P ), a theory used in defining the completion of P (see e.g. [Doets 1994]
or [Drabent and Milkowska 2005] for a definition).
Example 5.7. We show that program SPLIT from Ex. 4.2 is complete w.r.t. the specification from Ex. 4.4,
n ≥ 0,
s([t1 , . . . , t2n ], [t1 , · · · , t2n−1 ], [t2 , · · · , t2n ]),
,
SSPLIT =
s([t1 , . . . , t2n+1 ], [t1 , · · · , t2n+1 ], [t2 , · · · , t2n ]) t1 , . . . , t2n+1 ∈ HU
where [tk , · · · , tl ] denotes the list [tk , tk+2 , . . . , tl ], for k, l being both odd or both even.
Atom s([ ], [ ], [ ]) ∈ SSPLIT is covered by clause (1). For n > 0, any atom
A = s([t1 , . . . , t2n ], [t1 , · · · , t2n−1 ], [t2 , · · · , t2n ]) is covered by an instance of (2) with
a body B = s([t2 , . . . , t2n ], [t2 , · · · , t2n ], [t3 , · · · , t2n−1 ]). Similarly, for n ≥ 0 and any
atom A = s([t1 , . . . , t2n+1 ], [t1 , · · · , t2n+1 ], [t2 , · · · , t2n ]), the corresponding body is B =
s([t2 , . . . , t2n+1 ], [t2 , · · · , t2n ], [t3 , · · · , t2n+1 ]). In both cases, B ∈ SSPLIT . (To see this, rename each ti as t0i−1 .) So SSPLIT is covered by SPLIT. Thus SPLIT is semi-complete w.r.t.
SSPLIT , by Th. 5.6.
Now by Prop. 5.4 the program is complete, as it is recurrent under the level mapping
| s(t, t1 , t2 )| = |t|,
where
| [h|t] | = 1 + |t|,
|f (t1 , . . . , tn )| = 0 where n ≥ 0 and f is not [ | ],
(for any ground terms h, t, t1 , . . . , tn , and any function symbol f ).
0
from Ex. 3.1. We
Example 5.8. Consider program APPEND and specification SAPPEND
0
encourage the reader to check that each A ∈ SAPPEND is covered (by APPEND w.r.t.
0
0
). Now its
SAPPEND
). Hence by Th. 5.6 the program is semi-complete (w.r.t. SAPPEND
completeness follows by Prop. 5.4, as the program is recurrent under the level mapping
|app(t, t1 , t2 )| = |t|, where |t| is as in the Ex. 5.7.
The notion of semi-completeness is tailored for finite programs. An SLD-tree for an infinite
program may be infinite, but with all branches finite. Then a slight generalization of Th. 5.6
applies (Appendix, Prop. B.2, Ex. B.3).
5.3. Proving completeness directly
Semi-completeness and completeness are declarative properties of programs. However the
sufficient conditions for completeness of Prop. 5.4(1) and 5.4(3) are not declarative. The
former refers explicitly to an operational notion of SLD-tree. The latter employs the notion
of acceptable program, which depends on the order of atoms in program clauses, and is related to LD-resolution; this should not be considered as a purely logical reading of formulae.
Only condition (2) of Prop. 5.4 can be considered declarative, as it solely depends on logical
reading of program clauses. However, this condition is often inapplicable, as many practical programs are not recurrent. Note that the sufficient condition for semi-completeness of
Th. 5.6 is declarative.
We are interested in declarative reasoning about logic programs. So we present a declarative sufficient condition for completeness. Moreover, it is applicable in cases where completeness does not follow from semi-completeness. We first generalize the notion of level
10
24-04-2015
mapping by allowing that it is a partial function | |: HB ,→ N assigning natural numbers to
some atoms.
Definition 5.9. A ground atom H is recurrently covered by a program P w.r.t. a
specification S and a level mapping | |: HB ,→ N if H is the head of a ground instance
H ← B1 , . . . , Bn (n ≥ 0) of a clause of the program, such that |H|, |B1 |, . . . |Bn | are defined,
B1 , . . . , Bn ∈ S, and |H| > |Bi | for all i = 1, . . . , n.
For instance, given a specification S = { p(si (0)) | i ≥ 0 }, atom p(s(0)) is recurrently
covered by a program { p(s(X)) ← p(X). p(X) ← p(X).} under a level mapping for which
|p(si (0))| = i. No atom is recurrently covered by a program { p(X) ← p(X).}. Obviously, if
H is recurrently covered by P then it is covered by P .
The following theorem is a reformulation of Th. 6.1 from [Deransart and Maluszyński
1993] (derived there within a sophisticated theory relating attribute grammars and logic
programs).
Theorem 5.10 (Completeness 2). If, under some level mapping | |: HB ,→ N, all the
atoms from a specification S are recurrently covered by a program P w.r.t. S then P is
complete w.r.t. S.
Proof. Note first that |A| is defined for all A ∈ S. We show that A ∈ MP for each
A ∈ S, by induction on |A|. Let n ≥ 0, assume that B ∈ MP whenever B ∈ S and
|B| < n. Take an A ∈ S with |A| = n. As A is recurrently covered, there is a clause instance
A ← B1 , . . . , Bm (m ≥ 0) with Bi ∈ S and |Bi | < n for each i. So B1 , . . . , Bm ∈ MP , and
A ∈ MP .
Example 5.11. Consider a directed graph E. As a specification for a program describing
reachability in E, take S = Sp ∪Se , where Sp = { p(t, u) | there is a path from t to u in E },
and Se = { e(t, u) | (t, u) is an edge in E }. Let P consist of a procedure p:
p(X, X).
p(X, Z) ← e(X, Y ), p(Y, Z).
and a procedure e, which is a set of unary clauses describing the edges of the graph. Assume
the latter is complete w.r.t. Se . Notice that when E has cycles then infinite SLD-trees cannot
be avoided, and completeness of P cannot be shown by Prop. 5.4.
To apply Th. 5.10, let us define a level mapping for the elements of S such that |e(t, u)| = 0
and |p(t, u)| is the length of a shortest path in E from t to u (so |p(t, t)| = 0). Consider
an atom p(t, u) ∈ S where t 6= u. Let t = t0 , t1 , . . . , tn = u be a shortest path from t to u.
Then e(t, t1 ), p(t1 , u) ∈ S, |p(t, u)| = n, |e(t, t1 )| = 0, and |p(t1 , u)| = n − 1. Thus p(t, u) is
recurrently covered by P w.r.t. S and | |. The same trivially holds for the remaining atoms
of S, as they are instances of unary clauses of P . By Th. 5.10, P is complete w.r.t. S.
Note that P is not correct w.r.t. S, unless HU is the set of nodes of E (as p(t, t) 6∈ S when
t is not a node of the graph). This, again, illustrates usefulness of approximate specifications.
Completeness of the presented completeness proof methods is discussed in Appendix A. In
particular it is shown that, in a certain sense, the sufficient condition for semi-completeness
of Th. 5.6 is a necessary condition for completeness.
6. PRUNING
Pruning some parts of SLD-trees is often used to improve efficiency of programs. In Prolog
it is implemented by using the cut, the if-then-else construct, or by built-ins, like once/1.
Pruning preserves the correctness of a logic program, it also preserves termination under a
given selection rule, but may violate the program’s completeness. Consequently, the answers
of a program P executed with pruning may be correct w.r.t. a specification S for which P is
24-04-2015
11
not correct; a related programming technique is known as “red cuts” [Sterling and Shapiro
1994].
This section deals with reasoning about program answers in the presence of pruning.
The focus is on proving completeness. We begin with a formalization of pruning. Then we
introduce a sufficient condition for completeness, and present examples. The next subsection
identifies a class of cases where the sufficient condition does not hold (despite pruned trees
being complete), and shows how a completeness proof can be obtained anyway. Then we
present a sufficient condition for completeness dealing also with infinite pruned trees. We
conclude with an approach to prove correctness of the answers of pruned trees. In some
examples it is assumed that the reader is familiar with the operational semantics of the cut
in Prolog.
6.1. Pruned SLD-trees
By a pruned SLD-tree for a program P and a query Q we mean a tree with the root Q
which is a connected subgraph of an SLD-tree for P and Q. By an answer of a pruned SLDtree we mean the computed answer of a successful SLD-derivation which is a branch of the
tree. We will say that a pruned SLD-tree T with root Q is complete w.r.t. a specification S
if, for any ground Qθ, S |= Qθ implies that Qθ is an instance of an answer of T . Informally,
such a tree produces all the answers for Q required by S.
To facilitate reasoning about the answers of pruned SLD-trees, we will view pruning as
applying only certain clauses while constructing the children of a node. So we introduce
subsets Π1 , . . . , Πn of P . The intention is that for each node the clauses of one Πi are used.
Programs Π1 , . . . , Πn may be not disjoint.
. . . ,A ,. . .
P
Πi
···
not pruned
···
pruned
Definition 6.1. Given programs Π1 , . . . , Πn (n > 0), a c-selection rule is a function
assigning to a query Q0 an atom A in Q0 and one of the programs ∅, Π1 , . . . , Πn .
A csSLD-tree (cs for clause selection) for a query Q and programs Π1 , . . . , Πn , via a cselection rule R, is constructed as an SLD-tree, but for each node its children are constructed
using the program selected by the c-selection rule. An answer of the csSLD-tree is the answer
of a successful derivation which is a branch of the tree.
A c-selection rule may choose the empty program, thus making a given node a leaf. Note
that program Πi selected for a node Q0 may contain clauses whose heads do not unify
with the selected atom of Q0 (thus inapplicable in constructing the
S children of the node). A
csSLD-tree for Q and Π1 , . . . , Πn is a pruned SLD-tree for Q and i Πi . Conversely, for each
pruned SLD-tree T for Q and a (finite) program P there exist n > 0, and Π1 , . . . , Πn ⊆ P
such that T is a csSLD-tree for Q and Π1 , . . . , Πn .
It may be needed that different atoms or programs are selected for repeated occurrences
of the same query. So, formally, a c-selection rule is a function of a sequence of queries (a
path in a tree) Q, . . . , Q0 . Also, strictly speaking, not an atom, but an occurrence of an
atom in Q0 is selected. In some cases it may be convenient to treat a c-selection rule as a
function of not only the sequence of queries, but also of (the not pruned part of) the subtree
rooted at Q0 (as often the decision about pruning is taken after a part of the subtree has
been computed).
12
24-04-2015
Example 6.2. Consider a Prolog program
c(L1, L2) :- m(X, L1), m(X, L2), !.
m(X, [X| ] ).
m(X, [ |L] ) :- m(X, L).
which is a logic program P with the cut added. Let Π1 = P and Π2 = {m(X, [X| ] ).}.
Consider a query Q = c([t1 , . . . , tk ], [u1 , . . . , un ]), the SLD-tree T for Q via the Prolog
selection rule, and the csSLD-tree T 0 which is T pruned due to the cut. If Q does not
succeed then T 0 = T . Otherwise T 0 is T without all the nodes to the right of the leftmost
successful branch. So in T 0 , for a node Q0 with an m-atom selected, Π2 is selected whenever
there is a success in the subtree rooted in the first child of Q0 ; otherwise Π1 is selected.
Note that the single cut may prune children of many nodes of T .
The c-selection rule of T 0 can be expressed as a function of a node: Π2 is selected for a
node Qi = m(X, [ti , . . . , tk ]), m(X, [u1 , . . . , un ]) in T 0 if ti is unifiable with some element of
[u1 , . . . , un ]. Otherwise Π1 is selected for Qi . For a node of the form Q0j = m(t, [uj , . . . , un ]),
program Π2 is selected if t is unifiable with uj , otherwise Π1 is selected.
In this example, determining the c-selection rule corresponding to the cut was not straightforward. Simpler cases are provided by Examples 6.6, 6.7, 6.8 below.
6.2. Completeness of pruned trees
We begin with an example illustrating some of the difficulties related to completeness of
pruned csSLD-trees. Then, after introducing some auxiliary notions, a sufficient condition
for completeness is presented. A few example proofs follow. We conclude with an example
of a completeness proof in a case for which the sufficient condition is not applicable directly.
Example 6.3. We show that completeness of each of Π1 , . . . , Πn is not sufficient for completeness of a csSLD-tree for Π1 , . . . , Πn . Consider a program P :
q(X) ← p(Y, X).
p(Y, 0).
(3)
(4)
p(a, s(X)) ← p(a, X).
p(b, s(X)) ← p(b, X).
(5)
(6)
and programs Π1 = {(3), (4), (6)}, Π2 = {(3), (4), (5)}, As a specification for completeness
consider S0 = { q(sj (0)) | j ≥ 0 }. Each of the programs Π1 , Π2 , P is complete w.r.t. S0 .
Assume a c-selection rule R choosing alternatively Π1 , Π2 along each branch of a tree. Then
the csSLD-tree for q(sj (0)) ∈ S0 via R (where j > 2) has no answers, thus the tree is not
complete w.r.t. S0 . Similarly the csSLD-tree for q(X) is not complete.
Sn
Consider
Snprograms P, Π1 , . . . , Πn and specifications S, S1 , . . . , Sn , such that P ⊇ i=1 Πi
and S = i=1 Si . The intention is that each Si describes which answers are to be produced
by using Πi in the first resolution step. We will call Π1 , . . . , Πn , S1 , . . . , Sn a split (of P
and S).
S
Definition 6.4. Let S = Π1 , . . . , Πn , S1 , . . . , Sn be a split, and S = Si . Specification
Si is suitable for an atom A w.r.t. S when no instance of A is in S \ Si . (In other words,
when ground(A) ∩ S ⊆ Si .) We also say that a program Πi is suitable for A w.r.t. S when
Si is.
A c-selection rule is compatible with S if for each non-empty query Q0 it selects an atom
A and a program Π, such that
– Π ∈ {Π1 , . . . , Πn } is suitable for A w.r.t. S, or
– none of Π1 , . . . , Πn is suitable for A w.r.t. S and Π = ∅ (so Q0 is a leaf).
A csSLD-tree for Π1 , . . . , Πn via a c-selection rule compatible with S is said to be weakly
compatible with S. The tree is compatible with S when for each its nonempty node some
Πi is selected.
The intuition is that when Πi is suitable for A then Si is a fragment of S sufficient to
deal with A. It describes all the answers for query A required by S.
24-04-2015
13
The reason of incompleteness of the trees in Ex. 6.3 may be understood as selecting a Πi not suitable for the selected atom. Take S = Π1 , Π2 , S0 ∪ S10 , S0 ∪ S20 , where
S10 = { p(b, si (0)) | i ≥ 0 } and S20 = { p(a, si (0)) | i ≥ 0 }. In the incomplete trees, Π1
is selected for an atom A = p(a, u), or Π2 is selected for an atom B = p(b, u) (where
u ∈ T U). However Π1 is not suitable for A whenever A has an instance in S (as then
ground(A) ∩ S ⊆ S20 , thus ground(A) ∩ S 6⊆ S0 ∪ S10 ); similarly for Π2 and B.
When Πi is suitable for A then if each atom of Si is covered by Πi (w.r.t. S) then using
for A only the clauses of Πi does not impair completeness w.r.t. S:
Sn
Sn
Theorem 6.5. Let P ⊇ i=1 Πi (where n > 0) be a program, S = i=1 Si a specification, and T a csSLD-tree for Π1 , . . . , Πn . If
(1 ) for each i = 1, . . . , n, all the atoms from Si are covered by Πi w.r.t. S, and
(2 ) T is compatible with Π1 , . . . , Πn , S1 , . . . , Sn ,
(3 ) (a) T is finite, or
(b) P is recurrent, or
(c) P is acceptable and T is built under the Prolog selection rule
then T is complete w.r.t. S.
See Appendix B for a proof. Note that in 3c the program may be acceptable w.r.t. a
specification distinct from S.
Example 6.6. A Prolog program [Clocksin and Mellish 1981]
nop(adam, 0) :- !.
nop(eve, 0) :- !.
nop(X, 2).
is an example of difficulties and dangers of using the cut in Prolog. (For its (in)correctness,
see Ex. 6.14.) Due to the two occurrences of the cut, for an atomic query A only the first
clause with the head unifiable with A will be used. The program can be seen as logic program
P = Π1 ∪ Π2 ∪ Π3 executed with pruning, where (for i = 1, 2, 3) Πi is the i-th clause of
the program with the cut removed. The intended meaning
is S = S1 ∪ S2 ∪ S3 , where
S1 = {nop(adam, 0)}, S2 = {nop(eve, 0)}, and S3 = nop(t, 2) ∈ HB | t 6∈ {adam, eve} .
Note that all the atoms from Si are covered by Πi (for i = 1, 2, 3).
Let S be Π1 , Π2 , Π3 , S1 , S2 , S3 . Consider a query A = nop(t, Y ) with a ground t. If t =
adam then only Π1 is suitable for A w.r.t. S, as ground(A) ∩ S = S1 . If t = eve then
only Π2 is suitable. For t 6∈ {adam, eve} the suitable program is Π3 . So for a query A the
pruning due to the cuts in the program results in selecting a suitable Πi , and the obtained
csSLD-tree is compatible with S. By Th. 6.5 the tree is complete w.r.t. S.
For a query nop(X, Y ) or nop(X, 0) only the first clause, i.e. Π1 , is used. However Π1
is not suitable for the query (w.r.t. S), and the csSLD-tree is not compatible with S. The
premises of Th. 6.5 do not hold, and the tree is not complete (w.r.t. S).
Example 6.7. The following program SAT0 is a simplification of a fragment of the SAT
solver of [Howe and King 2012] discussed in [Drabent 2012a; 2012b]. Pruning is crucial for
the efficiency and usability of the original program.
p(P -P, [ ]).
(7)
p(V -P, [B|T ]) ← q(V -P, [B|T ]). (8)
p(V -P, [B|T ]) ← q(B, [V -P |T ]). (9)
q(V -P, ) ← V = P. (10)
q( , [A|T ]) ← p(A, T ). (11)
P = P.
(12)
The program is complete w.r.t. a specification
p(t0 -u0 , [t1 -u1 , . . . , tn -un ]), n ≥ 0, t0 , . . . , tn , u0 , . . . , un ∈ T,
S=
∪ S=
q(t0 -u0 , [t1 -u1 , . . . , tn -un ])
ti = ui for some i ∈ {0, . . . , n}
14
24-04-2015
where T = {false, true} ⊆ HU, and S= = { t=t | t ∈ HU }. We omit a completeness
proof, mentioning only that SAT0 is recurrent w.r.t. a level mapping |p(t, u)| = 2|u| + 2,
|q(t, u)| = 2|u| + 1, |=(t, u)| = 0, where |u| for u ∈ HU is defined in Ex. 5.7.
The first case of pruning is related to redundancy within (8), (9); both Π1 = SAT0 \ {(9)}
and Π2 = SAT0 \ {(8)} are complete w.r.t. S. For any selected atom at most one of (8),
(9) is used, and the choice is dynamic. As the following reasoning is independent from this
choice, we omit further details.
So in such pruned SLD-trees the children of each node are constructed using one of
programs Π1 , Π2 . Thus they are csSLD-trees for Π1 , Π2 . They are compatible with S =
Π1 , Π2 , S, S (as Π1 , Π2 are trivially suitable for any A, due to S \ Si = ∅ in Df. 6.4). Each
atom of S is covered w.r.t. S both by Π1 and Π2 . As SAT0 is recurrent, by Th. 6.5 each
such tree is complete w.r.t. S.
Example 6.8. We continue with program SAT0 and specification S from the previous
example, and add a second case of pruning. When the selected atom is of the form A =
q(s1 , s2 ) with a ground s1 then only one of clauses (10), (11) is needed – (10) when s1 is
of the form t-t, and (11) otherwise. The other clause can be abandoned without losing the
completeness w.r.t. S.4
Actually, SAT0 is included in a bigger program, say P = SAT0 ∪ Π0 . We skip the details
of Π0 , let us only state that P is recurrent, Π0 does not contain any clause for p or for q,
and that P is complete w.r.t. a specification S 0 = S ∪ S0 where S0 does not contain any por q-atom. (Hence each atom of S0 is covered by Π0 w.r.t. S 0 .)
To formally describe the trees for P resulting from both cases of pruning, consider the
following programs and specifications:
Π1
Π3
Π4
Π5
= {(7), (8)}, Π2 = {(7), (9)},
= {(10)},
= {(11)},
= {(12)},
S1
S3
S4
S5
= S2 = S ∩ { p(s, u) | s, u ∈ HU },
= S ∩ { q(t-t, s) | t, s ∈ HU },
= S ∩ { q(t-u, s) | t, u, s ∈ HU, t 6= u },
= S= .
Let S be Π0 , . . . , Π5 , S0 , . . . , S5 . Each atom from Si is covered by Πi w.r.t. S 0 (for i =
0, . . . , 5). For each q-atom A with its first argument ground, Π3 or Π4 is suitable. (Both
Π3 , Π4 are suitable when ground(A) ∩ S = ∅.) For each remaining atom from T B (at least)
one of programs Π0 , Π1 , Π2 , Π5 is suitable.
Consider a pruned SLD-tree T for P (employing the two cases of pruning described
above). Assume that each q-atom selected in T has its first argument ground. Then T is a
csSLD-tree compatible with S. From Th. 6.5 it follows that T is complete w.r.t. S 0 .
The restriction on the selected q-atoms is implemented by means of Prolog delays. This
is done in such a way that, for the intended initial queries, floundering is avoided [Howe
and King 2012] (i.e. an atom is selected in each query). So the obtained pruned trees are as
T above, and the pruning preserves completeness of the program.
6.2.1. Sufficient condition not applicable directly. Th. 6.5 has an additional consequence. Not
only T but also any its subtree is complete w.r.t. S (as the subtree satisfies the premises
of Th. 6.5). So the theorem is not (directly) applicable when a csSLD-tree is complete, but
pruning removes some answers for a query within the tree (cf. Ex. 6.9 below). In Ex. 6.10 we
show how, at least in some such cases, completeness can be proved by employing a family
of specifications.
4 The same holds for A of the form q(s -s , s ), or q(s -s , s ) with non-unifiable s , s . The pruning
11 11 2
11 12 2
11
12
is implemented using the if-then-else construct in Prolog: q(V-P,[A|T]) :- V=P -> true ; p(A,T). (And
p(V-P,[B|T]) :- nonvar(V) -> q(V-P,[B|T]) ; q(B,[V-P|T]) implements the first case of pruning.)
24-04-2015
15
Example 6.9. Consider the logic program P from Ex. 6.2. Predicate c of P is complete
w.r.t. a specification
l1 , l2 are lists with
.
Sc = c(l1 , l2 ) ∈ HB
a common element
The pruning does not impair completeness for ground queries of the form Q = c(l1 , l2 ).
To apply Th. 6.5 to prove the completeness, each atom of Sc has to be covered by
the first clause of P w.r.t. a specification S. Hence S ⊇ Sc ∪ Sm is needed where
Sm = { m(ti , [t1 , . . . , tn ]) ∈ HB | n > 0, 1 ≤ i ≤ n }.
Consider now a pruned tree for such Q, and its subtree T 00 rooted at Q0 =
m(X, l1 ), m(X, l2 ). The subtree is not complete whenever l1 , l2 have more than one common
element. Thus, for each such case, Th. 6.5 is inapplicable for any superset of Sc .5
Example 6.10. The previous example shows a case where completeness of pruned trees
w.r.t. a specification cannot be shown by Th. 6.5. However the completeness can be proved
by employing a family of specifications. Consider the same program P and specification
Sc . For each c(l1 , l2 ) ∈ Sc we construct a specification S 0 for which P is semi-complete,
c(l1 , l2 ) ∈ S 0 , and Th. 6.5 is applicable. Informally, S 0 may be understood as the subset of
MP needed to obtain c(l1 , l2 ), by means of the clauses involved in a pruned tree for c(l1 , l2 ).
Let l1 = [t1 , . . . , tk ], l2 = [u1 , . . . , un ]. So there exist a term t and indices g, h such that
t = tg = uh , for i = 1, . . . , g − 1 no ti is a member of l2 , and uj 6= t for j = 1, . . . , h − 1. Let
0
S 0 = { c(l1 , l2 ) } ∪ Sm
where
0
Sm
= { m(t, [ti , . . . , tk ]) | 1 ≤ i ≤ g } ∪ { m(t, [uj , . . . , un ]) | 1 ≤ j ≤ h }.
Let S = Π1 , Π2 , S1 , S2 , where Πi are as in Ex. 6.2, S1 = S 0 , S2 = { m(t, [tg , . . . , tk ]),
m(t, [uh , . . . , un ]) }. All the atoms from Si are covered by Πi w.r.t. S 0 (for i = 1, 2). Consider
the csSLD-tree T 0 for the (ground) query Q = c(l1 , l2 ). From the description in Ex. 6.2 of
the c-selection rule, it follows that whenever Π2 is selected together with an atom A in T 0
then A = m(X, [tg , . . . , tk ]) (for some variable X) or A = m(t, [uh , . . . , un ]). In both cases
ground(A) ∩ S 0 ⊆ S2 . Thus T 0 is compatible with S and, as it is finite, by Th. 6.5 T 0 is
complete w.r.t. S 0 .
We showed that, for each Q ∈ Sc , the pruned tree is complete w.r.t. {Q}. Hence for each
ground query of the form Q = m(s1 , s2 ) the csSLD-tree for Q is complete w.r.t. Sc .
6.2.2. Another sufficient condition for completeness. By generalizing Th. 5.10 to pruned SLDtrees, we obtain a sufficient condition for completeness, which does not require that the
trees are finite. See Appendix B for a proof.
Sn
Sn
Proposition 6.11. Let P ⊇ i=1 Πi ( n > 0) be a program, S = i=1 Si a specification,
| |: HB ,→ N a level mapping, and T a csSLD-tree for Π1 , . . . , Πn . If
(1 ) T is compatible with Π1 , . . . , Πn , S1 , . . . , Sn ,
(2 ) each atom from Si is recurrently covered by Πi w.r.t. S and | |, for each i = 1, . . . , n,
then T is complete w.r.t. S.
Example 6.12. Consider the program P , specification S and the level mapping | | from
Ex. 5.11. Consider a c-selection rule R which selects the first atom of the query, and if the
5
It is not applicable in a more general case of l1 , l2 having a common element, which is not the last element
of l1 . The pruned tree for such Q contains a node with A = m(Y, [ti , . . . , tk ]) and Π2 = { m(X, [X| ] ).}
selected (where ti , . . . , tk are ground, and ti is a member of l2 ). For such atom A, a suitable specification
S2 is a superset of ground(A) ∩ S ⊇ { m(tj , [ti , . . . , tk ]) | i ≤ j ≤ k }. But such S2 cannot be covered by Π2 ,
unless all ti , . . . , tk are equal. So the premises of Th. 6.5 are not satisfied.
Note that this reasoning holds for any representation of the considered pruned trees as csSLD-trees,
possibly employing some other n, Π1 , . . . , Πn .
16
24-04-2015
atom is of the form p(t, t) it selects Π1 = { p(X, X) }, if it is of the form p(t, u) with distinct
ground t, u then it selects Π2 = { p(X, Z) ← e(X, Y ), p(Y, Z) }, otherwise the whole program is selected, Π3 = P . Let S1 = { p(t, t) | t is a node of E }, S2 = { p(t, u) ∈ S | t 6= u },
S3 = S. Note that R is compatible with S = Π1 , Π2 , Π3 , S1 , S2 , S3 (as in the first case
ground(p(t, t)) ∩ S ⊆ S1 , and in the second case {p(t, u)} ∩ S ⊆ S2 ).
From Ex. 5.11 it follows that each atom from Si is recurrently covered by Πi w.r.t. S, for
i = 1, 2, 3. Thus by Prop. 6.11, any csSLD-tree for P via R is complete w.r.t. S. Note that
the tree may have infinite branches.
Assume that the initial queries are ground. Then, under the Prolog selection rule, all
the selected p-atoms are ground. So the c-selection rule R can be implemented in Prolog
by adding the cut to the first clause of P : p(X,X):-!. However when the csSLD-tree is
infinite, not the whole tree may searched by Prolog due to the depth-first search strategy.
6.3. Correctness of answers of pruned trees
The answers of pruned trees are, obviously, logical consequences of the program. However,
it may happen that, due to pruning, not all such consequences are obtained. This feature is
sometimes used in Prolog programming; it is called “red cuts” [Sterling and Shapiro 1994].
In this section we show how to characterize more precisely the answers in such cases. We
present a sufficient condition for the answers of csSLD-trees to be correct (cf. Df. 2.4), w.r.t.
a specification for which the program may be not correct. Informally, the idea is to apply
the condition of Th. 4.1 only to those instances of program clauses which were employed in
the pruned trees.
Theorem 6.13. Let T be a csSLD-tree for a query Q and programs Π1 , . . . , Πn , and S
be a specification. If for each i = 1, . . . n
B1 , . . . , Bm ∈ S implies H ∈ S, for each (H ← B1 , . . . , Bm ) ∈ ground(Πi ) (m ≥ 0)
where H is an instance of some atom A selected together with Πi in T ,
then each answer Qθ of T is correct w.r.t. S (i.e. S |= Qθ).
Proof. Assume the premises of the theorem. We have to show that S |= Qθ. Answer
Qθ is obtained out of a successful branch of the tree. The branch represents a derivation
with queries Q0 , . . . , Qk , mgu’s θ1 , . . . , θk , and renamed clauses C1 , . . . , Ck , where Qk is
empty, Q0 = Q, and Qθ = Qθ1 · · · θk . By induction on k − i we obtain S |= Qi θi+1 · · · θk for
i = k, k−1, . . . , 0. The base case is obvious. For the inductive step, assume S |= Qi θi+1 · · · θk .
Let Q0 ∈ ground(Qi−1 θi · · · θk ). Let an atom A and a program Πj be selected for Qi−1 . So
Ci ∈ Πj . A ground instance Q00 of Qi θi+1 · · · θk can be obtained out of Q0 by replacing an
instance H of A by a sequence of atoms B1 , . . . , Bm (m ≥ 0), where (H ← B1 , . . . , Bm ) ∈
ground(Ci ). As each atom of Q00 is in S, so is each atom of Q0 .
Example 6.14. Let us continue Ex. 6.6. The program is obviously not correct w.r.t. S.
However, we show that for queries of the form A = nop(t, Y ), where t is ground, the answers
of pruned trees are correct w.r.t. S.
The condition of Th. 6.13 obviously holds for Π1 , Π2 . Consider Π3 , it is selected together
with A = nop(t, Y ) only if t 6∈ {adam, eve}. Then any ground instance H of both A and
the head nop(X, 2) of the clause from Π3 is in S. Thus the premise of Th. 6.13 is satisfied.
On the other hand, it is not satisfied for e.g. a query A0 = nop(Y, 2) (due to its instance
H = nop(adam, 2) 6∈ S, H ∈ ground(Π3 ) and Π3 is selected together with A0 ). Note that
the pruned tree for such A0 is not correct w.r.t. S.
7. RELATIONS WITH DECLARATIVE DIAGNOSIS
We now discuss the relationship between program diagnosis, and proving correctness and
completeness of programs. We first introduce declarative diagnosis of logic programs, and
24-04-2015
17
compare it with the proof methods of Sections 4, 5; details irrelevant for this paper are
omitted. Then we show how diagnosis can be performed by attempting a correctness or
completeness proof of the program concerned. Finally we discuss a major drawback of
declarative diagnosis, and show how it can be avoided by employing approximate specifications.
7.1. Declarative diagnosis
Declarative diagnosis, called sometimes algorithmic, or declarative, debugging, was introduced by Shapiro [1983] (see also [Drabent et al. 1989; Naish 2000] and references therein).
It provides methods to locate the reasons of incorrectness or incompleteness of programs. A
diagnosis algorithm starts with a symptom, obtained from testing the program; the symptom
shows incorrectness or incompleteness of the program w.r.t. its specification S. The symptom is an answer Q such that S 6|= Q, or a query Q for which the computation terminates
but some answers required by S are not produced. (A less general notion of incompleteness
symptom is an atom A ∈ S for which the program finitely fails.) The located error turns
out to be a fragment of the program which violates our sufficient condition for correctness
or, respectively, semi-completeness.
Incorrectness diagnosis. In declarative diagnosis, the reason for incorrectness is an incorrect instance of a program clause. An incorrect clause instance is one which violates the
sufficient condition of Th. 4.1. Obviously, by Th. 4.1 if the program is incorrect then such
clause must exist. On the other hand, a violation of the sufficient condition does not imply
that the program is incorrect (cf. Ex. A.1 in the Appendix). Incorrectness diagnosis cannot
be made more precise; we cannot say which part of the incorrect clause is wrong, as different
modifications of various parts of the incorrect clause may lead to a correct one.
Technically, an incorrectness diagnosing algorithm works by constructing a proof tree for
an atomic symptom Q, and then searching the tree for an incorrect clause instance. Notice
that the actions performed by the algorithm boil down to checking the sufficient condition
for correctness of Th. 4.1, but only for some clause instances – those involved in producing
the symptom.
Example 7.1 (Incorrectness diagnosis). Consider a buggy insertion sort program
[Shapiro 1983, Sect. 3.2.3]:
isort([],[]).
isort([X|Xs],Ys) :- isort(Xs,Zs), insert(X,Zs,Ys).
insert(X,[],[X]).
insert(X,[Y|Ys],[Y|Zs]) :- Y > X, insert(X,Ys,Zs).
insert(X,[Y|Ys],[X,Y|Ys]) :- X =< Y.
(where >, =< are arithmetic built-ins of Prolog [Apt 1997]). Procedure isort should define the
sorting relation for lists of integers. Procedure insert should describe inserting an element
into a sorted list, producing a sorted list as a result. The program computes an incorrect
answer A = isort([2, 1, 3], [2, 3, 1]). (Formally, S 6|= A, where S is a specification known by
the user; we do not formalize S here.) Thus A is a symptom of incorrectness. The diagnoser
constructs a proof tree
isort([2, 1, 3], [2, 3, 1])
isort([1, 3], [3, 1])
insert(2, [3, 1], [2, 3, 1])
...
isort([3], [3])
insert(1, [3], [3, 1])
...
3>1
insert(1, [ ], [1])
18
24-04-2015
Starting from the root of the proof tree, the diagnosis algorithm asks the user whether the
children of the current node are correct w.r.t. the user’s specification. When all the answers
are YES, the current node together with its children are an incorrect clause instance – the
error is found. An answer NO means that a new incorrect node is found, and it becomes
the current node.
For the tree above, the questions are first asked about the two children of the root.
The child A0 = isort([1, 3], [3, 1]) is found incorrect, formally S 6|= A0 . (The question about
the second child is discussed in the next example.) Queries about the children of A0 lead
to identifying insert(1, [3], [3, 1]) as incorrect. As both its children are found correct, the
algorithm returns insert(1, [3], [3, 1]) ← 3 > 1, insert(1, [ ], [1]) as an incorrect clause instance.
(The nodes found incorrect are marked in the diagram.)
Note that more precise locating of the error is impossible, The wrong clause may be
corrected for instance by swapping X and Y in Y > X, or in the first and last atom of the
clause. Thus we cannot determine which fragment of the clause is erroneous.
Incompleteness diagnosis. There exist a few versions of incompleteness diagnosis algorithms ([Naish 1992] provides a comparison). Some of algorithms start with a symptom
which is a finitely failed specified atom A ∈ S. Some use, instead, a query Q for which the
SLD-tree is finite, and some answers required by the specification S are missing. As the reason for incompleteness, a not covered specified atom is found, say p(. . .) ∈ S. Alternatively,
the result of the diagnosis is an atom B = p(. . .) which has a not covered specified instance
Bθ ∈ S. In both cases procedure p is identified as erroneous.
The existence of a not covered specified atom violates the sufficient condition for semicompleteness of Th. 5.6. Conversely, if the program is not complete for a query Q with a
finite SLD-tree then it is not semi-complete and, by Th. 5.6, there must exist a not covered
specified atom. Again, violating the sufficient condition for completeness does not imply that
the program is not complete. Roughly speaking, the specification may be not sufficiently
general, cf. Ex. A.1. Similarly to incorrectness, incompleteness diagnosis cannot be made
more precise: A whole procedure is located, and it cannot be determined which fragment of
the procedure is wrong, or whether a clause is missing.
We skip an example, and details of incompleteness diagnosis algorithms, see e.g. [Shapiro
1983; Drabent et al. 1989; Naish 1992]. We only point out a relation between completeness
proving and some of such algorithms (like that of [Shapiro 1983]). The actions performed
by such algorithm boil down to checking the sufficient condition for semi-completeness of
Th. 5.6. In contrast to completeness proofs, the condition is checked only for some specified
atoms, they are instances of atoms selected in the SLD-tree producing the symptom.
Diagnosis by proof attempts. In a sense, the proof methods presented in this paper supersede the declarative diagnosis methods. An attempt to prove a buggy program to be correct
(complete) results in violating the corresponding sufficient condition for some clause (specified atom). Any error located by diagnosis will also be found by a proof attempt; moreover
no symptom is needed, and all the errors are found. For instance, in this way the author
found an error in an early version of one of the examples from [Drabent 2012a]. However
the sufficient condition has to be checked for all the (instances of the) clauses of the program (when proving correctness), or for all specified atoms (in the case of completeness). In
contrast, diagnosis takes into account only the clauses (respectively, atoms) related to the
computation that produced the symptom.
7.2. A main drawback of declarative diagnosis
Declarative diagnosis requires the user to know an exact specification (a single intended
model) of the program. The program has to be (made) correct and complete w.r.t. this
specification. This is a serious difficulty in diagnosing actual programs [Drabent et al. 1989,
Sect. 26.8], [Naish 2000]. Some diagnoser queries, like “is append([a], b, [a|b]) correct”, may
24-04-2015
19
be confusing and hardly possible to answer, as the programmer often does not know some
details of the intended model, like those related to applying append to non lists. In the
author’s opinion this was a main reason of the lack of acceptance of declarative diagnosis
in practice.
Example 7.2 (Exact specification unfeasible). The previous example provides a diagnoser query which is hardly possible to answer. Determining correctness of B =
insert(2, [3, 1], [2, 3, 1]) is confusing, as the user does not know how insertion into unsorted
lists should behave. From a formal point of view, the user does not know an exact specification of insert.
Actually, both cases of atom B being correct or incorrect are reasonable. They result in
locating various errors, and lead to various versions of the final, corrected program. If the
specification for insert is 6
l , l are sorted lists of integers,
0
,
Sinsert
= insert(n, l1 , l2 ) ∈ HB 1 2
elms(l2 ) = {n} ∪ elms(l1 )
0
where elms(l) is the multiset of the elements of list l, then Sinsert
6|= B and an erroneous
clause instance is eventually found within the subtree rooted in B. Correcting the program
0
but inefficient, as it assures
would lead to a procedure insert which is correct w.r.t. Sinsert
that two of its arguments are sorted lists of integers.
If the specification is
if n ∈ Z and l1 is a sorted lists of integers,
Sinsert = insert(n, l1 , l2 ) ∈ HB
0
then insert(n, l1 , l2 ) ∈ Sinsert
then Sinsert |= B, and the diagnosis continues as in Ex. 7.1. Each of the alternative corrections described there leads to a program with insert correct w.r.t. Sinsert . One of
the programs is the “right” one – Program 3.21 of [Sterling and Shapiro 1994] (with
insert(X, [Y |Y s], [Y |Zs]) ← X > Y , insert(X, Y s, Zs) as the corrected clause).
0
is an exact specification of procedure isort in the
Notice that neither Sinsert nor Sinsert
“right” program. Providing its exact specification is not so obvious. The “right” procedure
0
is correct w.r.t. Sinsert , and complete w.r.t. Sinsert
. Any such procedure, combined with
procedure isort from Ex. 7.1, would implement a correct and complete predicate isort.
The programmer should have the freedom of choosing (and changing) the actual semantics
of insert, guided by e.g. the efficiency of the program. An exact specification would be
counterproductive, at least at earlier stages of program development.
So for this procedure one should take Sinsert as the specification for correctness and
0
as the specification for completeness, and apply them respectively in incorrectness
Sinsert
and incompleteness diagnosis.
The discussion in the example suggests a natural way to overcome the difficulty due to
declarative diagnosis being based on a single intended model of a program:
Remark 7.3. Declarative diagnosis should employ approximate specifications. In the incorrectness diagnosis the specification for correctness should be used, and in the incompleteness diagnosis – the specification for completeness.
The problem with exact specifications was first dealt with in [Pereira 1986]. In that approach a specification also describes which atoms are admissible goals, i.e. are allowed to be
selected in SLD-trees. Thus specifications and diagnosis are not declarative, as they refer
6
An (exact) specification for the remaining predicates is obvious:
l1 is a list of integers,
>(i1 , i2 ) ∈ HB,
sort(l1 , l2 ) ∈ HB
∪
l2 is a sorted permutation of l1
=<(j1 , j2 ) ∈ HB
i1 , i2 , j1 , j2 are integers,
i1 > i 2 , j1 ≤ j2
.
20
24-04-2015
to the operational semantics. Naish [2000] introduces a three valued approach, in which
a specification classifies each ground atom as correct, erroneous or inadmissible. For such
specifications, three-valued declarative debugging algorithms are presented. From our point
of view, the set of non-erroneous atoms can be understood as a specification for correctness, and the set of correct atoms as a specification for completeness. However introducing
debugging algorithms based on a three-valued logic seems to be an unnecessary complication. In the approach proposed here the standard two-valued diagnosis algorithms remain
unchanged.
8. RELATED WORK
Approximate specifications. The difficulties caused by a restriction to exact specifications
have been recognized by a few authors. Naish and Søndergaard [2014, and the references
therein] propose employing a 3-valued or 4-valued logic. The logical value i (inadmissible)
is given to the atoms that, in our approach, are in the difference Scorr \ Scompl between the
specification for correctness and that for completeness.
Apt [1997] and Deville [1990] present approaches in which it is possible to deal with a
suitable subset of MP instead of MP itself. Apt [1997] employs preconditions and postconditions,7 and a sufficient condition that, for any atomic query A, if A ∈ pre then Aθ ∈ post
for any computed answer Aθ (where pre and post is, respectively, a pre- and postcondition).
In this way MP is in a sense replaced by post, more precisely by the set MP ∩ pre, called
M(pre,post) . (See also Proof methods below.) Eg. for program APPEND the precondition
may be pre = { app(k, l, m) ∈ T B | l or m is a list }, and the postcondition may describe
0
exactly the list concatenation relation: post = SAPPEND
, in the notation of Ex. 3.1. This
0
gives M(pre,post) = SAPPEND . See [Drabent and Milkowska 2005] for further comparison.
In [Deville 1990] the notions of correctness and completeness of programs involve, for each
predicate of the program, the intended relation (on ground terms) and the so-called domain.
Let us represent them as Herbrand interpretations, the intended relations as rel ⊆ HB
and the domains as dom ⊆ HB. Then correctness (called there partial correctness) of a
definite clause program P is equivalent to MP ∩ dom ⊆ rel ∩ dom, and completeness
0
means inclusion in the other direction. For the APPEND example, rel = SAPPEND
and
dom = { app(k, l, m) ∈ HB | k, l, m are lists }.
Sometimes it is claimed that programs like APPEND should be modified, so that the
resulting program defines “the right” relation [Naish 1996; Apt 1993; Apt 1995]. For the
case of APPEND, the new program should define exactly the list concatenation. This can
be done by modifying its non-recursive clause so that its variable is restricted to be bound
to a list. The clause becomes app([ ], L, L) ← list(L), and an appropriate definition of list
is added. The program is believed to be simpler to reason about, but it is impractical due
to its inefficiency. Naish [1996, p. 187] proposes that the meaning of the original program
should be defined as the standard meaning of the modified program.
The reader is encouraged to compare the treatment of APPEND in Examples 3.1, 4.3,
5.8 to that in the approaches outlined here. To sum up, for reasoning about correctness and
completeness of definite programs, the approach based on approximate specifications seems
preferable, as it is simpler than the approaches outlined above, refers only to the the basic
notions of logic programming, and seems at least as powerful.
Proof methods. Below we review previous work on proving correctness and completeness
of logic programs. We omit most of approaches based on operational semantics.
Our definitions of correctness and completeness (Df. 2.1) are basically the same as those
in [Sterling and Shapiro 1994]. However no particular method of reasoning on correctness
and completeness is given there, and the example informal proofs are made by analyzing
possible proof trees. The correctness proving method of [Clark 1979] used here (Th. 4.1)
7
Technically, pre- and postconditions are sets of atoms closed under substitution.
24-04-2015
21
should be well-known, but is often neglected. For instance, an important monograph [Apt
1997] uses a method,8 which is more complicated and not declarative. It proves a certain
property of LD-derivations, from which the declarative property of program correctness
follows. See [Drabent and Milkowska 2005] for comparison and argumentation that Clark’s
method is sufficient, and for its generalization to programs with negation.
Surprisingly little has been done on reasoning about completeness. For instance, it is
absent from [Apt 1997]9 and [Hogger 1984].10 Kowalski [1985] discusses completeness, but
the example proofs concern only correctness. As a sufficient condition for completeness of
a program P he suggests P ` TS , where TS is a specification in a form of a logical theory.
(A symmetrical criterion TS ` P is used for correctness, notice its similarity to that of
Th. 4.1.) The condition seems impractical as it fails when TS contains auxiliary predicates,
not occurring in P . It also requires that all the models of P (including HB) are models of
the specification. But it seems that such specifications often have a substantially restricted
class of models, maybe a single Herbrand model, cf. [Deville 1990].
Stärk [1998] presented an elegant method of reasoning about a broad class of properties
of programs with negation, executed under LDNF-resolutions. A tool to verify proofs mechanically was provided. The approach involves a rather complicated induction scheme, so
it seems impossible to apply the method informally by programmers.
Deville [1990] introduces a systematic approach to constructing programs. Correctness
and completeness of a program follows from construction. No direct sufficient criteria, applicable to arbitrary programs, are given. The underlying semantics of Prolog programs is that
of LDNF-resolution. Note that the last two approaches are based on operational semantics,
and depend on the order of literals in clauses. So they are not declarative.
Deransart and Maluszyński [1993] present criteria for definite program completeness, in a
sophisticated framework of relating logic programming and attribute grammars. We present,
as Th. 5.10, their sufficient (and complete) criterion for completeness in a simplified setting
and with a short direct proof.
The method introduced here (Th. 5.6 with Prop. 5.4) is a simplification of that from
[Drabent and Milkowska 2005]. Due to a restriction to specifications over Herbrand domains, we could deal here with ground(P ) and covered ground atoms, instead of a theory
ONLY-IF(P ) and its truth in an interpretation, as in the former work. Another difference
is employing here a new notion of semi-completeness. Also, the former work uses a slightly
different notion of completeness (cf. a remark in Sect. 2), and provides a generalization for
programs with negation, with a declarative semantics of [Kunen 1987].
Some ideas of the current work appeared in [Drabent 2012a; 2012b] (see also Applications
in Sect. 9). This includes a weaker version of Th. 6.5 on completeness of pruned SLD-trees.
The author is not aware of any other former work on proving completeness in presence of
pruning, or on proving correctness of the answers of pruned SLD-trees (w.r.t. specifications
for which the program is not correct). Part of this work was presented in a conference paper
[Drabent 2015a].
For the related work on declarative diagnosis see Sect. 7. Dealing with the occur-check
problem (see [Apt 1997; Deransart and Maluszyński 1993] and the references therein) is
outside of the scope of this paper.
8
It is basically the method of [Bossi and Cocco 1989], (stemming from [Drabent and Maluszyński 1988],
a paper focused on non-declarative properties of programs). The employed notion of correctness says that,
given a precondition pre and a postcondition post, if the initial atomic query A is in pre then in each
LD-derivation each selected atom is in pre and each corresponding computed answer is in post. Note that
it depends on the order of atoms in a clause. It implies a declarative property that, in our terminology, the
program is correct w.r.t. (HB \ pre) ∪ (post ∩ HB).
9 Instead, for a program P and an atomic query A, a characterization of the set of computed instances of
A is studied, in a special case of the set being finite and the answers ground [Apt 1997, Sect. 8.4]. This is
based on finding the least Herbrand model (of P or of a certain subset of ground(P )).
10 The notion of completeness is defined in [Hogger 1984], but no sufficient condition is discussed.
22
24-04-2015
9. REMARKS
Declarativeness. In this paper we are mainly interested in declarative reasoning about
logic programs. Correctness and completeness of logic programs are declarative properties.
However non declarative methods for dealing with correctness are often suggested, cf. the
previous section. If it were necessary to reason non declaratively about correctness or completeness of programs then logic programming would not deserve to be called a declarative
programming paradigm. This paper presents declarative sufficient conditions for correctness and semi-completeness (Th. 4.1, 5.6). Some of the conditions for completeness are also
declarative (Prop. 5.4(2), Th. 5.10).
The remaining sufficient conditions for completeness of Prop. 5.4 are not declarative.
Condition (1) involves termination, condition (3) depends on the order of atoms in clause
bodies (and implies termination). Notice that declarative completeness proofs either imply
termination (when the program is shown to be recurrent, Prop. 5.4(2)), or require reasoning
similar to that in termination proofs (when showing that covered atoms are recurrently
covered, Th. 5.10). Notice also that in most practical cases termination has to be established
anyway. So proving completeness in two steps – a declarative proof of semi-completeness
and a proof of termination – may be a reasonable compromise between declarative and
non-declarative reasoning. Moreover, in many cases termination proofs can be obtained
automatically (see [Codish and Taboch 1999; Mesnard and Bagnara 2005; Schneider-Kamp
et al. 2009; Nguyen et al. 2011] and references therein). Semi-completeness alone may be a
useful property, as it guarantees that whenever the computation terminates, all the answers
required by the specification have been computed.
Note that the sufficient conditions of Section 6 for completeness of pruned trees and
correctness of their answers also are declarative, except for conditions 3a, 3c of Th.6.5 (which
are similar to conditions (1) and (3) of Prop. 5.4, discussed above). Proving condition 3a,
i.e. termination, can be partly automated, as mentioned above.
Granularity of proofs. The sufficient conditions presented in this paper impose a certain
granularity of proofs. Correctness proofs deal with separate program clauses. Proofs of semicompleteness deal with whole program procedures (to check that an atom p(. . .) is covered
one has to consider all the clauses for p). For completeness, a certain level mapping has to
be found, or program termination is to be considered. In both cases the whole program has
to be taken into account.
Interpretations as specifications. Specifications which are interpretations are somewhat
limited [Drabent and Milkowska 2005]. Some program properties cannot be expressed. The
problem also concerns e.g. the approaches of [Apt 1997], [Hogger 1984], [Kowalski 1985],
[Deville 1990], and the declarative diagnosis. The limitation is due to considering completeness w.r.t. a specification which describes a single relation for a predicate symbol. Such
specification cannot express that e.g. for a given t there exists a u such that p(t, u) is the
program’s answer. It has to explicitly state some (one or more) particular u.
In our framework an approximate specification describes a set { I | Scompl ⊆ I ⊆ Scorr }
of specifications; its elements are the least Herbrand models of programs compatible with
the specification; the set has the least and the greatest element. We have just shown a case
where an approximate specification should describe a set with many minimal elements. (A
symmetric case – related to correctness – is e.g. describing that at most one of p(t, u), p(t, s)
is allowed to be an answer of the program.) A possible solution may be to introduce specifications in a form of logical theories. Such a theory may include axioms like ∀t∃u. p(t, u),
and describe a set inexpressible in our framework.
Applications. We agree with the opinion of [Apt 1993] that “unless the verification method
is easy and amenable to informal use, it will be ignored”. We want to stress the simplicity
and naturalness of the sufficient conditions for correctness (Th. 4.1) and semi-completeness
24-04-2015
23
(Th. 5.6, the condition is a part of each discussed sufficient condition for completeness).
Informally, the former states that the clauses of a program should produce only correct
conclusions, given correct premises. The latter states that each ground atom that should be
produced by P can be produced by a clause of P out of atoms which should be produced
by P . The author believes that this is a way a competent programmer reasons about (the
declarative semantics of) a logic program.
The practical applicability of the methods presented here is illustrated by a larger example
in [Drabent 2012a; 2012b]. The example demonstrates a systematic construction of a nontrivial Prolog program (the SAT solver of [Howe and King 2012]). Starting from a formal
specification, a definite clause logic program is constructed hand in hand with proofs of
its correctness, completeness, and termination under any selection rule. The final Prolog
program is obtained by adding control to the logic program (delays and pruning SLDtrees). Adding control preserves correctness and termination. However completeness may
be violated by pruning, and by floundering related to delays. By Th. 6.5, the program with
pruning remains complete.11 Reasoning about floundering is outside of the scope of the
current paper; non-floundering of this program can be confirmed by a program analysis
algorithm [Genaim and King 2008].
The example shows how much of the programming task can be done declaratively, without
considering the operational semantics; how “logic” could be separated from “control.” It
is important that all the considerations and decisions about the program execution and
efficiency are independent from those related to the declarative semantics: to the correctness
and completeness of the final program. For the role of approximate specifications in this
example, see Comments in Sect. 3.
Future work. A natural continuation of the work presented here is generalization of the
proof methods for programs with negation. There are three semantics to deal with: 3valued completion semantics of Kunen (Prolog with additional checks for sound negation
[Kunen 1987; Doets 1994]), the well-founded semantics (the same with tabulation, as in XSB
[Van Gelder et al. 1991; Swift and Warren 2012]), and the answer set semantics [Brewka
et al. 2011, and the references therein]. The approach for Kunen semantics could possibly
stem from that of [Drabent and Milkowska 2005]. Also generalization to constraint logic
programming and to CHR (constraint handling rules) could be of interest.
For reasoning about pruning, it should be useful to provide methods which directly refer
to the pruning constructs (like the cut) used in a program, instead of employing the cselection rule. An initial attempt was presented in [Drabent 2015a]. An interesting issue is
introducing another form of specifications, to overcome the limitation discussed earlier in
this section.
An important task is formalization and automatizing of correctness and completeness
proofs; a first step is formalization of specifications. On the other hand, further examples
of proofs for practical programs are due, as are experiments with teaching programmers to
informally use the proof methods in practice.
10. CONCLUSIONS
This paper presents proof methods for proving correctness and completeness of definite
clause programs. The method for correctness [Clark 1979] is simple and natural. It should
be well-known, but is often neglected. Little work has been done on proving completeness,
and this is the main subject of this paper (Sect. 5). A simplification of the approach of
[Drabent and Milkowska 2005] is presented. We introduce a notion of semi-completeness,
for which the corresponding sufficient condition deals with program procedures separately,
11
In the former paper a weaker version of Th. 6.5 has been used, and one case of pruning is discussed
informally. A proof covering both cases of pruning is illustrated here in Ex. 6.8.
24
24-04-2015
while for completeness the whole program has to be taken into account. Semi-completeness
and termination imply completeness. In practice this means that if a semi-complete program has terminated then it has produced all the answers required by the specification.
The presented sufficient condition for semi-completeness corresponds to a natural, intuitive
way of reasoning about programs. It is a necessary condition for program completeness (in
a sense made precise in Appendix A). We propose a few sufficient conditions for program
completeness; all of them involve the above mentioned condition for semi-completeness.
Also, sufficient conditions are given for completeness being preserved under SLD-tree pruning (Sect. 6). This is augmented by a sufficient condition for correctness of the answers of
pruned SLD-trees (applicable in describing the effects of so called “red cuts” [Sterling and
Shapiro 1994]).
Logic programming can’t be considered a declarative programming paradigm unless there
exist declarative ways of reasoning about program correctness and completeness (i.e. reasoning which abstracts from the operational semantics). Regrettably, non-declarative methods
are usually suggested (cf. Sect. 8). The presented methods for proving correctness and semicompleteness are purely declarative (Sect. 9), however some of the sufficient conditions for
completeness are not, as they refer, perhaps indirectly, to program termination. Such methods may nevertheless be useful in practical declarative reasoning about programs, as usually
termination has to be established anyway.
A larger example of applying the proof methods described here is provided in [Drabent
2012a; 2012b]. That paper presents a construction of a non-trivial program hand in hand
with its correctness and completeness proofs. The construction shows how “logic” can be
separated from “control”; how the reasoning about correctness and completeness can be
separated from that related to the operational semantics, efficiency, etc (cf. Applications in
Sect. 9).
We point out advantages of approximate specifications (Sect. 3). They are crucial for
avoiding unnecessary complications in constructing specifications and proofs. It is often
cumbersome and superfluous to exactly describe the relations computed by a program. Approximate specifications are natural in the process of program development: when starting
construction of a program, the relations it should compute are often known only approximately. This suggests an extension of the well-known paradigm of program development by
transformations that preserve program semantics (see [Pettorossi et al. 2010] for references);
one should also consider transformations which preserve correctness and completeness w.r.t.
an approximate specification (cf. Comments in Sect. 3).
In Sect. 7 we compared the proof methods with declarative diagnosis (algorithmic debugging). Similarity was demonstrated between two ways of locating errors in programs: by
means of declarative diagnosis, and by a failure of an attempt to construct a correctness or
completeness proof. We have also shown how approximate specifications help to remove a
serious drawback of declarative diagnosis.
We argue that the proof methods presented here are simple, and reflect a natural way of
declarative thinking about programs (Sect. 9). We believe that they can be actually used –
at least at an informal level – in practical programming; this is supported by examples.
APPENDIX
A. COMPLETENESS OF PROOF METHODS
An important feature of a proof method is its completeness. Here we show in which sense
the methods of proving correctness (Th. 4.1), semi-completeness (Th. 5.6), and completeness
(Th. 5.10) are complete.
Obviously, we cannot require that if a program P is correct (respectively, complete)
w.r.t. a specification S then a proof method applied to P and S shows the correctness
(completeness). The specification may provide insufficient information.
24-04-2015
25
Example A.1 (Specifications too weak (strong) for a correctness (completeness) proof ).
Program P = { p(X) ← q(X). q(a). } is correct w.r.t. S = { p(a), q(a), q(b) }. However the
sufficient condition from Th. 4.1 is violated: clause instance p(b) ← q(b) has its body atom
in S and the head not in S. We need a stronger specification S 0 = { p(a), q(a) } ⊆ S, for
which P is correct and the sufficient condition holds.
In Ex. 6.3, program Π1 is complete w.r.t. S0 = { q(sj (0)) | j ≥ 0 }, but the sufficient
condition for semi-completeness of Th. 5.6 does not hold: some atoms of S0 (actually all) are
not covered by P w.r.t. S0 . To prove the completeness, we need a more general specification
S = S0 ∪{ p(b, sj (0)) | j ≥ 0 }. Now each atom S is covered by Π1 w.r.t. S, and the sufficient
condition holds. (Similar reasoning applies to the completeness of P and Π2 w.r.t. S0 .)
Theorems A.2, A.4 below are from [Deransart and Maluszyński 1993].
Theorem A.2 (Completeness of the method for correctness). If a program
P is correct w.r.t. a specification S then there exists a stronger specification S 0 ⊆ S such
that the sufficient condition from Th. 4.1 holds for P and S 0 .
Proof. Take S 0 = MP .
Theorem A.3 (Completeness of the method for semi-completeness). If a program P is semi-complete w.r.t. a specification S then there exists a more general specification
S 0 ⊇ S such that the sufficient condition from Th. 5.6 holds for P and S 0 .
Proof. For P and each atom A ∈ S there exists a successful or infinite SLD-derivation.
We consider ground instances of such derivations. Let D be a set of derivations for
ground(P ), containing a successful or infinite derivation for each A ∈ S. Let S 0 be the
set of atoms selected in the derivations of D. Each atom of S 0 is covered by P .
Proving completeness of programs based on Prop. 5.4 and Th. 5.6 is not complete. It fails
e.g. when a program contains a void clause A ← A. However the method based on Th. 5.10
is complete.
Theorem A.4 (Completeness of the method for completeness, Th. 5.10 ). If
a program P is complete w.r.t. a specification S then there exists a more general specification
S 0 ⊇ S, and a level mapping | |, such that the sufficient condition from Th. 5.10 holds for
P , S 0 and | |.
Proof. For each A ∈ S ⊆ MP there exists a proof tree for ground(P ) with the root
A. Let T be the set of such trees and their subtrees. Let S 0 be the set of their nodes. For
a B ∈ S 0 , let |B| be the minimal height of a tree from T with the root B. Consider such a
minimal height tree. Let B1 , . . . , Bn be the children of the root B in the tree. So |B| > |Bi |
for each i. As B1 , . . . , Bn ∈ S 0 and B ← B1 , . . . , Bn ∈ ground(P ), atom B is recurrently
covered by P w.r.t. S 0 and | |.
From the last theorem it follows that if P is complete w.r.t. S then some S 0 ⊇ S is covered
by P . Thus the sufficient condition of Th. 5.6 (applied to some S 0 ⊇ S) is a necessary
condition for program completeness (w.r.t. S).
B. PROOFS
This appendix contains the remaining proofs (for sections 2, 5 and 6). Let us note that the
results of sections 2–6 hold also for infinite programs, unless stated otherwise.
B.1. The least Herbrand model and the logical consequences of a program
This section contains a proof of Lemma 2.2.
Lemma 2.2. Let P be a program, and Q a query such that
26
24-04-2015
(1) Q contains exactly k ≥ 0 (distinct) variables, and the underlying language has (at least)
k constants not occurring in P, Q, or a non-constant function symbol not occurring
in P, Q.
Then MP |= Q iff P |= Q.
Proof. The “if” case is obvious. Assume MP |= Q. Let V1 , . . . , Vk be the variables
occurring in the query Q. Consider a substitution ρ = {V1 /t1 , . . . , Vk /tk } where t1 , . . . , tk
are distinct ground terms whose main function symbols f1 , . . . , fk do not occur in P ∪ {Q}.
Note that none of f1 , . . . , fk occurs in any computed answer for P and Q. As MP |= Qρ
and Qρ is ground, P |= Qρ by Th. 4.30 of [Apt 1997].
By completeness of SLD-resolution, Qρ is an instance of some computed answer Qϕ for
P and Q: Qρ = Qϕσ. So (for i = 1, . . . , k) ti = Vi ϕσ, ti = fi (. . .) and fi does not occur in
Vi ϕ. Thus each Vi ϕ is a variable. As t1 , . . . , tk are distinct, V1 ϕ, . . . , Vk ϕ are distinct. Thus
Qϕ is a variant of Q and, by soundness of SLD-resolution, P |= Q.
B.2. Sufficient conditions for completeness
Here we prove Th. 5.6, 6.5, and Prop. 6.11. We begin with a slightly more general version
of Th. 6.5, from which Th. 6.5 follows immediately. The only difference is condition 3a.
Sn
Sn
Lemma B.1. Let P ⊇ i=1 Πi (where n > 0) be a program, S = i=1 Si a specification,
and T a csSLD-tree for Π1 , . . . , Πn . If
(1 ) for each i = 1, . . . , n, all the atoms from Si are covered by Πi w.r.t. S, and
(2 ) T is compatible with Π1 , . . . , Πn , S1 , . . . , Sn ,
(3 ) (a) all the branches of T are finite, or
(b) P is recurrent, or
(c) P is acceptable and T is built under the Prolog selection rule
then T is complete w.r.t. S.
Proof. Let Q0 be a node of T which has a ground instance Q0 σ such that S |= Q0 σ.
Let an atom A of Q0 and a program Πi be selected in Q0 . Now Aσ ∈ S, Aσ 6∈ S \ Si , thus
Aσ ∈ Si , and Aσ is covered by Πi w.r.t. S. Let Aσ ← B1 , . . . , Bm be a ground instance of
a clause from Πi , with B1 , . . . , Bm ∈ S. Let Q00 be Q0 σ with Aσ replaced by B1 , . . . , Bm .
Then, by the lifting theorem [Doets 1994, Th. 5.37], Q00 is an instance of a child Q+ of Q0
in T . Obviously, S |= Q00 .
Assume the root of T is Q. Let S |= Qθ for a ground instance Qθ of Q. By induction,
using the previous paragraph as the inductive step, there exists a branch ∆ in T which (i) is
infinite or its last node is the empty query, (ii) each its nonempty node has a ground instance
consisting of atoms from S, (iii) the sequence of these ground instances is a derivation Γ
for Qθ and ground (P ), and (iv) ∆ is a lift [Doets 1994] of Γ.
We show that ∆ is finite. This is immediately implied by condition 3a of the Lemma.
Condition 3b or 3c implies that ground(P ) is recurrent or acceptable. Hence Γ is finite (by
[Apt 1997, Corollaries 6.10, 6.25]); thus ∆ is finite too. By (i), ∆ is a successful derivation
(for Q and P ), thus so is Γ, by (iv). In the notation of [Doets 1994], the respective answers of
Γ, ∆ are Qθ = res(Γ) and res(∆) (the resultants of the derivations). By the lifting theorem,
Qθ is an instance of res(∆), which is an answer of T .
Now, let us take a program P and a specification S, and let n = 1, Π1 = P, S1 = S.
P is suitable for any atom, as S \ S1 = ∅. Thus each SLD-tree T for P is a csSLD-tree
compatible with P, S, and from Lemma B.1 it follows:
24-04-2015
27
Proposition B.2. If all the atoms from a specification S are covered w.r.t. S by a
program P , and there exists an SLD-tree for P and a query Q with no infinite branches
then P is complete for Q.
Th. 5.6 follows immediately from Prop. B.2. Now we show an example where Th. 5.6 is
not applicable, but Prop. B.2 is.
Example B.3. Program P = { p(s(X)) ← p(X). p(0). q(X) ← p(s(Y )).} is complete
w.r.t. S = {p(si (0)) | i ≥ 0} ∪ {q(0)}, but its completeness does not follow from semicompleteness. The program loops for any instance of Q = q(X) (the SLD-tree has an
infinite branch). However all the branches of the SLD-tree for q(X) and ground(P ) are
finite (although the tree has infinite height). Hence by Prop. B.2, ground(P ) is complete for
Q. Thus so is P , as both programs have the same least Herbrand model.
Sn
Sn
Proposition 6.11. Let P ⊇ i=1 Πi (n > 0) be a program, S = i=1 Si a specification,
| |: HB ,→ N a level mapping, and T a csSLD-tree for Π1 , . . . , Πn . If
(1) T is compatible with Π1 , . . . , Πn , S1 , . . . , Sn ,
(2) each atom from Si is recurrently covered by Πi w.r.t. S and | |, for each i = 1, . . . , n,
then T is complete w.r.t. S.
Proof. For a ground query Q = A1 , . . . , An , let us define the level of Q as the multiset
|Q| = bag(|A1 |, . . . , |An |), whenever |A1 |, . . . , |An | are defined. Let ≺m be the multiset
ordering [Apt 1997].
We follow the proof of Lemma B.1. The first part of the proof is basically the same, with
an additional restriction made possible by the atom Aσ ∈ S being recurrently covered by
Πi w.r.t. S. Namely for the ground clause Aσ ← B1 , . . . , Bm we additionally require that
|Aσ|, |B1 |, . . . |Bn | are defined, and and |Aσ| > |Bi | for all i = 1, . . . , n. Hence |Q00 | ≺m |Q0 σ|.
As in the proof of Lemma B.1, there is a branch ∆ in T and a derivation Γ for Qθ and
ground(P ), satisfying conditions (i),. . . ,(iv). For each query Q00 of Γ, S |= Q00 , and |Q00 |
is defined. As the levels of the queries in Γ are a ≺m -decreasing sequence, and ≺m is well
founded, Γ is finite. Hence Γ and ∆ are successful, and Qθ is an instance of an answer
of T .
REFERENCES
K. R. Apt. 1993. Declarative Programming in Prolog. In Logic Programming, Proc. of the 1993 International
Symposium, D. Miller (Ed.). The MIT Press, 12–35.
K. R. Apt. 1995. Program Verification and Prolog. In Specification and Validation Methods for Programming
Languages and Systems, E. Börger (Ed.). Oxford University Press, 55–95.
K. R. Apt. 1997. From Logic Programming to Prolog. Prentice-Hall.
K. R. Apt and D. Pedreschi. 1993. Reasoning about Termination of Pure Prolog Programs. Information
and Computation 106, 1 (1993), 109–157.
M. Bezem. 1993. Strong Termination of Logic Programs. J. Log. Program. 15, 1&2 (1993), 79–97.
A. Bossi and N. Cocco. 1989. Verifying Correctness of Logic Programs. In TAPSOFT, Vol.2 (Lecture Notes
in Computer Science), J. Dı́az and F. Orejas (Eds.), Vol. 352. Springer, 96–110.
G. Brewka, T. Eiter, and M. Truszczyński. 2011. Answer set programming at a glance. Commun. ACM 54,
12 (2011), 92–103. DOI:http://dx.doi.org/10.1145/2043174.2043195
K. L. Clark. 1979. Predicate logic as computational formalism. Technical Report 79/59. Imperial College,
London.
W. F. Clocksin and C. S. Mellish. 1981. Programming in Prolog. Springer.
M. Codish and C. Taboch. 1999. A Semantic Basis for the Termination Analysis of Logic Programs. J. Log.
Program. 41, 1 (1999), 103–123. DOI:http://dx.doi.org/10.1016/S0743-1066(99)00006-0
P. Deransart and J. Maluszyński. 1993. A grammatical view of logic programming. The MIT Press. I–XXXI,
1–454 pages.
28
24-04-2015
Y. Deville. 1990. Logic Programming: Systematic Program Development. Addison-Wesley.
Y. Deville and K. K. Lau. 1994. Logic Program Synthesis. J. Log. Program. 19/20 (1994), 321–350.
K. Doets. 1994. From Logic to Logic Programming. The MIT Press, Cambridge, MA.
W. Drabent. 2012a. Logic + control: An example. In Technical Communications of the 28th International Conference on Logic Programming (ICLP’12) (Leibniz International Proceedings in Informatics
(LIPIcs)), Agostino Dovier and Vı́tor Santos Costa (Eds.), Vol. 17. 301–311. http://drops.dagstuhl.de/
opus/volltexte/2012/3631.
W. Drabent. 2012b. Logic + control: An example of program construction. CoRR abs/1110.4978 (2012).
http://arxiv.org/abs/1110.4978.
W. Drabent. 2015a. On completeness of logic programs. In Logic Based Program Synthesis and Transformation, LOPSTR 2014. Revised Selected Papers (Lecture Notes in Computer Science), Vol. 8981.
Springer. Extended version in CoRR abs/1411.3015 (2014). http://arxiv.org/abs/1411.3015.
W. Drabent. 2015b. On definite program answers and least Herbrand models. CoRR abs/1503.03324 (2015).
http://arxiv.org/abs/1503.03324.
W. Drabent and J. Maluszyński. 1988. Inductive Assertion Method for Logic Programs. Theoretical Computer Science 59 (1988), 133–155.
W. Drabent and M. Milkowska. 2005. Proving correctness and completeness of normal programs – a declarative approach. Theory and Practice of Logic Programming 5, 6 (2005), 669–711.
W. Drabent, S. Nadjm-Tehrani, and J. Maluszyński. 1989. Algorithmic Debugging with Assertions. In
Meta-Programming in Logic Programming, H. Abramson and M. H. Rogers (Eds.). The MIT Press,
501–522.
S. Genaim and A. King. 2008. Inferring Non-Suspension Conditions for Logic Programs with Dynamic
Scheduling. ACM Trans. Comput. Log. 9, 3 (2008). http://www.cs.kent.ac.uk/pubs/2008/2456
C. J. Hogger. 1984. Introduction to Logic Programming. Academic Press, London.
J. M. Howe and A. King. 2012. A Pearl on SAT and SMT Solving in Prolog. Theor. Comput. Sci. 435
(2012), 43–55.
R. A. Kowalski. 1985. The Relation between Logic Programming and Logic Specification. In Mathematical
Logic and Programming Languages, C. Hoare and J. Shepherdson (Eds.). Prentice-Hall, 11–27. Also in
Phil. Trans. R. Soc. Lond. A, Vol 312, 1984, 345–361.
K. Kunen. 1987. Negation in Logic Programming. J. Log. Program. 4, 4 (1987), 289–308.
F. Mesnard and R. Bagnara. 2005. cTI: A constraint-based termination inference tool for ISO-Prolog. TPLP
5, 1-2 (2005), 243–257. DOI:http://dx.doi.org/10.1017/S1471068404002017
L. Naish. 1992. Declarative Diagnosis of Missing Answers. New Generation Comput. 10, 3 (1992), 255–286.
L. Naish. 1996. A Declarative View of Modes. In Proceedings of JICSLP’96. The MIT Press, 185–199.
L. Naish. 2000. A Three-Valued Declarative Debugging Scheme. In 23rd Australasian Computer Science
Conference (ACSC 2000). IEEE Computer Society, 166–173.
L. Naish. 2006. A three-valued semantics for logic programmers. TPLP 6, 5 (2006), 509–538.
L. Naish and H. Søndergaard. 2014. Truth versus information in logic programming. TPLP 14, 6 (2014),
803–840. DOI:http://dx.doi.org/10.1017/S1471068413000069
M. T. Nguyen, D. De Schreye, J. Giesl, and P. Schneider-Kamp. 2011. Polytool: Polynomial interpretations as a basis for termination analysis of logic programs. TPLP 11, 1 (2011), 33–63.
DOI:http://dx.doi.org/10.1017/S1471068410000025
A. Paschke. 2011. Rules and Logic Programming for the Web. In Reasoning Web. Semantic Technologies
for the Web of Data - 7th International Summer School 2011, Galway, Ireland, August 23-27, 2011,
Tutorial Lectures (Lecture Notes in Computer Science), Axel Polleres et al. (Ed.), Vol. 6848. Springer,
326–381. DOI:http://dx.doi.org/10.1007/978-3-642-23032-5 6
L. M. Pereira. 1986. Rational Debugging in Logic Programming. In ICLP (Lecture Notes in Computer
Science), E. Y. Shapiro (Ed.), Vol. 225. Springer, 203–210.
A. Pettorossi, M. Proietti, and V. Senni. 2010. The Transformational Approach to Program Development.
In A 25-Year Perspective on Logic Programming: Achievements of the Italian Association for Logic
Programming, GULP (Lecture Notes in Computer Science), A. Dovier and E. Pontelli (Eds.), Vol.
6125. Springer, 112–135. DOI:http://dx.doi.org/10.1007/978-3-642-14309-0 6
F. Pfenning (Ed.). 1992. Types in Logic Programming. The MIT Press.
P. Schneider-Kamp, J. Giesl, A. Serebrenik, and R. Thiemann. 2009. Automated termination
proofs for logic programs by term rewriting. ACM Trans. Comput. Log. 11, 1 (2009).
DOI:http://dx.doi.org/10.1145/1614431.1614433
E. Shapiro. 1983. Algorithmic Program Debugging. The MIT Press.
24-04-2015
29
R. F. Stärk. 1998. The Theoretical Foundations of LPTP (A Logic Program Theorem Prover). J. Log.
Program. 36, 3 (1998), 241–269.
L. Sterling and E. Shapiro. 1994. The Art of Prolog (2 ed.). The MIT Press.
T. Swift and D. S. Warren. 2012. XSB: Extending Prolog with Tabled Logic Programming. TPLP 12, 1-2
(2012), 157–187. DOI:http://dx.doi.org/10.1017/S1471068411000500
A. Van Gelder, K. A. Ross, and J. S. Schlipf. 1991. The Well-Founded Semantics for General Logic Programs.
J. ACM 38, 3 (1991), 620–650. DOI:http://dx.doi.org/10.1145/116825.116838
| 6cs.PL
|
Lightweight User-Space Record And Replay
Robert O’Callahan, Chris Jones, Nathan Froyd, Kyle Huey, Albert Noll, Nimrod Partush
arXiv:1610.02144v1 [cs.PL] 7 Oct 2016
Abstract
The ability to record and replay program executions with
low overhead enables many applications, such as reverseexecution debugging, debugging of hard-to-reproduce test
failures, and “black box” forensic analysis of failures in
deployed systems. Existing record-and-replay approaches
rely on recording an entire virtual machine (which is
heavyweight), modifying the OS kernel (which adds deployment and maintenance costs), or pervasive code instrumentation (which imposes significant performance
and complexity overhead). We investigated whether it
is possible to build a practical record-and-replay system
avoiding all these issues. The answer turns out to be
yes — if the CPU and operating system meet certain
non-obvious constraints. Fortunately modern Intel CPUs,
Linux kernels and user-space frameworks meet these constraints, although this has only become true recently. With
some novel optimizations, our system RR records and replays real-world workloads with low overhead with an entirely user-space implementation running on stock hardware and operating systems. RR forms the basis of an
open-source reverse-execution debugger seeing significant use in practice. We present the design and implementation of RR, describe its performance on a variety of
workloads, and identify constraints on hardware and operating system design required to support our approach.
1
Introduction
The ability to record a program execution with low overhead and play it back precisely has many applications
[13, 14, 18] and has received significant attention in the
research community. It has even been implemented in
products such as VMware Workstation [27], Simics [19],
UndoDB [1] and TotalView [21]. Unfortunately, deployment of these techniques has been limited, for various
reasons. Some approaches [16, 19, 27] require recording
and replaying an entire virtual machine, which is heavyweight. Other approaches [6, 13, 25] require running a
modified OS kernel, which is a major barrier for users
and adds security and stability risk to the system. Some
approaches [23, 28, 32] require custom hardware not yet
available. Many approaches [1, 7, 21, 31] require pervasive instrumentation of code, which adds complexity and
overhead, especially for self-modifying code (commonly
used in polymorphic inline caching [24] and other implementation techniques in modern just-in-time compilers). A performant dynamic code instrumentation engine
is also expensive to build and maintain.
We set out to build a system that maximizes deployability by avoiding all these issues: to record and replay
unmodified user-space applications on stock Linux kernels and x86/x86-64 CPUs, with a fully user-space implementation running without special privileges, and without using pervasive code instrumentation. We assume
that RR should run unmodified applications, and they will
have bugs (including data races) that we wish to faithfully
record and replay, but these applications will not maliciously try to subvert recording or replay. We combine
techniques already known, but not previously demonstrated working together in a practical system: primarily,
using ptrace to record and replay system call results
and signals, avoiding non-deterministic data races by running only one thread at a time, and using CPU hardware
performance counters to measure application progress so
asynchronous signal and context-switch events are delivered at the right moment [30]. Section 2 describes our
approach in more detail.
With the basic functionality in place, we discovered
that the main performance bottleneck for our applications
was the context switching required by using ptrace
to monitor system calls. We implemented a novel
system-call buffering optimization to eliminate those context switches, dramatically reducing recording and replay overhead on important real-world workloads. This
optimization relies on modern Linux kernel features:
seccomp-bpf to selectively suppress ptrace traps for
certain system calls, and perf context-switch events to
detect recorded threads blocking in the kernel. Section
3 describes this work, and Section 4 gives some performance results, showing that on important application
workloads RR recording and replay slowdown is less than
a factor of two.
We rely on hardware and OS features designed for other
goals, so it is surprising that RR works. In fact, it skirts
the edge of feasibility, and in particular it cannot be implemented on ARM CPUs. Section 5 summarizes RR’s
hardware and software requirements, which we hope will
influence system designers.
RR is in daily use by many developers as the foundation of an efficient reverse-execution debugger that works
on complex applications such as Samba, Firefox, QEMU,
LibreOffice and Wine. It is free software, available at
https://github.com/mozilla/rr. This paper
makes the following research contributions:
are not performed. Instead the recorded user-space-visible
effects of those operations, and future related operations,
are replayed. We do create one replay thread per recorded
thread (not strictly necessary), and we create one replay
address space (i.e. process) per recorded address space,
along with matching memory mappings.
With this design, our recording boundary is the interface between user-space and the kernel. The inputs and
sources of nondeterminism are mainly the results of system calls, and the timing of asynchronous events.
• We show that record and replay of Linux user-space
processes on modern, stock hardware and kernels 2.2 Avoiding Data Races
and without pervasive code instrumentation is posIf we allow threads to run on multiple cores simultanesible and practical.
ously and share memory, racing read-write or write-write
• We introduce the system-call buffering optimization accesses to the same memory location by different threads
and show that it dramatically reduces overhead.
would be a source of nondeterminism. It seems impractical to record such nondeterminism with low overhead
• We show that recording and replay overhead is low
on stock hardware, so instead we avoid such races by alin practice, for applications that don’t use much parlowing only one recorded thread to run at a time. RR preallelism.
emptively schedules these threads, so context switches are
• We identify hardware and operating system design a source of nondeterminism that must be recorded. Data
race bugs can still be observed if a context switch occurs at
constraints required to support our approach.
the right point in the execution (though bugs due to weak
memory models cannot be observed).
2 Design
This approach is simple and efficient for applications
without much parallelism. It imposes significant slow2.1 Summary
down for applications with a consistently high degree of
Most low-overhead record-and-replay systems depend on parallelism, but those are still relatively uncommon.
the observation that CPUs are mostly deterministic. We
identify a boundary around state and computation, record
all sources of nondeterminism within the boundary and all
inputs crossing into the boundary, and reexecute the computation within the boundary by replaying the nondeterminism and inputs. If all inputs and nondeterminism have
truly been captured, the state and computation within the
boundary during replay will match that during recording.
To enable record and replay of arbitrary Linux applications, without requiring kernel modifications or a virtual
machine, RR records and replays the user-space execution of a group of processes. To simplify invariants, and
to make replay as faithful as possible, replay preserves
almost every detail of user-space execution. In particular, user-space memory and register values are preserved
exactly, with a few exceptions noted later in the paper.
This implies CPU-level control flow is identical between
recording and replay, as is memory layout.
While replay preserves user-space state and execution,
only a minimal amount of kernel state is reproduced during replay. For example, file descriptors are not opened,
signal handlers are not installed, and filesystem operations
2.3
System Calls
System calls return data to user-space by modifying registers and memory, and these changes must be recorded.
The ptrace system call allows a process to supervise the
execution of other “tracee” processes and threads, and to
be synchronously notified when a tracee thread enters or
exits a system call. When a tracee thread enters the kernel
for a system call, it is suspended and RR is notified. When
RR chooses to run that thread again, the system call will
complete, notifying RR again, giving it a chance to record
the system call results. RR contains a model of most Linux
system calls describing the user-space memory they can
modify, given the system call input parameters and result.
As noted above, RR normally avoids races by scheduling only one thread at a time. However, if a system
call blocks in the kernel, RR must try to schedule another application thread to run while the blocking system call completes. It’s possible (albeit unlikely) that
the running thread could access the system call’s output
buffer and race with the kernel’s writes to that buffer. To
2
Tracee thread T
read(fd, buf, size)
ptrace_notify
sys_read
ptrace_notify
2.4.1
Recorder thread
waitpid(T)
redirect arg2 to scratch
ptrace(T, CONT_SYSCALL)
Nondeterministic Performance Counters
Measuring user-space progress with a hardware performance counter for record-and-replay requires that every
time we execute a given sequence of user-space instructions, the counter value changes by an amount that depends only on the instruction sequence, and not any system state unobservable from user space (e.g. the contents
of caches, the state of page tables, or speculative CPU
state). As noted in previous work [16, 37], this property
(commonly described as “determinism”) does not hold for
most CPU performance counters in practice. For example
it does not hold for any “instructions retired” counter on
any known x86 CPU model, because (among other reasons) an instruction triggering a page fault is restarted and
counted twice.
Fortunately, modern Intel CPUs have exactly one deterministic performance counter: “retired conditional
branches” (“RCB”), so we use that. We cannot just count
the number of RCBs during recording and deliver the signal after we have executed that number of RCBs during
replay, because the RCB count does not uniquely determine the execution point to deliver the signal at. Therefore we pair the RCB count with the complete state of
general-purpose registers (including the program counter)
to identify an execution point. In theory that still does not
uniquely identify an execution point: consider, for example, the infinite loop label: inc [global var];
jmp label;. In practice applications do not do that sort
of thing and we have found it works reliably.
d
waitpid(T)
N = syscall_result_reg
save N scratch bytes to trace
copy N scratch bytes to buf
ptrace(T, CONT_SYSCALL)
Figure 1: Recording simple system call
avoid this, we redirect system call output buffers to perthread temporary “scratch memory” which is otherwise
unused by the application. When we get a ptrace event
for a blocked system call completing, RR copies scratch
buffer contents to the real user-space destination(s) while
no other threads are running, eliminating the race.
Figure 1 illustrates the flow of control recording a simple read system call. The gray box represents kernel
code.
During replay, when the next event to be replayed is
an intercepted system call, we set a temporary breakpoint
at the address of the system call instruction (recorded in
the trace). We use ptrace to run the tracee thread until it hits the breakpoint, remove the breakpoint, advance
the program counter past the system call instruction, and
apply the recorded register and memory changes. This
approach is simple and minimizes the number of context
switches between RR and the tracee thread. (Occasionally it is unsafe and we fall back to a more complicated 2.4.2 Late Interrupt Firing
mechanism.)
The other major problem is that, although CPUs can be
programmed to fire an interrupt after a specified number
of performance events have been observed, the interrupt
does not fire immediately. In practice we often observe it
2.4 Asynchronous Events
firing after dozens more instructions have retired. To comWe need to support two kinds of asynchronous events: pensate for this, during replay, we program the interrupt to
preemptive context switches and signals. The former is trigger some number of events earlier than the actual RCB
handled as a special case of the latter; we force a context count we are expecting. Then we set a temporary breakswitch by sending a signal to a running tracee thread. In point at the program counter value for the state we’re trygeneral we need to ensure that during replay, a signal is ing to reach, and repeatedly run to the breakpoint until the
delivered when the program is in exactly the same state as RCB count and the general-purpose register values match
it was when the signal was delivered during recording.
their recorded values.
Our approach is very similar to ReVirt’s [16]: use CPU
hardware performance counters. Ideally we would count 2.5 Shared Memory
the number of retired instructions leading up to a signal
delivery during recording, and during replay, program the By scheduling only one thread at a time, RR avoids issues
CPU to fire an interrupt after that many instructions have with races on shared memory as long as that memory is
been retired. However, this simple approach needs two written only by tracee threads. It is possible for recorded
processes to share memory with other processes, and even
modifications to work in practice.
3
on that core during replay.
We could easily avoid most of these issues in wellbehaved programs if we could just trap-and-emulate the
CPUID instruction, since then we could mask off the feature bits indicating support for RDRAND, hardware transactions, etc. Modern Intel CPUs support this (“CPUID
faulting”); we are in the process of adding an API for this
to Linux.
kernel device drivers, where that non-recorded code can
perform writes that race with accesses by tracee threads.
Fortunately, this is rare for applications running in common Linux desktop environments, occurring in only four
common cases: applications sharing memory with the
PulseAudio daemon, applications sharing memory with
the X server, applications sharing memory with kernel
graphics drivers and GPUs, and vdso syscalls. We avoid
the first three problems by automatically disabling use of
shared memory with PulseAudio and X (falling back to a
socket transport in both cases), and disabling direct access
to the GPU from applications.
vdso syscalls are a Linux optimization that implements some common read-only system calls (e.g.
gettimeofday) entirely in user space, partly by reading memory shared with the kernel and updated asynchronously by the kernel. We disable vdso syscalls by
patching their user-space implementations to perform the
equivalent real system call instead.
Applications could still share memory with nonrecorded processes in problematic ways, though this is
rare in practice and can often be solved just by enlarging
the scope of the group of processes recorded by RR.
2.6
2.7
Reducing Trace Sizes
For many applications the bulk of their input is memorymapped files, mainly executable code. Copying all executables and libraries to the recorded trace on every execution would impose significant time and space overhead.
RR creates hard links to memory-mapped executable files
instead of copying them; as long as a system update or
recompile replaces executables with new files, instead of
writing to the existing files, the links retain the old file
data. This works well in practice.
Even better, modern filesystems such as XFS and Btrfs
offer copy-on-write logical copies of files (and even block
ranges within files), ideal for our purposes. When a
mapped file is on the same filesystem as the recorded
trace, and the filesystem supports cloning, RR clones
mapped files into the trace. These clone operations are
essentially free in time and space, until/unless the original
file is modified or deleted.
RR compresses all trace data, other than cloned files and
blocks, with the zlib “deflate” method.
With these optimizations, in practice trace storage is a
non-issue. Section 4.4 presents some results.
Nondeterministic Instructions
Almost all CPU instructions are deterministic, but some
are not. One common nondeterministic x86 instruction is
RDTSC, which reads a time-stamp counter. This particular
instruction is easy to handle, since the CPU can be configured to trap on an RDTSC and Linux exposes this via
a prctl API, so we can trap, emulate and record each
RDTSC.
Other relatively recent x86 instructions are harder to
handle. RDRAND generates random numbers and hopefully is not deterministic. We have only encountered it
being used in one place in GNU libstdc++, so RR
patches that explicitly. XBEGIN and associated instructions support hardware transactional memory. These are
nondeterministic from the point of view of user space,
since a hardware transaction can succeed or fail depending on CPU cache state. Fortunately so far we have only
found these being used by the system pthreads library,
and we dynamically apply custom patches to that library
to disable use of hardware transactions.
The CPUID instruction is mostly deterministic, but one
of its features returns the index of the running core, which
affects behavior deep in glibc and can change as the kernel migrates a process between cores. We use the Linux
sched setaffinity API to force all tracee threads to
run on a particular fixed core, and also force them to run
2.8
Other Details
Apart from those major issues, many other details are required to build a complete record-and-replay system, too
many to mention here. Some system calls (e.g. execve)
are especially complex to handle. Recording and replaying signal delivery are complex, partly because signal delivery has poorly-documented side effects on userspace memory. Advanced Linux kernel features such as
unshare (kernel namespaces) and seccomp require
thoughtful handling. Many of these details are interesting, but they do not impact the overall approach.
3
System Call Buffering
The approach described in the previous section works,
but overhead is disappointingly high (see Figure 5 below). The core problem is that for every tracee system
4
call, as shown in Figure 1 the tracee performs four conTracee thread T
text switches: two blocking ptrace notifications, each
read(fd, buf, size)
syscall_hook()
requiring a context switch from the tracee to RR and back.
redirect arg2 to syscall buffer
For common system calls such as gettimeofday or
seccomp-bpf filter → ALLOW
read from cached files, the cost of even a single context
sys_read
switch dwarfs the cost of the system call itself. To signifN = syscall_result_reg
icantly reduce overhead, we must avoid context-switches
copy N syscall buffer bytes to buf
write system-call record to syscall buffer
to RR when processing these common system calls.
Therefore, we inject into the recorded process a library
that intercepts common system calls, performs the system
Figure 2: Recording with system-call buffering
call without triggering a ptrace trap, and records the results to a dedicated buffer shared with RR. RR periodically
flushes the buffer to its trace. The concept is simple but system call performed by the target process, the kernel
there are problems to overcome.
runs the filter. The filter has access to incoming user-space
register values, including the program counter. The filter
result directs the kernel to either allow the system call,
3.1 Intercepting System Calls
fail with a given errno, kill the target process, or trigger
When the tracee makes a kernel-entering system call, RR a ptrace trap. The overhead of filter execution is negis notified via a ptrace trap. It tries to rewrite the ligible since filters run directly in the kernel and are even
system-call instruction to call into the interception library compiled to native code on most architectures.
instead. This is tricky because on x86 a system call inFigure 2 illustrates recording a simple read system
struction is two bytes long, but we need to replace it with
call with system-call buffering. The solid-border box repa five-byte call instruction. (On x86-64, to ensure we
resents code in the interception library and the grey box
can call from anywhere in the address space to the interrepresents kernel code.
ception library, we also need to also allocate trampolines
RR injects a special page of memory into every tracee
within 2GB of the patched code.) The vast majority of
process at a fixed address (immediately after execve).
executed system call instructions are followed by one of
That page contains a system call instruction — the “unonly five different instructions; for example, very many
traced instruction”. RR applies a seccomp-bpf filter to
system call instructions are followed by a particular cmp
each recorded process that triggers a ptrace trap for evinstruction testing the syscall result. So we added five
ery system call — except when the program counter is at
hand-written stubs to our interception library that execute
the untraced instruction, in which case the call is allowed.
those post-system-call instructions before returning to the
Whenever the interception library needs to make an unpatched code, and on receipt of a ptrace system-call
traced system call, it uses that instruction.
notification, RR replaces the system call instruction and
its following instruction with a call to the corresponding
stub.
3.3 Detecting Blocked System Calls
For simplicity we (try to) redirect all system call instructions to the interception library, but the library only Some common system calls sometimes block (e.g. read
contains wrappers for the most frequently called system on an empty pipe). Because RR runs tracee threads one
calls. For other system calls it falls back to doing a regu- at a time, if a thread enters a blocking system call without notifying RR, it will hang and could cause the enlar ptrace-trapping system call.
tire recording to deadlock (e.g. if another tracee thread
is about to write to the pipe). We need the kernel to no3.2 Selectively Trapping System Calls
tify RR and suspend the tracee thread whenever it blocks
The classic ptrace system-call interception API trig- in an untraced system call, to ensure we get a chance to
gers traps for all system calls, but our interception li- schedule a different tracee thread.
We can do this using the Linux perf event system to
brary needs to avoid traps for selected system calls. Fortunately, modern Linux kernels provide a facility for monitor PERF COUNT SW CONTEXT SWITCHES. One
selectively generating ptrace traps: seccomp-bpf. of these events occurs every time the kernel deschedules a
seccomp-bpf was designed primarily for sandboxing. thread from a CPU core. The interception library opens a
A process can apply a seccomp-bpf filter function, ex- file descriptor for each thread to monitor these events for
pressed in bytecode, to another process; then, for every that thread, and requests that the kernel send a signal to the
5
library is part of the recording and replay and therefore
should execute the same code in both cases.
For this reason (and because we want to avoid races of
the sort discussed in Section 2.3), the interception library
redirects system call outputs so that results are written directly to the trace buffer. After the system call completes,
the interception library copies the output data from the
trace buffer to the original output buffer(s). During replay the untraced system call instruction is replaced with
a no-op, so the system call does not occur; the results
are already present in the trace buffer so the post-systemcall copy from the trace buffer to the output buffer(s) does
what we need.
During recording, each untraced system call sets a result register and the interception library writes it to the
trace buffer. During replay we must read the result register from the trace buffer instead. We accomplish this
with a conditional move instruction so that control flow
is perfectly consistent between recording and replay. The
condition is loaded from an is replay global variable,
so the register holding the condition is different over a
very short span of instructions (and explicity cleared afterwards).
Another tricky issue is handling system call memory
parameters that are “in-out”. During recording we must
copy the input buffer to the trace buffer, pass the system
call a pointer to the trace buffer, then copy the trace buffer
contents back to the input buffer. If we perform the first
copy during replay, we’ll overwrite the trace buffer values
holding the system call results we need. During replay we
turn that copy into a no-op using a conditional move to set
the source address copy to the destination address.
We could allow the replay behavior of the interception
library to diverge further from its recording behavior, but
it would have to be done very carefully. At least we’d
have to ensure the number of retired conditional branches
was identical along both paths, and that register values
were consistent whenever we exit the interception library
or trap to RR within the interception library. It’s simplest
to minimize the divergence.
Tracee thread T
read(fd, buf, size)
syscall_hook()
redirect arg2 to syscall buffer
enable SWITCHES event
seccomp-bpf → ALLOW
sys_read
SWITCHES event;
interruption by signal
ptrace_notify
Tracee thread T2
futex_wait()
ptrace_notify
sys_futex
ptrace_notify
Recorder thread
waitpid(T)
determine T is blocked
in syscall buffering
T2 = waitpid(-1)
record T2 system call
ptrace(T2,
CONT_SYSCALL)
Figure 3: Recording a blocking system call
thread every time the event occurs. These signals trigger
ptrace notifications to RR while preventing the thread
from executing further. To avoid spurious signals (e.g.
when the thread is descheduled due to normal timeslice
expiration), event counting is normally disabled and explicitly enabled during an untraced system call that might
block. It is still possible for spurious SWITCHES to occur at any point between enabling and disabling the event;
we handle these edge cases with careful inspection of the
tracee state.
Figure 3 illustrates recording a blocking read system
call with system-call buffering. The kernel deschedules
the thread, triggering a perf event which sends a signal
to the thread, rescheduling it, interrupting the system call,
and sending a ptrace notification to the recorder. The
recorder does bookkeeping to note that a buffered system
call was interrupted in thread T, then checks whether any
tracee threads in blocking system calls have progressed
to a system-call exit and generated a ptrace notification. In this example T2 has completed a (not systemcall-buffered) blocking futex system call, so we resume
executing T2.
Resuming a buffered system call that was interrupted
by a signal (e.g. T’s read call in Figure 3) is more complicated and requires understanding Linux system call 3.5 Optimizing Reads With Block Cloning
restart semantics, which are too nasty to cover in this paWhen an input file is on the same filesystem as the
per.
recorded trace and the filesystem supports copy-on-write
cloning of file blocks, for large block-aligned reads the
3.4 Handling Replay
system call buffering code clones the data to a per-thread
Conceptually, during recording we need to copy system “cloned-data” trace file, bypassing the normal system-call
call output buffers to a trace buffer, and during replay we recording logic. This greatly reduces space and time overneed to copy results from the trace buffer to system call head for file-read-intensive workloads; see the next secoutput buffers. This is a problem because the interception tion.
6
This optimization works by cloning the input blocks
and then reading the input data from the original input
file. This opens up a possible race: between the clone
and the read, another process could overwrite the input
file data, in which case the data read during replay would
differ from the data read during recording, causing replay
to fail. However, when a file read races with a write under Linux, the reader can receive an arbitrary mix of old
and new data, so such behavior would almost certainly
be a severe bug, and in practice such bugs do not seem
to be common. The race could be avoided by reading
from the cloned-data file instead of the original input file,
but that performs very poorly because it defeats Linux’s
readahead optimizations (since the data in the cloned-data
file is never available until just before it’s needed).
htmltest runs the Mozilla Firefox HTML forms
tests (Mercurial revision 9bd900888753).
The
harness is excluded from recording (using mach
mochitest -f plain --debugger R R
dom/html/test/forms).
This is an example
from real-world usage. About 30% of user-space CPU
time is in the harness.
sambatest
runs
a
Samba
(git
revision
9ee4678b) UDP echo test via make test
TESTS=samba4.echo.udp. This is an example
from real-world usage.
All tests run on a Dell XPS15 laptop with a quad-core
Intel Skylake CPU (8 SMT threads), 16GB RAM and a
512GB SSD using Btrfs in Fedora Core 23 Linux.
4.2
4
Results
Table 1 shows the wall-clock run time of various configurations, normalized to the run time of the baseline configuration. For octane, because the benchmark is designed
to run for a fixed length of time and report a score, we
report the ratio of the baseline score to the configurationunder-test score instead — except for replay tests, where
the reported score will necessarily be the same as the score
during recording. For octane replay tests, we report the ratio of the baseline score to the recorded score, multiplied
by the ratio of replay run time to recording run time. Each
test was run six times, discarding the first result and reporting the geometric mean of the other five results. Thus
the results represent warm-cache performance.
“Single core” reports the overhead of just restricting all
threads to a single core using Linux taskset.
“Record no-syscallbuf” and “Replay no-syscallbuf” report overhead with the system-call buffering optimization
disabled (which also disables block cloning). “Record
no-cloning” reports overhead with just block cloning disabled.
“DynamoRio-null” reports the overhead of running the
tests under the DynamoRio [8] (version 6.1.0, the latest
stable version at time of writing) “null tool”, to estimate
a lower bound for the overhead of using dynamic code
instrumentation as an implementation technique. (DynamoRio is reported to be among the fastest dynamic code
instrumentation engines.)
Our results address the following questions:
• What is the run-time overhead of RR recording and
replay, across different kinds of workloads?
• How much of the overhead is due to the single-core
limitation?
• What are the impacts of the system-call buffering and
block cloning optimizations?
• What is the impact of not having to instrument code?
• How much space do RR traces consume?
4.1
Overhead
Workloads
We present a variety of workloads to illuminate RR’s
strengths and weaknesses. Benchmarks were chosen to
fit in system memory (to minimize the impact of I/O on
test results), and to run for about 30 seconds each (except
for cp where a 30s run time would require it to not fit in
memory).
cp duplicates a git checkout of glibc (revision
2d02fd07) using cp -a (15200 files constituting 732MB
of data, according to du -h). cp is single-threaded,
making intensive use of synchronous reads and a variety
of other filesystem-related system calls.
make builds DynamoRio [8] (version 6.1.0) with make
-j8 (-j8 omitted when restricting to a single core). This
tests potentially-parallel execution of many short-lived
processes.
octane runs the Google Octane benchmark under the
Mozilla Spidermonkey Javascript engine (Mercurial revision 9bd900888753). This illustrates performance on
CPU-intensive code in a complex language runtime.
4.3
Observations
Overhead on make is significantly higher than for the
other workloads. Just forcing make onto a single core imposes major slowdown. Also, make creates many shortlived processes; it forks and execs 2430 processes. (The
next most prolific workload is sambatest with 89.) Our
7
Baseline
duration
1.04s
20.99s
32.73s
23.74s
31.75s
Workload
cp
make
octane
htmltest
sambatest
Record
1.49×
7.85×
1.79×
1.49×
1.57×
Replay
0.72×
11.93×
1.56×
1.01×
1.23×
Single
core
0.98×
3.36×
1.36×
1.07×
0.95×
Record
no-syscallbuf
24.53×
11.32×
2.65×
4.66×
2.23×
Replay
no-syscallbuf
15.39×
14.36×
2.43×
3.43×
1.74×
Record
no-cloning
3.68×
7.84×
1.80×
1.50×
1.57×
DynamoRionull
1.24×
10.97×
crash
14.03×
1.43×
Table 1: Run-time overhead
14 24.53
Record
Replay
2.0
Overhead relative to baseline
Overhead relative to baseline
2.5
1.5
1.0
0.5
0.0
cp
octane
htmltest
Workload
10
8
6
4
2
0
sambatest
Record
Record-no-syscallbuf
Record-no-cloning
12
cp
make
octane htmltest sambatest
Workload
Figure 4: Run-time overhead excluding make
Figure 5: Impact of optimizations
system-call buffering optimization only starts working in
a process once RR’s preload library has been loaded, but
typically at least 80 system calls are performed before that
completes, so the effectiveness of system-call buffering is
limited with short-lived processes.
Figure 4 shows the overall recording and replay overhead for workloads other than make. Error bars in figures
show 95% confidence intervals; these results are highly
stable across runs.
Excluding make, RR’s recording slowdown is less than
a factor of two. Excluding make, RR’s replay overhead
is lower than its recording overhead. Replay can even be
faster than normal execution, in cp because system calls
do less work. For interactive applications, not represented
here, replay can take much less time than the original execution because idle periods are eliminated.
octane is the only workload here other than make making significant use of multiple cores, and this accounts for
the majority of RR’s overhead on octane.
Figure 5 shows the impact of system-call buffering and
blocking cloning on recording. The system-call buffering
optimization produces a large reduction in recording (and
replay) overhead. Cloning file data blocks is a major im-
provement for cp recording but has essentially no effect
on the other workloads.
Figure 6 compares RR recording overhead with the
overhead of DynamoRio’s “null tool”, which runs all
code through the DynamoRio instrumentation engine but
does not modify the code beyond whatever is necessary to maintain supervised execution; this represents
a minimal-overhead code instrumentation configuration.
DynamoRio crashed on octane 1 cp executes very little
user-space code and DynamoRio’s overhead is low on
that workload. On make and sambatest DynamoRio overhead is similar to RR recording, even though on make
DynamoRio can utilize multiple cores. On htmltest DynamoRio’s overhead is very high, possibly because that
test runs a lot of Javascript with dynamically generated
and modified machine code. Implementing record-andreplay on top of dynamic instrumentation would incur significant additional overhead, so we would expect the resulting system to have significantly higher overhead than
RR .
1 We reported DynamoRio’s crash on our “octane” workload
to the developers at https://github.com/DynamoRIO/
dynamorio/issues/1930.
8
Workload
cp
make
octane
htmltest
sambatest
14
12
10
Record
DynamoRio-null
8
6
4
cp
make
htmltest
Workload
800
sambatest
Figure 6: Comparison with DynamoRio-null
Workload
cp
make
octane
htmltest
sambatest
Baseline
Record
Replay
1000
2
0
Compressed
MB/s
19.03
15.82
0.08
0.79
6.85
deflate
ratio
4.87×
8.32×
8.33×
5.94×
21.87×
Baseline
0.51
510.51
419.47
690.81
298.68
Record
34.54
327.19
610.48
692.06
400.49
Replay
9.11
314.16
588.01
324.71
428.79
Single core
0.51
288.65
392.95
689.75
303.03
Table 3: Memory usage (peak PSS MB)
Peak PSS (MB)
Overhead relative to baseline
16
Cloned blocks
MB/s
586.14
5.50
0.00
0.00
0.00
Single core
Supervisor process
600
400
200
0
cp
make
octane htmltest sambatest
Workload
Figure 7: Memory usage
Table 2: Storage space usage
baseline configuration of the workload. The space consumed by each trace shows very little variation between
runs.
RR traces contain three kinds of data:
Like cloned files, cloned file blocks do not consume
space as long as the underlying data they’re cloned from
• Cloned (or hard-linked) files used for memory-map
persists.
operations
The different workloads have highly varying space consumption
rates, but several megabytes per second is easy
• Cloned file blocks
for modern systems to handle. Compression ratios vary
• All other trace data, especially event metadata and depending on the kind of data being handled by the workload. In any case, in real-world usage trace storage has
the results of general system calls
not been a concern.
Memory-mapped files are almost entirely just the executables and libraries loaded by tracees. As long as the
4.5 Memory Usage
original files don’t change and are not removed, which
is usually true in practice, their clones take up no addi- Table 3 shows the memory usage of each workload. Evtional space and require no data writes. For simplicity, ery 10ms we sum the proportional-set-size (“PSS”) values
RR makes no attempt to eliminate duplicate file clones, so of all workload processes (including RR if running); we
most traces contain many duplicates and reporting mean- determine the peak values for each run and take their geoingful space usage for these files is both difficult and metric mean. In Linux, each page of memory mapped into
unimportant in practice.
a process’s address space contributes 1/n pages to that
Table 2 shows the space usage of each workload, in process’s PSS, where n is the number of processes mapmegabytes per second, for the general trace data and the ping the page; thus it is meaningful to sum PSS values
cloned file blocks. We compute the geometric mean of the over processes which share memory. The same data are
data usage for each trace and divide by the run-time of the shown in Figure 7. In the figure, the fraction of PSS used
4.4
Storage Space Usage
9
by the RR process is shown in orange. Memory usage data
was gathered in separate runs from the timing data shown
above, to ensure the overhead of gathering memory statistics did not impact those results.
Given these experiments ran on an otherwise unloaded
machine with 16GB RAM and all data fits in cache, none
of these workloads experienced any memory pressure.
cp uses almost no memory. In make, just running on a
single core reduces peak PSS significantly because not
as many processes run simultaneously. In octane memory usage is volatile (highly sensitive to small changes
in GC behavior) but recording significantly increases application memory usage; recording also increases application memory usage a small amount in sambatest but
slightly decreases it in htmltest. (We would expect to
see a small increase in application memory usage due to
system-call buffers and scratch buffers mapped into application processes.) These effects are difficult to explain
due to the complexity of the applications, but could be
due to changes in timing and/or effects on application or
operating system memory management heuristics.
Replay memory usage is similar to recording except in
htmltest, where it’s dramatically lower because we’re not
replaying the test harness.
The important point is that RR’s memory overhead is
relatively modest and is not an issue in practice.
5
5.1
Hardware and Software Design
Constraints
work in KVM and VMware (Xen not tested). Interestingly the VMware VM exit clustering optimization [4],
as implemented, breaks the determinism of the retiredconditional-branch counter and must be manually disabled. Given a pair of close-together CPUID instructions separated by a conditional branch, the clustering optimization translates the code to an instruction sequence
that runs entirely in the host, reducing the number of VM
exits from two (one per CPUID) to one, but performance
counter effects are not emulated so the conditional branch
is not counted. In some runs of the instruction sequence
the clustering optimization does not take effect (perhaps
because of a page fault?), so the count of retired conditional branches varies.
5.2
Nondeterminisic Instructions
Most CPU instructions are deterministic. As discussed
in Section 2.6, the nondeterministic x86(-64) instructions
can currently be handled or (in practice) ignored, but this
could change, so we need some improvements in this area.
Extending Linux (and hardware, where necessary)
to support trapping on executions of CPUID, as we
can for RDTSC, would let us mask off capability bits
so that well-behaved software does not use the newer
non-deterministic instructions. Trapping and emulating CPUID is also necessary to support trace portability
across machines with different CPU models and would let
us avoid pinning recording and replay to a specific core.
Peformance Counters
We would like to support record-and-replay of
programs using hardware transactional memory
As discussed in Section 2.4.1, RR uses hardware perfor(XBEGIN/XEND). This could be done efficiently if
mance counters to measure application progress. We need
the hardware and OS could be configured to raise a signal
a counter that increments very frequently so that the pair
on any failed transaction.
(counter, general purpose registers) uniquely identifies a
Trapping on all other nondetermnistic instructions (e.g.
program execution point in practice, and it must be “deRDRAND)
would be useful.
terministic” as described earlier. Performance counters
are typically used for statistical analysis, which does not
Porting RR to ARM failed because all ARM
need this determinism property, so unsurprisingly it does atomic memory operations use the “load-linked/storenot hold for most performance counter implementations conditional” approach, which is inherently nondetermin[37]. We are indeed fortunate that there is one satisfac- stic. The conditional store can fail because of nontory counter in Intel’s modern Core models. The ideal user-space-observable activity, e.g. hardware interrupts,
performance counter for our purposes would count the so counts of retired instructions or conditional branches
exact number of instructions retired as observed in user- for code performing atomic memory operations are nonspace (e.g., counting an instruction interrupted by a page determinstic. These operations are inlined into very many
fault just once); then the counter value alone would al- code locations, so it appears patching them is not feasible
ways uniquely identify an execution point.
except via pervasive code instrumentation. On x86(-64),
Being able to run RR in virtual machines is helpful atomic operations (e.g. compare-and-swap) are determinfor deployability. KVM, VMware and Xen support vir- istic in terms of user-space state, so there is no such probtualization of performance counters, and RR is known to lem.
10
5.3
Use Of Shared Memory
As noted in Section 2.5, RR does not support applications using shared memory written to by non-recorded
processes or kernel code — unless there’s some defined
protocol for preventing read-after-write races that RR can
observe, for example as is the case for Video4Linux.
Fortunately, popular desktop Linux frameworks (X11,
PulseAudio) can be automatically configured per-client to
minimize usage of shared memory to a few easily-handled
edge cases.
via deterministic reexecution [19]. There have been multiple independent efforts to add some amount of recordand-replay support to QEMU [14, 15, 35]. Whole-system
record-and-replay can be useful, but for many users it is
inconvenient to have to hoist the application into a virtual
machine. Many applications of record-and-replay also require cheap checkpointing, and checkpointing an entire
system is generally more expensive than checkpointing
one or a few processes.
6.2
5.4
Operating System Features
RR performance depends on leveraging relatively modern
Linux features. The seccomp-bpf API, or something
like it, is essential to selectively trapping system calls.
The Linux PERF COUNT SW CONTEXT SWITCHES
performance event is essential for handling blocking system calls. The Linux copy-on-write file and block cloning
APIs reduce overhead for I/O intensive applications.
5.5
Operating System Design
Any efficient record-and-replay system depends on
clearly identifying a boundary within which code is replayed deterministically, and recording and replaying the
timing and contents of all inputs into that boundary. In RR,
that boundary is mostly the interface between the kernel
and user-space. This works well with Linux: most of the
Linux user/kernel interface is stable across OS versions,
relatively simple, and well-documented, and it’s easy to
count hardware performance events occurring within the
boundary (i.e. all user-space events for a specific process).
This is less true in other operating systems. For example,
in Windows, the user/kernel interface is not publicly documented, and is apparently more complex and less stable
than in Linux. Implementing and maintaining the RR approach for an operating system like Windows would be
considerably more challenging than for Linux, at least for
anyone other than the OS vendor.
6
6.1
Related Work
Whole-System Replay
ReVirt [16] was an early project that recorded and
replayed the execution of an entire virtual machine.
VMware [27] used the same approach to support recordand-replay debugging in VMware Workstation, for a time,
but discontinued the feature. The full-system simulator
Simics has long supported reverse-execution debugging
Replaying User-Space With Kernel
Support
Scribe [25], dOS [6] and Arnold [13] replay a process or group of processes by extending the OS kernel
with record-and-replay functionality. Requiring kernel
changes makes maintenance and deployment more difficult — unless record-and-replay is integrated into the base
OS. But adding invasive new features to the kernel has
risks, so if record-and-replay can be well implemented
outside the kernel, moving it into the kernel may not be
desirable.
6.3
Pure User-Space Replay
Approaches based on user-space interception of system
calls have existed since at least MEC [10], and later
Jockey [33] and liblog [20]. Those systems did not handle
asynchronous event timing. Also, techniques relying on
library injection, such as liblog and Jockey, do not handle
complex scenarios such as ASAN [34] or Wine [3] with
their own preloading requirements. PinPlay [31], iDNA
[7], UndoDB [1] and TotalView ReplayEngine [21] use
code instrumentation to record and replay asynchronous
event timing. Unlike UndoDB and RR, PinPlay and iDNA
instrument all loads, thus supporting parallel recording in
the presence of data races and avoiding having to compute the effects of system calls, but this gives them higher
overhead than the other systems.
As far as we know RR is the only pure user-space
record-and-replay system to date that eschews code instrumentation in favour of hardware performance counters for asynchronous event timing, and it has the lowest
overhead of the pure user-space systems.
6.4
Higher-Level Replay
Record-and-replay features have been integrated into
language-level virtual machines. DejaVu [11] added
record-and-replay capabilities to the Jalapeño Java VM.
Microsoft IntelliTrace [2] instruments CLR bytecode to
11
record high-level events and the parameters and results
of function calls; it does not produce a full replay. Systems such as Chronon [12] for Java instrument bytecode
to collect enough data to provide the appearance of replaying execution for debugging purposes, without actually doing a replay. Dolos [9] provides record-and-replay
for JS applications in Webkit by recording and replaying the nondeterministic inputs to the browser. R2 [22]
provides record-and-replay by instrumenting library interfaces with assistance from the application developer. Such
systems are all significantly narrower in scope than the
ability to replay general user-space execution.
6.5
Parallel Replay
Recording application threads running concurrently on
multiple cores, with the possibility of data races, with
low overhead, is extremely challenging. PinPlay [31]
and iDNA [7] instrument shared-memory loads and report high overhead. SMP-ReVirt [17] tracks page ownership using hardware page protection and reports high
overhead on benchmarks with a lot of sharing. DoublePlay [36] runs two instances of the application and thus
has high overhead when the application alone could saturate available cores. ODR [5] has low recording overhead
but replay can be extremely expensive and is not guaranteed to reproduce the same program states.
The best hope for general low-overhead parallel recording seems to be hardware support. Projects such as FDR
[38], BugNet [29], Rerun [23], DeLorean [28] and QuickRec [32] have explored low-overhead parallel recording
hardware.
— or, if they share memory, techniques inspired by SMPReVirt [17] or dthreads [26] may perform adequately for
some workloads.
The applications of record-and-replay are perhaps more
interesting and important than the base technology. For
example, one can perform high-overhead dynamic analysis during replay [13, 14, 31], potentially parallelized
over multiple segments of the execution. With RR’s noinstrumentation approach, one could collect performance
data such as sampled stacks and performance counter values during recording, and correlate that data with rich
analysis generated during replay (e.g. cache simulation).
Always-on record-and-replay would make finding and fixing bugs in the field much easier.
Supporting recording and replay of parallel sharedmemory applications with potential races, while keeping
overhead low, will be very difficult without hardware support. Demonstrating compelling applications for recordand-replay will build the case for including such support
in commodity hardware.
8
Conclusions
The current state of Linux on commodity x86 CPUs enables single-core user-space record-and-replay with low
overhead, without pervasive code instrumentation — but
only just. This is fortuitous; we use software and hardware features for purposes they were not designed to
serve. It is also a recent development; five years ago
seccomp-bpf and the Linux file cloning APIs did not
exist, and commodity architectures with a deterministic hardware performance counter usable from user-space
had only just appeared (Intel Westmere)2 . By identifying
the utility of these features for record-and-replay, we hope
7 Future Work
that they will be supported by an increasingly broad range
RR perturbs execution, especially by forcing all threads of future systems. By providing an open-source, easy-toonto a single core, and therefore can fail to reproduce deploy, production-ready record-and-replay framework
bugs that manifest outside RR. We have addressed this we hope to enable more compelling applications of this
problem by introducing a “chaos mode” that intelligently technology.
adds randomness to scheduling decisions, enabling us to
reproduce many more bugs, but that work is beyond the
scope of this paper. There are many more opportunities to References
enhance the recorder to find more bugs.
[1] Reversible debugging tools for C/C++ on Linux &
Putting record-and-replay support in the kernel has perAndroid. http://undo-software.com. Acformance benefits, e.g. reducing the cost of recording concessed: 2016-04-16.
text switches. We may be able to find reusable primitives
that can be added to kernels to improve the performance [2] Understanding IntelliTrace part I: What the @#$%
of user-space record-and-replay while being less invasive
is IntelliTrace?
https://blogs.msdn.
than a full kernel implementation.
2 Performance counters have been usable for kernel-implemented reRecording multiple processes running in parallel on play [16, 30] for longer, because kernel code can observe and compenmultiple cores seems feasible if they do not share memory sate for events such as interrupts and page faults.
12
microsoft.com/zainnab/2013/02/12/
[13] D. Devecsery, M. Chow, X. Dou, J. Flinn, and P. M.
understanding-intellitrace-part-iChen. Eidetic systems. In Proceedings of the 11th
what-the-is-intellitrace.
Accessed:
USENIX Symposium on Operating Systems Design
2016-04-16.
and Implementation, October 2014.
[3] Wine windows-on-posix framework. https:// [14] B. Dolan-Gavitt, J. Hodosh, P. Hulin, T. Leek, and
R. Whelan. Repeatable reverse engineering with
www.winehq.org. Accessed: 2016-09-20.
PANDA. In Proceedings of the 5th Program Protec[4] O. Agesen, J. Mattson, R. Rugina, and J. Sheldon.
tion and Reverse Engineering Workshop, December
Software techniques for avoiding hardware virtual2015.
ization exits. In Proceedings of the 2012 USENIX
[15] P. Dovgalyuk. Deterministic replay of systems exAnnual Technical Conference, June 2012.
ecution with multi-target QEMU simulator for dy[5] G. Altekar and I. Stoica. ODR: Output-deterministic
namic analysis and reverse debugging. 2012.
replay for multicore debugging. In Proceedings of
the ACM SIGOPS 22nd Symposium on Operating [16] G. Dunlap, S. King, S. Cinar, M. A. Basrai, and P. M.
Chen. ReVirt: Enabling intrusion analysis through
Systems Principles, October 2009.
virtual-machine logging and replay. In Proceedings
of the 5th USENIX Symposium on Operating Sys[6] T. Bergan, N. Hunt, L. Ceze, and S. D. Gribble. Detems Design and Implementation, December 2002.
terministic process groups in dOS. In Proceedings of
the 9th USENIX Symposium on Operating Systems
[17] G. W. Dunlap, D. G. Lucchetti, M. A. Fetterman,
Design and Implementation, October 2010.
and P. M. Chen. Execution replay of multiprocessor virtual machines. In Proceedings of the 4th ACM
[7] S. Bhansali, W.-K. Chen, S. de Jong, A. Edwards,
SIGPLAN/SIGOPS International Conference on VirR. Murray, M. Drinić, D. Mihočka, and J. Chau.
tual Execution Environments, March 2008.
Framework for instruction-level tracing and analysis of program executions. In Proceedings of the
[18] J. Engblom. A review of reverse debugging. In Sys2nd International Conference on Virtual Execution
tem, Software, SoC and Silicon Debug Conference,
Environments, June 2006.
September 2012.
[8] D. Bruening, Q. Zhao, and S. Amarasinghe. Trans- [19] J. Engblom, D. Aarno, and B. Werner. Full-system
parent dynamic instrumentation. In Proceedings of
simulation from embedded to high-performance systhe 8th International Conference on Virtual Executems. In Processor and System-on-Chip Simulation,
tion Environments, March 2012.
2010.
[9] B. Burg, R. Bailey, A. Ko, and M. Ernst. Interac- [20] D. Geels, G. Altekar, S. Shenker, and I. Stoica.
tive record/replay for web application debugging. In
Replay debugging for distributed applications. In
Proceedings of the 26th ACM Symposium on User
Proceedings of the 2006 USENIX Annual Technical
Interface Software and Technology, October 2013.
Conference, June 2006.
[10] M. E. Chastain. https://lwn.net/1999/ [21] C. Gottbrath. Reverse debugging with the TotalView
0121/a/mec.html, January 1999. Accessed:
debugger. In Cray User Group Conference, May
2016-04-16.
2008.
[11] J.-D. Choi, B. Alpern, T. Ngo, M. Sridharan, and [22] Z. Guo, X. Wang, J. Tang, X. Liu, Z. Xu, M. Wu,
J. Vlissides. A perturbation-free replay platform for
M. F. Kaashoek, and Z. Zhang. R2: An applicationcross-optimized multithreaded applications. In Prolevel kernel for record and replay. In Proceedings of
ceedings of the 15th International Parallel and Disthe 8th USENIX Symposium on Operating Systems
tributed Processing Symposium, April 2001.
Design and Implementation, December 2008.
[12] P. Deva.
http://chrononsystems.com/ [23] D. Hower and M. Hill. Rerun: Exploiting episodes
blog/design-and-architecture-offor lightweight memory race recording. In Proceedthe-chronon-record-0, December 2010.
ings of the 35th Annual International Symposium on
Accessed: 2016-04-16.
Computer Architecture, June 2008.
13
[24] U. Hlzle, C. Chambers, and D. Ungar. Optimizing
dynamically-typed object-oriented languages with
polymorphic inline caches. In Proceedings of the
1991 European Conference on Object-Oriented Pro[33]
gramming, July 1991.
record and replay of multithreaded programs. In
Proceedings of the 40th Annual International Symposium on Computer Architecture, June 2013.
Y. Saito. Jockey: A user-space library for recordreplay debugging. In Proceedings of the 6th International Symposium on Automated Analysis-driven
Debugging, September 2005.
[25] O. Laadan, N. Viennot, and J. Nieh. Transparent,
lightweight application execution replay on commodity multiprocessor operating systems. In Proceedings of the ACM International Conference on [34] K. Serebryany, D. Bruening, A. Potapenko, and
D. Vyukov. Addresssanitizer: A fast address sanity
Measurement and Modeling of Computer Systems,
checker. In Proceedings of the 2012 USENIX Annual
June 2010.
Technical Conference, June 2012.
[26] T. Liu, C. Curtsinger, and E. Berger. Dthreads: Ef[35] D. Srinivasan and X. Jiang. Time-traveling forensic
ficient deterministic multithreading. In Proceedings
analysis of VM-based high-interaction honeypots.
of the ACM SIGOPS 23rd Symposium on Operating
In Security and Privacy in Communication NetSystems Principles, October 2011.
works: 7th International ICST Conference, September 2011.
[27] V. Malyugin, J. Sheldon, G. Venkitachalam,
B. Weissman, and M. Xu. ReTrace: Collecting ex[36] K. Veeraraghavan, D. Lee, B. Wester, J. Ouyang,
ecution trace with virtual machine deterministic reP. M. Chen, J. Flinn, and S. Narayanasamy. Doubleplay. In Proceedings of the Workshop on Modeling,
play: Parallelizing sequential logging and replay. In
Benchmarking and Simulation, June 2007.
Proceedings of the 16th International Conference on
Architectural Support for Programming Languages
[28] P. Montesinos, L. Ceze, and J. Torrellas. DeLorean:
and Operating Systems, March 2011.
Recording and deterministically replaying sharedmemory multiprocessor execution efficiently. In [37] V. Weaver, D. Terpstra, and S. Moore. NonProceedings of the 35th Annual International Symdeterminism and overcount on modern hardware
posium on Computer Architecture, June 2008.
performance counter implementations. In Proceedings of the IEEE International Symposium on Per[29] S. Narayanasamy, G. Pokam, and B. Calder. Bugnet:
formance Analysis of Systems and Software, April
Continuously recording program execution for de2013.
terministic replay debugging. In Proceedings of the
32nd Annual International Symposium on Computer [38] M. Xu, R. Bodik, and M. D. Hill. A “flight data
Architecture, June 2005.
recorder” for enabling full-system multiprocessor
deterministic replay. In Proceedings of the 30th An[30] M. Olszewski, J. Ansel, and S. Amarasinghe.
nual International Symposium on Computer ArchiKendo: Efficient deterministic multithreading in
tecture, June 2003.
software. In Proceedings of the 14th International
Conference on Architectural Support for Programming Languages and Operating Systems, March
2009.
[31] H. Patil, C. Pereira, M. Stallcup, G. Lueck, and
J. Cownie. PinPlay: A framework for deterministic replay and reproducible analysis of parallel programs. In Proceedings of the 8th Annual IEEE/ACM
International Symposium on Code Generation and
Optimization, April 2010.
[32] G. Pokam, K. Danne, C. Pereira, R. Kassa,
T. Kranich, S. Hu, J. Gottschlich, N. Honarmand,
N. Dautenhahn, S. King, and J. Torrellas. QuickRec: Prototyping an intel architecture extension for
14
| 6cs.PL
|
CLASSIFICATION OF LOCAL FIELD POTENTIALS USING GAUSSIAN SEQUENCE
MODEL
Taposh Banerjee? John Choi† Bijan Pesaran† Demba Ba? and Vahid Tarokh?
arXiv:1710.01821v4 [stat.ME] 27 Nov 2017
?
School of Engineering and Applied Sciences, Harvard University
†
Center for Neural Science, New York University
ABSTRACT
A problem of classification of local field potentials (LFPs), recorded
from the prefrontal cortex of a macaque monkey, is considered. An
adult macaque monkey is trained to perform a memory-based saccade. The objective is to decode the eye movement goals from the
LFP collected during a memory period. The LFP classification problem is modeled as that of classification of smooth functions embedded in Gaussian noise. It is then argued that using minimax function
estimators as features would lead to consistent LFP classifiers. The
theory of Gaussian sequence models allows us to represent minimax
estimators as finite dimensional objects. The LFP classifier resulting from this mathematical endeavor is a spectrum based technique,
where Fourier series coefficients of the LFP data, followed by appropriate shrinkage and thresholding, are used as features in a linear
discriminant classifier. The classifier is then applied to the LFP data
to achieve high decoding accuracy. The function classification approach taken in the paper also provides a systematic justification for
using Fourier series, with shrinkage and thresholding, as features for
the problem, as opposed to using the power spectrum. It also suggests that phase information is crucial to the decision making.
Index Terms— Brain-machine interface (BMI), brain signal
processing, local field potentials, function classification, minimax
estimators, Gaussian sequence model, Pinsker’s theorem, blockwise
James-Stein estimator, dimensionality reduction, linear discriminant
analysis.
1. INTRODUCTION
One of the most important challenges in the development of effective
brain-computer interfaces is the decoding of brain signals. Although
we are still very far from obtaining a complete understanding of the
brain, yet important contributions have been made. For example,
results are reported in the literature where local field potential (LFP)
data or spike data are collected from animal species, while the animal
is performing a task. The LFP or spike data are then used to make
various inferences regarding the task. See for example [1] and [2].
Spectrum-based techniques are popular in the literature. Fourier
series, wavelet transform or power spectrum of the recorded signals
are often used as features in the machine learning algorithms used for
inference. However, the choice for such spectrum based techniques
is either ad-hoc or is not well motivated.
In this paper, we propose a mathematical framework for classification of LFP signals into one of many classes or categories. The
framework is of Gaussian sequence model for nonparametric estimation; see [3] and [4]. The LFP data is collected from the preThe work at both Harvard and NYU is supported by the Army Research
Office MURI under Contract Number W911NF-16-1-0368.
frontal cortex of a macaque monkey, while the monkey is performing a memory-based task. In this task the monkey saccades to one of
eight possible target locations. The goal is to predict the target location using the LFP signal. More details about the task are provided
in Section 2 below.
The LFP signal is modeled as a smooth function corrupted by
Gaussian noise. The smooth signal is part of the LFP relevant to
the classification problem. The classification problem is then formulated as the problem of classifying the smooth functions into one
of finitely many classes (Section 3). We provide mathematical arguments to justify that using minimax function estimator for the
smooth signal as a feature leads to a consistent classifier (Section 4).
The theory of Gaussian sequence models allows us to obtain minimax estimators and also to represent them as finite dimensional vectors (Section 5). We propose two classification algorithms based on
the Gaussian sequence model framework (Section 6). The classification algorithms involve computing the Fourier series coefficients
of the LFP data and then using them as features after appropriate
shrinking and thresholding. The classifiers are then applied to the
LFP data collected from the monkey and provide high decoding accuracy of up to 88% (Section 7).
The Gaussian sequence model approach used in this paper leads
to the following insights:
1. A systematic modeling of the classification problem leads to
Fourier series coefficients, with shrinkage and thresholding,
to be used as features for the problem (as opposed to the use
of power spectrum for example).
2. Numerical results in Section 7 show that using the absolute
value of the Fourier series coefficients as features results in
a drop of 15% in accuracy. This suggests that phase information is crucial to the classification problem studied in this
paper.
2. A MEMORY-BASED SACCADE EXPERIMENT
An adult macaque monkey is trained to perform a memory-based
saccade. A single trial starts with the monkey looking at the center
of a square target board; see Fig. 1. The center is illuminated with an
LED light, and the monkey is trained to fixate at the center light at
the beginning of the trial. After a small delay, a target light at one of
eight corners of the target board (four vertices and four side centers)
is switched on. The location of the target is randomly chosen across
trials. After another short delay, the target light is switched off. The
switching off of the target light marks the beginning of a memory
period. After some time the center light is switched off. The latter
provides a cue to the monkey to saccade to the location of the target
light, the location where the target was switched on before switching off. The cue also marks the end of the memory period. If the
2
4
7
1
6
0
5
3
Fig. 2: Sample LFP waveform and its reconstruction using the first
10 Fourier series coefficients.
Fig. 1: Memory task and target locations from [2].
monkey correctly saccades to the target chosen for the trial, i.e., if
the monkey saccades to within a small area around the target, then
the trial is declared a success. Only data collected from successful
trials is utilized in this paper.
LFP and spike data are collected from electrodes embedded in
the cortex of the monkey throughout the experiment. The data is
collected using a 32 electrode array resulting in 32 parallel streams
of LFP data. The data used in this paper has been first collected
and analyzed in [2]. Further details about sampling rates, timing
information and the post-processing done on the data, can be found
in [2].
It is hypothesized that different target locations would correspond to different LFP and spike rate patterns. The objective is thus
to detect or predict one of eight target locations based on the LFP
data after the target is switched off and before the cue, i.e., the LFP
data collected during the memory period.
3. A NONPARAMETRIC REGRESSION FRAMEWORK
FOR LFP DATA
model this impertinent data as noise, and treat f (t) as the signal.
We assume that f ∈ F, where F is a class of smooth functions
in L2 [0, 1]. Precise assumptions about F will be made later; see
Section 6 below. In Fig. 2 we show a sample LFP waveform and its
reconstruction using first 10 Fourier coefficients. The figure shows
that a nonparametric regression framework to model the LFP data is
appropriate.
We hypothesize that the signal f will be different for all the eight
different target locations. In fact, even if the monkey saccades to the
same location over different trials, the signal f cannot be expected
to be exactly the same. Thus, we assume that for target location k,
k ∈ {1, 2, · · · , 8}, f ∈ Fk , where Fk is a subset of F, Fk ∩Fj = ∅,
for j 6= k. Thus, our classification problem can be stated as
Hk :
f ∈ Fk ,
k ∈ {1, 2, · · · , 8},
Fk ⊂ F , Fk ∩ Fj = ∅, for j 6= k.
(2)
Our objective is to find a classifier that maps the observed time series
{Y` } into one of the eight possible target locations.
4. CLASSIFICATION USING MINIMAX FUNCTION
ESTIMATORS
Let fˆN be an estimator for function f based on (Y0 , · · · , YN −1 )
such that its maximum estimation error over the function class F
goes to zero:
sup Ef [kfˆN − f k22 ] → 0, as N → ∞.
(3)
f ∈F
Let Y` , for ` ∈ {0, 1, · · · , N − 1}, represent the sampled version
of the LFP waveform collected from one of the 32 channels during
the memory period. We model the LFP data in a non-parametric
regression framework:
Y` = f (`/N ) + Z` ,
` ∈ {0, · · · , N − 1}.
(1)
Here, f : [0, 1] → R is a smooth function, and represent the part
of the LFP waveform that is relevant to the classification problem.
Thus, f (t) is the signal that contains information about where the
monkey will saccade, out of eight possible target locations, after the
i.i.d.
cue. Also, Z` ∼ N (0, 1), and correspond to neural firing data
either not pertinent to the memory experiment itself, or corresponds
to the experiment, but not relevant to the classification task. We
Here Ef is the expectation with respect to the law of
(Y0 , · · · , YN −1 ) when f is the true function, and kf − gk2 =
R1
(f (t) − g(t))2 dt is the L2 function norm.
0
Let δmd be the minimum distance decoder:
δmd (f ) = arg min kf − Fk k2 := arg min inf kf − gk2 .
k≤8
k≤8
g∈Fk
(4)
Of course, to implement this decoder we need to know the function
classes {Fk }8k=1 .
We now show that if the function classes {Fk }8k=1 are known,
then δmd (fˆN ), the minimum distance decoder using fˆN , is consistent. In the rest of this section, we use δ̂ to denote δmd (fˆN ) for
brevity.
Let Pe be the worst case classification error defined by
Pe =
max
5. GAUSSIAN SEQUENCE MODEL
sup Pf (δ̂ 6= k).
k∈{1,··· ,8} f ∈Fk
Also, let for s > 0, the classes are separated by a distance of at least
2s, i.e.,
min kFj − Fk k2 := min
inf
k6=j f ∈Fj , g∈Fk
k6=j
kf − gk2 > 2s.
(5)
Then, for f ∈ Fk and δ̂ = δmd (fˆN ),
Pf (δ̂ 6= k) ≤ Pf (kfˆN − Fk k2 > s)
≤ Pf (kfˆN − f k2 > s)
1
≤ 2 Ef [kfˆN − f k22 ]
s
1
≤ 2 sup Eg [kfˆN − gk22 ] → 0,
s g∈F
5.1. Nonparametric Estimation
as N → ∞,
(6)
where the limit to zero follows from (3). In fact, since the right most
term of (6) is not a function of k and f , we have
Pe =
max
sup Pf (δ̂ 6= k) ≤
k∈{1,··· ,8} f ∈Fk
1
sup Eg [kfˆN − gk22 ] → 0,
s2 g∈F
In this section, we provide a brief review of nonparametric or function estimation (Section 5.1) and its connection with the Gaussian
sequence model (Section 5.2). We also state the Pinsker’s theorem,
that is the fundamental result on minimax estimation on compact ellipsoids (Section 5.3). Finally, we discuss the blockwise James-Stein
estimator, that has adaptive minimaxity properties (Section 5.5). In
Section 6, we discuss the applications of these results to classification in neural data.
as N → ∞.
(7)
Thus, the worst case classification error goes to zero if we use
a function estimator with property (3). If we choose the minimax
∗
∗
satisfies
, i.e., fˆN
estimator fˆN
Consider the problem of estimating a function f : [0, 1] → R from
its noisy observation Y : [0, 1] → R, where Y (t) now does not
necessarily represent an LFP signal, but is a more abstract object.
The precise observation model(s) will be provided below. Let fˆ :
[0, 1] → R be our estimate. If the function and its estimate are in
R1
L2 ([0, 1]), then one can use 0 (f (t) − fˆ(t))2 dt as an estimate of
the error. This error estimate is random due to the randomness of the
noise. Thus, a more appropriate error criterion is
Z 1
MSE(fˆ) = Ef
(f (t) − fˆ(Y ; t))2 dt ,
(9)
0
where the dependence of estimate fˆ on the observation Y is made
explicit, and also we assume that the expectation is well defined for
the noise model in question.
Two popular models for the observation Y are as follows.
White Noise Model: Here the observation Y is given by the
stochastic differential equation [5]
Z t
Y (t) =
f (s) ds + Wt , t ∈ [0, 1],
(10)
0
where W is a standard Brownian motion, and > 0.
∗
− f k22 ], as N → ∞, (8)
inf sup Ef [kfˆ − f k22 ] ∼ sup Ef [kfˆN
fˆ f ∈F
f ∈F
then we will have the fastest rate of convergence to zero (fastest
among all classifiers of this type) for the misclassification error in
(7). Note further that the above consistency result is valid for any
choice of function classes {Fk }, as long as the classes are well separated, as in (5). This fact demonstrates a robust nature of the mini∗
max estimator based classifier δmd (fˆN
)
In practice, the function classes {Fk } are not known, and one
has to learn them from samples of data. In machine learning, a feature is often extracted from the observations, and the feature is then
used to train a classifier. The above discussion on minimax estimators suggests that if the class boundaries can be reliably learned,
then a classifier based on the minimax estimator and the minimum
distance decoder is consistent and robust. This motivates the use of
the minimax estimator itself as a feature. That is, given the data, use
the minimax estimators to learn the class boundaries.
A typical function estimator is an infinite-dimensional object,
and hence it is a complicated object to choose as a feature. However, the theory of Gaussian sequence models, to be discussed next,
allows us to map the minimax estimator to a finite-dimensional vector (under certain assumptions on F). We use the finite-dimensional
representation of the minimax estimator to learn the class boundaries
and train a classifier.
Regression Model: Here the observation Y is a discrete stochastic process corresponding to samples of the function f at discrete
points,
Yk = f (k/N ) + Zk ,
k ∈ {0, · · · , N − 1},
(11)
i.i.d.
where Zk ∼ N (0, 1), and N is the number of observations. Although we only use the regression model for classification of LFP in
this paper, the discussion of white noise model is important as it provided an exact map to a Gaussian sequence model (see Section 5.2
below for precise statements).
In the theory of nonparametric estimation [4], minimax optimality of estimators like kernel estimators, local polynomial estimators,
projection estimators, etc. is investigated. It is shown that if the function f is smooth enough, then carefully designed versions of these
type of estimators are asymptotitcally optimal in a minimax sense,
with N → ∞ in the regression model (11), and → 0 in the white
noise model (10).
5.2. Gaussian Sequence Model
One of the most remarkable observations in nonparametric estimation theory is that the estimation in the above two models, the white
noise model (10) and the regression model (11), are fundamentally equivalent. Moreover, these two are also equivalent to many
other popular estimation problems, including density estimation and
power spectrum estimation. The common thread that ties them together is the Gaussian sequence model [3].
A Gaussian sequence model is described by an observation sequence {yk }:
yk = θk + zk , k ∈ N.
(12)
One of the most important results in Gaussian sequence models
is the famous Pinsker’s theorem. Let Θ be an ellipsoid in `2 (N) such
that
(
)
X 2 2
Θ = Θ(a, C) = θ = (θ1 , θ2 , · · · ) :
ak θk ≤ C ,
The sequence {θk } is the unknown to be estimated using the observations {yk } corrupted by white Gaussian noise sequence {zk }, i.e.,
i.i.d.
zk ∼ N (0, 1).
To see the connection between the Gaussian sequence model and
say the white noise model, consider an orthonormal basis {φk } (e.g.,
Fourier Series or wavelet series) in L2 ([0, 1]), and define
with {ak } positive, nondecreasing sequence and ak → ∞. Consider the following diagonal linear shrinkage estimator of θ, θˆ∗ =
(θˆ1∗ , θˆ2∗ , · · · ):
ak
θ̂k∗ = 1 −
yk .
(17)
µ +
1
Z
yk =
1
Z
φk (t) dY (t) =
0
1
Z
φk (t) f (t) dt +
0
φk (t) dW (t).
0
(13)
where the integrals with respect to Y and W are stochastic integrals
[5]. Define
Z 1
θk =
φk (t) f (t) dt
0
and
1
Z
zk =
φk (t) dW (t)
0
to get the Gaussian sequence model (12).
Let y = (y1 , y2 , · · · ) be the observation sequence, and θ̂(y) be
an estimator of θ = (θ1 , θ2 , · · · ). If we are interested in squared
error loss as in (9), and θ̂ is in `2 (N), then
MSE(fˆ) = Ef
Z
1
(f (t) − fˆ(Y ; t))2 dt
0
h
i
= Eθ kθ − θ̂k22 = MSE(θ̂),
1
Z
the observations yk by an amount 1 − aµk if aµk < 1, otherwise
resets the observation to zero. It is linear because the amount of
shrinkage used is not a function of the observations, and is defined
beforehand. It is called diagonal, because the estimate of θk depends
only on yk , and not on other yj , j 6= k.
Theorem 5.1 (Pinsker’s Theorem, [3]) For the Gaussian sequence model (12), the linear diagonal estimator θ̂∗ (17) is
consistent, i.e.,
h
i
sup Eθ kθ − θˆ∗ k22 → 0 as → 0.
θ∈Θ(a,C)
Also, θ̂∗ is optimal among all linear estimators for each . Moreover,
it is asymptotically minimax over the ellipsoid Θ(a, C), i.e.,
h
i
h
i
sup Eθ kθ − θˆ∗ k22 ∼ inf sup Eθ kθ − θ̂k22 as → 0,
θ̂ θ∈Θ(a,C)
provided condition (5.17) in [3] is satisfied.
We make the following remarks on the implications of the theorem.
φk (t)fˆ(t)dt.
0
Thus, function estimation in the white noise model is equivalent to
sequence estimation in Gaussian sequence model.
For the regression problem (11), mapping to a finite Gaussian
sequence model
yk = θk + zk ,
Here, µ > 0 is a constant that is a function of the noise level and
the ellipsoid parameters {ak } and C;
see [3]. This estimator shrinks
θ∈Θ(a,C)
(14)
provided θ̂ is the transform of fˆ, i.e.,
θ̂k =
k
k ∈ {0, 1, · · · , N − 1}.
(15)
1. The Pinsker’s theorem states that a linear diagonal estimator
of the form
θ̂k = ck yk , k ∈ N,
(18)
is as good as any estimator, non-diagonal or nonlinear.
2. Define a Sobolev class of functions
n
Σ(α, C) = f ∈ L2 [0, 1] : f (α−1) is absolutely
Z 1
(α)
2
2
continuous and
[f (t)] dt ≤ C .
0
can be similarly defined using discrete Fourier transform. Another
option is to consider discrete approximation to Fourier series. In
both the cases, we need to take N → ∞ to recover the exact infinite
Gaussian sequence model (12).
The other problems, e.g., density estimation and spectrum estimation, can also be mapped to Gaussian sequence model by appropriate mappings. See [3] and [4] for details.
5.3. Pinsker’s Theorem
The discussion in the above two sections motivates us to restrict attention to the Gaussian sequence model
yk = θk + z k ,
k ∈ N.
(16)
(19)
It can be shown that a function is in the Sobolev function class
Σ(α, π α C) if and only if its Fourier series coefficients {θk }
are in an ellipsoid Θ(a, C) with
a1 = 0,
a2k = a2k+1 = (2k)α .
(20)
Thus, finding the minimax function estimator over
Σ(α, π α C) is equivalent to finding the minimax estimator
over Θ(a, C) in the Gaussian sequence model. Thus, to
obtain the minimax function estimator, one can take the
Fourier series of the observation Y , shrink the coefficients as
in Pinsker’s theorem, and then reconstruct a signal using the
modified Fourier series coefficients. See Lemma 3.3 in [3]
and the discussion surrounding it for further details.
3. Since {ak } is an increasing sequence, the optimal estimator
(17) has only a finite number of non-zero components. This
is the key to the LFP classification problem we are interested
in this paper.
5.4. Discussion on Minimax Estimators: Shrinkage and Thresholding
The parameters θ need not be the Fourier coefficients of the function to be estimated. One can also take wavelet coefficients, and
the Pinsker’s theorem is valid for them as well. The shrinkage and
reconstruction procedure has to be appropriately adjusted.
The optimal estimator is linear only because the risk is maximized over ellipsoids. For other families of Θs, the resulting optimal
estimators can be nonlinear. The most significant alternatives are the
nonlinear thresholding based estimators. The basic idea is that we
compute the Fourier or wavelet coefficients and set the coefficient
below a threshold to zero. Thresholding based estimators are optimal when the elements of Θ are sparse. Threshold-based estimators
and its applications to LFP classification will be discussed in detail
in a future version of this paper.
5.5. Adaptive Minimaxity: Blockwise James-Stein Estimator
A major drawback of the Pinsker estimator (17) is that the optimal estimator depends on the ellipsoid parameters. In practice, the
smoothness parameter α, constant C, noise level , and hence the
parameter µ, are not known. Thus, one does not know how many
Fourier coefficients to compute, and the amount of shrinkage to use.
Motivated by our discussion in Section 4, if we use Pinsker’s estimator as a feature for learning a classifier, then one has to optimize
over these choices of parameters using some error estimation techniques like cross-validation (see Section 6). If the number of training
samples is small, such an optimization can lead to overfitting. We
now discuss another estimator that does not need the information on
ellipsoid parameters, and that is also adaptively minimax over any
possible choice of ellipsoid parameters.
Let y = (y1 , y2 , · · · , yn ) ∼ Nn (θ, 2 I), where θ =
(θ1 , · · · , θn ), be a multivariate diagonal Gaussian random vector.
The James-Stein estimator of θ, defined for n > 2, is given by
(n − 2) 2
θ̂JS (y) = 1 −
y.
(21)
kyk22
+
Note that this estimator is non-linear, as the amount of shrinkage
used depends on the norm of the observation. Also note that here
θ̂JS is the estimate of the entire vector
θ, and is a n dimensional
2
vector. The shrinkage 1 − (n−2)
is applied coordinate wise
2
kyk
2
+
to each observation in the vector. It is well known that the JamesStein estimator is uniformly better than the maximum likelihood estimator y for this problem.
Now consider the infinite Gaussian sequence model (12). A
blockwise James-Stein estimator for the Gaussian sequence model is
defined by dividing the infinite observation sequence {yk } in blocks,
and by applying James-Stein estimator (21) to each block. To define
things more precisely, we need some notations.
Consider the partition of positive integers
N = ∪∞
j=0 Bj ,
where Bj are disjoint dyadic blocks
Bj = {2j , · · · , 2j+1 − 1}.
Thus, size of block j is |Bj | = 2j . Now define the observation for
the jth block as
y (j) = {yi : i ∈ Bj }.
Also, fix integers L and J. The blockwise James-Stein (BJS) estimator is defined as (j is used as a block index)
(j)
if j ≤ L
y ,
BJS
θ̂j
= θ̂JS (y (j) ), if L < j ≤ J
(22)
0
if j ≥ J,
where, as in (21),
θ̂JS (y (j) ) =
1−
(2j − 2) 2
ky (j) k22
y (j) .
(23)
+
Thus, the BJS estimator leaves the first L blocks untouched, applies
James-Stein shrinkage to the next J − L blocks, and shrinks the rest
of the observations to zero. The integer J is typically chosen to be
of the order of log −2 . Also, for regression problems ∼ √1N ; see
(11). Thus, the only free parameter is L. See Section 6 for more
details.
Consider the dyadic Sobolev ellipsoid
X
X
Θα
θ = (θ1 , θ2 , · · · ) :
22jα
θ`2 ≤ C 2 ,
D (C) =
j≥0
`∈Bj
and let
TD,2 = {Θα
D (C) : α, C > 0}
be the class of all such dyadic ellipsoids. The following theorem
states the adaptive minimaxity of BJS estimator over TD,2 .
Theorem 5.2 ( [3]) For the Gaussian sequence model (12), with any
fixed L and J = log −2 , the BJS estimator satisfies for each Θ ∈
TD,2
h
i
h
i
sup Eθ kθ − θ̂BJS k22 ∼ inf sup Eθ kθ − θ̂k22
as → 0.
θ̂ θ∈Θ
θ∈Θ
(24)
Thus, the BJS estimator adapts to the rate of the minimax estimator
for each class in TD,2 without using the parameters of the classes in
its design.
6. CLASSIFICATION ALGORITHMS BASED ON
GAUSSIAN SEQUENCE MODEL
In this section, we propose classification procedures based on the
ideas discussed in previous sections. We propose two classifiers:
the Pinsker’s classifier (Section 6.1), and the Blockwise James-Stein
Classifier (Section 6.2). Numerical results are reported in the next
section.
Recall from Section 3 that the sampled LFP waveform was modeled by a regression model
Y` = f (`/N ) + Z` ,
` ∈ {0, · · · , N − 1},
(25)
i.i.d.
where Z` ∼ N (0, 1), and f ∈ F, where F is a class of smooth
functions. Also, recall our classification problem
Hk :
f ∈ Fk ,
k ∈ {1, 2, · · · , 8},
(26)
Fk ⊂ F , Fk ∩ Fj = ∅, for j 6= k.
6.1. Pinsker Classifier
We assume that the function class is a Sololev class defined in (19)
F = Σ(α, C),
with α and C unknown. The classification subclasses {Fk } are to
be learned from multiple samples of LFP waveforms. As was motivated in Section 4, we use minimax estimator as a feature. Pinsker’s
theorem allows us to choose a finite dimensional representation for
the minimax estimator: use the Fourier series estimates
yk =
N −1
1 X
Y` φk (`/N ),
N
k ∈ {1, 2, · · · , T },
(27)
`=0
where {φk } are the trigonometric series
φ1 (x) ≡ 1
√
φ2k (x) = 2 cos(2πkx)
√
φ2k+1 (x) = 2 sin(2πkx),
4. Dimensionality Reduction (optional): If LFP data is collected
using multiple channels, then use principal component analysis to project the FS coefficients from all the channels (vectorized into a single high-dimensional vector) into a low dimensional subspace.
5. Train LDA: Train an LDA classifier.
6. Cross-validation: Estimate the generalization error using
cross-validation.
7. Optimize free parameters: Optimize over the choice of L and
c1 , · · · , cT fixed in step 1.
6.2. Blockwise James-Stein Classifier
The BJS classifier can be similarly defined.
(28)
k = {1, 2, · · · },
and T is a design parameter. Further, let {ck } be a sequence of
shrinkage values to be selected with ck ∈ [0, 1]. Then our representation for the minimax estimator is
θ̂k = ck yk ,
3. Use Shrinkage: Shrink the FS coefficients by {ck } to get feature vector (c1 y1 , c2 y2 , · · · , cT yT ).
k = {1, 2, · · · , T }.
Note that we need not select T and {ck } as in Pinsker’s theorem
because we do not have precise knowledge of F = Σ(α, L). Moreover, since the classes {Fk } and how they are separated are also
not known, it is not clear if the shrinkage mandated by the Pinsker’s
theorem is optimal for classification. We thus search over all possible shrinkage values (allowing for a low pass as well as a band
pass filtering) in order to design a classifier. Thus, given a sampled LFP waveform (Y0 , Y1 , · · · , YN −1 ) we map it to the vector
(c1 y1 , c2 y2 , · · · , cT yT ), where {ck }Tk=1 and T are now design parameters.
Once T and {ck }Tk=1 has been chosen, the vector
(c1 y1 , c2 y2 , · · · , cT yT ) is multivariate normal random vector.
Since a multivariate normal population is completely specified by
its mean and covariance, one can use the training data to learn
means and covariances of samples from different classes, and train a
classifier using these learned means and covariances. For example,
if we make the following assumptions:
1. We assume that the function classes {Fk } are
such that the means of the T -length random vector
(c1 y1 , c2 y2 , · · · , cT yT ) above is approximately the same for
all functions within each class and differ significantly across
classes.
2. We
also
assume
that
the
covariance
of
(c1 y1 , c2 y2 , · · · , cT yT ) is approximately the same for
all functions in all the classes.
Then it is clear that using linear discriminant analysis (LDA) for
classification would be optimal. For the LFP data we study in this
paper, LDA indeed provides the best classification performance.
The Pinsker classifier is summarized below:
1. Fix parameters: Fix T and c1 , · · · , cT .
2. Compute FS coefficients: Use the LFP time-series to compute
T Fourier series coefficients yk as in (27).
1. Fix parameters: Fix L√to say L = 2. Fix J = log −2 =
log N . Thus, = 1/ N . This relationship between and
N comes from an equivalence map between the white noise
model (10) and the regression model (11) (see [3] and [4] for
more details).
2. Compute FS coefficients: Use the LFP time-series to compute
N Fourier series coefficients yk as in (27). Thus, the number
of FS coefficients computed is equal to the sample size N .
3. Use Shrinkage: Use BJS shrinkage to get the feature vector
θ̂BJS
(j)
if j ≤ 2
y ,
BJS
θ̂j
= θ̂JS (y (j) ), if 2 < j ≤ log N
(29)
0
if j ≥ log N,
where, as in (21),
θ̂JS (y (j) ) =
(2j − 2)
y (j) .
1 − (j) 2
ky k2 N +
(30)
4. Dimensionality Reduction (optional): If LFP data is collected
using multiple channels, then use principal component analysis to project the FS coefficients from all the channels (vectorized into a single high-dimensional vector) into a low dimensional subspace.
5. Train LDA: Train an LDA classifier.
6. Cross-validation: Estimate the generalization error using
cross-validation.
Note that in this classifier no optimization step is needed.
7. CLASSIFICATION ALGORITHMS APPLIED TO LFP
DATA
We collected 736 samples of LFP data from 32 channels (each channel corresponds to a different electrode) across 9 recording sessions.
All the samples had the same depth profile. This means that the
32 electrodes were at different depths. But, those 32-length vectors
were the same for all the 736 samples. Each sample is a data matrix
of 32 rows (corresponding to 32 channels) and N columns. The integer N is smaller than the length of the memory period, and was a
design parameter for us.
For Pinsker classifier (see Section 6.1), we mapped each row (N
length data from each channel) to a T length Fourier series coefficients vector, and used shrinkage on the Fourier coefficients by ck
to get θ̂k . The 32 sets of Fourier coefficients were then appended to
form a single vector of length 32 × (2T + 1) (T sines and T cosines,
and one mean coefficient). A linear discriminant analysis was performed on the data after projecting the data onto P principal modes.
All the free parameters were optimized using leave-one-session-out
cross-validation. The optimal choice of parameters was found to be:
N = 500,
labels from 0 to 7 correspond to eight different target locations, also
shown in Fig. 4. Both the classifiers have better decoding accuracy
for the targets on the left. This phenomenon was also reported in [2].
T = 5,
ck = 1, k ≤ 5, and ck = 0, otherwise, and P = 165.
(31)
Thus, for the LFP data we analyzed, the best classification performance, of 88%, was obtained by low pass filtering the signal. In general, these choices of the free parameters would depend on the data,
and may not always correspond to low pass filtering. In Fig. 3, we
have plotted a sample LFP waveform and its reconstructions using 5
and 10 Fourier coefficients. Clearly, using 10 Fourier coefficients is
good for estimation, but the optimal performance for classification
was found using 5 coefficients.
2
4
6
1
0
7
3
5
Fig. 4: Classification accuracy (conditional probability of correct
decoding) per target. The average accuracy is 88% for the Pinsker
classifier and is 85% for the BJS classifier.
Also shown in the figure is the performance of another classifier, where the absolute values of the Fourier series coefficients were
used in the Pinsker’s classifier. This causes a drop in performance to
71%. The decoding accuracy of the targets on the right also drops by
a significant margin. This suggests that phase information is crucial
to this decoding task, at least for the targets on the right. We note that
power spectrum based techniques, popular in the literature, use absolute values of the Fourier coefficients to compute power spectrum
estimates.
Fig. 3: Sample LFP waveform and its reconstruction using first 5
and 10 Fourier coefficients. Estimation is better with 10 coefficients,
but classification performance was better with 5 coefficients.
For the BJS classifier (see Section 6.2), the process was similar
to the Pinsker classifier, except the shrinkage was different and fixed.
The parameter N was fixed to 500, and optimal P was found to be
P = 190. The BJS classifier achieved a performance of 85% but
had few free parameters. Due to the adaptive minimaxity properties
of the BJS estimator and few free parameters, we can expect robust
performance, and hence better generalization performance, of the
BJS classifier.
In Fig. 4, we have plotted target wise decoding performances
of the Pinsker classifier and the BJS classifier. The horizontal axis
8. CONCLUSIONS AND FUTURE WORK
We proposed a new framework for robust classification of local field
potentials and used it to obtain Pinsker’s and blockwise James-Stein
classifiers. These classifiers were then used to obtain high decoding
performance on the LFP data. For future work, we will consider the
following directions:
1. Testing with more data: The data chosen for the above numerical results were for a particular electrode depths and were
collected from a single monkey. Also, the data we used were
a sub-sampled version of the broadband data collected in the
experiment. We plan to test the performance of the proposed
algorithm on different data sets. We expect that the optimal
choice of {ck } and T , the number of Fourier coefficients to
choose and the amount of shrinkage to use, could be different.
2. Apply similar techniques to Spike Rate data: A spike rate data
can also be modeled as a nonparametric function, provided
the window for calculating the rate is sufficiently large, and
the window is slid by a small amount. We plan to test classification performance based on spike data. Alternative ways to
accommodate spike data for decoding will also be explored.
3. Apply Wavelet thresholding techniques: If the function
classes are not modeled using an ellipsoid, then the minimax
estimator for the class might have a different structure than
suggested by Pinsker’s theorem. For example, if the function class F has a sparse representation in an orthogonal basis (like wavelets), then under certain conditions, a nonlinear
thresholding based estimator is minimax. In another article,
we will study wavelet transform based classifiers that are robust over a larger class of functions than the Sobolev classes.
9. REFERENCES
[1] R. P. Rao, Brain-computer interfacing: an introduction. Cambridge University Press, 2013.
[2] D. A. Markowitz, Y. T. Wong, C. M. Gray, and B. Pesaran, “Optimizing the decoding of movement goals from local field potentials in macaque cortex,” Journal of Neuroscience, vol. 31,
no. 50, pp. 18412–18422, 2011.
[3] I. M. Johnstone, Gaussian estimation: Sequence and wavelet
models. Book Draft, 2017. Available for download from
http://statweb.stanford.edu/˜imj/GE_08_09_
17.pdf.
[4] A. B. Tsybakov, Introduction to nonparametric estimation.
Springer Series in Statistics. Springer, New York, 2009.
[5] L. C. G. Rogers and D. Williams, Diffusions, Markov processes
and martingales: Volumes 1 and 2. Cambridge university press,
2000.
| 7cs.IT
|
arXiv:1510.01290v2 [math.ST] 2 May 2016
TOTAL POSITIVITY IN MARKOV STRUCTURES∗
By Shaun Fallat† , Steffen Lauritzen‡ Kayvan Sadeghi§
Caroline Uhlerk Nanny Wermuth†† and Piotr Zwiernik‡‡
University of Regina† , University of Copenhagen‡ , University of
Cambridge§ , Massachusetts Institute of Technologyk , IST Austriak ,
Chalmers University of Technology†† , Johannes Gutenberg-University†† ,
and Universitat Pompeu Fabra‡‡
We discuss properties of distributions that are multivariate totally positive of order two (MTP2 ) related to conditional independence. In particular, we show that any independence model generated by an MTP2 distribution is a compositional semigraphoid which
is upward-stable and singleton-transitive. In addition, we prove that
any MTP2 distribution satisfying an appropriate support condition is
faithful to its concentration graph. Finally, we analyze factorization
properties of MTP2 distributions and discuss ways of constructing
MTP2 distributions; in particular we give conditions on the log-linear
parameters of a discrete distribution which ensure MTP2 and characterize conditional Gaussian distributions which satisfy MTP2 .
1. Introduction. This paper discusses a special form of positive dependence. Positive dependence may refer to two random variables that have
a positive covariance, but other definitions of positive dependence have
been proposed as well; see [24] for an overview. Random variables X =
(X1 , . . . , Xd ) are said to be associated if cov{f (X), g(X)} ≥ 0 for any
two non-decreasing functions f and g for which E|f (X)|, E|g(X)|, and
E|f (X)g(X)| all exist [13]. This notion has important applications in probability theory and statistical physics; see, for example, [28, 29].
However, association may be difficult to verify in a specific context. The
celebrated FKG theorem, formulated by Fortuin, Kasteleyn, and Ginibre in
[14], introduces an alternative notion and establishes that X are associated if
∗
SF was supported in part by an NSERC Discovery Research Grant, KS by grant
#FA9550-12-1-0392 from the U.S. Air Force Office of Scientific Research (AFOSR) and
the Defense Advanced Research Projects Agency (DARPA), CU by the Austrian Science
Fund (FWF) Y 903-N35, and PZ by the European Union Seventh Framework Programme
PIOF-GA-2011-300975.
MSC 2010 subject classifications: Primary 60E15, 62H99; secondary 15B48
Keywords and phrases: Association, concentration graph, conditional Gaussian distribution, faithfulness, graphical models, log-linear interactions, Markov property, positive
dependence
1
2
S. FALLAT ET AL.
their joint densityQfunction is multivariate totally positive of order 2 : A function f over X = v∈V Xv , where each Xv is totally ordered, is multivariate
totally positive of order two (MTP2 ) if
f (x)f (y)
≤
f (x ∧ y)f (x ∨ y)
for all x, y ∈ X ,
where x ∧ y and x ∨ y denote the element-wise minimum and maximum.
These inequalities are often easier to check. Furthermore, most other
known definitions of positive dependence are implied by the MTP2 constraints; see for example [7] for a recent overview. Note that the conditions
are on the probabilities or the density and not on other types of traditional
measures of dependence. But as we shall see, the above inequality constraints
combined with conditional independence restrictions specify positive associations along edges in undirected graphs, named and studied as dependence
graphs or concentration graphs; see for instance [22, 44].
MTP2 distributions have also played an important role in the study of
ferromagnetic Ising models, i.e. distributions of binary variables where all
interaction potentials are pairwise and non-negative. It has been noted in [34]
that the block Gibbs sampler is monotonic if the target distribution is MTP2 ,
and hence particularly efficient in this setting. Bartolucci and Besag [3,
Section 5] showed that much of this work can in fact be extended to arbitrary
binary Markov fields. See also [11] for an optimization viewpoint. The special
case of Gaussian distributions was studied by Karlin and Rinott [21] and very
recently by Slawski and Hein [39] from a machine learning perspective.
Consequences of MTP2 distributions with respect to marginal and mutual
independences were studied by Lebowitz [23] and Newman [29]. They showed
in particular that independence of two components of a random vector with
an MTP2 distribution, is equivalent to a block-diagonal structure in the covariance matrix and that mutual independence of several components can
be inferred from a block-diagonal covariance matrix (see also Theorem 3.7
and Theorem 5.5 below). This is remarkable because covariances and correlations are the weakest types of measures of dependence, see [48]; although
they identify independence in Gaussian distributions, this is often not the
case for other types of distribution.
In this paper, we discuss implications of the MTP2 constraints for conditional independence and vice versa. There is some related work in the
context of copulae [27]. Our paper can be seen as a continuation of work by
Sarkar [38] and, in particular, by Karlin and Rinott [20, 21]. They noted that
the family of MTP2 distributions is stable with respect to forming marginal
and conditional distributions. At least as important is that they give constraints on different types of measures of dependence needed to verify the
TOTAL POSITIVITY IN MARKOV STRUCTURES
3
MTP2 property of a joint distribution for discrete and for Gaussian random
variables.
The MTP2 property may appear extremely restrictive when higher order
interactions are needed to capture the relevant types of conditional dependence or when distributions are studied which do not satisfy any conditional
independence constraints. However, as we shall see, the MTP2 constraints
become less restrictive when imposing an additional Markov structure. For
example, all finite dimensional distributions of a Markov chain are MTP2
whenever all 2×2 minors of its transition matrix are non-negative [20, Proposition 3.10]. This result holds true also for non-homogeneous Markov chains.
Moreover, models with latent, that is hidden or unobserved, variables may
be MTP2 . For example, factor analysis models with a single factor are MTP2
when each observed variable has an, albeit unobserved, positive dependence
on the single, hidden factor [46]. Similar statements apply to binary latent
class models [2, 16, 46] and to latent tree models, both in the Gaussian and
in the binary setting [40, 50]. Furthermore, many data sets can be well explained or modelled assuming that the generating distribution is MTP2 or
nearly MTP2 ; see Section 4 for some examples and also the discussion in
Section 8.
The paper is organized as follows: In Section 2 we introduce our notation
and provide the main definitions. In Section 3 we review basic properties of
MTP2 distributions and discuss the link to positive dependence and independence structures. In Section 4 we concentrate on the MTP2 condition in
the Gaussian and binary setting and provide several data examples where
the MTP2 property appears naturally. Section 5 analyzes MTP2 distributions with respect to conditional independence relations. One of the main
results in this paper is Theorem 5.3, which shows that any independence
model generated by an MTP2 distribution is a singleton-transitive compositional semigraphoid which is also upward-stable; the latter means that new
arbitrary elements can be added to the conditioning set of every existing independence statement without violating independence. Theorem 5.5 gives a
complete characterization of the marginal independence structures of MTP2
distributions. In Section 6, we study Markov properties of MTP2 distributions and show that such distributions are always faithful to their concentration graph. In Section 7, we analyze factorization properties of MTP2
distributions, show how to use these properties to build MTP2 distributions
from smaller MTP2 distributions, briefly discuss log-linear expansions of
discrete MTP2 distributions, and give conditions for conditional Gaussian
distributions to satisfy the MTP2 constraints. We conclude our paper with
a short discussion in Section 8.
4
S. FALLAT ET AL.
2. Preliminaries and notation. Let V be a finite set and let X =
(X
Q v , v ∈ V ) be a random vector. We consider the product space X =
v∈V Xv , where Xv ⊆ R is the state space of Xv , inheriting the order from
R. In this paper, the state spaces are either discrete (finite sets) or open
intervals on the real line. Hence, we can partition the set of variables as
V = ∆ ∪ Γ, where Xv is discrete if v ∈ ∆, and Xv is an open interval if
v ∈ Γ. For A ⊆ V we further write XA = (Xv )v∈A , XA = ×v∈A Xv and so
on.
All distributions are assumed to have densities with respect to the product
measure µ = ⊗v∈V µv , where µv is the counting measure for v ∈ ∆, and µv
is the Lebesgue measure giving length 1 to the unit interval for v ∈ Γ. We
shall refer to µ as the standard base measure. Similarly to the above, we
write µA = ⊗v∈A µv .
Finally, we introduce some definitions related to graphs: An undirected
graph G = (V, E) (sometimes also called concentration graph in the literature of graphical models) consists of a nonempty set of vertices or nodes V
and a set of undirected edges E. Our graphs are simple meaning that they
have no self-loops and no multiple edges. We write uv for an edge between
u and v and say that the vertices u and v are adjacent. A path in G is a
sequence of nodes (v0 , v1 , . . . , vk ) such that vi vi+1 ∈ E for all i = 0, . . . , k − 1
and no node is repeated, i.e., vi 6= vj for all i, j ∈ {0, 1, . . . , k} with i 6= j.
Thus an edge is the shortest type of path. A cycle is a path with the modification that v0 = vk . Furthermore, we say that two distinct nodes u, v ∈ V
are connected if there is a path between u and v; a graph is connected if
all pairs of distinct nodes are connected. A graph is complete if all possible edges are present. In addition, two subsets A, B ⊆ V are separated by
S ⊂ V \ (A ∪ B) if every path between A and B passes through a node in S.
A subgraph of G, induced by a set A ⊂ V , consists of the nodes in A and of
the edges in G between nodes in A. Finally, a maximal complete subgraph
is a clique.
3. Basic properties and positive dependence. We start this section by formally introducing MTP2 distributions and discuss basic properties
of these. We define the coordinate-wise minimum and maximum as
x ∧ y = (min(xv , yv ), v ∈ V ),
x ∨ y = (max(xv , yv ), v ∈ V ).
A function f on X is said to be multivariate totally positive of order two
(MTP2 ) if
(3.1)
f (x)f (y)
≤
f (x ∧ y)f (x ∨ y)
for all x, y ∈ X .
TOTAL POSITIVITY IN MARKOV STRUCTURES
5
For |V | = 2, a function that is MTP2 is simply called totally positive of order
two (TP2 ) [20]. Let X = (Xv , v ∈ V ) have density function f with respect
to the standard base measure µ. Then we say that X or the distribution
of X is MTP2 if its density function f is MTP2 . Note that this concept is
well-defined since Xv is either discrete or an open interval on the real line.
A basic property of MTP2 distributions is that it is preserved under increasing coordinatewise transformations. We begin with a simple result for
strictly increasing functions.
Proposition 3.1. Let X be a random vector taking values in X . Let φ =
(φv , v ∈ V ) be such that φv : R → R are strictly increasing and differentiable
for all v ∈ V . If the distribution of X on X is MTP2 , then the distribution
of Y = φ(X) is MTP2 .
Proof. We use the following fact from [20, Equation (1.13)]: Let av :
R → R be positive and let bv : R → R be non-decreasing. If f : RV → R is
MTP2 , then the function
Y
(3.2)
g(y) = f {bv (yv ), v ∈ V }
av (yv ),
v∈V
′
−1
′
is MTP2 . Let bv (yv ) = φ−1
v (yv ) and let av (yv ) = 1/φv (φv (yv )), where φv (yv )
denotes the first derivative of φv . Then g(y) is the density of Y = φ(X) and
we obtain from (3.2) that Y is MTP2 .
We say that a function φv (x) : Xv → R is piecewise constant if φv (Xv ) is
finite and we can then similarly show that the MTP2 property is preserved
under transformations which are piecewise constant and non-decreasing.
Proposition 3.2. Let X be a random vector taking values in X as
before. For A ⊆ V , let φ = (φv , v ∈ V ) be such that φv : Xv → R is piecewise
constant and non-decreasing for all v ∈ A and φv (xv ) = xv for v 6∈ A. If the
distribution of X is MTP2 , then the distribution of Y = φ(X) is MTP2 .
Proof. If f denotes the density function for X and g the density function for Y , both w.r.t. a standard base measure, we have that
Z
f (x) dµA (xA ).
g(y) =
φ−1
A (yA )
Since φ is non-decreasing, we have
2
−1 1
1
−1 2
φ−1
A (yA ) ∧ φA (yA ) = φA (yA ∧ yA ),
2
−1 1
1
−1 2
φ−1
A (yA ) ∨ φA (yA ) = φA (yA ∨ yA ),
6
S. FALLAT ET AL.
where for two sets A, B,
A ∧ B = {a ∧ b | a ∈ A, b ∈ B},
A ∨ B = {a ∨ b | a ∈ A, b ∈ B}.
Hence, we can apply [20, Corollary 2.1] to obtain that Y is MTP2 .
Corollary 3.3. Let X be a random vector taking values in X as before
and φ = (φv , v ∈ V ) be such that φv : Xv → R is piecewise constant and nondecreasing for all v ∈ A and φv (xv ) is strictly increasing and differentiable
for v 6∈ A. If the distribution of X is MTP2 , then the distribution of Y =
φ(X) is MTP2 .
Proof. Just combine Proposition 3.1 and Proposition 3.2 by first transforming to Zv = Xv for v ∈ A and Zv = φv (Xv ) for v 6∈ A; then subsequently
letting Yv = φv (Zv ) for v ∈ A and Yv = Zv for v 6∈ A.
The following result establishes that the MTP2 property is preserved under conditioning, marginalization, and monotone coarsening. A monotone
coarsening is an operation on a finite discrete state space Xi that identifies
a collection of neighbouring (in the given total order) states. For example,
if Xi = {i1 , . . . , ip } then Xi′ = {{i1 , . . . , ij }, ij+1 , . . . , ik , {ik+1 , . . . , ip }} is a
monotone coarsening.
Proposition 3.4. The MTP2 property is closed under conditioning,
marginalization, and monotone coarsening. More precisely,
(i) If X has an MTP2 distribution, then for every C ⊆ V , the conditional
distribution of XC | XV \C = xV \C is MTP2 for almost all xV \C ;
(ii) If X has an MTP2 distribution, then for every A ⊆ V , the marginal
distribution XA of X is MTP2 ;
(iii) If X is MTP2 and discrete, and Y is obtained from X by monotone
coarsening, then Y is MTP2 .
Proof. Property (i) follows directly from the definition of MTP2 . Property (ii) is shown in [20, Proposition 3.2]. Property (iii) is an instance of
a non-decreasing and piecewise constant transformation and follows from
Proposition 3.2.
As we will see in the following, the properties (i) and (ii) are the fundamental building blocks for understanding the implications of MTP2 on
Markov properties and vice versa. Property (iii) has direct relevance for applications. In the statistical literature it is often warned that dependence
TOTAL POSITIVITY IN MARKOV STRUCTURES
7
relations may get distorted when combining neighboring levels of discrete
variables, see for instance [36]. This may still be true for MTP2 distributions, see Example 6.2 below, but the coarsening property (iii) implies that
associations cannot become negative by such a process.
Another interesting fact about the MTP2 property is that, under suitable
support conditions, it is a pairwise property meaning that it can be checked
on the level of two variables only, when the remaining variables are fixed.
We say that f has interval support if for any x, y ∈ X the following holds
(3.3)
f (x)f (y) 6= 0
implies
f (z) 6= 0 for any x ∧ y ≤ z ≤ x ∨ y.
Note that having interval support is equivalent to having full support over
a restricted state space that is a product of intervals. In this setting, Karlin
and Rinott [20, Proposition 2.1] prove the following result.
Proposition 3.5. If f has interval support and f : X → R is TP2 in
every pair of arguments when the remaining arguments are held constant,
then f is MTP2 .
We conjecture that this result holds also under a weaker support condition, namely that the support is coordinate-wise connected [33], meaning
that the connected components of the support can be joined by axis-parallel
lines. We now provide such an instance in the binary 2 × 2 × 2 setting.
Example 3.6. Consider a binary 2×2×2 distribution, where the support
only misses the entries (1, 0, 0) and (1, 0, 1). In this example there are only
two non-trivial pairwise TP2 constraints to consider, namely for x1 = 0 and
for x2 = 1, i.e.
(3.4)
p000 p011 ≥ p010 p001
and
p010 p111 ≥ p110 p011
as the remaining six pairwise inequalities reduce to 0 ≥ 0.
So assume p satisfies (3.4). We need to show that the nine inequalities in
(4.1) below are satisfied. Again, six of them are trivial, and two of them are
exactly those in (3.4). The remaining inequality
p000 p111 ≥ p001 p110 ,
follows from multiplying the two inequalities in (3.4) to get
(p000 p011 )(p010 p111 ) ≥ (p010 p001 )(p110 p011 ),
and dividing both sides by p010 p011 . Hence, in this case pairwise TP2 constraints imply the MTP2 property even though the distribution does not
have interval support.
8
S. FALLAT ET AL.
As mentioned in Section 1, if X is MTP2 , then the variables in X are
associated, i.e.,
(3.5)
cov{f (X), g(X)} ≥ 0
for any coordinatewise non-decreasing functions f and g for which the covariance exists. For discrete distributions this follows by the FKG theorem [14], or, more generally, by the four functions theorem by Ahlswede
and Daykin [1]; the general case was proved by Sarkar [38]. The following
result, first proven by Lebowitz [23], shows that the independence structure
for associated vectors is encoded in the covariance matrix; see also [17, 29].
Theorem 3.7 (Corollary 3, [29]). If X are associated and E|Xv |2 < ∞
for all v ∈ V , then XA is independent of XB if and only if cov(Xu , Xv ) = 0
for all u ∈ A and v ∈ B.
In Section 5, we study conditional independence models for MTP2 distributions. Interestingly, we will show in Theorem 5.5 that for MTP2 distributions a stronger result holds, namely that every MTP2 random vector can be
decomposed into independent components such that within each component
all variables are mutually marginally dependent. This means in particular
that for MTP2 distributions, all marginal independence relations also hold
when conditioning on further variables; the general version of this property
will be termed upward-stability, see Section 5.
4. Examples of Gaussian and binary MTP2 distributions. Many
examples of MTP2 random variables are discussed in the literature; see, e.g.,
[20, 21]. In this section we focus on binary and multivariate Gaussian MTP2
distributions. Although the MTP2 property may appear restrictive, we want
to suggest that MTP2 distributions are important in practice and in fact
appear in real data sets.
4.1. Multivariate Gaussian MTP2 distributions. Consider a multivariate
Gaussian random vector X with mean µ and covariance matrix Σ. Denote
by K the inverse of Σ. Then, the distribution of X is MTP2 if and only if
K is an M-matrix; see [21], i.e.,
(i) kvv > 0 for all v ∈ V ,
(ii) kuv ≤ 0 for all u, v ∈ V with u 6= v.
Properties and consequences of M-matrices were studied by Ostrowski [30]
who chose the name to honour H. Minkowski who had considered aspects of
TOTAL POSITIVITY IN MARKOV STRUCTURES
9
such matrices earlier. The connection to multivariate Gaussian distributions
was established by Bølviken [5].
In the previous section we mentioned that if X is MTP2 , then its constituent variables are associated. Therefore, for MTP2 Gaussian distributions it holds that σuv ≥ 0 for all u, v ∈ V . More precisely, the covariance
matrix has a block diagonal structure and each block has only strictly positive elements; see also Theorem 5.5 below. Note, however, that this condition
is necessary but not sufficient for the MTP2 property.
We now analyze by simulation how restrictive the MTP2 constraint is for
Gaussian distributions. We quantify this by studying the ratio of the volume
of all correlation matrices that satisfy the MTP2 constraint to the volume
of all correlation matrices. Since no closed-form formula for these volumes
is known, we use a simple Monte Carlo simulation. We uniformly sample
correlation matrices using the method suggested by Joe [18], which is implemented in the R package clusterGeneration. We performed simulations
for |V | = 3, 4, 5 and we here report how many correlation matrices out of
100,000 samples satisfy the MTP2 constraint.
|V |
MTP2
3
5004
4
90
5
0
These simulation results show that the relative volume of MTP2 Gaussian
distributions drops dramatically with increasing |V | when no conditional
independences are taken into account. However, the picture changes when
imposing conditional independence relations. For example, if |V | = 3, then
by the above simulations about 5% of all Gaussian distributions correspond
to MTP2 distributions. If 1 ⊥⊥ 2 | 3, then by a symmetry argument precisely
25% of such distributions are MTP2 . If, in addition, we impose 1 ⊥⊥ 3 | 2 —
which implies also 1 ⊥⊥ (3, 2) — the ratio of MTP2 distributions increases to
50%. Finally, all distributions that are fully independent are MTP2 .
We next discuss a prominent data set consisting of the examination marks
of 88 students in five different mathematical subjects. The data were reported in [25] and analyzed, for example, in [12, 15, 47]. The inverse of the
sample covariance matrix, together with the corresponding partial correlations ρuv·V \{u,v} , are displayed in Table 1. This matrix is very close to being
an M-matrix with only one negative partial correlation equal to −0.00001.
Furthermore, when fitting reasonable graphical models to the data, all fitted
distributions are MTP2 .
4.2. Binary MTP2 distributions. Suppose that X is a binary random
vector with X = {0, 1}|V | and we denote its distribution by P = [px ] for
10
S. FALLAT ET AL.
Table 1
Empirical partial correlations (below the diagonal) and entries of the inverse of the
sample covariance matrix (×1000, on and above the diagonal) for the examination marks
in five mathematical subjects.
Mechanics
Vectors
Algebra
Analysis
Statistics
Mechanics
5.24
0.33
0.23
-0.00
0.02
Vectors
−2.44
10.43
0.28
0.08
0.02
Algebra
−2.74
−4.71
26.95
0.43
0.36
Analysis
0.01
−0.79
−7.05
9.88
0.25
Statistics
−0.14
−0.17
−4.70
−2.02
6.45
x ∈ X . For example, if |V | = 3, binary MTP2 distributions must satisfy the
following nine inequalities
p011 p000 ≥ p010 p001 p101 p000 ≥ p100 p001 p110 p000 ≥ p100 p010
(4.1)
p111 p100 ≥ p110 p101 p111 p010 ≥ p110 p011 p111 p001 ≥ p101 p011
p111 p000 ≥ p100 p011 p111 p000 ≥ p010 p101 p111 p000 ≥ p001 p110 .
The first two rows correspond to the inequalities px∧y px∨y ≥ px py as in (3.1),
where x and y differ only in two entries. These inequalities are equivalent to
requesting that the six possible conditional log-odds ratios are non-negative.
By Proposition 3.5, the inequalities in the last row are implied by the remaining ones in the case when P > 0, or more generally, if P has interval
support. For P > 0 this can be seen from identities of the form
p000
p101
p111 p000 − p010 p101 =
(p111 p001 − p101 p011 ) +
(p011 p000 − p010 p001 ).
p001
p001
For positive binary distributions we can verify MTP2 for any pair of variables with the remaining variables fixed. In the binary case this gives a single
constraint for any choiceof a pair and values for the remaining |V | − 2 variables, in other words |V2 | ·2|V |−2 inequalities. For binary MTP2 distributions
there is a nice description in terms of log-linear parameters in [4], see also
Corollary 7.7 below.
The MTP2 hypothesis is rather restrictive in the binary setting when no
further conditional independence restrictions are assumed. Note, however,
that binary models can become more complex than in the Gaussian case,
since log-linear interactions of higher-order than pairwise may be present.
In the following, we study the volume of MTP2 distributions with respect
to the volume of the whole probability simplex. Similarly as in the Gaussian
setting, we sample uniformly from the probability simplex. We here report
how many samples out of 100,000 satisfy the MTP2 constraints for |V | = 3, 4.
Note that already for |V | = 4 we did not find a single instance although the
volume of the set of MTP2 distributions is always positive:
TOTAL POSITIVITY IN MARKOV STRUCTURES
|V |
MTP2
3
2195
11
4
0
Like in the Gaussian case the relative volume of MTP2 distributions is
higher when imposing additional conditional independence restrictions. By
the same symmetry argument as in the Gaussian setting we obtain that
for |V | = 3 precisely 25% of all binary distributions satisfying 1 ⊥⊥ 2 | 3 are
MTP2 . If, in addition, we have 1 ⊥⊥ 3 | 2 then half of these distributions are
MTP2 . Finally, all binary full independence distributions are MTP2 .
This interplay with conditional independence might explain in part why
binary MTP2 distributions do arise in practice. See [46, Section 5] for examples of datasets that are MTP2 or nearly MTP2 . In the following, we discuss
two such examples.
Example 4.1. We first consider a dataset on EPH-gestosis, collected 40
years ago in a study on “Pregnancy and Child Development” by the German
Research Foundation and recently analyzed in [46, Section 5.1]. EPH-gestosis
represents a disease syndrome for pregnant women. The three symptoms are
edema (high body water retention), proteinuria (high amounts of urinary
proteins) and hypertension (elevated blood pressure). The observed counts
N = (nx ) are
n000 n010 n001 n011
3299 107 1012 58
=
.
n100 n110 n101 n111
78
11
65 19
If untreated, EPH-gestosis is a major cause of death of mother and child
during birth [42, p. 65]. However, treatment of the symptoms prevents negative consequences and the symptoms occur rarely after the first pregnancy.
The observed counts have odds-ratios larger than one for each pair at the
fixed level of the third variable, hence the empirical distribution is MTP2 .
Equivalently, the sample distribution satisfies all the constraints in (4.1). The
three symptoms do not occur more frequently jointly than in pairs and the
observed conditional odds-ratios are nearly equal given the third symptom.
Possible interpretations are that physicians intervened at the latest when
two symptoms occurred and that a single common cause, though unknown
and unobserved, may have generated the marginal dependences between the
symptom pairs.
Example 4.2. Next we discuss an example on five binary random variables. This is a subset of data from a Polish case-control study on laryngeal
cancer [49]. Details on the study design, our selection criteria for cases and
controls, and the analysis will be given elsewhere.
12
S. FALLAT ET AL.
In case-control studies the observations are implicitly obtained conditionally on the values of at least one response variable and on relevant explanatory variables. For such designs, the class of concentration graph models are
appropriate for studying dependence structure among the variables.
In this study, we have 185 laryngeal cancer cases in urban residential areas
(coded 1; 35.7%) and 308 controls, coded 0. Four further 0, 1 variables are
defined so that 1 indicates the level known to carry the higher cancer risk,
namely heavy vodka drinking (1:= regularly for 2 or more years; 21.3%),
heavy cigarette smoking (1:= 30 or more cigarettes per day; 13.8%, and 0:=
6 to 29 cigarettes per day), age at study entry (1:= 54 to 65 years; 51.5%
and 0:= 46 to 53 years), and level of formal education (1:= less than 8 years;
57.8 % and 0:=8 to 11 years).
A well-fitting log-linear model for these data is determined by the sufficient margins {{1, 2}, {1, 3}, {2, 3}, {1, 4, 5}}, in other words by permitting
log-linear interaction terms only among variable groups that are subsets of
these sets. This model yields an overall likelihood-ratio χ2 of 13.6 with 19
degrees of freedom and corresponds to a concentration graph with cliques:
{1, 2, 3} and {1, 4, 5}. The corresponding observed and fitted counts are:
00000
00100
00010
00110
00001
00101
00011
00111
10000
10100
10010
10110
10001
10101
10011
10111
01000
01100
01010
01110
01001
01101
01011
01111
11000
11100
11010
11110
11001
11101
11011
11111
:
85
10
46
7
51
4
73
5
11
1
15
2
27
6
36
9
5
1
3
2
7
1
5
3
6
2
7
5
18
4
30
6
85.88
9.27
47.59
5.32
,
51.70
5.78
70.57
7.89
9.87
1.70
14.31
2.47
27.13
4.68
39.96
6.89
7.30
1.55
4.19
0.89
4.55
0.97
6.22
1.32
6.35
2.08
9.21
3.02
17.46
5.73
25.72
8.43
.
For pairs within the cliques, the fitted two-way margins must coincide with
the observed bivariate tables of counts; here we report marginal observed
and fitted odds-ratios, or(I, J), and fitted conditional odds-ratios given the
remaining variables, or(I, J | R):
variable pair:
1,2
1,3
1,4
1,5
2,3
2,4
2,5
3,4
3,5
4,5
observed or(I, J):
fitted or(I, J):
7.6
1.9
1.7
3.0
2.3
1.4
1.3
2.0
1.6
1.3
1.1
0.9
1.2
2.0
fitted or(I, J | R):
7.3
1.5
2.5
4.4
1.9
1
1
1
1
∗)
∗)
2.4 for controls and 1.02 for cases.
Because the observed (3, 5) odds-ratio is smaller than 1 and hence the logodds-ratio is negative, the observed distribution is not MTP2 . In addition,
21 of the observed 80 conditional log-odds ratios, or(I, J | R), are less than
1. However in the well-fitting model described above we have or(I, J | R) ≥ 1
TOTAL POSITIVITY IN MARKOV STRUCTURES
13
for all 80 odds-ratios, so that the fitted distribution is MTP2 . This implies
that each possible marginal table — here of two, three, or four variables —
shows positive or vanishing pairwise dependence for all variable pairs.
From the concentration graph it follows that prediction of drinking and
smoking habits cannot be improved by using information about age or level
of formal education for the studied cases or controls and that there is no
log-linear interaction involving more than three factors. The set of minimal
sufficient tables tells that the only three-factor interaction is in the {1, 4, 5}table. From the above change in the conditional odds-ratio for pair (4, 5)
from 2.4 to 1.02, it follows that the expected improvement in education for
younger – compared to older participants – only shows for controls but not
for the cases. In combination with the fact that or(1, 5 | R) = 4.4, this implies
that level of formal education should be explicitly included in comparisons
of results across countries and in future studies on laryngeal cancer.
5. Conditional independence models and total positivity. An
independence model J over a finite set V is a set of triples hA, B | Ci (called
independence statements), where A, B, and C are disjoint subsets of V ; C
may be empty, and h∅, B | Ci and hA, ∅ | Ci are always included in J . The
independence statement hA, B | Ci is read as “A is independent of B given
C”. Independence models do not necessarily have a probabilistic interpretation; for a discussion on general independence models, see [41].
An independence model J over a set V is a semi-graphoid if it satisfies
the following four properties for disjoint subsets A, B, C, and D of V :
(S1) hA, B | Ci ∈ J if and only if hB, A | Ci ∈ J (symmetry);
(S2) if hA, B ∪ D | Ci ∈ J , then hA, B | Ci ∈ J and hA, D | Ci ∈ J (decomposition);
(S3) if hA, B ∪ D | Ci ∈ J , then hA, B | C ∪ Di ∈ J and hA, D | C ∪ Bi ∈ J
(weak union);
(S4) hA, B | C ∪ Di ∈ J and hA, D | Ci ∈ J if and only if hA, B ∪ D | Ci ∈ J
(contraction).
A semi-graphoid for which the reverse implication of the weak union property holds is said to be a graphoid that is, it also satisfies
(S5) if hA, B | C ∪ Di ∈ J and hA, D | C ∪ Bi ∈ J then hA, B ∪ D | Ci ∈ J
(intersection).
Furthermore, a graphoid or semi-graphoid for which the reverse implication
of the decomposition property holds is said to be compositional, that is it
also satisfies
14
S. FALLAT ET AL.
(S6) if hA, B | Ci ∈ J and hA, D | Ci ∈ J then hA, B ∪ D | Ci ∈ J (composition).
Some independence models have additional properties; below we write singleton sets {u}, {v} compactly as u, v, etc.
(S7) if hu, v | Ci ∈ J and hu, v | C∪wi ∈ J , then hu, w | Ci ∈ J or hv, w | Ci ∈
J (singleton-transitivity);
(S8) if hA, B | Ci ∈ J and D ⊆ V \ (A ∪ B), then hA, B | C ∪ Di ∈ J
(upward-stability).
The properties above are not independent and upward-stability is a very
strong property. For example, we have the following simple lemma:
Lemma 5.1.
tion.
Any upward stable semi-graphoid satisfies (S6) composi-
Proof. If hA, B | Ci ∈ J , (S8) yields hA, B | C ∪ Di ∈ J ; hence from
(S4) we get that hA, D | Ci ∈ J implies hA, B∪D | Ci ∈ J , which is (S6).
A fundamental example of an independence model is induced by separation in an undirected graph G = (V, E), denoted by J (G):
hA, B | Si ∈ J (G) ⇐⇒ S separates A from B
in the sense that all paths between A and B intersect S. The independence
model J (G) satisfies all of the above properties (S1)–(S8).
Consider a set V and associated random variables X = (Xv )v∈V . For
disjoint subsets A, B, and C of V we use the short notation A ⊥⊥ B | C to
denote that XA is conditionally independent of XB given XC [8, 22], i.e.
that for any measurable Ω ⊆ XA and P -almost all xB and xC ,
P (XA ∈ Ω | XB = xB , XC = xC ) = P (XA ∈ Ω | XC = xC ).
We can now induce an independence model J (P ) by letting
hA, B | Ci ∈ J (P ) if and only if A ⊥⊥ B | C w.r.t. P .
Probabilistic independence models are always semi-graphoids [32]. If, for
example, P has a strictly positive density f , the induced independence model
is always a graphoid; see e.g. Proposition 3.1 in [22]. More generally, if
f is continuous, Peters [33] showed that the induced independence model
is a graphoid if and only if the support is coordinate-wise connected, i.e.,
all connected components of the support of the density can be connected
TOTAL POSITIVITY IN MARKOV STRUCTURES
15
by axis-parallel lines. In particular, this applies to the discrete case since
any function over a discrete space is continuous. See also [9] for general
discussions and [26] for necessary and sufficient conditions under which the
intersection property holds in joint Gaussian or binary distributions.
Examples of discrete distributions violating one of (S5), (S6), or (S7)
have been given in [43]. We will prove in this section that, under weak
assumptions, independence models generated by MTP2 distributions satisfy
all of the properties (S1)-(S8).
First, note that by Proposition 3.4 the MTP2 property is closed under
marginalization and conditioning. Applying this to the conditional distribution of (Xu , Xv ) given XC for all u, v ∈ V with u 6= v, C ⊆ V \ {u, v},
implies that the following conditional covariances must be nonnegative
(5.1)
cov{φ(Xu ), ψ(Xv ) | XC } ≥ 0
a.s.
for non-decreasing functions φ, ψ : R → R for which the covariance exists.
Recall that a function of several variables is non-decreasing if it is nondecreasing in each coordinate. The following related result was first proved
in [38, Section 3.1] (see also [20, Theorem 4.1]).
Proposition 5.2. Let X be MTP2 . Then for any subset A ⊆ V and
any non-decreasing function ϕ : XA → R for which E|ϕ(XA )| < ∞, the
conditional expectation
E ϕ(XA ) | XV \A = xV \A
is non-decreasing in xV \A .
We first show that any induced independence model of an MTP2 distribution is an upward-stable and singleton-transitive compositional semigraphoid, i.e. (S1)–(S4) and (S6)–(S8) all hold.
Theorem 5.3. Any independence model J (P ) induced by an MTP2 distribution P is an upward-stable and singleton-transitive compositional semigraphoid.
Proof. We first note that any probabilistic independence model is a
semi-graphoid [31]. Next, we establish upward-stability. For this, it suffices
to prove that u ⊥⊥ v | C implies u ⊥⊥ v | C ∪ {w} for all w ∈ V \ (C ∪ {u, v}).
Since the MTP2 property is closed under marginalization, it follows that the
marginal distribution of XC∪{u,v,w} is MTP2 . Further, because the MTP2
property is closed under conditioning, after conditioning on C, it suffices to
16
S. FALLAT ET AL.
consider only 3 variables and prove the following statement: If the distribution of X = (X1 , X2 , X3 ) is MTP2 , then 1 ⊥⊥ 2 implies 1 ⊥⊥ 2 | 3.
Recall that 1 ⊥⊥ 2 if and only if P(X1 > a1 , X2 > a2 ) = P(X1 > a1 )P(X2 >
a2 ) for all a1 , a2 ∈ R. Equivalently, defining Yi = 11(Xi > ai ) this translates
into Y1 ⊥⊥ Y2 , which now must be satisfied for any choice of a1 , a2 ∈ R. A
similar condition holds for the conditional independence 1 ⊥⊥ 2 | 3, in which
case we require that Y1 ⊥⊥ Y2 | X3 for all a1 , a2 ∈ R.
The advantage of working with the indicator functions is that they define bounded random variables, which implies the existence of moments. By
Proposition 3.2 the vector (Y1 , Y2 , X3 ) is MTP2 . Independence of X1 and X2 ,
and the law of total covariance, implies that for every choice of a1 , a2 ∈ R
0 = cov(Y1 , Y2 )
= cov(E(Y1 | X3 ), E(Y2 | X3 )) + E(cov(Y1 , Y2 | X3 )).
By Proposition 5.2, E(Y1 | X3 = x3 ) and E(Y2 | X3 = x3 ) are almost everywhere non-decreasing and bounded functions of x3 and hence their covariance exists and is non-negative by (3.5) and the fact that univariate
random variables are always associated. Moreover, it follows from (5.1)
that cov(Y1 , Y2 | X3 = x3 ) ≥ 0 for almost all x3 , and thus its expectation is non-negative. This means that we expressed zero as a sum of two
non-negative terms and thus both terms must be zero. This implies that
cov(Y1 , Y2 | X3 = x3 ) = 0 for almost all x3 . Hence, by Theorem 3.7, we obtain that Y1 ⊥⊥ Y2 | X3 . Now by varying a1 , a2 we conclude that X1 ⊥⊥ X2 | X3 .
Having established upward-compatibility, composition now follows from
Lemma 5.1.
We finally prove singleton-transitivity. Using upward-stability this property can be rephrased in a simpler form as
1 ⊥⊥ 2
=⇒
1 ⊥⊥ 3
or
2 ⊥⊥ 3.
So we assume that 1 ⊥⊥ 2. By the same argument as in the previous paragraph, cov(Y1 , Y2 ) = 0 implies that cov(E(Y1 | X3 ), E(Y2 | X3 )) = 0. By
Theorem 5.2, both f (X3 ) := E(Y1 | X3 ) and g(X3 ) := E(Y2 | X3 ) are nondecreasing functions. Their covariance is zero and can be rewritten as
Z
f (x′ ) − f (x) g(x′ ) − g(x) d(µ3 (x) ⊗ µ3 (x′ )).
cov(f (X3 ), g(X3 )) =
{x′ >x}
Note that {x′ > x} ⊇ {f (x′ ) > f (x)} ∩ {g(x′ ) > g(x)} and so the integral
on the right-hand side, which is equal to zero, is bounded from below by
Z
f (x′ ) − f (x) g(x′ ) − g(x) d(µ3 (x) ⊗ µ3 (x′ )),
{f (x′ )>f (x)}∩{g(x′ )>g(x)}
TOTAL POSITIVITY IN MARKOV STRUCTURES
17
which is strictly positive unless the set {f (x′ ) > f (x)} ∩ {g(x′ ) > g(x)} has
measure zero. This set is the set of all pairs (x, x′ ) such that x′ > x and
f (x′ ) > f (x), g(x′ ) > g(x), and its measure is half the measure of the set
of all (x, x′ ) such that f (x′ ) 6= f (x) and g(x′ ) 6= g(x). This measure is zero
only if either f or g is constant almost everywhere.
It follows that for every a1 , a2 ∈ R either E(Y1 | X3 ) = P(X1 > a1 ) or
E(Y2 | X3 ) = P(X2 > a2 ) . Let U ⊆ R2 be the set of all (a1 , a2 ) such that
the former equality holds and V be the set such that the latter holds. Let πi
denote the projection on the i-th coordinate in R2 . We have U ∪ V = R2 and
so if π1 (U ) 6= R then π2 (V ) = R. This implies that π1 (U ) = R or π2 (V ) = R.
For simplicity assume that the latter holds but the general argument is
the same. If π2 (V ) = R then for every a2 ∈ R there exists a1 such that
E(Y2 | X3 ) = P(X2 > a2 ) , or equivalently P(X2 > a2 |X3 ) = P(X2 > a2 ). We
conclude that 2 ⊥⊥ 3. This, up to symmetry, implies that 1 ⊥⊥ 3 or 2 ⊥⊥ 3.
We now analyze the intersection property. It is important to note that an
MTP2 independence model is not necessarily a graphoid, as the following
simple example shows.
Example 5.4.
Consider the binary MTP2 distribution with
p000 = p111 =
1
.
2
Then 1 ⊥⊥ 2 | 3 and 1 ⊥⊥ 3 | 2, but 1 ⊥
6⊥ (2, 3), and therefore, the intersection
property does not hold.
As a consequence of the earlier mentioned result by Peters [33], any MTP2
distribution with continuous density and coordinate-wise connected support
is an upward-stable and singleton-transitive compositional graphoid.
We conclude this section with the following property of MTP2 distributions.
Theorem 5.5. Let the distribution of X be MTP2 with none of Xv
having a degenerate distribution. Then X can be decomposed into independent components such that within each component all variables are mutually
marginally dependent.
Proof. As in the previous proof we define Yi = 11(Xi > ai ) and by
Theorem 3.7 it suffices to prove that the covariance matrix of Y is block
diagonal with strictly positive entries in each block. We write u ∼ v if the
covariance between Yu and Yv is non-zero and we show that u ∼ v is an
18
S. FALLAT ET AL.
equivalence relation and thus induces a partition of V into independent
blocks. It is clear that u ∼ u and u ∼ v whenever v ∼ u. It remains to show
that u ∼ v and v ∼ w imply u ∼ w. But if u 6∼ w, we have σuw = 0 and
thus u ⊥⊥ w. Using upward-stability from Theorem 5.3 yields u ⊥⊥ w | v and
singleton-transitivity yields u ⊥⊥ v or v ⊥⊥ w, which contradicts that u ∼ v
and v ∼ w.
6. Faithfulness and total positivity. In the following we shall write
A ⊥G B | C for the graph separation hA, B | Ci ∈ J (G) and A ⊥⊥ B | C for
the relation hA, B | Ci ∈ J (P) in the independence model generated by P .
For a graph G = (V, E), an independence model J defined over V satisfies
the global Markov property w.r.t. a graph G, if for disjoint subsets A, B, and
C of V the following holds
A ⊥G B | C =⇒ hA, B | Ci ∈ J .
If J (P ) satisfies the global Markov property w.r.t. a graph G, we also say
that P is Markov w.r.t. G.
We say that an independence model J is probabilistic if there is a distribution P such that J = J (P ). We then also say that P is faithful to J . If
P is faithful to J (G) for a graph G, then we also say that P is faithful to
G. Thus, if P is faithful to G it is also Markov w.r.t. G.
In this section we examine the faithfulness property for MTP2 distributions. Let P denote a distribution on X . The pairwise independence graph
of P is the undirected graph G(P ) = (V, E(P )) with
uv ∈
/ E(P )
⇐⇒
u ⊥⊥ v | V \ {u, v}.
A distribution P is said to satisfy the pairwise Markov property w.r.t. an
undirected graph G = (V, E) if
uv ∈
/E
=⇒
u ⊥⊥ v | V \ {u, v}.
Thus any distribution P satisfies the pairwise Markov property w.r.t. its
pairwise independence graph G(P ); indeed, G(P ) is the smallest graph that
makes P pairwise Markov.
Generally, a distribution may be pairwise Markov w.r.t. a graph without
being globally Markov. However, if an MTP2 distribution P satisfies the
coordinate-wise connected support condition, in particular if it is strictly
positive, and since these are sufficient conditions for the intersection property
to hold, then pairwise and global Markov properties are equivalent; see [37].
We prove in the following result that if the pairwise-global equivalence is
already established, then P is in fact faithful to G(P ).
TOTAL POSITIVITY IN MARKOV STRUCTURES
19
Theorem 6.1. If P is MTP2 and its independence model is a graphoid,
then P is faithful to its pairwise independence graph G(P ).
Proof. By Theorem 3.7 in [22] it follows that P is globally Markov
w.r.t. G(P ).
To establish faithfulness, we consider disjoint subsets A, B, and C so that
C does not separate A from B in G(P ). We need to show that A ⊥
6⊥ B | C.
First, let uv ∈ E. Then u ⊥
6⊥ v | V \ {u, v} and hence by upward-stability
6⊥ v | C for any C ⊂ V \ {u, v}.
as shown in Theorem 5.3, u ⊥
Since C does not separate A from B, there exists u ∈ A and v ∈ B
and a path u = v1 , v2 , . . . , vr = v such that vk ∈
/ C for all k = 1, . . . , r, and
vk vk+1 ∈ E for all k = 1, . . . , r − 1. By the previous argument we obtain that
vk ⊥
6⊥ vk+1 | C for all k = 1, . . . , r − 1. By singleton-transitivity, v1 ⊥
6⊥ v2 | C and
v2 ⊥
6⊥ v3 | C imply that v1 ⊥
6⊥ v3 | C. Repeating this argument yields u ⊥
6⊥ v | C
and hence A ⊥
6⊥ B | C.
Note that this was also shown by Slawski and Hein [39] in the case of a
Gaussian MTP2 distribution. In fact in the Gaussian case it follows readily
since conditional covariances between any pair of variables can be obtained
through adding (non-negative) partial correlations along paths in G(P ) [19],
see also [45].
Notice that if X has coordinate-wise connected support, then Theorem 5.5
is a direct corollary of Theorem 6.1. This is because if a distribution is
faithful to an undirected graph, then the statement of Theorem 5.5 obviously
holds. However, the latter theorem is still interesting as it covers cases that
Theorem 6.1 does not; such as that of Example 5.4.
In Section 3 we postponed to show that the stability of the MTP2 property under coarsening as established in (iii) of Proposition 3.4 does not imply
that coarsening preserves conditional independence relations for MTP2 distributions. We demonstrate this in the following example.
Example 6.2. Consider the trivariate discrete distributions of (I, J, K)
where I and K are binary taking values in {0, 1} whereas J is ternary with
state space {0, 1, 2} given as follows
pijk = θi|j φk|j ψj ,
where ψj = 1/3 for all j, θ1|0 = φ1|0 = 1 − θ0|0 = 1 − φ0|0 = 1/4, θ1|1 =
φ1|1 = 1 − θ0|1 = 1 − φ0|1 = 1/3, and θ1|2 = φ1|2 = 1 − θ0|2 = 1 − φ0|2 =
1/2. This distribution is easily seen to be MTP2 which also follows from
Proposition 7.1 below. By construction, it also satisfies I ⊥⊥ K | J so that its
concentration graph has edges IJ and JK.
20
S. FALLAT ET AL.
Now define the binary variable L by monotone coarsening of J so that
L = 0 if J = 0 and L = 1 if J ∈ {1, 2}. Letting qilk denote the joint
distribution of (I, L, K) we get for example
q010 = p010 + p020 = (θ0|1 φ0|1 + θ0|2 φ0|2 )/3 = (4/9 + 1/4)/3 = 25/108,
and similarly
q011 = q110 = 17/108,
q111 = 13/108
so that the odds-ratio between I and J conditional on {L = 1} becomes
θ=
13 × 25
325
q010 q111
=
=
> 1.
q110 q011
172
289
Hence, after coarsening, the conditional association between I and J given
the third variable changes from absent to positive. Note that the MTP2
property ensures non-negativity of the distorted association.
Clearly, the distribution after coarsening remains faithful to its concentration graph, but coarsening changes the latter to become the complete
graph on V = {I, L, K}.
For completeness of Theorem 6.1 it is important to show that any concentration graph is realizable by an MTP2 distribution. We prove this fact
in the special case of Gaussian distributions.
Proposition 6.3. Any undirected graph G is realizable as the concentration graph G(P ) of some MTP2 Gaussian distribution.
Proof. Let A be the adjacency matrix of G, that is, Aij = 1 if and
only if (i, j) is an edge of G. Because G is undirected, A is symmetric.
Since the set of positive definite matrices is open and the identity matrix
is positive definite, it follows that K = I − ǫA is positive definite if ǫ > 0
is sufficiently small. By construction, K is an M-matrix and its non-zero
elements correspond to the edges of the graph G.
7. Special instances of total positivity. We conclude this paper
with a section on how to construct MTP2 distributions from a collection of
smaller MTP2 distributions, a brief discussion of conditions for the MTP2
property of discrete distributions in terms of log-linear interaction parameters, and characterizing conditional Gaussian distributions which are MTP2 .
TOTAL POSITIVITY IN MARKOV STRUCTURES
21
7.1. Singleton separators. Let A, B ⊂ V . We then say that two random
variables XA and XB with distributions PA and PB are consistent if the
distribution of XA∩B is the same under PA as under PB . Then one can
define a new distribution denoted by PA ⋆ PB and known as the Markov
combination of PA with density f and PB with density g (see [10]). Its
density is denoted by f ⋆ g and given by
(f ⋆ g)(xA∪B ) =
f (xA )g(xB )
.
h(xA∩B )
Here, h denotes the density of XA∩B , common to PA and PB . In the following, we show that the Markov combination of two MTP2 distributions is
again MTP2 as long as they are glued together over a 1-dimensional margin.
Proposition 7.1. Suppose that |A ∩ B| = 1. Then the Markov combination PA ⋆ PB of a consistent pair of distributions PA and PB is MTP2 if
and only if PA and PB are both MTP2 .
Proof. Since PA and PB are marginal distributions of PA ⋆ PB and the
MTP2 condition is preserved under marginalization, we only need to prove
one direction. Assume that PA and PB are MTP2 . The product of MTP2
functions is an MTP2 function; see, e.g. Proposition 3.3 in [20] for this basic
result. This implies that (f g)(x) = f (xA )g(xB ) is MTP2 . Now, if A ∩ B is
a singleton, then also f ⋆ g is MTP2 because multiplying by functions of a
single variable preserves the MTP2 property; c.f. (3.2).
For example, Proposition 7.1 implies that for the fitted model in Example 4.2 we only need to check the MTP2 condition in each of the two clique
marginals {1, 2, 3} and {1, 4, 5} to verify that the fitted distribution is MTP2 .
Since the fitted distribution is positive, this involves only 6 + 6 = 12 logodds-ratios, see the discussion of (4.1) in Section 4.2. In addition, as there
are only pairwise interactions in the {1, 2, 3}-marginal, conditional log-oddsratios for any pair of these variables are constant in the third variable and
hence we actually only need to check 3 + 6 = 9 such ratios to verify the
MTP2 property for the model fitted to the laryngeal cancer data; see also
Theorem 7.5 below.
Unfortunately, the conclusion in Proposition 7.1 does not hold in general
if |A ∩ B| > 1, as we show in the following example.
Example 7.2. Suppose that A = {1, 2, 3} and B = {2, 3, 4}, and let
X = (X1 , X2 , X3 , X4 ) ∈ {0, 1}4 . Consider the following distribution:
[p0000 , p0001 , p0010 , p0011 , p0100 , p0101 , p0110 , p0111 ] = [1, 2, 2, 20, 2, 20, 20, 400]/Z,
[p1000 , p1001 , p1010 , p1011 , p1100 , p1101 , p1110 , p1111 ] = [2, 4, 20, 200, 20, 200, 400, 8000]/Z,
22
S. FALLAT ET AL.
where the normalizing constant Z = 9313. It is easy to check that for every
i, j, k, l ∈ {0, 1} the following holds
pijkl =
pijk+p+jkl
.
p+jk+
Hence, the distribution P = [pijkl ] can be obtained as the Markov combination of two distributions, namely pijk+ over {1, 2, 3} and p+jkl over {2, 3, 4}.
One can also easily check that both these distributions are MTP2 . However,
since
p(1,1,0,1)∧(1,0,1,1) p(1,1,0,1)∨(1,0,1,1) − p(1,1,0,1) p(1,0,1,1) =
p1001 p1111 − p1101 p1011 = −
8000
,
93132
P is not MTP2 .
As a direct consequence of Proposition 7.1 we obtain the following result
for decomposable graphs, which are graphs where there is no cycle of length
more than three such that all its non-neighboring nodes (on the cycle) are
not adjacent (see e.g. [22] for a review).
Corollary 7.3. Let G be a decomposable graph such that the intersection of any two cliques is either empty or a singleton. Let P be a distribution
that is Markov w.r.t. G. Then P is MTP2 if and only if the marginal distribution over each clique is MTP2 .
Proof. The proof follows by induction over the number of cliques.
As we show in the following example, Corollary 7.3 cannot be extended
directly to non-decomposable graphs. It does not hold in general that a distribution is MTP2 if the margins over all cliques in the graph are MTP2 and
the cliques intersect in singletons only. However, as we show in Theorem 7.5
such a result does hold if all clique potentials are MTP2 functions.
Example 7.4.
P = [pijkl ] with
Consider the following 4-dimensional binary distribution
pijkl =
1
Aij Bjk Ckl Dil ,
Z
where
Z = 243,
6 5
A=
4 3
and
2 1
B=C=D=
.
1 2
TOTAL POSITIVITY IN MARKOV STRUCTURES
23
This distribution is Markov w.r.t. the 4-cycle. We now show that the marginal
distributions over each edge are MTP2 . For this, note that a binary 2dimensional random vector is MTP2 if and only if its covariance is nonnegative. To see this, observe that in the bivariate case there is only one
inequality p00 p11 − p01 p10 ≥ 0. Using the fact that p00 + p01 + p10 + p11 =
1, it is seen that this inequality is equivalent to p11 − p1+ p+1 ≥ 0; but
cov(X1 , X2 ) = p11 − p1+ p+1 .
In this example,
cov(X1 , X2 ) =
148
,
2432
cov(X2 , X3 ) =
4812
,
2432
4842
4632
, cov(X1 , X4 ) =
,
2
243
2432
and hence all edge-marginals are MTP2 . However, the full distribution P is
not MTP2 , since
32
p0011 p1111 − p0111 p1011 = −
,
2432
which completes the proof by a similar argument as in Example 7.2.
cov(X3 , X4 ) =
We now show how to overcome these limitations and build MTP2 distributions over non-decomposable graphs, namely by using MTP2 potentials
over the edges instead of MTP2 marginal distributions over the edges.
Theorem 7.5.
A distribution of the form
1 Y
p(x) =
ψuv (xu , xv ),
Z
uv∈E
where ψuv are positive functions and Z is a normalizing constant, is MTP2
if and only if each ψuv is an MTP2 function.
Proof. Since the distribution p is strictly positive, by Proposition 3.5
p satisfies MTP2 if and only if it does so for x, y ∈ X that differ in two
coordinates, say with indices u, v. Write Eu for the set of edges that contain u
but not v and Ev for the set of edges that contain v but not u. First, consider
the case where uv ∈ E. Then we have that p(x ∧ y)p(x ∨ y) − p(x)p(y) ≥ 0
if and only if
Y
ψst ((x ∧ y)st ) ψst ((x ∨ y)st ) ≥
ψuv ((x ∧ y)uv ) ψuv ((x ∨ y)uv )
st∈Eu ∪Ev
ψuv (xuv ) ψuv (yuv )
Y
st∈Eu ∪Ev
ψst (xst ) ψst (yst ).
24
S. FALLAT ET AL.
All other terms cancel out because of the assumption that xw = yw for
all w ∈ V \ {u, v}. Now note that for st ∈ Eu ∪ Ev we have {xst , yst } =
{(x ∧ y)st , (x ∨ y)st } and so the above inequality holds if and only if
ψuv ((x ∧ y)uv )ψuv ((x ∨ y)uv ) ≥ ψuv (xuv )ψuv (yuv ),
that is, if and only if ψuv is MTP2 .
Next consider the case where uv ∈
/ E. By the same argument one concludes that in this case the above inequalities are in fact equalities, which
completes the proof.
As a final remark note that Theorem 7.5 can directly be extended to
distributions of the form
1 Y
p(x) =
ψC (xC ),
Z
C∈C
where C is a family of subsets of V such that for any two C, C ′ ∈ C we have
|C ∩ C ′ | ∈ {0, 1}.
7.2. Log-linear interactions. We next give a short discussion of interaction representations for discrete MTP2 distributions, as they typically are
used in log-linear models for contingency
Q tables. Suppose that X = (Xv )v∈V
is a random vector with values in X = v∈V Xv where each Xv is finite. Let
D denote the set of subsets of V . Any function h : X → Rn of X can be
expanded as
X
(7.1)
h(x) =
θD (x),
D∈D
where θD are functions on X that only depend on x through xD , i.e. θD
satisfy that θD (x) = θD (xD ). In the case where h(x) = log p(x) where p is a
positive probability distributions over X , the functions θD (x) are known as
interactions among variables in D and we shall also use this expression for
a general function h.
Without loss of generality we may assume that min Xv = 0 for all v ∈ V
and to assure that the representation is unique, we may require that θD (x) =
0 whenever xd = 0 for some d ∈ D. With this convention, the sum in (7.1)
can be rewritten so it only extends over such D ∈ D which are contained
in the support S(x) of x where d ∈ S(x) ⇐⇒ xd 6= 0. In the binary case,
when dv = 1, this allows us to use a simpler notation, namely θD (1D ) := θD
for all D ∈ D.
TOTAL POSITIVITY IN MARKOV STRUCTURES
25
For any such interaction expansion and a fixed pair u, w ∈ V , we define
a function γuw on X by
X
γuw (x) =
θD (x).
D:{u,w}⊆D⊆S(x)
Observe that then γuw (x) = 0 unless u, w ∈ S(x) and thus in particular
whenever |S(x)| ≤ 1; further, γuw is a linear combination of interaction
terms.
Let Z ⊆ X be a subset of X which is closed under ∧ and ∨. A function
g : Z → R is supermodular if
g(x ∧ y) + g(x ∨ y) ≥ g(x) + g(y)
for all x, y ∈ Z.
Thus g is supermodular on X if and only if exp(g) is MTP2 on X . A function
g is modular if both of g and −g are supermodular.
Denote by X A the set of all x ∈ X with S(x) = A. Clearly, X A is closed
under ∧ and ∨. Then we obtain the following result.
Theorem 7.6. Let P be a strictly positive distribution of X. Then P is
MTP2 if and only if for all A ⊆ V with |A| ≥ 2 and any given u, w ∈ V the
function γuw is non-negative, non-decreasing, and supermodular over X A .
Proof. By Proposition 3.5 p is MTP2 if and only if
(7.2)
log p(x ∧ y) + log p(x ∨ y) − log p(x) − log p(y) ≥ 0
for all x, y ∈ X that differ only in two entries. Let u, w ∈ V and take
x, y ∈ X satisfying xv = yv for all v ∈ V \ {u, w}. Without loss of generality
we can assume xu < yu and yw < xw for otherwise the inequality is trivially
satisfied. Using the expansion (7.1), the inequality (7.2) becomes
X
θD ((x ∧ y)D ) + θD ((x ∨ y)D ) − θD (xD ) − θD (yD ) ≥ 0.
(7.3)
D∈D
For every D ⊆ V \ {w} we have (x ∧ y)D = xD and (x ∨ y)D = yD .
Similarly, for every D ⊆ V \ {u} we have (x ∧ y)D = yD and (x ∨ y)D = xD ,
and thus, in both cases the corresponding summands in (7.3) are zero. It
follows that (7.3) is equivalent to
X
θD ((x∧y)D )+θD ((x∨y)D )−θD (xD )−θD (yD ) ≥ 0,
(7.4)
D:{u,w}⊆D⊆A
26
S. FALLAT ET AL.
where A ⊆ V is the support of x ∨ y (if D is not contained in A all terms
θD are zero by our convention).
We now show that the fact that (7.4) must hold for all u, w ∈ V and
x, y ∈ X as above is equivalent to the fact that all γuw satisfy the conditions
of the theorem.
Consider three possible cases:
(a) S(x) = A \ {u}, S(y) = A \ {w},
(b) either S(x) = A \ {u}, S(y) = A or S(x) = A, S(y) = A \ {w};
(c) S(x) = S(y) = A.
In other words: in case (a) we have xu = yw = 0; in case (b) either xu = 0,
yw > 0 or xu > 0, yw = 0; and in case (c) we have xu , yw > 0.
In case (a) we have θD ((x ∧ y)D ) = θD (xD ) = θD (yD ) = 0 for every D
containing {u, w} so (7.4) becomes
X
θD ((x ∨ y)D ) ≥ 0.
D:{u,w}⊆D⊆A
By choosing different pairs x, y, this can be equivalently rewritten as
X
θD (xD ) ≥ 0 for all x ∈ X
D:{u,w}⊆D⊆S(x)
and the sum on the left is precisely γuw (x).
In case (b), if S(x) = A \ {u}, S(y) = A, (7.4) becomes
X
(θD ((x ∨ y)D ) − θD (yD )) ≥ 0,
D:{u,w}⊆D⊆A
where A = S(x ∨ y) = S(y), which is equivalent to γuw being nondecreasing
on X A .
Finally, in case (c), all x ∨ y, x ∧ y, x, y have the same support. Thus γuw
must be supermodular over each X A .
As a special case we recover the characterization of binary MTP2 distributions in [4].
Corollary 7.7.
Let P be a binary distribution with
X
log p(x) =
θD
D:D⊆S(x)
TOTAL POSITIVITY IN MARKOV STRUCTURES
27
using the convention that θD = θD (1D ). Then P is MTP2 if and only if for
all A with |A| ≥ 2 and all {u, w} ⊆ V we have
X
θD ≥ 0.
D:{u,w}⊆D⊆A
Proof. In the binary case each X A has only one element and so the
only constraint from Theorem 7.6 is the non-negativity constraint.
Example 7.8. Let X = (X1 = 1A , X2 = 1B , X3 = 1C ) be the vector of
binary indicator functions of events A, B, C. Reichenbach [35, p. 190] (using
a different notation) says that an event B is causally between A and C if
P (C | B ∧ A) = P (C | B) and further
1 > P (C | B) > P (C | A) > P (C) > 0,
1 > P (A | B) > P (A | C) > P (A) > 0.
Equivalently, as defined in [6], B is causally between A and C if the following
hold:
(7.5)
P (A ∧ C) > P (A)P (C)
(7.6)
P (A | B) > P (A | C)
(7.7)
P (C | B) > P (C | A)
(7.8)
(7.9)
P (A ∧ C | B) = P (A | B)P (C | B),
P (¬A ∧ B) > 0,
P (¬C ∧ B) > 0.
In general, causal betweenness does not imply MTP2 ; if we let p101 = 0,
p000 = 4/10, and pijk = 1/10 for the remaining six possibilities, B is causally
between A and C, but X is not MTP2 since 0 = p101 p000 < p100 p001 .
However, if P (X = x) > 0 for all x and B is causally between A and
C, then P is MTP2 . To see this, we expand P in log-linear interaction
parameters to get
(7.10)
1A ⊥⊥ 1C | (1B = 1) ⇐⇒ θAC + θABC = 0.
Further, a simple but somewhat tedious calculation using (7.10) yields that
(7.6) and (7.7) hold if and only if
θAB > θAC ,
θBC > θAC ,
which in combination with (7.10) gives that (7.6)–(7.8) hold if and only if
(7.11)
θAB + θABC > 0,
θBC + θABC > 0,
θAC + θABC = 0.
28
S. FALLAT ET AL.
Thus, Corollary 7.7 ensures that P is MTP2 .
Conversely, if P (X = x) > 0 for all x and P is MTP2 , the condition
1A ⊥⊥ 1C | (1B = 1) implies “weak causal betweenness”, i.e. P satisfies the
inequalities (7.5)–(7.7) with > replaced by ≥; this is true because the weak
form of (7.6) and (7.7) follows from the weak form of (7.11), and (7.5) in its
weak form expresses that cov(X1 , X3 ) ≥ 0, which is also a consequence of
MTP2 .
Finally, if P (X = x) > 0 for all x, P is MTP2 , and the independence
graph of P is 1—2—3, then B is causally between A and C, as then P is
faithful by Theorem 6.1, which ensures that inequalities are strict.
7.3. Conditional Gaussian distributions. In this section, we study CGdistributions satisfying the MTP2 property. The density of a CG-distribution
is given by specifying a strictly positive distribution p(i) over the discrete
variables for i ∈ X∆ . Then the joint density f (x) = f (i, y) is determined by
specifying f (y | i) to be the density of a Gaussian distribution NΓ (ξ(i), Σ(i)),
where ξ(i) ∈ RΓ is the mean vector and Σ(i) is the covariance matrix. CGdistributions can also be represented by the set of canonical characteristics
(g, h, K) where
1
log f (x) = log f (y, i) = g(i) + h(i)T y − y T K(i)y;
2
see [22]. Here K(i) = Σ−1 (i) is the conditional concentration matrix. We
shall say that a function u(i) is additive if it has the form
X
u(i) =
αδ (iδ ).
δ∈∆
Before we characterize CG-distribution with the MTP2 property, we need a
small lemma.
Lemma 7.9.
modular.
A function u : X∆ → R is additive if and only if it is
Proof. If u is additive, then it is clearly modular. To show the converse,
we make a log-linear expansion of u as in (7.1)
X
u(i) =
ηD (i).
D∈D
We shall show that if u is modular, then ηD (i) = 0 whenever |D| ≥ 2. If for
C ⊆ V we let
wC (i) = u(iC , 0V \C ),
TOTAL POSITIVITY IN MARKOV STRUCTURES
29
it follows from the Möbius inversion lemma, also known as inclusion–exclusion
( e.g. p. 239 of [22]) that
X
ηD (i) =
(−1)|D\A| wA (i).
A:A⊆D
If |D| ≥ 2 we can for distinct u, v ∈ D rewrite this as
ηD (i) =
X
(−1)|D\A| wA∪{u,v} (i) − wA∪{u} (i) − wA∪{v} (i) + wA (i) .
A:A⊆D\{u,v}
If u is modular, all terms inside the curly brackets are zero and hence u is
additive.
Proposition 7.10. A CG-distribution P with canonical characteristics
(g, h, K) is MTP2 if and only if
(i) g(i) is supermodular;
(ii) h(i) is additive and non-decreasing;
(iii) K(i) = K for all i where K is an M -matrix.
Proof. By Proposition 3.5, a CG distribution is MTP2 if and only if it
satisfies
f (y ∧ z, i ∧ j)f (y ∨ z, i ∨ j) ≥ f (y, i)f (z, j)
for cases where (y, i) and (z, j) differ on two coordinates. Suppose first that
i = j and y, z differ on two coordinates. Then we equivalently need to check
whether
f (y ∧ z | i)f (y ∨ z | i) ≥ f (y | i)f (z | i).
Since f (y | i) is the density of a Gaussian distribution, this inequality holds
for every y, z ∈ RΓ and i if and only if each K(i) is an M -matrix.
If i, j and y, z both differ on one coordinate then without loss of generality
we can assume i < j and y > z so that i = i ∧ j and z = y ∧ z. In this case
we need to show that
(7.12)
log f (z, i) + log f (y, j) ≥ log f (y, i) + log f (z, j).
Write y = z + tek for some t > 0, where ek is a unit vector in RΓ . Then
equivalently
1
1
(log f (z + tek , j) − log f (z, j)) ≥ (log f (z + tek , i) − log f (z, i)).
t
t
30
S. FALLAT ET AL.
Since this holds for every t > 0, we can take the limit t → 0, which implies
that necessarily
∇z log f (z, j) ≥ ∇z log f (z, i)
for all z ∈ RΓ , i < j ∈ ∆.
Since ∇y log f (y, i) = h(i) − K(i)y, this is equivalent to
h(j) − h(i) − (K(j) − K(i))z ≥ 0.
The function on the left-hand side is linear in z and thus this holds for every
z if and only if K(j) = K(i) for every i, j and h(i) is non-decreasing in i.
If y = z and i, j differ on two coordinates, using all the conditions that
have been already proven to be necessary we need to check that
(7.13) (g(i∧j)+g(i∨j)−g(i)−g(j))+(h(i∧j)+h(i∨j)−h(i)−h(j))T z ≥ 0.
This can hold for every z only if h is modular, i.e.
(7.14)
h(i ∧ j) + h(i ∨ j) − h(i) − h(j) = 0 for all i, j.
Now if (7.14) holds, (7.13) holds if and only if g(i) is super-modular. By
Lemma 7.9, h is additive, which concludes the proof.
Proposition 7.10 gives a simple condition for CG distributions to be MTP2
in terms of their canonical characteristics. This also implies that the moment
characteristics (p, ξ, Σ) have simple properties.
Proposition 7.11. If a CG-distribution is MTP2 , its moment characteristics (p, ξ, Σ) satisfy
(i) p(i) is MTP2 ;
(ii) ξ(i) is additive and non-decreasing;
(iii) Σ(i) = Σ for all i and all elements of Σ are non-negative.
Proof. If the CG distribution is MTP2 , (iii) follows directly from Proposition 7.10. The condition (i) follows since marginals of MTP2 distributions
are MTP2 and (ii) follows from (ii) of Proposition 7.10 since ξ(i) = Σh(i)
and Σ has only non-negative elements.
Thus, MTP2 CG-distributions are in particular homogeneous — Σ(i) constant in i — and mean-additive [12, 22]. Note that the converse of Proposition 7.11 is not true since ξ(i) can be non-increasing and h = Kξ(i) decreasing, even when K is an M-matrix.
TOTAL POSITIVITY IN MARKOV STRUCTURES
Finally we make expansions of g(i) as in Section 7.2:
X
X
g
g(i) =
λD (i),
γuw
(i) =
D∈D
31
λD (i)
D:{u,w}⊆D⊆S(i)
and recall that I A = {i : S(i) = A}. We then have the following alternative
formulation of Proposition 7.10:
Corollary 7.12.
A CG-distribution is MTP2 if and only if
g
(i) For all A ⊆ V with |A| ≥ 2 and all u, w ∈ V , the functions γuw
(i) are
A
supermodular and non-decreasing over each I ;
(ii) The function h(i) is additive
X
h(i) =
αδ (iδ )
δ∈∆
with non-decreasing components αδ (iδ )v ;
(iii) There exists an M -matrix K such that K(i) = K for all i.
Proof. The proof of Theorem 7.6 does not use that log p is the logarithm of a probability distribution; hence the corresponding conclusions also
apply to the expansion of g.
Note that if i is binary, condition (i) of the Corollary simplifies as in
Corollary 7.7 to the condition that for all A ⊆ V and all u 6= w ∈ A we have
X
λD ≥ 0.
D:{u,w}⊆D⊆A
8. Discussion. In this paper, we showed that MTP2 distributions enjoy many important properties related to conditional independence; in particular, an independence model generated by an MTP2 distribution is an
upward-stable and singleton-transitive compositional semigraphoid which is
faithful to its concentration graph if it has coordinate-wise connected support.
We illustrated with several examples that MTP2 models are useful for
data analysis. The MTP2 constraint seems restrictive when no conditional
independences are taken into account. However, the picture changes and
the MTP2 constraint becomes less restrictive when imposing conditional
independence constraints in the form of Markov properties.
An important property of MTP2 models, which is of practical relevance, is
that the positive conditional dependence of two variables given all remaining
32
S. FALLAT ET AL.
variables implies a positive dependence given any subset of the remaining
variables. This is a highly desirable feature, especially when results of followup empirical studies are to be compared with an earlier comprehensive study,
and the studies coincide only in a subset of core variables.
More generally, in any MTP2 concentration graph model, dependence reversals cannot occur. This undesirable, worrying feature has been described
and studied under different names depending on the types of the involved
variables, for instance as near multicollinearity for continuous variables, as
the Yule-Simpson paradox for discrete variables, or as the effects of highly
unbalanced experimental designs with explanatory discrete variables for continuous responses. It is a remarkable feature of MTP2 distributions that such
dependence reversals are absent.
These observations suggest that it would be desirable to develop further
methods for hypothesis testing and estimation under the MTP2 constraint
as done for the binary case in [4]. Our results may also be applied not to
the joint distribution of all variables, but only to the joint distribution of a
subset of the variables, given a fixed set of level combinations of the remaining variables. This is particularly interesting in empirical studies, where a
set of possible regressors and background variables is manipulated to make
the studied groups of individuals as comparable as possible; for instance
by selecting equal numbers of persons for fixed level combinations of some
features, by proportional allocation of patients to treatments, by matching or by stratified sampling. In such situations, not much can be inferred
from the study results about the conditional distribution of the manipulated
variables given the responses. However, the MTP2 property of the joint conditional distribution of the responses given the manipulated variables could
be essential.
Acknowledgements. This article represents research initiated at the
American Institute of Mathematics workshop “Positivity, Graphical Models,
and the Modeling of Complex Multivariate Dependencies” in October 2014.
The authors thank the AIM and NSF for their support and are indebted to
other participants of the workshop for useful and constructive discussions.
We are especially grateful to Hélène Massam for her constructive remarks
and encouragement. Thanks are also due to Heiko Becher for permission to
use his data in Example 4.2, and to Antonio Forcina and two anonymous
referees for their helpful comments.
TOTAL POSITIVITY IN MARKOV STRUCTURES
33
References.
[1] Ahlswede, R. and Daykin, D. E. (1978). An inequality for the weights of two
families of sets, their unions and intersections. Probability Theory and Related Fields
43 183–185.
[2] Allman, E. S., Rhodes, J. A., Sturmfels, B. and Zwiernik, P. (2015). Tensors
of nonnegative rank two. Linear Algebra Appl. 473 37–53. MR3338324
[3] Bartolucci, F. and Besag, J. (2002). A recursive algorithm for Markov random
fields. Biometrika 89 724–730.
[4] Bartolucci, F. and Forcina, A. (2000). A likelihood ratio test for MTP2 within
binary variables. Ann. Statist. 28 1206–1218. MR1811325 (2003b:62121)
[5] Bølviken, E. (1982). Probability Inequalities for the Multivariate Normal with Nonnegative Partial Correlations. Scand. J. Statist. 9 49–58.
[6] Chvátal, V. and Wu, B. (2011). On Reichenbach’s Causal Betweenness. Erkenntnis
76 41–48.
[7] Colangelo, A., Scarsini, M. and Shaked, M. (2005). Some notions of multivariate
positive dependence. Insurance Math. Econom. 37 13–26.
[8] Dawid, A. P. (1979). Conditional independence in statistical theory (with discussion). J. Roy. Stat. Soc. Ser. B 41 1–31.
[9] Dawid, A. P. (1980). Conditional Independence for Statistical Operations. Annals
of Statistics 8 598–617.
[10] Dawid, A. P. and Lauritzen, S. L. (1993). Hyper Markov Laws in the Statistical
Analysis of Decomposable Graphical Models. Ann. Statist. 21 1272–1317.
[11] Djolonga, J. and Krause, A. (2015). Scalable Variational Inference in Logsupermodular Models. arXiv preprint arXiv:1502.06531.
[12] Edwards, D. (2000). Introduction to Graphical Modelling, second ed. Springer Texts
in Statistics. Springer-Verlag, New York. MR1880319
[13] Esary, J. D., Proschan, F. and Walkup, D. W. (1967). Association of random
variables, with applications. Ann. Math. Stat. 38 1466–1474.
[14] Fortuin, C. M., Kasteleyn, P. W. and Ginibre, J. (1971). Correlation inequalities
on some partially ordered sets. Comm. Math. Phys. 22 89–103.
[15] Højsgaard, S. and Lauritzen, S. L. (2008). Graphical Gaussian models with edge
and vertex symmetries. J. R. Stat. Soc. Ser. B 70 1005–1027. MR2530327
[16] Holland, P. W. and Rosenbaum, P. R. (1986). Conditional Association and Unidimensionality in Monotone Latent Variable Models. Ann. Statist. 14 1523–1543.
[17] Joag-Dev, K. (1983). Independence Via Uncorrelatedness Under Certain Dependence Structures. Ann. Probab. 11 1037–1041.
[18] Joe, H. (2006). Generating random correlation matrices based on partial correlations.
J. Multiv. Anal. 97 2177 - 2189.
[19] Jones, B. and West, M. (2005). Covariance Decomposition in Undirected Gaussian
Graphical Models. Biometrika 92 779-786.
[20] Karlin, S. and Rinott, Y. (1980). Classes of Orderings of Measures and Related
Correlation Inequalities. I. Multivariate Totally Positive Distributions. J. Multiv.
Anal. 10 467–498. MR599685 (83j:60020a)
[21] Karlin, S. and Rinott, Y. (1983). M-Matrices as covariance matrices of multinormal distributions. Linear Algebra Appl. 52 419 - 438.
[22] Lauritzen, S. L. (1996). Graphical Models. Clarendon Press, Oxford, United Kingdom.
[23] Lebowitz, J. L. (1972). Bounds on the correlations and analyticity properties of
ferromagnetic Ising spin systems. Comm. Math. Phys. 28 313–321.
34
S. FALLAT ET AL.
[24] Lehmann, E. L. (1966). Some Concepts of Dependence. Ann. Math. Statist. 37
1137–1153.
[25] Mardia, K. V., Kent, J. T. and Bibby, J. M. (1979). Multivariate Analysis. Academic Press [Harcourt Brace Jovanovich, Publishers], London-New York-Toronto,
Ont. Probability and Mathematical Statistics: A Series of Monographs and Textbooks. MR560319 (81h:62003)
[26] Martı́n, E. S., Mouchart, M. and Rolin, J.-M. (2005). Ignorable Common Information, Null Sets and Basu’s First Theorem. Sankhyā: The Indian Journal of
Statistics 67 674-698.
[27] Müller, A. and Scarsini, M. (2005). Archimedean copulae and positive dependence. Journal of Multivariate Analysis 93 434–445.
[28] Newman, C. M. (1983). A general central limit theorem for FKG systems. Comm.
Math. Phys. 91 75–80.
[29] Newman, C. M. (1984). Asymptotic Independence and Limit Theorems for Positively and Negatively Dependent Random Variables. Lecture Notes-Monograph Series
5 127-140.
[30] Ostrowski, A. (1937). Über die Determinanten mit überwiegender Hauptdiagonale.
Commentarii Mathematici Helvetici 10 69–96.
[31] Pearl, J. (1988). Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann Publishers, San Mateo, CA, USA.
[32] Pearl, J. and Paz, A. (1987). Graphoids: A Graph Based Logic for Reasoning about
Relevancy Relations. In Advances in Artificial Intelligence, (B. D. Boulay, D. Hogg
and L. Steel, eds.) II 357–363. North-Holland, Amsterdam.
[33] Peters, J. (2015). On the Intersection Property of Conditional Independence and
its Application to Causal Discovery. J. Causal Inference 3 97–108.
[34] Propp, J. G. and Wilson, D. B. (1996). Exact sampling with coupled Markov
chains and applications to statistical mechanics. Random Structures Algorithms 9
223–252.
[35] Reichenbach, H. (1956). The Direction of Time. University of California Press,
Berkeley, CA.
[36] Royston, P., Altman, D. G. and Sauerbrei, W. (2006). Dichotomizing continuous
predictors in multiple regression: a bad idea. Stat Med 25 127–141.
[37] Sadeghi, K. and Lauritzen, S. L. (2014). Markov properties for mixed graphs.
Bernoulli 20 676–696.
[38] Sarkar, T. K. (1969). Some lower bounds of reliability Tech. Report, No. 124, Department of Operations Research and Department of Statistics, Stanford University.
[39] Slawski, M. and Hein, M. (2015). Estimation of positive definite M-matrices and
structure learning for attractive Gaussian Markov random field. Linear Algebra Appl.
473 145–179.
[40] Steel, M. and Faller, B. (2009). Markovian log-supermodularity, and its applications in phylogenetics. Appl. Math. Lett. 22 1141 - 1144.
[41] Studený, M. (2005). Probabilistic Conditional Independence Structures. SpringerVerlag, London, United Kingdom.
[42] Štembera, Z., Znamenáček, K. and Poláček, K. (2012). High risk pregnancy and
child. Springer Science & Business Media.
[43] Wermuth, N. (2012). Traceable regressions. Int. Stat. Rev. 80 415–438.
[44] Wermuth, N. (2015). Graphical Markov Models, unifying results and their interpretation. In Wiley Statsref: Statistics Reference Online; also on arXiv:1505.02456.
[45] Wermuth, N. and Cox, D. R. (1998). On Association Models Defined over Independence Graphs. Bernoulli 4 477-495.
TOTAL POSITIVITY IN MARKOV STRUCTURES
35
[46] Wermuth, N. and Marchetti, G. M. (2014). Star graphs induce tetrad correlations: for Gaussian as well as for binary variables. Electron. J. Stat. 8 253–273.
[47] Whittaker, J. (1990). Graphical Models in Applied Multivariate Statistics. Wiley
Series in Probability and Mathematical Statistics. John Wiley & Sons, Ltd., Chichester. MR1112133 (93f:62002)
[48] Xie, X., Ma, Z. and Geng, Z. (2008). Some association measures and their collapsibility. Statist. Sinica 18 1165–1183.
[49] Zatoński, W., Becher, H., Lissowska, J. and Wahrendorf, J. (1991). Tobacco,
alcohol, and diet in the etiology of laryngeal cancer: a population-based case-control
study. Cancer Causes & Control 2 3–10.
[50] Zwiernik, P. (2015). Semialgebraic Statistics and Latent Tree Models. Monographs
on Statistics and Applied Probability 146. Chapman & Hall.
Department of Mathematics and Statistics
University of Regina
Regina, SK, Canada
E-mail: [email protected]
Statistical Laboratory
University of Cambridge
Cambridge, UK
E-mail: [email protected]
Chalmers University of Technology
Gothenburg, Sweden
and Johannes Gutenberg-University
Mainz, Germany
E-mail: [email protected]
Department of Mathematical Sciences
University of Copenhagen
Copenhagen, Denmark
E-mail: [email protected]
Massachusetts Institute of Technology
Cambridge, USA
and Institute of Science and Technology Austria
Klosterneuburg, Austria
E-mail: [email protected]
Universitat Pompeu Fabra
and Barcelona GSE
Barcelona, Spain
E-mail: [email protected]
| 10math.ST
|
arXiv:1209.5429v4 [cs.NE] 1 Jul 2014
copulaedas: An R Package for Estimation of
Distribution Algorithms Based on Copulas
Yasser Gonzalez-Fernandez
Marta Soto
Institute of Cybernetics,
Mathematics and Physics
Institute of Cybernetics,
Mathematics and Physics
Abstract
The use of copula-based models in EDAs (estimation of distribution algorithms) is
currently an active area of research. In this context, the copulaedas package for R provides
a platform where EDAs based on copulas can be implemented and studied. The package
offers complete implementations of various EDAs based on copulas and vines, a group
of well-known optimization problems, and utility functions to study the performance of
the algorithms. Newly developed EDAs can be easily integrated into the package by
extending an S4 class with generic functions for their main components. This paper
presents copulaedas by providing an overview of EDAs based on copulas, a description of
the implementation of the package, and an illustration of its use through examples. The
examples include running the EDAs defined in the package, implementing new algorithms,
and performing an empirical study to compare the behavior of different algorithms on
benchmark functions and a real-world problem.
Keywords: black-box optimization, estimation of distribution algorithm, copula, vine, R.
1. Introduction
The field of numerical optimization (see e.g., Nocedal and Wright 1999) is a research area with
a considerable number of applications in engineering, science, and business. Many mathematical problems involve finding the most favorable configuration of a set of parameters that
achieve an objective quantified by a function. Numerical optimization entails the case where
these parameters can take continuous values, in contrast with combinatorial optimization,
which involves discrete variables. The mathematical formulation of a numerical optimization
problem is given by min f (x), where x ∈ Rn is a real vector with n ≥ 1 components and
x
f : Rn → R is the objective function (also known as the fitness, loss or cost function).
In particular, we consider within numerical optimization a black-box (or direct search) scenario where the function values of evaluated search points are the only available information
on f . The algorithms do not assume any knowledge of the function f regarding continuity, the
existence of derivatives, etc. A black-box optimization procedure explores the search space by
generating solutions, evaluating them, and processing the results of this evaluation in order to
generate new promising solutions. In this context, the performance measure of the algorithms
is generally the number of function evaluations needed to reach a certain value of f .
Algorithms that have been proposed to deal with this kind of optimization problems can be
2
copulaedas: An R Package for EDAs Based on Copulas
classified in two groups according to the approach followed for the generation of new solutions. On the one hand, deterministic direct search algorithms, such as the Hooke-Jeeves
(Hooke and Jeeves 1961) and Nelder-Mead (Nelder and Mead 1965) methods, perform transformations to one or more candidate solutions at each iteration. Given their deterministic
approach, these algorithms may have limited global search capabilities and can get stuck in local optima, depending on an appropriate selection of the initial solutions. On the other hand,
randomized optimization algorithms offer an alternative to ensure a proper global exploration
of the search space. Examples of these algorithms are simulated annealing (Kirkpatrick et al.
1983), evolution strategies (see e.g., Beyer and Schwefel 2002), particle swarm optimization
(Kennedy and Eberhart 1995), and differential evolution (Storn and Price 1997).
In this paper, we focus on EDAs (estimation of distribution algorithms; Mühlenbein and Paaß
1996; Baluja 1994; Larrañaga and Lozano 2002; Pelikan et al. 2002), which are stochastic
black-box optimization algorithms characterized by the explicit use of probabilistic models
to explore the search space. These algorithms combine ideas from genetic and evolutionary
computation, machine learning, and statistics into an optimization procedure. The search
space is explored by iteratively estimating and sampling from a probability distribution built
from promising solutions, a characteristic that differentiates EDAs among other randomized
optimization algorithms. One key advantage of EDAs is that the search distribution may
encode probabilistic dependences between the problem variables that represent structural
properties of the objective function, performing a more effective optimization by using this
information.
Due to its tractable properties, the normal distribution has been commonly used to model
the search distributions in EDAs for real-valued optimization problems (Bosman and Thierens
2006; Kern et al. 2003). However, once a multivariate normal distribution is assumed, all the
margins are modeled with the normal density and only linear correlation between the variables
can be considered. These characteristics could lead to the construction of incorrect models
of the search space. For instance, the multivariate normal distribution cannot represent
properly the fitness landscape of multimodal objective functions. Also, the use of normal
margins imposes limitations on the performance when the sample of the initial solutions is
generated asymmetrically with respect to the optimum of the function (see Soto et al. in press
for an illustrative example of this situation).
Copula functions (see e.g., Joe 1997; Nelsen 2006) offer a valuable alternative to tackle these
problems. By means of Sklar’s Theorem (Sklar 1959), any multivariate distribution can be
decomposed into the (possibly different) univariate marginal distributions and a multivariate
copula that determines the dependence structure between the variables. EDAs based on copulas inherit these properties, and consequently, can build more flexible search distributions that
may overcome the limitations of a multivariate normal probabilistic model. The advantages
of using copula-based search distributions in EDAs extend further with the possibility of factorizing the multivariate copula with the copula decomposition in terms of lower-dimensional
copulas. Multivariate dependence models based on copula factorizations, such as nested
Archimedean copulas (Joe 1997) and vines (Joe 1996; Bedford and Cooke 2001; Aas et al.
2009), provide great advantages in high dimensions. Particularly in the case of vines, a more
appropriate representation of multivariate distributions having pairs of variables with different
types of dependence is possible.
Although various EDAs based on copulas have been proposed in the literature, as far as we
know there are no publicly available implementations of these algorithms (see Santana 2011
Yasser Gonzalez-Fernandez, Marta Soto
3
for a comprehensive review of EDA software). Aiming to fill this gap, the copulaedas package
(Gonzalez-Fernandez and Soto 2014a) for the R language and environment for statistical computing (R Core Team 2014) has been published on the Comprehensive R Archive Network at
http://CRAN.R-project.org/package=copulaedas. This package provides a modular platform where EDAs based on copulas can be implemented and studied. It contains various
EDAs based on copulas, a group of well-known benchmark problems, and utility functions to
study EDAs. One of the most remarkable features of the framework offered by copulaedas is
that the components of the EDAs are decoupled into separated generic functions, which promotes code factorization and facilitates the implementation of new EDAs that can be easily
integrated into the framework.
The remainder of this paper provides a presentation of the copulaedas package organized
as follows. Section 2 continues with the necessary background on EDAs based on copulas.
Next, the details of the implementation of copulaedas are described in Section 3, followed by
an illustration of the use of the package through examples in Section 4. Finally, concluding
remarks are given in Section 5.
2. Estimation of distribution algorithms based on copulas
This section begins by describing the general procedure of an EDA, according to the implementation in copulaedas. Then, we present an overview of the EDAs based on copulas
proposed in the literature with emphasis on the algorithms implemented in the package.
2.1. General procedure of an EDA
The procedure of an EDA is built around the concept of performing the repeated refinement of
a probabilistic model that represents the best solutions of the optimization problem. A typical
EDA starts with the generation of a population of initial solutions sampled from the uniform
distribution over the admissible search space of the problem. This population is ranked
according to the value of the objective function and a subpopulation with the best solutions
is selected. The algorithm then constructs a probabilistic model to represent the solutions in
the selected population and new offspring are generated by sampling the distribution encoded
in the model. This process is repeated until some termination criterion is satisfied (e.g., when
a sufficiently good solution is found) and each iteration of this procedure is called a generation
of the EDA. Therefore, the feedback for the refinement of the probabilistic model comes from
the best solutions sampled from an earlier probabilistic model.
Let us illustrate the basic EDA procedure with a concrete example. Figure 1 shows the
steps performed to minimize the two-dimensional objective function f (x1 , x2 ) = x21 + x22
using a simple continuous EDA that assumes independence between the problem variables.
Specifically, we aim to find the global optimum of the function f (0, 0) = 0 with a precision of
two decimal places.
The algorithm starts by generating an initial population of 30 candidate solutions from a
continuous uniform distribution in [−10, 10]2 . Out of this initial sampling, the best solution
found so far is f (−2.20, −0.01) = 4.85. Next, the initial population is ranked according
to their evaluation in f (x1 , x2 ), and the best 30% of the solutions is selected to estimate
the probabilistic model. This EDA factorizes the joint probability density function (PDF)
of the best solutions as φ1,2 (x1 , x2 ) = φ1 (x1 , µ1 , σ12 )φ2 (x2 , µ2 , σ22 ), which describes mutual
4
copulaedas: An R Package for EDAs Based on Copulas
i←1
repeat
if i = 1 then
Generate an initial population P1 using a seeding method.
Evaluate the solutions in the population P1 .
If required, apply a local optimization method to the population P1 .
else
Select a population PiSelected from Pi−1 according to a selection method.
Learn a probabilistic model Mi from PiSelected using a learning method.
Sample a new population PiSampled from Mi using a sampling method.
Evaluate the solutions in the population PiSampled .
If required, apply a local optimization method to the population PiSampled .
Create the population Pi from Pi−1 and PiSampled using a replacement method.
end if
If required, report progress information using a reporting method.
i ←i+1
until A criterion of the termination method is met.
Algorithm 1: Pseudocode of an EDA.
independence, and where φ1 denotes the univariate normal PDF of x1 with mean µ1 and
variance σ12 , and φ1 denotes the univariate normal PDF of x2 with mean µ1 and variance σ22 .
In the first generation, the parameters of the probabilistic model are µ1 = −0.04, σ1 = 3.27,
µ2 = −0.66 and σ2 = 3.81. The second generation starts with the simulation of a new
population from the estimated probabilistic model. Afterwards, the same selection procedure
is repeated and the resulting selected population is used to learn a new probabilistic model.
These steps are then repeated for a third generation.
Notice how in the first three generations the refinement of the probabilistic model that represents the best solutions is evidenced in the reduction of the variance of the marginal distributions towards a mean value around zero. Also, the convergence of the algorithm is reflected
in the reduction of the value of the objective function from one generation to another. Ultimately, the simulation of the probabilistic model estimated at the third generation produces
f (−0.02, −0.07) = 0.00, which satisfies our requirements and the algorithm terminates.
In practice, EDAs include other steps in addition to the ones illustrated in the previous
basic example. The general procedure of an EDA implemented in copulaedas is outlined in
Algorithm 1. In the following, we provide a description of the purpose of the main steps of
the algorithm, which are highlighted in italics in the pseudocode:
• The first step is the generation of an initial population of solutions following a seeding
method, which is usually random, but it can use a particular heuristic when a priori
information about the characteristics of the problem is available.
• The results of global optimization algorithms such as EDAs can often be improved if
combined with local optimization methods that look for better solutions in the neighborhood of each candidate solution. Local optimization methods can also be used to
implement repairing methods for constrained problems where the simulated solutions
may be unfeasible and a strategy to repair these solutions is available.
5
Yasser Gonzalez-Fernandez, Marta Soto
x1
x2
f (x1 , x2 )
5.21
7.72
-0.87
-6.67
-1.95
4.14
2.89
-2.20
.
.
.
-5.47
-6.24
3.63
-2.59
-2.76
-9.13
-5.73
4.649
-0.01
.
.
.
-1.28
66.25
72.86
7.50
52.18
87.17
50.16
29.97
4.85
.
.
.
31.59
Selection
x1
x2
f (x1 , x2 )
-2.20
-0.87
-0.75
4.55
-2.17
-0.30
2.89
-5.47
3.97
-0.01
-2.59
-3.57
2.34
-4.80
-5.26
4.64
-1.28
4.59
4.85
7.50
13.34
26.25
27.83
27.85
29.97
31.59
36.88
µ1 = −0.04, σ1 = 3.27
µ2 = −0.66, σ2 = 3.81
Sampling
Learning
10
5
5
0
x2
-5
-5
-10
Initial Population
x1
x2
f (x1 , x2 )
2.61
7.15
6.66
-2.58
-0.02
0.04
0.07
-7.83
.
.
.
-1.71
4.90
-6.05
1.50
-0.52
0.18
-3.22
-1.79
1.97
.
.
.
4.35
30.95
87.79
46.74
6.94
0.03
10.37
3.21
65.28
.
.
.
21.86
Selected Population
Selection
x1
x2
f (x1 , x2 )
1.27
-0.18
1.4
0.72
2.97
-1.02
0.78
0.52
.
.
.
1.28
1.61
0.07
3.41
10.56
8.86
4.74
6.96
12.25
.
.
.
2.24
Simulated Population
x1
x2
f (x1 , x2 )
0.06
0.33
0.05
0.19
-0.02
0.53
0.71
0.58
.
.
.
0.69
-0.56
-0.47
0.29
-0.10
-0.07
0.49
0.60
-0.09
.
.
.
-0.33
0.32
0.33
0.09
0.04
0.00
0.52
0.87
0.34
.
.
.
0.59
x2
f (x1 , x2 )
-0.02
0.00
-1.42
0.07
2.23
-2.58
1.56
0.04
-2.16
0.18
0.47
0.48
-1.79
0.31
-0.52
-2.72
-3.22
2.66
0.03
0.22
2.27
3.21
5.11
6.94
9.87
10.37
11.74
µ1 = −0.25, σ1 = 1.59
µ2 = −0.46, σ2 = 1.83
Sampling
Learning
10
5
5
0
x2
Selected Population
Selection
x1
x2
f (x1 , x2 )
-0.19
0.45
0.94
0.41
0.50
-0.01
0.27
1.09
1.13
-0.18
-0.01
0.01
1.18
-1.15
1.27
-1.38
0.98
-0.94
0.07
0.21
0.89
1.56
1.60
1.61
1.98
2.16
2.18
Selected Population
x1
Search Distribution
10
0
-5
-5
-10
Simulated Population
-0.01
-0.19
-1.18
-3.16
-0.16
-1.92
2.52
3.46
.
.
.
-0.77
x1
10
0
x1
Search Distribution
µ1 = 0.51, σ1 = 0.46
µ2 = −0.02, σ2 = 1.01
Sampling
Learning
10
5
5
0
x2
10
0
-5
-5
-10
x1
Search Distribution
Simulated Population
Figure 1: Steps performed to minimize the function f (x1 , x2 ) = x21 + x22 using a continuous
EDA that assumes independence between the variables. The search distribution models each
variable with the normal distribution and, since mutual independence is assumed, the joint
PDF is factorized as φ1,2 (x1 , x2 ) = φ1 (x1 , µ1 , σ12 )φ2 (x2 , µ2 , σ22 ). The simulation of the probabilistic model estimated at the third generation produces a solution f (−0.02, −0.07) = 0.00
that approximates the global optimum of the function with the required precision.
6
copulaedas: An R Package for EDAs Based on Copulas
• A selection method is used to determine the most promising solutions of the population. An example selection method is truncation selection, which creates the selected
population with a percentage of the best solutions of the current population.
• The estimation and simulation of the search distribution are the essential steps of an
EDA. These steps are implemented by learning and sampling methods, which are tightly
related. Learning methods estimate the structure and parameters of the probabilistic
model used by the algorithm from the selected population, while sampling methods are
used to generate new solutions from the estimated probabilistic model.
• A replacement method is used to incorporate a new group of solutions into the current
population. For example, a replacement strategy is to substitute the current population with the newly sampled population. Other replacement strategies retain the best
solutions found so far or try to maintain the diversity of solutions.
• Reporting methods provide progress information during the execution of the EDA. Relevant progress information can be the number of evaluations of the objective function
and the best solution found so far.
• A termination method determines when the algorithm stops according to certain criteria;
for example, when a fixed number of function evaluations are realized or a certain value
of the objective function is reached.
Although it was possible to locate with the required precision the optimum of the simple
function presented in this section, it is not always possible to perform a successful search by
considering only the marginal information. As we show later in this paper, the assumption
of independence between the variables constitutes a strong limitation that may compromise
the convergence of an EDA. The use of information about the relationships between the
variables allows searching efficiently for the best solutions and it constitutes one of the main
advantages of EDAs. Among the algorithms that consider dependences between the variables,
we are particularly interested in EDAs whose learning and sampling steps involve probabilistic
models based on copulas. The next section provides an overview of such algorithms.
2.2. Overview of EDAs based on copulas
To the best of our knowledge, the technical report (Soto et al. 2007) and the theses (Arderı́
2007; Barba-Moreno 2007) constitute the first attempts to incorporate copulas into EDAs.
Since then, a considerable number of EDAs based on copula theory have been proposed in
the literature and, as evidence of its increasing popularity, the use of copulas in EDAs has
been identified as an emerging approach for the solution of real-valued optimization problems
(Hauschild and Pelikan 2011).
The learning step of copula-based EDAs consists of two tasks: the estimation of the marginal
distributions and the estimation of the probabilistic dependence structure. In general, these
tasks have been performed by following one of the two-step estimation procedures known in
the copula literature as the IFM (inference functions for margins; Joe and Xu 1996; Joe 2005)
and the semiparametric estimation method (Genest et al. 1995). Firstly, the marginal distributions are estimated and the selected population is transformed into uniformly distributed
variables in (0, 1) by means of the evaluation of each marginal cumulative distribution function. Secondly, the transformed population is used to estimate a copula-based model of the
Yasser Gonzalez-Fernandez, Marta Soto
7
dependence structure among the variables. Usually, a particular parametric distribution (e.g.,
normal or beta) is assumed for each margin and its parameters are estimated by maximum
likelihood (see e.g., Soto et al. 2007; Salinas-Gutiérrez et al. 2009). In other cases, empirical
marginal distributions or kernel density estimation have been used (see e.g., Soto et al. 2007;
Gao 2009; Cuesta-Infante et al. 2010). The simulation step typically starts with the generation of a population of uniformly distributed variables in (0, 1) with the dependence structure
described by the copula-based model that was estimated in the learning step. Finally, this
uniform population is transformed to the domain of the variables through the evaluation of
the inverse of each marginal cumulative distribution function.
According to the copula model being used, EDAs based on copulas can be classified as EDAs
based on either multivariate or factorized copulas. In the rest of this section we give an overall
description of representative algorithms belonging to each group that have been proposed in
the literature.
EDAs based on multivariate copulas
The research on EDAs based on multivariate copulas has focused on the use of multivariate
elliptical copulas (Abdous et al. 2005; Fang et al. 2002) and Archimedean copulas (Joe 1997;
McNeil and Nešlehová 2009). The algorithms described in Soto et al. (2007); Arderı́ (2007)
and Barba-Moreno (2007) are both based on the multivariate normal copula and theoretically similar, but they present differences in the estimation of the marginal distributions and
the use of techniques such as variance scaling. Wang et al. (2009b) present an EDA based
on the bivariate normal copula and, since only normal marginal distributions are used, the
proposed algorithm is equivalent to EMNA (estimation of multivariate normal algorithm;
Larrañaga et al. 2001). On the other hand, the algorithms presented in Wang et al. (2009a)
and Gao (2009) use exchangeable Archimedean copulas. Wang et al. (2009a) propose two
algorithms that use Clayton and Ali-Mikhail-Haq copulas with fixed parameters, while Gao
(2009) does not state which particular members of the family of Archimedean copulas are
used.
Two EDAs based on multivariate copulas are implemented in copulaedas, one is based on the
product or independence copula and the other on the normal copula. The first algorithm is
UMDA (univariate marginal distribution algorithm) for continuous variables (Larrañaga et al.
1999, 2000), which can be integrated into the framework of copula-based EDAs although
originally it was not defined in terms of copulas. A consequence of Sklar’s Theorem is that
random variables are independent if and only if the underlying copula is the product copula.
Thus, UMDA can be described as an EDA that models the dependence structure between
the variables using a multivariate product copula.
The second EDA based on a multivariate copula implemented in copulaedas is GCEDA (Gaussian copula estimation of distribution algorithm; Soto et al. 2007; Arderı́ 2007). This algorithm is based on the multivariate normal copula, which allows the construction of multivariate
distributions with normal dependence structure and non-normal margins. The dependence
structure of the multivariate normal copula is determined by a positive-definite correlation
matrix. If the marginal distributions are not normal, the correlation matrix is estimated
through the inversion of the non-parametric estimator of Kendall’s tau for each pair of variables (see e.g., Genest and Favre 2007; Hult and Lindskog 2002). If the resulting matrix is
not positive-definite, the transformation proposed by Rousseeuw and Molenberghs (1993) is
8
copulaedas: An R Package for EDAs Based on Copulas
applied. GCEDA is equivalent to EMNA when all the marginal distributions are normal.
EDAs based on copula factorizations
The use of multivariate copulas to model the dependence structure between the variables offers
a considerable number of advantages over the use of the multivariate normal distribution;
nevertheless, it presents limitations. The number of tractable copulas available when more
than two variables are involved is limited, most available copulas are just investigated in the
bivariate case. In addition, the multivariate elliptical copulas might not be appropriate when
all pairs of variables do not have the same dependence structure. Another limitation is that
some multivariate extensions, such as exchangeable Archimedean copulas or the multivariate
t copula, have only one parameter to describe certain aspects of the overall dependence. This
characteristic can be a serious limitation when the type and strength of the dependence is
not the same for all pairs of variables. One alternative to these limitations is to use copula
factorizations that build high-dimensional probabilistic models by using lower-dimensional
copulas as building blocks. Several EDAs based on copula factorizations, such as nested
Archimedean copulas (Joe 1997) and vines (Joe 1996; Bedford and Cooke 2001; Aas et al.
2009), have been proposed in the literature.
The EDA introduced in Salinas-Gutiérrez et al. (2009) is an extension of MIMIC (mutual
information maximization for input clustering) for continuous domains (Larrañaga et al. 1999,
2000) that uses bivariate copulas in a chain structure instead of bivariate normal distributions.
Two instances of this algorithm were presented, one uses normal copulas and the other Frank
copulas. In Section 4.2, we illustrate the implementation of this algorithm using copulaedas.
The exchangeable Archimedean copulas employed in Wang et al. (2009a) and Gao (2009)
represent highly specialized dependence structures (Berg and Aas 2007; McNeil 2008). Within
the domain of Archimedean copulas, nested Archimedean copulas provide a more flexible
alternative to build multivariate copula distributions. In particular, hierarchically nested
Archimedean copulas present one of the most flexible solutions among the different nesting
structures that have been studied (see e.g., Berg and Aas 2007 for a review). Building from
these models, Ye et al. (2010) propose an EDA that uses a representation of hierarchically
nested Archimedean copulas based on Lévy subordinators (Hering et al. 2010).
Cuesta-Infante et al. (2010) investigate the use of bivariate empirical copulas and a multivariate extension of Archimedean copulas. The EDA based on bivariate empirical copulas
is completely nonparametric: it employs empirical marginal distributions and a construction
based on bivariate empirical copulas to represent the dependence between the variables. The
marginal distributions and the bivariate empirical copulas are defined through the linear interpolation of the sample in the selected population. The EDA based on Archimedean copulas
uses a construction similar to a fully nested Archimedean copula and uses copulas from one of
the families Frank, Clayton or HRT (i.e., heavy right tail copula or Clayton survival copula).
The parameters of the copulas are fixed to a constant value, i.e., not estimated from the selected population. The marginal distributions are modeled as in the EDA based on bivariate
empirical copulas.
The class of VEDAs (vine EDAs) is introduced in Soto and Gonzalez-Fernandez (2010) and
Gonzalez-Fernandez (2011). Algorithms of this class model the search distributions using regular vines, which are graphical models that represent a multivariate distribution by decomposing the corresponding multivariate density into conditional bivariate copulas, unconditional
Yasser Gonzalez-Fernandez, Marta Soto
9
bivariate copulas and univariate densities. In particular, VEDAs are based on the simplified
pair-copula construction (Haff et al. 2010), which assumes that the bivariate copulas depend
on the conditioning variables only through their arguments. Since all bivariate copulas do
not have to belong to the same family, regular vines model a rich variety of dependences by
combining bivariate copulas from different families.
A regular vine on n variables is a set of nested trees T1 , . . . , Tn−1 , where the edges of tree Tj
are the nodes of the tree Tj+1 with j = 1, . . . , n − 2. The edges of the trees represent
the bivariate copulas in the decomposition and the nodes their arguments. Moreover, the
proximity condition requires that two nodes in tree Tj+1 are joined by an edge only if the
corresponding edges in Tj share a common node. C-vines (canonical vines) and D-vines
(drawable vines) are two particular types of regular vines, each of which determines a specific
decomposition of the multivariate density. In a C-vine, each tree Tj has a unique root node
that is connected to n − j edges. In a D-vine, no node is connected to more than two edges.
Two EDAs based on regular vines are presented in Soto and Gonzalez-Fernandez (2010) and
Gonzalez-Fernandez (2011): CVEDA (C-vine EDA) and DVEDA (D-vine EDA) based on
C-vines and D-vines, respectively. Since both algorithms are implemented in copulaedas, we
describe them in more detail in the rest of this section.
The general idea of the simulation and inference methods for C-vines and D-vines was developed by Aas et al. (2009). The simulation algorithm is based on the conditional distribution
method (see e.g., Devroye 1986), while the inference method should consider two main aspects:
the selection of the structure of the vines and the choice of the bivariate copulas. In the rest
of this section we describe how these aspects are performed in the particular implementation
of CVEDA and DVEDA.
The selection of the structure of C-vines and D-vines is restricted to the selection of the
bivariate dependences explicitly modeled in the first tree. This is accomplished by using
greedy heuristics, which use the empirical Kendall’s tau assigned to the edges of the tree. In
a C-vine, the node that maximizes the sum of the weights of its edges to the other nodes is
chosen as the root of the first tree and a canonical root node is assumed for the rest of the
trees. In a D-vine, the construction of the first tree consists of finding the maximum weighted
sequence of the variables, which can be transformed into a TSP (traveling salesman problem)
instance (Brechmann 2010). For efficiency reasons, in copulaedas we find an approximate
solution of the TSP by using the cheapest insertion heuristic (Rosenkrantz et al. 1977).
The selection of each bivariate copula in both CVEDA and DVEDA starts with an independence test (Genest and Rémillard 2004; Genest et al. 2007). The product copula is selected
when there is not enough evidence against the null hypothesis of independence at a given
significance level. Otherwise, the parameters of a group of candidate copulas are estimated
and the copula that minimizes a Cramér-von Mises statistic of the empirical copula is selected
(Genest and Rémillard 2008).
The cost of the construction of C-vines and D-vines increases with the number of variables.
To reduce this cost, we apply the truncation strategy presented in Brechmann (2010), for
which the theoretical justification can be found in Joe et al. (2010). When a vine is truncated
at a certain tree during the tree-wise estimation procedure, all the copulas in the subsequent
trees are assumed to be product copulas. A model selection procedure based on either AIC
(Akaike information criterion; Akaike 1974) or BIC (Bayesian information criterion; Schwarz
1978) is applied to detect the required number of trees. This procedure expands the tree Tj+1
10
copulaedas: An R Package for EDAs Based on Copulas
if the value of the information criterion calculated up to the tree Tj+1 is smaller than the
value obtained up to the previous tree; otherwise, the vine is truncated at the tree Tj . At this
point, it is important to note that the algorithm presented in Salinas-Gutiérrez et al. (2010)
also uses a D-vine. In this algorithm only normal copulas are fitted in the first two trees
and conditional independence is assumed in the rest of the trees, i.e., the D-vine is always
truncated at the second tree.
The implementation of CVEDA and DVEDA included in copulaedas uses by default the
truncation procedure based on AIC and the candidate copulas normal, t, Clayton, Frank and
Gumbel. The parameters of all copulas but the t copula are estimated using the method of
moments. For the t copula, the correlation coefficient is computed as in the normal copula, and
the degrees of freedom are estimated by maximum likelihood with the correlation parameter
fixed (Demarta and McNeil 2005).
3. Implementation in R
According to the review presented by Santana (2011), the approach followed for the implementation of EDA software currently available through the Internet can be classified into three
categories: (1) implementation of a single EDA, (2) independent implementation of multiple
EDAs, and (3) common modular implementation of multiple EDAs. In our opinion, the third
approach offers greater flexibility for the EDA community. In these modular implementations,
the EDA components (e.g., learning and sampling methods) are independently programmed
by taking advantage of the common schema shared by most EDAs. This modularity allows the
creation and validation of new EDA proposals that combine different components, and promotes code factorization. Additionally, as the EDAs are grouped under the same framework,
it facilitates performing empirical studies to compare the behavior of different algorithms.
Existing members of this class are ParadisEO (Cahon et al. 2004; DOLPHIN Project Team
2012), LiO (Mateo and de la Ossa 2006, 2007), Mateda-2.0 (Santana et al. 2009, 2010) and
now copulaedas.
The implementation of copulaedas follows an object-oriented design inspired by the Mateda2.0 toolbox for MATLAB (The MathWorks, Inc. 2014). EDAs implemented in the package
are represented by S4 classes (Chambers 2008) with generic functions for their main steps.
The base class of EDAs in the package is ‘EDA’, which has two slots: name and parameters.
The name slot stores a character string with the name of the EDA and it is used by the
show method to print the name of the algorithm when it is called with an ‘EDA’ instance as
argument. The parameters slot stores all the EDA parameters in a list.
In copulaedas, each step of the general procedure of an EDA outlined in Algorithm 1 is represented by a generic function that expects an ‘EDA’ instance as its first argument. Table 1 shows
a description of these functions and their default methods. The help page of these generic
functions in the documentation of copulaedas contains information about their arguments,
return value, and methods already implemented in the package.
The generic functions and their methods that implement the steps of an EDA look at the
parameters slot of the ‘EDA’ instance received as first argument for the values of the parameters that affect their behavior. Only named members of the list must be used and reasonable
default values should be assumed when a certain component is missing. The help page of
each generic function describes the members of the list in the parameters slot interpreted by
Yasser Gonzalez-Fernandez, Marta Soto
Generic function
edaSeed
edaOptimize
edaSelect
edaLearn
edaSample
edaReplace
edaReport
edaTerminate
11
Description
Seeding method. The default method edaSeedUniform generates the
values of each variable in the initial population from a continuous uniform distribution.
Local optimization method. The use of a local optimization method is
disabled by default.
Selection method. The default method edaSelectTruncation implements truncation selection.
Learning method. No default method.
Sampling method. No default method.
Replacement method.
The default method edaReplaceComplete
completely replaces the current population with the new population.
Reporting method. Reporting progress information is disabled by
default.
Termination method. The default method edaTerminateMaxGen ends
the execution of the algorithm after a maximum number of generations.
Table 1: Description of the generic functions that implement the steps of the general procedure
of an EDA outlined in Algorithm 1 and their default methods.
Slot
eda
f
lower
upper
numGens
fEvals
bestSol
bestEval
cpuTime
Description
‘EDA’ instance.
Objective function.
Lower bounds of the variables of the objective function.
Upper bounds of the variables of the objective function.
Total number of generations.
Total number of evaluations of the objective function.
Best solution.
Evaluation of the best solution.
Run time of the algorithm in seconds.
Table 2: Description of the slots of the ‘EDAResult’ class.
each function and their default values.
The edaRun function implements the Algorithm 1 by linking together the generic functions for
each step. This function expects four arguments: the ‘EDA’ instance, the objective function
and two vectors specifying the lower and upper bounds of the variables of the objective
function. The length of the vectors with the lower and upper bounds should be the same,
since it determines the number of variables of the objective function. When edaRun is called,
it runs the main loop of the EDA until the call to the edaTerminate generic function returns
TRUE. Then, the function returns an instance of the ‘EDAResult’ class that encapsulates the
results of the algorithm. A description of the slots of this class is given in Table 2.
Two subclasses of ‘EDA’ are already defined in copulaedas: ‘CEDA’, that represents EDAs
based on multivariate copulas; and ‘VEDA’, that represents vine-based EDAs. The implementation of UMDA, GCEDA, CVEDA and DVEDA relies on the copula (Hofert et al. 2014),
vines (Gonzalez-Fernandez and Soto 2014b), mvtnorm (Genz et al. 2014; Genz and Bretz
12
copulaedas: An R Package for EDAs Based on Copulas
2009), and truncnorm (Trautmann et al. 2014) R packages. These packages implement the
techniques for the estimation and simulation of the probabilistic models used in these EDAs.
4. Using copulaedas
In this section, we illustrate how to use copulaedas through several examples. To begin
with, we show how to run the EDAs included in the package. Next, we continue with the
implementation of a new EDA by using the functionalities provided by the package, and
finally we show how to perform an empirical study to compare the behavior of a group of
EDAs on benchmark functions and a real-world problem.
The two well-known test problems Sphere and Summation Cancellation are used as the benchmark functions. The functions fSphere and fSummationCancellation implement these problems in terms of a vector x = (x1 , . . . , xn ) according to
fSphere (x) =
n
X
x2i ,
i=1
fSummation Cancellation (x) =
10−5
+
1
Pn
i=1 |yi |
,
y1 = x1 , yi = yi−1 + xi .
Sphere is a minimization problem and Summation Cancellation is originally a maximization problem but it is defined in the package as a minimization problem. Sphere has its
global optimum at x = (0, . . . , 0) with evaluation zero and Summation Cancellation at
x = (0, . . . , 0) with evaluation −105 . For a description of the characteristics of these functions
see Bengoetxea et al. (2002) and Bosman and Thierens (2006).
The results presented in this section were obtained using R version 3.1.0 with copulaedas
version 1.4.0, copula version 0.999-8, vines version 1.1.0, mvtnorm version 0.9-99991, and
truncnorm version 1.0-7. Computations were performed on a 64-bit Linux machine with an
Intel(R) Core(TM)2 Duo 2.00 GHz processor. In the rest of this section, we assume copulaedas
has been loaded. This can be attained by running the following command:
R> library("copulaedas")
4.1. Running the EDAs included in the package
We begin by illustrating how to run the EDAs based on copulas implemented in copulaedas.
As an example, we execute GCEDA to optimize Sphere in five dimensions. Before creating a
new instance of the ‘CEDA’ class for EDAs based on multivariate copulas, we set up the generic
functions for the steps of the EDA according to the expected behavior of GCEDA. The termination criterion is either to find the optimum of the objective function or to reach a maximum
number of generations. That is why we set the method for the edaTerminate generic function
to a combination of the functions edaTerminateEval and edaTerminateMaxGen through the
auxiliary function edaTerminateCombined.
R> setMethod("edaTerminate", "EDA",
+
edaTerminateCombined(edaTerminateEval, edaTerminateMaxGen))
Yasser Gonzalez-Fernandez, Marta Soto
13
The method for the edaReport generic function is set to edaReportSimple to make the
algorithm print progress information at each generation. This function prints one line at each
iteration of the EDA with the minimum, mean and standard deviation of the evaluation of
the solutions in the current population.
R> setMethod("edaReport", "EDA", edaReportSimple)
Note that these methods were set for the base class ‘EDA’ and therefore they will be inherited
by all subclasses. Generally, we find it convenient to define methods of the generic functions
that implement the steps of the EDA for the base class, except when different subclasses
should use different methods.
The auxiliary function ‘CEDA’ can be used to create instances of the class with the same
name. All the arguments of the function are interpreted as parameters of the EDA to be
added as members of the list in the parameters slot of the new instance. An instance
of ‘CEDA’ corresponding to GCEDA using empirical marginal distributions smoothed with
normal kernels can be created as follows:
R> gceda <- CEDA(copula = "normal", margin = "kernel", popSize = 200,
+
fEval = 0, fEvalTol = 1e-6, maxGen = 50)
R> gceda@name <- "Gaussian Copula Estimation of Distribution Algorithm"
The methods that implement the generic functions edaLearn and edaSample for ‘CEDA’ instances expect three parameters. The copula parameter specifies the multivariate copula
and it should be set to "normal" for GCEDA. The marginal distributions are determined
by the value of margin and all EDAs implemented in the package use this parameter for the
same purpose. As margin is set to "kernel", the algorithm will look for three functions
named fkernel, pkernel and qkernel already defined in the package to fit the parameters
of the margins and to evaluate the distribution and quantile functions, respectively. The
fkernel function computes the bandwidth parameter of the normal kernel according to the
rule-of-thumb of Silverman (1986) and pkernel implements the empirical cumulative distribution function. The quantile function is evaluated following the procedure described in
Azzalini (1981). The popSize parameter determines the population size while the rest of the
arguments of CEDA are parameters of the functions that implement the termination criterion.
Now, we can run GCEDA by calling edaRun. The lower and upper bounds of the variables
are set so that the values of the variables in the optimum of the function are located at
25% of the interval. It was shown in Arderı́ (2007) and Soto et al. (in press) that the use
of empirical marginal distributions smoothed with normal kernels improves the behavior of
GCEDA when the initial population is generated asymmetrically with respect to the optimum
of the function.
R> set.seed(12345)
R> result <- edaRun(gceda, fSphere, rep(-300, 5), rep(900, 5))
Generation
Minimum
Mean
Std. Dev.
1 1.522570e+05 1.083606e+06 5.341601e+05
2 2.175992e+04 5.612769e+05 3.307403e+05
3 8.728486e+03 2.492247e+05 1.496334e+05
14
copulaedas: An R Package for EDAs Based on Copulas
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
4.536507e+03
5.827775e+03
2.402107e+03
9.170485e+02
4.591915e+02
2.448265e+02
7.727107e+01
4.601731e+01
8.555769e+00
1.865639e+00
5.157326e+00
1.788793e+00
7.418832e-01
6.223596e-01
4.520045e-02
6.981399e-02
3.440069e-02
1.370064e-02
3.050774e-03
1.367716e-03
7.599946e-04
4.009605e-04
1.083879e-04
8.441887e-05
3.429462e-05
1.999004e-05
1.038719e-05
6.297005e-06
1.034002e-06
8.483830e-07
1.025119e+05
5.126260e+04
2.527349e+04
1.312806e+04
6.726731e+03
3.308515e+03
1.488859e+03
6.030030e+02
2.381415e+02
1.000919e+02
4.404530e+01
2.195864e+01
1.113184e+01
4.880880e+00
2.327805e+00
1.123582e+00
5.118243e-01
1.960786e-01
7.634156e-02
3.400907e-02
1.461478e-02
6.488932e-03
2.625759e-03
1.079075e-03
5.077934e-04
2.232605e-04
1.104123e-04
5.516721e-05
2.537823e-05
1.332463e-05
5.829982e+04
2.983622e+04
1.430142e+04
6.815822e+03
4.150888e+03
1.947486e+03
8.567864e+02
3.529036e+02
1.568382e+02
6.078611e+01
2.413589e+01
1.136284e+01
6.157461e+00
2.723950e+00
1.287697e+00
6.956201e-01
2.985175e-01
1.329600e-01
4.453917e-02
2.056747e-02
8.861180e-03
4.043431e-03
1.618058e-03
5.759307e-04
3.055568e-04
1.198675e-04
5.888948e-05
2.945027e-05
1.295004e-05
7.399488e-06
The result variable contains an instance of the ‘EDAResult’ class. The show method prints
the results of the execution of the algorithm.
R> show(result)
Results for Gaussian Copula
Best function evaluation
No. of generations
No. of function evaluations
CPU time
Estimation of Distribution Algorithm
8.48383e-07
33
6600
7.895 seconds
Due to the stochastic nature of EDAs, it is often useful to analyze a sequence of independent
runs to ensure reliable results. The edaIndepRuns function supports performing this task.
To avoid generating lot of unnecessary output, we first disable reporting progress information
on each generation by setting edaReport to edaReportDisabled and then we invoke the
edaIndepRuns function to perform 30 independent runs of GCEDA.
Yasser Gonzalez-Fernandez, Marta Soto
15
R> setMethod("edaReport", "EDA", edaReportDisabled)
R> set.seed(12345)
R> results <- edaIndepRuns(gceda, fSphere, rep(-300, 5), rep(900, 5), 30)
The return value of the edaIndepRuns function is an instance of the ‘EDAResults’ class. This
class is simply a wrapper for a list with instances of ‘EDAResult’ as members that contain
the results of an execution of the EDA. A show method for ‘EDAResults’ instances prints a
table with all the results.
R> show(results)
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
Run
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
Generations Evaluations Best Evaluation CPU Time
33
6600
8.483830e-07
7.583
38
7600
4.789448e-08
8.829
36
7200
4.798364e-07
8.333
37
7400
9.091651e-07
8.772
34
6800
5.554465e-07
7.830
35
7000
3.516341e-07
8.071
35
7000
9.325531e-07
8.106
35
7000
6.712550e-07
8.327
36
7200
8.725061e-07
8.283
37
7400
2.411458e-07
8.565
36
7200
4.291725e-07
8.337
35
7000
7.245520e-07
8.313
37
7400
2.351322e-07
8.538
36
7200
8.651248e-07
8.320
34
6800
9.422646e-07
7.821
35
7000
9.293726e-07
8.333
36
7200
6.007390e-07
8.274
38
7600
3.255231e-07
8.763
36
7200
6.012969e-07
8.353
35
7000
5.627017e-07
8.296
36
7200
5.890752e-07
8.259
35
7000
9.322505e-07
8.067
35
7000
4.822349e-07
8.084
34
6800
7.895408e-07
7.924
36
7200
6.970180e-07
8.519
34
6800
3.990247e-07
7.808
35
7000
8.876874e-07
8.055
33
6600
8.646387e-07
7.622
36
7200
9.072113e-07
8.519
35
7000
9.414666e-07
8.040
Also, the summary method can be used to generate a table with a statistical summary of the
results of the 30 runs of the algorithm.
R> summary(results)
16
copulaedas: An R Package for EDAs Based on Copulas
Generations Evaluations Best Evaluation
Minimum
33.000000
6600.0000
4.789448e-08
Median
35.000000
7000.0000
6.841365e-07
Maximum
38.000000
7600.0000
9.422646e-07
Mean
35.433333
7086.6667
6.538616e-07
Std. Dev.
1.250747
250.1494
2.557100e-07
CPU Time
7.5830000
8.2895000
8.8290000
8.2314667
0.3181519
4.2. Implementation of a new EDA based on copulas
In this section we illustrate how to use copulaedas to implement a new EDA based on copulas.
As an example, we consider the extension of MIMIC for continuous domains proposed in
Salinas-Gutiérrez et al. (2009). Similarly to MIMIC, this extension learns a chain dependence
structure, but it uses bivariate copulas instead of bivariate normal distributions. The chain
dependence structure is similar to a D-vine truncated at the first tree, i.e., a D-vine where
independence is assumed for all the trees but the first. Two instances of the extension of
MIMIC based on copulas were presented in Salinas-Gutiérrez et al. (2009), one uses bivariate
normal copulas while the other uses bivariate Frank copulas. In this article, the algorithm
will be denoted as Copula MIMIC.
Since the algorithm in question matches the general schema of an EDA presented in Algorithm 1, only the functions corresponding to the learning and simulation steps have to be
implemented. The first step in the implementation of a new EDA is to define a new S4 class
that inherits from ‘EDA’ to represent the algorithm. For convenience, we also define an auxiliary function CopulaMIMIC that can be used to create new instances of this class.
R> setClass("CopulaMIMIC", contains = "EDA",
+
prototype = prototype(name = "Copula MIMIC"))
R> CopulaMIMIC <- function(...)
+
new("CopulaMIMIC", parameters = list(...))
Copula MIMIC models the marginal distributions with the beta distribution. A linear transformation is used to map the sample of the variables in the selected population into the (0, 1)
interval to match the domain of definition of the beta distribution. Note that, since the copula
is scale-invariant, this transformation does not affect the dependence between the variables.
To be consistent with the margins already implemented in copulaedas, we define three functions with the common suffix betamargin and the prefixes f, p and q to fit the parameters
of the margins and for the evaluation of the distribution and quantile functions, respectively.
By following this convention, the algorithms already implemented in the package can use beta
marginal distributions by setting the margin parameter to "betamargin".
R> fbetamargin <- function(x, lower, upper) {
+
x <- (x - lower) / (upper - lower)
+
loglik <- function(s) sum(dbeta(x, s[1], s[2], log = TRUE))
+
s <- optim(c(1, 1), loglik, control = list(fnscale = -1))$par
+
list(lower = lower, upper = upper, a = s[1], b = s[2])
+ }
R> pbetamargin <- function(q, lower, upper, a, b) {
Yasser Gonzalez-Fernandez, Marta Soto
17
+
q <- (q - lower) / (upper - lower)
+
pbeta(q, a, b)
+ }
R> qbetamargin <- function(p, lower, upper, a, b) {
+
q <- qbeta(p, a, b)
+
lower + q * (upper - lower)
+ }
The ‘CopulaMIMIC’ class inherits methods for the generic functions that implement all the
steps of the EDA except learning and sampling. To complete the implementation of the
algorithm, we must define the estimation and simulation of the probabilistic model as methods
for the generic functions edaLearn and edaSample, respectively.
The method for edaLearn starts with the estimation of the parameters of the margins and
the transformation of the selected population to uniform variables in (0, 1). Then, the mutual information between all pairs of variables is calculated through the copula entropy
(Davy and Doucet 2003). To accomplish this, the parameters of each possible bivariate copula are estimated by the method of maximum likelihood using the value obtained through
the method of moments as an initial approximation. To determine the chain dependence
structure, a permutation of the variables that maximizes the pairwise mutual information
must be selected but, since this is a computationally intensive task, a greedy algorithm is
used to compute an approximate solution (Bonet et al. 1997; Larrañaga et al. 1999). Finally,
the method for edaLearn returns a list with three components that represents the estimated
probabilistic model: the parameters of the marginal distributions, the permutation of the
variables, and the copulas in the chain dependence structure.
R> edaLearnCopulaMIMIC <- function(eda, gen, previousModel,
+
selectedPop, selectedEval, lower, upper) {
+
margin <- eda@parameters$margin
+
copula <- eda@parameters$copula
+
if (is.null(margin)) margin <- "betamargin"
+
if (is.null(copula)) copula <- "normal"
+
fmargin <- get(paste("f", margin, sep = ""))
+
pmargin <- get(paste("p", margin, sep = ""))
+
copula <- switch(copula,
+
normal = normalCopula(0), frank = frankCopula(0))
+
n <- ncol(selectedPop)
+
# Estimate the parameters of the marginal distributions.
+
margins <- lapply(seq(length = n),
+
function(i) fmargin(selectedPop[, i], lower[i], upper[i]))
+
uniformPop <- sapply(seq(length = n), function(i) do.call(pmargin,
+
c(list(selectedPop[ , i]), margins[[i]])))
+
# Calculate pairwise mutual information by using copula entropy.
+
C <- matrix(list(NULL), nrow = n, ncol = n)
+
I <- matrix(0, nrow = n, ncol = n)
+
for (i in seq(from = 2, to = n)) {
+
for (j in seq(from = 1, to = i - 1)) {
+
# Estimate the parameters of the copula.
18
copulaedas: An R Package for EDAs Based on Copulas
+
data <- cbind(uniformPop[, i], uniformPop[, j])
+
startCopula <- fitCopula(copula, data, method = "itau",
+
estimate.variance = FALSE)@copula
+
C[[i, j]] <- tryCatch(
+
fitCopula(startCopula, data, method = "ml",
+
start = startCopula@parameters,
+
estimate.variance = FALSE)@copula,
+
error = function(error) startCopula)
+
# Calculate mutual information.
+
if (is(C[[i, j]], "normalCopula")) {
+
I[i, j] <- -0.5 * log(1 - C[[i, j]]@parameters^2)
+
} else {
+
u <- rcopula(C[[i, j]], 100)
+
I[i, j] <- sum(log(dcopula(C[[i, j]], u))) / 100
+
}
+
C[[j, i]] <- C[[i, j]]
+
I[j, i] <- I[i, j]
+
}
+
}
+
# Select a permutation of the variables.
+
perm <- as.vector(arrayInd(which.max(I), dim(I)))
+
copulas <- C[perm[1], perm[2]]
+
I[perm, ] <- -Inf
+
for (k in seq(length = n - 2)) {
+
ik <- which.max(I[, perm[1]])
+
perm <- c(ik, perm)
+
copulas <- c(C[perm[1], perm[2]], copulas)
+
I[ik, ] <- -Inf
+
}
+
list(margins = margins, perm = perm, copulas = copulas)
+ }
R> setMethod("edaLearn", "CopulaMIMIC", edaLearnCopulaMIMIC)
The edaSample method receives the representation of the probabilistic model returned by
edaLearn as the model argument. The generation of a new solution with n variables starts
with the simulation of an n-dimensional vector U having uniform marginal distributions in
(0, 1) and the dependence described by the copulas in the chain dependence structure. The
first step is to simulate an independent uniform variable Uπn in (0, 1), where πn denotes the
variable in the position n of the permutation π selected by the edaLearn method. The rest of
the uniform variables are simulated conditionally on the previously simulated variable by using
the conditional copula C(Uπk |Uπk+1 ), with k = n − 1, n − 2, . . . , 1. Finally, the new solution is
determined through the evaluation of the beta quantile functions and the application of the
inverse of the linear transformation.
R> edaSampleCopulaMIMIC <- function(eda, gen, model, lower, upper) {
+
popSize <- eda@parameters$popSize
+
margin <- eda@parameters$margin
Yasser Gonzalez-Fernandez, Marta Soto
19
+
if (is.null(popSize)) popSize <- 100
+
if (is.null(margin)) margin <- "betamargin"
+
qmargin <- get(paste("q", margin, sep = ""))
+
n <- length(model$margins)
+
perm <- model$perm
+
copulas <- model$copulas
+
# Simulate the chain structure with the copulas.
+
uniformPop <- matrix(0, nrow = popSize, ncol = n)
+
uniformPop[, perm[n]] <- runif(popSize)
+
for (k in seq(from = n - 1, to = 1)) {
+
u <- runif(popSize)
+
v <- uniformPop[, perm[k + 1]]
+
uniformPop[, perm[k]] <- hinverse(copulas[[k]], u, v)
+
}
+
# Evaluate the inverse of the marginal distributions.
+
pop <- sapply(seq(length = n), function(i) do.call(qmargin,
+
c(list(uniformPop[, i]), model$margins[[i]])))
+
pop
+ }
R> setMethod("edaSample", "CopulaMIMIC", edaSampleCopulaMIMIC)
The code fragments given above constitute the complete implementation of Copula MIMIC.
As it was illustrated with GCEDA in the previous section, the algorithm can be executed by
creating an instance of the ‘CopulaMIMIC’ class and calling the edaRun function.
4.3. Performing an empirical study on benchmark problems
We now show how to use copulaedas to perform an empirical study of the behavior of a
group of EDAs based on copulas on benchmark problems. The algorithms to be compared
are UMDA, GCEDA, CVEDA, DVEDA and Copula MIMIC. The first three algorithms are
included in copulaedas and the fourth algorithm was implemented in Section 4.2. The two
functions Sphere and Summation Cancellation described at the beginning of Section 4 are
considered as benchmark problems in 10 dimensions.
The aim of this empirical study is to assess the behavior of these algorithms when only linear
and independence relationships are considered. Thus, only normal and product copulas are
used in these EDAs. UMDA and GCEDA use multivariate product and normal copulas,
respectively. CVEDA and DVEDA are configured to combine bivariate product and normal
copulas in the vines. Copula MIMIC learns a chain dependence structure with normal copulas.
All algorithms use normal marginal distributions. Note that in this case, GCEDA corresponds
to EMNA and Copula MIMIC is similar to MIMIC for continuous domains. In the following
code fragment, we create class instances corresponding to these algorithms.
R>
R>
R>
R>
R>
umda <- CEDA(copula = "indep", margin = "norm")
umda@name <- "UMDA"
gceda <- CEDA(copula = "normal", margin = "norm")
gceda@name <- "GCEDA"
cveda <- VEDA(vine = "CVine", indepTestSigLevel = 0.01,
20
+
R>
R>
+
R>
R>
R>
copulaedas: An R Package for EDAs Based on Copulas
copulas = c("normal"), margin = "norm")
cveda@name <- "CVEDA"
dveda <- VEDA(vine = "DVine", indepTestSigLevel = 0.01,
copulas = c("normal"), margin = "norm")
dveda@name <- "DVEDA"
copulamimic <- CopulaMIMIC(copula = "normal", margin = "norm")
copulamimic@name <- "CopulaMIMIC"
The initial population is generated using the default edaSeed method, therefore, it is sampled
uniformly in the real interval of each variable. The lower and upper bounds of the variables are
set so that the values of the variables in the optimum of the function are located in the middle
of the interval. We use the intervals [−600, 600] in Sphere and [−0.16, 0.16] in Summation
Cancellation. All algorithms use the default truncation selection method with a truncation
factor of 0.3. Three termination criteria are combined using the edaTerminateCombined
function: to find the global optimum of the function with a precision greater than 10−6 , to
reach 300000 function evaluations, or to loose diversity in the population (i.e., the standard
deviation of the evaluation of the solutions in the population is less than 10−8 ). These
criteria are implemented in the functions edaTerminateEval, edaTerminateMaxEvals and
edaTerminateEvalStdDev, respectively.
The population size of EDAs along with the truncation method determine the sample available
for the estimation of the search distribution. An arbitrary selection of the population size
could lead to misleading conclusions of the results of the experiments. When the population
size is too small, the search distributions might not be accurately estimated. On the other
hand, the use of an excessively large population size usually does not result in a better behavior
of the algorithms but certainly in a greater number of function evaluations. Therefore, we
advocate for the use of the critical population size when comparing the performance of EDAs.
The critical population size is the minimum population size required by the algorithm to find
the global optimum of the function with a high success rate, e.g., to find the optimum in 30
of 30 sequential independent runs.
An approximate value of the critical population size can be determined empirically using a
bisection method (see e.g., Pelikan 2005 for a pseudocode of the algorithm). The bisection
method begins with an initial interval where the critical population size should be located
and discards one half of the interval at each step. This procedure is implemented in the
edaCriticalPopSize function. In the experimental study carried out in this section, the
initial interval is set to [50, 2000]. If the critical population size is not found in this interval,
the results of the algorithm with the population size given by the upper bound are presented.
The complete empirical study consists of performing 30 independent runs of every algorithm
on every function using the critical population size. We proceed with the definition of a list
containing all algorithm-function pairs.
R>
R>
R>
R>
+
+
+
edas <- list(umda, gceda, cveda, dveda, copulamimic)
fNames <- c("Sphere", "SummationCancellation")
experiments <- list()
for (eda in edas) {
for (fName in fNames) {
experiment <- list(eda = eda, fName = fName)
experiments <- c(experiments, list(experiment))
Yasser Gonzalez-Fernandez, Marta Soto
+
+
21
}
}
Now we define a function to process the elements of the experiments list. This function
implements all the experimental setup described before. The output of edaCriticalPopSize
and edaIndepRuns is redirected to a different plain text file for each algorithm-function pair.
R> runExperiment <- function(experiment) {
+
eda <- experiment$eda
+
fName <- experiment$fName
+
# Objective function parameters.
+
fInfo <- list(
+
Sphere = list(lower = -600, upper = 600, fEval = 0),
+
SummationCancellation = list(lower = -0.16, upper = 0.16,
+
fEval = -1e5))
+
lower <- rep(fInfo[[fName]]$lower, 10)
+
upper <- rep(fInfo[[fName]]$upper, 10)
+
f <- get(paste("f", fName, sep = ""))
+
# Configure termination criteria and disable reporting.
+
eda@parameters$fEval <- fInfo[[fName]]$fEval
+
eda@parameters$fEvalTol <- 1e-6
+
eda@parameters$fEvalStdDev <- 1e-8
+
eda@parameters$maxEvals <- 300000
+
setMethod("edaTerminate", "EDA",
+
edaTerminateCombined(edaTerminateEval, edaTerminateMaxEvals,
+
edaTerminateEvalStdDev))
+
setMethod("edaReport", "EDA", edaReportDisabled)
+
sink(paste(eda@name, "_", fName, ".txt", sep = ""))
+
# Determine the critical population size.
+
set.seed(12345)
+
results <- edaCriticalPopSize(eda, f, lower, upper,
+
eda@parameters$fEval, eda@parameters$fEvalTol, lowerPop = 50,
+
upperPop = 2000, totalRuns = 30, successRuns = 30,
+
stopPercent = 10, verbose = TRUE)
+
if (is.null(results)) {
+
# Run the experiment with the largest population size, if the
+
# critical population size was not found.
+
eda@parameters$popSize <- 2000
+
set.seed(12345)
+
edaIndepRuns(eda, f, lower, upper, runs = 30, verbose = TRUE)
+
}
+
sink(NULL)
+ }
We can run all the experiments by calling runExperiment for each element of the list.
R> for (experiment in experiments) runExperiment(experiment)
22
copulaedas: An R Package for EDAs Based on Copulas
Algorithm
Success
Sphere:
UMDA
30/30
GCEDA
30/30
CVEDA
30/30
DVEDA
30/30
Copula MIMIC 30/30
Summation Cancellation:
UMDA
0/30
GCEDA
30/30
CVEDA
30/30
DVEDA
30/30
Copula MIMIC
0/30
Pop.
81
310
104
111
172
Evaluations
3788.1
13102.6
4804.8
5080.1
7441.8
±
±
±
±
±
99.0
180.8
99.9
111.6
127.2
Best Evaluation
6.7e−07
6.8e−07
5.9e−07
6.5e−07
6.8e−07
±
±
±
±
±
2000 300000.0 ± 0.0
−7.2e+02 ±
325 38913.3 ± 268.9 −1.0e+05 ±
294 41228.6 ± 1082.7 −1.0e+05 ±
965 117022.3 ± 1186.8 −1.0e+05 ±
2000 300000.0 ± 0.0
−4.2e+04 ±
2.0e−07
2.0e−07
1.6e−07
2.3e−07
1.9e−07
CPU Time
0.2
0.4
4.5
4.4
65.6
±
±
±
±
±
4.5e+02
32.5 ±
1.0e−07
4.2 ±
1.0e−07 279.1 ±
1.1e−07 1617.5 ±
3.5e+04 1510.6 ±
0.0
0.0
0.8
0.8
1.2
0.3
0.1
5.1
39.1
54.9
Table 3: Summary of the results obtained in 30 independent runs of UMDA, GCEDA,
CVEDA, DVEDA and Copula MIMIC in the 10-dimensional Sphere (top) and Summation
Cancellation (bottom). Pop. denotes Population.
Running the complete empirical study sequentially is a computationally demanding operation.
If various processing units are available, it can be speeded up significantly by running the
experiments in parallel. The snow package (Tierney et al. 2013) offers a great platform to
achieve this purpose, since it provides a high-level interface for using a cluster of workstations
for parallel computations in R. The functions clusterApply or clusterApplyLB can be used
to call runExperiment for each element of the experiments list in parallel, with minimal
modifications to the code presented here.
A summary of the results of the algorithms with the critical population size is shown in
Table 3. Overall, the five algorithms are able to find the global optimum of Sphere in all the
30 independent runs with similar function values but only GCEDA, CVEDA and DVEDA
optimize Summation Cancellation. In the rest of this section we provide some comments
about results of the algorithms on each function.
UMDA exhibits the best behavior in terms of the number of function evaluations in Sphere.
There are no strong dependences between the variables of this function and the results suggest
that considering the marginal information is enough to find the global optimum efficiently. The
rest of the algorithms being tested require the calculation of a greater number of parameters to
represent the relationships between the variables and hence larger populations are needed to
compute them reliably (Soto et al. in press illustrate this issue in more detail with GCEDA).
CVEDA and DVEDA do not assume a normal dependence structure between the variables and
for this reason are less affected by this issue. The estimation procedure used by the vine-based
algorithms selects the product copula if there is not enough evidence of dependence.
Both UMDA and Copula MIMIC fail to optimize Summation Cancellation. A correct representation of the strong linear interactions between the variables of this function seems to be
essential to find the global optimum. UMDA completely ignores this information by assuming independence between the variables and it exhibits the worst behavior. Copula MIMIC
reaches better fitness values than UMDA but neither can find the optimum of the function.
The probabilistic model estimated by Copula MIMIC cannot represent important dependences
necessary for the success of the optimization. The algorithms GCEDA, CVEDA and DVEDA
Yasser Gonzalez-Fernandez, Marta Soto
23
do find the global optimum of Summation Cancellation. The results of GCEDA are slightly
better than the ones of CVEDA and these two algorithms achieve much better results than
DVEDA in terms of the number of function evaluations. The correlation matrix estimated
by GCEDA can properly represent the multivariate linear interactions between the variables.
The C-vine structure used in CVEDA, on the other hand, provides a very good fit for the
dependence structure between the variables of Summation Cancellation, given that it is possible to find a variable that governs the interactions in the sample (see Gonzalez-Fernandez
2011 for more details).
The running time of Copula MIMIC is considerably greater than the running time of the
other algorithms for all the functions. This result is due to the use of a numerical optimization algorithm for the estimation of the parameters of the copulas by maximum likelihood.
In the context of EDAs, where copulas are fitted at every generation, the computational
effort required to estimate the parameters of the copulas becomes an important issue. As
was illustrated with CVEDA and DVEDA, using a method of moments estimation is a viable alternative to maximum likelihood that requires much less CPU time. The empirical
investigation confirms the robustness of CVEDA and DVEDA in problems with both weak
and strong interactions between the variables. Nonetheless, the flexibility afforded by these
algorithms comes with an increased running time when compared to UMDA or GCEDA, since
the interactions between the variables have to be discovered during the learning step.
A general result of this empirical study is that copula-based EDAs should use copulas other
than the product only when there is evidence of dependence. Otherwise, the EDA will require
larger populations and hence a greater number of function evaluations to accurately determine
the parameters of the copulas that correspond to independence.
4.4. Solving the molecular docking problem
Finally, we illustrate the use of copulaedas for solving a so-called real-world problem. In particular, we use CVEDA and DVEDA to solve an instance of the molecular docking problem,
which is an important component of protein-structure-based drug design. From the point of
view of the computational procedure, it entails predicting the geometry of a small ligand that
binds to the active site of a large macromolecular protein receptor. Protein-ligand docking
remains being a highly active area of research, since the algorithms for exploring the conformational space and the scoring functions that have been implemented so far have significant
limitations (Warren et al. 2006).
In our docking simulations, the protein is treated as a rigid body while the ligand is fully
flexible. Thus, a candidate solution represents only the geometry of the ligand and it is
encoded as a vector of real values that represent its position, orientation and flexible torsion
angles. The first three variables of this vector represent the ligand position in the threedimensional space constrained to a box enclosing the receptor binding site. The construction
of this box is based on the minimum and maximum values of the ligand coordinates in its
crystal conformation plus a padding distance of 5Å added to each main direction of the space.
The remainder vector variables are three Euler angles that represent the ligand orientation
as a rigid body and take values in the intervals [0, 2π], [−π/2, π/2] and [−π, π], respectively;
and one additional variable restricted to [−π, π] for each flexible torsion angle of the ligand.
The semiempirical free-energy scoring function implemented as part of the suite of automated
docking tools AutoDock 4.2 (Morris et al. 2009) is used to evaluate each candidate ligand con-
24
copulaedas: An R Package for EDAs Based on Copulas
formation. The overall binding energy of a given ligand molecule is expressed as the sum of the
pairwise interactions between the receptor and ligand atoms (intermolecular interaction energy), and the pairwise interactions between the ligand atoms (ligand intramolecular energy).
The terms of the function consider dispersion/repulsion, hydrogen bonding, electrostatics, and
desolvation effects, all scaled empirically by constants determined through a linear regression
analysis. The aim of an optimization algorithm performing the protein-ligand docking is to
minimize the overall energy value. Further details of the energy terms and how the function
is derived can be found in Huey et al. (2007).
Specifically, we consider as an example here the docking of the 2z5u test system, solved
by X-ray crystallography and available as part of the Protein Data Bank (Berman et al.
2000). The protein receptor is lysine-specific histone demethylase 1 and the ligand is a
73-atom molecule (non-polar hydrogens are not counted) with 20 ligand torsions in a box
of 28Å×32Å×24Å. In order to make it easier for the readers of the paper to reproduce this
example, the implementation of the AutoDock 4.2 scoring function in C was extracted from
the original program and it is included (with unnecessary dependences removed) in the supplementary material as docking.c. During the evaluation of the scoring function, precalculated
grid maps (one for each atom type present in the ligand being docked) are used to make the
docking calculations fast. The result of these precalculations and related metadata for the
2z5u test system are contained in the attached ASCII file 2z5u.dat.
We make use of the system command R CMD SHLIB to build a shared object for dynamic
loading from the file docking.c. Next, we integrate the created shared object into R using the
function dyn.load and load the precalculated grids using the utility C function docking_load
as follows.
R> system("R CMD SHLIB docking.c")
R> dyn.load(paste("docking", .Platform$dynlib.ext, sep = ""))
R> .C("docking_load", as.character("2z5u.dat"))
The docking of the 2z5u test system results in a minimization problem with a total of 26
variables. Two vectors with the lower and upper bounds of these variables are defined using
utility functions provided in docking.c to compute the bounds of the variables that determine
the position of the ligand and the total number of torsions. For convenience, we also define
an R wrapper function fDocking for the C scoring function docking_score provided in the
compiled code that was loaded into R.
R> lower <- c(.C("docking_xlo", out = as.double(0))$out,
+
.C("docking_ylo", out = as.double(0))$out,
+
.C("docking_zlo", out = as.double(0))$out, 0, -pi/2, -pi,
+
rep(-pi, .C("docking_ntor", out = as.integer(0))$out))
R> upper <- c(.C("docking_xhi", out = as.double(0))$out,
+
.C("docking_yhi", out = as.double(0))$out,
+
.C("docking_zhi", out = as.double(0))$out, 2 * pi, pi/2, pi,
+
rep(pi, .C("docking_ntor", out = as.integer(0))$out))
R> fDocking <- function(sol)
+
.C("docking_score", sol = as.double(sol), out = as.double(0))$out
CVEDA and DVEDA are used to solve the minimization problem, since they are the most
robust algorithms among the EDAs based on copulas implemented in copulaedas. The pa-
Yasser Gonzalez-Fernandez, Marta Soto
25
rameters of these algorithms are set to the values reported by Soto et al. (2012) in the solution of the 2z5u test system. The population size of CVEDA and DVEDA is set to 1400
and 1200, respectively. Both algorithms use the implementation of the truncated normal
marginal distributions (Johnson et al. 1994) provided by Trautmann et al. (2014) to satisfy
the box constraints of the variables. The termination criterion of both CVEDA and DVEDA
is to reach a maximum of 100 generations, since an optimum value of the scoring function is
not known. Instances of the ‘VEDA’ class that follow the description given above are created
with the following code.
R> setMethod("edaTerminate", "EDA", edaTerminateMaxGen)
R> cveda <- VEDA(vine = "CVine", indepTestSigLevel = 0.01,
+
copulas = "normal", margin = "truncnorm", popSize = 1400, maxGen = 100)
R> dveda <- VEDA(vine = "DVine", indepTestSigLevel = 0.01,
+
copulas = "normal", margin = "truncnorm", popSize = 1200, maxGen = 100)
Now we proceed to perform 30 independent runs of each algorithm using the edaIndepRuns
function. The arguments of this function are the instances of the ‘VEDA’ class corresponding
to CVEDA and DVEDA, the scoring function fDocking, and the vectors lower and upper
that determine the bounds of the variables.
R> set.seed(12345)
R> cvedaResults <- edaIndepRuns(cveda, fDocking, lower, upper, runs = 30)
R> summary(cvedaResults)
Minimum
Median
Maximum
Mean
Std. Dev.
Generations Evaluations Best Evaluation CPU Time
100
140000
-30.560360 1421.9570
100
140000
-29.387815 2409.0680
100
140000
-23.076769 3301.5960
100
140000
-29.139028 2400.4166
0
0
1.486381 462.9442
R> set.seed(12345)
R> dvedaResults <- edaIndepRuns(dveda, fDocking, lower, upper, runs = 30)
R> summary(dvedaResults)
Minimum
Median
Maximum
Mean
Std. Dev.
Generations Evaluations Best Evaluation CPU Time
100
120000
-30.93501 1928.2030
100
120000
-30.70019 3075.7960
100
120000
-24.01502 4872.1820
100
120000
-30.01427 3053.9579
0
0
1.68583 695.0071
The results obtained in the docking of the 2z5u test system are summarized in Table 4. In
addition to the information provided by edaIndepRuns, we present a column for the RMSD
(root mean square deviation) between the coordinates of the atoms in the experimental crystal
structure and the predicted ligand coordinates of the best solution found at each run. These
values can be computed for a particular solution using the docking_rmsd function included
26
copulaedas: An R Package for EDAs Based on Copulas
Algorithm
CVEDA
DVEDA
Pop.
1400
1200
Evaluations
140000.0 ± 0.0
120000.0 ± 0.0
Lowest Energy
−29.13 ± 1.48
−30.01 ± 1.68
RMSD
0.58 ± 0.13
0.56 ± 0.14
CPU Time
2400.4 ± 462.9
3053.9 ± 695.0
Table 4: Summary of the results obtained in 30 independent runs of CVEDA and DVEDA
for the docking of the 2z5u test system. Pop. denotes Population.
DVEDA
CVEDA
Average Number of Normal Copulas
50
40
30
20
10
0
1
20
40
60
80
100
Generation
Figure 2: Average number of normal copulas selected at each generation of CVEDA and
DVEDA in 30 independent runs for the docking of the 2z5u test system. Since this is an
optimization problem with 26 variables, the C-vines and D-vines have a total of 325 copulas.
in docking.c. The RMSD values serve as a measure of the quality of the predicted ligand
conformations when an experimentally determined solution is known.
Generally, a structure with RMSD below 2Å can be considered as successfully docked. Therefore, both CVEDA and DVEDA achieve good results when solving the 2z5u test system.
DVEDA exhibits slightly lower energy and RMSD values, and it requires a smaller population size than CVEDA. On the other hand, DVEDA uses more CPU time than CVEDA, a
fact that might be related with more dependences being encoded in the D-vines estimated
by DVEDA than in the C-vines used by CVEDA. This situation is illustrated in Figure 2,
which presents the average number of normal copulas selected by CVEDA and DVEDA at
each generation in the 30 independent runs. In both algorithms, the number of normal copulas increases during the evolution, but DVEDA consistently selects more normal copulas
than CVEDA. Although the estimation procedure of the C-vines in CVEDA intends to represent the strongest dependences in the first tree, the constraint that only one variable can be
Yasser Gonzalez-Fernandez, Marta Soto
27
connected to all the others prevent strong correlations to be explicitly encoded in the C-vines.
It is also worth noting that because of the use of the truncation procedure based on AIC in
both CVEDA and DVEDA, the number of statistical tests required to estimate the vines was
dramatically reduced. The median number of vine trees fitted in the 30 independent runs was
4 in CVEDA and 5 in DVEDA out of a total of 25 trees. The interested reader is referred
to Soto et al. (2012) for a deeper study of the use of EDAs based on copulas for solving the
molecular docking problem.
5. Concluding remarks
We have developed copulaedas aiming at providing in a single package an open-source implementation of EDAs based on copulas and utility functions to study these algorithms. In this
paper, we illustrate how to run the copula-based EDAs implemented in the package, how to
implement new algorithms, and how to perform an empirical study to compare copula-based
EDAs on benchmark functions and practical problems. We hope that these functionalities
help the research community to improve EDAs based on copulas by getting a better insight of
their strengths and weaknesses, and also help practitioners to find new applications of these
algorithms to real-world problems.
References
Aas K, Czado C, Frigessi A, Bakken H (2009). “Pair-Copula Constructions of Multiple Dependence.” Insurance: Mathematics and Economics, 44(2), 182–198.
Abdous B, Genest C, Rémillard B (2005). “Dependence Properties of Meta-Elliptical Distributions.” In P Duchesne, B Rémillard (eds.), Statistical Modelling and Analysis for Complex
Data Problems, pp. 1–15. Kluwer Academic Publisher.
Akaike H (1974). “A New Look at Statistical Model Identification.” IEEE Transactions on
Automatic Control, 19(6), 716–723.
Arderı́ RJ (2007). Algoritmo con estimación de distribuciones con cópula Gaussiana. Bachelor’s thesis, University of Havana, Cuba.
Azzalini A (1981). “A Note on the Estimation of a Distribution Function and Quantiles by a
Kernel Method.” Biometrika, 68(1), 326–328.
Baluja S (1994).
“Population-Based Incremental Learning:
A Method for Integrating Genetic Search Based Function Optimization and Competitive Learning.” Technical Report CMU-CS-94-163, Carnegie Mellon University, USA. URL
http://www.esprockets.com/papers/CMU-CS-94-163.ps.gz.
Barba-Moreno S (2007). Una propuesta para algoritmos de estimación de distribución no
paramétricos. Master’s thesis, Center for Research in Mathematics, Mexico.
Bedford T, Cooke RM (2001). “Probability Density Decomposition for Conditionally Dependent Random Variables Modeled by Vines.” The Annals of Mathematics and Artificial
Intelligence, 32(1), 245–268.
28
copulaedas: An R Package for EDAs Based on Copulas
Bengoetxea E, Miquélez T, Lozano JA, Larrañaga P (2002). “Experimental Results in Function Optimization with EDAs in Continuous Domain.” In P Larrañaga, JA Lozano (eds.),
Estimation of Distribution Algorithms. A New Tool for Evolutionary Computation, pp.
181–194. Kluwer Academic Publisher.
Berg D, Aas K (2007).
“Models for Construction of Multivariate Dependence.”
Note SAMBA/23/07, Norwegian Computing Center, NR, Norway.
URL
http://www.nr.no/files/samba/bff/SAMBA2307.pdf.
Berman HM, Westbrook J, Feng Z, Gilliland G, Bhat TN, Weissig H, Shindyalov IN, Bourne
PE (2000). “The Protein Data Bank.” Nucleic Acids Research, 28(1), 235–242.
Beyer HG, Schwefel HP (2002). “Evolution Strategies: A Comprehensive Introduction.” Natural Computing, 1(1), 3–52.
Bonet JSD, Isbell CL, Viola P (1997). “MIMIC: Finding Optima by Estimating Probability
Densities.” In M Mozer, M Jordan, T Petsche (eds.), Advances in Neural Information
Processing Systems, volume 9, pp. 424–430. MIT Press.
Bosman PAN, Thierens D (2006). “Numerical Optimization with Real-Valued Estimation of
Distribution Algorithms.” In M Pelikan, K Sastry, E Cantú-Paz (eds.), Scalable Optimization via Probabilistic Modeling. From Algorithms to Applications, pp. 91–120. SpringerVerlag.
Brechmann EC (2010).
Truncated and Simplified Regular Vines and Their Applications.
Diploma thesis, University of Technology, Munich, Germany.
URL
http://mediatum.ub.tum.de/doc/1079285/1079285.pdf.
Cahon S, Melab N, Talbi E (2004). “ParadisEO: A Framework for the Reusable Design of
Parallel and Distributed Metaheuristics.” Journal of Heuristics, 10(3), 357–380.
Chambers J (2008). Software for Data Analysis: Programming with R. Springer-Verlag.
Cuesta-Infante A, Santana R, Hidalgo JI, Bielza C, naga PL (2010). “Bivariate Empirical and
n-Variate Archimedean Copulas in Estimation of Distribution Algorithms.” In Proceedings
of the IEEE Congress on Evolutionary Computation (CEC 2010), pp. 1355–1362.
Davy M, Doucet A (2003). “Copulas: A New Insight into Positive Time-Frequency Distributions.” IEEE Signal Processing Letters, 10(7), 215–218.
Demarta S, McNeil AJ (2005). “The t Copula and Related Copulas.” International Statistical
Review, 73(1), 111–129.
Devroye L (1986). Non-Uniform Random Variate Generation. Springer-Verlag.
DOLPHIN Project Team (2012). ParadisEO 2.0.1. Inria Lille – Nord Europe, Villeneuve
d’Ascq, France. URL http://paradiseo.gforge.inria.fr/.
Fang HB, Fang KT, Kotz S (2002). “The Meta-Elliptical Distributions with Given Marginals.”
Journal of Multivariate Analysis, 82(1), 1–16.
29
Yasser Gonzalez-Fernandez, Marta Soto
Gao Y (2009). “Multivariate Estimation of Distribution Algorithm with Laplace Transform
Archimedean Copula.” In W Hu, X Li (eds.), Proceedings of the International Conference
on Information Engineering and Computer Science (ICIECS 2009), pp. 1–5.
Genest C, Favre AC (2007). “Everything You Always Wanted to Know About Copula Modeling But Were Afraid to Ask.” Journal of Hydrologic Engineering, 12(4), 347–368.
Genest C, Ghoudi K, Rivest LP (1995). “A Semiparametric Estimation Procedure of Dependence Parameters in Multivariate Families of Distributions.” Biometrika, 82(3), 543–552.
Genest C, Quessy JF, Rémillard B (2007). “Asymptotic Local Efficiency of Cramér-von Mises
Tests for Multivariate Independence.” The Annals of Statistics, 35(1), 166–191.
Genest C, Rémillard B (2004). “Tests of Independence or Randomness Based on the Empirical
Copula Process.” Test, 13(2), 335–369.
Genest C, Rémillard B (2008). “Validity of the Parametric Bootstrap for Goodness-of-Fit
Testing in Semiparametric Models.” Annales de l’Institut Henri Poincaré: Probabilités et
Statistiques, 44(6), 1096–1127.
Genz A, Bretz F (2009). Computation of Multivariate Normal and t Probabilities. SpringerVerlag.
Genz A, Bretz F, Miwa T, Mi X, Leisch F, Scheipl F, Hothorn T (2014). mvtnorm: Multivariate Normal and t Distributions. R package version 0.9-99991., URL
http://CRAN.R-project.org/package=mvtnorm.
Gonzalez-Fernandez Y (2011). Algoritmos con estimación de distribuciones basados en cópulas
y vines. Bachelor’s thesis, University of Havana, Cuba.
Gonzalez-Fernandez Y, Soto M (2014a).
copulaedas:
tribution Algorithms Based on Copulas.
R package
http://CRAN.R-project.org/package=copulaedas.
Estimation of Disversion 1.4.0., URL
Gonzalez-Fernandez Y, Soto M (2014b). vines: Multivariate Dependence Modeling with
Vines. R package version 1.1.0., URL http://CRAN.R-project.org/package=vines.
Haff IH, Aas K, Frigessi A (2010). “On the Simplified Pair-Copula Construction – Simply
Useful or Too Simplistic?” Journal of Multivariate Analysis, 101(5), 1296–1310.
Hauschild M, Pelikan M (2011). “An Introduction and Survey of Estimation of Distribution
Algorithms.” Swarm and Evolutionary Computation, 1(3), 111–128.
Hering C, Hofert M, Mai JF, Scherer M (2010). “Constructing Hierarchical Archimedean
Copulas with Lévy Subordinators.” Journal of Multivariate Analysis, 101(6), 1428–1433.
Hofert M, Kojadinovic I, Maechler M, Yan J (2014). copula: Multivariate Dependence with
Copulas. R package version 0.999-8., URL http://CRAN.R-project.org/package=copula.
Hooke R, Jeeves TA (1961). “Direct Search Solution of Numerical and Statistical Problems.”
Journal of the Association for Computing Machinery, 8(2), 212–229.
30
copulaedas: An R Package for EDAs Based on Copulas
Huey R, Morris GM, Olson AJ, Goodsell DS (2007). “A Semiempirical Free Energy Force Field
with Charge-Based Desolvation.” Journal of Computational Chemistry, 28(6), 1145–1152.
Hult H, Lindskog F (2002). “Multivariate Extremes, Aggregation and Dependence in Elliptical
Distributions.” Advances in Applied Probability, 34(3), 587–608.
Joe H (1996). “Families of m-Variate Distributions with Given Margins and m(m − 1)/2
Bivariate Dependence Parameters.” In L Rüschendorf, B Schweizer, MD Taylor (eds.),
Distributions with Fixed Marginals and Related Topics, pp. 120–141.
Joe H (1997). Multivariate Models and Dependence Concepts. Chapman & Hall.
Joe H (2005). “Asymptotic Efficiency of the Two-Stage Estimation Method for the CopulaBased Models.” Journal of Multivariate Analysis, 94(2), 401–419.
Joe H, Li H, Nikoloulopoulos AK (2010). “Tail Dependence Functions and Vine Copulas.”
Journal of Multivariate Analysis, 101(1), 252–270.
Joe H, Xu JJ (1996). “The Estimation Method of Inference Functions for Margins for
Multivariate Models.” Technical Report 166, University of British Columbia. URL
http://www.stat.ubc.ca/~harry/ifm.pdf.
Johnson NL, Kotz S, Balakrishnan N (1994). Continuous Univariate Distributions, volume 1.
2nd edition. John Wiley & Sons.
Kennedy J, Eberhart RC (1995). “Particle Swarm Optimization.” In Proceedings of the IEEE
International Conference on Neural Networks, volume 4, pp. 1942–1948.
Kern S, Müller SD, Hansen N, Büche D, Ocenasek J, Koumoutsakos P (2003). “Learning
Probability Distributions in Continuous Evolutionary Algorithms – A Comparative Review.” Natural Computing, 3(1), 77–112.
Kirkpatrick S, Gelatt CD, Vecchi MP (1983). “Optimization by Simulated Annealing.” Science, 220(4598), 671–680.
Larrañaga P, Etxeberria R, Lozano JA, Peña JM (1999).
“Optimization by
Learning and Simulation of Bayesian and Gaussian Networks.”
Technical Report EHU-KZAA-IK-4/99, University of the Basque Country.
URL
http://www.sc.ehu.es/ccwbayes/publications/postscript/kzaa-ik-04-99.ps.
Larrañaga P, Etxeberria R, Lozano JA, Peña JM (2000). “Optimization in Continuous Domains by Learning and Simulation of Gaussian Networks.” In Proceedings of the Workshop
in Optimization by Building and Using Probabilistic Models in the Genetic and Evolutionary
Computation Conference (GECCO 2000), pp. 201–204.
Larrañaga P, Lozano JA (eds.) (2002). Estimation of Distribution Algorithms: A New Tool for
Evolutionary Computation, volume 2 of Genetic Algorithms and Evolutionary Computation.
Kluwer Academic Publisher.
Larrañaga P, Lozano JA, Bengoetxea E (2001). “Estimation of Distribution Algorithms Based
on Multivariate Normal and Gaussian Networks.” Technical Report EHU-KZAA-IK-1/01,
University of the Basque Country.
Yasser Gonzalez-Fernandez, Marta Soto
31
Mateo JL, de la Ossa L (2006).
LiO.
Intelligent Systems Group and Data
Mining (SIMD), Universidad de Castilla-La Mancha, Albacete, Spain.
URL
http://www.dsi.uclm.es/simd/SOFTWARE/LIO/.
Mateo JL, de la Ossa L (2007).
“LiO: An Easy and Flexible Library of Metaheuristics.”
Technical Report DIAB-06-04-1, Department of Computing Systems,
Escuela Politecnica Superior de Castilla La Mancha, Albacete, Spain.
URL
http://www.info-ab.uclm.es/descargas/thecnicalreports/DIAB-06-04-1/LiO.pdf.
McNeil AJ (2008). “Sampling Nested Archimedean Copulas.” Journal of Statistical Computation and Simulation, 78(6), 567–581.
McNeil AJ, Nešlehová (2009). “Multivariate Archimedean Copulas, d-Monotone Functions
and l1 -Norm Symmetric Distributions.” The Annals of Statistics, 37(5b), 3059–3097.
Morris GM, Huey R, Lindstrom W, Sanner MF, Belew RK, Goodsell DS, Olson, J A (2009).
“AutoDock4 and AutoDockTools4: Automated Docking with Selective Receptor Flexiblity.”
Journal of Computational Chemistry, 30(16), 2785–2791.
Mühlenbein H, Paaß G (1996). “From Recombination of Genes to the Estimation of Distributions I. Binary Parameters.” In Parallel Problem Solving from Nature — PPSN IV, pp.
178–187. Springer-Verlag.
Nelder JA, Mead R (1965). “A Simplex Method for Function Minimization.” Computer
Journal, 7(4), 308–313.
Nelsen RB (2006). An Introduction to Copulas. 2nd edition. Springer-Verlag.
Nocedal J, Wright SJ (1999). Numerical Optimization. Springer Series in Operations Research.
Springer-Verlag.
Pelikan M (2005). Hierarchical Bayesian Optimization Algorithm. Toward a New Generation
of Evolutionary Algorithms. Springer-Verlag.
Pelikan M, Goldberg DE, Lobo F (2002). “A Survey of Optimization by Building and Using
Probabilistic Models.” Computational Optimization and Applications, 21(1), 5–20.
R Core Team (2014). R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria. URL http://www.R-project.org/.
Rosenkrantz DJ, Stearns RE, Lewis II, M P (1977). “An Analysis of Several Heuristics for
the Traveling Salesman Problem.” SIAM Journal on Computing, 6(3), 563–581.
Rousseeuw P, Molenberghs G (1993). “Transformation of Nonpositive Semidefinite Correlation
Matrices.” Communications in Statistics – Theory and Methods, 22(4), 965–984.
Salinas-Gutiérrez R, Hernández-Aguirre A, Villa-Diharce E (2009). “Using Copulas in Estimation of Distribution Algorithms.” In Proceedings of the Eight Mexican International
Conference on Artificial Intelligence (MICAI 2009), pp. 658–668.
Salinas-Gutiérrez R, Hernández-Aguirre A, Villa-Diharce E (2010). “D-Vine EDA: A New
Estimation of Distribution Algorithm Based on Regular Vines.” In Proceedings of the
Genetic and Evolutionary Computation Conference (GECCO 2010), pp. 359–365.
32
copulaedas: An R Package for EDAs Based on Copulas
Santana R (2011). “Estimation of Distribution Algorithms: From Available Implementations
to Potential Developments.” In Proceedings of the Genetic and Evolutionary Computation
Conference (GECCO 2011), pp. 679–686.
Santana R, Bielza C, Larrañaga P, Lozano JA, Echegoyen C, Mendiburu A, Armañanzas R,
Shakya S (2010). “Mateda-2.0: A MATLAB Package For the Implementation and Analysis
of Estimation of Distribution Algorithms.” Journal of Statistical Software, 35(7), 1–30.
URL http://www.jstatsoft.org/v35/i07/.
Santana R, Echegoyen C, Mendiburu A, Bielza C, Lozano JA, Larrañaga P, Armañanzas R, Shakya S (2009). Mateda-2.0. University of the Basque Country. URL
http://www.sc.ehu.es/ccwbayes/members/rsantana/software/matlab/MATEDA.html.
Schwarz G (1978). “Estimating the Dimension of a Model.” The Annals of Statistics, 6(2),
461–464.
Silverman BW (1986). Density Estimation for Statistics and Data Analysis. Chapman &
Hall.
Sklar A (1959). “Fonctions de répartition à n dimensions et leurs marges.” Publications de
l’Institut de Statistique de l’Université de Paris, 8, 229–231.
Soto M, Gonzalez-Fernandez Y (2010). “Vine Estimation of Distribution Algorithms.” Technical Report ICIMAF 2010-561, Institute of Cybernetics, Mathematics and Physics, Cuba.
Soto M, Gonzalez-Fernandez Y, Ochoa A (in press). “Modeling with Copulas and Vines in
Estimation of Distribution Algorithms.” Investigación Operacional. Preprint available at
http://arxiv.org/abs/1210.5500.
Soto M, Ochoa A, Arderı́ RJ (2007). “Gaussian Copula Estimation of Distribution Algorithm.”
Technical Report ICIMAF 2007-406, Institute of Cybernetics, Mathematics and Physics,
Cuba.
Soto M, Ochoa A, Gonzalez-Fernandez Y, Milanés Y, Álvarez A, Carrera D, Moreno E (2012).
“Vine Estimation of Distribution Algorithms with Application to Molecular Docking.” In
S Shakya, R Santana (eds.), Markov Networks in Evolutionary Computation, volume 14 of
Adaptation, Learning, and Optimization, pp. 209–225. Springer-Verlag.
Storn R, Price K (1997). “Differential Evolution — A Simple and Efficient Heuristic for Global
Optimization Over Continuous Spaces.” Journal of Global Optimization, 11(4), 341–359.
The MathWorks, Inc (2014).
MATLAB – The Language of Technical Computing, Version R2014a.
The MathWorks, Inc., Natick, Massachusetts.
URL
http://www.mathworks.com/products/matlab/.
Tierney L, Rossini AJ, Li N, Sevcikova H (2013). snow: Simple Network of Workstations. R
package version 0.3-13., URL http://CRAN.R-project.org/package=snow.
Trautmann H, Steuer D, Mersmann O, Bornkamp B (2014).
truncnorm:
Truncated Normal Distribution.
R package version 1.0-7., URL
http://CRAN.R-project.org/package=truncnorm.
Yasser Gonzalez-Fernandez, Marta Soto
33
Wang L, Zeng J, Hong Y (2009a). “Estimation of Distribution Algorithm Based on
Archimedean Copulas.” In Proceedings of the First ACM/SIGEVO Summit on Genetic
and Evolutionary Computation (GEC 2009), pp. 993–996.
Wang L, Zeng J, Hong Y (2009b). “Estimation of Distribution Algorithm Based on Copula
Theory.” In Proceedings of the IEEE Congress on Evolutionary Computation (CEC 2009),
pp. 1057–1063.
Warren GL, Andrews CW, Capelli AM, Clarke B, LaLonde J, Lambert MH, Lindvall M,
Nevins N, Semus SF, Senger S, Tedesco G, Wall ID, Woolven JM, Peishoff CE, Head MS
(2006). “A Critical Assessment of Docking Programs and Scoring Functions.” Journal of
Medicinal Chemistry, 49(20), 5912–5931.
Ye B, Gao H, Wang X, Zeng J (2010). “Estimation of Distribution Algorithm Based on
Nested Archimedean Copulas Constructed with Lévy Subordinators.” In Proceedings of
the Eleventh International Conference on Computer-Aided Industrial Design & Conceptual
Design (CAIDCD 2010), pp. 1586–1590.
Affiliation:
Yasser Gonzalez-Fernandez
Department of Interdisciplinary Mathematics
Institute of Cybernetics, Mathematics and Physics
Calle 15 No. 551 e/ C y D, Vedado, La Habana, Cuba
Current affiliation:
School of Information Technology
York University
4700 Keele St, Toronto, Ontario, Canada
E-mail: [email protected]
Marta Soto
Department of Interdisciplinary Mathematics
Institute of Cybernetics, Mathematics and Physics
Calle 15 No. 551 e/ C y D, Vedado, La Habana, Cuba
E-mail: [email protected]
| 9cs.NE
|
APPROXIMATE LIKELIHOOD CONSTRUCTION FOR ROUGH
DIFFERENTIAL EQUATIONS
arXiv:1612.02536v1 [math.ST] 8 Dec 2016
ANASTASIA PAPAVASILIOU AND KASIA B. TAYLOR
Abstract. The paper is split in two parts: in the first part, we construct the
exact likelihood for a discretely observed rough differential equation, driven by
a piecewise linear path. In the second part, we use this likelihood in order to
construct an approximation of the likelihood for a discretely observed differential
equation driven by a general class of rough paths. Finally, we study the behaviour
of the approximate likelihood when the sampling frequency tends to infinity.
1. Introduction
Rough differential equations, introduced in the mid-1990s by Terry Lyons (see
[21], [24] and [22]), have been recently attracting more and more attention of the applied mathematics community. The theoretical foundations are well established and
evolved into a self-contained branch of theoretical pursuits, inspiring the inception
of new important developments such as regularity structures (see [14]). For more
information about rough paths theory see e.g. [23], [12], [11]. The rough paths theory allows to interpret solutions of differential equations of the type dYt = f (Yt )dXt
where X is only of a.s. finite p-variation for all p greater than some q > 2, intuitively
speaking when X exhibits rougher paths than the Brownian motion. The obtained
solutions are pathwise. Aside from that, the rough paths setting allows to encode
information about processes in an efficient way (see [20] and references therein).
The most widely applied process that is considered as signal X in rough differential equations is fractional Brownian motion (fBm) B H with Hurst parameter
H ∈ (0, 1). This was introduced in [18] and popularised by [26]. For H = 21 it
coincides with Brownian motion. Fractional Brownian motion is a Gaussian selfsimilar centred process with stationary increments and the covariance structure
E(BtH BsH ) = 12 (|t|2H + |s|2H − |t − s|2H ). It is of a.s. finite p-variation for all
p > H1 . The span of interdependence between the increments of fBm is infinite, in
other words the increments are not strongly mixing (with exception when H = 21 ).
Date: April 10, 2018.
Supported by the Leverhulme RPG-2013-270: “Statistical Inference of Complex Systems through
Rough Paths”.
1
2
ANASTASIA PAPAVASILIOU AND KASIA B. TAYLOR
An n-dimensional fBM is defined as an ordered set of n independent scalar fBMs;
discussion on multi-dimensional fBMs can be found in e.g. [34].
As a natural consequence of introduction of RDEs, a new class of parametric models arises. In this paper we suggest a novel method for construction of approximate
likelihood that allows for inference on parameters of RDEs. Brownian motion and
fractional Brownian motion can be considered in two settings, as signals in standard
stochastic differential equations (SDEs) and their lifts as signals in RDEs. Recent
surveys of methods for inference for SDEs driven by Brownian motion can be found
in e.g. [19], [13], [16], [3]. Reviews of parameter estimation methods for SDEs driven
by fractional Brownian can be found in e.g. [31], [27]. Recent papers on the topic
include [2], [32], [28], [5], [4], [15], [33]. This list is by no means exhaustive.
Now we proceed to discuss inference methods specifically designed for the rough
paths setting which entails considering convergence of processes in p-variation norm
and iterated integrals approach. Research in this direction was initiated in [29]. The
presented estimation method is based on matching the empirical expected signature
with the theoretical one; the signature is the set of all iterated integrals of a path.
Bayesian inference is conducted in the article [25] where rough paths approach is
adopted and data augmentation technique and Hybrid Monte Carlo are employed.
Stochastic filtering and MLEs in rough paths setting are investigated in [8] and
related papers [6], [9], [7].
2. Setting and Main Ideas
In the first part of the paper, we consider the following type of differential equations
(1)
dYtD = a(YtD ; θ)dt + b(YtD ; θ)dXtD , Y0 = y0 , t ≤ T,
where X D is a realisation of a random piecewise linear path in Rm corresponding to
partition D of [0, T ]. We also assume that θ ∈ Θ, where Θ is the parameter space.
Moreover, we request that a(·, θ) : Rd → Rd and b(·, θ) : Rd → L(Rm , Rd ) are Lip(1),
which are sufficient conditions for the existence and uniqueness of the solution Y D ,
which is a bounded variation path on Rd .
We will use Iθ to denote the Itô map defined by (1). That is, Iθ maps the path
D
X to the path Y D and we write
Y D = Iθ (X D ).
First, we develop a framework for performing statistical inference for differential
equation (1), assuming that we know the distribution of X D . More precisely, we will
aim to construct the likelihood of discrete observations of Y D on the grid D, which
we will denote by yD . The main idea is to use the observations to explicitly construct
the Itô map that maps a finite parametrization of Y D to a finite parametrization of
X D . Typically, Y D will be parametrized by the observations yD := {yti ; ti ∈ D}
APPROXIMATE LIKELIHOOD CONSTRUCTION FOR ROUGH DIFFERENTIAL EQUATIONS 3
and X D will be parametrized by the corresponding normalised increments (∆x)D :=
−xti
xt
; ti , ti+1 ∈ D}.
{ ti+1
i+1 −ti
In section 3, we study the existence and uniqueness for the pair (X D , Y D ) for
D
Y parametrised by the given dataset yD = {yti ; ti ∈ D}. We give conditions for
existence, which is necessary for the methodology to work. Then, we show that for
a and b in Lip(2) and b non singular, the solution will be unique for the case m = d
and it will have m − d degrees of freedom for the case m > d. Since existence will
not in generally be true for the case d > m, this case will not be considered.
In section 4, we explicitly construct the likelihood, treating separately the cases
where we have uniqueness and where we have one or more degrees of freedom.
In the second part of the paper, we consider equation
(2)
dYt = a(Yt ; θ)dt + b(Yt ; θ)dXt , Y0 = y0 , t ≤ T,
where X ∈ GΩp (Rm ) is the realisation of a random geometric p-rough path, defined
as the p-variation limit of a random sequence of nested piecewise linear paths. Let
us denote by D(n, T ) the sequence of nested partitions of [0, T ] and by πn (X) the
corresponding sequence of piecewise linear paths, such that dp (πn (X), X) → 0 as
n → ∞. We now assume that for each θ ∈ Θ, a(·, θ) and b(·, θ) are Lip(γ + 1), for
some γ > p, which are sufficient conditions for the existence and uniqueness of the
solution Y = Iθ (X) ∈ GΩp (Rd ). Moreover, as before, for b non-singular, the pair
(X, Y ) is unique. If we denote by Y (n) the response to the piecewise linear path
πn (X), i.e. Y (n) = Iθ (πn (X)), then the continuity of the Itô map in the p-variation
topology implies that dp (Y (n), Y ) → 0 as n → ∞.
To simplify notation, we will assume that the partitions D(n, T ) are the dyadic
partitions of [0, T ], i.e. they are homogeneous with interval size δ = 2−n . We write
D(n) = {k2−n ; k = 0, . . . , N}, where N = 2n T .
In section 5, we use the likelihood constructed before to construct an approximate
likelihood of observing a realisation of (2) on grid D(n) for some fixed n – denoted
by yD(n) . The main idea behind the construction is to replace the model (2) that
produces the data by (1), which is tractable and converges to (2) for n → ∞.
However, one also needs to normalise the likelihood appropriately, so that the limit
still depends on the parameter that we want to estimate.
In section 6, we demonstrate how the method works: we construct the likelihood
in the simple case of a discretely observed Ornstein-Uhlenbeck model driven by a
picewise linear approximation to fractional Brownian motion and we compute the
corresponding limiting likelihoods for h = 21 .
Finally, in section 7, we make precise in what sense the likelihood constructed in the
previous section is approximate. Replacing the complicated model by a simpler one
approximating the actual model, when we can construct the likelihood corresponding
4
ANASTASIA PAPAVASILIOU AND KASIA B. TAYLOR
to the simpler model exactly, is not an uncommon approach for performing statistical
inference for otherwise intractable models. For example, this is done in [30] where
the authors replace the actual multiscale model by its limiting diffusion and use
that to construct the likelihood. They show that the approximation error due to
the mismatch between data (coming from the multiscale model) and model (the
limiting equation) disappears in the limit. Following a similar approach, we show
that, under suitable conditions, an appropriate distance between the likelihood for
discrete observations on a grid D(n) of the corresponding process Y (n) and of the
limiting process Y respectively disappears, as n → ∞.
3. Existence and Uniqueness
We are given a set of points yD in Rd , where D is the fixed partition of [0, T ].
In this section, we study the existence and uniqueness of piecewise linear path X D ,
whose response Y D = Iθ (X D ) for given θ ∈ Θ goes through points yD , i.e. YtD
= yti
i
for each ti ∈ D.
First, we discuss how to express Y D in terms of X D . By construction, X D is linear
between grid points, i.e.
XtD = XtDi + ∆XtDi (t − ti ) , ∀t ∈ [ti , ti+1 ), ti , ti+1 ∈ D,
XtD
−XtD
i
where ∆XtDi = ti+1
. By definition, Y D = Iθ (X D ) which implies that for every
i+1 −ti
t ∈ [ti , ti+1 ), YtD satisfies
dYtD = a(YtD ; θ)dt + b(YtD ; θ)dXtD =
= a(YtD ; θ)dt + b(YtD ; θ)∆XtDi dt
with initial conditions YtD
= yti . This is an ODE and we have already assumed
i
sufficient regularity on a and b for existence and uniqueness of its solutions. The
general form of the ODE is given by
(3)
dỸt = a(Ỹt ; θ) + b(Ỹt ; θ) · c dt, Y0 = y0
and we will denote its solution by Ft (y0 , c; θ). Then,
(4)
YtD = Ft−ti (yti , ∆Xti ; θ), ∀t ∈ [ti , ti+1 ).
In order to fit Y D to the observed data yD , we need to solve for ∆Xti , using the
terminal value, i.e. solve
(5)
Fti+1 −ti (yti , ∆Xti ; θ) = yti+1
for ∆Xti (yti , yti+1 ; θ). So, for every interval [ti , ti+1 ), we need to solve an independent
system of d equations and m unknowns. That is, we need to study the existence and
APPROXIMATE LIKELIHOOD CONSTRUCTION FOR ROUGH DIFFERENTIAL EQUATIONS 5
uniqueness of solutions with respect to c of the system
(6)
Fδ (y0 , c; θ) = y1 ,
for every θ and for appropriate values of δ, y0 and y1 . We are going to assume
existence of solution, by requiring that y1 ∈ ∩θ∈Θ Mδ (y0 ; θ), where
Mδ (y0 ; θ) = {Fδ (y0 , c; θ); c ∈ Rm } .
(7)
Now, suppose that c1 and c2 are both solutions for a given θ ∈ Θ, i.e.
Fδ (y0 , c1 ; θ) = y1 = Fδ (y0 , c2 ; θ).
We can write the difference as
Fδ (y0 , c2 ; θ) − Fδ (y0 , c1 ; θ) =
Z
1
Dc Fδ (y0 , c1 + s(c2 − c1 ); θ)ds · (c2 − c1 ).
0
Thus, Fδ (y0 , c1 ; θ) = Fδ (y0 , c2 ; θ) implies
Z 1
Dc Fδ (y0 , c1 + s(c2 − c1 ); θ)ds · (c2 − c1 ) = 0.
0
So, it is sufficient to show that ∀ξ ∈ Rm , the rank of d × m matrix Dc Fδ (y0 , ξ; θ)
is d, which implies that the solution will have m − d degrees of freedom, i.e. given
m − d coordinates of c, the other coordinates are uniquely defined. In particular, for
d = m we get uniqueness.
Since the vector field of (3) is linear with respect to c, we know that Ft (y0, c; θ)
will be continuously differentiable with respect to c for every y0 , θ and t in the
appropriate bounded interval [17]. Thus, we define a new auxiliary process as Zt (c) =
Dc Ft (y0 , c; θ) ∈ Rd×m , or,
(8)
Zti,α (c) =
∂ i
F (y0 , c; θ), for i = 1, . . . , d, α = 1, . . . , m.
∂cα t
Then, assuming one additional degree of regularity, Zt (c) satisfies
d i,α
∂ d i
d ∂ i
Ft (y0 , c; θ) =
Zt (c) =
F (y0 , c; θ) =
dt
dt ∂cα
∂cα dt t
!
m
X
∂
cβ biβ (Ft (y0 , c; θ)) =
ai (Ft (y0, c; θ)) +
=
∂cα
β=1
!
d
m
X
X
=
∂j ai (Ft (y0 , c; θ)) +
cβ ∂j biβ (Ft (y0 , c; θ)) Z̄tjα(c) + biα (Ft (y0 , c; θ)),
j=1
β=1
6
ANASTASIA PAPAVASILIOU AND KASIA B. TAYLOR
where by Z̄tα (c) we denote column α ∈ {1, . . . , m} of matrix Zt (c). More concisely,
we write
d α
Z̄t (c) = ▽ (a + b · c) (Ft ; θ) · Z̄tα (c) + b̄α (Ft ; θ),
(9)
dt
where ▽f of a function f : Rd → Rd we denote the d × d matrix defined as
(▽f (y))i,j = ∂j fi (y).
Also, b̄α is column α of matrix b. Note that, for each fixed α, this is a linear equation
of Z̄ α (c) with non-homogeneous coefficients. Also note that the initial conditions
will be
∂ i
∂
Z0i,α (c) =
F0 (y0 , c; θ) =
y0 ≡ 0, ∀i = 1, . . . , d, α = 1, . . . , m.
∂cα
∂cα
Thus, the solution to this equation will be
Z t
α
(10)
Z̄t (c) =
exp (A)s,t b̄α (Fs ; θ)ds,
0
where by exp (A)s,t we denote the sum of iterated integrals
exp (A)s,t =
∞
X
Aks,t
k=0
and
Aks,t
for
(11)
=
Z
···
Z
A(Fu1 ; θ) · · · A(Fuk ; θ)du1 . . . duk
s<u1 <···<uk <t
A(y; θ) = ▽ (a + b · c) (y; θ)
This is a d × d matrix and for k = 0 we get the identity matrix, i.e. A0s,t = Id . Since
each vector Z̄δα is a column of the matrix Dc Fδ (y0 , c; θ), the condition that the rank
of this matrix is d is equivalent to d columns being linearly independent. Without
loss of generality, let’s consider the first d columns (d ≤ m) and let us assume that
λ1 Z̄δ1 + · · · + λd Z̄δd = 0̄,
(12)
for some λ1 , . . . , λd ∈ R. We need to find conditions such that (12) is equivalent to
λ1 = · · · = λd = 0. Using (10) we get that (12) is equivalent to
Z δ
exp (A)s,δ λ1 b̄1 (Fs ; θ) + · · · + λd b̄d (Fs ; θ) ds = 0̄.
0
APPROXIMATE LIKELIHOOD CONSTRUCTION FOR ROUGH DIFFERENTIAL EQUATIONS 7
Using the continuity of the integrated function with respect to s, we can deduce that
there exists a δ ′ ∈ [0, δ], such that we can write the above relationship as
exp (A)δ′ ,δ λ1 b̄1 (Fδ′ ; θ) + · · · + λd b̄d (Fδ′ ; θ) · δ = 0̄.
It is known that exp (A)δ′ ,δ is invertible, with inverse equal to exp (A)δ,δ′ . Consequently, the above relationship can only be true if
λ1 b̄1 (Fδ′ ; θ) + · · · + λd b̄d (Fδ′ ; θ) = 0̄.
Assuming that the rank of d × m matrix b(y; θ) is d for every y, this implies that
λ1 = · · · = λd = 0, which is what we required.
We have shown the following results:
Lemma 3.1. Suppose that rank (b(y, θ)) = d for every y and that a(·, θ) and b(·, θ)
are Lip(2). Then
rank (Zt (c)) = rank (Dc Ft (y0 , c; θ)) = d.
Note that the construction of the process Z can also be done for X ∈ GΩp (Rm ),
provided that its piecewise linear approximations converge in p-variation and that
the vector field functions a and b are now Lip(γ + 1). Uniqueness of the pair (X, Y )
for given Y follows by taking limits. We make this statement formal in the following
Corollary 3.2. Suppose that rank (b(y; θ)) = d for every y and that a(·; θ) and b(·; θ)
are Lip(γ + 1). Then, for a given Y , the solution (X, Y ) of (2) is unique.
4. Construction of the Likelihood
In this section, we construct the exact likelihood of observing the process Y D on
a fixed grid D, denoted by yD = YDD , where Y D is the response to a piecewise linear
path X D on D through (1). The key realisation is that the values of Y D on D actually
completely describe the process Y D .
First, we need to impose a probability structure to the space. Let (Ω, F , P) be
a probability space and let X D be a random variable, taking values in the space
of piecewise linear paths on D, equipped with the 1-variation topology. So, X D is
a random piecewise linear path on Rm corresponding to partition D. Thus, it is
fully described by the distribution of its values on the grid D, or, equivalently, its
increments. Let us denote that distribution by P∆XD .
The measure P∆XD is a distribution on the finite dimensional space Rm×N , with
N = |D| being the size of the partition. We will assume that this is absolutely
continuous with respect to Lebesgue.
By the continuity of Iθ , Y D = Iθ (X D ) is also an implicitly finite dimensional
random variable, whose distribution can be fully describe by the probability of its
values on the grid. Below, we construct the likelihood of observing a realisation of
Y D , corresponding to parametrisation yD .
8
ANASTASIA PAPAVASILIOU AND KASIA B. TAYLOR
4.1. Case I: Uniqueness. Let us first consider the case where we have existence
and uniqueness of solutions to system (6), so m = d. Then, for each dataset yD , the
set {∆Xti (yti , yti+1 ; θ), ti ∈ D} will be uniquely defined as the collection of solutions
of (6). This defines a map
−1
Iθ,D
(yD ) = {∆Xti (yti , yti+1 ; θ), ti ∈ D},
(13)
which can be viewed as a transformation of the observed random variable in terms
of the increments of the driving noise. Note that yD and {∆Xti , ti ∈ D} fully
parametrize processes Y D and X D . Thus, we can write the likelihood of observing
yD as
−1
−1
LY D (yD |θ) = L∆XD Iθ,D
(yD ) |DIθ,D
(yD )|,
where by L∆XD (∆xD ) we denote the Radon-Nikodym derivative of P∆XD with respect to Lebesque. This will be explicitly known since we assumed that we know
the distribution of X D . Finally, since ∆Xti only depends on yti and yti+1 and not
the whole path, it is not hard to see that the Jacobian matrix will be block lower
triangular and consequently, the determinant will be the product of the determinants
of the blocks on the diagonal:
Y
−1
▽∆Xti (yti , y; θ)|y=yti+1 .
(14)
|DIθ,D
(yD )| =
ti ∈D
Note that, by definition,
Fti+1 −ti (yti , ∆Xti (yti , y; θ); θ) ≡ y.
Thus,
Dc Fti+1 −ti (yti , c; θ)|c=∆Xti (yti ,yti+1 ;θ) · ▽∆Xti (yti , y; θ)|y=yti+1 ≡ Id
and, consequently,
−1
Dc Fti+1 −ti (yti , c; θ)|c=∆Xti (yti ,yti+1 ;θ)
=
−1
.
= Zti+1 −ti (∆Xti (yti , yti+1 ; θ))
▽∆Xti (yti , y; θ)|y=yti+1 =
So, the likelihood can be written as
(15)
−1
LY D (yD |θ) = L∆XD Iθ,D
(yD )
Y
ti ∈D
−1
Zti+1 −ti Iθ,D
(yD )ti
!−1
.
APPROXIMATE LIKELIHOOD CONSTRUCTION FOR ROUGH DIFFERENTIAL EQUATIONS 9
4.2. Case II: Degrees of Freedom. Now suppose that m > d. Without loss
of generality, let us assume that given coordinates cd+1 , . . . , cm , the remaining coordinates c1 , . . . , cd are uniquely defined. Similar to previous case, we denote by
−1
Iθ,D,c
(yD ) the map from data points yD to the first d increments, denoted
d+1 ,...,cm
by {∆Xti (yti , yti+1 ; θ, cd+1 , . . . , cm )i , ti ∈ D, i = 1, . . . , d}, for fixed cd+1 , . . . , cm . As
before, this can be viewed as a transformation of the observed random variable in
terms of the first d increments of the driving noise and we get a similar formula for
the likelihood:
LY D (yD |θ, cd+1 , . . . , cm ) =
Q
−1
−1
−1
L∆XD Iθ,D,cd+1,...,cm (yD ) ·
.
ti ∈D Zti+1 −ti Iθ,D,cd+1 ,...,cm (yD )ti
However, cd+1 , . . . , cm will not be known in general, so we have to consider all possible
values of them, leading to the formula
Z
LY D (yD |θ) =
LY D (yD |θ, xd+1 , . . . , xm ) Pcd+1 ,...,cm (dxd+1 , . . . , dxm ) =
Rm−d
!
Z
−1
Y
−1
−1
L∆XD Iθ,D,x
(yD ) ·
Zti+1 −ti Iθ,D,x
(yD )ti
·
d+1 ,...,xm
d+1 ,...,xm
Rm−d
ti ∈D
·Pcd+1 ,...,cm (dxd+1 , . . . , dxm ) ,
where Pcd+1,...,cm is the marginal distribution of P∆XD on R(m−d)×N .
5. The Limiting Case
In the first part of the paper, we assumed that we observe the response to a
differential equation driven by a piecewise linear path (1) and we constructed the
exact likelihood of the observations. In this second part of the paper, we lift the
assumption that the driver X is a piecewise linear path; instead, we assume that we
discretely observe the response to a differential equation (2) driven by a p-rough path
X. We aim to construct an approximate likelihood for the observations and we study
the behaviour of the approximate likelihood when the sampling size δ → 0. A crucial
assumption is that there exists a sequence of partitions D(n) (usually dyadic) such
that the corresponding piecewise linear interpolations πn (X) of the path X converge
in p-variation to the p-rough path X. This allows us to replace (2) by (1).
Let us denote by yD(n) the sequence of observations of the limiting equation (2) on
the grid D(n). We will use the likelihood LY D(n) constructed in (15) to construct an
approximate likelihood for the partially observed limiting equation – for simplicity,
we will now denote it by LY (n) . Also, to simplify the exposition, we will focus on the
case where we have uniqueness, i.e. m = d and b is non-singular.
10
ANASTASIA PAPAVASILIOU AND KASIA B. TAYLOR
A first idea would be to define the approximate likelihood as LY (n) yD(n) |θ . Then,
we would hope to show that, for n large, this will be close to LY (n) y(n)D(n) |θ in
a way that allows the estimators constructed using this likelihood to inherita lot of
the properties of those constructed using exact likelihood LY (n) y(n)D(n)|θ . Note
that the difference between yD(n) and y(n)D(n) is that the first is the response to
a realisation of the rough path x while the latter is the response to the piecewise
linear approximation of x on the grid D(n), i.e. y(n)D(n) = Iθ (πn (x))D(n) , making the
likelihood exact. Note that the two sequences converge in p-variation, for n → ∞.
So, we expect that the estimators constructed using the datasets yD(n) and y(n)D(n)
will be close, provided that the estimator is continuous in the p-variation topology.
However, when the model involves
more than one parameter, it is often the case
that, in the limit, LY (n) yD(n) |θ as a function of θ scales differently for different
coordinates of θ. In particular, this occurs because the drift component dt scales
differently than the ‘diffusion’ component dXt . Thus, we need to carefully normalise
the likelihood appropriately, depending on which coordinate of θ we want to estimate at any time. Actually, it is equivalent and more convenient to work with the
log-likelihood: normalising the log-likelihood involves adding functions to the loglikelihood that are independent of the parameters we want to estimate and thus
do not alter the estimation, as both the maximum with respect to the parameter
and the posterior on the parameter remain unaffected. So, we want to construct an
expansion of the log-likelihood of the form
(16)
ℓY (n) yD(n) |θ =
M
X
k=0
(k)
ℓY (n) yD(n) |θ n−αk + RM (yD(n) , θ)
(k)
for M ∈ N and −∞ < α0 < α1 < · · · < αM < ∞, where ℓY (n) yD(n) |θ converges
to a non-trivial limit (finite and non-zero) for every k = 0, . . . , M and the remainder RM (yD(n) , θ) satisfies limn→∞ nαM RM (yD(n) , θ) = 0. This will exist, assuming
sufficient smoothness of the log-likelihood function of ∆xD(n) . The construction is
based on a Taylor expansion around the two components of the inverse that scale
differently. More precisely, it is done as follows:
−1
1. Using (15), we express the log-likelihood in terms of Iθ,D(n)
(yD(n) ), which we
−1
will denote for simplicity by Iθ,n (yD(n) ). That is,
−1
X
N
−1
−1
ℓY (n) yD(n) |θ = ℓ∆XD(n) Iθ,n
log |Zti+1 −ti Iθ,n
(yD(n) ) −
(yD(n) ) |.
i=0
2. Assuming sufficient regularity of ℓ∆XD(n) , we expand ℓY (n) yD(n) |θ in terms
−1
−1
of the monomials of Iθ,n
(yD(n) ). Note that Zti+1 −ti Iθ,n
(yD(n) given by (10)
will always be smooth.
APPROXIMATE LIKELIHOOD CONSTRUCTION FOR ROUGH DIFFERENTIAL EQUATIONS11
3. We have assumed uniqueness, which is equivalent to b(y) being invertible.
−1
Thus, Iθ,n
(yD(n) ) can be expressed as
Z ti+1
−1
Iθ,n (yD(n) )ti ,ti+1 =
(17)
b−1 (Y (n, yD(n) )u ; θ)dY (n, yD(n) )u
ti
−
Z
ti+1
ti
= Iθ−1
b−1 (Y (n, yD(n) )u ; θ)a(Y (n, yD(n) )u ; θ)du
Y (n, yD(n) ) ti ,ti+1
where the process Y (n, yD(n) ) is the response to a piecewise linear path,
parametrised by its values on the grid D(n), given by yD(n) . Using this equiv−1
alence, we further expand the monomials of Iθ,n
(yD(n) ) in terms of monomials
of the two vectors formed by these integrals, i.e.
Z ti+1
N −1
−1
b (Y (n, yD(n) )u ; θ)dY (n, yD(n) )u
ti
i=0
and
Z
ti+1
−1
b (Y (n, yD(n) )u ; θ)a(Y (n, yD(n) )u ; θ)du
ti
N −1
,
i=0
where N = 2n T = Tδ , as before.
4. We classify each monomial to class k if the normalisation needed for it to
converge to something “meaningful” is nαk , i.e. the monomial multiplied by
nαk converges to a function that is not equivalent
to 0. We denote by M the
(k)
number of such classes and by ℓY (n) yD(n) |θ the sum of all class k monomials
multiplied by nαk .
Remark 5.1. Note that while (17) is very useful for studying limiting behaviour, it
is not useful for constructing the likelihood for fixed n, as constructing Y (n, yD(n) )
corresponding to observations yD(n) also requires the solution of (5).
Now, let θi be an arbitrary coordinate of the parameter θ and suppose that
(m)
yD(n) |θ are independent of θi for k = 0, . . . , m − 1 while ℓY (n) yD(n) |θ depends on θi , in a way to be made precise later, which will depend on the way the
constructed likelihood is used. Intuitively, we expect that the first m − 1 components
will be irrelevant to the estimation of the parameter and should be ignored, while
the remaining log-likelihood should be normalised by nαm . Thus, we will say that
coordinate θi of the parameter is of order m and, given observations of the limiting
(m)
equation, we will use the dominating term ℓY (n) yD(n) |θ for its estimation. We will
(k)
ℓY (n)
12
ANASTASIA PAPAVASILIOU AND KASIA B. TAYLOR
assume that all coordinates of the parameter are of finite order – otherwise, they
cannot be estimated!
Below, we use this framework to build estimators for parameter θ using the constructed likelihoods. We discuss separately the two most common approaches, corresponding to the Frequentist or Bayesian paradigm.
5.1. Frequentist Setting. In the frequentist setting, we use the likelihood constructed above in order to construct the Maximum Likelihood Estimator (MLE) of
the parameter θ ∈ Θ. We inductively define the MLEs of different co-ordinates of θ,
as follows:
1. We start with the lower order α0 . We say that co-ordinates of the parameter
θ are of order α0 and we denote them by θ0 if
ℓ(0) (yD(n) |θ) = ℓ(0) (yD(n)|θ0 ), ∀θ ∈ Θ.
Then, we define their estimate as
(0)
θ̂0 (yD(n) ) = argmaxθ0 ℓY (n) yD(n) |θ0 .
2. Suppose that we have defined parameters of order up to m and their MLEs,
for some m ≥ 0. Then, we say that the co-ordinates of θ that satisfy
ℓ(m+1) (yD(n) |θ) = ℓ(m+1) (yD(n) |θ̂0 (yD(n) ), . . . , θ̂m (yD(n)), θm+1 )
for all θ with coordinates of order less or equal to m equal to their MLE
estimates are of order m + 1, and we denote them by θm+1 . We define their
MLE as
(m+1)
θ̂m+1 (yD(n) ) = argmaxθm+1 ℓY (n) yD(n) |θ̂0 (yD(n) ), . . . , θ̂m (yD(n) ), θm+1 .
5.2. Bayesian Setting. In the Bayesian setting, we use the likelihood constructed
above together with a prior distribution on the parameter space that we will denote
by u, in order to construct the posterior distribution of the parameter θ ∈ Θ. We
inductively define the posterior distributions of different co-ordinates of θ, as follows:
1. First, we start with the lower order α0 and work our way up. We say that
a co-ordinate θi of the parameter is of order αk if the distance between the
marginal posterior on θi and the marginal prior (for an appropriate choice
of distance on the measure space) is non-zero for the first time when the
posterior is computed
using the scaling of the likelihood corresponding to
(k)
ℓY (n) yD(n) |θ0 . We will denote by θ(k) all the co-ordinates of the parameter
that are of order k and by r the maximum order.
APPROXIMATE LIKELIHOOD CONSTRUCTION FOR ROUGH DIFFERENTIAL EQUATIONS13
2. The posterior can be written as a product of the posteriors of parameters of
different orders as follows:
r
Y
P(θ|yD(n) ) =
P θ(k)|yD(n), θ(k − 1), . . . , θ(0) ,
k=0
where for each k, P θ(k)|yD(n), θ(k − 1), . . . , θ(0) is computed at the relevant
scale, i.e.
(k)
exp ℓY (n) yD(n) |θ(k), cθ(k) n−αk u(θ(k), cθ(k))
,
R
(k)
−α
k
exp
ℓ
u(
θ̃(k),
cθ(k))d
θ̃(k)
y
|
θ̃(k),
cθ(k)
n
D(n)
Y (n)
Θk
where Θk is the projection of the parameter space to the coordinates of order
k and by cθ(k) we denote all the parameters that are not of order k.
6. Example: The 1d fractional O.U. process
To demonstrate the methodology, we will apply the ideas described in the previous
section to a simple example. We consider the differential equation
(18)
dYtD = −λYtD dt + σXtD , Y0D = 0,
where XtD is the piecewise linear interpolation to a fractional Brownian path with
Hurst parameter h on a homogeneous grid D = {kδ; k = 0, . . . , N} where Nδ = T .
Our goal will be to construct the likelihood of discretely observing a realisation of
the solution Y D (ω) on the grid, for parameter values θ = (λ, σ) ∈ R+ × R+ .
Our first task is to explicitly construct the parametrization of Y D (ω) in terms of
its values on the grid yD , that completely determine the process. Let X D (ω) be
the piecewise linear interpolation on D of the corresponding realisation of a fractional Brownian path driving (18). We will denote by xti its values on the grid, i.e.
X D (ω)ti = xti , ∀ti ∈ D. Since X D (ω) is the piecewise linear path defined on these
points, Y D (ω) will be the solution to
x(k+1)δ − xkδ
dY D (ω)t = −λY D (ω)tdt + σ
dt,
δ
which is given by
σ x(k+1)δ − xkδ
1 − e−λ(t−kδ) , t ∈ [kδ, (k + 1)δ).
Y D (ω)t = Y D (ω)kδ e−λ(t−kδ) +
λ
δ
We now need to solve for the unknown ∆xk+1 := x(k+1)δ − xkδ : for t = (k + 1)δ. We
get
σ∆xk+1
(19)
y(k+1)δ = ykδ e−λδ +
1 − e−λδ
λδ
14
ANASTASIA PAPAVASILIOU AND KASIA B. TAYLOR
and, consequently,
(20)
−1
Iθ,D
(yD )k+1 := ∆xk+1
λδ y(k+1)δ − ykδ e−λδ
, k = 0, . . . , N − 1,
=
σ (1 − e−λδ )
with y0 = 0 and θ = (λ, σ). Thus, Y D (ω) is given by
(21)
Y D (ω)t = ykδ e−λ(t−kδ) +
y(k+1)δ − ykδ e−λδ
1 − e−λ(t−kδ) ,
−λδ
1−e
for t ∈ [kδ, (k + 1)δ) and y0 = 0.
Clearly, in this case, the solution of system (6) always exists and is unique under
the condition that σ 6= 0. Let us now compute the process Z defined in (8). In this
case, since d = 1, this is a scalar process. It is easy to compute Z directly but we
will use formula (10) instead, as a demonstration. First, we note that A defined in
(11) will be A(y) = ∂y (−λy + σc) = −λ. Thus, (10) becomes
Z t
σ
σ
(1 − e−λt ).
Zt =
exp(−λ(t − s)) ds =
δ
λδ
0
We now have all the elements we need to write down the likelihood: from (15), we
get
N
λδ
−1
LY D (yD | θ) = L∆XD Iθ,D (yD )
.
σ(1 − e−λδ )
Finally, we note that the likelihood of the increments ∆XD is a mean zero Gaussian
distribution with covariance matrix given by
δ 2h
|j − i + 1|2h + |j − i − 1|2h − 2|j − i|2h , i, j = 1, . . . , N,
ΣD
h ij =
2
where h is the Hurst parameter of the fractional Brownian motion. Thus, the likelihood becomes
N
1 −1
λδ
D − 21
D −1 −1
∗
(22)
|2πΣh | exp − Iθ,D (yD ) Σh
Iθ,D (yD )
,
2
σ(1 − e−λδ )
where we denote by z ∗ the transpose of a vector z. The corresponding log-likelihood
is proportional to
λδ
1 −1
D −1 −1
∗
.
Iθ,D (yD ) + N log
(23)
ℓY (yD | θ) ∝ − Iθ,D (yD ) Σh
2
σ(1 − e−λδ )
−1
Finally, we can replace Iθ,D
above with its exact expression, which gives
(24)
∗
λδ
λ2 δ 2
λ
D −1
λ
,
∆ y D Σh
∆ y D + N log
ℓY (yD | λ, σ) ∝ − 2
2σ (1 − e−λδ )2
σ(1 − e−λδ )
APPROXIMATE LIKELIHOOD CONSTRUCTION FOR ROUGH DIFFERENTIAL EQUATIONS15
where by ∆λ ykδ = y(k+1)δ − ykδ e−λδ .
Now, let us construct the corresponding limiting likelihoods derived from (24), for
h = 21 , i.e. the diffusion case. Then, (24) becomes
∗
λ2 δ
λδ
λ
λ
ℓY (yD | λ, σ) ∝ − 2
=
∆ y D ∆ y D + N log
2σ (1 − e−λδ )2
σ(1 − e−λδ )
N
−1
X
λδ
λ2 δ
−λδ 2
=
(y(k+1)δ − ykδ e ) + N log
= − 2
2σ (1 − e−λδ )2 k=0
σ(1 − e−λδ )
N
−1
X
T
λ2 δ
T
1 − e−λδ
(25) = − log(σ) − log
(y(k+1)δ − ykδ e−λδ )2 .
− 2
−λδ
2
δ
δ
λδ
2σ (1 − e ) k=0
Clearly, the first term of (25) is of order o( 1δ ). Noting that
1 − e−λδ
λ
log
= − δ + O(δ 2 ),
λδ
2
it follows that the second term of (25) is of order o(1). In particular,
λ
1 − e−λδ
T
= − T + O(δ).
log
δ
λδ
2
Now, let us consider the third term in (25). First, we note that
λ2 δ
1
λ
= 2 + 2 + O(δ).
2
−λδ
2
2σ (1 − e )
2σ δ 2σ
Then, we expand the sum as follows:
N
−1
X
(y(k+1)δ − ykδ e−λδ )2 =
k=0
N
−1
X
(y(k+1)δ − ykδ )2 +
k=0
+2
N
−1
X
2
ykδ
(1 − e−λδ )2
k=0
N
−1
X
ykδ (1 − e−λδ )(y(k+1)δ − ykδ ).
k=0
The first term above is of order O(1), since it converges to the Quadratic Variation
of the process, which is finite. The second term is of order O(δ), with the dominating
term being
N
−1
X
k=0
2
ykδ
(1
−λδ 2
−e
2
) =λ δ
N
−1
X
k=0
2
ykδ
δ
16
ANASTASIA PAPAVASILIOU AND KASIA B. TAYLOR
since the latter sum converges to the corresponding integral. Finally, the last term
of the sum expansion is also of order O(δ), with the dominating term being
2
N
−1
X
−λδ
ykδ (1 − e
)(y(k+1)δ − ykδ ) = 2λδ
N
−1
X
ykδ (y(k+1)δ − ykδ ),
k=0
k=0
with the sum again converging to the corresponding Itô integral. Putting everything
together, we get that the third term of (25) will be
N −1
N −1
N −1
N −1
λ X
λ2 X 2
λ X
1 X
2
2
(y(k+1)δ −ykδ ) + 2
(y(k+1)δ −ykδ ) + 2
y δ+
ykδ (y(k+1)δ −ykδ )+O(δ).
2σ 2 δ k=0
2σ k=0
2σ k=0 kδ σ 2 k=0
Finally, we get that the log-likelihood (25) can be expanded as
1
ℓY (yD | λ, σ) ∝
δ
N −1
T
1 X
2
− log σ − 2
(y(k+1)δ − ykδ )2
2
2σ k=0
!
N −1
λ X
λ
(y(k+1)δ − ykδ )2 +
T− 2
2
2σ k=0
!
+
N −1
N −1
λ X
λ2 X 2
y δ− 2
ykδ (y(k+1)δ − ykδ )
− 2
2σ k=0 kδ
σ k=0
(26)
!
+ O(δ).
Thus, we estimate the diffusion parameter σ 2 using the O( 1δ ) term of the expansion,
which leads to the Quadratic Variation estimate for σ 2 . We estimate the drift parameter λ using the O(1) term of the expansion. If σ 2 is estimated by the Quadratic
Variation, the first part cancels living just the second part which, in the limit, is the
log-likelihood function one gets using Girsanov.
7. Convergence of Approximate Likelihood
In this section, we study the behaviour of the approximate likelihoods constructed
in section 5. The main result of this section is the following:
(k)
Theorem 7.1. Let ℓY (n) (·|θ) be the scaled likelihoods constructed in section 5, so
that (16) holds. Let y be the response to a p-rough path x through (2) and y(n) be
the response to πn (x) through (1), where πn (x) is the piecewise linear interpolation
of x on grid D(n) = {k2−n T, k = 0, . . . , N} for N = 2n T . Then, assuming that the
determinant of b is uniformly bounded from below over both parameters y and θ,
(k)
(k)
(27)
lim sup ℓY (n) yD(n) |θ − ℓY (n) y(n)D(n) |θ = 0,
n→∞
θ
for all k such that αk ≤ 0.
APPROXIMATE LIKELIHOOD CONSTRUCTION FOR ROUGH DIFFERENTIAL EQUATIONS17
First, we will show that the log-likelihood function, which is the logarithm of (15),
is continuous with respect to the inverse Itô map of the data
−1
Iθ,n
(yD(n) ) = Iθ−1 (Y (n, yD(n) )),
(28)
where, as before, Y (n, yD(n) ) is the response to a piecewise linear path parametrised
by its values on the grid, yD(n) . So, Iθ−1 (Y (n, yD(n) )) will be exactly that piecewise
linear path driving the process. More precisely, we will show that, under certain
assumptions, the following holds:
−1
−1
ℓY (n) yD(n) |θ − ℓY (n) ỹD(n) |θ ≤ ω dp (Iθ,n
(29)
(yD(n) ), Iθ,n
(ỹD(n) ) ,
for some modulus of continuity function ω, independent of θ and n. This can be
further split into two parts, one corresponding to the log-likelihood of the inverse and
the other corresponding to the Jacobian correction. We start by deriving a bound
for the part of the log-likelihood error corresponding to the Jacobian correction.
−1
Lemma 7.2. For Zti+1 −ti and Iθ,n
defined as in (10) and (13) respectively and under
the additional assumption on b that
inf ||b(y; θ)|| =
y,θ
1
> 0,
Mb
for some Mb > 0, it holds that
P
P
−1
−1
ti ∈D(n) log |Zti+1 −ti Iθ,n (yD(n) )ti | −
ti ∈D(n) log |Zti+1 −ti Iθ,n (ỹD(n) )ti | ≤
−1
−1
C · ω(dp (Iθ,n
(yD(n) ), Iθ,n
(ỹD(n) )))
for some C ∈ R+ and modulus of continuity function ω.
Proof. We write
P
P
−1
−1 n
(30) ti ∈D(n) log |Zti+1 −ti Iθ,n (yD(n) )ti | − ti ∈D(n) log |Zti+1 −ti Iθ,n (ỹD(n))ti | =
=
P
ti ∈D(n) log
−1
|Zti+1 −ti (Iθ,n
(yD(n) )ti )|
−1 n
(ỹD(n) )ti |
|Zti+1 −ti Iθ,n
As before, using the continuity of integrated function within Zti+1 −ti with respect to
the time variable, we write
−1
−1
Zti+1 −ti Iθ,n
(yD(n) )ti = exp A(yti , Iθ,n
(yD(n) )ti ; θ) ζ ,t · b(Fζi ; θ) · (ti+1 − ti )
i i+1
and
−1 n
−1 n
Zti+1 −ti Iθ,n
(ỹD(n) )ti = exp A(ỹtni , Iθ,n
(ỹD(n) )ti ; θ) η ,t
i i+1
· b(Fηi ; θ) · (ti+1 − ti )
18
ANASTASIA PAPAVASILIOU AND KASIA B. TAYLOR
for some ζi , ηi ∈ [ti , ti+1 ]. So, (30) simplifies to
P
ti ∈D(n) log
P
ti ∈D(n) log
−1
−1
(yD(n) )ti )ζi ;θ)|
·b(F (yti ,Iθ,n
(yD(n) )ti ;θ))
| exp(A(yti ,Iθ,n
ζ ,t
i i+1
−1 n
−1 n
n
n
·b(F (ỹt ,Iθ,n (ỹD(n) )ti )ηi ;θ)|
| exp A(ỹt ,Iθ,n (ỹD(n) )ti ;θ)
i
ηi ,ti+1
−1
|
(yD(n) )ti ;θ))
| exp(A(yti ,Iθ,n
ζ ,t
i i+1
−1 n
|
| exp A(ỹtn ,Iθ,n
(ỹD(n) )ti ;θ)
i
ηi ,ti+1
+
P
=
i
|b(F (yt ,I −1 (yD(n) )ti )ζi ;θ)|
i θ,n
ti ∈D(n) log |b(F (ỹ n ,I −1 (ỹ n
ti
θ,n
) ) ;θ)|
D(n) ti ηi
.
Focusing on the second summand, we write
P
|b(F (yt ,I −1 (yD(n) )ti )ζi ;θ)|
|≤
| ti ∈D(n) log |b(F (ỹni ,Iθ,n
−1 n
ti θ,n (ỹD(n) )ti )ηi ;θ)|
P
−1
−1 n
n
≤
ti ∈D(n) log 1 + Mb |b(F (yti , Iθ,n (yD(n) )ti )ζi ; θ)| − |b(F (ỹti , Iθ,n (ỹD(n) )ti )ηi ; θ)|
P
−1
−1 n
Mb · ti ∈D(n) |b(F (yti , Iθ,n
(yD(n) )ti )ζi ; θ)| − |b(F (ỹtni , Iθ,n
(ỹD(n) )ti )ηi ; θ)| ≤
−1
−1 n
Mb · Cb · ω1 (dp (Iθ,n
(yD(n) ), Iθ,n
(ỹD(n) ))),
where we used the inequality log(1 + x) < x, assumption that inf y ||b(y)|| = M1b > 0,
the Lipschitz continuity of b and the universal limit theorem (see [1] for exact bound).
Similarly, we get that
−1
| exp A(yti , Iθ,n
(yD(n) )ti ; θ) ζ ,t |
X
−1
−1 n
i i+1 ≤ MA CA ·ω2 (dp (Iθ,n
(yD(n) ), Iθ,n
(ỹD(n) ))).
log
−1
n
| exp A(ỹtni , Iθ,n (ỹD(n)
)ti ; θ)
|
ti ∈D(n)
ηi ,ti+1
under the additional assumption that infy |exp(A)(y; θ)ti ,ti+1 | > 0, which is trivially
satisfied. Putting the two together, we get the result.
Since we have not specified the distribution of the driving process X, we will just
assume that the log-likelihood of the inverse satisfies a similar bound. Then, (29)
will hold under the additional assumption that
(31)
ℓ∆XD(n) ∆xD(n) − ℓ∆XD(n) ∆x̃D(n) ≤ ω (dp (x, x̃)) ,
where x and x̃ are p-rough paths with ∆xD(n) and ∆x̃D(n) denoting increments on
D(n) respectively.
To prove Theorem 7.1, it remains to show that for y and y(n) as described in
−1
−1
the theorem, the piecewise linear paths Iθ,n
(yD(n) ) and Iθ,n
(y(n)D(n) ) converge in pvariation in such a way that the different scalings also converge. Let us first look
into the different scalings and how these affect the requirements for the convergence
of the inverse piecewise linear paths.
APPROXIMATE LIKELIHOOD CONSTRUCTION FOR ROUGH DIFFERENTIAL EQUATIONS19
Lemma 7.3. Suppose that the log-likelihood has the form (16) and that it is uniformly continuous in the p-variation topology, with some modulus of continuity ω
independent of θ or n, i.e.
−1
−1
(32)
|ℓY (n) yD(n) |θ − ℓY (n) ỹD(n) |θ | ≤ ω dp (Iθ,n
(yD(n) ), Iθ,n
(ỹD(n) ))
Then, if for every k = 0, . . . , M
−1
−1
lim nαk ω dp (Iθ,n
(yD(n)), Iθ,n
(ỹD(n) )) = 0,
n→∞
it follows that
(k)
(k)
lim |ℓY (n) yD(n) |θ − ℓY (n) ỹD(n) |θ | = 0.
(33)
n→∞
Proof. Let’s denote by Rk (yD(n) , θ) the remainder of the log-likelihood expansion
truncated at αk for some k ≥ 0, i.e.
Rk (yD(n) , θ) =
M
X
j=k+1
(j)
ℓY (n) yD(n) |θ n−αj + RM (yD(n) , θ).
(j)
Since functions ℓY (n) are convergent, we get that
nαk Rk (yD(n) , θ) → 0, as n → ∞.
We first prove (33) for k = 0. We re-write (32) as
(0)
(0)
ℓY (n) yD(n) |θ − ℓY (n) ỹD(n) |θ + R0 (yD(n) , θ) − R0 (ỹD(n) , θ) nα0 ≤
−1
−1
≤ nα0 ω dp (Iθ,n
(yD(n) ), Iθ,n
(ỹD(n) ))
which implies that
(0)
(0)
ℓY (n) yD(n) |θ − ℓY (n) ỹD(n) |θ ≤
−1
−1
nα0 ω dp (Iθ,n
(yD(n) ), Iθ,n
(ỹD(n) )) + nα0 R0 (yD(n) , θ) − R0 (ỹD(n) , θ) .
Given that the nα0 times the remainder disappears in the limit, for every ǫ > 0, we
can find n0 such that ∀n ≥ n0
(0)
(0)
−1
−1
ℓY (n) yD(n) |θ − ℓY (n) ỹD(n)|θ ≤ nα0 ω dp (Iθ,n
(yD(n) ), Iθ,n
(ỹD(n) )) + ǫ.
proving the statement for k = 0. The proof of (33) for k > 0 is similar, building
inductively on k.
Finally, we show the following
20
ANASTASIA PAPAVASILIOU AND KASIA B. TAYLOR
Lemma 7.4. Let Iθ−1 be the inverse Itô map defined by (1), with b invertible and uniformly bounded away from zero. We denote by Y (n, Iθ0 (x)D(n) ) and Y (n, Iθ0 (πn (x))D(n) )
the responses to the piecewise linear map as in (2), parametrised by its values on the
grid D(n), given by Iθ0 (x)D(n) and Iθ0 (πn (x))D(n) respectively, where x is a fixed rough
path in GΩp (Rd ) and θ0 ∈ Θ. Then,
(34)
lim dp Iθ−1 Y (n, Iθ0 (x)D(n) ) , Iθ−1 Y (n, Iθ0 (πn (x))D(n) ) = 0,
n→∞
provided that dp (πn (x), x) → 0 as n → ∞.
Proof. Under the assumption that b is invertible, Iθ−1 is given by (17), which is an
integral of the rough path Y (n, ·), for any θ. Thus, it is sufficient to show that
Y (n, Iθ0 (πn (x))D(n) ) = 0,
Y (n, Iθ0 (x)D(n) ) , Iθ−1
(35)
lim dp Iθ−1
0
0
n→∞
as a consequence of the continuity of integration and the Itô map in p-variation
topology. By construction,
these are piecewise linear paths. Clearly, πn (x) =
Iθ−1
(π
(x))
)
and
πn (x)ti ,ti+1 = ∆xi . Let us denote by
Y
(n,
I
θ0
n
D(n)
0
x̃(n) = Iθ−1
Y
(n,
I
(x)
)
θ
D(n)
0
0
and let x̃(n)ti ,ti+1 = ∆x̃i be the corresponding increments. Then, (35) becomes
equivalent to
(36)
lim kx̃(n) − πn (x)kp = 0,
n→∞
where k · kp is the p-variation metric. The process x̃(n) − πn (x) is piecewise linear
on D(n), as a difference of two piecewise linear functions on D(n). We know that in
this case, the supremum of the p-variation norm is achieved at a subset of D(n) (see
[10]), so
p1
X
(37)
kx̃(n) − πn (x)kp =
kx̃(n)τi ,τi+1 − πn (x)τi ,τi+1 kp .
τi E(n)⊂D(n)
Now, by definition, we have that
yti+1 = F (yti , δ; ∆x̃i ),
where F denotes the solution of the ODE (3), as before. By expanding F around
∆xi , we get
yti+1 = F (yti , δ; ∆xi ) + Dc F (yti , δ; ξi )(∆x̃i − ∆xi ),
where Dc F is the derivative of the solution to the ODE with respect to constant c.
It follows that
∆x̃i − ∆xi = Dc F (yti , δ; ξi )−1 yti+1 − F (yti , δ; ∆xi ) .
APPROXIMATE LIKELIHOOD CONSTRUCTION FOR ROUGH DIFFERENTIAL EQUATIONS21
We have already seen that Dc F is given by the process Z(c) defined in (8). So, the
above relationship becomes
∆x̃i − ∆xi = Z(ξi )−1 yti+1 − F (yti , δ; ∆xi ) .
Moreover, it follows by (10) that if b is uniformly bounded away from zero, i.e.
kbk > M1 , then so is Z and consequently, Z −1 is bounded from above. Thus
k∆x̃i − ∆xi k ≤ Mkyti+1 − F (yti , δ; ∆xi )k.
Finally, noting that yti+1 and F (yti , δ; ∆xi ) are both the solution of (2) for the same
i
initial conditions yti , driven by x and ∆x
t respectively. Since the first iterated inδ
tegral (increment) of these two drives is the same, if we expand yti+1 and F (yti , δ; ∆xi )
in terms of the iterated integrals of the driver and initial conditions, we find that their
difference comes from the difference of the second iterated integrals, which should be
2
of order δ p , since dp (πn (x), x) → 0. Thus
2
k∆x̃i − ∆xi k ≤ MO(δ p ).
The result follows from (37) and the above estimate.
The three lemmas together prove the theorem.
References
[1] Baudoin,
F.
Rough
Path
Theory.
Online
notes
available
at
http://www.math.purdue.edu/ fbaudoin/Rough.pdf
[2] Beskos, A and Dureau, J and Kalogeropoulos, K. Bayesian inference for partially observed
stochastic differential equations driven by fractional Brownian motion. Biometrika, 102(4):
809–827, 2015.
[3] Bishwal, JPN. Parameter estimation in stochastic differential equations, Springer, 2008.
[4] Brouste, A and Iacus, SM. Parameter estimation for the discretely observed fractional
Ornstein–Uhlenbeck process and the Yuima R package. Computational Statistics, 28(4): 1529–
1547, 2013.
[5] Chronopoulou, A and Tindel, S. On inference for fractional differential equations. Statistical
Inference for Stochastic Processes, 16(1): 29–61, 2013.
[6] Crisan, D and Diehl, J and Friz, PK and Oberhauser, H. Robust filtering: correlated noise and
multidimensional observation. Annals of Applied Probability, 23(5): 2139–2160, 2013.
[7] Diehl, J and Friz, P and Mai, H. Pathwise stability of likelihood estimators for diffusions via
rough paths. arXiv preprint arXiv:1311.1061.
[8] Diehl, J, Friz, PK, Mai, H, Oberhauser, H, Riedel, S and Stannat, W. Robustness in stochastic filtering and maximum likelihood estimation for SDEs. In Extraction of Quantifiable
Information from Complex Systems, pp161–178, 2014.
[9] Diehl, J, Oberhauser, H and Riedel, S. A Levy-area between Brownian motion and rough paths
with applications to robust non-linear filtering and RPDEs. arXiv preprint arXiv:1301.3799.
22
ANASTASIA PAPAVASILIOU AND KASIA B. TAYLOR
[10] Driver,
B.
Rough
Path
Analysis.
Online
notes
available
at
http://www.math.ucsd.edu/ bdriver.
[11] Friz, PK and Hairer, M. A course on rough paths: with an introduction to regularity structures,
Springer, 2014.
[12] Friz, PK and Victoir, NB. Multidimensional stochastic processes as rough paths: theory and
applications, CUP, 2010.
[13] Fuchs, C. Inference for Diffusion Processes: With Applications in Life Sciences. Springer Science & Business Media, 2013.
[14] Hairer, M. A theory of regularity structures. Inventiones mathematicae, 198(2): 269–504, 2014.
[15] Hu, Y and Nualart, D. Parameter estimation for fractional Ornstein–Uhlenbeck processes.
Statistics & Probability Letters, 80(11): 1030–1038, 2010.
[16] Iacus, SM. Simulation and inference for stochastic differential equations: with R examples,
Springer, 2009.
[17] Kelley, W G and Peterson, A C. The Theory of Differential Equations (Second Edition),
Springer, 2010
[18] Kolmogorov, AN. Wienersche Spiralen und einige andere interessante Kurven im Hilbertschen
Raum. CR (Dokl.) Acad. Sci. URSS, 26(2): 115–118, 1940.
[19] Kutoyants, YA. Statistical inference for ergodic diffusion processes, Springer, 2013.
[20] Levin, D and Lyons, T and Ni, H. Learning from the past, predicting the statistics for the
future, learning an evolving system. arXiv preprint arXiv:1309.0260.
[21] Lyons, T. Differential equations driven by rough signals (I): An extension of an inequality of
L. C. Young. Math. Res. Lett, 4(1): 451–464, 1994.
[22] Lyons, T. Differential equations driven by rough signals. Rev. Mat. Iberoamericana, 14(2),
1998.
[23] Lyons, T and Qian, Z. System control and rough paths, OUP, 2002.
[24] Lyons, T. The interpretation and solution of ordinary differential equations driven by rough
signals. Proc. Symposia in Pure Math, 57: 115–128, 1995.
[25] Lysy, M and Pillai, NS. Statistical Inference for Stochastic Differential Equations with Memory.
arXiv preprint arXiv:1307.1164
[26] Mandelbrot, BB and Van Ness, JW. Fractional Brownian motions, fractional noises and applications. SIAM review, 10(4): 422–437, 1968.
[27] Mishura, Y. Stochastic calculus for fractional Brownian motion and related processes. Springer,
2008.
[28] Neuenkirch, A and Tindel, S. A least square-type procedure for parameter estimation in stochastic differential equations with additive fractional noise. Statistical Inference for Stochastic
Processes, 17(1):99–120, 2014.
[29] Papavasiliou, A and Ladroue, C. Parameter estimation for rough differential equations. The
Annals of Statistics, 39(4): 2047–2073, 2011.
[30] Papavasiliou, A and Pavliotis, GA and Stuart, AM. Maximum likelihood drift estimation for
multiscale diffusions. Stochastic Processes and their Applications, 119(10): 3173–3210, 2009.
[31] Rao, BLSP. Statistical inference for fractional diffusion processes, John Wiley & Sons, 2011.
[32] Saussereau, B. Nonparametric inference for fractional diffusion. Bernoulli, 20(2): 878–918,
2014.
[33] Tudor, CA and Viens, Frederi G. Statistical aspects of the fractional stochastic calculus. Annals
of Statistics, 35(3): 1183–1212, 2007.
APPROXIMATE LIKELIHOOD CONSTRUCTION FOR ROUGH DIFFERENTIAL EQUATIONS23
[34] Unterberger, J. A rough path over multidimensional fractional Brownian motion with arbitrary
Hurst index by Fourier normal ordering. Stochastic Processes and their Applications, 120(8):
1444–1472, 2010.
Anastasia Papavasiliou, Department of Statistics, University of Warwick, Coventry, CV4 7AL, UK
Kasia B. Taylor, Department of Statistics, University of Warwick, Coventry,
CV4 7AL, UK
| 10math.ST
|
arXiv:1705.03673v1 [cs.CC] 10 May 2017
The Complexity of Routing with Few Collisions∗
Till Fluschnik†,1 , Marco Morik1 , and Manuel Sorge‡,1,2
1
Institut für Softwaretechnik und Theoretische Informatik, TU Berlin, Germany,
[email protected],[email protected]
2
Ben Gurion University of the Negev, Be’er Sheva, Israel, [email protected]
Abstract
We study the computational complexity of routing multiple objects through
a network in such a way that only few collisions occur: Given a graph G with
two distinct terminal vertices and two positive integers p and k, the question is
whether one can connect the terminals by at least p routes (e.g. paths) such
that at most k edges are time-wise shared among them. We study three types
of routes: traverse each vertex at most once (paths), each edge at most once
(trails), or no such restrictions (walks). We prove that for paths and trails the
problem is NP-complete on undirected and directed graphs even if k is constant
or the maximum vertex degree in the input graph is constant. For walks, however, it is solvable in polynomial time on undirected graphs for arbitrary k and
on directed graphs if k is constant. We additionally study for all route types
a variant of the problem where the maximum length of a route is restricted by
some given upper bound. We prove that this length-restricted variant has the
same complexity classification with respect to paths and trails, but for walks it
becomes NP-complete on undirected graphs.
1
Introduction
We study the computational complexity of determining bottlenecks in networks. Consider a network in which each link has a certain capacity. We want to send a set
of objects from point s to point t in this network, each object moving at a constant
rate of one link per time step. We want to determine whether it is possible to send
our (predefined number of) objects without congestion and, if not, which links in the
network we have to replace by larger-capacity links to make it possible.
Apart from determining bottlenecks, the above-described task arises when securely
routing very important persons [15], or packages in a network [2], routing container
transporting vehicles [18], and generally may give useful insights into the structure
and robustness of a network. A further motivation is congestion avoidance in routing
∗ Major
parts of this work done while all authors were with TU Berlin.
by the DFG, project DAMM (NI 369/13-2).
‡ Supported by the DFG, project DAPA (NI 369/12-2), the People Programme (Marie Curie Actions) of the European Union’s Seventh Framework Programme (FP7/2007-2013) under REA grant
agreement number 631163.11 and by the Israel Science Foundation (grant no. 551145/14).
† Supported
1
fleets of vehicles, a problem treated by recent commercial software products (e.g.
http://nunav.net/) and poised to become more important as passenger cars and
freight cars become more and more connected. Assume that we have many requests
on computing a route for a set of vehicles from a source location to a target location,
as it happens in daily commuting traffic. Then the idea is to centrally compute these
routes, taking into account the positions in space and time of all other vehicles. To
avoid congestion, we try to avoid that on two of the routes the same street appears at
the same time.
A first approximation to determine such bottlenecks would be to compute the set
of minimum cuts between s and t. However, by daisy chaining our objects, we may
avoid such “bottlenecks” and, hence, save on costs for improving the capacity of our
links. Apart from the (static) routes we have to take into account the traversals in
time that our objects take.
Formally, we are given an undirected or directed graph with marked source and
sink vertex. We ask whether we can construct routes between the source and the sink
in such a way that these routes share as few edges as possible. By routes herein we
mean either paths, trails, or walks, modeling different restrictions on the routes: A
walk is a sequence of vertices such that for each consecutive pair of vertices in the
sequence there is an edge in the graph. A trail is a walk where each edge of the graph
appears at most once. A path is a trail that contains each vertex at most once. We
say that an edge is shared by two routes, if the edge appears at the same position
in the sequence of the two routes. The sequence of a route can be interpreted as the
description of where the object taking this route is at which time. So we arrive at the
following core problem:
Routing with Collision Avoidance (RCA)
Input:
A graph G = (V, E), two distinct vertices s, t ∈ V , and two integers p ≥ 1 and k ≥ 0.
Question: Are there p s-t routes that share at most k edges?
This definition is inspired by the Minimum Shared Edges (MSE) problem [6,
15, 20], in which an edge is already shared if it occurs in two routes, regardless of the
time of traversal. Finally, note that finding routes from s to t also models the general
case of finding routes between a set of sources and a set of sinks.
Considering our introductory motivating scenarios, it is reasonable to restrict the
maximal length of the routes. For instance, when routing vehicles in daily commuting
traffic while avoiding congestion, the routes should be reasonably short. Motivated by
this, we study the following variant of RCA.
Fast Routing with Collision Avoidance (FRCA)
Input:
A graph G = (V, E), two distinct vertices s, t ∈ V , and three integers
p, α ≥ 1 and k ≥ 0.
Question: Are there p s-t routes each of length at most α that share at most
k edges?
In the problem variants Path-RCA, Trail-RCA, and Walk-RCA, the routes are
restricted to be paths, trails, or walks, respectively (analogously for FRCA).
2
Table 1: Overview of our results: DAGs abbreviates directed acyclic graphs; NPc., W[2]-h., P abbreviate NP-complete, W[2]-hard, and containment in the class P ,
respectively; ∆ denotes the maximum degree; ∆i/o denotes the maximum over the inand outdegrees. a (Thm. 1) b (Thm. 3) c (Cor. 2) d (even on planar graphs)
Undirected, with k
constant
Path-(F)RCA
Directed, with k
arbitrary
constant arbitrary
DAGs, with k
const. arbitrary
NP-c.d , Thm. 4 (Cor. 3) NP-c.d , Thm. 4 (Cor. 3) Pa
each k ≥ 0
each ∆ ≥ 4
each k ≥ 0
d
each ∆i/o ≥ 4
Trail-(F)RCA NP-c. , Thm. 5 (Cor. 4) NP-c. , Thm. 6 (Cor. 5) Pa
each k ≥ 0
Walk-RCA
Walk-FRCA
d
each ∆ ≥ 5
each k ≥ 0
NP-c.,
W[2]-hb/c
P (Thm. 9) NP-c.,
W[2]-h.b
Pa
NP-c.,
W[2]-h.b
NP-c., W[2]-h. P (Thm. 9) NP-c.,
W[2]-h.c
(Thm. 8)
Pa
NP-c.,
W[2]-h.c
P (Thm. 7)
open
each ∆i/o ≥ 3
NP-c.,
W[2]-hb/c
Our Contributions. We give a full computational complexity classification of
RCA and FRCA (except Walk-FRCA) with respect to the three mentioned route
types; with respect to undirected, directed, and directed acyclic input graphs; and distinguishing between constant and arbitrary budget. Table 1 summarizes our results.
To our surprise, there is no difference between paths and trails in our classification.
Both Path-RCA (Section 4) and Trail-RCA (Section 5) are NP-complete in all of
our cases except on directed acyclic graphs when k ≥ 0 is constant (Section 3). We
show that the problems remain NP-complete on undirected and directed graphs even
if k ≥ 0 is constant or the maximum degree is constant. We note that the Minimum
Shared Edges problem is solvable in polynomial time when the number of shared
edges is constant, highlighting the difference to its time-variant Path-RCA.
The computational complexity of the length-restricted variant FRCA for paths and
trails equals the one of the variant without length restrictions. The variant concerning
walks (Section 6) however differs from the other two variants as it is tractable in more
cases, in particular on undirected graphs. (We note that almost all of our tractability
results rely on flow computations in time-expanded networks (see, e.g., Skutella [19]).)
Remarkably, the tractability does not transfer to the length-restricted variant WalkFRCA, as it becomes NP-complete on undirected graphs. This is the only case where
RCA and FRCA differ with respect to their computational complexity.
Related Work. As mentioned, Minimum Shared Edges inspired the definition
of RCA. MSE is NP-hard on directed [15] and undirected [5, 6] graphs. In contrast to RCA, if the number of shared edges equals zero, then MSE is solvable in
polynomial time. Moreover, MSE is W[2]-hard with respect to the number of shared
edges and fixed-parameter tractable with respect to the number of paths [6]. MSE is
3
polynomial-time solvable on graphs of bounded treewidth [20, 1].
There are various tractability and hardness results for problems related to RCA
with k = 0 in temporal graphs, in which edges are only available at predefined time
steps [3, 10, 14, 13]. The goal herein is to find a number of edge or vertex-disjoint
time-respecting paths connecting two fixed terminal vertices. Time-respecting means
that the time steps of the edges in the paths are nondecreasing. Apart from the fact
that all graphs that we study are static, the crucial difference is in the type of routes:
vehicles moving along time-respecting paths may wait an arbitrary number of time
steps at each vertex, while we require them to move at least one edge per time step
(unless they already arrived at the target vertex).
Our work is related to flows over time, a concept already introduced by Ford and
Fulkerson [7] to measure the maximal throughput in a network over a fixed time period.
This and similar problems were studied continually, see Skutella [19] and Köhler et
al. [12] for surveys. In contrast, our throughput is fixed, our flow may not stand
still or go in circles arbitrarily, and we want to augment the network to allow for our
throughput.
2
Preliminaries
We define [n] := {1, . . . , n} for every n ∈ N. Let G = (V, E) be an undirected (directed)
graph. Let the sequence P = (v1 , . . . , vℓ ) of vertices in G be a walk, trail, or path. We
call v1 and vℓ the start and end of P . For i ∈ [ℓ], we denote by P [i] the vertex vi at
position i in P . Moreover, for i, j ∈ [ℓ], i < j, we denote by P [i, j] the subsequence
(vi , . . . , vj ) of P . By definition, P has an alternative representation as sequence of
edges (arcs) P = (e1 , . . . , eℓ−1 ) with ei := {vi , vi+1 } (ei := (vi , vi+1 )) for i ∈ [ℓ − 1].
Along this representation, we say that P contains/uses edge (arc) e at time step i if
edge (arc) e appears at the ith position in P represented as sequence of edges (arcs)
(analogue for vertices). We call an edge/arc shared if two routes uses the edge/arc at
the same time step. We say that a walk/trail/path Q is an s-t walk/trail/path, if s
is the start and t is the end of Q. The length of a walk/trail/path is the number of
edges (arcs) contained, where we also count multiple occurrences of an edge (arc) (we
refer to a path of length m as an m-chain). (We define the maximum over in- and
outdegrees in G by ∆i/o (G) := maxv∈V (G) {outdeg(v), indeg(v)}.)
A parameterized problem P is a set of tuples (x, ℓ) ∈ Σ∗ × N, where Σ denotes a
finite alphabet. A parameterized problem P is fixed-parameter tractable if it admits an
algorithm that decides every input (x, ℓ) in f (ℓ) · |x|O(1) time (FPT-time), where f is a
computable function. The class FPT is the class of fixed-parameter tractable problems.
The classes W[q], q ≥ 1, contain parameterized problems that are presumably not
fixed-parameter tractable. For two parameterized problems P and P ′ , a parameterized
reduction from P to P ′ is an algorithm that maps each input (x, ℓ) to (x′ , ℓ′ ) in FPTtime such that (x, ℓ) ∈ P if and only if (x′ , ℓ′ ) ∈ P ′ , and ℓ′ ≤ g(ℓ) for some function g.
A parameterized problem P is W[q]-hard if for every problem contained in W[q] there
is a parameterized reduction to P .
4
Preliminary Observations on RCA and FRCA. We state some preliminary
observations on RCA and FRCA. If there is a shortest path between the terminals s
and t of length at most k, then routing any number of paths along the shortest path
introduces at most k shared edges. Hence, we obtain the following.
Observation 1. Let (G, s, t, p, k) be an instance of RCA with dist(s, t) < ∞. If
k ≥ distG (s, t), then (G, s, t, p, k) is a yes-instance.
If we consider walks, the length of an s-t walk in a graph can be arbitrarily large.
We prove, however, that for paths, trails, and walks, RCA and FRCA are contained
in NP, that is, each variant allows for a certificate of size polynomial in the input size
that can be verified in time polynomial in the input size.
Lemma 1. RCA and FRCA on undirected and on directed graphs are contained
in NP.
Proof. Given an instance (G, s, t, p, k) of Path-RCA and a set of p s-t paths, we
can check in polynomial time whether they share at most k edges. The same holds
for Trail-RCA and Walk-RCA (the latter follows from Theorem 7). This is still
true for all variants on directed graphs (for walks we refer to Lemma 12). Moreover,
we can additionally check in linear time whether the length of each path/trail/walk is
at most some given α ∈ N. Hence, the statement follows.
3
Everything is Equal on DAGs
Note that on directed acyclic graphs, every walk contains each edge and each vertex
at most once. Hence, every walk is a path in DAGs, implying that all three types of
routes are equivalent in DAGs.
We prove that RCA is solvable in polynomial time if the number k of shared arcs
is constant, but NP-complete if k is part of the input. Moreover, we prove that the
same holds for the length-restricted variant FRCA. We start the section with the case
of constant k ≥ 0.
3.1
Constant Number of Shared Arcs
Theorem 1. RCA and FRCA on n-vertex m-arc DAGs are solvable in O(mk+1 ·
n3 ) time and O(mk+1 · α2 · n) time, respectively.
We prove Theorem 1 as follows: We first show that RCA and FRCA on DAGs
are solvable in polynomial time if k = 0 (Theorem 2 below). We then show that an
instance of RCA and FRCA on directed graphs is equivalent to deciding, for all ksized subsets K of arcs, the instance with k = 0 and a modified input graph in which
each arc in K has been copied p times:
Theorem 2. If k = 0, RCA on n-vertex m-arc DAGs is solvable in O(n3 · m) time.
We need the notion of time-expanded graphs.
5
Definition 1. Given a directed graph G, we denote a directed graph H the (directed)
τ -time-expanded graph of G if V (H) = {v i | v ∈ V (G), i = 0, . . . , τ } and A(H) =
{(v i−1 , wi ) | i ∈ [τ ], (v, w) ∈ A(G)}.
Note that for every directed n-vertex m-arc graph the τ -time-expanded graph can
be constructed in O(τ · (n + m)) time. We prove that we can decide RCA and FRCA
by flow computation in the time-expanded graph of the input graph:
Lemma 2. Let G = (V, A) be a directed graph with two distinct vertices s, t ∈ V .
Let p ∈ N and τ := |V |. Let H be the τ -time-expanded graph of G with p additional
arcs (ti−1 , ti ) between the copies of t for each i ∈ [τ ]. Then, G allows for at least p
s-t walks of length at most τ not sharing any arc if and only if H allows for an s0 -tτ
flow of value at least p.
Proof. (⇒) Let G allow for p s-t walks W1 , . . . , Wp not sharing any arc. We construct
an s0 -tτ flow of value p in H as follows. Observe that Wi = (v0 , . . . , vℓ ) corresponds
to a path P = (v00 , v11 , . . . , vℓℓ ) in H. If ℓ < τ , then extend this path to the path P =
(v00 , . . . , vℓℓ , tℓ+1 , , . . . , tτ ) (observe that vℓℓ = tℓ ). Set the flow on the arcs of P to one.
From the fact that W1 , . . . , Wp are not sharing any arc in G, we extend the flow as
described above for each walk by one, hence obtaining an s0 -tτ flow of value p.
(⇐) Let H allow an s0 -tτ flow of value p. It is well-known that any s0 -tτ flow of value p
in H can be turned into p arc-disjoint s0 -tτ paths in H [11]. Let P = (v00 , v11 , . . . , vℓℓ )
′
′
be an s0 -tτ path in H. Let ℓ′ be the smallest index such that vℓℓ′ = tℓ . Then
W = (v1 , v2 , . . . , vℓ′ ) is an s-t walk in G. Let P be the set of p s0 -tτ paths in H
obtained from an s0 -tτ flow of value p, and let W be the set of p s-t walks in G
obtained from P as described above. As every pair of paths in P is arc-disjoint, no
pair of walks in W share any arc in G.
Proof of Theorem 2. Let (G = (V, A), s, t, p, 0) be an instance of Walk-RCA with G
being a directed acyclic graph. Let n := |V |. We construct the directed n-timeexpanded graph H of G with p additional arcs (ti−1 , ti ) for each i ∈ [τ ]. Note that
any s-t path in G is of length at most n − 1 due to G being directed and acyclic. The
statement then follows from Lemma 2.
Lemma 2 is directly applicable to FRCA, by constructing an α-expanded graph.
Corollary 1. If k = 0, then FRCA on n-vertex m-arc DAGs is solvable in O(α2 · n ·
m) time.
Let G = (V, A) be a directed graph and let K ⊆ A and x ∈ N. We denote
by G(K, x) the graph obtained from G by replacing each arc (v, w) ∈ K in G by x
copies (v, w)1 , . . . , (v, w)x .
Lemma 3. Let (G = (V, A), s, t, p, k) be an instance of Walk-RCA with G being
a directed graph. Then, (G, s, t, p, k) is a yes-instance of Walk-RCA if and only if
there exists a set K ⊆ A with |K| ≤ k such that (G(K, p), s, t, p, 0) is a yes-instance
of Walk-RCA. The same statement holds true for Walk-FRCA.
6
Proof. (⇒) Let G allow for a set of p s-t walks W = {W1 , . . . , Wp } sharing at most
k edges. Let K ⊆ A denote the set of at most k arcs shared by the walks in W. We
construct a set of p s-t walks W ′ = {W1′ , . . . , Wp′ } in G(K, p) from W as follows. For
each i ∈ [p], let Wi′ = Wi . Whenever an arc (v, w) ∈ K appears in Wi′ , we replace
the arc by its copy (v, w)i . Observe that (i) Wi′ forms an s-t walk in G(K, p), (ii) the
positions of the arcs in A \ K in the walks remain unchanged, and (iii) for each arc
(v, w) ∈ K, no walk contains the same copy of the arc. As the arcs in K are the only
shared arcs of the walks in W, the walks in W ′ do not share any arc in G(K, p).
(⇐) Let K ⊆ A be a subset of arcs in G with |K| ≤ k such that G(K, p) allows for
a set of p s-t walks W ′ = {W1′ , . . . , Wp′ } with no shared arc. We construct a set of p
s-t walks W = {W1 , . . . , Wp } in G from W ′ as follows. For each i ∈ [p], let Wi = Wi′ .
Whenever an arc (v, w)x , for some x ∈ [p], with (v, w) ∈ K appears in Wi′ , we replace
the arc by its original (v, w). Observe that (i) Wi forms an s-t walk in G, (ii) the
positions of the arcs in A \ K in the walks remain unchanged. As the arcs in the set K
of at most |K| ≤ k arcs can appear at the same positions in any pair of two walks
in W, the s-t walks in P share at most k arcs in G.
Note that as the length of the walks do not change in the proof, the statement of the
lemma also holds for Walk-FRCA.
Proof of Theorem 1. Let (G = (V, A), s, t, p, k) be an instance of Walk-RCA with G
being a directed acyclic graph. For each k-sized subset K ⊆ A of arcs in G, we decide
the instance (G(K, p), s, t, p, 0). The statement for RCA then follows from Lemma 3
and Theorem 2. We remark that the value of a maximum flow between two terminals
in an n-vertex m-arc graph can be computed in O(n·m) time [16]. The running time of
the algorithm is in O(|A|k · (|V |3 · |A|)). The statement for FRCA follows analogously
with Lemma 3 and Corollary 1.
3.2
Arbitrary Number of Shared Arcs
If the number k of shared arcs is part of the input, then both RCA and FRCA are
NP-complete and W[2]-hard with respect to k.
Theorem 3. RCA on DAGs is NP-complete and W[2]-hard with respect to k.
The construction in the reduction for Theorem 3 is similar to the one used by
Omran et al. [15, Theorem 2]. Herein, we give a (parameterized) many-one reduction
from the NP-complete [9] Set Cover problem: given a set U = {u1 , . . . , un }, a set of
subsets F = {F1 , . . . , Fm } with Fi ⊆ U for
S all i ∈ [m], and an integer ℓ ≤ m, is there
a subset F ′ ⊆ F with |F ′ | ≤ ℓ such that F ∈F ′ F = U . We say that F ′ is a set cover
and we say that the elements in F ∈ F are covered by F . Note that Set Cover
is W[2]-complete with respect to the solution size ℓ in question [4]. In the following
Construction 1, given a Set Cover instance, we construct the DAG in an equivalent
RCA or FRCA instance.
Construction 1. Let a set U = {u1 , . . . , un }, a set of subsets F = {F1 , . . . , Fm } with
Fi ⊆ U for all i ∈ [m], and an integer ℓ ≤ m be given. Construct a directed acyclic
graph G = (V, A) as follows. Initially, let G be the empty graph. Add the vertex sets
7
VU = {v1 , . . . , vn } and VF = {w1 , . . . , wm }, corresponding to U and F , respectively.
Add the edge (vi , wj ) to G if and only if ui ∈ Fj . Next, add the vertex s to G. For
each w ∈ VF , add an (ℓ + 2)-chain to G connecting s with w, and direct all edges in
the chain from s towards w. For each v ∈ VU , add an (ℓ + 1)-chain to G connecting
s with v, and direct all edges in the chain from s towards v. Finally, add the vertex t
to G and add the arcs (w, t) for all w ∈ VF .
Lemma 4. Let U , F , ℓ, and G as in Construction 1. Then there are at most ℓ sets
in F such that their union is U if and only if G admits n + m s-t walks sharing at
most ℓ arcs in G.
S
Proof. (⇒) Suppose there are ℓ sets F1′ , . . . , Fℓ′ ∈ F such that i∈[ℓ] Fi′ = U . Let
w1′ , . . . , wℓ′ be the vertices in VF corresponding to F1′ , . . . , Fk′ . We construct n + m
s-t walks as follows. Each outgoing chain on s corresponds to exactly one s-t walk.
Those walks that start with the chains connecting s with a w ∈ VF are extended
directly to t (there is no other choice). For all the other walks, as F1′ , . . . , Fk′ cover
U , for each vi ∈ VU there is at least one arc towards {w1′ , . . . , wk′ }. Route the walks
arbitrarily towards one out of {w1′ , . . . , wk′ } and then forward to t. Observe that all
walks contain exactly one vertex in VF at time step ℓ + 3. Moreover, only the arcs
(wi′ , t) for i ∈ [ℓ] are contained in more than one walk. As they are at most ℓ, the
claim follows.
(⇐) Suppose G admits a set W of n+m s-t walks sharing at most ℓ arcs in G. Observe
first that the arcs of the form (w, t), w ∈ VF , are the only arcs that can be shared
whenever at most ℓ arcs are shared, due to the fact that each outgoing chain on s is
of length longer than k. Moreover, each arc (w, t), w ∈ VF is contained in at least one
s-t walk in W, because no two walks in W can share a chain outgoing from s, and
for each w ∈ VF the only outgoing arc on w has endpoint t. Denote by W ⊆ VF the
set of vertices such that the set {(w, t) | w ∈ W } is exactly the set of shared arcs by
the n + m s-t walks in W. Observe that |W | ≤ ℓ. We claim that the set of sets F ′ ,
containing the sets that by construction correspond to the vertices in W , forms a set
cover for U . We show that for each element u ∈ U there is a w ∈ W such that the set
corresponding to w is containing u.
Let u ∈ U be an arbitrary element of U . Consider the walk P ∈ W containing the
vertex v ∈ VU corresponding to element u. As P forms an s-t walk in G, walk P contains a vertex w′ ∈ VF . As discussed before, there is a walk P ′ containing the chain
from s to w′ not containing any vertex in VU . By construction, w′ is at time step ℓ + 3
in both P and P ′ . As the only outgoing arc on w′ is (w′ , t), both P and P ′ use
the arc (w′ , t) at time step ℓ + 3, and hence (w, t) is shared by P and P ′ . It follows
that w′ ∈ W , and hence u is covered by the set in F ′ corresponding to w′ .
Proof of Theorem 3. We give a (parameterized) many-one reduction from Set Cover
to RCA. Let (U, F , ℓ) be an instance of Set Cover. We construct the instance
(G, s, t, p, k), where G is obtained by applying Construction 1, p = |U | + |F |, and
k = ℓ. The correctness of the reduction then follows from Lemma 4. Finally, note
that as k = ℓ and Set Cover is W[2]-hard with respect to the size ℓ of the set cover,
it follows that RCA is W[2]-hard with respect to the number k of shared arcs.
8
Observe that each s-t walk in the graph obtained from Construction 1 is of length
at most ℓ + 3.
Hence, in the proof of Theorem 3, we can instead reduce to an
instance (G, s, t, p, k, α) of FRCA, where G is obtained by applying Construction 1,
p = |U | + |F |, and k = ℓ, and α = ℓ + 3. Therewith, we obtain the following.
Corollary 2. FRCA on DAGs is NP-complete and W[2]-hard with respect to k + α.
4
Path-RCA
In this section, we prove the following theorem.
Theorem 4. Path-RCA both on undirected planar and directed planar graphs is
NP-complete, even if k ≥ 0 is constant or ∆ ≥ 4 is constant.
In the proof of Theorem 4, we reduce from the following NP-complete [8] problem
(a cubic graph is a graph where every vertex has degree exactly three):
Planar Cubic Hamiltonian Cycle (PCHC)
Input:
An undirected, planar, cubic graph G.
Question: Is there a cycle in G that visits each vertex exactly once?
Roughly, the instance of Path-RCA obtained in the reduction consists of the original
graph G connected to the terminals s, t via a bridge (see Figure 1). We ask for constructing roughly n paths connecting the terminals, where n is the number of vertices
in the input graph of PCHC. All but one of these paths will use the bridge to t in
the constructed graph for n time steps in total, each in a different time step. Thus,
this bridge is occupied for roughly n time steps, and the final path is forced to stay
in the input graph of PCHC for n time steps. For a path, this is only possible by
visiting each of the n vertices in the graph exactly once, and hence it corresponds to
a Hamiltonian cycle.
The reduction to prove Theorem 4 uses the following Construction 2.
Construction 2. Let G = (V, E) be an undirected, planar, cubic graph with n = |V |.
Construct in time polynomial in the size of G an undirected planar graph G′ as follows
(refer to Figure 1 for an illustration of the constructed graph). Let initially G′ be the
empty graph. Add a copy of G to G′ . Denote the copy of G in G′ by H. Next,
add the new vertices s, t, v, w to G. Connect s with v, and w with t by an edge. For
each m ∈ {4, 5, . . . , n+ 1}, add an m-chain connecting s with w. Next, consider a fixed
plane embedding φ(G) of G. Let x1 denote a vertex incident to the outer face in φ(G).
Then, there are two neighbors x2 and x3 of x1 also incident to the outer face in φ(G).
Add the edges {v, x1 }, {x2 , w} and {x3 , w} to G′ completing the construction of G′ .
We remark that G′ is planar as it allows a plane embedding (see Figure 1) using φ as
an embedding of H.
Lemma 5. Let G and G′ be as in Construction 2. Then G admits a Hamiltonian
cycle if and only if G′ allows for at least n − 1 s-t paths with no shared edge.
9
hain
1)-c
..
+
(n
n .
i
a
h
5 -c
in x2
4-cha
s
v
w
t
H
x1
x3
Figure 1: Graph G′ obtained in Construction 2. The gray part represents the graph H.
Dashed lines represent chains.
Proof. (⇐) Let P denote a set of n − 1 s-t paths in G′ with no shared edge. Note
that the degree of s is equal to n − 1. As no two paths in P share any edge in G′ ,
each path in P uses a different edge incident to s. This implies that n − 2 paths in P
uniquely contain each of the chains connecting s with w, and one path P ∈ P contains
the edge {s, v}. Note that each of the n − 2 paths contain the vertex w at most once,
and since they contain the chains connecting s with w, the edge {w, t} appears at
the time steps {5, 6, . . . , n + 2} in these n − 2 paths P. Hence, the path P has to
contain the edge {w, t} at a time step smaller than five or larger than n + 2. Observe
that, by construction, the shortest path between s and w is of length 4 and, thus, P
cannot contain the edge {w, t} on any time step smaller than five. Hence, P has to
contain the edge at time step at least n + 3. Since the distance between s and x1 is
two, and the distance from x2 , x3 to w is one, P has to visit each vertex in H exactly
once, starting at x1 , and ending at one of the two neighbors x2 or x3 of x1 . Hence,
P restricted to H describes a Hamiltonian path in H, which can be extended to an
Hamiltonian cycle by adding the edge {x1 , x2 } in the first or {x1 , x3 } in the second
case.
(⇒) Let G admit a Hamiltonian cycle C. Since C contains every vertex in G exactly
once, it contains x1 and its neighbors x2 and x3 . Since C forms a cycle in G and G
is cubic, at least one of the edges {x1 , x2 } or {x1 , x3 } appears in C. Let C ′ denote an
ordering of the vertices in C such that x1 appears first and the neighbor x ∈ {x2 , x3 }
of x1 with {x1 , x} contained in C appears last. We construct n − 1 s-t paths without
sharing an edge. First, we construct n − 2 s-t paths, each containing a different chain
connecting s with w and the edge {w, t}. Observe that since the lengths of each chain
is unique, no edge (in particular, not {w, t}) is shared. Finally, we construct the one
remaining s-t path P as follows. We lead P from s to x1 via v, then following C ′
in H to x, and then from x to t via w. Observe that P has length n + 3 and contains
the edge {w, t} at time step n + 3. Hence, no edge is shared as the path containing
the (n + 1)-chain contains the edge {w, t} at time step (n + 2). We constructed n − 1
s-t paths in G′ with no shared edge.
Note that the maximum degree of the graph obtained in the Construction 2 depends on the number of vertices in the input graph. In what follows, we give a second
construction where the obtained graph has constant maximum degree ∆ = 4.
10
(η + η)-chain
..
.
(η + 2)-chain
Ts
(η + 1)-chain
v
x2
s
u
(η + log(η) + (η − n))-chain
x1
Tw
H
w
t
log(η)-chain
x3
Figure 2: Ts and Tw refer to complete binary trees with η leaves.
Construction 3. Let G = (V, E) be an undirected, planar, cubic graph with n = |V |.
Construct in time polynomial in the size of G an undirected planar graph G′ as follows
(refer to Figure 2 for an illustration of the constructed graph). Let initially G′ be the
graph obtained from Construction 2. Remove s, w, and the chains connecting s with w
from G′ . Add a vertex u to G′ and add the edges {x2 , u} and {x3 , u} to G′ . Let η be
the smallest power of two larger than n (note that n ≤ η ≤ 2n − 2). Add a complete
binary tree Ts with η leaves to G′ , and denote its root by s. Denote the leaves by
a1 , . . . , aη , ordered by a post-order traversal on Ts . Next add a copy of Ts to G′ , and
denote the copy by Tw and its root by w. If ai is a leaf of Ts , denote by a′i its copy
in Tw . Next, for each i ∈ [η], connect ai and a′i via a (η + i)-chain. Finally, connect s
with v via an (η + log(η) + (η − n))-chain, connect u with w via a log(η)-chain, and
add the edge {w, t} to G′ , which completes the construction of G′ . Note that Ts
and Tw allow plane drawings, and the chains connecting the leaves can be aligned as
illustrated in Figure 2. It follows that G′ allows for a plane embedding.
Lemma 6. Let G and G′ be as in Construction 3. Then G admits a Hamiltonian
cycle if and only if G′ allows for at least η + 1 s-t paths with at most η − 2 shared
edges.
Proof. (⇒) Let G admit a Hamiltonian cycle C. As discussed in the proof of Lemma 5,
there is an ordering C ′ of the vertices in C such that x1 appears first and x ∈ {x2 , x3 }
appears last in C ′ . We construct η s-t paths as follows. We route them from s in Ts
to the leaves of Ts in such a way that each path contains a different leaf of Ts . Herein,
η − 2 edges are shared. Next, route each of them via the chain connecting the leaf to
the corresponding leaf in Tw , then to w, and finally to t. In this part, no edge is shared,
as the lengths of the chains connecting the leaves of Ts and Tw are pairwise different.
Hence, the η s-t paths contain the edge {w, t} at the time steps 2 log(η) + η + i + 1
11
for each i ∈ [η]. We construct the one remaining s-t path P as follows. The path P
contains the chain connecting s with v, the edge {v, x1 }. Then P follows C ′ in H to
x ∈ {x2 , x3 }, via the edge {x, u} to u, via the chain connecting u with w to w, and
finally to t via the edge {w, t}. Observe that P contains the edge {w, t} at the time
step 2η + 2 log(η) + 2, and hence no sharing any further edge in G′ .
(⇐) Let P be a set of η + 1 s-t paths in G′ sharing at most k := η − 2 edges. First
observe that no two paths contain the chain connecting s with v, as otherwise more
that k edges are shared. Hence, at most one s-t path leaves s via the chain to v. It
follows that at least η paths leave s via the edges in Ts . By the definition of paths,
observe that each of s-t paths arrive at a leaf of Ts at same time step. Suppose at
least two s-t paths contain the same leaf of Ts . As each leaf is of degree two, the
s-t paths follow the chain towards a leaf of Tw simultaneously. This introduces at
least η + 1 > k shared edges, contradicting the choice of P. It follows that exactly η
s-t paths leave s via Ts (denote the set by P ′ ), and they arrive each at a different
leaf of Ts at time step log(η). Moreover, by construction, each path in P ′ arrives at a
different leaf of Tw at the time steps log(η) + η + i + 1 for every i ∈ [η].
We next discuss why no path in P ′ contains more than one chain connecting a
pair (a, a′ ) of leaves, where a and a′ are leaves of Ts and Tw , respectively. Assume
that there is a path P ′ ∈ P ′ containing at least two chains connecting the pairs (a, a′ )
and (b, b′ ) of leaves, where a, b and a′ , b′ are leaves of Ts and Tw , respectively, and
vertex a appears at smallest time step over all such leaves of Ts in P ′ and b′ appears
at smallest time step over all such leaves of Tw in P ′ (recall that P ′ must contain a
leaf of Ts at smaller time step than every leaf in Tw ). By construction, a′ and b′ are
the copies of a and b in Tw . Let r denote the vertex in Ts such that r is the root of
the subtree of minimum height in Ts containing a and b as leaves. Let r′ denote its
copy in Tw . Observe that by construction, r′ is the root of the subtree of minimum
height in Tw containing a′ and b′ as leaves. As P ′ starts at vertex s, and the path
from s to a in Ts is unique, P ′ contains the vertex r at smallest time step among the
vertices in X := {r, a, a′ , r′ , b′ , b}. As vertex a appears at smallest time step over all
leaves of Ts and a is of degree two, P ′ contains the vertices a and a′ at second and
third smallest time step, respectively, among the vertices in X. As in any tree, the
unique path between every two leaves contains the root of the subtree of minimum
height containing the leaves, P ′ contains the vertex r′ at fourth smallest time step
among the vertices in X. Finally, as vertex b′ appears at smallest time step over all
leaves of Tw and b′ is of degree two, P ′ contains the vertices b′ and b at fifth and sixth
smallest time steps, respectively, among the vertices in X. In summary, the vertices
in X appear in P ′ in the order (r, a, a′ , r′ , b′ , b). Now, observe that {r, r′ } forms a b-t
separator in G′ , that is, there is no b-t path in G′ − {r, r′ }. As P ′ contains r and r′ at
smaller time steps than b, P ′ contains a vertex different to t at last time step. This
contradict the fact that P ′ is an s-t path in G′ . It follows that no path in P ′ contains
more than one chain connecting a pair consisting of leaf of Ts and a leaf of Tw .
It follows that the paths in P ′ contain the edge {w, t} at the time steps 2 log(η)+η+
i + 1 for each i ∈ [η]. Hence, the remaining s-t path containing the chain connecting s
with v, denoted by P , has to contain the edge {w, t} at time step at most 2 log(η) +
η + 1 or at least 2 log(η) + 2η + 2. At the earliest P can contain the edge {w, t} at
time step 2 log(η) + η + (η − n) + 3 > 2 log(η) + η + 1, and thus, path P has to contain
12
edge {w, t} at time step 2 log(η) + 2η + 2. This is only possible if P forms a path in H
that visits each vertex in H, starting at x1 and ending at vertex x ∈ {x2 , x3 }. As the
edge {x1 , x} is contained in H, it follows that P restricted to H forms a Hamiltonian
cycle in G.
Proof of Theorem 4. We provide a many-one reduction from PCHC to Path-RCA
on undirected graph via Construction 2 (for constant number k of shared edges) on
the one hand, and Construction 3 (for constant maximum degree ∆) on the other. Let
(G) be an instance of PCHC with n = |V (G)|.
Via Construction 2. Let (G′ , s, t, p, 0) be an instance of Path-RCA where G′ is obtained from G by applying Construction 2 and p = n − 1. Note that (G′ , s, t, p, 0) is
constructed in polynomial time and, by Lemma 5, G is a yes-instance of PCHC if and
only if (G′ , s, t, p, 0) is a yes-instance of Path-RCA.
The case of constant k > 0. Reduce (G′ , s, t, p, 0) to an equivalent instance
′
(Gk , s′ , t, p, k) of Path-RCA with k > 0 as follows. Let G′k denote the graph obtained from G′ by the following modification: Add a chain of length k to G′ , and
identify one endpoint with s and denote by s′ the other endpoint. Set s′ as the new
source. Observe that any s′ -t path in G′k contains the k-chain appended on s, and
hence, any solution introduces exactly k shared edges.
The directed case. Direct the edges in G′ as follows. Direct each chain connecting s
with w from s towards t. (In the case of k > 0, also direct the chain from s′ towards s.)
Direct the edges {s, v}, {v, x1 }, {x2 , w}, {x3 , w}, and {w, t} as (s, v), (v, x1 ), (x2 , w),
(x3 , w), and (w, t). Finally, replace each edge {a, b} in H by two (anti-parallel) arcs
(a, b), (b, a) to obtain the directed variant of H. The correctness follows from the fact
that we consider paths that are not allowed to contain vertices more than once. Note
that the planarity is not destroyed.
Via Construction 3. Let (G′ , s, t, p, k) be an instance of Path-RCA where G′ is
obtained from G by applying Construction 3, p = η + 1, and k = η − 2. Note
that (G′ , s, t, p, k) is constructed in polynomial time and, by Lemma 6, G is a yesinstance of PCHC if and only if (G′ , s, t, p, k) is a yes-instance of Path-RCA.
The directed case. Direct the edges in G′ as follows. Direct the edges in Ts from s
towards the leaves, and the edges in Tw from the leaves towards w. Direct each chain
connecting Ts with Tw from Ts towards Tw . Direct the edges {v, x1 }, {x2 , u}, {x3 , u},
and {w, t} as (v, x1 ), (x2 , u), (x3 , u), and (w, t). Direct the chain connecting s with v
from s towards v, and the chain connecting u with w from u towards w. Finally,
replace each edge {a, b} in H by two (anti-parallel) arcs (a, b), (b, a) to obtain the
directed variant of H. The correctness follows from the fact that we consider paths
that are not allowed to contain vertices more than once. Note that the planarity is
not destroyed.
As the length of every s-t path is upper bounded by the number of vertices in the
graph, we immediately obtain the following.
Corollary 3. Path-FRCA both on undirected planar and directed planar graphs is
NP-complete, even if k ≥ 0 is constant or ∆ ≥ 4 is constant.
13
n−1
H′
...
s
x
w
t
v
Figure 3: Graph G′ obtained in Construction 4.
graph H ′ .
5
The gray part represents the
Trail-RCA
We now show that Trail-RCA has the same computational complexity fingerprint
as Path-RCA. That is, Trail-RCA (Trail-FRCA) is NP-complete on undirected
and directed planar graphs, even if the number k ≥ 0 of shared edges (arcs) or the
maximum degree ∆ ≥ 5 (∆i/o ≥ 3) is constant. The reductions are slightly more
involved, because it is harder to force trails to take a certain way.
5.1
On Undirected Graphs
In this section, we prove the following.
Theorem 5. Trail-RCA on undirected planar graphs is NP-complete, even if k ≥ 0
is constant or ∆ ≥ 5 is constant.
We provide two constructions supporting the two subresults for constants k, ∆.
The reductions are again from Planar Cubic Hamiltonian Cycle (PCHC).
Construction 4. Let G = (V, E) be an undirected planar cubic graph with n = |V |.
Construct an undirected planar graph G′ as follows (refer to Figure 3 for an illustration
of the constructed graph). Initially, let G′ be the empty graph. Add a copy of G
to G′ and denote the copy by H. Subdivide each edge in H and denote the resulting
graph H ′ . Note that H ′ is still planar. Consider a plane embedding φ(H ′ ) of H ′ and
and let x ∈ V (H ′ ) be a vertex incident to the outer face in the embedding. Next, add
the vertex set {s, v, w, t} to G. Add the edges {s, x}, {s, v}, {v, w}, and {w, t} to G.
Finally, add n − 1 vertices B = {b1 , . . . , bn−1 } to G and connect each of them with s
by two edges (in the following, we distinguish these edges as {s, bi }1 and {s, bi }2 , for
each i ∈ [n − 1]). Note that the graph is planar (see Figure 3 for an embedding, where
H ′ is embedded as φ(H ′ )) but not simple.
Lemma 7. Let G and G′ as in Construction 4. Then G admits a Hamiltonian cycle
if and only if G′ admits 2n s-t trails with no shared edge.
Proof. (⇒) Let G admit a Hamiltonian cycle C. Observe that H ′ allows for a cycle C ′
in H ′ that contains each vertex corresponding to a vertex in H exactly once. We
construct 2n trails in G′ as follows.
We group the trails in two groups. The first group of trails first visits some of
the vertices b1 , . . . , bn−1 by each time first using the edge {s, bi }1 , i ∈ [n − 1], and
14
then proceeding to t via v. The second group of trails first visits some of the vertices b1 , . . . , bn−1 by each time first using the edge {s, bi }2 , i ∈ [n − 1], and then
proceeding via x, then following the cycle C ′ , and finally again via x towards t. Let
T1i , . . . , Tni denote the trails of group i ∈ {1, 2}. For each j < n, the trail Tji first visits
the vertices bj , . . . , bn−1 in that order before proceeding as described above. The trails
Tni , i ∈ {1, 2}, do not contain any of the vertices b1 , . . . , bn−1 , and directly approach t
as described above.
Observe that, within each of the two groups, no two trails share an edge. Between
trails of different groups, only edge {w, t} can possibly be shared. Note that any cycle
in H ′ is of even length. Hence, the trails of group 1 contain the edge {w, t} at each of
the time steps 2j + 1 for every j ∈ [n]. The trails of group 2 contain the edge {w, t}
at each of the time steps 2n + 2j + 1 for every j ∈ [n]. Hence, no two trails share an
edge.
(⇐) Let G′ admit 2n s-t trails with no shared edge. First note that s has exactly 2n
incident edges. Observe that, for each β = 0, . . . , |B|, no more than two trails contain
β vertices of B, as otherwise any of the edges {s, v} or {s, x} would be shared. By
the pigeon hole principle it follows that, for each β = 0, . . . , |B|, there are exactly
two trails that contain β vertices of B. Hence, for each even time step, there are two
trails leaving s via the edges {s, v} and {s, x}, respectively. Observe that those trails
that proceed towards t via v use the edges {w, t} exactly at the time steps 2j + 1
for every j ∈ [n]. Because each trial in H ′ that starts and ends at the same vertex
has even length, those trials that proceed towards t via x can use {w, t} only at odd
time steps. Hence, since the edge {w, t} is not shared, the trails proceeding towards t
via x need to stay in H ′ for 2n time steps. As H ′ − x has maximum degree three, no
vertex in H ′ beside x is contained more than once in all of these trails. As the length
between every two vertices in H ′ corresponding to vertices in H, it follows that every
of these trails visits the vertices in H ′ corresponding to the vertices in H. It follows
that each of these trails forms a Hamiltonian cycle C ′ in H ′ . As C ′ can easily turned
into a Hamiltonian cycle C in G (consider the sequence when deleting all vertices that
do not correspond to a vertex in H), the statement follows.
To deal with the parallel edges in graph G′ in Construction 4, we now subdivide
edges, maintaining an equivalent statement as in Lemma 7.
Lemma 8. Let G be an undirected graph (not necessarily simple) with two distinct
vertices s and t. Obtain graph G′ from G by replacing each edge {u, v} ∈ E in G by
a path of length three, identifying its endpoints with u and v. Then G admits p ∈ N
s-t trails with no shared edge if and only if G′ admits p s-t trails with no shared edge.
Proof. For each edge e ∈ E(G) in G denote by P (e) the corresponding path of length
three in G′ . By definition, for all e, f ∈ E(G′ ) it holds that e 6= f if and only if
P (e) 6= P (f ).
(⇒) Let P = {Pi | i ∈ [p]} be a set of p s-t trails in G with no shared edge. Each Pi
as an edge sequence representation Pi = (ei1 , . . . , eiℓi ), where ℓi is the number of edges
in Pi . For each Pi , consider the corresponding trail Pi′ = (P (ei1 ), . . . , P (eiℓi )) in G′ ,
and the set P ′ = {Pi′ | i ∈ [p]}. Suppose that two trails Pi′ and Pj′ share an edge.
15
n−1
un−1
u2
u1
b0
bn−1 bn−2 b2
b1
...
s
ℓn−1
ℓ2
ℓ1
H ′′
2n-chain
2n-chain
Figure 4: Graph G′ obtained in Construction 5.
graph H ′ .
x
w
t
v
The gray part represents the
Then the shared edge is contained in subpaths P (eix ) and P (ejx ). As P (eix ) and P (ejx )
are not edge-disjoint (as they share an edge), it follows that eix = ejx , and hence Pi
and Pj share the edge eix in G. This contradicts the fact that P = {Pi | i ∈ [p]} is a
set of p s-t trails in G with no shared edge. It follows that P ′ is a set of p s-t trails
in G′ with no shared edge.
(⇐) Let P ′ = {Pi′ | i ∈ [p]} be a set of p s-t trails in G′ with no shared edge. Observe
that, by the construction of G′ , each s-t trail in G′ is composed of paths of length
three with endpoints corresponding to vertices in G. Hence, for each i ∈ [p], let Pi′ be
represented as Pi′ = (P (ei1 ), . . . , P (eiℓi )), where 3 · ℓi is the number of edges in Pi′ . For
each Pi′ , consider the corresponding trail Pi = (ei1 , . . . , eiℓi ) in G, and the set P = {Pi |
i ∈ [p]}. Suppose that two trails Pi and Pj share an edge, that is, there is an index x
such that eix = ejx . It follows that P (eix ) = P (ejx ). Let eix = {v, w} =: e. If both
trails Pi′ and Pj′ traverse P (e) in the same “direction”, i.e. either from v to w or from w
to v, then Pi′ and Pj′ share at least three edges (all edges in P (e)). This contradicts the
definition of P ′ . Consider the case that the trails Pi′ and Pj′ traverse P (e) in opposite
“directions”, i.e. one from v to w and the other from w to v. As P (e) is of length three,
the edge in P (e) with no endpoint in {v, w} is then used by Pi′ and Pj′ at the same
time step, yielding that the edge is shared. This contradicts the definition of P ′ . It
follows that P = {Pi | i ∈ [p]} is a set of p s-t trails in G with no shared edge.
We now show how to modify Construction 4 for maximum degree five, giving up,
however, a constant upper bound on the number of shared edges.
Construction 5. Let G = (V, E) be an undirected planar cubic graph with n = |V |.
Construct an undirected planar graph G′ as follows (see Figure 4 for an illustration of
the constructed graph). Let initially G′ be the graph obtained from Construction 4.
Subdivide each edge in H ′ and denote the resulting graph by H ′′ . Observe that the
distance in H ′′ between any two vertices in V (H ′′ ) ∩ V (H ′ ) is divisible by four. Next,
delete all edges incident with vertex s. Connect s with v via a 2n-chain, and connect s
with x via a 2n-chain. Connect s with b1 via two P2 ’s. Denote the two vertices on
the P2 ’s by ℓ1 and u1 . Finally, for each i ∈ [n − 2], connect bi with bi+1 via two P2 ’s.
For each i ∈ [n − 2], denote the two vertices on the P2 ’s between bi and bi+1 by ℓi+1
and ui+1 . For an easier notation, we denote vertex s also by b0 .
16
Lemma 9. Let G and G′ as in Construction 5. Then G admits a Hamiltonian cycle
if and only if G′ has 2n s-t trails with at most 2n − 4 shared edges.
Proof. (⇐) Let G′ admit a set P of 2n s-t trails with at most 2n − 4 shared edges. At
each time step, at most two trails leave s towards v and x. Otherwise, all the edges in at
least one of the 2n-chains connecting s with v and s with x are shared, contradicting
the fact that the trails in P share at most 2n − 4 edges. Note that every s-t trail
contains vertex s at the first time step and at most once more at time step 4j + 1, for
some j ∈ N (indeed, we will show that j ∈ [n − 1]). This follows on the one hand from
the fact that s has degree four and hence every trail can contain s at most twice, and
on the other hand from the fact that for each i ∈ [n − 1], every s-bi path is of even
length.
We show that at each time step 4j + 1, 0 ≤ j ≤ n − 1, exactly one s-t trial
leaves s towards v and exactly one s-t trial leaves s towards x. First, observe that
|{bi , ui , ℓi | i ∈ [n − 1]}| = 3(n − 1) and each trail can contain each vertex in {ui , ℓi |
i ∈ [n − 1]} ∪ {bn−1 } at most once (as each vertex in this set is of degree two)
and each vertex bi , i ∈ [n − 2], at most twice (as they are of degree four). Hence,
any trail starting on s and returning to s after visiting the vertices in B contains at
most 3(n−1)+(n−2)+2 = 4(n−1)+1 vertices. It follows that every s-t trail contains s
at the first time step and at most once more at time step 4j + 1 for some j ∈ [n − 1].
As there are 2n s-t trails and at each time step at most two trails leave s towards v
and x, together with the pigeon hole principle it follows that exactly one s-t trial
leaves s towards v and exactly one s-t trial leaves s towards x at each time step 4j + 1,
0 ≤ j ≤ n − 1. Moreover, note that each bi , i ∈ [n − 1], appears in at least two
s-t trails.
We claim that there are exactly 2n − 4 shared edges and that every shared edge is
incident with a vertex in {ui , ℓi | i ∈ [n − 1]}. This follows from the fact that at least
three trails going at the same time from bi to bi+1 , 0 ≤ i ≤ n − 3, share at least two
edges. As four trails contain bn−2 (those four which leave s towards v and x at the
time steps 4j + 1 with j ∈ {n − 2, n − 1}), it follows that at least 2(n − 2) edges are
shared. Hence, no two trails share an edge after they have left s for v or x.
There is a trail P ∈ P that contains the vertex x and that contains vertex s only
once at the first time step, because at each time step, two trails leave s for v or x.
Observe that vertex w is contained in the trails containing v at the time steps 4j+2n+2
for all j ∈ [n − 1] ∪ {0} whence edge {w, t} is occupied at time steps 4j + 2n + 3 for
all j ∈ [n − 1] ∪ {0}. Hence, the edge {w, t} is contained at time step 2n + 2 in a trail
different to P and, thus, trail P contains at least one vertex in H ′′ . Furthermore, P
can contain x a second time only at time steps of the form 4j + 2n + 1, 3 ≤ j ≤ n,
because each path in H ′′ between two vertices that correspond to vertices in G has
length four. However, as mentioned, {w, t} is occupied at time steps 4j + 2n + 3,
j ∈ [n − 1]. Hence, P has to stay in H ′′ for 4n time steps. Recall that G is cubic, and
hence no vertex in H ′′ − {x} corresponding to a vertex in G appears more than once
in any trail. That is, P follows a cycle in H ′′ containing each vertex corresponding to
vertex in G exactly once. It follows that G admits a Hamiltonian cycle.
(⇒) Let G admit a Hamiltonian cycle C. Let C ′ be C, ordered such that x is the
first and last vertex in C ′ . Let C ′′ denote the cycle in H ′′ following the order of the
17
vertices in C ′ . We construct 2n s-t trails in G′ sharing at most 2n − 4 edges as follows.
We denote the trails by Pix , 0 ≤ i ≤ n − 1, x ∈ {u, ℓ}. The trails are divided into two
groups according to their superscript x ∈ {u, ℓ}. For i ≥ 1, the trails Piu and Piℓ start
with the sequences
Piu :
(s, ℓ1 , b1 , . . . , ℓi−1 , bi−1 , ui , bi , ℓi , bi−1 , ℓi−1 , . . . , b1 , ℓ1 , s),
Piℓ
(s, ℓ1 , b1 , . . . , ℓi−1 , bi−1 , ℓi , bi , ui , bi−1 , ℓi−1 , . . . , b1 , ℓ1 , s).
:
Trails P0u and P0ℓ do not visit any vertex in B and simply start at s. Then, for each
i = 0, . . . , n − 1, trail Piu follows the chain to v, the edge to w, and then to t. For each
i = 0, . . . , n − 1, trail Piℓ follows the chain to x, then the cycle C ′′ in H ′′ , then the edge
from x to w, then to t. Observe that trail Pix , x ∈ {u, ℓ}, contains s at time step one
and 4i + 1. Hence, Piu contains the vertex w at time step 4i + 2n + 2. Moreover, Piℓ
contains the vertex x at time steps 4i + 2n + 1 and 4i + 2n + 4n + 1. From the latter
it follows that Piℓ contains the vertex w at time step 4i + 2n + 4n + 2. Altogether, the
edge {w, t} is not shared by any pair of trails.
Next, we count the number of edges shared between the two visits of s. Denote by
X ⊆ E(G′ ) the set {{bi , ℓi } | 1 ≤ i ≤ n − 2} ∪ {{bi , ℓi+1 } | 0 ≤ i < n − 2}. Observe
that |X| = n − 2 + n − 2 = 2(n − 2). We claim that the edges in X are the only
ℓ
u
shared edges by the trails Pix , 0 ≤ i ≤ n − 1, x ∈ {u, ℓ}. As Pn−1
and Pn−1
contain
the set X at the same time steps, every edge in X is shared. For each i ≥ 1, the
edges {ui , bi−1 } and {ui , bi } are only contained in the trails Pix , x ∈ {u, ℓ}. Recall
that Piu and Piℓ contain bi exactly once and at the same time step. The subsequence
around bi of Piℓ and Piu is (bi−1 , ℓi , bi , ui , bi−1 ) and (bi−1 , ui , bi , ℓi , bi−1 ), respectively.
It follows that both edges {ui , bi−1 } and {ui , bi } appear at two different time steps
in Piu and Piℓ . The same argument holds for the edges {bn−2 , ℓn−1 } and {bn−1 , ℓn−1 }
ℓ
u
are the only trails containing the two edges.
and Pn−1
as Pn−1
Altogether, it follows that X is the set of shared edges of the s-t trails Pix , x ∈
{u, ℓ}, and the claim follows. Finally, as |X| = 2n − 4, the statement follows.
Proof of Theorem 5. We provide a many-one reduction from Planar Cubic Hamilton Circuit (PCHC) to Trail-RCA on undirected graph via Construction 4 on
the one hand, and Construction 5 on the other. Let (G = (V, E)) be an instance of
PCHC and let n := |V | vertices.
Via Construction 4. Let (G′ , s, t, p, 0) an instance of Trail-RCA where G′ is obtained
from G by applying Construction 4 and p = 2n. Note that (G′ , s, t, p, 0) can be
constructed in polynomial time and by Lemma 7, (G) is a yes-instance of PCHC if
and only if (G′ , s, t, p, 0) is a yes-instance of Trail-RCA. However, G′ is not simple
in general. Hence, replace each edge {u, v} ∈ E(G′ ) in G′ by a path of length three
and identify its endpoints with u and v. Denote by G′′ the obtained graph. Due
to Lemma 8, (G′′ , s, t, p, 0) is a yes-instance of Trail-RCA if and only if (G′ , s, t, p, 0)
is a yes-instance of Trail-RCA.
The case of constant k > 0 works analogously as in the proof of Theorem 4.
Via Construction 5. Let (G′ , s, t, p, k) an instance of Trail-RCA where G′ is obtained from G by applying Construction 5 and p = 2n. Instance (G′ , s, t, p, k) can be
18
(n + 2)-chain
..
.
4-chain
3-chain
x
s
v
w
t
H
Figure 5: Sketch of the graph G′ obtained in Construction 6. The enclosed gray part
represents the graph H. Dashed lines represent directed chains.
constructed in polynomial time. By Lemma 9, (G) is a yes-instance of PCHC if and
only if (G′ , s, t, p, k) is a yes-instance of Trail-RCA.
As the length of each s-t trail is upper bounded by the number of edges in the
graph, we immediately obtain the following.
Corollary 4. Trail-FRCA on undirected planar graphs is NP-complete, even if k ≥
0 is constant or ∆ ≥ 5 is constant.
5.2
On Directed Graphs
We know that Trail-RCA and Trail-FRCA are NP-complete on undirected graphs,
even if the number of shared edges or the maximum degree is constant. In what
follows, we show that this is also the case for Trail-RCA and Trail-FRCA on
directed graphs.
Theorem 6. Trail-RCA on directed planar graphs is NP-complete, even if k ≥ 0 is
constant or ∆i/o ≥ 3 is constant.
To prove Theorem 6, we reduce from the following NP-complete [17] problem.
Directed Planar 2/3-In-Out Hamiltonian Circuit (DP2/3HC)
Input:
A directed, planar graph G = (V, A) such that, for each v ∈ V ,
max{outdeg(v), indeg(v)} ≤ 2 and outdeg(v) + indeg(v) ≤ 3.
Question: Is there a directed Hamiltonian cycle in G?
Construction 6. Let G = (V, A) be a directed, planar graph where for each vertex
v ∈ V holds max{outdeg(v), indeg(v)} ≤ 2 and outdeg(v)+indeg(v) ≤ 3, and n = |V |.
Construct a directed graph G′ as follows (refer to Figure 5for an illustration of the
constructed graph). Initially, let G′ be the empty graph. Add a copy of the graph G to
G′ and denote the copy by H. Add the vertex set {s, t, v, w} to G′ . Consider a plane
embedding φ(G) and choose a vertex x ∈ V (H) incident to the outer face. Add the
arcs (s, v), (v, x), (x, w), and (w, t) to G′ . Moreover, add n chains connecting s with v
of lengths 3, 4, . . . , n+2 respectively to G′ , and direct the edges from s towards v. Note
that G′ is planar (see Figure 5 for an embedding where H is embedded as φ(H)).
19
(η + η)-chain
..
.
Ts↑
(η + 2)-chain
(η + 1)-chain
v
Tw↓
x
s
(log(η) + η + (η − n))-chain
log(η)-chain
w
t
H
Figure 6: Sketch of the graph G′ obtained in Construction 7. The enclosed gray part
represents the graph H. Dashed lines represent directed chains. Ts↑ and Tw↓ refer to
the complete binary (directed) trees with η leaves and rooted at s and w, respectively.
Lemma 10. Let G and G′ as in Construction 6. Then G admits a Hamiltonian cycle
if and only if G′ admits n + 1 s-t trails with no shared arc.
Proof. (⇒) Let G admit a Hamiltonian cycle C. We construct n s-t trails in G′ , each
using a chain connecting s with w, where no two use the same chain. By construction,
the n s-t trails do not introduce any shared arc. Moreover, the trails contain the
arc (w, t) at every time step in {4, . . . , n + 3}. The remaining trail contains no chain,
but the vertices v, x, w as well as C. As C is a Hamiltonian cycle, the trail uses the
arc (w, t) at time step n + 4.
(⇐) Let G′ admit a set P of n+1 s-t trails with no shared arc. As s has outdegree n+1,
in P n trails contain a chain connecting s with w, where no two contain the same
chain. As there is no shared arc, the remaining trail cannot use the arc (w, t) before
time step n+3. As the shortest s-w path containing v is of length three, the remaining
trail has to contain n arcs in the copy H of G. As for each vertex v ∈ V (G) holds
that max{outdeg(v), indeg(v)} ≤ 2 and outdeg(v) + indeg(v) ≤ 3, no vertex despite x
is visited twice by the trail. Hence, the trail restricted to the copy H of G forms an
Hamiltonian cycle in G.
We provide another construction where the obtained graph has constant maximum
in- and out-degree.
Construction 7. Let G = (V, A) be a directed, planar graph where for each vertex
v ∈ V holds max{outdeg(v), indeg(v)} ≤ 2 and outdeg(v) + indeg(v) ≤ 3, and let
n := |V |. Construct a directed graph G′ as follows (refer to Figure 6 for an illustration
of the constructed graph). Let initially G′ be the graph obtained from Construction 6.
Remove s, w, and the directed chains connecting s with w from G′ . Let η be the
smallest power of two larger than n (note that n ≤ η ≤ 2n − 2). Add a complete
binary tree Ts with η leaves to G′ , and denote its root by s. Denote the leaves by
a1 , . . . , aη , ordered by a post-order traversal on Ts . Next add a copy of Ts to G′ ,
and denote the copy by Tw and its root by w. For each leaf ai of Ts , denote by a′i
20
its copy in Tw . Next, for each i ∈ [η], connect ai and a′i via a (η + i)-chain. Direct
all edges in Ts away from s towards the leaves of Ts . Direct all edges in Tw away
from the leaves of Tw towards w. Next, direct all chains connecting the leaves of Ts
and Tw from the leaves of Ts towards the leaves of Tw . To complete the construction
of G′ , connect s with v via a (log(η) + η + (η − n))-chain, and connect x with w via
a log(η)-chain. Note that Ts and Tw allow plane drawings, and the chains connecting
the leaves can be aligned as illustrated in Figure 2. It follows that G′ allows for an
plane embedding.
Lemma 11. Let G and G′ as in Construction 7. Then G admits a Hamiltonian cycle
if and only if G′ admits η + 1 s-t trails with at most η − 2 shared arcs.
Proof. (⇒) Let G admit a Hamiltonian cycle C. First, we construct η s-t trails in G′
as follows. For each leaf of Ts , there is a trail containing the unique path from s to
the leaf in Ts . This part introduces η − 2 shared arcs. Next, each trail follows the
chain connecting the leaf of Ts with a leaf of Tw , then the unique path from the leaf
of Tw to w, and finally the arc (w, t). Observe that the trails contain the leaves of Tw
at different time steps log(η) + η + i + 1, i ∈ [η], and hence no shared arc is introduced
in this part. Moreover, the arc (w, t) appears in the time steps 2 log(η) + η + i + 1,
i ∈ [η]. The remaining trail P contains the chain connecting s with v, the edge {v, x},
follows the cycle C in H, starting and ending at vertex x. Trail P then contains the
chain connecting x with w, and the arc (w, t). Observe that the arc (w, t) appears
in P at time step 2 log(η) + 2η + 2 (recall that C is a Hamiltonian cycle in H), and
hence (w, t) is not shared.
(⇐) Let G′ admit η + 1 s-t trails with at most η − 2 shared arcs.
First observe that the chain connecting s with v is not contained in more than one st trail. Hence, at least η trails leave s through Ts . Note that no chain connecting the
leaves of Ts with the leaves of Tw is contained in more than one s-t trail. It follows
that exactly η s-t trails (denote the set by P ′ ) leave s via Ts and each contains a
different leaf of Ts . Herein, η − 2 arcs are shared by the trails in P ′ . Note that the
path from a leaf of Ts to t is unique, each trail in P ′ follows the unique path to t. The
arc (w, t) appears in the trails in P ′ at time steps 2 log(η) + η + i + 1 for every i ∈ [η].
The remaining s-t trail P 6∈ P contains the chain connecting s with v. Note that
arc (w, t) is not shared, as all shared arcs are contained in Ts . As the shortest st path via x is of length 2 log(η) + η + (η − n) + 2 ≥ 2 log(η) + η + 2, trail P has
to contain a cycle C in H. As the arc (w, t) is not shared and appears in the trails
in P ′ at the time steps 2 log(η) + η + i + 1, for every i ∈ [η], the cycle C must be of
length n. As for each vertex v ∈ V (G) holds that max{outdeg(v), indeg(v)} ≤ 2 and
outdeg(v) + indeg(v) ≤ 3, no vertex in H despite x is visited twice by the trail P .
Hence, trail P restricted to the copy H of G forms a Hamiltonian cycle in G.
Proof of Theorem 6. We provide a many-one reduction from DP2/3HC to TrailRCA on directed graphs via Construction 6 on the one hand, and Construction 7 on
the other. Let (G) be an instance of DP2/3HC where G consists of n vertices.
Via Construction 6. Let (G′ , s, t, p, 0) an instance of Trail-RCA where G′ is obtained from G by applying Construction 6 and p = n + 1. Note that (G′ , s, t, p, 0) is
21
constructed in polynomial time and by Lemma 10, (G) is a yes-instance of DP2/3HC
if and only if (G′ , s, t, p, 0) is a yes-instance of Trail-RCA.
The case of constant k > 0 works analogously as in the proof of Theorem 4.
Via Construction 7. Let (G′ , s, t, p, k) an instance of Trail-RCA where G′ is obtained
from G by applying Construction 7, p = η+1, and k = η−2. Note that (G′ , s, t, p, k) is
constructed in polynomial time and maxv∈V (G′ ) {outdeg(v)+indeg(v)} ≤ 5. By Lemma 11,
(G) is a yes-instance of DP2/3HC if and only if (G′ , s, t, p, k) is a yes-instance of
Trail-RCA.
As the length of each s-t trail is upper bounded by the number of edges in the
graph, we immediately obtain the following.
Corollary 5. Trail-FRCA on directed planar graphs is NP-complete, even if k ≥ 0
is constant or ∆i/o ≥ 3 is constant.
6
Walk-RCA
Regarding their computational complexity fingerprint, Path-RCA and Trail-RCA
are equal. In this section, we show that Walk-RCA differs in this aspect. We prove
that the problem is solvable in polynomial time on undirected graphs (Section 6.1)
and on directed graphs if k ≥ 0 is constant(Section 6.2).
6.1
On Undirected Graphs
On a high level, the tractability on undirected graphs is because a walk can alternate
arbitrarily often between two vertices. Hence, we can model a queue on the source
vertex s, where at distinct time steps the walks leave s via a shortest path towards t.
However, if the time of staying in the queue is upper bounded, that is, if the lengthrestricted variant Walk-FRCA is considered, the problem becomes NP-complete.
Theorem 7. Walk-RCA on undirected graphs is solvable in linear time.
Proof. Let I := (G, s, t, p, k) be an instance of Walk-RCA with G being connected.
Let P be a shortest s-t path in G. We assume that p ≥ 2, since otherwise P witnesses
that I is a yes-instance. We can assume that the length of P is at least k +1, otherwise
we can output that I is a yes-instance. Let {s, v} be the edge in P incident to the
endpoint s. We distinguish the two cases whether k is positive or k = 0. In this proof,
we represent a walk as a sequence of edges.
Case k > 0: We can construct p s-t walks P1 , . . . , Pp sharing at most one edge as
follows. We set P1 := P and Pi = ({s, v}, . . . , {s, v}, P ) for i ∈ [p], that is, the s{z
}
|
2i-times
t walk Pi alternates between s and v i times. We show that the set P := {P1 , . . . , Pp }
share exactly edge {s, v}. It is easy to see that {s, v} is shared by all of the walks. Let
us consider an arbitrary edge e = {x, y} 6= {s, v} in P , which appears in P at time step
ℓ > 1 (ordered from s to t). By construction, e appears in Pi at position 2i + ℓ. Thus,
no two walks in P contain an edge in P , that is on time step ℓ > 1 in P , at the same
22
time step. Since each walk in P only contains edges in P , it follows that {s, v} is
exactly the shared edges by all walks in P. As k ≥ 1, it follows that we can output
that I is a yes-instance.
Case k = 0: Let v1 , . . . , vℓ be the neighbors of s, and suppose that v1 = v (that is, the
vertex incident to s appearing in P ). If ℓ < p, then we can immediately output that I
is a no-instance, as by the pigeon hole principle at least one edge has to appear in at
least two walks at time step one in any set of p s-t walks in G. If ℓ ≥ p, we construct
ℓ s-t walks P1 , . . . , Pℓ that do not share any edge in G as follows. We set P1 = P , and
Pi = ({s, vi }, . . . , {s, vi }, P ) for i ∈ [ℓ], that is, the s-t walk Pi alternates between s
{z
}
|
2i-times
and vi i times. Following the same argumentation as in preceding case, it follows that
no edge is shared by the constructed walks P1 , . . . , Pp .
In summary, if k > 0, then we can output that I is a yes-instance. If k = 0, then
we first check the degree of s in linear time, and then output that I is a yes-instance
if deg(s) ≥ p, and that I is a no-instance, otherwise.
The situation changes for Walk-FRCA, that is, when restricting the length of the
walks.
Theorem 8. Walk-FRCA on undirected graphs is NP-complete and W[2]-hard with
respect to k + α.
Given a directed graph G, we call an undirected graph H the undirected version
of G if H is obtained from G by replacing each arc with an undirected edge.
Proof. We give a (parameterized) many-one reduction from Set Cover. Let (U, F , ℓ)
be an instance of Set Cover. Let G′ the graph obtained from applying Construction 1
given (U, F , ℓ). Moreover, let G be the undirected version of G′ . Let (G, s, t, p, k, α)
be the instance of Walk-FRCA, where p = n + m, k = ℓ, and α = ℓ + 3. Note that
any shortest s-t path in G is of length α, and hence every s-t walk of length at most α
behaves as in the directed acyclic case. That is, each walk contains a vertex of VF at
time step k + 2. The correctness follows then analogously as in the proof of Lemma 4.
It remains open whether Walk-FRCA is NP-complete when k is constant.
6.2
On Directed Graphs
Due to Theorems 2 and 3, we know that Walk-RCA is NP-complete on directed
graphs and is solvable in polynomial time on directed acyclic graphs when k = 0,
respectively. In this section, we prove that if k ≥ 0 is constant, then Walk-RCA
remains tractable on directed graphs (this also holds true for Walk-FRCA). Note
that for Path-RCA and Trail-RCA the situation is different, as both become NPcomplete on directed graphs, even if k ≥ 0 is constant.
Theorem 9. Walk-RCA and Walk-FRCA on directed n-vertex m-arc graphs is
solvable in O(mk+1 · n · (p · n)2 ) time and O(mk+1 · n · α2 ) time, respectively.
23
Our proof of Theorem 9 follows the same strategy as our proof of Theorem 1. That
is, we try to guess the shared arcs, make them infinite capacity in some way, and then
solve the problem with zero shared arcs via a network flow formulation in the timeexpanded graph. The crucial difference is that here we do not have at first an upper
bound on the length of the walks in the solution.
Theorem 10. If k = 0, then Walk-RCA on directed n-vertex m-arc graphs is solvable in O(n · m · (p · n)2 ) time.
Lemma 12. Every yes-instance (G, s, t, p, k) of Walk-RCA on directed graphs admits a solution in which the longest walk is of length at most p · dt , where dt =
maxv∈V : distG (v,t)<∞ distG (v, t).
Observe that dt is well-defined on every yes-instance of Walk-RCA. Moreover, it
holds that dt ≤ |V (G)|. In the subsequent proof, we use the following notation: For
two walks P1 = (v1 , . . . , vℓ ) and P2 = (w1 , . . . , wℓ′ ) with vℓ = w1 , denote by P1 ◦ P2
the walk (v1 , . . . , vℓ , w2 , . . . , wℓ′ ) obtained by the concatenation of the two walks.
Proof of Lemma 12. Let P be a solution to (G, s, t, p, k) with |P| = p where the sum of
the lengths of the walks in P is minimum among all solutions to (G, s, t, p, k). Suppose
towards a contradiction that the longest walk P ∗ ∈ P is of length |P ∗ | > p · dt . Then,
by the pigeon hole principle, there is an i ∈ [p] such that there is no walk in P of
length ℓ with (i − 1) · dt < ℓ ≤ i · dt .
Let v = P ∗ [(i − 1) · dt + 1], that is, v is the ((i − 1) · dt + 1)th vertex on P ∗ , and
let S be a shortest v-t path. Observe that the length of S is at most dt . Consider the
walk P ′ := P ∗ [1, (i−1)·dt +1]◦S, that is, we concatenate the length-((i−1)·dt ) initial
subpath of P ∗ with S to obtain P ′ . Observe that (i−1)·dt < |P ′ | ≤ i·dt . If P \P ∗ ∪P ′
forms a solution to (G, s, t, p, k), then, since |P ′ | < |P ∗ |, P \ P ∗ ∪ P ′ is a solution of
smaller sum of the lengths of the walks, contradicting the choice of P. Otherwise,
P ′ introduce additional shared arcs and let A′ ⊆ A(G) denote the corresponding set.
Observe that A′ is a subset of the arcs of S. Let a = (x, y) ∈ A′ be the shared arc such
that distS (y, t) is minimum among all shared arcs in A′ , and let P ′ [j] = y. Let P ∈ P
be a walk sharing the arc with P ′ . Then P ′′ := P [1, j] ◦ P ′ [j + 1, |P ′ |] is a walk of
shorter length than P .Moreover, P \P ∪P ′′ is a solution to (G, s, t, p, k). As |P ′′ | < |P |,
P \ P ∪ P ′′ is a solution of smaller sum of the lengths of the walks, contradicting the
choice of P. As either case yields a contradiction, it follows that |P ∗ | ≤ p · dt .
The subsequent proof of Theorem 10 relies on time-expanded graphs. Due to
Lemma 12, we know that the time-horizon is bounded polynomially in the input size.
Proof of Theorem 10. Let (G, s, t, p, 0) be an instance of Walk-RCA where G =
(V, A) is an directed graph. We first compute dt in linear time. Let τ := p · dt . Next,
we compute the τ -time-expanded (directed) graph H = (V ′ , A′ ) of G with p additional
arcs (ti−1 , ti ) for each i ∈ [τ ]. We compute in O(τ 2 · (|V | · |A|) ⊆ O(p2 · (|V |3 · |A|) time
the value of a maximum s0 -tτ flow in H. Due to Lemma 12 together with Lemma 2,
the theorem follows.
Restricting to α-time-expanded graphs yields the following.
24
Corollary 6. If k = 0, Walk-FRCA on directed n-vertex m-arc graphs is solvable
in O(n · m · α2 ) time.
Proof of Theorem 9. Let (G = (V, E), s, t, p, k) be an instance of Walk-RCA with G
being a directed graph. For each k-sized subset K ⊆ A of arcs in G, we decide the instance (G(K, p), s, t, p, 0). The statement for Walk-RCA then follows from Lemma 3
and Theorem 10. The running time of the algorithm is in O(|A|k · p2 · (|V |3 · |A|)).
The statement for Walk-FRCA then follows from Lemma 3 and Corollary 6.
7
Conclusion and Outlook
Some of our results can be seen as a parameterized complexity study of RCA focusing
on the number k of shared edges. It is interesting to study the problem with respect to
other parameters. Herein, the first natural parameterization is the number of routes.
Recall that the Minimum Shared Edges problem is fixed-parameter tractable with
respect to the number of path [6]. A second parameterization we consider as interesting
is the combined parameter maximum degree plus k. In our NP-completeness results for
Path-RCA and Trail-RCA it seemed difficult to achieve constant k and maximum
degree at the same time.
Another research direction is to further investigate on which graph classes PathRCA and Trail-RCA become tractable. We proved that that both problems remain
NP-complete even on planar graphs. Do both Path-RCA and Trail-RCA remain
NP-complete on graphs of bounded treewidth? Recall that the Minimum Shared Edges
problem is tractable on this graph class [20, 1].
Finally, we proved that on undirected graphs, Walk-RCA is solvable in polynomial time while Walk-FRCA is NP-complete. However, we left open whether WalkFRCA on undirected graphs is NP-complete or polynomial-time solvable when k is
constant.
References
[1] Yusuke Aoki, Bjarni V. Halldórsson, Magnús M. Halldórsson, Takehiro Ito, Christian Konrad, and Xiao Zhou. The minimum vulnerability problem on graphs. In
Proc. 8th International Conference on Combinatorial Optimization and Applications (COCOA ’14), volume 8881 of LNCS, pages 299–313. Springer, 2014. 4,
25
[2] Sepehr Assadi, Ehsan Emamjomeh-Zadeh, Ashkan Norouzi-Fard, Sadra Yazdanbod, and Hamid Zarrabi-Zadeh. The minimum vulnerability problem. Algorithmica, 70(4):718–731, 2014. 1
[3] Kenneth A. Berman. Vulnerability of scheduled networks and a generalization of
Menger’s theorem. Networks, 28(3):125–134, 1996. 4
[4] Rodney G. Downey and Michael R. Fellows. Fundamentals of Parameterized
Complexity. Texts in Computer Science. Springer, 2013. 7
25
[5] Till Fluschnik. The parameterized complexity of finding paths with shared
edges. Master thesis, Institut für Softwaretechnik und Theoretische Informatik,
TU Berlin, 2015. 3
[6] Till Fluschnik, Stefan Kratsch, Rolf Niedermeier, and Manuel Sorge. The parameterized complexity of the minimum shared edges problem. In Proceedings of the
35th IARCS Annual Conference on Foundation of Software Technology and Theoretical Computer Science (FSTTCS’15), volume 45 of LIPIcs, pages 448–462.
Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2015. 2, 3, 25
[7] Lester Randolph Ford and D. R. Fulkerson. Flows in Networks. Princeton University Press, 1962. 4
[8] M. R. Garey, David S. Johnson, and Robert Endre Tarjan. The planar Hamiltonian circuit problem is NP-complete. SIAM Journal on Computing, 5(4):704–714,
1976. 9
[9] Richard M. Karp. Reducibility among combinatorial problems. In Proceedings of
a symposium on the Complexity of Computer Computations, The IBM Research
Symposia Series, pages 85–103. Plenum Press, New York, 1972. 7
[10] David Kempe, Jon M. Kleinberg, and Amit Kumar. Connectivity and inference
problems for temporal networks. Journal of Computer and System Sciences,
64(4):820–842, 2002. 4
[11] Jon M. Kleinberg and Éva Tardos. Algorithm Design. Addison-Wesley, 2006. 6
[12] Ekkehard Köhler, Rolf H. Möhring, and Martin Skutella. Traffic networks and
flows over time. In Algorithmics of Large and Complex Networks - Design, Analysis, and Simulation, volume 5515 of LNCS, pages 166–196. Springer, 2009. 4
[13] George B. Mertzios, Othon Michail, Ioannis Chatzigiannakis, and Paul G. Spirakis. Temporal Network Optimization Subject to Connectivity Constraints. In
Automata, Languages, and Programming, pages 657–668. Springer, 2013. 4
[14] Othon Michail. An Introduction to Temporal Graphs: An Algorithmic Perspective. Internet Mathematics, 12(4):239–280, 2016. 4
[15] Masoud T. Omran, Jörg-Rüdiger Sack, and Hamid Zarrabi-Zadeh. Finding paths
with minimum shared edges. Journal of Combinatorial Optimization, 26(4):709–
722, 2013. 1, 2, 3, 7
[16] James B. Orlin. Max flows in O(nm) time, or better. In Proceedings of the 45th
ACM Symposium on Theory of Computing (STOC’13), pages 765–774. ACM,
2013. 7
[17] Ján Plesník. The NP-completeness of the Hamiltonian cycle problem in planar
digraphs with degree bound two. Information Processing Letters, 8(4):199–201,
1979. 19
26
[18] Hassan Rashidi and Edward Tsang. Vehicle Scheduling in Port Automation: Advanced Algorithms for Minimum Cost Flow Problems, Second Edition. CRC Press,
2015. 1
[19] Martin Skutella. An introduction to network flows over time. In Research Trends
in Combinatorial Optimization: Bonn 2008, pages 451–482. Springer, 2009. 3, 4
[20] Zhi-Qian Ye, Yi-Ming Li, Hui-Qiang Lu, and Xiao Zhou. Finding paths with
minimum shared edges in graphs with bounded treewidths. In Proc. Frontiers of
Computer Science (FCS ’13), pages 40–46, 2013. 2, 4, 25
27
| 8cs.DS
|
Recognition of feature curves on 3D shapes using an algebraic
approach to Hough transforms ∗
arXiv:1709.10177v1 [cs.CV] 28 Sep 2017
Maria-Laura Torrente, Silvia Biasotti, Bianca Falcidieno
October 2, 2017
Abstract
Feature curves are largely adopted to highlight shape features, such as sharp lines, or to
divide surfaces into meaningful segments, like convex or concave regions. Extracting these
curves is not sufficient to convey prominent and meaningful information about a shape. We
have first to separate the curves belonging to features from those caused by noise and then
to select the lines, which describe non-trivial portions of a surface. The automatic detection
of such features is crucial for the identification and/or annotation of relevant parts of a given
shape. To do this, the Hough transform (HT) is a feature extraction technique widely used
in image analysis, computer vision and digital image processing, while, for 3D shapes, the
extraction of salient feature curves is still an open problem.
Thanks to algebraic geometry concepts, the HT technique has been recently extended to
include a vast class of algebraic curves, thus proving to be a competitive tool for yielding an
explicit representation of the diverse feature lines equations. In the paper, for the first time we
apply this novel extension of the HT technique to the realm of 3D shapes in order to identify
and localize semantic features like patterns, decorations or anatomical details on 3D objects
(both complete and fragments), even in the case of features partially damaged or incomplete.
The method recognizes various features, possibly compound, and it selects the most suitable
feature profiles among families of algebraic curves.
1
Introduction
Due to the intuitiveness and meaningful information conveyed in human line drawings, feature
curves have been largely investigated in shape modelling and analysis to support several processes,
ranging from non-photorealistic rendering to simplification, segmentation and sketching of graphical information [16, 42, 19, 27].
These curves can be represented as curve segments identified by a set of vertices, splines interpolating the feature points [7], L1 medial skeletons [74] or approximated with known curves, like
spirals [29, 30].
Traditional methods proposed for identifying feature curves on 3D models can be divided into
view dependent and view independent methods. The first ones extract feature curves from a
projection of the 3D model onto a plane perpendicular to the view direction, while view independent
techniques extract feature points by computing curvatures or other differential properties of the
model surface.
View independent feature curves assume various names according to the criteria used for their
characterization (ridge/valleys, crest lines, sharp lines, demarcating curves, etc.). Furthermore,
they are possibly organized into curve networks [19, 27] and filtered to omit short and non-salient
curves [14].
∗ Mathematics Subject Classification 8U05, 65D18.
Keywords and phrases. Feature curve recognition, Hough transform, curve identification on surfaces, robust line
detection
1
However, extracting feature curves is not sufficient to convey prominent and meaningful information about a shape. We have first to separate the curves belonging to features from those caused
by noise and then, among the remaining curves, to select the lines which describe non-trivial portions of a surface. These salient curves usually correspond to high level features that characterize
a portion of a shape. They can be represented by a multitude of similar occurrences of similar
simple curves (like the set of suction cups of an octopus tentacle), or by the composition of different
simple curves (like the eye and pupil contours). In addition, for some applications, it is necessary
to develop methods applicable also in the case of curves partially damaged or incomplete, as in
the case of archaeological artefacts.
To the best of our knowledge, the literature has not yet fully addressed the problem of identifying on surfaces similar occurrences of high level feature curves, of different size and orientation, and
in the case they are degraded. This is not true for 2D feature curves, where the Hough Transform
(HT) is commonly used for detecting lines as well as parametrized curves or 2D shapes in images.
Our attempt is then to study an HT-based approach suitable for extracting feature curves on 3D
shapes. We have found the novel generalization of the HT introduced in [9] convenient for the
detection of curves on 3D shapes. Using this technique we can identify suitable algebraic curves
exploiting computations in the parameters space, thus providing an explicit representation of the
equation of the feature curves. Moreover, we recognize which curves of the family are on the shape
and how many occurrences of the same curve are there (see Section 5 for various illustrative examples). Applying this framework to curves in the 3D space is not a trivial task: spatial algebraic
curves can be represented as the intersection of two surfaces and theoretical foundations for their
detection via HT has already been laid using Gröbner bases theory (see [9]). Nevertheless, the
atlas of known algebraic surfaces is not as wide as that of algebraic plane curves, therefore working
directly with curves could end with some limitations. Further, similarly to [42, 29, 30, 34] our
point of view is local since we are interested to the problem of the extraction of features contours
that can be locally flattened onto a plane without any overlap.
Contribution In this paper we describe a new method to identify and localize feature curves,
which characterize semantic features like patterns, decorations, reliefs or anatomical features on the
digital models of 3D objects, even if the features are partially damaged or incomplete. The focus
is on the extraction of feature curves from a set of potentially significant points using the cited
generalization of the HT [9]. This technique takes advantage of a rich family of primitive curves that
are flexible to meet the user needs. The method recognizes various features, possibly compound,
and selects the most suitable profile among families of algebraic curves. Deriving from the HT,
our method inherits the robustness to noise and the capability of dealing with data incompleteness
as for the degraded and broken 3D artefacts on which we realized our first experiments [69]. Our
main contributions can be summarized as follows:
• To the best of our knowledge this is the first attempt to systematically apply the HT to the
recognition of curves on 3D shapes.
• The method is independent of how the feature points are detected, e.g. variations of curvature, colour, or both; in general, we admit a multi-modal characterization of the feature lines
to be identified, see Section 3.1.
• The method is independent of the model representation, we tested it on point clouds and
triangle meshes but the same framework applies to other representations like quad meshes.
Details on the algorithms are given in Section 3 and in the Appendix.
• A vast catalogue of functions is adopted, which is richer than previous ones, and it is shown
how to modify the parameters to include families of curves instead of a single curve, details
are in Section 4.
• The set of curves is open and it can be enriched with new ones provided that they have an
algebraic representation.
2
• We introduce the use of curves represented also in polar coordinates, like the Archimedean
spiral.
• Our framework includes also compound curves, see Section 4.1.
• As a proof of concept, we apply this method to real 3D scans, see Section 5.
If compared to the previous methods, we think that our approach, conceived under the framework
of the Hough Transform technique, can be used and tuned for a larger collection of curves. Indeed,
we will show how spirals, geometric petals and other algebraic curves can be gathered using our
recognition technique.
2
Related work
The literature on the extraction of feature curves and the Hough transform is vast and we cannot
do justice to it here. In this section we limit our references only to the methods relevant to our
approach, focusing on HT-based curve detection and feature curve characterization.
Hough transform We devote this section to a brief introduction to the HT technique, while we
refer to recent surveys (for instance [49, 38]) for a detailed overview.
HT is a standard pattern recognition technique originally used to detect straight lines in images,
[32, 21]. Since its original conception, HT has been extensively used and many generalizations have
been proposed for identifying instances of arbitrary shapes over images [8], more commonly circles
or ellipses. The first very popular extension concerning the detection of any parametric analytic
curve is usually referred to as the Standard Hough Transform (SHT). In spite of the robustness
of SHT to discontinuity or missing data on the curve, its use has been limited by a number
of drawbacks, like the need of a parametric expression, or the dependence of the computation
time and the memory requirements on the number of curve parameters, or even the need of finer
parameters quantization for a higher accuracy of results.
To overcome some of these limitations, other variants have been proposed, one of the most
popular being the Generalized Hough Transform (GHT) by Ballard [8]. Since its conception, it
proved to be very useful for detecting and locating translated two-dimensional objects, without
requiring a parametric analytic expression. Thus, GHT is more general than SHT, as it is able to
detect a larger class of rigid objects, still retaining the robustness of SHT. Nevertheless, GHT often
requires brute force to enumerate all the possible orientations and scales of the input shape, thus
the number of parameters needs to be increased in its process. Further, GHT cannot adequately
handle shapes that are more flexible, as in the case of different instances of the same shape, which
are similar but not identical, e.g. petals and leaves.
Recently, thanks to algebraic geometry concepts, theoretical foundations have been laid to
extend the HT technique to the detection of algebraic objects of codimension greater than one
(for instance algebraic space curves) taking advantage of various families of algebraic plane curves
(see [9] and [10]). Being so general, such a method allows to deal with different shapes, possibly
compound, and to get the most suitable approximating profile among a large vocabulary of curves.
In 3D, other variants of HT have been introduced and used but as far as we know none of them
exploits the huge variety of algebraic plane curves (for this we refer again to the surveys [49] and
[38]). For instance, in [50] the HT has been employed to identify recurring straight line elements
on the walls of buildings. In that application, the HT is applied only to planar point sets and line
elements are clustered according to their angle with respect to a main wall direction; in this sense,
the Hough aggregator is used to select the feature line directions (horizontal, vertical, slanting)
one at a time.
Feature characterization Overviews on methods for extracting feature points are provided in
[16, 43]. In the realm of feature characterization it is possible to distinguish between features
3
that are view-dependent, like silhouettes, suggestive contours and principal highlights [20], mainly
useful for rendering purposes [44], or those that are independent of the spatial embedding and,
therefore, more suitable for feature recognition and classification processes. In the following we
sketch some of the methods that are relevant for our approach, i.e., characterizations that do not
depend on the spatial embedding of the surface.
A popular choice to locate features is to estimate the curvature, either on meshes [42, 72] or point
clouds [28, 18]). When dealing with curvature estimation, parameters have to be tuned according
to the target feature scale and the underlying noise. The Moving Least Square method [54] and its
variations are robust to scale variations [18, 36] and does not incorporate smoothness effects in the
estimation. Alternative approaches to locate curvature extrema use discrete differential operators
[31] or probabilistic methods, such as random walks [48]. For details on the comparison of methods
for curvature estimation, we refer to a recent benchmark [70].
Partial similarity and, in particular, self-similarity, is the keyword used to detect repeated
features over a surface. For instance, the method [26] is able to recognize repeated surface features
(circles or stars) over a surface. However, being based on geometry hashing, the method is scale
dependent and does not provide the exact parameters that characterize such features.
Despite the consolidate literature for images, feature extraction based on colour information
is less explored for 3D shapes, and generally used as a support to the geometric one [11, 13].
Examples of descriptions for textured objects adopt a 3D feature-vector description, where the
colour is treated as a general property without considering its distribution over the shape, see
for instance [64, 58, 62]. Another strategy is to consider local image patches that describe the
behaviour of the texture around a group of pixels. Examples of these descriptions are the Local
Binary Patterns (LPB) [52], the Scale Invariant Feature Transform (SIFT) [47], the Histogram of
Oriented Gradients (HOG) [17] and the Spin Images [35]. The generalization of these descriptors
to 3D textured models has been explored in several works, such as the VIP description [71],
the meshHOG [73] and the Textured Spin-Images [53]. Further examples are the colour-CHLAC
features computed on 3D voxel data proposed in [37]; the sampling method [46] used to select points
in regions of either geometry-high variation or colour-high variation, and to define a signature
based on feature vectors computed at these points; the CSHOT descriptor [66], meant to solve
point-to-point correspondences coding geometry- and colour-based local invariant descriptors of
feature points. However, these descriptors are local, sensitive to noise and, similarly to [26], scale,
furthermore they do not provide the parameters that characterize the features detected.
Feature curves identification The extraction of salient features from surfaces or point clouds
has been addressed either in terms of curves [28], segments (i.e. regions) [59] and shape descriptions [12]. Thanks to their illustrative power, feature curves are a popular tool for visual shape
illustration [42] and perception studies support feature curves as a flexible choice for representing
the salient parts of a 3D model [16, 29].
Feature curves are often identified as ridges and valleys, thus representing the extrema of
principal curvatures [51, 72, 14] or sharp features [43]. Other types of lines used for feature curve
representation are parabolic ones. They partition the surface into hyperbolic and elliptic regions,
and zero-mean curvature curves, which classify sub-surfaces into concave and convex shapes [40].
Parabolic lines correspond to the zeros of the Gaussian and mean curvature, respectively. Finally,
demarcating curves are the zero-crossings of the curvature in its gradient direction [42, 41]. In
general, all these curves, defined as the zero set of a scalar function, do not consider curve with
knots, fact that an algebraic curve like the Cartesian Folium (see [61]) could arrange.
In general, given a set of (feature) points, the curve fitting problem is largely addressed in the
literature, [23, 60, 56, 18]. Among the others, we mention [7] that recently grouped the salient
points into a curve skeleton that is fitted with a quadratic spline approximation. Being based on a
local curve interpolation, such a class of methods is not able to recognize entire curves, to complete
missing parts and it is difficult to assess if a feature is repeated at different scales.
Besides interpolating approaches such as splines, it is possible to fit the feature curve set with
some specific family of curves, for instance the natural 3D spiral [29] and the 3D Euler spiral
4
[30] have been proposed as a natural way to describe line drawings and silhouettes showing their
suitability for shape completion and repair. However, using one family of curves at a time implies
the need of defining specific solutions and algorithms for settings the curve parameters during the
reconstruction phase.
Recently, feature curve identification has been addressed with co-occurrence analysis approaches
[63, 45]. In this case, the curve identification is done in two steps: first, a local feature characterization is performed, for instance computing the Histogram of Oriented Curvature (HOC) [39]
or a depth image of the 3D model [63]; second, a learning phase is applied to the feature characterization. The learning phase is interactive and requires 2-3 training examples for every type
of curve to be identified and sketched; in case of multiple curves it is necessary also to specify
salient nodes for each curve. Feature lines are poly-lines (i.e. connected sequences of segments)
and do not have any global equation. These methods are adopted mainly for recognizing parts of
buildings (such as windows, doors, etc.) and features in architectural models that are similar to
strokes. Main limitations of these methods are the partial tolerance to scale variance, the need of
a number of training curves for each class of curves, the non robustness to missing data and the
fact that compound features can be addressed only one curve at a time [63].
In conclusion, we observe that the existing feature curve identification methods on surfaces
do not satisfy all the good properties typical of the Hough transforms, such as the robustness to
noise, the ability to deal with partial information and curve completion, the accurate evaluation of
the curve parameters and the possibility of identifying repeated or compound curves. Moreover,
the vocabulary of possible curves is generally limited to straight lines, circles, spirals, while the
HT-based framework we are considering encompasses all algebraic curves.
3
Overview of the method
Our approach to the extraction of peculiar curves from feature points of a given 3D model is
general, it can be applied to identify anatomical features, extract patterns, localize decorations,
etc. Our point of view is local since we are interested to the extraction of features contours that
locally can be projected on a plane without any overlap. From the mathematical point of view,
every surface can be locally projected onto a plane using an injective map and, if locally regular,
it can be expressed in local coordinates as (x, y, z(x, y)) [57].
We assume that the geometric model of an object is available as a triangulated mesh or a point
cloud, possibly equipped with colour. Nevertheless, our methodology can be applied also to other
model representations like quad meshes. Moreover, the photometric information, which, if present,
contains rich information about the real appearance of objects (see [65]), can be exploited alone
or in combination with the shape properties for extracting the feature points sets.
Since the straight application of the HT technique to curves in the 3D space is not trivial
(indeed, a curve is represented by the intersection of two surfaces), here we describe the steps
necessary to identify the set of points that are candidate to belong to a feature curve, to simplify
their representation using a local projection and to approximate the feature curve. Basically, we
identify three main steps:
Step 1: Potential feature points recognition: using different shape properties it extracts the sets of
feature points from the input model; then, points are aggregated into smaller dense subsets;
it works in the 3D space, see details in Section 3.1.
Step 2: Projection of the feature sets onto best fitting planes: it computes a projection of each set
of points obtained in the step 1 onto a best fitting plane; see Section 3.2.
Step 3: Feature curve approximation: based on a generalization of the HT it computes an approximation of the feature curve; it is applied to each set resulting from step 2; see Section 3.3.
While the first step is done only once, the second and the third ones run over each set of potential
feature points.
5
For the sake of clarity, a synthetic flowchart of our method is shown in Figure 1. In the boxes
the pseudo-code algorithms corresponding to the different actions are referred.
We provide the outline of our feature recognition method in the Main Algorithm 1 and we refer
to Sections 3.1-3.3 for a detailed description of the procedures. For the convenience of the reader
we sum up the variables of input/output and the notation we use in the Main Algorithm 1. The
algorithm requires four inputs:
• a given 3D model denoted by M;
• a property of the shape, such as curvature or photometric information, denoted by prop and
used to extract the feature points set;
• a threshold denoted by p and used for filtering the feature points;
• an HT-regular family of planar curves denoted by F.
Figure 1: A synthetic flowchart of our method.
In the body of the Main Algorithm 1 the following variables are used: X, which denotes the set of
feature points extracted from the model M by means of the Feature Points Recognition Algorithm
2; Yj , with j = 1 . . . m, which denotes the subset of points of X sharing some similar properties
obtained applying the Aggregation Algorithm 6; Zj , with j = 1 . . . m, which is the local projection
of Yj to its best fitting plane obtained applying Projection Algorithm 7; Cj , which is the curve of
the family F that best approximates Zj . Cj is obtained applying the Curve Detection Algorithm
4 to Zj with respect to a family of functions F in a region T of the parameter space which is
discretized with a step d.
6
3.1
Potential feature points recognition
Using various shape properties (curvatures, parabolic points identification or photometric information) the feature points set is extracted from the input model. In this paper we concentrate on
features highlighted by means of curvature functions and/or colorimetric attributes.
The geometric properties are derived using classical curvatures, like minimum, maximum, mean,
Gaussian or total curvatures. We denote these well-known geometric quantities by Cmin , Cmax ,
Cmean , CGauss , Ctot respectively. Several methods for estimating curvatures over triangle meshes
and point clouds exist but, unfortunately, there is not a universal solution that works best for
every kind of input [70]. We decided to keep our implementation flexible, adopting in alternative a
discretization of the normal cycles [15] proposed in the Toolbox graph [55] and the statistic-based
method presented in [36].
Main Algorithm 1: Feature curve identification for a given 3D model M with respect to a
property (different types of curvatures/colour) and a family F of curves
Input : a 3D model M, the property prop (the type of curvature and/or colour), the
threshold p for feature points filtering, an HT-regular family F of planar curves
Output: a list of curves belonging to the family F
1 begin
/* extracts the feature points set from the model M
*/
2
X ← Feature Points Recognition Algorithm 2 applied to M, prop and p;
/* groups the points of X into smaller dense subsets
*/
3
Y = [Y1 , . . . , Ym ] ← Aggregation Algorithm 6 applied to X;
4
for j = 1, . . . , m do
/* projects the points of Yj onto the best fitting plane
*/
5
Zj ←Projection Algorithm 7 (see Appendix) applied to Yj ;
6
t ← number of parameters of the curves of F;
/* initializes the region T of the parameter space and its
discretization step d
*/
7
T ← [a1 , b1 ] × . . . × [at , bt ] ∈ Rt ; d ← (d1 , . . . , dt ) ∈ Rt>0 ;
/* computes the curve of F that best approximates Zj
*/
8
Cj ← Curve Detection Algorithm 4 applied to Zj , F, T , d;
9
end
10
return C = [C1 , . . . , Cm ]
11 end
The photometric properties can be represented in different colour spaces, such as RGB, HSV,
and CIELab spaces. Our choice is to work in the CIELab space [6], which has been proved to
approximate human vision in a good way. In such a space, tones and colours are distinct: the L
channel is used for the luminosity, which closely matches the human perception of light (L = 0
yields black and L = 100 yields diffuse white), whereas the a and b channels specify colours [33].
The different types of curvatures Cmin , Cmax , Cmean , CGauss , Ctot , and the luminosity L are
used as scalar real functions defined over the surface vertices. One (or more) of these properties is
given as input (stored in the variable prop) in the Feature Points Recognition Algorithm 2. The
feature points X of the model M are extracted by selecting the vertices at which the property prop
is significant (e.g high maximal curvature and/or low minimal curvature and/or low luminosity).
This is automatically achieved by filtering the distribution of the function that we decide to use
by means of a filtering threshold p (see Appendix, Algorithm 5). Note that p is given in input in
the Main Algorithm 1. Its value varies according to the precision threshold set for the property
used to extract the feature points (e.g. in the case of maximum curvature a typical value of p is
80%). We sum up the described procedure in the Algorithm 2.
As an illustrative example we show the steps performed by the extraction of the feature points,
when applied to a 3D model M given as a triangulated mesh (made up of approximately 106
7
vertices and 2 · 106 faces) without photometric properties, see Figure 2(a).
Feature Points Recognition Algorithm 2: Extracts the feature points sets X ⊂ R3 from
the input model M
Input : 3D model M, the property prop, that is the type of curvature (Cmin , Cmax ,
Cmean , CGauss , Ctot ) and/or the L channel, the filtering threshold p
Output: Set of feature points X ⊂ R3
1 begin
/* loads M; faces are optional (available only for meshes)
*/
2
[vertices, faces]= load(M);
/* evaluates curvatures and colour (if available)
*/
3
if prop is a curvature then
4
[Cmin, Cmax] = EvaluateCurvatures(vertices, faces);
5
switch prop do
6
case Cmean do Cmean=(Cmin+Cmax)/2;
7
case CGauss do CGauss=Cmin.*Cmax;
8
case Ctot do Ctot = abs(Cmin)+ abs(Cmax);
9
end
10
else
11
[L, a, b] = get Lab(M);
12
end
13
f ← function according to the property prop;
/* builds the histogram of f and filters it using p
*/
14
h = histogram(f);
15
v = Filtering(h, p) (see Filtering Algorithm 5 in Appendix);
/* constructs the feature points set X
*/
16
for j = 1 . . .size(vertices) do
17
if f[j] > v then add vertices[j] to X;
18
end
return : X
19 end
A visual representation with colours of the different curvature functions computed on the model
M is shown in Figure 2(b)-(f); their histograms are shown in Figure 3. In Figure 4(b) the output
of Algorithm 2 when applied to M using the maximum curvature and filtering the corresponding
histogram at 90% is shown.
Once detected, the set X of feature points is subdivided into smaller clusters (that is, groups of
points sharing some similar properties, such as curvature values and/or chromatic attributes) by
using classical methods of cluster analysis. Here, we adopt a very well-known density model, the
Density-Based Spatial Clustering of Applications with Noise (DBSCAN) method [22], which groups
together points that lie closeby marking as outliers isolated points in low-density regions. The
DBSCAN algorithm requires two additional parameters: a real positive number ε, the threshold
used as the radius of the density region, and a positive integer M inP oints, the minimum number
of points required to form a dense region. In order to estimate the density of the feature set X
necessary to automatically relate the choice of the threshold ε to the context, we use the K-Nearest
Neighbor (KNN), a very efficient non parametric method that computes the k closest neighbors of
a point in a given dataset [24].
We sum up the described procedure in the Appendix (Algorithm 6). Figure 4(c) represents,
with different colours, the outcome of the aggregation algorithm when applied to the feature points
X in Figure 4(b).
8
(a)
(b)
(c)
(d)
(e)
(f)
Figure 2: A 3D model M (a) and the visualization (colours range from blue (low) to red (high))
of the values of different curvatures: (b) minimum curvature, (c) maximum curvature, (d) mean
curvature, (e) Gaussian curvature and (f) total curvature.
3.2
Projection of the feature points sets onto best fitting planes
This step performs a local projection of each feature points set to its best fitting plane. This
operation does not represent a strong restriction on our method since we apply such a projection
separately to each group (obtained from the aggregation of the feature points) and also because
we have already assumed that the kind of features that we are looking for can be locally projected
onto a plane. Let Y be a 3-dimensional set of points that can be injectively projected onto a plane.
During this phase the following operations are performed: firstly, the points of the set Y are shifted
to move their centroid onto the origin. Secondly, for the points of Y a best fitting plane Π is found
by computing the multiple linear regression using the least squares method. We compute it as
follows. We denote by s the cardinality of Y, and we consider the matrix XY of size s × 2 and the
column vector Z of size s × 1 whose columns respectively contain the x-and y-coordinates of the
points of Y and the the z-coordinates of the same points. We apply a linear regression function
to the pair (XY, Z) and get the real values b1 and b2 used to construct the best fitting plane Π,
whose equation is Π : z − b1 x − b2 y = 0. Finally, the orthogonal transformation ϕ moving the
plane Π onto the plane z = 0 is defined and applied to the points of Y to get the new set Z. We
sum up the described procedure in the Appendix, Algorithm 7. Figure 5(b) represents the best
fitting plane of the subset Y8 depicted in Figure 5(a); the new set Z8 is shown in Figure 5(c).
9
(1) - min curvature
(2) - max curvature
(4) - Gaussian curvature
(5) - total curvature
(3) - mean curvature
Figure 3: Histograms of different curvature functions on the 3D model M in Figure 2(a)
(a)
(b)
(c)
Figure 4: (a) The model M, (b) the feature points X resulting from Algorithm 2 applied to M
with parameters prop = 2 and p = 0.9, and (c) the groups obtained applying Algorithm 6 to X
with parameters K = 50 and M inP ts = 5.
3.3
Feature curve approximation
Once we have projected every set of points onto its best fitting plane, we apply to the single group
elements the generalization of the HT technique (see [10] and [67]) aimed at approximating the
feature curve. With respect to the previous application of [10] to images, our approach is novel
since we apply the method to each projected 3D feature points set by exploiting a vast catalogue
of curves (see Section 4 for a detailed description of the families of curves used in this paper).
Further, we propose a method that does not undergo to any grid approximation of the coordinates
of the given points.
We provide here some details on the curve detection algorithm which is the core of the HTbased technique. In its classical version [32, 21], HT permits to recover the equation of a straight
line exploiting a very easy mathematical principle: starting from points lying on a straight line,
defined using the common Cartesian coordinates x and y (or equivalently the polar coordinates ρ
and θ) and following the usual slope-intercept parametrization with parameters a and b we end
with a collection of straight lines in the parameters’ space (or equivalently sinusoidal curves in the
variables ρ and θ) that all intersect in exactly one point. The equation of each straight line is also
called Hough transform of the point, usually denoted by Γp (F) (where F denotes the chosen family
of curves, the straight lines in this particular case). The coordinates of the unique intersection
point of all the Hough transforms identify the original line.
10
(a)
(b)
(c)
Figure 5: Representation of (a) the cluster Y8 (see also Figure 4(c)), (b) its best fitting plane Π8
and the subset Z8 obtained by projecting Y8 onto Π8 .
Following the approach of [9], we consider a collection of curves which includes more complex
shapes beyond the commonly used lines, e.g. circles and ellipses (see Section 4). Under the
assumption of Hough regularity on the chosen family of curves F (see [10]), the detection procedure
can be detailed as follows. Let p1 , . . . , ps be the set of feature points; in the parameter space we find
the (unique) intersection of the Hough transforms (the hypersurfaces depending on the parameters)
Γp1 (F), . . . , Γps (F) corresponding to the points p1 , . . . , ps ; that is, we compute λ = ∩i=1...s Γpi (F).
Finally, we return the curve of the family F uniquely determined by the parameter λ.
From a computational viewpoint, the burden of the outlined methodology is represented by the
computation of the intersection of the Hough transforms, which is usually implemented using a
so called “voting procedure”. Following [67] the voting procedure can be detailed in the following
four steps:
1. Fix a bounded region T of the parameter space.
This is achieved exploiting typical characteristics of the chosen family F of curves, like the
bounding box properties or the presence of salient points. In Section 4 we detail how to
derive the bounding box of the parameter space from the algebraic or polar representation
of various families of curves.
2. Fix a discretization of T .
This is nontrivial yet fundamental for the detection result which is valid up to the chosen
discretization step. The choice of the size of the discretization step is currently done as a
percentage of the range interval of each parameter. For some recent results on this topic we
refer to [68]. Let T = [a1 , b1 ] × . . . × [at , bt ] be the fixed region of the parameters space and
d = (d1 , . . . , dt ) ∈ Rt>0 be the discretization step. For each k = 1, . . . , t, we define
&
'
bk − ak − d2k
+ 1,
(1)
Jk :=
dk
where dxe = min{z ∈ N | z ≥ x} and
λk,jk := ak + jk dk
(2)
with jk = 0, . . . , Jk − 1. Here Jk denotes the number of considered samples for each component, and jk the index of the sample. We denote by j the multi-index (j1 , . . . , jt ), by
λj := (λ1,j1 , . . . , λt,jt ) the j-th sampling point, and by
dk
dk
t
C(j) := (λ1 , . . . , λt ) ∈ R λk ∈ λk,jk − , λk,jk +
, k = 1, . . . , t
(3)
2
2
the cell centered at (and represented by) the point λj . The discretization of T is given by
the J1 × · · · × Jn cells of type C(j) which are a covering of the region T .
11
P
3. Construct an accumulator function A : T → N, which is defined as A = 1=1...s fpi where
fpi : T → N is defined as follows:
1 if the Hough Transform Γpi (F) crosses the cell c;
fpi (c) =
0 otherwise.
For the evaluation of each fpi we adopt the Crossing Cell algorithm (detailed in Algorithm 3)
which is based on bounds of the evaluation of fpi theoretically proved in [67]. The function
EvalF irstBound, resp. EvalSecondBound, in the pseudocode represents the evaluation of
the bound B1 (resp. B2 ) of a given polynomial f in n variables w.r.t a cell centered at p
with radius ε. Such bounds depend on the Jacobian and the Hessian matrices of f , denoted
by Jacf and Hf respectively. They are defined as follows:
B1
= kJacf (p)t k1 ε +
n 2
Hε
2
(4)
2R
(5)
J(c + n5/2 HJR)
o
n
√
kJacf (p)k1
, c = max{2, n} and
where H = max{x∈Rn :kx−pk∞ ≤ε} kHf (x)k∞ , R < min ε,
H
B2
=
J = sup{x∈Rn :kx−pk∞ <R} kJac†f (x)k∞ , with Jac†f denoting the Moore-Penrose pseudo-inverse
of Jacf . Since the above quantities depend on the Jacobian and the Hessian matrices, p must
be a point for which these values are non-trivial. Note that the evaluation is made over a
symbolic representation of the Jacobian matrix, its pseudo-inverse and the Hessian matrix.
This implies that the computational cost of this operation is constant while their symbolic
representation is computed only once, in the overall Curve Detection Algorithm 4. In our
implementation we use the system CoCoA [5] for the symbolic manipulation of polynomials
and matrices.
4. Identify the cell corresponding to the maximum value of the accumulator function and return
the coordinates of its center.
Following the general theory, the HT regularity guarantees that the maximum of the accumulator function is unique. In addition, being based on local maxima, the voting strategy
permits to identify curves also from partial and incomplete data, even if the missing part is
significant, see the examples in Figure 19(I.c-d) and in Figure 19(II.c-d).
The Curve Detection Algorithm 4 sums up the curve detection’s procedure described in the
previous steps 1-4. Note that a prototype implementation in CoCoA of the Curve Detection
Algorithm 4 is freely available1 .
To give an idea of the outcome of the Algorithm 4, we have run it on the set Z8 (represented
in Figure 5(c)) made of 981 points. In this example we use the geometric petal curve, an HTregular family of curves presented both in polar and in cartesian form, see Section 4 for details. By
exploiting the bounding box of the set Z8 and properties of the geometric petal curve, we consider
T = [121, 122] × [0.43, 0.45] and d = (0.025, 0.005). The Algorithm 4 applied to this framework
computes the geometric petal curve depicted in Figure 6(a); the feature curve is shown on the 3D
model as the red line in Figure 6(b).
We outline the steps of the Main Algorithm 1 when applied to the 3D model (made up of
152850 vertices and 305695 faces) represented in Figure 7(a). In this case, in order to detect the
spiral-like curves we work with the generalized family F of Archimedean spirals (see Section 4 for
more details). The Feature Points Recogniton Algorithm 2 is applied using the mean curvature
and returns a set X made up of 12689 points. Then, the Aggregation Algorithm 6 (in Appendix)
subdivides the points of X into 8 groups Y1 , . . . , Y8 . For shortness, we give details for the first group,
Y1 , which is made up of 2382 points. Exploiting some geometrical properties of the Archimedean
spirals (see again Section 4), we fix and discretize a suitable region of the parameter space, and
1 http://www.dima.unige.it/
torrente/recognitionAlgorithm.cocoa5
12
Crossing Cell Algorithm 3: Returns 1 or 0 if f = 0 crosses or not the cell of radius ε
centered at p. If not decidable it returns undetermined.
Input : The symbolic representation of f in the variables x1 , . . . , xn , a point p ∈ Rn , a
tolerance ε, the symbolic expression of the Jacobian Jacf of f , the Moore-Penrose
pseudo-inverse Jac†f of Jacf , and the Hessian matrix Hf of f .
Output: an element of {0, 1, undetermined }
1 begin
2
B1 ← EvalF irstBound(f, p, ε, Jacf ) (see formula (4));
3
B2 ← EvalSecondBound(f, p, ε, Jac†f , Hf ) (see formula (5));
4
if abs(f (p)) > B1 then
5
returnValue=0
6
else
7
if abs(f (p)) < B2 then returnValue=1
8
else returnValue= undetermined end
9
end
return : returnValue
10 end
Curve Detection Algorithm 4: Computes the curve C of the family F best detecting the
profile highlighted by the points of the set X
Input : A finite set X = {p1 , . . . , ps } ⊂ R2 , an HT-regular family F = F (x, y, λ1 1, . . . , λt )
of curves, a region T = [a1 , b1 ] × . . . × [at , bt ] ⊂ Rt of the parameter space, a
discretizaton step d = (d1 , . . . , dt ) ∈ Rt>0
Output: a curve C of the family F
1 begin
/* discretizes the region T with step d
*/
2
Initialize Jk , with k = 1, . . . t, (see formula (1));
3
Initialize λk,jk , with k = 1, . . . t and jk = 0, . . . , Jk − 1 (see formula (2));
4
Initialize λj and C(j), with j ∈ J1 × . . . × Jt (see formula (3));
/* constructs of the multi-matrix A
*/
5
A ← zero matrix of size J1 × . . . × Jt ;
6
Jac ← Jacobian matrix of F (x, y, λ1 , . . . , λt ) w.r.t λ1 , . . . , λt ;
7
Jac†f ← Moore-Penrose pseudo-inverse of Jac;
8
Hf ← Hessian matrix of F (x, y, λ1 , . . . , λt ) w.r.t λ1 , . . . , λt ;
9
for i = 1, . . . , s do
10
for each j ∈ J1 × . . . × Jt do
11
A(j) ← A(j) + Crossing Cell(F (pi ), λj , d2 , Jacf (pi ), Jac†f (pi ), Hf (pi )) (Alg. 3)
12
end
13
end
/* computation of the maximum of A
*/
14
j̄ ← max(A);
15
return C = C(λj̄ ), the curve of F with parameters λj̄
16 end
13
(a)
(b)
Figure 6: Representation of the set Z8 and the geometric petal curve (in red) computed by Algorithm 4 (a) and visualization of the curve (in red) on the original model of Figure 2(b).
(a)
(b)
(c)
(d)
Figure 7: A 3D model (a), recognition of a spiral (b), two views of the spiral-like curves detected
by our method (c,d).
apply the Curve Detection Algorithm 4. The corresponding spiral curve is depicted in Figure 7(b)
and shown on the 3D model as the red line in Figure 7(c-d). Other detected spirals are marked
using different colours in Figure 7(c-d).
3.4
Computational complexity
Here we briefly analyse the computational complexity of the method. The feature points recognition
function (Algorithm 2) depends on the chosen procedure, in case we adopt the curvature estimation
proposed in [15] the computational complexity is O(n log n), where n represents the number of
points of the model. The computational cost for converting RGB coordinates into CIELab ones is
linear (O(n)). Denoting nf the number of the feature points (in general nf n), the aggregation
into groups using DBSCAN (see the Algorithm 6 in the Appendix) takes O(nf 2 ) operations in the
worst case [22] (on average it takes O(nf log nf ) operations, and thus it is overcome by the KNN
search operation detailed in the Algorithm 6 that costs O(knf log nf ), see [25].
After the aggregation of the feature points, the projection (Algorithm 7 in the Appendix) and
the curve detection (Algorithm 4) algorithms are applied to each group, separately (note that the
sum sf of the elements in the groups is, in general, smaller than nf because of the presence of
outliers and small groups).
The cost of the curve detection algorithm is dominated by the size of the discretization
of the
Qt
region T , as detailed in the Algorithm 4. Such a discretization consists of M = k=1 Jk elements,
where t is the number of parameters (in the curves proposed in this paper, t = 2, 3) and Jk is
the number of subdivisions for the kth parameter, see formula (1). The evaluation of the HT
on each cell is constant and the Jacobian, pseudo-inverse and Hessian matrices are symbolically
computed once for each curve; therefore, the cost of fitting a curve to each group with our HT
method is O(M ). Since the curve fitting is repeated for m groups, the overall cost of our method
is O(max(n log n, msf 2 , mM )) where n, m, sf and M represent, respectively, the number of points
of the 3D model, the number of groups of feature points and their maximum size, the size of the
14
discrete space on which we evaluate the accumulator function.
4
Families of curves
In this section we list the families of curves used in our experiments, focusing in each case on
their main properties and characteristics (parameter dependency, boundedness, computation of the
bounding box). For every family of curves we explicitly describe how to derive from its algebraic
representation the parameters used as input for the Curve Detection Algorithm 4. The selected
curves form an atlas, which is both flexible and open: in fact, these curves are modifiable (for
instance, by adding, stretching or scaling parameters) and the insertion of new families of curves is
always possible. Note that the atlas comprises families of curves defined using Cartesian or polar
coordinates, since our framework works on both cases.
Our collection also includes some elementary, but likewise interesting, families of algebraic
curves like straight lines, circles and ellipses, which can be exploited to detect for instance eyes
contours, pupils shapes and lips lines. Their equation are linear (straight lines) and quadratic
(ellipses and circles) and depend on 3 parameters at most.
In the following, we list the other families of the collection; these curves mainly come from [61]
that contains a rich vocabulary of curves many of which are suitable for our approach, too. The
curves employed in our experiments (see Section 5) are: the curve of Lamet, the citrus curve, the
Archimedean spiral, the curve with m-convexities and the geometric petal curve.
The curve of Lamet is a curve of degree m, with m an even positive integer, whose outline is
a rectangle with rounded corners. Its Cartesian equation is:
ym
xm
+
=1
am
b
or equivalently in polynomial form: bxm + am y m = am b, with a, b ∈ R>0 and it is a bounded
connected closed curve with two axes of symmetry (the x and the y axes). The curve of Lamet is
contained in the rectangular region [−a, a] × [−b1/m , b1/m ]. Some examples are provided in Figure
8 using different values of the parameters a, b and m.
-2,4
-2
-1,6
-1,2
-0,8
-0,4
1,2
1,2
1,2
0,8
0,8
0,8
0,4
0,4
0
0,4
0,8
1,2
1,6
2
2,4
-2,4
-2
-1,6
-1,2
-0,8
-0,4
0
0,4
0,4
0,8
1,2
1,6
2
2,4
-2,4
-2
-1,6
-1,2
-0,8
-0,4
0
-0,4
-0,4
-0,4
-0,8
-0,8
-0,8
-1,2
-1,2
-1,2
(a)
(b)
0,4
0,8
1,2
1,6
2
2,4
(c)
Figure 8: Curve of Lamet with a = 2, b = 1 and: (a) m = 4, (b) m = 8, and (c) m = 16.
Another interesting shape (for instance when looking for a mouth, an eye feature or a leaf like
decoration) is given by the sextic surface of equation
a4 (x2 + z 2 ) + (y − a)3 y 2 = 0
with a ∈ R, called the zitrus (or citrus) surface by Herwig Hauser [1]. The citrus surface has
1
πa3 .
bounding box [− a8 , a8 ] × [0, a] × [− a8 , a8 ], centroid at (0, a2 , 0) and volume 140
We derive the citrus curve of equation fa (x, y) = 0 as the intersection of a rotation of π/2 of
the citrus surface with the plane z = 0, where fa (x, y) is the following sextic polynomial
a 3
a 3
fa (x, y) = a4 y 2 + x −
x+
2
2
15
with a ∈ R (see Figure 9(a)). The citrus curve is a symmetric bounded curve with bounding box
[− a2 , a2 ] × [− a8 , a8 ].
To include shapes with a different ratio, we introduce another citrus curve whose equation is
fa,c (x, y) = 0 (which is simply stretched or shortened along the y-axis) where fa,c (x, y) is given by:
a 3
a 3
f (x, y) = a4 c2 y 2 + x −
x+
2
2
with a, c ∈ R (see Figure 9(b)). Note that this is again a symmetric bounded curve with bounding
a a
box [− a2 , a2 ] × [− 8c
, 8c ].
-1,25
-1
-0,75
-0,5
-0,25
0,75
0,75
0,5
0,5
0,25
0,25
0
0,25
0,5
0,75
1
1,25
-1,25
-1
-0,75
-0,5
-0,25
0
-0,25
-0,25
-0,5
-0,5
-0,75
-0,75
(a)
0,25
0,5
0,75
1
1,25
(b)
Figure 9: The citrus curve of equation: (a) fa (x, y) = 0 with a = 2, c = 1/4, (b) fa,c (x, y) = 0
with a = 2, c = 1/2.
The Archimedean spiral (or arithmetic spiral ) can be found in human artefact decorations as
well as in nature. Its polar equation is:
ρ = a + bθ
with a, b ∈ R. The Archimedean spiral is an unbounded connected curve with a single singular
point (the cessation point (a, 0)). Two consecutive turnings of the spiral have a constant separation
distance equal to 2πb, hence the name arithmetic spiral. Another peculiar aspect is that, though
its unboundedness nature, the kth turning of the spiral is contained in a region bounded by two
concentric circles of radii a + 2(k − 1)πb and a + 2kπb. In particular, the first turning is contained
in the circular annulus of radii a and a + 2πb. An example is provided in Figure 10(a).
We extend the Archimedean spiral by weakening its constant pitch property, introducing an
extra parameter c; its polar equation becomes:
ρ = a + bθ + cθ2
with a, b, c ∈ R. Analogously to the Archimedean spiral, this generalized family is still an unbounded connected curve with a single singular point (the cessation point (a, 0)). An example is
provided in Figure 10(b).
In order to show the behaviour of the Main Algorithm 1, the generalized version of the
Archimedean spirals has been employed in Section 3. Some computational details have been
provided in the case of the set Y1 , represented in Figure 7(b). Exploiting the Cartesian coordinates of the cessation point (0.3915, 0.0048) of the set Y1 , and the fact that the first turning of the
spiral is contained in the circumference of radius 1.1542 centered at the origin, we considered the
region of the parameter space T = [0.35, 0.45] × [0.1, 0.15] × [0.0032, 0.0048] and the discretization
step (0.01, 0.01, 0.004). Figure 7 presents a set of feature curves recognized with the extended
Archimedean spiral. With reference to the example shown in Figure 7(b), the maximum of the
accumulator function is 44 (the second maximum has value 29) and corresponds to the cell with
7
1
2
center ( 20
, 10
, 625
).
16
2,4
2,4
1,6
1,6
0,8
-4,8
-4
-3,2
-2,4
-1,6
-0,8
0,8
0
0,8
1,6
2,4
3,2
4
4,8
-4,8
-4
-3,2
-2,4
-1,6
-0,8
0
-0,8
-0,8
-1,6
-1,6
-2,4
-2,4
(a)
0,8
1,6
2,4
3,2
4
4,8
(b)
Figure 10: Archimedean spirals with parameters: (a) a = 1/10, b = 1/2π and (b) a = 1/10,
b = 1/2π, c = 1/100.
Figure 11 shows another family of curves. The curve with m-convexities defined by the polar
equation
a
ρ=
1 + b cos(mθ)
with a, b ∈ R>0 , b < 1, and m ∈ N+ , m ≥ 2. The curve with m-convexities is a bounded connected
π
π
closed curve with m axes of symmetry (the straight lines of equation x sin m
k − y cos m
k = 0,
with k = 0, . . . , m − 1). This curve is contained in a region bounded by two concentric circles of
a
a
radii 1+b
and 1−b
. The shape of the curve with m-convexities strongly depends on the values of
its parameters. In particular, the parameter a plays the role of a scale factor, while the values
of b tune the convexities’ sharpness. Some examples of curves with m-convexities are provided in
Figure 11 where in the first row parameter m = 3, and in the second row parameter m = 5.
(a)
(b)
(c)
7,5
2,4
2,5
1,6
5
2,5
0,8
-4
-3,2
-2,4
-1,6
-0,8
0
0,8
1,6
2,4
3,2
4
-5
-2,5
0
-0,8
-1,6
5
-12,5
-10
-7,5
-5
-2,5
0
2,5
5
7,5
10
12,5
-2,5
-2,5
-2,4
(d)
2,5
-5
-7,5
(e)
(f)
Figure 11: Curves with m-convexities with a = 2 and (a) b = 1/4, m = 3, (b) b = 1/2, m = 3, (c)
b = 3/4, m = 3, (d) b = 1/4, m = 5, (e) b = 1/2, m = 5, (f) b = 3/4, m = 5.
The so-called geometric petal curve resembles an eye contour line, for particular values of the
parameters. Its polar equation is:
ρ = a + b cos2n θ
17
with n ∈ N+ and a, b ∈ R. The geometrical petal is a bounded symmetric curve with a singularity
at the origin. Some examples are provided in Figure 12(a)-(b) where the values of the parameters
are set as follows: a = 2, b = −2 and n = 1, 10.
For our purposes, we can restrict to the case b √
= −a. In this case, we observe that the curve
pass to the Cartesian equation using
is completely contained inside thepcircle of radius 2a. Wep
2 and cos θ = x/ x2 + y 2 ; further, in order to lower
the standard substitutions ρ = x2 + y√
the parameters degree, we replace a by a. The Cartesian equation of the geometric petal is
ga (x, y) = 0 where
ga (x, y) = (x2 + y 2 )2n+1 − a[(x2 + y 2 )n − x2n ]2
From an analytich intersectionqof the curve with
axes, we compute the bounding box
qthe Cartesian
i
√ √
1
2n √ 2n
1
2n √ 2n
of the curve as: − 2n+1 a
2n+1 , 2n+1 a
2n+1 × [− a, a].
Most of times, like in the extraction/localization of eyes contours, we need a shape which is
more stretched along the x-axis
√ (see Figure 12(c)-(d)). We stretch the geometric petal by scaling
the x-variable by a factor of c, where c ∈ R>0 . The new Cartesian equation of the curve is
ga,c (x, y) = 0 where
ga,c (x, y) = (cx2 + y 2 )2n+1 − a[(cx2 + y 2 )n − cn x2n ]2 .
(6)
q
q
i
h
p a 2n
p a 2n
√ √
1
1
2n
2n
,
with bounding box − 2n+1
c
2n+1 2n+1
c
2n+1 × [− a, a].
To give an idea of the outcome of Algorithm 4, the stretched version of the geometric petal curve
(equation 6) has been employed in Section 3. In that example, we fixed n = 50 (indeed the shape
of the geometric petal curve changes with this parameter, see again Figure 12) and, exploiting the
bounding box of the set Z8 (see Figure 5(c)), we considered a region T = [121, 122] × [0.43, 0.45]
of the parameter space.
In the following, we show how to reduce the number of parameters of this curve. We observed
that the value of the exponent parameter n is related to the bounding box of the curve; indeed it
has to satisfy the following condition:
2n
2n + 1
r
1−
n
1
2n + 1
!1/2
=
yB
yA
where yA and yB are the y-coordinates values of the points A and B (see Figure 12 (c)-(d)). By
using the previous relation, it is possible to estimate the value of n, thus working with a curve
which depends on the two parameters a and c.
4.1
Compound curves
To grasp compound shapes we use more families of curves simultaneously. We represent compound
curves as a combination of different families of curves and define their equation simply as the
product of (two or more) curves’ equations. In the following we show some examples of possible
compositions.
Combining a citrus curve and a circle we get the new family of curves of equation:
a 3
a2
a 3
2
2
4 2
x+
x +y −
=0
a y + x−
2
2
64
with a ∈ R> 0. An example is provided in Figure 13(a) where the parameter is a = 2.
Another possibility is to couple a citrus curve with a line (we choose one of the two axes of
symmetry, for instance the x axis) whose equation is:
a 3
a 3
4 2
y a y + x−
x+
=0
2
2
18
2,8
2,8
2,4
2,4
2
2
1,6
1,6
1,2
1,2
0,8
0,8
0,4
-2,4
-2
-1,6
-1,2
-0,8
-0,4
0,4
0
0,4
0,8
1,2
1,6
2
2,4
-2,4
-2
-1,6
-1,2
-0,8
-0,4
(a)
3
3
2,5
2,5
A
A
1,5
1,5
1
-1,5
-1
-0,5
0,8
0,5
1
1,2
1,6
2
2,4
2
1
B
0,5
-2
0,4
(b)
2
-2,5
0
0
0,5
1
1,5
2
B
0,5
2,5
3
-2,5
-2
(c)
-1,5
-1
-0,5
0
1,5
2
2,5
3
(d)
Figure 12: Geometric petal curves with parameters a = 2, b = −2 and (a) n = 1, (b) n = 10.
Geometric petal curves defined by ga,c (x, y) = 0 with parameters a = 4, n = 50 and (c) c = 1, (d)
c = 1/2.
with a ∈ R> 0. An example is provided in Figure 13(b) where the parameter is a = 2. An
application of the use of this compound curve is given in Figure 17(b).
It is also possible to associate a curve with 5-convexities with a circumference; the resulting
polar equation is:
a
ρ−
(ρ − r) = 0
1 + b cos(5θ)
a
with a, b, r ∈ R>0 , b < 1 and r ≤ 1+b
. An example is provided in Figure 13(c).
Finally, starting with three ellipses we construct a family of algebraic whose equation is:
2
2
x
y2
16x2
(2y − b)2
4x
y2
+
−
1
+
−
1
+
−
1
=0
a2
b2
9a2
b2
b2
b2
with a, b ∈ R>0 . An example is provided in Figure 13(d).
5
Examples and conclusive remarks
Our method has been tested on a collection of artefacts and models collected from the web, the
AIM@SHAPE repository [3], the STARC repository [2] and the 3D dataset of the EPSRC project
[4].
Most of the models are triangulated meshes; an exception is represented by the model in Figure
19(I.a), which is given as a point cloud. Further, in the examples in Figure 16(IV.a) and 19(I.a)
the colorimetric information is available and combined with the curvature; in the models in Figure
15(IV.a) and Figure 18(II.a) features are only characterized by the dark colour of the decoration;
the other models come without any colorimetric information and we adopt the maximum, minimum
and mean curvature properties. In all the examples, the model embedding in the 3-dimensional
19
-1
-0,75
-0,5
-0,25
0,5
0,5
0,25
0,25
0
0,25
0,5
0,75
1
-1
-0,75
-0,5
-0,25
0
-0,25
-0,25
-0,5
-0,5
(a)
0,25
0,5
0,75
1
(b)
7,5
2,4
5
1,6
2,5
0,8
-4,8
-4
-3,2
-2,4
-1,6
-0,8
0
0,8
1,6
2,4
3,2
4
-10
4,8
-7,5
-5
-2,5
0
2,5
5
7,5
10
-0,8
-2,5
-1,6
-5
-2,4
-7,5
(c)
(d)
Figure 13: Compound curves: (a) citrus curve with a = 2 coupled with a circumference; (b)
citrus curve with a = 2 coupled with the line y = 0; (c) curve with 5-convexities coupled with a
circumference with a = 2, b = 1/4, r = 1; (d) three ellipses with a = 10, b = 6.
space is completely random and the “best” view shown in the pictures is artificially reported for the
user’s convenience. Furthermore, the features we identify are view-independent and represented
by curves that can be locally projected onto a plane without any overlap.
For the feature identification, we assume to know in advance the class of features that are
present in an object (for instance because there exists an archeological description of the artwork),
thus converting the problem into a recognition of definite curves that identifies specific parts such
as mouth, eyes, pupils, decorations, buttons, etc. on that model. Once aggregated and projected
onto a plane, each single feature group of points is fitted with a potential feature curve and the
value of the HT aggregation function is kept as the voting for that. After all curves are run, we
keep the highest vote to select the curve that better fits with that feature. Sometimes more than
one curve potentially fits the feature point set; in this case we have selected the curve with the
highest value of the HT aggregation function.
Figure 14 presents an overview of multiple feature curves obtained by our method. These
examples are shown in a portion of the model but are valid for all the instances of the same
features. Multiple instances of the same family are shown in Figure 14(I) and Figure 14(II), in
which the curves are circles with different radii, in Figure 14(III) and Figure 14(IV), where the
detected curves belong to the family of curves with 5-convexities, in Figure 14(V) and Figure
14(VI), where different Archimedean spirals have been used.
It is also possible to recognize different families of curves on the same surface, as it happens in
the models in Figure 15(I), where circles are combined with two curves of Lamet, in Figure 15(II),
where a curve with 5-convexities is used in combination with a circle, in Figure 15(III), where a
curve with 8-convexities is contained in a region delimited by two concentric circles, and in Figure
15(IV), where the decoration is a repeated pattern of leaves, each one identified with a specific
group and approximated by a citrus or a geometric petal curve.
Figure 16 shows four examples of eye contours detection from 3D models. The eyes in the first
20
I.
(a)
(b)
(c)
(a)
(b)
(c)
(a)
(b)
(c)
(a)
(b)
(c)
(a)
(b)
(c)
(a)
(b)
(c)
II.
III.
IV.
V.
VI.
Figure 14: Examples of recognition of various feature curves; I. circles on the Gargoyle model
(the AIM@SHAPE repository [3]); circles on a tentacle of an octopus model; III. stars (curves
with 5-convexities) on the knot model (the AIM@SHAPE repository [3]); IV. stars (curves with
5-convexities) of different sizes on the trim-star model (the AIM@SHAPE repository [3]); V. and
VI. spirals on architectural ornamental artefacts (from the 3D dataset of the EPSRC project [4]).
21
I.
(a)
(b)
(c)
(a)
(b)
(c)
(a)
(b)
(c)
(a)
(b)
(c)
II.
III.
IV.
Figure 15: I. circles and curves of Lamet on a phone model; II. a circle and a curve with 5convexities on an architectural ornamental artefact (from the 3D dataset of the EPSRC project
[4]); III. a curve with 8-convexities contained in a region delimited by two concentric circles on an
architectural ornamental artefact (from the 3D dataset of the EPSRC project [4]); IV. decorations
on a bust (from the STARC repository [2]).
22
row are better approximated by the citrus curve while in the other three cases the geometric petal
is the best fitting curve.
I.
(a)
(b)
(c)
(d)
(a)
(b)
(c)
(d)
(a)
(b)
(c)
(d)
(a)
(b)
(c)
(d)
II.
III.
IV.
Figure 16: Detection of eye contours and pupils on collection of artefacts collected from the
AIM@SHAPE repository [3] and the STARC repository [2].
The automatic detection of multiple instances of the same family of curves, possibly with
parameter changes as in the case of circles of different radii (e.g. the suction cups of the octopus
model, see again Figure 14.II), is automatically done in the space of the parameters by the HT
procedure. As a side effect, this immediately yields the equation of the curve that represents these
feature points.
An important advantage of our approach is the flexible choice of the family of algebraic curves
used to approximate the desired features, thus being adaptive to approximate various shapes. Our
set of primitives includes generic algebraic curves and it can be extended to all curves with an
implicit representation, see Section 4.
As discussed in Section 4.1, another interesting point is the possibility of recognizing compound
features as a whole, as for the eye contour and the pupil, see examples in Figures 16.I and 17(a).
Figure 17(b) detects a mouth contour combining a citrus curve with a line. Such an option opens
23
the method to a wide range of curves that includes also repeated patterns like bundles of straight
lines or quite complex decorations, see examples in Figure 15.
Another major benefit of using an HT-based method is the HT well known robustness to
noise and outliers. Moreover, by selecting the curve with the highest score of the HT aggregation
function, our method is able to keep a good recognition power also in the case of degraded (Figure
18) and partial features (Figure 19). Moreover, we are able to work on both 3D meshes and point
clouds, thus paving the road to the application of the method to object completion and model
repairing.
In addition, it is important to point out that we can similarly parametrize features that are
comparable. As we extract a feature curve and its parameters, we can use them to build a template
useful for searching similar features in the artefacts, even if heavily incomplete, see Figure 18(II.c-d)
and Figure 19(I.c-d) and Figure 19(II.c-d).
(a)
(b)
Figure 17: Combined anatomical shapes: (a) a circumference and a geometric petal curve detecting
an eye; (b) a citrus curve with a line detecting a mouth.
I.
(a)
(b)
(c)
(d)
(a)
(b)
(c)
(d)
II.
Figure 18: Feature recognition on models with degraded features.
With reference to Figure 20, we use different colours to show the feature curves identified by
different curves/parameters. In Figure 20(a), the blue lines represent the circles, all with the same
radius, while the red and green lines depict the curves of Lamet. In particular, the red and the green
lines are used for Lamet curves which differ for the values of the parameters a and b. Similarly, in
Figure 20(b) which represents a tentacle of an octopus model, the red graduate shadings are used
to highlight the different radii of the detected circles.
Different colours are also used in Figure 21 where two concentric spirals (red and blue) with
different parameters, are detected. Because of the proximity of the two features curves and the
presence of degradation and holes on the model, this example can be considered as a borderline
case. Nevertheless, our method results in a fairly good curve detection.
24
I.
(a)
(b)
(c)
(d)
(a)
(b)
(c)
(d)
II.
Figure 19: Feature recognition on models (from the STARC repository [2]) with partial features.
(a)
(b)
Figure 20: Colours represent different parameters of the HT.
To give a concrete idea of the computational performances of our software prototype we report
the running time of some of the examples discussed in the paper. The proposed method is relatively fast, a significant complexity being induced only by the Curve Detection Algorithm 4. We
implemented the whole pipeline in MATLAB 2016a, while for the HT detection routine we used
the CoCoA library [5]. All experiments were performed on an Intel Core i5 processor (at 2.7 GHz).
Indeed, the Feature Points Recognition Algorithm 2 generally takes an average of 20 seconds on
a model of 150000 vertices and a few seconds are taken by the Aggregation Algorithm 6 and the
Projection Algorithm 7. The complexity of the Curve Detection Algorithm 4 strongly depends on
the family of curves, ranging from less than a minute for the curve with 5-convexities (example
in Figure 14-III), to a couple of minutes for the circle, the Archimedean spiral and the citrus
curve (examples in Figures 14-I, 7 and 16-I), to approximately three minutes for the Lamet curve
(example in Figure 15-I), to a dozen of minutes for the geometric petal (example in Figure 2).
The reasoning on parameters can be extended and possibly generalized. From our experiments,
(a)
(b)
(c)
Figure 21: Detection of two extremely nearby spirals on a degraded architectural ornament artefact
(from the 3D dataset of the EPSRC project [4]).
25
we noticed that eyes of the same collection, such as the ones in the STARC repository [2], share
similar parameters. For instance, the eyes in Figure 16(IV), 18(II), 19(I) and 19(II) have nearly the
same ratio among the parameters a and c (whose square roots represent the height and eccentricity
of the curve) while in the case of the models in Figure 16(II) and 16(III) these values considerably
differ. We plan to deepen these aspects on larger datasets aiming at automatically inferring a
specific style from the models and supporting automatic model annotation.
As a minor drawback, we point out that the use of more complex algebraic curves may involve
more than three parameters which has consequences in the definition and manipulation of the
accumulator function, thus becoming computationally expensive (see Section 3.4), although adhoc methods have been introduced to solve it (see [67]).
It is our opinion that the proposed method is particularly appropriate to drive the recognition
of feature curves and the annotation of shape parts that are somehow expected to be in a model.
As an example, we refer to the case of archaeological artefacts that are always equipped with a
textual description carrying information on the presence of features, e.g. eyes, mouth, decorations,
etc. In our experience, this important extra information allowed us to recognize the eyes in the
models represented in Figure 18(II), presenting a deep level of erosion, and 19(II), where the feature
is only partially discernible. Analogously, we expect the method to be essential in situations where
a taxonomic description of peculiar curve shapes exists (like architectural artefacts) or a standard
reference shape is available, or when it is possible to infer a template of the feature curve to be
identified.
Acknowledgements
Work developed in the CNR research activity DIT.AD004.028.001, and partially supported by the
GRAVITATE European project, “H2020 REFLECTIVE”, contract n. 665155, (2015-2018).
References
[1] IMAGINARY - open mathematics. https://imaginary.org.
[2] STARC repository. http://public.cyi.ac.cy/starcRepo/.
[3] The Shape Repository. http://visionair.ge.imati.cnr.it/ontologies/shapes/, 2011–2015.
[4] Automatic
Semantic
Analysis
of
http://www.ornament3d.org/, 2014–2016.
3D
Content
in
Digital
Repositories.
[5] J. Abbott, A. M. Bigatti, and G. Lagorio. CoCoA-5: a system for doing Computations in
Commutative Algebra. Available at http://cocoa.dima.unige.it.
[6] E. Albuz, E. D. Kocalar, and A. A. Khokhar. Quantized cielab* space and encoded spatial
structure for scalable indexing of large color image archives. In Acoustics, Speech, and Signal
Processing, volume 6, pages 1995–1998, 2000.
[7] A. Andreadis, G. Papaioannou, and P. Mavridis. Generalized digital reassembly using geometric registration. In Digital Heritage, volume 2, pages 549–556, 2015.
[8] Dana H Ballard. Generalizing the Hough transform to detect arbitrary shapes. Pattern
recognition, 13(2):111–122, 1981.
[9] Mauro C Beltrametti and Lorenzo Robbiano. An algebraic approach to Hough transforms. J.
of Algebra, 37:669–681, 2012.
[10] MC Beltrametti, AM Massone, and M Piana. Hough transform of special classes of curves.
SIAM J. Imaging Sci., 6(1):391–412, 2013.
26
[11] S. Biasotti, A. Cerri, A. Bronstein, and M. Bronstein. Recent trends, applications, and
perspectives in 3D shape similarity assessment. Computer Graphics Forum, 35(6):87–119,
2016.
[12] S. Biasotti, L. De Floriani, B. Falcidieno, P. Frosini, D. Giorgi, C. Landi, L. Papaleo, and
M. Spagnuolo. Describing shapes by geometrical-topological properties of real functions. ACM
Computing Surveys, 40(4):1–87, 2008.
[13] Silvia Biasotti, Andrea Cerri, Bianca Falcidieno, and Michela Spagnuolo. 3D artifacts similarity based on the concurrent evaluation of heterogeneous properties. J. Comput. Cult. Herit.,
8(4):19:1–19:19, 2015.
[14] Yuanhao Cao, Dong-Ming Yan, and Peter Wonka. Patch layout generation by detecting feature
networks. Computers & Graphics, 46:275 – 282, 2015.
[15] David Cohen-Steiner and Jean-Marie Morvan. Restricted Delaunay triangulations and normal
cycle. In Proc. of the 9th Ann. Symp. on Computational Geometry, SCG ’03, pages 312–321,
New York, NY, USA, 2003. ACM.
[16] Forrester Cole, Aleksey Golovinskiy, Alex Limpaecher, Heather Stoddart Barros, Adam
Finkelstein, Thomas Funkhouser, and Szymon Rusinkiewicz. Where do people draw lines?
ACM Trans. Graph., 27(3):1–11, August 2008.
[17] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In Computer
Vision and Pattern Recognition (CVPR), 2005 IEEE Conference on, volume 1, pages 886–893,
2005.
[18] Joel Daniels II, Tilo Ochotta, K. Linh Ha, and T. Cláudio Silva. Spline-based feature curves
from point-sampled geometry. The Visual Computer, 24(6):449–462, 2008.
[19] Fernando de Goes, Siome Goldenstein, Mathieu Desbrun, and Luiz Velho. Exoskeleton: Curve
network abstraction for 3d shapes. Computers & Graphics, 35(1):112 – 121, 2011.
[20] Doug DeCarlo and Szymon Rusinkiewicz. Highlight lines for conveying shape. In International
Symposium on Non-Photorealistic Animation and Rendering (NPAR), pages 63–70. ACM,
August 2007.
[21] Richard O. Duda and Peter E. Hart. Use of the Hough transformation to detect lines and
curves in pictures. Commun. ACM, 15(1):11–15, 1972.
[22] Martin Ester, Hans P. Kriegel, Jorg Sander, and Xiaowei Xu. A density-based algorithm
for discovering clusters in large spatial databases with noise. In 2nd Int. Conf. Knowledge
Discovery and Data Mining, pages 226–231. AAAI Press, 1996.
[23] Gerald Farin. Curves and Surfaces for Computer Aided Geometric Design (3rd Ed.): A
Practical Guide. Academic Press Professional, Inc., San Diego, CA, USA, 1993.
[24] Jerome H. Friedman, Jon Louis Bentley, and Raphael Ari Finkel. An algorithm for finding
best matches in logarithmic expected time. ACM Trans. Math. Softw., 3(3):209–226, 1977.
[25] Jerome H. Friedman, Jon Louis Bentley, and Raphael Ari Finkel. An algorithm for finding best
matches in logarithmic expected time. ACM Trans. Math. Softw., 3(3):209–226, September
1977.
[26] Ran Gal and Daniel Cohen-Or. Salient geometric features for partial shape matching and
similarity. ACM Transactions on Graphics (TOG), 25(1):130–150, 2006.
[27] Anne Gehre, Isaak Lim, and Leif Kobbelt. Adapting Feature Curve Networks to a Prescribed
Scale. Computer Graphics Forum, 35(2):319–330, 2016.
27
[28] Stefan Gumhold, Xinlong Wang, and Rob Macleod. Feature extraction from point clouds. In
10th Int. Meshing Roundtable, pages 293–305, 2001.
[29] Gur Harary and Ayellet Tal. The Natural 3D Spiral. Computer Graphics Forum, 30(2):237–
246, 2011.
[30] Gur Harary and Ayellet Tal. 3D Euler spirals for 3D curve completion. Computational
Geometry, 45(3):115 – 126, 2012.
[31] Klaus Hildebrandt, Konrad Polthier, and Max Wardetzky. Smooth feature lines on surface
meshes. In Eurographics Symposium on Geometry Processing. The Eurographics Association,
2005.
[32] Paul. V. C Hough. Method and means for recognizing complex patterns, 1962. US Patent
3,069,654.
[33] R. W. G. Hunt and M. R. Pointer. Measuring Colour, Fourth Edition. Wiley, 2011.
[34] Arik Itskovich and Ayellet Tal. Surface partial matching and application to archaeology.
Computers & Graphics, 35(2):334 – 341, 2011.
[35] Andrew E. Johnson and Martial Hebert. Using spin images for efficient object recognition in
cluttered 3d scenes. IEEE T. Pattern Anal., 21(5):433–449, 1999.
[36] Evangelos Kalogerakis, Patricio Simari, Derek Nowrouzezahrai, and Karan Singh. Robust
statistical estimation of curvature on discretized surfaces. In Proc. of the 5th EG Symp. on
Geometry Processing, pages 13–22. Eurographics Association, 2007.
[37] Asako Kanezaki, Tatsuya Harada, and Yasuo Kuniyoshi. Partial matching of real textured
3d objects using color cubic higher-order local auto-correlation features. Visual Comput.,
26(10):1269–1281, 2010.
[38] A.A Kassim, T Tan, and K.H Tan. A comparative study of efficient generalised Hough
transform techniques. Image and Vision Computing, 17(10):737 – 748, 1999.
[39] J. Kerber, M. Bokeloh, M. Wand, and H.-P. Seidel. Scalable symmetry detection for urban
scenes. Computer Graphics Forum, 32(1):3–15, 2013.
[40] Jan J Koenderink. What does the occluding contour tell us about solid shape? Perception,
13(3):321–330, 1984.
[41] M. Kolomenkin, I. Ilan Shimshoni, and A. Tal. Prominent field for shape processing of archaeological artifacts. Int. J. Comput. Vision, 94(1):89–100, 2011.
[42] Michael Kolomenkin, Ilan Shimshoni, and Ayellet Tal. Demarcating curves for shape illustration. ACM Trans. Graph., 27(5):157:1–157:9, 2008.
[43] Y. K. Lai, Q. Y. Zhou, S. M. Hu, J. Wallner, and H. Pottmann. Robust feature classification
and editing. IEEE Transactions on Visualization and Computer Graphics, 13(1):34–45, Jan
2007.
[44] K. Lawonn, E. Trostmann, B. Preim, and K. Hildebrandt. Visualization and extraction of
carvings for heritage conservation. IEEE Transactions on Visualization and Computer Graphics, 23(1):801–810, Jan 2017.
[45] C. Li, M. Wand, X. Wu, and H. P. Seidel. Approximate 3d partial symmetry detection using
co-occurrence analysis. In 2015 International Conference on 3D Vision, pages 425–433, Oct
2015.
28
[46] Yong-Jin Liu, Yi-Fu Zheng, Lu Lv, Yu-Ming Xuan, and Xiao-Lan Fu. 3D model retrieval
based on color + geometry signatures. Visual Comput., 28(1):75–86, 2012.
[47] David G. Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput.
Vision, 60(2):91–110, 2004.
[48] Tao Luo, Renju Li, and Hongbin Zha. 3D line drawing for archaeological illustration. Int. J.
Comput. Vision, 94(1):23–35, 2010.
[49] Priyanka Mukhopadhyay and Bidyut B. Chaudhuri. A survey of Hough transform. Pattern
Recognition, 48(3):993 – 1010, 2015.
[50] Sven Oesau, Florent Lafarge, and Pierre Alliez. Indoor Scene Reconstruction using Feature
Sensitive Primitive Extraction and Graph-cut. ISPRS Journal of Photogrammetry and Remote
Sensing, 90:68–82, March 2014.
[51] Yutaka Ohtake, Alexander Belyaev, and Hans-Peter Seidel. Ridge-valley lines on meshes via
implicit surface fitting. ACM Trans. Graph., 23(3):609–612, 2004.
[52] Timo Ojala, Matti Pietikäinen, and David Harwood. A comparative study of texture measures
with classification based on featured distributions. Pattern Recognition, 29(1):51–59, 1996.
[53] Giuliano Pasqualotto, Pietro Zanuttigh, and Guido M. Cortelazzo. Combining color and shape
descriptors for 3D model retrieval. Signal Process-Image, 28(6):608 – 623, 2013.
[54] Mark Pauly, Richard Keiser, and Markus Gross. Multi-scale feature extraction on pointsampled surfaces. Comput. Graph. Forum, 22(3):281–289, 2003.
[55] G. Peyre.
Toolbox graph - A toolbox to process graph and triangulated meshes.
http://www.ceremade.dauphine.fr/ peyre/matlab/graph/content.html.
[56] Les Piegl and Wayne Tiller. The NURBS Book (2Nd Ed.). Springer-Verlag New York, Inc.,
New York, NY, USA, 1997.
[57] W. Rudin. Principles of Mathematical Analysis. McGraw-Hill,, Singapore:, 3rd edition edition,
1976.
[58] C.R. Ruiz, R. Cabredo, L.J. Monteverde, and Zhiyong Huang. Combining Shape and Color
for Retrieval of 3D Models. In INC, IMS and IDC (NCM’09). Fifth International Joint
Conference on, pages 1295–1300, 2009.
[59] Ariel Shamir. Segmentation and Shape Extraction of 3D Boundary Meshes. In Brian Wyvill
and Alexander Wilkie, editors, Eurographics 2006 - State of the Art Reports. The Eurographics
Association, 2006.
[60] E. V. Shikin and A. I. Plis. Handbook on Splines for the User. CRC Press, Boca Raton, FL,
1995.
[61] Eugene V Shikin. Handbook and atlas of curves. CRC, 1995.
[62] J. Starck and A Hilton. Correspondence labelling for wide-timeframe free-form surface matching. In Computer Vision (ICCV), 2007 IEEE International Conference on, pages 1–8, 2007.
[63] Martin Sunkel, Silke Jansen, Michael Wand, Elmar Eisemann, and Hans-Peter Seidel. Learning line features in 3D geometry. Computer Graphics Forum (Proc. EUROGRAPHICS), 30(2),
April 2011.
[64] M.T. Suzuki. A Web-based retrieval system for 3D polygonal models. In IFSA World Congress
and 20th NAFIPS International Conference. Joint 9th, volume 4, pages 2271–2276, 2001.
29
[65] James Tanaka, Daniel Weiskopf, and Pepper Williams. The role of color in high-level vision.
Trends in cognitive sciences, 5(5):211–215, 2001.
[66] F. Tombari, S. Salti, and L. Di Stefano. A combined texture-shape descriptor for enhanced
3d feature matching. In Image Processing (ICIP), 2011 IEEE International Conference on,
pages 809–812, 2011.
[67] Maria-Laura Torrente and Mauro C Beltrametti. Almost vanishing polynomials and an application to the Hough transform. J. of Algebra and Its Applications, 13(08):1450057, 2014.
[68] Maria-Laura Torrente, Mauro C. Beltrametti, and Juan Rafael Sendra. Perturbation of polynomials and applications to the hough transform. Journal of Algebra, 486:328 – 359, 2017.
[69] Maria-Laura Torrente, Silvia Biasotti, and Bianca Falcidieno. Feature Identification in Archaeological Fragments Using Families of Algebraic Curves. In Chiara Eva Catalano and Livio De
Luca, editors, Eurographics Workshop on Graphics and Cultural Heritage. The Eurographics
Association, 2016.
[70] Libor Váša, Petr Vaněček, Martin Prantl, Věra Skorkovská, Petr Martı́nek, and Ivana
Kolingerová. Mesh Statistics for Robust Curvature Estimation. Computer Graphics Forum,
35(5):271–280, 2016.
[71] Changchang Wu, B. Clipp, Xiaowei Li, J.-M. Frahm, and M. Pollefeys. 3D model matching with Viewpoint-Invariant Patches (VIP). In Computer Vision and Pattern Recognition
(CVPR), 2008 IEEE Conference on, pages 1–8, 2008.
[72] Shin Yoshizawa, Alexander Belyaev, Hideo Yokota, and Hans-Peter Seidel. Fast, robust, and
faithful methods for detecting crest lines on meshes. Comput. Aided Geom. Des., 25(8):545–
560, 2008.
[73] Andrei Zaharescu, Edmond Boyer, and Radu Horaud. Keypoints and local descriptors of
scalar functions on 2D manifolds. Int. J. Comput. Vision, 100(1):78–98, 2012.
[74] Yuhe Zhang, Guohua Geng, Xiaoran Wei, Shunli Zhang, and Shanshan Li. A statistical
approach for extraction of feature lines from point clouds. Computers & Graphics, 56:31 – 45,
2016.
30
Appendix
We list the pseudocode of some procedures mentioned into Sections 3.1-3.3.
Filtering Algorithm 5: Computes the value v of the histogram h corresponding to the
threshold p.
Input : histogram h, cut percentage p
Output: cut value v
1 begin
2
Values = get values(h);
3
V=0; k=0;
4
while k ≤ size(h) and V < p · size(h) do
5
k = k + 1; V = V+Values[k];
6
end
7
m = min bin value(h);
8
width= get bin width(h);
return : v = m+k · width
9 end
Aggregation Algorithm 6: Groups the points of the set X into smaller dense subsets
Input : set of feature points X ⊂ R3
Output: Y, a set of groups Yj of elements of X
1 begin
/* searches the average distance from each point of X to its K-nearest
neighbours; the parameter K is fixed and refers to how many neighbours
we want to consider
*/
2
K = 50;
3
[IDX, D] = knnsearch(X, X, ‘k 0 , K);
/* computes an estimate for the threshold ε
*/
4
ε = mean(D(:, K));
/* aggregates the points of X using the DBSCAN algorithm
*/
5
M inP ts = 5;
6
IDX = DBSCAN(X, ε, MinPts);
7
for j = 1, . . . , size(IDX) do
8
Yj ← the points of X of indices IDX[j];
9
add Yj to the list Y
10
end
return : Y
11 end
Projection Algorithm 7: Projects a set of feature points Y onto a best fitting plane
Input : a set of points Y
Output: the projection Z of Y onto a best fitting plane
1 begin
/* shifts the points of Y to move the centroid onto (0, 0, 0)
*/
2
Y ← Y − mean(Y);
/* computes the best fitting plane Π of Y with linear regression
*/
3
XY ← Y[1, 2]; Z ← Y[3]; Π ← regress(Z, XY );
/* defines the orthogonal transformation which moves Π to the plane z = 0,
and apply it to the points of Y
*/
4
n ← [Π(1), Π(2), −1]; v1 ← [1, 0, Π(1)]; v2 ← CrossProduct(v1 , n);
5
R ← [v1 /norm(v1 ), v2 /norm(v2 ), n/norm(n)];
6
Z ← R(Y);
return : Z[1, 2]
7 end
31
Maria-Laura Torrente, Istituto di Matematica Applicata e Tecnologie Informatiche “E. Magenes”
CNR, Genova, Italy. e-mail [email protected]
Silvia Biasotti, Istituto di Matematica Applicata e Tecnologie Informatiche “E. Magenes” CNR,
Genova, Italy. e-mail [email protected]
Bianca Falcidieno, Istituto di Matematica Applicata e Tecnologie Informatiche “E. Magenes” CNR,
Genova, Italy. e-mail [email protected]
32
| 1cs.CV
|
High-throughput genome-wide association analysis
for single and multiple phenotypes
Diego Fabregat-Traver
Yurii S. Aulchenko
Paolo Bientinesi
arXiv:1207.2169v2 [cs.CE] 10 Nov 2012
Aachen Institute for Advanced Study
Institute of Cytology and Genetics
Aachen Institute for Advanced Study
in Computational Engineering Science SD RAS, Novosibirsk, 630090, Russia in Computational Engineering Science
RWTH Aachen, Aachen, 52062, Germany Email: [email protected] RWTH Aachen, Aachen, 52062, Germany
Email: [email protected]
Email: [email protected]
Abstract—The variance component tests used in genomewide association studies of thousands of individuals become
computationally exhaustive when multiple traits are analysed in
the context of omics studies. We introduce two high-throughput
algorithms —CLAK-C HOL and CLAK-E IG— for single and multiple phenotype genome-wide association studies (GWAS). The
algorithms, generated with the help of an expert system, reduce
the computational complexity to the point that thousands of traits
can be analyzed for association with millions of polymorphisms in
a course of days on a standard workstation. By taking advantage
of problem specific knowledge, CLAK-C HOL and CLAK-E IG
significantly outperform the current state-of-the-art tools in both
single and multiple trait analysis.
Current biomedical research is experiencing a large boost
in the amount of data generated. Individual genomes are
being characterized at increased level of details using single
nucleotide polymorphism (SNP) arrays, and, more recently,
exome and whole-genome re-sequencing. At the same time,
technologies for high-throughput characterization of tens of
thousands of molecular omics phenotypes in thousands of
people are becoming increasingly affordable [1] [2], [3] [4].
Genome-Wide Association Studies (GWAS) is a powerful
tool for identifying loci involved in the control of complex
traits [5]. In such studies, thousands of individuals are measured for the trait of interest and their genomes are characterized by re-sequencing or by genotyping of millions of genetic
markers using SNP arrays. The association between genetic
markers and a phenotype of interest is studied, with significant association highlighting the genomic regions harboring
functional variants involved in the control of the trait.
Even in carefully designed population-based studies, some
degree of relatedness and population stratification is expected.
One of the most flexible and powerful methods of accounting
for such substructure is the Variance Components (VC) approach based on linear mixed models [6], [7] (Supplementary
Note, Section 1). Recent FaST-LMM [8] implementation of
the VC model can complete genome-wide scan of association
between 36 millions of genetic markers and a phenotype
in 1,000 individuals in about 4 hours using a multicore(12)
processor. While these results are impressive and represent a
breakthrough compared to older algorithms and implementations, such computational throughput is still prohibitively low
for analysis of omics data. For example, for transcriptome
(30,000 phenotypes), the estimated time to complete the analy-
sis would be approximately 13 years. Currently, such amount
of computations would only be tractable if large computing
facilities with many thousands of cores are used.
In this work, we consider the problem of genome-wide
association analysis using the mixed models, with special
emphasis on analysis of thousands of phenotypes (Fig. 1). We
demonstrate how utilization of problem-specific knowledge,
coupled with innovative techniques of automatic generation
of linear algebra algorithms and hardware-tailored implementation, allows reducing the computation time from years to
days. For the aforementioned analysis of transcriptome, our
algorithm takes 26 days on a single node, or equivalently, 20
hours on a standard 32-node cluster.
In the context of linear algebra computations, we recently
developed a symbolic system, CLAK, that closely mirrors the
reasoning of a human expert for the generation and analysis
of algorithms [9]. The idea is to first decompose a target
linear algebra operation in terms of library-supported kernels,
and then apply optimizations aimed at reducing redundant
calculations. Since the decomposition is not unique, CLAK
returns not one but a family of algorithms, together with their
corresponding cost estimates.
Figure 1 illustrates how multi-trait analysis consists of t separate single-trait analyses, each of which, in turn, consists of
m generalized least-squares (GLS) problems. The key to fast
algorithms is the realization that such problems are correlated,
both in the m and t directions; a naive approach that only aims
at optimizing one GLS in isolation will never be competitive
with methods that tackle sequences as a whole. With the
help of CLAK, we generated more than 20 algorithms for
performing genome-wide association analysis. Based on the
cost estimates, we assessed the potential of these algorithms
for single-trait and multiple-trait studies. Interestingly, for the
two scenarios, the best theoretical performance was attained
by two different algorithms, CLAK-C HOL and CLAK-E IG,
respectively (Supplementary Note, Sections 2 and 3).
Both algorithms incorporate a number of techniques for
increasing the computational efficiency and save intermediate
results across adjacent problems. In CLAK-C HOL, differently
than in the other existing methods (such as EMMAX [10] and
FaST-LMM), the positive-definiteness of the covariance matrix
is exploited to utilize a faster factorization; moreover, the
−1
=
bγ,β
XγT
Mβ
−1
Xγ
−1
=
XγT
Mβ
yβ
XδT
bδ,β
γ
−1
=
bγ,α
XγT
Mα
−1
Xγ
Mβ
−1
Xδ
XδT
Mβ
δ
yβ
XL
XRδ
β
XγT
Mα
yα
α
B
t
p
m
Figure 1: Interpretation of GWAS as a 2-dimensional sequence of generalized least-squares problems (b :=
(X T M −1 X)−1 X T M −1 y). GWAS with multiple phenotypes requires the solution of m×t correlated GLS problems, originating
a three-dimensional object B of size m × t × p. Along the t direction, the covariance matrix M and the phenotype y vary,
while the design matrix X does not; conversely, in the m direction, M and y are fixed while X varies. Specifically, X can
be viewed as consisting of two parts, XL and XR , where the former is constant across the entire grid and the latter changes
along m. The figure also captures GWAS with single phenotype, in which case the dimension t reduces to 1.
computation is rearranged to fully benefit from the potential
of the underlying optimized kernels. In CLAK-E IG instead,
redundant computation is avoided by exploiting the fact that
the relationship (kinship) matrix is constant even across different traits, and only the heritability and total variance of the
trait change. When compared with the current state-of-theart implementations such as FaST-LMM, both our algorithms
achieve a lower computational complexity (for a comparison,
see Supplementary Table 1).
The size of the data sets involved in genome-wide association studies is considerably larger than the memory capacity
of current processors; input and output data can only be
stored in disk devices. Since the penalty for accessing a
piece of data residing on disk is enormous –several order of
magnitude greater than the cost for performing one arithmetic
operation– it is imperative to efficiently handle the data. From
our experiments, in fact, we observed that the time spent on
data movement adds about 30% to the time spent on arithmetic
calculations. Instead, through asynchronous transfers between
memory and disk, our algorithms achieve a perfect overlap of
computation and data movement. As long as the covariance
matrix fits in main memory, and regardless of the size of
the data sets –both in terms of SNPs and phenotypes–, the
processor never idles waiting for data, thus computing at
maximum efficiency.
To demonstrate the practical advantages of CLAK-E IG and
CLAK-C HOL, we implemented routines and compared their
execution time with that of well-established methods: EMMAX, FaST-LMM (two-step approximation), and GWFGLS
(implementation of the mmscore method of ProbABEL [11]
in the MixABEL-package). In the experiments we considered
three different scenarios, varying the sample size, the number
of SNPs, and the number of traits, while keeping the other
two values constant (Fig. 2). A description of the experimental
setup is provided in the Supplementary Note, Section 4.
In the first scenario (single trait and 10,000,000 SNPs), even
though all methods exhibit a quadratic behaviour, CLAKC HOL is the only algorithm that completed all tests within
1.5 days. For the largest problem considered (sample size
n = 40,000), the speedup over FaST-LMM is 4.53: 158 hours
vs. 35; for n = 1,000 instead, the speedups over GWFGLS,
FaST-LMM and EMMAX are 15, 28 and 106, respectively:
38, 68 and 257 minutes vs. 2.5 minutes.
The second scenario (single trait and sample size of 10,000)
shows a linear dependence on the number of genetic markers
for all methods. Again, CLAK-C HOL attains the best timings,
outperforming FaST-LMM, GWFGLS and EMMAX by a
factor of 6.3, 56.8 and 112, respectively.
Thanks to CLAK-E IG’s linear complexity with respect to
sample size, SNPs, and traits, the advantage in the analysis
of multiple phenotypes (third scenario: sample size of 1, 000
and 1,000,000 genetic markers) becomes most apparent: when
thousands and more traits are considered, CLAK-E IG outperforms GWFGLS, FaST-LMM and EMMAX by a factor of
105, 177, and 418, respectively, bringing the execution time
from several months down to two days.
Establishing the functional roles of genetic variants and
finding the biological link between genetic variation and
complex phenotypes remain significant challenges in the postgenomic era. Genome-wide association studies are increasingly applied to understand the regulation of human and
animal transcriptome [1], metabolome [2], [3], glycome [4]
and other types of omics data; they are also used to uncover
the link between these molecular phenotypes and high-level
complex traits, including common diseases [12]. In GWAS, it
is well recognized that genetic (sub)structure can act as confounder and lead to false-positive discoveries, unless correctly
accounted for; one of the most flexible and powerful methods
for such a correction is provided by the mixed models [6].
However, this powerful method comes at a high computational
price: Even when using the most advanced methods and
tools [8], it takes a few hours for the GWAS to complete.
80
EMMAX
GWFGLS
FaSTLMM
CLAK-Chol
60
40
Time ( hours )
Time ( hours )
80
40 hours
20
10
10,000
20,000
60
40
20
8.5 hours
106
40,000
107
3.6*107
n
m
(a)
(b)
300
800
600
14.4 months
400
37 weeks
200
58 hours
2
4
10
EMMAX: 418x
2.78 years
EMMAX
FaSTLMM
GWFGLS
CLAK-Eig
Ratio over CLAK-Eig
1000
Time ( days )
53 hours
10
9 hours
1,000
EMMAX
GWFGLS
FaSTLMM
CLAK-Chol
5 * 10
t
4
250
FaST-LMM: 177x
200
150
GWFGLS: 105x
100
50
CLAK-Eig:
1
5
10
(c)
0
10
1
10
2
3
10
10
4
10
1x
105
t
(d)
Figure 2: Timing comparison. Panels (a) and (b) include timings for EMMAX, GWFGLS, FaST-LMM, and CLAK-C HOL,
relative to single trait analysis; (c) and (d) present a comparison of EMMAX, GWFGLS, FaST-LMM, and CLAK-E IG in the
case of multiple traits. In (a), the number of SNPs is fixed to m = 10,000,000 and the sample size n ranges from 1,000 to
40,000. In panel (b), the sample size is fixed to n = 10,000 and the number of SNPs m ranges between 106 and 3.6 × 107 .
In (c) and (d), n = 1,000, m = 106 and t ranges from 1 to 100,000.
While this time scale might not seem a big hindrance when
only a relatively small number of phenotypes are analyzed (for
instance in studies of complex traits and common diseases), the
issues become apparent as soon as omics data are considered.
The computational time becomes prohibitively large (up to
several years), and in order to complete the studies within
reasonable time one has to count on supercomputers with
thousands of cores. This solution delays the research and
increases its cost.
In this work, we clearly demonstrate that use of problemspecific knowledge, coupled with innovative techniques for
algorithm generation and with hardware-tailored implementations, leads to both a decrease in computational complexity and
an increase in algorithmic efficiency. Specifically, for the analysis of omics data we were able to attain remarkable speedups, making it feasible to analyze —on a single standard multicore computer, as opposed to supercomputing facilities— tens
and even hundreds of thousands of phenotypes in the course
of few days rather than years.
Further optimizations are possible, for instance by exploiting the structure of the kinship matrix. A compressed MLM
approach was proposed for decreasing the effective sample
size of datasets by clustering individuals into groups [13];
similarly, the fast decaying and possibly sparse structure of
the kinship matrix can be exploited to decrease the algorithmic
complexity.
We believe that the approach outlined here —integrating
problem-specific knowledge with automatic algorithm generation and hardware-tailored implementation—- has many
applications in high-throughput analysis of biomedical data.
I. ACKNOWLEDGEMENTS
The work of YSA was funded by grants from the Russian Foundation of Basic Research (RFBR), the Helmholtz
society (RFBR-Helmholtz Joint Research Groups), and the
MIMOmics project supported by FP7.
PB and DFT gratefully acknowledge the support received
from the Deutsche Forschungsgemeinschaft (German Research
Association) through grant GSC 111.
The authors wish to thank the Center for Computing
and Communication at RWTH Aachen for the computing
resources.
R EFERENCES
[1] Goring, H. H. H. et al. Discovery of expression qtls using large-scale
transcriptional profiling in human lymphocytes. Nat Genet 39, 1208–
1216 (2007). URL http://dx.doi.org/10.1038/ng2119.
[2] Hicks, A. A. et al. Genetic determinants of circulating sphingolipid
concentrations in european populations. PLoS Genet 5, e1000672
(2009). URL http://dx.doi.org/10.1371%2Fjournal.pgen.1000672.
[3] Demirkan, A. et al. Genome-wide association study identifies novel loci
associated with circulating phospho- and sphingolipid concentrations.
PLoS Genet 8, e1002490 (2012). URL http://dx.doi.org/10.1371%
2Fjournal.pgen.1002490.
[4] Lauc, G. et al. Genomics meets glycomics –the first gwas study of
human n-glycome identifies hnf1 as a master regulator of plasma protein
fucosylation. PLoS Genet 6, e1001256 (2010). URL http://dx.doi.org/
10.1371%2Fjournal.pgen.1001256.
[5] Hindorff, L. A. et al. Potential etiologic and functional implications of
genome-wide association loci for human diseases and traits. Proc. Natl.
Acad. Sci. 106, 9362–9367 (2009).
[6] Yu, J. et al. A unified mixed-model method for association mapping
that accounts for multiple levels of relatedness. Nat. Genet. 38, 203–
208 (2006).
[7] Astle, W. & Balding, D. Population structure and cryptic relatedness
in genetic association studies. Statist. Sci. 24, 451–471 (2009). URL
http://dx.doi.org/10.1214/09-STS307.
[8] Lippert, C. et al. Fast linear mixed models for genome-wide association
studies. Nat Meth 8, 833–835 (2011). URL http://dx.doi.org/10.1038/
nmeth.1681.
[9] Fabregat-Traver, D. & Bientinesi, P. A domain-specific compiler for
linear algebra operations. In Proceedings of VECPAR 2012, Lecture
Notes in Computer Science (Springer Berlin / Heidelberg, 2012).
[10] Kang, H. M. et al. Variance component model to account for sample
structure in genome-wide association studies. Nat. Genet. 42, 348–354
(2010). URL http://dx.doi.org/10.1038/ng.548.
[11] Aulchenko, Y., Struchalin, M. & van Duijn, C. Probabel package for
genome-wide association analysis of imputed data. BMC Bioinformatics
11, 134 (2010).
[12] Cookson, W., Liang, L., Abecasis, G., Moffatt, M. & Lathrop, M.
Mapping complex disease traits with global gene expression. Nat. Rev.
Genet. 10, 184–194 (2009). URL http://dx.doi.org/10.1038/nrg2537.
[13] Zhang, Z. et al. Mixed linear model approach adapted for genomewide association studies. Nat. Genet. 42, 355–360 (2010). URL http:
//dx.doi.org/10.1038/ng.546.
arXiv:1207.2169v2 [cs.CE] 10 Nov 2012
Supplementary Notes
Diego Fabregat, Yurii Aulchenko, Paolo Bientinesi
February 17, 2018
1
Mixed Models for GWAS
The Variance Components model for a quantitative trait can be formulated as
Y = Xβ + R,
where Y is the vector containing the phenotypes for n individuals, X is the design
matrix, and β and R are the vectors of fixed and random effects, respectively. The
partitioning X = [1|L|g] indicates that the design matrix is composed of three parts:
1 denotes a column-vector (corresponding to the intercept) containing ones, L is an
n × p matrix corresponding to fixed covariates such as age and sex, and g typically
consists of a single column-vector containing genotypes. The vector of random effect R
is assumed to be distributed as a Multivariate Normal with mean zero and variancecovariance matrix M = σ 2 · (h2 Φ + (1 − h2 )I); here, σ 2 is the total variance of the
trait, h2 (in the range [0, 1]) is the heritability coefficient –which defines the strength of
the correlation between phenotypes of relatives–, I is the identity matrix, and Φ is the
matrix containing the relationship coefficients for all pairs of studied individuals. The
relationship coefficient is defined as the proportion of the genome identical-by-descent
that a pair of individuals share; for example, in the case of identical twins (they have
the same DNA) the relationship coefficient is 1, while since a parent transmits half of
its genome to the offspring, the relationship coefficient for a parent with its offspring is
0.5. The relationship matrix Φ can be estimated from the pedigree or from the genomic
data [1]. In GWAS, the quantity of interest is the effect of the genotype, that is, the
element(s) of β corresponding to g. Technically, a GWAS with mixed model consists of
traversing all measured polymorphic sites in the genome, substituting the corresponding
g into X, and fitting the above model; the result is millions of estimates of genetic effect
together with their p-values.
1
One of the most used mixed model-based approaches used in GWAS relies on a twostep analysis methodology [2, 3, 4, 5, 6]. In the first step, the reduced model (with X =
[1|L]) is fit to the data, thus obtaining the estimates {σˆ2 , hˆ2 }; the variance-covariance
matrix corresponding to such estimates is denoted by M̂ = σ̂ 2 · (ĥ2 Φ + (1 − ĥ2 )I). In
the second step, for each gi and corresponding Xi = [1|L|gi ], the estimates of the effects
and the variance-covariance matrix are respectively obtained as
β̂i = (XiT M̂ −1 Xi )−1 XiT M̂ −1 y,
(1)
and
V ar(β̂i ) = σ̂ 2 · (XiT M̂ −1 Xi )−1 ,
with i = 1, . . . , m, and m is the number of genetic markers considered.
In this work, we consider an extended formulation of this problem to the case of
multiple phenotypes, that is, Y is a collection of t vectors, with yj (j = 1, . . . , t) being a
vector corresponding to a specific trait. In this case, trait-specific estimates σ̂j2 , ĥ2j need
to be obtained, resulting in t different M̂j s. As the result of the analysis, m × t vectors
of estimates of β̂ij and corresponding V ar(β̂ij ) are generated. In summary, the problem
we are facing is (see Fig. 1 of the main manuscript text for a visual description)
β̂ij = (XiT M̂j−1 Xi )−1 XiT M̂j−1 yj
with 1 ≤ i ≤ m
˙ T M̂ −1 Xi )−1
V ar(β̂ij ) = σ̂j2 (X
(2)
i
j
and 1 ≤ j ≤ t.
M̂j = σ̂j2 (ĥ2j Φ + (1 − ĥ2j )I)
2
Single-instance Algorithms
In this section we provide an overview of a simplified version of CLAK-Chol and
CLAK-Eig to solve a single GLS; the versions tailored for sequences of GLS’s are then
introduced in Section 3. In the following, we use b to indicate β̂; in both our algorithms,
the computation of V ar(b) represents an intermediate result towards b.
CLAK-Chol The approach for CLAK-Chol (Alg. 1) is to first reduce the initial
GLS b := (X T M̂ −1 X)−1 X T M̂ −1 y to a linear least-squares problem, and then solve this
via normal equations. Specifically, the algorithm starts by forming M̂ = σˆ2 · (hˆ2 Φ +
(1 − hˆ2 )I), which is known to be symmetric positive definite, and by computing its
Cholesky factor L. This leads to the expression b := (X T L−T L−1 X)−1 X T L−T L−1 y,
in which two triangular linear systems can be identified and solved — X 0 := L−1 X
and y 0 := L−1 y — thus completing the reduction to the standard least-squares problem
b := (X 0T X 0 )−1 X 0T y 0 . Numerical considerations allow us to safely rely on the Cholesky
factorization of S := X 0T X 0 without incurring instabilities. The algorithm completes
by computing b0 := X 0T y 0 and solving the linear system b := S −1 b0 , for a total cost of
1 3
n + O(n2 p) for a single GLS.
3
Algorithm 1: CLAK-Chol’s approach
for a single GLS problem
M̂ := σˆ2 · (hˆ2 Φ + (1 − hˆ2 )I)
LLT = M̂
X := L−1 X
y := L−1 y
S := X T X
b := X T y
b := S −1 b
1
2
3
4
5
6
7
CLAK-Eig Instead of forming the matrix M̂ , Algorithm CLAK-Eig (Alg. 2) operates on the matrix Φ: At first, it diagonalizes Φ as ZW Z T , leading to the expression
M̂ := σˆ2 (hˆ2 ZW Z T + (1 − h2 )I),
with diagonal W . By orthogonality of Z, the inverse of M̂ can be represented as
M̂ −1 := Z(σˆ2 (hˆ2 W + (1 − h2 )I))−1 Z T
and easily computed via
D := (σˆ2 (hˆ2 W + (1 − h2 )I))−1 ;
the solution to the GLS is thus given by
b := (X T ZD−1 Z T X)−1 X T ZD−1 Z T y.
(3)
Moreover, since D is symmetric positive definite, Eq. 3 can be rewritten as
b := (X T ZKK T Z T X)−1 X T ZKK T Z T y,
and the algorithm proceeds by computing V := X T ZK (matrix-matrix multiplication
and scaling), obtaining b := (V V T )−1 V K T Z T y. Similar to CLAK-Chol, b is finally
obtained through matrix-vector multiplications and a linear system, for a total cost of
10 3
n + O(n2 p).
3
Algorithm 2: CLAK-Eig’s approach
for a single GLS problem
1
2
3
4
5
6
7
3
ZW Z T = Φ
D := (σˆ2 (hˆ2 W + (1 − h2 )I))−1
KK T = D
V := X T ZK
S := V V T
b := V K T Z T y
b := S −1 b
Sequences of GLS’
As shown in Table1, the strength of CLAK-Chol and CLAK-Eig becomes apparent
in the context of 1D and 2D sequences of GLS’, corresponding to GWAS with single
and multiple phenotypes, respectively. The straightforward approach, which is the only
alternative provided by current general-purpose numerical libraries, lies in a loop that
utilizes the best performing algorithm for one GLS. No matter how optimized the GLS
solver, such an approach is prohibitively expensive for either single or multiple phenotypes, due to the unmanageable complexity of O(mn3 ) and O(tmn3 ), respectively. By
contrast, the versions of CLAK-Chol and CLAK-Eig shown in Fig. ?? are the product of a number of optimizations aimed at saving intermediate results across successive
problems, thus avoiding redundant calculations. For instance, the matrix X is logically
split as [XL |XR ], where XL includes the intercept and the covariates ([1|L]), while XR
is the collection of all the genetic markers gi . Thanks to these savings, both algorithms
achieve a lower overall complexity (Table 1).
Algorithm 3: CLAK-Chol
single-trait analysis
1
2
3
4
5
6
7
8
9
10
11
M̂ := σˆ2 · (hˆ2 Φ + (1 − hˆ2 )I)
LLT = M̂
XL := L−1 XL
XR := L−1 XR
y := L−1 y
ST L := XLT XL
bT := XLT y
f or i = 1 :m
ST L
?
S :=
T
T
XRi XL XRi XRi
!
bT
bi :=
T
y
XR
i
−1
bi := Si bi
Algorithm 4: CLAK-Eig for
multi-trait analysis
1
2
3
4
5
6
7
!
8
9
10
11
12
13
14
15
16
ZW Z T = Φ
XL := Z T XL
XR := Z T XR
Y := Z T Y
fo r j = 1 : t
2
2
D := (σˆj 2 (hˆj W + (1 − hˆj )I))−1
KK T = D
Yj := K T Yj
WL := K T XL
ST L := WLT WL
bT := WLT Yj
fo r i = 1 :m
WR := K T XRi
!
S :=
ST L
WRT WL
bT
WRT Yj
bij := S −1 bij
bij :=
?
WTW
! R R
Unfortunately, the reduced complexity of algorithms CLAK-Chol and CLAK-Eig
is not enough to guarantee high-performance implementations. It is well known that
in terms of execution time, a straightforward translation of algorithms into code and
a carefully assembled routine often differ by at least one order of magnitude. In other
words, the benefits inherent to our new algorithms might go unnoticed unless paired
with state-of-the-art implementation techniques. In this section we detail our strategy
to attain high-performance routines.
Both CLAK-Chol and CLAK-Eig are entirely expressed in terms of standard
linear algebra operations such as matrix products and matrix factorizations, provided
by the BLAS [7] and LAPACK [8] libraries. Since LAPACK itself is built in terms of
BLAS kernels, these are the main responsible for the overall performance of an algorithm. BLAS consists of a relatively small set of highly optimized kernels, organized
in three levels, corresponding to vector, vector-matrix, and matrix-matrix operations,
respectively. A common misconception is that all the BLAS kernels, across the three
levels, attain a comparable (and high) level of efficiency. Instead, it is only BLAS-3
–when operating on large matrices– that fully exploits the processors’ potential; as an
example, the matrix-vector multiplication, matrix-matrix multiplication on small matrices, and matrix-matrix-multiplication on large matrices attain an efficiency of 10%,
≈ 40%, and more than 95%, respectively. In this context, the linear systems X := L−1 X
in CLAK-Chol (Line 3 of Algorithm 1), to be solved for each individual SNP, should
ideally be aggregated into a single —very large— linear system XR := L−1 XR , in which
XR is the collection of the genetic markers of all SNPs (Line 3 of Algorithm 3). An
analogous comment is valid for CLAK-Eig (see the multiplications at Lines 3 and 4 of
Algorithm 4), in which the number of markers accessed at once is a user-configurable
input parameter.
Since all the current computing platforms include multicore processors, we now
briefly discuss how to take advantage of this architecture. An effective practice is to
invoke a multi-threaded version of the BLAS library. Both in Algorithm 3 and 4, we
rely on such a solution for the sections leading up to the outer loop (Lines 1–6 and
1–4, respectively). In the remainder of the algorithms (Lines 7–11 and 5–13), due to
the lack of matrix-matrix operations, we instead apply a thread-based parallelization
in conjunction with single-threaded BLAS. This mixed use of multi-threading becomes
more and more effective as the number of available computing cores increases, leading
to speedups up to 10 or even 20%.
3.1
Space requirement
CLAK-Chol To form and factor the variance matrix, the algorithm uses n2 memory
(Lines 1–2 in Algorithm 3). The overall space requirement is determined by the triangular solve (Line 4), which necessitates the full L and a portion of X to reside in memory
at the same time. This operation is performed in a streaming fashion —operating on k
SNPs at the time— and overwriting X. All the other instructions do not require any
extra space. In total, CLAK-Chol uses about n2 + knp memory.
CLAK-Eig The initial eigendecomposition (Line 1 in Algorithm 4) needs 2n2 memory. The following matrix-matrix multiplications (Lines 2, 3 and 8) overwrite the input
matrices and again are performed in a streaming fashion; in terms of space, the analysis
is similar to that for the triangular solve in CLAK-Chol. The remaining instructions
do not affect the overall memory usage. In total, the space requirement is 2n2 + knp.
If memory is a limiting factor, one can set k to 1 in either algorithm, possibly at
the cost of exposed data movement. Moreover, by using a different eigensolver, the
eigenvectors Z can overwrite the input matrix Φ, effectively saving n2 memory. These
considerations indicate that our algorithms are capable of solving GWAS of any size, as
long as the kinship matrix and one SNP fit in RAM.
3.2
Time complexity
As it was described in Sec. 1, before the sequence of GLS can be solved, one needs to
estimate the parameters related to the random part of the model, namely the heritability
(h2 ) and the total variance (σ 2 ). Software packages such as GenABEL [9], EMMAX, and
FaST-LMM use algorithms similar to our CLAK-Eig. First, the eigendecomposition
of the kinship matrix is performed, for a computational cost of O(n3 ). Then, the model
parameters are estimated using an iterative procedure based on the Maximum Likelihood
(GenABEL, FaST-LMM) or Restricted Maximum Likelihood (EMMAX) methods, for a
cost of vnp2 , where v is the average number of iterations required to reach convergence.
n3 + O(vnp2 ) operations.
In total, the parameter estimation has a complexity of 10
3
By contrast, when our CLAK-Chol algorithm is used, the initial estimation requires a Cholesky factorization of the kinship matrix at each of the v iterations, for a
total cost of 31 vn3 operations. Due to the reduced cost and better performance of such
a factorization, this strategy is advantageous when v <≈ 30, a condition commonly
verified in practice.
In the second step, a 1D or 2D sequence of GLS is solved, corresponding to single
and multiple phenotype analysis. For 1D sequences, all the considered methods share
the same asymptotic time complexity, but the constant factor for CLAK-Chol is the
lowest, yielding at least a 4-fold speedup. In 2D sequences, EMMAX, FaST-LMM
and GenABEL simply tackle each individual trait independently, one after another,
for a total complexity of tmpn2 . By exploiting the common structure of the variancecovariance matrix of different phenotypes, CLAK-Eig reduces the complexity by a
factor of n, down to tmpn. As a result, our algorithm outperforms the other methods
by a factor of 100 and more.
4
Computing environment
All the computing tests were run on a SMP system consisting of two Intel Xeon X5675
6-core processors, operating at a frequency of 3.06 GHz. The system is equipped with
96GB of RAM and 1TB of disk as secondary memory.1 The routines were compiled with
the GNU C Compiler (gcc v4.4.5) and linked to Intel’s MKL multi-threaded library
(v10.3). For double-buffering, our out-of-core routines make use of the AIO library,
one of the standard libraries on UNIX systems. The available multi-core parallelism is
exploited through MKL’s multi-threaded BLAS and OpenMP’s pragma directives.
1
Recall that our algorithms do not require large amounts of available memory: As long as the kinship
matrix fits in memory, they will complete.
Estimation of σ 2 and h2
1D Sequence
2D Sequence
Naive
—
O(mn3 )
O(mtn3 )
CLAK-Chol
chol-based
O(mpn2 )
—
CLAK-Eig
eigen-based
—
O(tmpn)
FaST-LMM
eigen-based
O(mpn2 )
O(tmpn2 )
GWFGLS
eigen-based
O(mpn2 )
O(tmpn2 )
EMMAX
eigen-based
O(mpn2 )
O(tmpn2 )
Table 1: Computational costs for single GLS, as well as 1-dimensional and 2-dimensional
sequences of GLSs. The variables n, m and t denote the sample size, the number of
genetic markers, and the number of traits, respectively. p is the width of the X matrix,
which is determined by the number of fixed covariates used (e.g. age and sex) and the
genetic model assumed.
4.1
Simulated data
A data set for GWAS can be characterized by the number of individuals in the sample
(n), the number of measured and/or imputed SNPs (m) to be tested, the number of
outcomes to be analyzed (t), and the number of covariates (p) to be included in the
model. In current GWAS, a typical scenario consists of a few covariates (for example,
two, such as sex and age), 105 − −107 SNPs, and thousands or tens of thousands individuals; also, only one or a limited number of outcomes are studied. In our experiments,
we assumed that the number of covariates is two (p=2), and we varied the three other
characteristics (n, m, t) of the data set leading to these three scenarios.
(A) The number of SNPs was fixed to m = 10,000,000 and one single outcome (t = 1)
was studied. As sample size, we used 1,000, 5,000, 10,000, 20,000, and 40,000. The
latter test represents a scenario with large number of individuals.
(B) The number of individuals was fixed to n = 10, 000 and one single outcome (t =
1) was studied. The number of markers m to be analyzed was set to 1,000,000,
10,000,000, and 36,000,000. The latter test is a scenario that represents a whole
genome resequencing.
(C) The number of individuals and markers were fixed to n = 1,000 and m = 1,000,000,
respectively. The number of outcomes t studied varied (1, 10, 100, 1,000, 10,000,
and 100,000), corresponding to an Omics analysis.
For testing purposes, we generated artificial data sets which met pre-specified values
of t, m, p, and n.
References
[1] Astle, W. & Balding, D. Population structure and cryptic relatedness in genetic
association studies. Statist. Sci. 24, 451–471 (2009). URL http://dx.doi.org/10.
1214/09-STS307.
[2] Chen, W. & Abecasis, G. Family-based association tests for genomewide association
scans. The American Journal of Human Genetics 81, 913–926 (2007).
[3] Aulchenko, Y., Struchalin, M. & van Duijn, C. Probabel package for genome-wide
association analysis of imputed data. BMC Bioinformatics 11, 134 (2010).
[4] Kang, H. M. et al. Variance component model to account for sample structure
in genome-wide association studies. Nat. Genet. 42, 348–354 (2010). URL http:
//dx.doi.org/10.1038/ng.548.
[5] Zhang, Z. et al. Mixed linear model approach adapted for genome-wide association
studies. Nat. Genet. 42, 355–360 (2010). URL http://dx.doi.org/10.1038/ng.
546.
[6] Lippert, C. et al. Fast linear mixed models for genome-wide association studies. Nat.
Meth. 8, 833–835 (2011). URL http://dx.doi.org/10.1038/nmeth.1681.
[7] Dongarra, J., Croz, J. D., Hammarling, S. & Duff, I. S. A set of level 3 basic linear
algebra subprograms. ACM Trans. Math. Softw. 16, 1–17 (1990).
[8] Anderson, E. et al. LAPACK Users’ Guide (SIAM, Philadelphia, PA, 1999), third
edn.
[9] Aulchenko, Y. S., Ripke, S., Isaacs, A. & van Duijn, C. M. Genabel: an R library
for genome-wide association analysis. Bioinformatics 23, 1294–6 (2007).
| 5cs.CE
|
Published as a conference paper at ICLR 2017
S TRUCTURED ATTENTION N ETWORKS
Yoon Kim∗
Carl Denton∗
Luong Hoang
Alexander M. Rush
{yoonkim@seas,carldenton@college,lhoang@g,srush@seas}.harvard.edu
School of Engineering and Applied Sciences
Harvard University
Cambridge, MA 02138, USA
arXiv:1702.00887v3 [cs.CL] 16 Feb 2017
A BSTRACT
Attention networks have proven to be an effective approach for embedding categorical inference within a deep neural network. However, for many tasks we
may want to model richer structural dependencies without abandoning end-to-end
training. In this work, we experiment with incorporating richer structural distributions, encoded using graphical models, within deep networks. We show that
these structured attention networks are simple extensions of the basic attention
procedure, and that they allow for extending attention beyond the standard softselection approach, such as attending to partial segmentations or to subtrees. We
experiment with two different classes of structured attention networks: a linearchain conditional random field and a graph-based parsing model, and describe
how these models can be practically implemented as neural network layers. Experiments show that this approach is effective for incorporating structural biases,
and structured attention networks outperform baseline attention models on a variety of synthetic and real tasks: tree transduction, neural machine translation,
question answering, and natural language inference. We further find that models trained in this way learn interesting unsupervised hidden representations that
generalize simple attention.
1
I NTRODUCTION
Attention networks are now a standard part of the deep learning toolkit, contributing to impressive
results in neural machine translation (Bahdanau et al., 2015; Luong et al., 2015), image captioning
(Xu et al., 2015), speech recognition (Chorowski et al., 2015; Chan et al., 2015), question answering
(Hermann et al., 2015; Sukhbaatar et al., 2015), and algorithm-learning (Graves et al., 2014; Vinyals
et al., 2015), among many other applications (see Cho et al. (2015) for a comprehensive review).
This approach alleviates the bottleneck of compressing a source into a fixed-dimensional vector by
equipping a model with variable-length memory (Weston et al., 2014; Graves et al., 2014; 2016),
thereby providing random access into the source as needed. Attention is implemented as a hidden
layer which computes a categorical distribution (or hierarchy of categorical distributions) to make a
soft-selection over source elements.
Noting the empirical effectiveness of attention networks, we also observe that the standard attentionbased architecture does not directly model any structural dependencies that may exist among the
source elements, and instead relies completely on the hidden layers of the network. While one might
argue that these structural dependencies can be learned implicitly by a deep model with enough data,
in practice, it may be useful to provide a structural bias. Modeling structural dependencies at the
final, output layer has been shown to be important in many deep learning applications, most notably
in seminal work on graph transformers (LeCun et al., 1998), key work on NLP (Collobert et al.,
2011), and in many other areas (Peng et al., 2009; Do & Artiéres, 2010; Jaderberg et al., 2014; Chen
et al., 2015; Durrett & Klein, 2015; Lample et al., 2016, inter alia).
In this work, we consider applications which may require structural dependencies at the attention
layer, and develop internal structured layers for modeling these directly. This approach generalizes
categorical soft-selection attention layers by specifying possible structural dependencies in a soft
∗
Equal contribution.
1
Published as a conference paper at ICLR 2017
manner. Key applications will be the development of an attention function that segments the source
input into subsequences and one that takes into account the latent recursive structure (i.e. parse tree)
of a source sentence.
Our approach views the attention mechanism as a graphical model over a set of latent variables. The
standard attention network can be seen as an expectation of an annotation function with respect to a
single latent variable whose categorical distribution is parameterized to be a function of the source.
In the general case we can specify a graphical model over multiple latent variables whose edges
encode the desired structure. Computing forward attention requires performing inference to obtain
the expectation of the annotation function, i.e. the context vector. This expectation is computed over
an exponentially-sized set of structures (through the machinery of graphical models/structured prediction), hence the name structured attention network. Notably each step of this process (including
inference) is differentiable, so the model can be trained end-to-end without having to resort to deep
policy gradient methods (Schulman et al., 2015).
The differentiability of inference algorithms over graphical models has previously been noted by
various researchers (Li & Eisner, 2009; Domke, 2011; Stoyanov et al., 2011; Stoyanov & Eisner,
2012; Gormley et al., 2015), primarily outside the area of deep learning. For example, Gormley
et al. (2015) treat an entire graphical model as a differentiable circuit and backpropagate risk through
variational inference (loopy belief propagation) for minimium risk training of dependency parsers.
Our contribution is to combine these ideas to produce structured internal attention layers within
deep networks, noting that these approaches allow us to use the resulting marginals to create new
features, as long as we do so a differentiable way.
We focus on two classes of structured attention: linear-chain conditional random fields (CRFs) (Lafferty et al., 2001) and first-order graph-based dependency parsers (Eisner, 1996). The initial work
of Bahdanau et al. (2015) was particularly interesting in the context of machine translation, as the
model was able to implicitly learn an alignment model as a hidden layer, effectively embedding
inference into a neural network. In similar vein, under our framework the model has the capacity
to learn a segmenter as a hidden layer or a parser as a hidden layer, without ever having to see a
segmented sentence or a parse tree. Our experiments apply this approach to a difficult synthetic reordering task, as well as to machine translation, question answering, and natural language inference.
We find that models trained with structured attention outperform standard attention models. Analysis of learned representations further reveal that interesting structures emerge as an internal layer of
the model. All code is available at http://github.com/harvardnlp/struct-attn.
2
BACKGROUND : ATTENTION N ETWORKS
A standard neural network consist of a series of non-linear transformation layers, where each layer
produces a fixed-dimensional hidden representation. For tasks with large input spaces, this paradigm
makes it hard to control the interaction between components. For example in machine translation,
the source consists of an entire sentence, and the output is a prediction for each word in the translated
sentence. Utilizing a standard network leads to an information bottleneck, where one hidden layer
must encode the entire source sentence. Attention provides an alternative approach.1 An attention
network maintains a set of hidden representations that scale with the size of the source. The model
uses an internal inference step to perform a soft-selection over these representations. This method
allows the model to maintain a variable-length memory and has shown to be crucially important for
scaling systems for many tasks.
Formally, let x = [x1 , . . . , xn ] represent a sequence of inputs, let q be a query, and let z be a
categorical latent variable with sample space {1, . . . , n} that encodes the desired selection among
these inputs. Our aim is to produce a context c based on the sequence and the query. To do so, we
assume access to an attention distribution z ∼ p(z | x, q), where we condition p on the inputs x and
a query q. The context over a sequence is defined as expectation, c = Ez∼p(z | x,q) [f (x, z)] where
f (x, z) is an annotation function. Attention of this form can be applied over any type of input,
however, we will primarily be concerned with “deep” networks, where both the annotation function
1
Another line of work involves marginalizing over latent variables (e.g. latent alignments) for sequence-tosequence transduction (Kong et al., 2016; Lu et al., 2016; Yu et al., 2016; 2017).
2
Published as a conference paper at ICLR 2017
and attention distribution are parameterized with neural networks, and the context produced is a
vector fed to a downstream network.
For example, consider the case of attention-based neural machine translation (Bahdanau et al., 2015).
Here the sequence of inputs [x1 , . . . , xn ] are the hidden states of a recurrent neural network (RNN),
running over the words in the source sentence, q is the RNN hidden state of the target decoder
(i.e. vector representation of the query q), and z represents the source position to be attended to
for translation. The attention distribution p is simply p(z = i | x, q) = softmax(θi ) where θ ∈ Rn
is a parameterized potential typically based on a neural network, e.g. θi = MLP([xi ; q]). The
annotation function is defined to simply return the selected hidden state, f (x, z) = xz . The context
vector can then be computed using a simple sum,
c = Ez∼p(z | x,q) [f (x, z)] =
n
X
p(z = i | x, q)xi
(1)
i=1
Other tasks such as question answering use attention in a similar manner, for instance by replacing
source [x1 , . . . , xn ] with a set of potential facts and q with a representation of the question.
In summary we interpret the attention mechanism as taking the expectation of an annotation function
f (x, z) with respect to a latent variable z ∼ p, where p is parameterized to be function of x and q.
3
S TRUCTURED ATTENTION
Attention networks simulate selection from a set using a soft model. In this work we consider generalizing selection to types of attention, such as selecting chunks, segmenting inputs, or even attending
to latent subtrees. One interpretation of this attention is as using soft-selection that considers all possible structures over the input, of which there may be exponentially many possibilities. Of course,
this expectation can no longer be computed using a simple sum, and we need to incorporate the
machinery of inference directly into our neural network.
Define a structured attention model as being an attention model where z is now a vector of discrete
latent variables [z1 , . . . , zm ] and the attention distribution is p(z | x, q) is defined as a conditional
random field (CRF), specifying the independence structure of the z variables. Formally, we assume
an undirected graph structure with m vertices. The CRF is parameterized with clique (log-)potentials
θC (zC ) ∈ R, where the zC indicates the subset of z given by
P clique C. Under this definition, the
attention probability is defined as, p(z | x, q; θ) = softmax( C θC (zC )), whereP
for symmetry we
use softmax in a general sense, i.e. softmax(g(z)) = Z1 exp(g(z)) where Z = z0 exp(g(z 0 )) is
the implied partition function. In practice we use a neural CRF, where θ comes from a deep model
over x, q.
In structured attention, we also P
assume that the annotation function f factors (at least) into clique
annotation functions f (x, z) = C fC (x, zC ). Under standard conditions on the conditional independence structure, inference techniques from graphical models can be used to compute the forwardpass expectations and the context:
X
c = Ez∼p(z | x,q) [f (x, z)] =
Ez∼p(zC | x,q) [fC (x, zC )]
C
3.1
E XAMPLE 1: S UBSEQUENCE S ELECTION
Suppose instead of soft-selecting a single input, we wanted to explicitly model the selection of contiguous subsequences. We could naively apply categorical attention over all subsequences, or hope
the model learns a multi-modal distribution to combine neighboring words. Structured attention
provides an alternate approach.
Concretely, let m = n, define z to be a random
Pn vector z = [z1 , . . . , zn ] with zi ∈ {0, 1}, and define
our annotation function to be, f (x, z) =
i=1 fi (x, zi ) where fi (x, zi ) = 1{zi = 1}xi . The
explicit expectation is then,
Ez1 ,...,zn [f (x, z)] =
n
X
i=1
3
p(zi = 1 | x, q)xi
(2)
Published as a conference paper at ICLR 2017
q
q
z1
x1
x2
x3
(a)
x4
q
z1
z2
z3
z4
z1
z2
z3
z4
x1
x2
x3
x4
x1
x2
x3
x4
(b)
(c)
Figure 1: Three versions of a latent variable attention model: (a) A standard soft-selection attention network,
(b) A Bernoulli (sigmoid) attention network, (c) A linear-chain structured attention model for segmentation.
The input and query are denoted with x and q respectively.
Equation (2) is similar to equation (1)—both are a linear combination of the input representations
where the scalar is between [0, 1] and represents how much attention should be focused on each
input. However, (2) is fundamentally different in two ways: (i) it allows for multiple inputs (or no
inputs) to be selected for a given query; (ii) we can incorporate structural dependencies across the
zi ’s. For instance, we can model the distribution over z with a linear-chain CRF with pairwise edges,
!
n−1
X
p(z1 , . . . , zn | x, q) = softmax
θi,i+1 (zi , zi+1 )
(3)
i=1
where θk,l is the pairwise potential for zi = k and zi+1 = l. This model is shown in Figure 1c.
Compare this model to the standard attention in Figure 1a, or to a simple Bernoulli (sigmoid) selection method, p(zi = 1 | x, q) = sigmoid(θi ), shown in Figure 1b. All three of these methods can
use potentials from the same neural network or RNN that takes x and q as inputs.
In the case of the linear-chain CRF in (3), the marginal distribution p(zi = 1 | x) can be calculated
efficiently in linear-time for all i using message-passing, i.e. the forward-backward algorithm. These
marginals allow us to calculate (2), and in doing so we implicitly sum over an exponentially-sized
set of structures (i.e. all binary sequences of length n) through dynamic programming. We refer to
this type of attention layer as a segmentation attention layer.
Note that the forward-backward algorithm is being used as parameterized pooling (as opposed to
output computation), and can be thought of as generalizing the standard attention softmax. Crucially
this generalization from vector softmax to forward-backward is just a series of differentiable steps,2
and we can compute gradients of its output (marginals) with respect to its input (potentials). This
will allow the structured attention model to be trained end-to-end as part of a deep model.
3.2
E XAMPLE 2: S YNTACTIC T REE S ELECTION
This same approach can be used for more involved structural dependencies. One popular structure
for natural language tasks is a dependency tree, which enforces a structural bias on the recursive
dependencies common in many languages. In particular a dependency tree enforces that each word
in a source sentence is assigned exactly one parent word (head word), and that these assignments do
not cross (projective structure). Employing this bias encourages the system to make a soft-selection
based on learned syntactic dependencies, without requiring linguistic annotations or a pipelined
decision.
A dependency parser can be partially formalized as a graphical model with the following cliques
(Smith & Eisner, 2008): latent variables zij ∈ {0, 1} for all i 6= j, which indicates that the i-th
word is the parent of the j-th word (i.e. xi → xj ); and a special global constraint that rules out
configurations of zij ’s that violate parsing constraints (e.g. one head, projectivity).
The parameters to the graph-based CRF dependency parser are the potentials θij , which reflect
the score of selecting xi as the parent of xj . The probability of a parse tree z given the sentence
2
As are other dynamic programming algorithms for inference in graphical models, such as (loopy and nonloopy) belief propagation.
4
Published as a conference paper at ICLR 2017
procedure F ORWARD BACKWARD(θ)
α[0, hti] ← 0
β[n + 1, hti] ← 0
for i = 1, . . . ,L
n; c ∈ C do
α[i, c] ← y α[i − 1, y] ⊗ θi−1,i (y, c)
for i = n, . . . L
, 1; c ∈ C do
β[i, c] ← y β[i + 1, y] ⊗ θi,i+1 (c, y)
procedure BACKPROP F ORWARD BACKWARD(θ, p, ∇L
p)
L
∇L
α ← log p ⊗ log ∇p ⊗ β ⊗ −A
L
∇L
β ← log p ⊗ log ∇p ⊗ α ⊗ −A
α̂[0, hti] ← 0
β̂[n + 1, hti] ← 0
for i = n, . . . 1; c ∈ C do
L
β̂[i, c] ← ∇L
α [i, c] ⊕
y θi,i+1 (c, y) ⊗ β̂[i + 1, y]
A ← α[n + 1, hti]
for i = 1, . . . , n; c ∈ C do
p(zi = c | x) ← exp(α[i, c] ⊗ β[i, c]
⊗ − A)
return p
for i = 1, . . . , n; c ∈ C do
L
α̂[i, c] ← ∇L
β [i, c] ⊕
y θi−1,i (y, c) ⊗ α̂[i − 1, y]
for i = 1, . . . , n; y, c ∈ C do
∇L
θi−1,i (y,c) ← signexp(α̂[i, y] ⊗ β[i + 1, c]
⊕ α[i, y] ⊗ β̂[i + 1, c]
⊕ α[i, y] ⊗ β[i + 1, c] ⊗ −A)
return ∇L
θ
Figure 2: Algorithms for linear-chain CRF: (left) computation of forward-backward tables α, β, and marginal
probabilities p from potentials θ (forward-backward algorithm); (right) backpropagation of loss gradients with
respect to the marginals ∇L
p . C denotes the state space and hti is the special start/stop state. Backpropagation
log p
L
L
uses the identity ∇L
=
p
∇L
, where is the element-wise multiplication.
p to calculate ∇θ = ∇log p ∇θ
log p
Typically the forward-backward with marginals is performed in the log-space semifield R ∪ {±∞} with binary
operations ⊕ = logadd and ⊗ = + for numerical precision. However, backpropagation requires working with
the log of negative values (since ∇L
p could be negative), so we extend to a field [R ∪ {±∞}] × {+, −} with
special +/− log-space operations. Binary operations applied to vectors are implied to be element-wise. The
signexp function is defined as signexp(la ) = sa exp(la ). See Section 3.3 and Table 1 for more details.
x = [x1 , . . . , xn ] is,
p(z | x, q) = softmax 1{z is valid}
X
1{zij = 1}θij
(4)
i6=j
where z is represented as a vector of zij ’s for all i 6= j. It is possible to calculate the marginal
probability of each edge p(zij = 1 | x, q) for all i, j in O(n3 ) time using the inside-outside algorithm
(Baker, 1979) on the data structures of Eisner (1996).
Pn
The parsing contraints ensure that each word has exactly one head (i.e. i=1 zij = 1). Therefore if
we want to utilize the soft-head selection of a position j, the context vector is defined as:
fj (x, z) =
n
X
1{zij = 1}xi
cj = Ez [fj (x, z)] =
i=1
n
X
p(zij = 1 | x, q)xi
i=1
Note that in this case the annotation function has the subscript j to produce a context vector for
each word in the sentence. Similar types of attention can be applied for other tree properties (e.g.
soft-children). We refer to this type of attention layer as a syntactic attention layer.
3.3
E ND - TO -E ND T RAINING
Graphical models of this form have been widely used as the final layer of deep models. Our contribution is to argue that these networks can be added within deep networks in place of simple attention
layers. The whole model can then be trained end-to-end.
The main complication in utilizing this approach within the network itself is the need to backpropagate the gradients through an inference algorithm as part of the structured attention network. Past
work has demonstrated the techniques necessary for this approach (see Stoyanov et al. (2011)), but
to our knowledge it is very rarely employed.
Consider the case of the simple linear-chain CRF layer from equation (3). Figure 2 (left) shows the
standard forward-backward algorithm for computing the marginals p(zi = 1 | x, q; θ). If we treat the
forward-backward algorithm as a neural network layer, its input are the potentials θ, and its output
5
Published as a conference paper at ICLR 2017
after the forward pass are these marginals.3 To backpropagate a loss through this layer we need to
compute the gradient of the loss L with respect to θ, ∇L
θ , as a function of the gradient of the loss with
4
respect to the marginals, ∇L
.
As
the
forward-backward
algorithm consists of differentiable steps,
p
this function can be derived using reverse-mode automatic differentiation of the forward-backward
algorithm itself. Note that this reverse-mode algorithm conveniently has a parallel structure to the
forward version, and can also be implemented using dynamic programming.
However, in practice, one cannot simply
use current off-the-shelf tools for this task.
⊕
⊗
For one, efficiency is quite important for
la+b
sa+b
la·b
sa·b
sa sb
these models and so the benefits of hand+ + la + log(1 + d)
+
la + lb
+
optimizing the reverse-mode implementa+ − la + log(1 − d)
+
la + lb
−
tion still outweighs simplicity of automatic
− + la + log(1 − d)
−
la + lb
−
differentiation. Secondly, numerical pre− − la + log(1 + d)
−
la + lb
+
cision becomes a major issue for structured attention networks. For computing
Table 1: Signed log-space semifield (from Li & Eisthe forward-pass and the marginals, it is im- ner (2009)). Each real number a is represented as a pair
portant to use the standard log-space semi- (la , sa ) where la = log |a| and sa = sign(a). Therefore
field over R ∪ {±∞} with binary opera- a = sa exp(la ). For the above we let d = exp(lb − la ) and
tions (⊕ = logadd, ⊗ = +) to avoid un- assume |a| > |b|.
derflow of probabilities. For computing the
backward-pass, we need to remain in logspace, but also handle log of negative values (since ∇L
p could be negative). This requires extending
to the signed log-space semifield over [R ∪ {±∞}] × {+, −} with special +/− operations. Table 1, based on Li & Eisner (2009), demonstrates how to handle this issue, and Figure 2 (right)
describes backpropagation through the forward-backward algorithm. For dependency parsing, the
forward pass can be computed using the inside-outside implementation of Eisner’s algorithm (Eisner, 1996). Similarly, the backpropagation parallels the inside-outside structure. Forward/backward
pass through the inside-outside algorithm is described in Appendix B.
4
E XPERIMENTS
We experiment with three instantiations of structured attention networks on four different tasks: (a)
a simple, synthetic tree manipulation task using the syntactic attention layer, (b) machine translation
with segmentation attention (i.e. two-state linear-chain CRF), (c) question answering using an nstate linear-chain CRF for multi-step inference over n facts, and (d) natural language inference with
syntactic tree attention. These experiments are not intended to boost the state-of-the-art for these
tasks but to test whether these methods can be trained effectively in an end-to-end fashion, can yield
improvements over standard selection-based attention, and can learn plausible latent structures. All
model architectures, hyperparameters, and training details are further described in Appendix A.
4.1
T REE T RANSDUCTION
The first set of experiments look at a tree-transduction task. These experiments use synthetic data
to explore a failure case of soft-selection attention models. The task is to learn to convert a random
formula given in prefix notation to one in infix notation, e.g.,
( ∗ ( + ( + 15 7 ) 1 8 ) ( + 19 0 11 ) ) ⇒ ( ( 15 + 7 ) + 1 + 8 ) ∗ ( 19 + 0 + 11 )
The alphabet consists of symbols {(, ), +, ∗}, numbers between 0 and 20, and a special root symbol
$. This task is used as a preliminary task to see if the model is able to learn the implicit tree structure
on the source side. The model itself is an encoder-decoder model, where the encoder is defined
below and the decoder is an LSTM. See Appendix A.2 for the full model.
3
Confusingly, “forward” in this case is different than in the forward-backward algorithm, as the marginals
themselves are the output. However the two uses of the term are actually quite related. The forward-backward
algorithm can be interpreted as a forward and backpropagation pass on the log partition function. See Eisner
(2016) for further details (appropriately titled “Inside-Outside and Forward-Backward Algorithms Are Just
Backprop”). As such our full approach can be seen as computing second-order information. This interpretation
is central to Li & Eisner (2009).
4
In general we use ∇ab to denote the Jacobian of a with respect to b.
6
Published as a conference paper at ICLR 2017
Figure 3: Visualization of the source self-attention distribution for the simple (left) and structured (right)
attention models on the tree transduction task. $ is the special root symbol. Each row delineates the distribution
over the parents (i.e. each row sums to one). The attention distribution obtained from the parsing marginals are
more able to capture the tree structure—e.g. the attention weights of closing parentheses are generally placed
on the opening parentheses (though not necessarily on a single parenthesis).
Training uses 15K prefix-infix pairs where the maximum nesting depth is set to be between 2-4 (the
above example has depth 3), with 5K pairs in each depth bucket. The number of expressions in each
parenthesis is limited to be at most 4. Test uses 1K unseen sequences with depth between 2-6 (note
specifically deeper than train), with 200 sequences for each depth. The performance is measured
as the average proportion of correct target tokens produced until the first failure (as in Grefenstette
et al. (2015)).
For experiments we try using different forms of self -attention over embedding-only encoders. Let
xj be an embedding for each source symbol; our three variants of the source representation x̂j are:
(a) no atten, just symbol embeddings by themselves, i.e. x̂j = xj ; (b) simple
Pn attention, symbol
embeddings and soft-pairing for each symbol, i.e. x̂j = [xj ; cj ] where cj = i=1 softmax(θij )xi
is calculated using soft-selection;
Pn (c) structured attention, symbol embeddings and soft-parent, i.e.
x̂j = [xj ; cj ] where cj = i=1 p(zij = 1 | x)xi is calculated using parsing marginals, obtained
from the syntactic attention layer. None of these models use an explicit query value—the potentials
come from running a bidirectional LSTM over the source, producing hidden vectors hi , and then
computing
θij = tanh(s> tanh(W1 hi + W2 hj + b))
where s, b, W1 , W2 are parameters (see Appendix A.1).
Depth
No Atten
Simple
Structured
2
3
4
5
6
7.6
4.1
2.8
2.1
1.5
87.4
49.6
23.3
15.0
8.5
99.2
87.0
64.5
30.8
18.2
The source representation [x̂1 , . . . , x̂n ] are attended
over using the standard attention mechanism at each
decoding step by an LSTM decoder.5 Additionally,
symbol embedding parameters are shared between the
parsing LSTM and the source encoder.
Results Table 2 has the results for the task. Note
that this task is fairly difficult as the encoder is quite
Table 2: Performance (average length to fail- simple. The baseline model (unsurprisingly) performs
ure %) of models on the tree-transduction task. poorly as it has no information about the source ordering. The simple attention model performs better, but
is significantly outperformed by the structured model
with a tree structure bias. We hypothesize that the
model is partially reconstructing the arithmetic tree. Figure 3 shows the attention distribution for the
simple/structured models on the same source sequence, which indicates that the structured model is
able to learn boundaries (i.e. parentheses).
5
Thus there are two attention mechanisms at work under this setup. First, structured attention over the
source only to obtain soft-parents for each symbol (i.e. self-attention). Second, standard softmax alignment
attention over the source representations during decoding.
7
Published as a conference paper at ICLR 2017
4.2
N EURAL M ACHINE T RANSLATION
Our second set of experiments use a full neural machine translation model utilizing attention over
subsequences. Here both the encoder/decoder are LSTMs, and we replace standard simple attention
with a segmentation attention layer. We experiment with two settings: translating directly from
unsegmented Japanese characters to English words (effectively using structured attention to perform
soft word segmentation), and translating from segmented Japanese words to English words (which
can be interpreted as doing phrase-based neural machine translation). Japanese word segmentation
is done using the KyTea toolkit (Neubig et al., 2011).
The data comes from the Workshop on Asian Translation (WAT) (Nakazawa et al., 2016). We
randomly pick 500K sentences from the original training set (of 3M sentences) where the Japanese
sentence was at most 50 characters and the English sentence was at most 50 words. We apply the
same length filter on the provided validation/test sets for evaluation. The vocabulary consists of all
tokens that occurred at least 10 times in the training corpus.
The segmentation attention layer is a two-state CRF where the unary potentials at the j-th decoder
step are parameterized as
hi Whj , k = 1
θi (k) =
0,
k=0
Here [h1 , . . . , hn ] are the encoder hidden states and h0j is the j-th decoder hidden state (i.e. the
query vector). The pairwise potentials are parameterized linearly with b, i.e. all together
θi,i+1 (zi , zi+1 ) = θi (zi ) + θi+1 (zi+1 ) + bzi ,zi+1
Therefore the segmentation attention layer requires just 4 additional parameters. Appendix A.3
describes the full model architecture.
We
Pn experiment with three attention configurations: (a) standard simple attention, i.e. cj =
i )hi ; (b) sigmoid attention: multiple selection with Bernoulli random variables,
i=1 softmax(θ
Pn
i.e. cj = i=1 sigmoid(θi )hi ; (c) structured attention, encoded with normalized CRF marginals,
cj =
n
X
p(zi = 1 | x, q)
hi
γ
i=1
n
γ=
1X
p(zi = 1 | x, q)
λ i=1
The normalization term γ is not ideal but we found it to be helpful for stable training.6 λ is a
hyperparameter (we use λ = 2) and we further add an l2 penalty of 0.005 on the pairwise potentials
b. These values were found via grid search on the validation set.
Results Results for the translation task on the test
set are given in Table 3. Sigmoid attention outperSimple Sigmoid Structured
forms simple (softmax) attention on the character-toC HAR
12.6
13.1
14.6
word task, potentially because it is able to learn manyW ORD
14.1
13.8
14.3
to-one alignments. On the word-to-word task, the opposite is true, with simple attention outperforming sigTable 3: Translation performance as mea- moid attention. Structured attention outperforms both
sured by BLEU (higher is better) on character- models on both tasks, although improvements on the
to-word and word-to-word Japanese-English word-to-word task are modest and unlikely to be statranslation for the three different models.
tistically significant.
For further analysis, Figure 4 shows a visualization of
the different attention mechanisms on the character-to-word setup. The simple model generally
focuses attention heavily on a single character. In contrast, the sigmoid and structured models are
able to spread their attention distribution on contiguous subsequences. The structured attention
learns additional parameters (i.e. b) to smooth out this type of attention.
Pn
With standard expectation (i.e. cj =
i=1 p(zi = 1 | x, q)hi ) we empirically observed the marginals
to quickly saturate. We tried various strategies to overcome this, such as putting an l2 penalty on the unary
potentials and initializing with a pretrained sigmoid attention model, but simply normalizing the marginals
proved to be the most effective. However, this changes the interpretation of the context vector as the expectation
of an annotation function in this case.
6
8
Published as a conference paper at ICLR 2017
Figure 4: Visualization of the source attention distribution for the simple (top left), sigmoid (top right), and
structured (bottom left) attention models over the ground truth sentence on the character-to-word translation
task. Manually-annotated alignments are shown in bottom right. Each row delineates the attention weights
over the source sentence at each step of decoding. The sigmoid/structured attention models are able learn an
implicit segmentation model and focus on multiple characters at each time step.
4.3
Q UESTION A NSWERING
Our third experiment is on question answering (QA) with the linear-chain CRF attention layer for
inference over multiple facts. We use the bAbI dataset (Weston et al., 2015), where the input is a set
of sentences/facts paired with a question, and the answer is a single token. For many of the tasks
the model has to attend to multiple supporting facts to arrive at the correct answer (see Figure 5 for
an example), and existing approaches use multiple ‘hops’ to greedily attend to different facts. We
experiment with employing structured attention to perform inference in a non-greedy way. As the
ground truth supporting facts are given in the dataset, we are able to assess the model’s inference
accuracy.
The baseline (simple) attention model is the End-To-End Memory Network (Sukhbaatar et al.,
2015) (MemN2N), which we briefly describe here. See Appendix A.4 for full model details. Let
x1 , . . . , xn be the input embedding vectors for the n sentences/facts and let q be the query embedding. In MemN2N, zk is the random variable for the sentence to select at the k-th inference step
(i.e. k-th hop), and thus zk ∈ {1, . . . , n}. The probability distribution over
Pnzk is given by p(zk =
i | x, q) = softmax((xki )> qk ), and the context vector is given by ck = i=1 p(zk = i | x, q)oki ,
where xki , oki are the input and output embedding for the i-th sentence at the k-th hop, respectively.
The k-th context vector is used to modify the query qk+1 = qk + ck , and this process repeats for
k = 1, . . . , K (for k = 1 we have xki = xi , qk = q, ck = 0). The K-th context and query vectors
are used to obtain the final answer. The attention mechanism for a K-hop MemN2N network can
therefore be interpreted as a greedy selection of a length-K sequence of facts (i.e. z1 , . . . , zK ).
For structured attention, we use an n-state, K-step linear-chain CRF.7 We experiment with two
different settings: (a) a unary CRF model with node potentials
θk (i) = (xki )> qk
7
Note that this differs from the segmentation attention for the neural machine translation experiments described above, which was a K-state (with K = 2), n-step linear-chain CRF.
9
Published as a conference paper at ICLR 2017
MemN2N
Ans % Fact %
Binary CRF
Ans % Fact %
Unary CRF
Ans % Fact %
Task
K
TASK 02 - T WO S UPPORTING FACTS
TASK 03 - T HREE S UPPORTING FACTS
TASK 07 - C OUNTING
TASK 08 - L ISTS S ETS
TASK 11 - I NDEFINITE K NOWLEDGE
TASK 13 - C OMPOUND C OREFERENCE
TASK 14 - T IME R EASONING
TASK 15 - BASIC D EDUCTION
TASK 16 - BASIC I NDUCTION
TASK 17 - P OSITIONAL R EASONING
TASK 18 - S IZE R EASONING
TASK 19 - PATH F INDING
2
3
3
3
2
2
2
2
3
2
2
2
87.3
52.6
83.2
94.1
97.8
95.6
99.9
100.0
97.1
61.1
86.4
21.3
46.8
1.4
−
−
38.2
14.8
77.6
59.3
91.0
23.9
3.3
10.2
84.7
40.5
83.5
93.3
97.7
97.0
99.7
100.0
97.9
60.6
92.2
24.4
81.8
0.1
−
−
80.8
36.4
98.2
89.5
85.6
49.6
3.9
11.5
43.5
28.2
79.3
87.1
88.6
94.4
90.5
100.0
98.0
59.7
92.0
24.3
22.3
0.0
−
−
0.0
9.3
30.2
51.4
41.4
10.5
1.4
7.8
AVERAGE
−
81.4
39.6
81.0
53.7
73.8
17.4
Table 4: Answer accuracy (Ans %) and supporting fact selection accuracy (Fact %) of the three QA models
on the 1K bAbI dataset. K indicates the number of hops/inference steps used for each task. Task 7 and 8 both
contain variable number of facts and hence they are excluded from the fact accuracy measurement. Supporting
fact selection accuracy is calculated by taking the average of 10 best runs (out of 20) for each task.
and (b) a binary CRF model with pairwise potentials
θk,k+1 (i, j) = (xki )> qk + (xki )> xk+1
+ (xk+1
)> qk+1
j
j
The binary CRF model is designed to test the model’s ability P
to perform sequential reasoning. For
both (a) and (b), a single context vector is computed: c =
z1 ,...,zK p(z1 , . . . , zK | x, q)f (x, z)
(unlike MemN2N which computes K context vectors). Evaluating c requires summing over all nK
possible sequences of length K, which may not be practical for large values of K. However, if
PK
f (x, z) factors over the components of z (e.g. f (x, z) = k=1 fk (x, zk )) then one can rewrite the
PK Pn
above sum in terms of marginals: c = k=1 i=1 p(zk = i | x, q)fk (x, zk ). In our experiments,
we use fk (x, zk ) = okzk . All three models are described in further detail in Appendix A.4.
Results We use the version of the dataset with 1K questions for each task. Since all models reduce
to the same network for tasks with 1 supporting fact, they are excluded from our experiments. The
number of hops (i.e. K) is task-dependent, and the number of memories (i.e. n) is limited to be
at most 25 (note that many question have less than 25 facts—e.g. the example in Figure 5 has 9
facts). Due to high variance in model performance, we train 20 models with different initializations
for each task and report the test accuracy of the model that performed the best on a 10% held-out
validation set (as is typically done for bAbI tasks).
Results of the three different models are shown in Table 4. For correct answer seletion (Ans %),
we find that MemN2N and the Binary CRF model perform similarly while the Unary CRF model
does worse, indicating the importance of including pairwise potentials. We also assess each model’s
ability to attend to the correct supporting facts in Table 4 (Fact %). Since ground truth supporting
facts are provided for each query, we can check the sequence accuracy of supporting facts for each
model (i.e. the rate of selecting the exact correct sequence of facts) by taking the highest probability
sequence ẑ = argmax p(z1 , . . . , zK | x, q) from the model and checking against the ground truth.
Overall the Binary CRF is able to recover supporting facts better than MemN2N. This improvement
is significant and can be up to two-fold as seen for task 2, 11, 13 & 17. However we observed that
on many tasks it is sufficient to select only the last (or first) fact correctly to predict the answer,
and thus higher sequence selection accuracy does not necessarily imply better answer accuracy (and
vice versa). For example, all three models get 100% answer accuracy on task 15 but have different
supporting fact accuracies.
Finally, in Figure 5 we visualize of the output edge marginals produced by the Binary CRF model
for a single question in task 16. In this instance, the model is uncertain but ultimately able to select
the right sequence of facts 5 → 6 → 8.
10
Published as a conference paper at ICLR 2017
Figure 5: Visualization of the attention distribution over supporting fact sequences for an example question
in task 16 for the Binary CRF model. The actual question is displayed at the bottom along with the correct
answer and the ground truth supporting facts (5 → 6 → 8). The edges represent the marginal probabilities
p(zk , zk+1 | x, q), and the nodes represent the n supporting facts (here we have n = 9). The text for the
supporting facts are shown on the left. The top three most likely sequences are: p(z1 = 5, z2 = 6, z3 =
8 | x, q) = 0.0564, p(z1 = 5, z2 = 6, z3 = 3 | x, q) = 0.0364, p(z1 = 5, z2 = 2, z3 = 3 | x, q) = 0.0356.
4.4
NATURAL L ANGUAGE I NFERENCE
The final experiment looks at the task of natural language inference (NLI) with the syntactic attention layer. In NLI, the model is given two sentences (hypothesis/premise) and has to predict their
relationship: entailment, contradiction, neutral.
For this task, we use the Stanford NLI dataset (Bowman et al., 2015) and model our approach off
of the decomposable attention model of Parikh et al. (2016). This model takes in the matrix of
word embeddings as the input for each sentence and performs inter-sentence attention to predict the
answer. Appendix A.5 describes the full model.
As in the transduction task, we focus on modifying the input representation to take into account soft
parents via self-attention (i.e. intra-sentence attention). In addition to the three baselines described
for tree transduction (No Attention, Simple, Structured), we also explore two additional settings: (d)
hard pipeline parent selection, i.e. x̂j = [xj ; xhead(j) ], where head(j) is the index of xj ’s parent8 ;
(e) pretrained structured attention: structured attention where the parsing layer is pretrained for one
epoch on a parsed dataset (which was enough for convergence).
Results Results of our models are shown in Table 5. Simple attention improves upon the no
attention model, and this is consistent with improvements observed by Parikh et al. (2016) with
their intra-sentence attention model. The pipelined model with hard parents also slightly improves
upon the baseline. Structured attention outperforms both models, though surprisingly, pretraining
the syntactic attention layer on the parse trees performs worse than training it from scratch—it is
possible that the pretrained attention is too strict for this task.
We also obtain the hard parse for an example sentence by running the Viterbi algorithm on the
syntactic attention layer with the non-pretrained model:
$
The
men
are
fighting
outside
a
deli
.
8
The parents are obtained from running the dependency parser of Andor et al. (2016), available at
https://github.com/tensorflow/models/tree/master/syntaxnet
11
Published as a conference paper at ICLR 2017
Accuracy %
Model
Handcrafted features (Bowman et al., 2015)
LSTM encoders (Bowman et al., 2015)
Tree-Based CNN (Mou et al., 2016)
Stack-Augmented Parser-Interpreter Neural Net (Bowman et al., 2016)
LSTM with word-by-word attention (Rocktäschel et al., 2016)
Matching LSTMs (Wang & Jiang, 2016)
Decomposable attention over word embeddings (Parikh et al., 2016)
Decomposable attention + intra-sentence attention (Parikh et al., 2016)
Attention over constituency tree nodes (Zhao et al., 2016)
Neural Tree Indexers (Munkhdalai & Yu, 2016)
Enhanced BiLSTM Inference Model (Chen et al., 2016)
Enhanced BiLSTM Inference Model + ensemble (Chen et al., 2016)
78.2
80.6
82.1
83.2
83.5
86.1
86.3
86.8
87.2
87.3
87.7
88.3
No Attention
No Attention + Hard parent
Simple Attention
Structured Attention
Pretrained Structured Attention
85.8
86.1
86.2
86.8
86.5
Table 5: Results of our models (bottom) and others (top) on the Stanford NLI test set. Our baseline model has
the same architecture as Parikh et al. (2016) but the performance is slightly different due to different settings
(e.g. we train for 100 epochs with a batch size of 32 while Parikh et al. (2016) train for 400 epochs with a batch
size of 4 using asynchronous SGD.)
Despite being trained without ever being exposed to an explicit parse tree, the syntactic attention
layer learns an almost plausible dependency structure. In the above example it is able to correctly
identify the main verb fighting, but makes mistakes on determiners (e.g. head of The should be
men). We generally observed this pattern across sentences, possibly because the verb structure is
more important for the inference task.
5
C ONCLUSION
This work outlines structured attention networks, which incorporate graphical models to generalize
simple attention, and describes the technical machinery and computational techniques for backpropagating through models of this form. We implement two classes of structured attention layers: a
linear-chain CRF (for neural machine translation and question answering) and a more complicated
first-order dependency parser (for tree transduction and natural language inference). Experiments
show that this method can learn interesting structural properties and improve on top of standard models. Structured attention could also be a way of learning latent labelers or parsers through attention
on other tasks.
It should be noted that the additional complexity in computing the attention distribution increases
run-time—for example, structured attention was approximately 5× slower to train than simple attention for the neural machine translation experiments, even though both attention layers have the
same asymptotic run-time (i.e. O(n)).
Embedding differentiable inference (and more generally, differentiable algorithms) into deep models is an exciting area of research. While we have focused on models that admit (tractable) exact
inference, similar technique can be used to embed approximate inference methods. Many optimization algorithms (e.g. gradient descent, LBFGS) are also differentiable (Domke, 2012; Maclaurin
et al., 2015), and have been used as output layers for structured prediction in energy-based models
(Belanger & McCallum, 2016; Wang et al., 2016). Incorporating them as internal neural network
layers is an interesting avenue for future work.
ACKNOWLEDGMENTS
We thank Tao Lei, Ankur Parikh, Tim Vieira, Matt Gormley, André Martins, Jason Eisner, Yoav
Goldberg, and the anonymous reviewers for helpful comments, discussion, notes, and code. We
additionally thank Yasumasa Miyamoto for verifying Japanese-English translations.
12
Published as a conference paper at ICLR 2017
R EFERENCES
Daniel Andor, Chris Alberti, David Weiss, Aliaksei Severyn, Alessandro Presta, Kuzman Ganchev,
Slav Petrov, and Michael Collins. Globally Normalized Transition-Based Neural Networks. In
Proceedings of ACL, 2016.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural Machine Translation by Jointly
Learning to Align and Translate. In Proceedings of ICLR, 2015.
James K. Baker. Trainable Grammars for Speech Recognition. Speech Communication Papers for
the 97th Meeting of the Acoustical Society, 1979.
David Belanger and Andrew McCallum. Structured Prediction Energy Networks. In Proceedings of
ICML, 2016.
Samuel R. Bowman, Christopher D. Manning, and Christopher Potts. Tree-Structured Composition
in Neural Networks without Tree-Structured Architectures. In Proceedings of the NIPS workshop
on Cognitive Computation: Integrating Neural and Symbolic Approaches, 2015.
Samuel R. Bowman, Jon Gauthier, Abhinav Rastogi, Raghav Gupta, Christopher D. Manning, and
Christopher Potts. A Fast Unified Model for Parsing and Sentence Understanding. In Proceedings
of ACL, 2016.
William Chan, Navdeep Jaitly, Quoc Le, and Oriol Vinyals.
arXiv:1508.01211, 2015.
Listen, Attend and Spell.
Liang-Chieh Chen, Alexander G. Schwing, Alan L. Yuille, and Raquel Urtasun. Learning Deep
Structured Models. In Proceedings of ICML, 2015.
Qian Chen, Xiaodan Zhu, Zhenhua Ling, Si Wei, and Hui Jiang. Enhancing and Combining Sequential and Tree LSTM for Natural Language Inference. arXiv:1609.06038, 2016.
Kyunghyun Cho, Aaron Courville, and Yoshua Bengio. Describing Multimedia Content using
Attention-based Encoder-Decoder Networks. In IEEE Transactions on Multimedia, 2015.
Jan Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, and Yoshua Bengio.
Attention-Based Models for Speech Recognition. In Proceedings of NIPS, 2015.
Ronan Collobert, Jason Weston, Leon Bottou, Michael Karlen, Koray Kavukcuoglu, and Pavel
Kuksa. Natural Language Processing (almost) from Scratch. Journal of Machine Learning Research, 12:2493–2537, 2011.
Trinh-Minh-Tri Do and Thierry Artiéres. Neural Conditional Random Fields. In Proceedings of
AISTATS, 2010.
Justin Domke. Parameter Learning with Truncated Message-Passing. In Proceedings of CVPR,
2011.
Justin Domke. Generic methods for optimization-based modeling. In AISTATS, pp. 318–326, 2012.
John Duchi, Elad Hazan, and Yoram Singer. Adaptive Subgradient Methods for Online Learning
and Stochastic Optimization. Journal of Machine Learning Research, 12:2021–2159, 2011.
Greg Durrett and Dan Klein. Neural CRF Parsing. In Proceedings of ACL, 2015.
Jason M. Eisner. Three New Probabilistic Models for Dependency Parsing: An Exploration. In
Proceedings of ACL, 1996.
Jason M. Eisner. Inside-Outside and Forward-Backward Algorithms are just Backprop. In Proceedings of Structured Prediction Workshop at EMNLP, 2016.
Matthew R. Gormley, Mark Dredze, and Jason Eisner. Approximation-Aware Dependency Parsing
by Belief Propagation. In Proceedings of TACL, 2015.
Alex Graves, Greg Wayne, and Ivo Danihelka. Neural Turing Machines. arXiv:1410.5401, 2014.
13
Published as a conference paper at ICLR 2017
Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwinska, Sergio Gomez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou,
Adria Puigdomenech Badia, Karl Moritz Hermann, Yori Zwols, Georg Ostrovski, Adam Cain,
Helen King, Christopher Summerfield, Phil Blunsom, Koray Kavukcuoglu, and Demis Hassabis.
Hybrid Computing Using a Neural Network with Dynamic External Memory. Nature, October
2016.
Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to
Transduce with Unbounded Memory. In Proceedings of NIPS, 2015.
Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa
Suleyman, and Phil Blunsom. Teaching Machines to Read and Comprehend. In Proceedings of
NIPS, 2015.
Max Jaderberg, Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep Structured Output
Learning for Unconstrained Text Recognition. In Proceedings of ICLR, 2014.
Diederik Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In Proceedings of
ICLR, 2015.
Eliyahu Kipperwasser and Yoav Goldberg. Simple and Accurate Dependency Parsing using Bidirectional LSTM Feature Representations. In TACL, 2016.
Lingpeng Kong, Chris Dyer, and Noah A. Smith. Segmental Recurrent Neural Networks. In Proceedings of ICLR, 2016.
John Lafferty, Andrew McCallum, and Fernando Pereira. Conditional Random Fields: Probabilistic
Models for Segmenting and Labeling Sequence Data. In Proceedings of ICML, 2001.
Guillaume Lample, Miguel Ballesteros, Sandeep Subramanian, Kazuya Kawakami, and Chris Dyer.
Neural Architectures for Named Entity Recognition. In Proceedings of NAACL, 2016.
Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based Learning Applied
to Document Recognition. In Proceedings of IEEE, 1998.
Zhifei Li and Jason Eisner. First- and Second-Order Expectation Semirings with Applications to
Minimum-Risk Training on Translation Forests. In Proceedings of EMNLP 2009, 2009.
Liang Lu, Lingpeng Kong, Chris Dyer, Noah A. Smith, and Steve Renals. Segmental Recurrent
Neural Networks for End-to-End Speech Recognition. In Proceedings of INTERSPEECH, 2016.
Minh-Thang Luong, Hieu Pham, and Christopher D. Manning. Effective Approaches to Attentionbased Neural Machine Translation. In Proceedings of EMNLP, 2015.
Dougal Maclaurin, David Duvenaud, and Ryan P. Adams. Gradient-based Hyperparameter Optimization through Reversible Learning. In Proceedings of ICML, 2015.
Lili Mou, Rui Men, Ge Li, Yan Xu, Lu Zhang, Rui Yan, and Zhi Jin. Natural language inference by
tree-based convolution and heuristic matching. In Proceedings of ACL, 2016.
Tsendsuren Munkhdalai and Hong Yu.
arxiv:1607.04492, 2016.
Neural Tree Indexers for Text Understanding.
Toshiaki Nakazawa, Manabu Yaguchi, Kiyotaka Uchimoto, Masao Utiyama, Eiichiro Sumita, Sadao
Kurohashi, and Hitoshi Isahara. Aspec: Asian scientific paper excerpt corpus. In Nicoletta Calzolari (Conference Chair), Khalid Choukri, Thierry Declerck, Marko Grobelnik, Bente Maegaard,
Joseph Mariani, Asuncion Moreno, Jan Odijk, and Stelios Piperidis (eds.), Proceedings of the
Ninth International Conference on Language Resources and Evaluation (LREC 2016), pp. 2204–
2208, Portoro, Slovenia, may 2016. European Language Resources Association (ELRA). ISBN
978-2-9517408-9-1.
Graham Neubig, Yosuke Nakata, and Shinsuke Mori. Pointwise Prediction for Robust, Adaptable
Japanese Morphological Analysis. In Proceedings of ACL, 2011.
14
Published as a conference paper at ICLR 2017
Ankur P. Parikh, Oscar Tackstrom, Dipanjan Das, and Jakob Uszkoreit. A Decomposable Attention
Model for Natural Language Inference. In Proceedings of EMNLP, 2016.
Jian Peng, Liefeng Bo, and Jinbo Xu. Conditional Neural Fields. In Proceedings of NIPS, 2009.
Jeffrey Pennington, Richard Socher, and Christopher D. Manning. GloVe: Global Vectors for Word
Representation. In Proceedings of EMNLP, 2014.
Tim Rocktäschel, Edward Grefenstette, Karl Moritz Hermann, Tomas Kocisky, and Phil Blunsom.
Reasoning about Entailment with Neural Attention. In Proceedings of ICLR, 2016.
John Schulman, Nicolas Heess, Theophane Weber, and Pieter Abbeel. Gradient estimation using
stochastic computation graphs. In Advances in Neural Information Processing Systems, pp. 3528–
3536, 2015.
David A. Smith and Jason Eisner. Dependency Parsing as Belief Propagation. In Proceedings of
EMNLP, 2008.
Veselin Stoyanov and Jason Eisner. Minimum-Risk Training of Approximate CRF-based NLP Systems. In Proceedings of NAACL, 2012.
Veselin Stoyanov, Alexander Ropson, and Jason Eisner. Empirical Risk Minimization of Graphical
Model Parameters Given Approximate Inference, Decoding, and Model Structure. In Proceedings
of AISTATS, 2011.
Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-To-End Memory Networks. In Proceedings of NIPS, 2015.
Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer Networks. In Proceedings of NIPS,
2015.
Shenlong Wang, Sanja Fidler, and Raquel Urtasun. Proximal Deep Structured Models. In Proceedings of NIPS, 2016.
Shuohang Wang and Jing Jiang. Learning Natural Language Inference with LSTM. In Proceedings
of NAACL, 2016.
Jason Weston, Sumit Chopra, and Antoine Bordes. Memory Networks. arXiv:1410.3916, 2014.
Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M Rush, Bart van Merriënboer, Armand
Joulin, and Tomas Mikolov. Towards Ai-complete Question Answering: A Set of Prerequisite
Toy Tasks. arXiv preprint arXiv:1502.05698, 2015.
Kelvin Xu, Jimma Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov,
Richard Zemel, and Yoshua Bengio. Show, Attend and Tell: Neural Image Caption Generation
with Visual Attention. In Proceedings of ICML, 2015.
Lei Yu, Jan Buys, and Phil Blunsom. Online Segment to Segment Neural Transduction. In Proceedings of EMNLP, 2016.
Lei Yu, Phil Blunsom, Chris Dyer, Edward Grefenstette, and Tomas Kocisky. The Neural Noisy
Channel. In Proceedings of ICLR, 2017.
Kai Zhao, Liang Huang, and Minbo Ma. Textual Entailment with Structured Attentions and Composition. In Proceedings of COLING, 2016.
15
Published as a conference paper at ICLR 2017
APPENDICES
A
A.1
M ODEL D ETAILS
S YNTACTIC ATTENTION
The syntactic attention layer (for tree transduction and natural language inference) is similar to the
first-order graph-based dependency parser of Kipperwasser & Goldberg (2016). Given an input sentence [x1 , . . . , xn ] and the corresponding word vectors [x1 , . . . , xn ], we use a bidirectional LSTM
to get the hidden states for each time step i ∈ [1, . . . , n],
hfwd
= LSTM(xi , hfwd
i
i−1 )
hbwd
= LSTM(xi , hbwd
i
i+1 )
bwd
hi = [hfwd
i ; hi ]
where the forward and backward LSTMs have their own parameters. The score for xi → xj (i.e. xi
is the parent of xj ), is given by an MLP
θij = tanh(s> tanh(W1 hi + W2 hj + b))
These scores are used as input to the inside-outside algorithm (see Appendix B) to obtain the probability of each word’s parent p(zij = 1 | x), which is used to obtain the soft-parent cj for each word
xj . In the non-structured case we simply have p(zij = 1 | x) = softmax(θij ).
A.2
T REE T RANSDUCTION
Let [x1 , . . . , xn ], [y1 , . . . , ym ] be the sequence of source/target symbols, with the associated embeddings [x1 , . . . , xn ], [y1 , . . . , ym ] with xi , yj ∈ Rl . In the simplest baseline model we take the source
representation to be the matrix of the symbol embeddings. The decoder is a one-layer LSTM which
produces the hidden states h0j = LSTM(yj , h0j−1 ), with h0j ∈ Rl . The hidden states are combined
with the input representation via a bilinear map W ∈ Rl×l to produce the attention distribution used
to obtain the vector mi , which is combined with the decoder hidden state as follows,
exp xi Wh0j
αi = Pn
0
k=1 exp xk Whj
mi =
n
X
αi xi
ĥj = tanh(U[mi ; h0j ])
i=1
Here we have W ∈ Rl×l and U ∈ R2l×l . Finally, ĥj is used to to obtain a distribution over the next
symbol yj+1 ,
p(yj+1 | x1 , . . . , xn , y1 , . . . , yj ) = softmax(Vĥj + b)
For structured/simple models, the j-th source representation are respectively
"
#
"
#
n
n
X
X
x̂i = xi ;
p(zki = 1 | x) xk
x̂i = xi ;
softmax(θki ) xk
k=1
k=1
where θij comes from the bidirectional LSTM described in A.1. Then αi and mi changed accordingly,
exp x̂i Wh0j
αi = Pn
0
k=1 exp x̂k Whj
mi =
n
X
αi x̂i
i=1
Note that in this case we have W ∈ R2l×l and U ∈ R3l×l . We use l = 50 in all our experiments.
The forward/backward LSTMs for the parsing LSTM are also 50-dimensional. Symbol embeddings
are shared between the encoder and the parsing LSTMs.
Additional training details include: batch size of 20; training for 13 epochs with a learning rate
of 1.0, which starts decaying by half after epoch 9 (or the epoch at which performance does not
improve on validation, whichever comes first); parameter initialization over a uniform distribution
U [−0.1, 0.1]; gradient normalization at 1 (i.e. renormalize the gradients to have norm 1 if the l2
norm exceeds 1). Decoding is done with beam search (beam size = 5).
16
Published as a conference paper at ICLR 2017
A.3
N EURAL M ACHINE T RANSLATION
The baseline NMT system is from Luong et al. (2015). Let [x1 , . . . , xn ], [y1 , . . . , ym ] be the
source/target sentence, with the associated word embeddings [x1 , . . . , xn ], [y1 , . . . , ym ]. The encoder is an LSTM over the source sentence, which produces the hidden states [h1 , . . . , hn ] where
hi = LSTM(xi , hi−1 )
and hi ∈ R . The decoder is another LSTM which produces the hidden states h0j ∈ Rl . In the simple
attention case with categorical attention, the hidden states are combined with the input representation
via a bilinear map W ∈ Rl×l and this distribution is used to obtain the context vector at the j-th
time step,
l
θi = hi Wh0j
cj =
n
X
softmax(θi )hi
i=1
The Bernoulli attention network has the same θi but instead uses a sigmoid to obtain the weights of
the linear combination, i.e.,
cj =
n
X
sigmoid(θi )hi
i=1
And finally, the structured attention model uses a bilinear map to parameterize one of the unary
potentials
hi Wh0j , k = 1
θi (k) =
0,
k=0
θi,i+1 (zi , zi+1 ) = θi (zi ) + θi+1 (zi+1 ) + bzi ,zi+1
where b are the pairwise potentials. These potentials are used as inputs to the forward-backward
algorithm to obtain the marginals p(zi = 1 | x, q), which are further normalized to obtain the context
vector
cj =
n
X
p(zi = 1 | x, q)
hi
γ
i=1
n
γ=
1X
p(zi = 1 | x, q)
λ i
We use λ = 2 and also add an l2 penalty of 0.005 on the pairwise potentials b. The context vector
is then combined with the decoder hidden state
ĥj = tanh(U[cj ; h0j ])
and ĥj is used to obtain the distribution over the next target word yj+1
p(yj+1 | x1 , . . . , xn , y1 , . . . yj ) = softmax(Vĥj + b)
The encoder/decoder LSTMs have 2 layers and 500 hidden units (i.e. l = 500).
Additional training details include: batch size of 128; training for 30 epochs with a learning rate of
1.0, which starts decaying by half after the first epoch at which performance does not improve
on validation; dropout with probability 0.3; parameter initialization over a uniform distribution
U [−0.1, 0.1]; gradient normalization at 1. We generate target translations with beam search (beam
size = 5), and evaluate with multi-bleu.perl from Moses.9
A.4
Q UESTION A NSWERING
Our baseline model (MemN2N) is implemented following the same architecture as described in
Sukhbaatar et al. (2015). In particular, let x = [x1 , . . . , xn ] represent the sequence of n facts with
the associated embeddings [x1 , . . . , xn ] and let q be the embedding of the query q. The embeddings
9
https://github.com/moses-smt/mosesdecoder/blob/master/scripts/generic/
multi-bleu.perl
17
Published as a conference paper at ICLR 2017
are obtained by simply adding the word embeddings in each sentence or query. The full model with
K hops is as follows:
p(zk = i | x, q) = softmax((xki )> qk )
n
X
k
c =
p(zk = i | x, q)oki
i=1
k+1
q
= q k + ck
p(y | x, q) = softmax(W(qK + cK ))
where p(y | x, q) is the distribution over the answer vocabulary. At each layer, {xki } and {oki } are
computed using embedding matrices Xk and Ok . We use the adjacent weight tying scheme from
the paper so that Xk+1 = Ok , WT = OK . X1 is also used to compute the query embedding at the
first hop. For k = 1 we have xki = xi , qk = q, ck = 0.
For both the Unary and the Binary CRF models, the same input fact and query representations are
computed (i.e. same embedding matrices with weight tying scheme). For the unary model, the
potentials are parameterized as
θk (i) = (xki )> qk
and for the binary model we compute pairwise potentials as
θk,k+1 (i, j) = (xki )> qk + (xki )> xk+1
+ (xk+1
)> qk+1
j
j
The qk ’s are updated simply with a linear mapping, i.e.
qk+1 = Qqk
In the case of the Binary CRF, to discourage the model from selecting the same fact again we
additionally set θk,k+1 (i, i) = −∞ for all i ∈ {1, . . . , n}. Given these potentials, we compute the
marginals p(zk = i, zk+1 = j | x, q) using the forward-backward algorithm, which is then used to
compute the context vector:
c=
X
p(z1 , . . . , zK | x, q)f (x, z)
f (x, z) =
z1 ,...,zK
K
X
fk (x, zk )
fk (x, zk ) = okzk
k=1
Note that if f (x, z) factors over the components of z (as is the case above) then computing c only
requires evaluating the marginals p(zk | x, q).
Finally, given the context vector the prediction is made in a similar fashion to MemN2N:
p(y | x, q) = softmax(W(qK + c))
Other training setup is similar to Sukhbaatar et al. (2015): we use stochastic gradient descent with
learning rate 0.01, which is divided by 2 every 25 epochs until 100 epochs are reached. Capacity
of the memory is limited to 25 sentences. The embedding vectors are of size 20 and gradients are
renormalized if the norm exceeds 40. All models implement position encoding, temporal encoding,
and linear start from the original paper. For linear start, the softmax(·) function in the attention
layer is removed at the beginning and re-inserted after 20 epochs for MemN2N, while for the CRF
models we apply a log(softmax(·)) layer on the qk after 20 epochs. Each model is trained separately
for each task.
A.5
NATURAL L ANGUAGE I NFERENCE
Our baseline model/setup is essentially the same as that of Parikh et al. (2016).
Let
[x1 , . . . , xn ], [y1 , . . . , ym ] be the premise/hypothesis, with the corresponding input representations
[x1 , . . . , xn ], [y1 , . . . , ym ]. The input representations are obtained by a linear transformation of
the 300-dimensional pretrained GloVe embeddings (Pennington et al., 2014) after normalizing the
GloVe embeddings to have unit norm.10 The pretrained embeddings remain fixed but the linear layer
10
We use the GloVe embeddings pretrained over the 840 billion word Common Crawl, publicly available at
http://nlp.stanford.edu/projects/glove/
18
Published as a conference paper at ICLR 2017
(which is also 300-dimensional) is trained. Words not in the pretrained vocabulary are hashed to one
of 100 Gaussian embeddings with mean 0 and standard deviation 1.
We concatenate each input representation with a convex combination of the other sentence’s input
representations (essentially performing inter-sentence attention), where the weights are determined
through a dot product followed by a softmax,
"
#
m
n
X
X
exp
e
exp
e
ij
ij
Pm
Pn
eij = f (xi )> f (yj ) x̄i = xi ;
yj ȳj = yj ;
xi
k=1 exp eik
k=1 exp ekj
j=1
i=1
Here f (·) is an MLP. The new representations are fed through another MLP g(·), summed, combined
with the final MLP h(·) and fed through a softmax layer to obtain a distribution over the labels l,
x̄ =
n
X
g(x̄i )
ȳ =
i=1
m
X
g(ȳj )
j=1
p(l | x1 , . . . , xn , y1 , . . . , ym ) = softmax(Vh([x̄; ȳ]) + b)
All the MLPs have 2-layers, 300 ReLU units, and dropout probability of 0.2. For structured/simple
models, we first employ the bidirectional parsing LSTM (see A.1) to obtain the scores θij . In the
structured case each word representation is simply concatenated with its soft-parent
"
#
n
X
x̂i = xi ;
p(zki = 1 | x)xk
k=1
and x̂i (and analogously ŷj ) is used as the input to the above model. In the simple case (which
closely corresponds to the intra-sentence attention model of Parikh et al. (2016)), we have
#
"
n
X
exp θki
Pn
xk
x̂i = xi ;
l=1 exp θli
k=1
The word embeddings for the parsing LSTMs are also initialized with GloVe, and the parsing layer
is shared between the two sentences. The forward/backward LSTMs for the parsing layer are 100dimensional.
Additional training details include: batch size of 32; training for 100 epochs with Adagrad (Duchi
et al., 2011) where the global learning rate is 0.05 and sum of gradient squared is initialized to
0.1; parameter intialization over a Gaussian distribution with mean 0 and standard deviation 0.01;
gradient normalization at 5. In the pretrained scenario, pretraining is done with Adam (Kingma &
Ba, 2015) with learning rate equal to 0.01, and β1 = 0.9, β2 = 0.999.
B
F ORWARD /BACKWARD THROUGH THE I NSIDE -O UTSIDE A LGORITHM
Figure 6 shows the procedure for obtaining the parsing marginals from the input potentials. This
corresponds to running the inside-outside version of Eisner’s algorithm (Eisner, 1996). The intermediate data structures used during the dynamic programming algorithm are the (log) inside tables
α, and the (log) outside tables β. Both α, β are of size n × n × 2 × 2, where n is the sentence length.
First two dimensions encode the start/end index of the span (i.e. subtree). The third dimension
encodes whether the root of the subtree is the left (L) or right (R) index of the span. The fourth
dimension indicates if the span is complete (1) or incomplete (0). We can calculate the marginal
distribution of each word’s parent (for all words) in O(n3 ) using this algorithm.
Backward pass through the inside-outside algorithm is slightly more involved, but still takes O(n3 )
time. Figure 7 illustrates the backward procedure, which receives the gradient of the loss L with
respect to the marginals, ∇L
p , and computes the gradient of the loss with respect to the potentials
∇L
.
The
computations
must
be performed in the signed log-space semifield to handle log of negative
θ
values. See section 3.3 and Table 1 for more details.
19
Published as a conference paper at ICLR 2017
procedure I NSIDE O UTSIDE(θ)
α, β ← −∞
. Initialize log of inside (α), outside (β) tables
for i = 1, . . . , n do
α[i, i, L, 1] ← 0
α[i, i, R, 1] ← 0
β[1, n, R, 1] ← 0
for k = 1, . . . , n do
. Inside step
for s = 1, . . . , n − k do
t←s+k
L
α[s, t, R, 0] ← u∈[s,t−1] α[s, u, R, 1] ⊗ α[u + 1, t, L, 1] ⊗ θst
L
α[s, t, L, 0] ← u∈[s,t−1] α[s, u, R, 1] ⊗ α[u + 1, t, L, 1] ⊗ θts
L
α[s, t, R, 1] ← u∈[s+1,t] α[s, u, R, 0] ⊗ α[u, t, R, 1]
L
α[s, t, L, 1] ← u∈[s,t−1] α[s, u, L, 1] ⊗ α[u, t, L, 0]
for k = n, . . . , 1 do
. Outside step
for s = 1, . . . , n − k do
t←s+k
for u = s + 1, . . . , t do
β[s, u, R, 0] ←⊕ β[s, t, R, 1] ⊗ α[u, t, R, 1]
β[u, t, R, 1] ←⊕ β[s, t, R, 1] ⊗ α[s, u, R, 0]
if s > 1 then
for u = s, . . . , t − 1 do
β[s, u, L, 1] ←⊕ β[s, t, L, 1] ⊗ α[u, t, L, 0]
β[u, t, L, 0] ←⊕ β[s, t, L, 1] ⊗ α[s, u, L, 1]
for u = s, . . . , t − 1 do
β[s, u, R, 1] ←⊕ β[s, t, R, 0] ⊗ α[u + 1, t, L, 1] ⊗ θst
β[u + 1, t, L, 1] ←⊕ β[s, t, R, 0] ⊗ α[s, u, R, 1] ⊗ θst
if s > 1 then
for u = s, . . . , t − 1 do
β[s, u, R, 1] ←⊕ β[s, t, L, 0] ⊗ α[u + 1, t, L, 1] ⊗ θts
β[u + 1, t, L, 1] ←⊕ β[s, t, L, 0] ⊗ α[s, u, R, 1] ⊗ θts
A ← α[1, n, R, 1]
. Log partition
for s = 1, . . . , n − 1 do
. Compute marginals. Note that p[s, t] = p(zst = 1 | x)
for t = s + 1, . . . , n do
p[s, t] ← exp(α[s, t, R, 0] ⊗ β[s, t, R, 0] ⊗ −A)
if s > 1 then
p[t, s] ← exp(α[s, t, L, 0] ⊗ β[s, t, L, 0] ⊗ −A)
return p
Figure 6: Forward step of the syntatic attention layer to compute the marginals, using the inside-outside
algorithm (Baker, 1979) on the data structures of Eisner (1996). We assume the special root symbol is the first
element of the sequence, and that the sentence length is n. Calculations are performed in log-space semifield
with ⊕ = logadd and ⊗ = + for numerical precision. a, b ← c means a ← c and b ← c. a ←⊕ b means
a ← a ⊕ b.
20
Published as a conference paper at ICLR 2017
procedure BACKPROP I NSIDE O UTSIDE(θ, p, ∇L
p)
log p
for s, t = 1, . . . , n; s 6= t do
. Backpropagation uses the identity ∇L
∇L
p )∇θ
θ = (p
L
δ[s, t] ← log p[s, t] ⊗ log ∇p [s, t]
. δ = log(p ∇L
p)
L
L
L
L
L
∇α , ∇β , log ∇θ ← −∞
. Initialize inside (∇α ), outside (∇β ) gradients, and log of ∇L
θ
L
for s = 1, . . . , n − 1 do
. Backpropagate δ to ∇L
α and ∇β
for t = s + 1, . . . , n do
L
∇L
α [s, t, R, 0], ∇β [s, t, R, 0] ← δ[s, t]
L
∇α [1, n, R, 1] ←⊕ −δ[s, t]
if s > 1 then
L
∇L
α [s, t, L, 0], ∇β [s, t, L, 0] ← δ[t, s]
L
∇α [1, n, R, 1] ←⊕ −δ[s, t]
for k = 1, . . . , n do
. Backpropagate through outside step
for s = 1, . . . , n − k do
t←s+k
ν ← ∇L
. ν, γ are temporary values
β [s, t, R, 0] ⊗ β[s, t, R, 0]
for u = t, . . . , n do
L
∇L
β [s, u, R, 1], ∇α [t, u, R, 1] ←⊕ ν ⊗ β[s, u, R, 1] ⊗ α[t, u, R, 1]
if s > 1 then
ν ← ∇L
β [s, t, L, 0] ⊗ β[s, t, L, 0]
for u = 1, . . . , s do
L
∇L
β [u, t, L, 1], ∇α [u, s, L, 1] ←⊕ ν ⊗ β[u, t, L, 1] ⊗ α[u, s, L, 1]
L
ν ← ∇β [s, t, L, 1] ⊗ β[s, t, L, 1]
for u = t, . . . , n do
L
∇L
β [s, u, L, 1], ∇α [t, u, L, 0] ←⊕ ν ⊗ β[s, u, L, 1] ⊗ α[t, u, L, 1]
for u = 1, . . . , s − 1 do
γ ← β[u, t, R, 0] ⊗ α[u, s − 1, R, 1] ⊗ θut
L
L
∇L
β [u, t, R, 0], ∇α [u, s − 1, R, 1], log ∇θ [u, t] ←⊕ ν ⊗ γ
γ ← β[u, t, L, 0] ⊗ α[u, s − 1, R, 1] ⊗ θtu
L
L
∇L
β [u, t, L, 0], ∇α [u, s − 1, R, 1], log ∇θ [t, u] ←⊕ ν ⊗ γ
L
ν ← ∇β [s, t, R, 1] ⊗ β[s, t, R, 1]
for u = 1, . . . , s do
L
∇L
β [u, t, R, 1], ∇α [u, s, R, 0] ←⊕ ν ⊗ β[u, t, R, 1] ⊗ α[u, s, R, 0]
for u = t + 1, . . . , n do
γ ← β[s, u, R, 0] ⊗ α[t + 1, u, L, 1] ⊗ θsu
L
L
∇L
β [s, u, R, 0], ∇α [t + 1, u, L, 1], log ∇θ [s, u] ←⊕ ν ⊗ γ
γ ← β[s, u, L, 0] ⊗ α[t + 1, u, L, 1] ⊗ θus
L
L
∇L
β [s, u, L, 0], ∇α [t + 1, u, L, 1], log ∇θ [u, s] ←⊕ ν ⊗ γ
for k = n, . . . , 1 do
. Backpropagate through inside step
for s = 1, . . . , n − k do
t←s+k
ν ← ∇L
α [s, t, R, 1] ⊗ α[s, t, R, 1]
for u = s + 1, . . . , t do
L
∇L
α [u, t, R, 0], ∇α [u, t, R, 1] ←⊕ ν ⊗ α[s, u, R, 0] ⊗ α[u, t, R, 1]
if s > 1 then
ν ← ∇L
α [s, t, L, 1] ⊗ α[s, t, L, 1]
for u = s, . . . , t − 1 do
L
∇L
α [s, u, L, 1], ∇α [u, t, L, 0] ←⊕ ν ⊗ α[s, u, L, 1] ⊗ α[u, t, L, 0]
ν ← ∇L
α [s, t, L, 0] ⊗ α[s, t, L, 0]
for u = s, . . . , t − 1 do
γ ← α[s, u, R, 1] ⊗ α[u + 1, t, L, 1] ⊗ θts
L
L
∇L
α [s, u, R, 1], ∇α [u + 1, t, L, 1], log ∇θ [t, s] ←⊕ ν ⊗ γ
L
ν ← ∇α [s, t, R, 0] ⊗ α[s, t, R, 0]
for u = s, . . . , t − 1 do
γ ← α[s, u, R, 1] ⊗ α[u + 1, t, L, 1] ⊗ θst
L
L
∇L
α [s, u, R, 1], ∇α [u + 1, t, L, 1], log ∇θ [s, t] ←⊕ ν ⊗ γ
L
return signexp log ∇θ
. Exponentiate log gradient, multiply by sign, and return ∇L
θ
Figure 7: Backpropagation through the inside-outside algorithm to calculate the gradient with respect to the
input potentials. ∇ab denotes the Jacobian of a with respect to b (so ∇L
θ is the gradient with respect to θ).
a, b ←⊕ c means a ← a ⊕ c and b ← b ⊕ c.
21
| 9cs.NE
|
arXiv:1702.07697v2 [cs.IT] 15 Nov 2017
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE
POLYNOMIALS
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
Abstract. We consider the class of Rudin-Shapiro-like polynomials, whose
L4 norms on the complex unit circle were studied by Borwein and Mossinghoff. The polynomial f (z) = f0 + f1 z + · · · + fd z d is identified with the
sequence (f0 , f1 , . . . , fd ) of its coefficients. From the L4 norm of a polynomial, one can easily calculate the autocorrelation merit factor of its associated sequence, and conversely. In this paper, we study the crosscorrelation
properties of pairs of sequences associated to Rudin-Shapiro-like polynomials. We find an explicit formula for the crosscorrelation merit factor. A
computer search is then used to find pairs of Rudin-Shapiro-like polynomials
whose autocorrelation and crosscorrelation merit factors are simultaneously
high. Pursley and Sarwate proved a bound that limits how good this combined autocorrelation and crosscorrelation performance can be. We find
infinite families of polynomials whose performance approaches quite close
to this fundamental limit.
1. Introduction
This paper concerns the discovery of Rudin-Shapiro-like polynomials that
have exceptionally good correlation properties. Shapiro [21] recursively constructed a family of polynomials with coefficients in {−1, 1} that are flat√on
the complex unit circle: the ratio of their L∞ to L2 norm never exceeds 2.
Around the same time, Golay [3] constructed binary sequences following the
same recursion as the coefficients of Shapiro’s polynomials. Shapiro’s polynomials were subsequently rediscovered by Rudin [16]. Littlewood used the L4
norm on the complex unit circle in his investigation [12] into the flatness of
polynomials with coefficients in {−1, 1}, which are now known as Littlewood
polynomials. He calculated the ratio of L4 to L2 norm of the Rudin-Shapiro
polynomials in [13, Problem 19].
Golay [4] independently developed the merit factor, a normalized average of
the mean squared magnitude of the aperiodic autocorrelation of sequences used
in remote sensing and communications networks. Eventually it was discovered
that determining the L4 norm on the unit circle of a polynomial is tantamount
to determining Golay’s merit factor for the sequence of coefficients of that
polynomial (see [7, eq. (4.1)]). We shall soon make precise this connection
Date: 15 November 2017.
This paper is based on work of the three authors supported in part by the National Science
Foundation under Grant DMS 1500856.
1
2
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
between the analytic behavior of polynomials on the complex unit circle and
the correlation behavior of their associated sequences.
Inspired by the work of Littlewood on the Rudin-Shapiro polynomials, Borwein and Mossinghoff [2] recursively define sequences f0 (z), f1 (z), . . . of polynomials, where f0 (z) is any Littlewood polynomial and the rest of the polynomials
are obtained via a recursion of the form
(1)
fn+1 (z) = fn (z) + σn z 1+deg fn fn† (−z),
where σn ∈ {−1, 1} represents an arbitrary sign that can differ at each stage of
the recursion, and where for any polynomial a(z) = a0 + a1 z + · · · + ad z d ∈ C[z]
of degree d, the polynomial a† (z) denotes the conjugate reciprocal polynomial
of a(z), which is ad + ad−1 z + · · · + a0 z d .1 If one sets f0 (z) = 1, σ0 = 1, and
σn = (−1)n+1 for all positive n, then f0 , f1 , f2 , . . . become Shapiro’s original
polynomials [21, Theorem 5(ii)].
In this paper, we relax the condition that the initial polynomial f0 (z) be
a Littlewood polynomial. We allow f0 (z) to be a polynomial in C[z], and
only impose the condition that f0 (z) have a nonzero constant coefficient. This
ensures that f0† (z) has the same degree as f0 (z), so that when we construct
f0 (z), f1 (z), . . . via recursion (1), a straightforward induction shows that every
fn has a nonzero constant coefficient and
(2)
1 + deg fn = 2n (1 + deg f0 )
for every n. We call the sequence σ = σ0 , σ1 , . . . of numbers in {−1, 1} that
occur in our recursion the sign sequence for that recursion. We call f0 (z) the
seed, and the sequence f0 (z), f1 (z), . . . of polynomials obtained from the seed
by applying the recursion is called the stem associated to seed f0 (z) and sign
sequence σ. Any stem obtained from a seed f0 (z) ∈ C[z] with nonzero constant
coefficient is also called a sequence of Rudin-Shapiro-like polynomials.
Borwein and Mossinghoff [2] study the L4 norm of Rudin-Shapiro-like Littlewood polynomials on the complex unit circle, or equivalently, the autocorrelation merit factor of these polynomials. We now describe the correspondence
between polynomials and sequences, and the relation of Lp norms on the complex unit circle to correlation.
In this paper, by a sequence of length ℓ we mean some (a0 , a1 , . . . , aℓ−1 ) ∈ Cℓ .
Most researchers are especially interested in sequences that are unimodular,
that is, whose terms are all of unit magnitude, and are most of all interested
in sequences that are binary, that is, whose terms lie in {−, 1, 1}. We always
identify the polynomial a(z) = a0 + a1 z + · · · + aℓ−1 z ℓ−1 ∈ C[z] of degree ℓ − 1
with the sequence a = (a0 , a1 , . . . , aℓ−1 ) of length ℓ. With this identification,
binary sequences correspond to Littlewood polynomials. Usually it is easier
to work with sequence length rather than polynomial degree, and using our
1In fact, Borwein-Mossinghoff use the reciprocal polynomial a∗ (z) = a + a
d
d−1 z + · · · + a0 ,
but since they are working with polynomials with real coefficients, this is the same as a† (z).
In this paper, it was found that using a† (z) instead of a∗ (z) gives the natural generalization
of their recursion for polynomials with non-real coefficients.
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
3
identification of sequences with polynomials, we define the length of a polynomial to be the length of the sequence associated with the polynomial, that is,
len a = 1 + deg a. Notice that when we have a stem f0 (z), f1 (z), . . . generated
from a seed f0 (z) ∈ C[z] with nonzero constant coefficient via recursion (1),
using length rather than degree simplifies relation (2) to
(3)
len fn = 2n len f0 .
If f = (f0 , f1 , . . . , fℓ−1 ) and g = (g0 , g1 , . . . , gℓ−1 ) are two sequences of length
ℓ and s ∈ Z, then we define the aperiodic crosscorrelation of f and g at shift s
to be
X
Cf,g (s) =
fj+s gj ,
j∈Z
where we take fj = gj = 0 whenever j 6∈ {0, 1, . . . , ℓ − 1}.
Autocorrelation is crosscorrelation of a sequence with itself, so the aperiodic
autocorrelation of f at shift s is just Cf,f (s). If the terms of f are complex
numbers of unit magnitude, then Cf,f (0) = len f .
Autocorrelation and crosscorrelation are studied extensively because of their
importance in communications networks: see [18, 17, 5, 9, 8, 6, 19] for some
overviews. It is desirable to have sequences whose autocorrelation values at
all nonzero shifts are small in magnitude, and it is desirable to have pairs of
sequences whose crosscorrelation values at all shifts are small in magnitude.
For a polynomial a(z) = a0 + a1 z + · · · + aℓ−1 z ℓ−1 ∈ C[z], we define a(z) to be
the Laurent polynomial a0 + a1 z −1 + · · · + aℓ−1 z −(ℓ−1) . (We identify z with z −1
because we are concerned with the properties of our polynomials on the complex
unit circle.) If f (z) and g(z) are polynomials in C[z], then it is not hard to
show that the values of the crosscorrelation between their associated sequences
at all shifts are recorded in the following product of Laurent polynomials:
X
Cf,g (s)z s .
f (z)g(z) =
s∈Z
The crosscorrelation demerit factor of f and g is defined to be
P
2
s∈Z |Cf,g (s)|
.
CDF(f, g) =
|Cf,f (0)| · |Cg,g (0)|
Its reciprocal, the crosscorrelation merit factor, is defined as CMF(f, g) =
1/ CDF(f, g). A low demerit factor (or equivalently, high merit factor) indicates
a sequence pair whose crosscorrelation values are collectively low, hence desirable. The autocorrelation demerit factor of f is much like the crosscorrelation
demerit factor, but omits |Cf,f (0)|2 in the numerator:
P
2
s∈Z,s6=0 |Cf,f (s)|
= CDF(f, f ) − 1,
(4)
ADF(f ) =
|Cf,f (0)|2
and the autocorrelation merit factor is its reciprocal AMF(f ) = 1/ ADF(f ).
The autocorrelation merit factor defined here is Golay’s original merit factor,
introduced in [4].
4
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
If f (z) ∈ C[z, z −1 ] is a Laurent polynomial, and p is a real number with
p ≥ 1, then we define the Lp norm of f (z) on the complex unit circle to be
1/p
Z 2π
1
.
|f (eiθ )|p dθ
kf kp =
2π 0
One can show (see [10, Section V]) that
(5)
CDF(f, g) =
and
kf gk22
,
kf k22 kgk22
kf k44
− 1.
kf k42
Borwein and Mossinghoff [2, Theorem 1 and Corollary 1] explicitly calculate
the autocorrelation demerit factors for Rudin-Shapiro-like Littlewood polynomials and determine their asymptotic behavior.
ADF(f ) = CDF(f, f ) − 1 =
Theorem 1.1 (Borwein-Mossinghoff, 2000). If f0 , f1 , f2 , . . . is a sequence of
Rudin-Shapiro-like polynomials generated from any Littlewood polynomial f0 (z)
via recursion (1), then
lim ADF(fn ) = −1 +
n→∞
where fe0 (z) is the polynomial f0 (−z).
1
2 kf0 k44 + kf0 fe0 k22
·
≥ ,
3
3
kf0 k42
Borwein and Mossinghoff [2, Section 3] go on to find examples where the
limiting autocorrelation demerit factor reaches the lower bound of 1/3, so that
well-chosen families of Rudin-Shapiro-like polynomials can reach asymptotic
autocorrelation merit factors as high as 3.
We are interested in both autocorrelation and crosscorrelation merit factors.
It turns out that there are limits to how good one can simultaneously make
autocorrelation and crosscorrelation performance. Pursley and Sarwate [15,
eqs. (3),(4)] proved a bound that relates autocorrelation and crosscorrelation
demerit factors:
p
(6)
| CDF(f, g) − 1| ≤ ADF(f ) ADF(g).
We define the Pursley-Sarwate Criterion of f and g to be
p
(7)
PSC(f, g) = ADF(f ) ADF(g) + CDF(f, g),
so that (6) implies that
(8)
PSC(f, g) ≥ 1.
Since we want sequence pairs with low mutual crosscorrelation and where both
sequences individually have low autocorrelation, we would like to find f and g
with PSC(f, g) as close to 1 as possible.
In pursuit of this goal, we found a formula for the asymptotic crosscorrelation
demerit factor of a family of pairs of Rudin-Shapiro-like sequences. This formula
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
5
specializes to give information about autocorrelation that generalizes the results
of Theorem 1.1 to embrace polynomials with coefficients other than −1 and 1.
Theorem 1.2. Let f0 , g0 ∈ C[z] be polynomials of equal length having nonzero
constant coefficients. If f0 , f1 , . . . and g0 , g1 , . . . are sequences of Rudin-Shapirolike polynomials generated from f0 and g0 via recursion (1), then
lim ADF(fn ) = −1 +
n→∞
1
2 kf0 k44 + kf0 fe0 k22
≥ ,
·
4
3
3
kf0 k2
g0 k22
1
2 kg0 k44 + kg0 e
·
≥ ,
3
3
kg0 k42
R
2kf0 g0 k2 + kf0 ge0 k2 + Re f0 fe0 g0 e
g0
lim ADF(gn ) = −1 +
n→∞
lim CDF(fn , gn ) =
n→∞
2
2
3kf0 k22 kg0 k22
,
where fe0 (z) and e
g0 (z) are respectively the polynomials f0 (−z) and g0 (−z), and
Z
1
f0 fe0 g0 ge0 =
2π
Z
0
2π
g0 (eiθ )dθ.
f0 (eiθ )fe0 (eiθ )g0 (eiθ )e
This theorem is proved in Corollary 2.7 in Section 2. We then use the formula
in Theorem 1.2 and computational searches (see Section 4) to find families of
pairs of Rudin-Shapiro-like Littlewood polynomials whose asymptotic PursleySarwate Criterion is as low as 331/300 = 1.10333 . . ., which is quite close to the
absolute lower bound in (8). In contrast, the typical Pursley-Sarwate Criterion
of randomly selected long binary sequences is about 2, and high-performance
sequence pairs constructed from finite field characters have been found with
asymptotic Pursley-Sarwate Criterion of 7/6 (see [10, §II.E, §IV.D] and [1,
eq.(6)]).
The rest of this paper is organized as follows. The goal of Section 2 is to
prove Theorem 1.2 above. This is accomplished by finding recursive relations
between various Lp norms associated with our Rudin-Shapiro-like polynomials
that arise from the original recursion (1).
Section 3 examines groups of symmetries that preserve the asymptotic correlation behavior when applied to our Rudin-Shapiro-like polynomials. These are
helpful in abbreviating computational searches (reported later in Section 4) for
polynomials with good autocorrelation and crosscorrelation performance. We
organize the good polynomials that we find into orbits modulo the action of
our symmetry groups, which makes our reports shorter and more intelligible.
In Section 4, we present some examples of Rudin-Shapiro-like Littlewood
polynomials f0 , f1 , . . . and g0 , g1 , . . . such that limn→∞ PSC(fn , gn ) is low, which
implies simultaneously good autocorrelation and crosscorrelation performance.
This includes the families of polynomials with the exceptionally low asymptotic
Pursley-Sarwate Criterion value reported above.
6
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
2. Asymptotic Crosscorrelation Formula
In this section we prove Theorem 1.2, the main theoretical result of this
paper. First we set down some notational conventions.
Throughout this paper, we let C[z, z −1 ] denote the ring of Laurent polynomials with coefficients from C. Because
we are working with polynomials on
P
the complex unit circle, if a(z) = j∈Z aj z j , then we use a(z) as a shorthand
P
for j∈Z aj z −j , Re(a(z)) as a shorthand for 21 (a(z) + a(z)), and |a(z)|2 as a
shorthand for a(z)a(z). We also use e
a(z) as a shorthand for a(−z). Also recall
from the Introduction that if a(z) = a0 + a1 z + · · · + ad z d is a polynomial of
degree d in C[z], then a† (z) denotes the conjugate reciprocal polynomial of a(z),
that is, a† (z) = ad + ad−1 z + · · · + a0 z d .
We first note how the transformations a 7→ e
a, a 7→ a, and a 7→ a† relate to
and interact with each other.
Lemma 2.1. If f (z) ∈ C[z], then
(i). f † (z) = z deg f f (z),
f† )(z) = (−z)deg f fe(z) = (−1)deg f · (fe)† (z), and
(ii). (f
(iii). f † (z) = z − deg f f (z).
Proof. For part (i), note that if d = deg f and f (z) = f0 + f1 z + · · · + fd z d ,
then by the definition of the conjugate reciprocal, we have
f †(z) = fd + · · · + f1 z d−1 + f0 z d
= z d (fd z −d + · · · + f1 z −1 + f0 )
= z d f (z).
For part (ii), use part (i) to see that
g
† )(z) = (−z)deg f f (−z)
(f
and then
= (−z)deg f fe(z),
e
(−z)deg f fe(z) = (−1)deg f · z deg f fe(z)
= (−1)deg f · (fe)† (z),
where we used part (i) again in the second equality.
For part (iii), use part (i) to see that
f † (z) = z deg f f (z)
= z − deg f f (z).
R
P
If a(z) ∈ C[z, z −1 ], say a(z) = j∈Z aj z j , then we use a(z) as a shorthand
for a0 . This is because if we actually perform integration on the complex unit
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
circle, we obtain a0 :
1
2π
Z
In particular note that
ka(z)k22
=
2π
7
a(eiθ )dθ = a0 .
0
Z
a(z)a(z) =
X
j∈Z
|aj |2 .
It will be important to know that replacing a with e
a changes neither integrals
nor norms.
Lemma 2.2.
R
R
(i). For any f (z) ∈ C[z, z −1 ], we have fe(z) = f (z).
(ii). For any f (z) ∈ C[z, z −1 ] and any p ∈ R with p ≥ 1, we have kfe(z)kp =
kf (z)kp .
Proof. If g(z) is any function that is integrable on the complex unit circle, then
Z 2π
Z π
Z 2π
1
1
1
iθ
i(θ+π)
g(−e )dθ =
g(e
)dθ +
g(ei(θ−π) )dθ
2π 0
2π 0
2π π
Z 2π
Z π
1
1
iη
g(e )dη +
g(eiη )dη
=
2π π
2π 0
Z 2π
1
g(eiθ )dθ.
=
2π 0
This proves parts (i) and (ii), where we set g(z) = f (z) or g(z) = |f (z)|p ,
respectively.
If we have two sequences f0 , f1 , . . . and g0 , g1 , . . . of Rudin-Shapiro-like polynomials constructed via recursion (1), then the following lemma tells us how
kfn+1 gn+1 k22 is related to kfn gn k22 . In view of (5), this is telling us how
CDF(fn , gn ) changes in one step of the recursion.
Lemma 2.3. Suppose fn (z), gn (z) ∈ C[z] are polynomials of length ℓ, and
let fn+1 (z) = fn (z) + σn z ℓ fn† (−z) and gn+1 (z) = gn (z) + τn z ℓ gn† (−z), where
σn , τn ∈ {−1, 1}. Then kfn+1 k22 = 2kfn k22 , kgn+1 k22 = 2kgn k22 , and if we define
(i). uj = kfj gj k22 ,
(ii). vj = kfj gej k22 ,
R
gj
(iii). wj = Re fj fej gj e
for j ∈ {n, n + 1}, then
(i). un+1 = 2un + 2vn + 2σn τn wn ,
(ii). vn+1 = 2un + 2vn − 2σn τn wn , and
(iii). wn+1 = 2σn τn un − 2σn τn vn + 2wn .
Proof. First observe that Lemma 2.1(ii) shows that
(9)
fn+1 (z) = fn (z) + σn (−1)ℓ−1 z 2ℓ−1 fen (z)
gn+1 (z) = gn (z) + τn (−1)ℓ−1 z 2ℓ−1 gen (z),
8
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
and so
kfn+1 k22
=
Z
Z
fn + σn (−1)ℓ−1 z 2ℓ−1 fen
2
|fn |2 + |fen |2 + 2 Re(σn (−1)ℓ−1 z 1−2ℓ fn fen )
Z
2
ℓ−1
= 2kfn k2 + 2σn (−1)
Re z 1−2ℓ fn fen
=
where the last equality uses Lemma 2.2(ii). Now observe that z 1−2ℓ fn fen is a
Laurent polynomial whose terms all have negative powers of z (because fn is
a polynomial of degree ℓ − 1), and so the last integral is zero. Thus we obtain
the desired result that kfn+1 k22 = 2kfn k22 . If one replaces every instance of f
with g in the above, one obtains a proof that kgn+1 k22 = 2kgn k22 .
Now we prove the recursions involving un , vn , and wn . First of all,
un+1 = kfn+1 gn+1 k22
1
1
gn+1 k22
= kfn+1 gn+1 k22 + kfen+1 e
2
2
2
1
fn + σn (−1)ℓ−1 z 2ℓ−1 fen gn + τn (−1)ℓ−1 z 2ℓ−1 gen
=
2
2
1 e
ℓ−1 2ℓ−1
ℓ−1 2ℓ−1
gn − τn (−1) z
fn e
gn
+
fn − σn (−1) z
2
Z
2
2
= 2kfn gn k2 + 2kfn e
gn k2 + 2σn τn Re fn fen gn e
gn
Z
+ 2σn τn Re z 2−4ℓ fn fen gn e
gn
Z
= 2un + 2vn + 2σn τn wn + 2σn τn Re z 2−4ℓ fn fen gn e
gn ,
2
2
where the first equality is the definition of un+1 , the second equality uses Lemma
2.2(ii), the third equality uses (9), and the fourth equality uses technical Lemma
2.10, which appears at the end of this section. Then note that z 2−4ℓ fn fen gn gen
is a Laurent polynomial whose terms all have negative powers of z (because fn
and gn are polynomials of degree ℓ − 1), so that the last integral in our chain
of equalities is zero, giving us the desired result.
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
9
We also have
vn+1 = kfn+1 e
gn+1 k22
1
1
gn+1 k22 + kfen+1 gn+1 k22
= kfn+1 e
2
2
2
1
fn + σn (−1)ℓ−1 z 2ℓ−1 fen gen − τn (−1)ℓ−1 z 2ℓ−1 gn
=
2
2
1 e
+
fn − σn (−1)ℓ−1 z 2ℓ−1 fn gn + τn (−1)ℓ−1 z 2ℓ−1 gen
2
Z
= 2kfn e
gn k22 + 2kfn gn k22 + 2σn (−τn ) Re fn fen gen gn
Z
+ 2σn (−τn ) Re z 2−4ℓ fn fen gen gn
Z
= 2un + 2vn − 2σn τn wn − 2σn τn Re z 2−4ℓ fn fen gn e
gn ,
2
2
where the first equality is the definition of vn+1 , the second equality uses Lemma
2.2(ii), the third equality uses (9), and the fourth equality uses technical Lemma
2.10, which appears at the end of this section. Then note that z 2−4ℓ fn fen gn gen
is a Laurent polynomial whose terms all have negative powers of z (because fn
and gn are polynomials of degree ℓ − 1), so that the last integral in our chain
of equalities is zero, giving us the desired result.
Finally, we use similar arguments to obtain
Z
wn+1 = Re fn+1 fen+1 gn+1 gen+1
Z
fn + σn (−1)ℓ−1 z 2ℓ−1 fen fen − σn (−1)ℓ−1 z 2ℓ−1 fn
= Re
× gn + τn (−1)ℓ−1 z 1−2ℓ gen gen − τn (−1)ℓ−1 z 1−2ℓ gn
Z h
i
fn fen − z 4ℓ−2 fn fen + σn (−1)ℓ−1 z 2ℓ−1 |fen |2 − |fn |2
= Re
h
i
× gn e
gn − z 2−4ℓ gn gen + τn (−1)ℓ−1 z 1−2ℓ |e
gn |2 − |gn |2
= I1 + σn τn I2 + τn (−1)ℓ−1 I3 + σn (−1)ℓ−1 I4 ,
where
I1 = Re
Z
Z
fn fen − z 4ℓ−2 fn fen
gn gen − z 2−4ℓ gn e
gn
|e
gn |2 − |gn |2
Z
fn fen − z 4ℓ−2 fn fen z 1−2ℓ |e
I3 = Re
gn |2 − |gn |2
Z
I4 = Re z 2ℓ−1 |fen |2 − |fn |2 gn e
gn − z 2−4ℓ gn gen .
I2 = Re
|fen |2 − |fn |2
10
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
Now we compute each of these four integrals.
Z
fn fen − z 4ℓ−2 fn fen gn e
gn − z 2−4ℓ gn gen
I1 = Re
Z
Z
= 2 Re fn fen gn gen − 2 Re z 2−4ℓ fn fen gn gen ,
and note that z 2−4ℓ fn fen gn e
gn is a Laurent polynomial whose terms all have
negative powers of z (because fn and gn are of length ℓ − 1), so that the last
integral is zero, and thus
I1 = 2wn .
Then
I2 = kfen e
gn k22 + kfn gn k22 − kfen gn k22 − kfn e
gn k22
= 2un − 2vn ,
where the second equality is due to Lemma 2.2(ii).
Let us examine the integrand in the definition of I3 , which is
h = z 1−2ℓ fn fen − z 2ℓ−1 fn fen |e
gn |2 − |gn |2 .
If one conjugates this, one obtains
gn |2 − |gn |2 ,
z 2ℓ−1 fn fen − z 1−2ℓ fn fen |e
which is Rjust −h. So h has purely imaginary values on the unit circle, and thus
I3 = Re h = 0. The same argument shows that I4 = 0, and so, putting all
our results together, we have
wn+1 = I1 + σn τn I2 + τn (−1)ℓ−1 I3 + σn (−1)ℓ−1 I4
= 2wn + σn τn (2un − 2vn ) + 0 + 0.
The above lemma allows us to compute crosscorrelation demerit factors for
pairs of Rudin-Shapiro-like polynomials constructed via recursion (1).
Theorem 2.4. Let f0 , g0 ∈ C[z] be polynomials of equal length having nonzero
constant coefficients. Let σ0 , σ1 , . . . be a sequence of values from {−1, 1}, and
suppose that fn (z) and gn (z) are defined recursively for all n ∈ N by
fn+1 (z) = fn (z) + σn z len fn fn† (−z)
gn+1 (z) = gn (z) + σn z len gn gn† (−z).
Then
R
2kf0 g0 k22 + kf0 ge0 k22 + Re f0 fe0 g0 ge0
kfn gn k22
=
kfn k22 kgn k22
3kf0 k22 kg0 k22
R
n
kf0 g0 k22 − kf0 e
g0 k22 − Re f0 fe0 g0 e
g0
1
+ −
.
2
2
2
3kf0 k2 kg0 k2
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
11
Proof. Since f0 and g0 have nonzero constant coefficients and are of the same
length, induction shows that for every n, the polynomials fn and gn have
nonzero constant coefficients and both are of length 2n len f0 = 2n len g0 , as
observed in (3) in the Introduction. Thus we may apply Lemma 2.3 repeatedly
to the pairs (fn , gn ) for every n.
Let uj , vj , and wj be as defined in Lemma 2.3. We want to calculate un ,
and the lemma says that
un+1
un
vn+1 = A vn ,
wn+1
wn
where
2 2
2
A = 2 2 −2 .
2 −2 2
Now A = BΛB −1 , where
−1 1 1
B = 1 1 0
1 0 1
So
and
−2 0 0
Λ = 0 4 0 .
0 0 4
un
u0
vn = An v0
wn
w0
u0
= BΛn B −1 v0 ,
w0
and so
kfn gn k22 = un =
4n (2u0 + v0 + w0 ) + (−2)n (u0 − v0 − w0 )
.
3
Repeated use of Lemma 2.3 also shows that kfn k22 = 2n kf0 k22 and kgn k22 =
and these norms are nonzero since f0 and g0 are nonzero, so that
2n kg0 k22 ,
kfn gn k22
(2u0 + v0 + w0 ) + (−1/2)n (u0 − v0 − w0 )
=
,
2
2
kfn k2 kgn k2
3kf0 k22 kg0 k22
and when one substitutes the values of u0 , v0 , and w0 as defined in Lemma 2.3,
then one obtains the desired result.
If f0 = g0 , we are considering autocorrelation. When we specialize to this
case and also specialize to the case where f0 is a Littlewood polynomial, then
we recover the results of Borwein and Mossinghoff [2, Theorem 1 and Corollary
1].
12
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
Corollary 2.5 (Borwein-Mossinghoff (2000)). Suppose that f0 (z) is a Littlewood polynomial and that fn (z) is defined recursively for all n ∈ N by
fn+1 (z) = fn (z) + z len fn fn† (−z).
Then
n
1 kf0 k44 − 2kf0 fe0 k22
kfn k44
2 kf0 k44 + kf0 fe0 k22
1
· ·
·
=
+
−
.
4
4
3
2
3
kfn k2
kf0 k2
kf0 k42
∗ 2
e 2
Remark 2.6. Borwein and Mossinghoff use kf0 ff
0 k2 instead of our kf0 f0 k2 , but it
is not hard to see that these are equal because f ∗ (−z) = fe†(z) = (−z)deg f fe(z)
0
for any Littlewood polynomial (see Lemma 2.1(ii)).
Our results now allow us to compute limiting autocorrelation and crosscorrelation demerit factors. The following corollary contains all the results that
we presented in Theorem 1.2 in the Introduction.
Corollary 2.7. Let f0 , g0 ∈ C[z] be polynomials of equal length having nonzero
constant coefficients. Let σ0 , σ1 , . . . be a sequence of values from {−1, 1}, and
suppose that fn (z) and gn (z) are defined recursively for all n ∈ N by
fn+1 (z) = fn (z) + σn z len fn fn† (−z)
gn+1 (z) = gn (z) + σn z len gn gn† (−z).
Then
lim ADF(fn ) = −1 +
n→∞
1
2 kf0 k44 + kf0 fe0 k22
≥ ,
·
3
3
kf0 k42
g0 k22
1
2 kg0 k44 + kg0 e
·
≥ ,
4
3
3
kg0 k2
R
2kf0 g0 k2 + kf0 ge0 k2 + Re f0 fe0 g0 e
g0
lim ADF(gn ) = −1 +
n→∞
lim CDF(fn , gn ) =
n→∞
so that
2
2
3kf0 k22 kg0 k22
,
R
g0 k22 + Re f0 fe0 g0 ge0
2kf0 g0 k22 + kf0 e
lim PSC(fn , gn ) =
n→∞
3kf0 k22 kg0 k22
r
2kf0 k44 + 2kf0 fe0 k22 − 3kf0 k42 2kg0 k44 + 2kg0 e
g0 k22 − 3kg0 k42
.
+
3kf0 k22 kg0 k22
Proof. The limiting crosscorrelation demerit factor is clear from Theorem 2.4
since the ratio of norms calculated there is the crosscorrelation demerit factor by
(5). For the limiting autocorrelation demerit factors, one again uses Theorem
2.4, but now one sets fn = gn for all n in that theorem, and combines the result
thus obtained with the fact from (4) that ADF(fn ) = CDF(fn , fn ) − 1 along
R
with the observations that kf0 f0 k2 = kf0 k4 and Re f0 fe0 f0 fe0 = kf0 fe0 k2 . The
2
4
2
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
13
limiting Pursley-Sarwate Criterion follows immediately from the definition in
(7) and the limits on the autocorrelation and crosscorrelation demerit factors.
To obtain the lower bounds on the limiting autocorrelation demerit factors,
one notes that for any f (z) ∈ C[z], we have
1
1
kf k44 + kf fek22 = kf k44 + kfek44 + kf fek22
2
2
1
2
= k(|f (z)| + |fe(z)|2 )k22
2
1
≥ k(|f (z)|2 + |fe(z)|2 )k21
2
2
1
kf k22 + kfek22
=
2
2
1
=
2kf k22
2
= 2kf k42 ,
where we use Lemma 2.2(ii) in the first and penultimate equalities, and the
inequality uses the fact that the L2 norm is always at least as large as the L1
norm (by Jensen’s
inequality)
because we are working on a space of measure 1.
2
4
e
Therefore, kf k4 + kf f k2 /kf k42 ≥ 2, which proves our lower bounds on the
limiting values autocorrelation demerit factors.
Notice that although Corollary 2.7 has lower bounds on limiting autocorrelation demerit factors, no lower bound for the crosscorrelation demerit factor is
given. This is because the only lower bound that could have been given is the
trivial lower bound of 0, for the following example shows that one can obtain
pairs of stems whose limiting crosscorrelation demerit factors are arbitrarily
close to 0.
Proposition 2.8. Let k be a positive integer, let f0 (z) = (1 − z 4k )/(1 − z),
and g0 (z) = (1 − z)(1 − z 2 )(1 − z 4k )/(1 − z 4 ). If fn (z) and gn (z) are defined
recursively for all n ∈ N by
fn+1 (z) = fn (z) + z len fn fn† (−z)
gn+1 (z) = gn (z) + z len gn gn† (−z),
then
1
.
3k
Remark 2.9. Note that f0 in Proposition 2.8 is a Littlewood polynomial representing a sequence of length 4k whose terms are all 1. And g0 is Littlewood
polynomial representing a sequence of length 4k consisting of k repetitions of
the smaller sequence (1, −1, −1, 1).
Proposition 2.8 shows that we can find a pair of stems whose limiting crosscorrelation demerit factor is as close to 0 as we like simply by choosing a sufficiently high value of k when we define f0 and g0 .
lim CDF(fn , gn ) =
n→∞
14
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
In view of the Pursley-Sarwate bound (6), we know that the the autocorrelation performance for such stems cannot be exceptionally good, and in fact, one
can use Corollary 2.7 to calculate the limiting autocorrelation demerit factors
for the stems from seeds f0 and g0 described in Proposition 2.8. It is easy to
calculate the sums of squares of the autocorrelation values for f0 and for g0 and
also to compute the sums of squares of the crosscorrelation values for f0 with fe0
and for g0 with ge0 to show that kf0 k44 = 4k(32k2 +1)/3, kg0 k44 = 4k(16k2 +5)/3,
kf0 fe0 k22 = 4k, and kg0 ge0 k22 = 4k(16k2 − 1)/3. Since f0 and g0 are Littlewood polynomials of length 4k, we have kf0 k22 = kg0 k22 = 4k, so that if
f0 , f1 , . . . and g0 , g1 , . . . are the stems obtained from seeds f0 and g0 by recursion
(1), then Corollary 2.7 tells us that limn→∞ ADF(fn ) = limn→∞ ADF(gn ) =
(16k2 − 9k + 2)/(9k), which is strictly increasing from a value of 1 (when k = 1)
to ∞ in the limit as k → ∞.
Proof of Proposition 2.8. We shall use Corollary 2.7 to calculate the limiting
crosscorrelation demerit factor. To that end, we calculate
(1 − z 4k )2 (1 − z)(1 − z 2 )
(1 − z)(1 − z 4 )
1 − z 4k
=
(1 − z 2 )(1 − z 4k )
1 − z4
f0 (z)g0 (z) =
= (1 − z 2 + z 4 − z 6 + · · · + z 4k−4 − z 4k−2 )(1 − z 4k ),
which is a polynomial with 4k nonzero coefficients, every of one of which is
either 1 or −1, so then kf0 g0 k22 = 4k.
And then we calculate
(1 − z 4k )2 (1 + z)(1 − z 2 )
(1 − z)(1 − z 4 )
1 − z 4k
=
(1 + z)2 (1 − z 4k )
1 − z4
f0 (z)e
g0 (z) =
= (1 + 2z + z 2 + · · · + z 4k−4 + 2z 4k−3 + z 4k−2 )(1 − z 4k ),
which is a polynomial with 4k coefficients of magnitude 1 and 2k coefficients of
magnitude 2, so then kf0 ge0 k22 = 4k + 4 · 2k = 12k.
And then we calculate
(1 − z 4k )2 (1 − z −4k )2 (1 − z −2 )2 (1 − z −1 )(1 + z −1 )
f0 (z)fe0 (z)g0 (z)e
g0 (z) =
(1 − z)(1 + z)(1 − z −4 )2
=
(1 − z 4k )2 (1 − z −4k )2 (1 − z −2 )2 (−z −1 )(z −1 )
(1 − z −4 )2
=−
=−
(1 − z 4k )2 (1 − z −4k )2 (1 − z −2 )2 z 2
(1 − z −4 )2 z 4
(1 − z 4k )2 (1 − z −4k )2 (1 − z 2 )(1 − z −2 )
.
(1 − z 4 )(1 − z −4 )
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
15
R
Thus f0 (z)fe0 (z)g0 (z)e
g0 (z) = −k(1 − z 4k )2 (1 − z 2 )/(1 − z 4 )k22 , and we have
R
already calculated the norm to be 4k, and so f0 (z)fe0 (z)g0 (z)e
g0 (z) = −4k.
Finally, kf0 k22 = kg0 k22 = 4k because f0 and g0 are Littlewood polynomials of
length 4k. Now Corollary 2.7 says that if fn (z) and gn (z) are defined recursively
for all n ∈ N by
fn+1 (z) = fn (z) + z len fn fn† (−z)
gn+1 (z) = gn (z) + z len gn gn† (−z),
then
R
2kf0 g0 k22 + kf0 e
g0 k22 + Re f0 fe0 g0 ge0
lim CDF(fn , gn ) =
n→∞
3kf0 k22 kg0 k22
2 · 4k + 12k + Re(−4k)
=
3(4k)2
1
.
=
3k
We close this section with the technical lemma used in the proof of Lemma
2.3 above.
Lemma 2.10. If a(z), b(z) ∈ C[z, z −1 ], k ∈ Z, and σ, τ ∈ {−1, 1}, and if
I=
then
1
a b + τ z keb
a + σz k e
2
I=
2kabk22
2
2
+ 2kaebk22 + 2στ Re
+
Z
2
1
,
e
a − σz k a eb − τ z k b
2
2
ae
abeb + 2στ Re
Z
z −2k ae
abeb.
Proof. Note that
Z
1 2
I=
|a| + |z k e
a|2 + 2σ Re az k e
a
|b|2 + |z keb|2 + 2τ Re bz ke
b
2
Z
1 2
az k a
|eb|2 + |z k b|2 − 2τ Re e
bz k b ,
|e
a| + |z k a|2 − 2σ Re e
+
2
and since we are integrating on the complex unit circle, we may omit terms of
the form |z k | and replace z k with z −k to obtain
I=
Z
2
2
−k
2
2
−k e
e
|a| + |e
a| + 2σ Re z ae
a
|b| + |b| + 2τ Re z bb
Z
1
|e
a|2 + |a|2 − 2σ Re z −k e
aa
|eb|2 + |b|2 − 2τ Re z −kebb ,
+
2
1
2
16
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
from which one obtains
Z
Z
a Re z −k beb
|a|2 + |e
a|2 |b|2 + |eb|2 + 4στ Re z −k ae
I=
Z
2
2
2
2
e
e
= kabk2 + ke
abk2 + kabk2 + ke
abk2 + 4στ Re z −k ae
a Re z −k beb
Z
2
2
e
= 2kabk2 + 2kabk2 + 4στ Re z −k ae
a Re z −k beb
Z h
i
2
2
e
Re z −k ae
az −k beb + Re z −k ae
az −k beb ,
= 2kabk2 + 2kabk2 + 2στ
where the third equality uses Lemma 2.2(ii) and the fourth equality uses the
observation that 2 Re(u) Re(v) = Re(uv) + Re(uv). The desired result now
readily follows.
3. Symmetry Groups
The expressions in Theorem 1.2 for the limiting autocorrelation and crosscorrelation demerit factors are invariant under certain symmetries. This helps
abbreviate computational searches for sequences and sequence pairs with optimum performance. These symmetries are based on negation of polynomials,
replacement of z by −z in polynomials, and transformation of polynomials to
their conjugate reciprocals. One should recall the notational conventions fe(z)
and f (z) for f (z) ∈ C[z, z −1 ] and the definition of the conjugate reciprocal
f †(z) for f (z) ∈ C[z] from the second paragraph of Section 2. One should note
that ffg(z) = fe(z)e
g (z) and f (z)g(z) = f (z)·g(z) for every f (z), g(z) ∈ C[z, z −1 ]
†
and (f (z)g(z)) = f † (z)g† (z) for every f (z), g(z) ∈ C[z]. (The third relation
follows easily from Lemma 2.1(i) and the second relation.) We shall also need
the following observation.
Lemma 3.1. For any f (z) ∈ C[z] and any p ∈ R with p ≥ 1, we have
kf † (z)kp = kf (z)kp .
Proof. By Lemma 2.1(i), we have
Z 2π
Z 2π
p
1
1
|f † (eiθ )|p dθ =
(eiθ )deg f f (eiθ ) dθ
2π 0
2π 0
Z 2π
1
=
|f (eiθ )|p dθ.
2π 0
Now we introduce a group of symmetries that, when applied to a polynomial
f0 (z) ∈ C[z] with nonzero constant coefficient, will preserve the autocorrelation
properties of the stem f0 , f1 , . . . of Rudin-Shapiro-like polynomials obtained
from seed f0 via recursion (1). First we describe our group and how it affects
certain norms, and then we show its effect on autocorrelation as a corollary.
Proposition 3.2. Let ℓ be a nonnegative integer, and let Pℓ be the set of all
polynomials of length ℓ in C[z] that have nonzero constant coefficient. We define
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
17
three maps from Pℓ to itself:
n(f ) = −f
h(f ) = fe
r(f ) = f † .
These maps generate a group Gℓ = hn, h, ri of permutations of Pℓ .
(i). If ℓ = 1, then G1 is the internal direct product of the two cyclic groups
hni and hri, each of order 2.
(ii). If ℓ is odd with ℓ > 1, then Gℓ is the internal direct product of the three
cyclic groups hni, hhi, and hri, each of order 2.
(iii). If ℓ is even, then Gℓ is isomorphic to the dihedral group of order 8
generated by rh and h, where rh is of order 4, h is of order 2, and
h(rh)h−1 = (rh)−1 .
For any t ∈ Gℓ , any f ∈ Pℓ , and any p ≥ 1, we have
kt(f )kp = kf kp
g)kp = kf fekp .
kt(f )t(f
Proof. It is clear that each of n, h, and r is an involution on Pℓ (except that
h is the identity element when ℓ = 1), so these maps generate a group of
permutations of Pℓ . If ℓ = 1, then it is not hard to show that eπi/4 ∈ P1 that
has four distinct images under Gℓ = hn, h, ri, so Gℓ has order at least 4. If
ℓ > 1 and ℓ is odd, then it is not hard to show that 1 + z + iz ℓ−1 ∈ Pℓ has
eight distinct images under Gℓ = hn, h, ri. If ℓ is even, then it is not hard to
show that eπi/6 + eπi/3 z ℓ−1 ∈ Pℓ has eight distinct images under Gℓ = hn, h, ri.
So Gℓ has order at least 8. Furthermore n commutes with both h and r, and
Lemma 2.1(ii) shows that hr = rh when ℓ is odd, but hr = nrh when ℓ is even.
Thus if ℓ = 1, then G1 = hn, h, ri = hn, ri is a group of order at least
4, generated by commuting involutions n and r. So G1 is the internal direct
product of hni and hri, which are both cyclic groups of order 2.
If ℓ is odd and greater than 1, our group Gℓ = hn, h, ri is a group or order at
least 8 generated by commuting involutions n, h, and r. So Gℓ is the internal
direct product of hni, hhi, and hri, which are three cyclic groups each of order
2.
On the other hand, if ℓ is even, then rh can be shown to have (rh)2 = n,
(rh)3 = nrh = hr, and (rh)4 the identity, and since the powers of rh take the
element 1+iz ℓ−1 ∈ Pℓ to four distinct elements, we see that rh has order 4. Thus
Gℓ = hn, h, ri = hh, ri = hh, rhi. Then note that h(rh)h−1 = hr = (rh)3 =
(rh)−1 , and so it can be seen that Gℓ = hh, rhi is generated by an element h or
order 2 and an element y = rh of order 4 that satisfy the relation hyh−1 = y −1 .
These are the relations satisfied by the generators of the dihedral group D
of order 8, that is, the group of symmetries of a square (with a 90◦ rotation
corresponding to y and a flip corresponding to h). So Gℓ is a homomorphic
image of D, but since Gℓ has order at least 8, we must have Gℓ ∼
= D.
18
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
To verify that kt(f )kp = kf kp for any f ∈ Pℓ , t ∈ Gℓ , and p ≥ 1, it suffices
to check that it holds when t is one of the generators n, h, and r. When t = n,
this is clear, and when t = h or r, it is a consequence, respectively, of Lemma
2.2(ii) or Lemma 3.1.
g)kp = kf fekp for any f ∈ Pℓ , t ∈ Gℓ , and
Similarly, to verify that kt(f )t(f
p ≥ 1, it suffices to check that it holds when t is one of the generators n, h, and
r. When t = n or h, this is clear, and when t = r, then
g
†
g)kp = kf † (f
kr(f )r(f
0 0 )kp
= kf0† (fe0 )† kp or k−f0†(fe0 )† kp
= k(f0 fe0 )† kp
= kf0 fe0 kp ,
where the second equality uses Lemma 2.1(ii), and the fourth equality uses
Lemma 3.1.
Corollary 3.3. Let f0 ∈ C[z] be a polynomial of length ℓ with nonzero constant
coefficient, let t be an element of the group Gℓ described in Proposition 3.2, and
let a0 = t(f0 ). If f0 , f1 , . . . and a0 , a1 , . . . are sequences of Rudin-Shapiro-like
polynomials generated from f0 and a0 via recursion (1), then
lim ADF(an ) = lim ADF(fn ).
n→∞
n→∞
Proof. By Theorem 1.2, we have
lim ADF(fn ) = −1 +
n→∞
2 kf0 k44 + kf0 fe0 k22
,
·
3
kf0 k42
but Proposition 3.2 shows that the values of the three norms occurring on the
right hand side do not change if we replace every instance of f0 with t(f0 ) = a0 ,
which changes the right hand side to limn→∞ ADF(an ) by Theorem 1.2.
Now we introduce a group of symmetries that, when applied to a pair of
polynomials (f0 (z), g0 (z)) from C[z], will preserve the crosscorrelation properties of the stems f0 , f1 , . . . and g0 , g1 , . . . of Rudin-Shapiro-like polynomials
obtained from seeds f0 and g0 via recursion (1). First we describe the group
and how it affects certain norms and integrals, and then we show its effect on
crosscorrelation as a corollary.
Proposition 3.4. Let ℓ be a nonnegative integer, and let Pℓ be the set of all
polynomials of length ℓ in C[z] that have nonzero constant coefficient. We define
four maps from Pℓ × Pℓ to itself:
s(f, g) = (g, f )
n(f, g) = (−f, g)
h(f, g) = (fe, ge)
r(f, g) = (f † , g† ).
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
19
These maps generate a group Gℓ,ℓ = hs, n, h, ri of permutations of Pℓ × Pℓ . Gℓ,ℓ
contains a dihedral subgroup D of order 8 generated ns and s, where ns has
order 4, s has order 2, and s(ns)s−1 = (ns)−1 .
(i). If ℓ = 1, then Gℓ,ℓ is the internal direct product of the dihedral group D
of order 8 and the cyclic group hri of order 2.
(ii). If ℓ is odd and ℓ > 1, Gℓ,ℓ is the internal direct product of the dihedral
group D of order 8, the cyclic subgroup hhi of order 2, and the cyclic
subgroup hri of order 2.
(iii). If ℓ is even, then Gℓ,ℓ is the internal central product of D and another
dihedral subgroup ∆ of order 8 generated by rh and h, where rh has order
4, h has order 2, and h(rh)h−1 = (rh)−1 . Thus Gℓ,ℓ is isomorphic to the
extraspecial group of order 25 of + type, which is also the inner holomorph
of the dihedral group of order 8.
For any t ∈ Gℓ,ℓ , any f, g ∈ Pℓ , and any p ≥ 1, let (a, b) = t(f, g), and then we
have
kakp kbkp = kf kp kgkp
kabkp = kf gkp
kaebkp = kf e
g kp
Z
e
g.
Re ae
abb = Re f fege
Z
Proof. It is clear that each of s, n, h, and r is an involution on Pℓ × Pℓ (except
that h is the identity element when ℓ = 1), so these maps generate a group
of permutations of Pℓ × Pℓ . Furthermore, s commutes with h and r, and n
also commutes with h and r. Thus we can better understand our group Gℓ,ℓ
by focusing on two subgroups, hs, ni and hh, ri, with the knowledge that every
element from the former subgroup commutes with every element of the latter
subgroup.
Let us first focus on the subgroup hs, ni of Gℓ,ℓ . We note that (sn)(f, g) =
(g, −f ) but (ns)(f, g) = (−g, f ), so that s and n do not commute. We define
x = ns, and then hs, ni = hs, nsi = hs, xi. We note that x is an element of
order 4 with x2 (f, g) = (−f, −g) = −(f, g) and x3 (f, g) = (g, −f ) = (sn)(f, g).
Then note that sxs−1 = snss−1 = sn = x3 = x−1 . Thus hs, ni = hs, xi must be
a homomorphic image of a dihedral group of order 8, the group of symmetries
of a square (with a 90◦ rotation corresponding to x and a flip corresponding to
s). And in fact, one can show that the (1 + z ℓ−1 , i + iz ℓ−1 ) ∈ Pℓ × Pℓ has eight
distinct images under the action of hs, ni, so hs, xi ∼
= D.
Now let us focus on the subgroup hh, ri of Gℓ,ℓ . If ℓ = 1, then h is the identity
element, so hh, ri = hri is a cyclic group of order 2. Since the elements of hs, ni
commute with the elements of hh, ri, this means that G1,1 is a homomorphic
image of a direct product of a dihedral group D of order 8 and a cyclic group
C of order 2. If ℓ = 1, then it is not hard to show that (1, eπi/4 ) ∈ P1 × P1 that
has 16 distinct images under G1,1 = hn, h, ri, so G1,1 ∼
= D × C.
20
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
Now suppose that ℓ > 1. Lemma 2.1(ii) shows that hr = rh when ℓ is odd,
but that (hr)(f, g) = −(rh)(f, g) when ℓ is even. Note that the group element
that maps (f, g) to (−f, −g) is x2 , described in the previous paragraph. So h
and r commute when ℓ is odd, but hr = x2 rh when ℓ is even.
So if ℓ > 1 and ℓ is odd, then hh, ri is a homomorphic image of a Klein
four-group, that is, of C × C with C a cyclic group of order 2. And in fact one
can show that (1 + z − z ℓ−1 , 1 + z − z ℓ−1 ) has four distinct images under the
action of hh, ri, so hh, ri ∼
= C × C.
On the other hand, if ℓ is even, then rh can be shown to have (rh)2 = x2 ,
(rh)3 = x2 rh = hr, and (rh)4 the identity, and since the powers of rh take the
element (1 + iz ℓ−1 , 1 + iz ℓ−1 ) ∈ Pℓ × Pℓ to four distinct elements, we see that rh
has order 4. Note that h(rh)h−1 = hr = (rh)3 = (rh)−1 , and so it can be seen
that hh, ri = hh, rhi is generated by an element h or order 2 and an element
y = rh of order 4 that satisfy the relation hyh−1 = y −1 . Thus if ℓ is even, then
hr, hi is a homomorphic image of the dihedral group D of order 8, the group
of symmetries of a square (with a 90◦ rotation corresponding to y and a flip
corresponding to h). One can show that (1 + iz ℓ−1 , 1 + iz ℓ−1 ) has eight distinct
images under the action of hh, ri, so that hh, ri ∼
= D.
Now we assemble what we have learned about the subgroups hs, ni and hh, ri
of Gℓ,ℓ using the fact that every element in the former subgroup commutes with
every element in the latter. We again separate into cases depending on the
parity of ℓ.
If ℓ is odd and greater than 1, we saw that hh, ri is a Klein four-group
generated by the elements h and r, each of order 2. We saw that hs, ni is a
dihedral group of order 8. So then the group Gℓ,ℓ = hs, n, h, ri is a homomorphic
image of D ×C ×C, where D is the dihedral group of order 8 and C is the cyclic
group of order 2. It is not hard to show that (eπi/6 + eπi/3 z ℓ−1 , 1 + z + z ℓ−1 ) ∈
Pℓ × Pℓ has 32 distinct images under Gℓ,ℓ = hn, h, ri, so Gℓ,ℓ has order at least
32. But since Gℓ,ℓ has at least 32 elements, we must have Gℓ,ℓ ∼
= D × C × C.
On the other hand, if ℓ is even, then hh, ri is a dihedral group of order 8
generated by element y = rh of order 4 and element h of order 2. We saw
that hs, ni is a dihedral group of order 8 generated by element x = ns of order
4 and element s of order 2. The group hs, ni = hx, si has a center hx2 i of
order 2. The group hh, ri = hy, hi has a center hy 2 i, and we observed above
that y 2 = (rh)2 = x2 , so the centers of hn, si and hh, ri completely overlap
with each other. So the group Gℓ,ℓ = hs, n, h, ri is a homomorphic image of
the central product of two dihedral groups of order 8, which makes Gℓ,ℓ a
homomorphic image of the extraspecial group of order 25 of + type. It is not
hard to show that (eπi/6 + eπi/3 z ℓ−1 , 1 + z ℓ−1 ) ∈ Pℓ × Pℓ has 32 distinct images
under Gℓ,ℓ = hn, h, ri, so Gℓ,ℓ has order at least 32. But since Gℓ,ℓ has at least
32 elements, Gℓ,ℓ must be isomorphic to the the extraspecial group of order 25
of + type, which is also the inner holomorph of the dihedral group of order 8.
Now we verify the four invariance relations for Gℓ,ℓ in the statement of this
proposition. It suffices to check these relations when the group element t ∈ Gℓ,ℓ
is one of the four generators s, n, h, or r of the group. For the rest of this
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
21
proof, we let (f, g) be an arbitrary pair in Pℓ × Pℓ , let p be a real number with
p ≥ 1, let t be an element of {s, n, h, r}, and we set (a, b) = t(f, g).
It is clear that kakp kbkp = kf kp kgkp when t = s or n, and when t = h or r,
this is a consequence, respectively, of Lemma 2.2(ii) or Lemma 3.1.
It is clear that kabkp = kf gkp when t = s or n, and when t = h or r, then
this is a consequence, respectively, of Lemma 2.2(ii) or Lemma 3.1.
Now we verify that kaebkp = kf e
g kp . This is clear when t = n, and when
f
e
e
t = s, then kabkp = kg f kp = kf e
gkp , which equals kf e
gkp by Lemma 2.2(ii).
When t = r, then
g
† )k
kaebkp = kf † (g
p
= kf † (e
g )† kp or k−f †(e
g )† kp
= k(f e
g)† kp
= kf e
gkp ,
where Lemmata 2.1(ii) and 3.1 are used in the second and fourth equalities.
R
R
g . When t = n or h, this is
Now we verify that Re ae
abeb = Re f fege
R
R
e
e which is equal
clear, and when t = s, we see that Re ae
abb = Re ge
g f f,
R
to Re f fege
g , since conjugation of the integrand does not change the real part
of the integral. Finally, if t = r, then
Z
Z
e
Re ae
abb = Re f † fe† g † ge†
Z
= Re f † (fe)† g† (e
g )†
Z
= Re (f fe)† (ge
g )†
Z
= Re
z −2 deg f f fe z −2 deg g ge
g
Z
g
= Re f fege
Z
= Re f fege
g,
where the second equality uses Lemma 2.1(ii) (and the fact that f and g are
assumed to have the same degree), the fourth equality uses Lemma 2.1(iii), the
fifth equality uses the fact that z = z −1 on the complex unit circle, and the
last equality uses the fact that conjugation of the integrand does not change
the real part of the integral.
Corollary 3.5. Let f0 , g0 ∈ C[z] be a polynomials of length ℓ with nonzero constant coefficients, let t be an element of the group Gℓ,ℓ be the group described
in Proposition 3.4, and let (a0 , b0 ) = t(f0 , g0 ). If f0 , f1 , . . . and g0 , g1 , . . . and
22
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
a0 , a1 , . . . and b0 , b1 , . . . are sequences of Rudin-Shapiro-like polynomials generated from f0 , g0 , a0 , and b0 via recursion (1), then
lim CDF(an , bn ) = lim CDF(fn , gn ) and
n→∞
n→∞
lim PSC(an , bn ) = lim PSC(fn , gn ).
n→∞
n→∞
Proof. By Theorem 1.2, we have
R
g0
2kf0 g0 k22 + kf0 ge0 k22 + Re f0 fe0 g0 e
lim CDF(fn , gn ) =
,
2
2
n→∞
3kf0 k2 kg0 k2
but Proposition 3.4 shows that the values of the three terms in the numerator
and the value of the denominator of the right hand side do not change if we
replace every instance of f0 with a0 and every instance of g0 with b0 . These
replacements change the right hand side to limn→∞ CDF(an , bn ) by Theorem
1.2.
Because of the structure of the groups Gℓ and Gℓ,ℓ described in Propositions
3.2 and 3.4 above, one can say that there exist u, v ∈ Gℓ such that either
(a0 , b0 ) = (u(f0 ), v(g0 )) or (a0 , b0 ) = (v(g0 ), u(f0 )). Thus by Corollary 3.3,
lim ADF(an ) ADF(bn ) = lim ADF(fn ) ADF(gn ),
n→∞
n→∞
and so, considering the formula (7) for the Pursley-Sarwate Criterion, we see
that
lim PSC(an , bn ) = lim PSC(fn , (gn ).
n→∞
n→∞
4. Some Examples of Pairs of Rudin-Shapiro-Like Sequences with
Low Correlation
For each ℓ ≤ 52, we considered every possible Littlewood polynomial f0
of length ℓ, and used computers, including opportunistic use of distributed
resources through the Open Science Grid [14, 20], to calculate via Corollary 2.5
(a result originally due to Borwein and Mossinghoff [2, Theorem 1]) the limiting
autocorrelation demerit factor of the stem f0 , f1 , . . . constructed from f0 via
our recursion (1). For each length ℓ, we report on Table 1 the lowest limiting
autocorrelation demerit factor achieved, and indicate how many seeds achieve
this minimum value. Seeds that are equivalent modulo the action of the group
Gℓ described in Proposition 3.2 always have the same limiting autocorrelation
demerit factor by Corollary 3.3, and so we group seeds into orbits under the
action of Gℓ and report how many distinct orbits there are on Table 1. For each
length ℓ, we give one example of a seed f0 whose stem achieves the smallest
limiting autocorrelation demerit factor.
Example seeds are reported using a hexadecimal code. To decode, expand
each hexadecimal digit into binary form (0 → 0000, 1 → 0001, . . ., F → 1111)
and, if necessary, remove initial 0 symbols to obtain a binary sequence of the
appropriate length. Then convert each 0 to +1 and each 1 to −1 to obtain the
list of coefficients of the seed f0 . For example, Table 2 reports for length ℓ = 14
that one seed of interest is 149B. Expand to 0001 0100 1001 1011 and delete
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
23
the initial two zeroes to obtain a sequence 01 0100 1001 1011 of length ℓ = 14.
Convert from 0, 1 to ±1 to obtain the coefficients of
g0 (z) = 1 − z + z 2 − z 3 + z 4 + z 5 − z 6 + z 7 + z 8 − z 9 − z 10 + z 11 − z 12 − z 13 .
Borwein-Mossinghoff [2, Corollary 1] proved that the limiting autocorrelation demerit factor for any sequence f0 , f1 , . . . of Rudin-Shapiro-like Littlewood
polynomials generated from a seed f0 using recursion (1) can never be less than
1/3 (see also our Theorem 1.2). Their computer experiments show that for
seeds of length ℓ ≤ 40, there exist seeds whose stems achieve limiting autocorrelation demerit factor 1/3 if and only if ℓ ∈ {1, 2, 4, 8, 16, 20, 32, 40}. We
also discovered seeds of length 52 whose stems achieve limiting autocorrelation
demerit factor 1/3. The first and third authors have now proved [11] that a
seed f0 of length ℓ > 1 produces a stem with limiting autocorrelation demerit
factor 1/3 if and only if f0 is the interleaving of the two sequences of some Golay complementary pair. This explains why both Borwein and Mossinghoff’s
searches and ours produced seeds with optimal asymptotic autocorrelation at
the lengths that we have observed. For lengths ℓ ≤ 40, where there are seeds
with optimal asymptotic correlation, Borwein and Mossinghoff also indicate
how many distinct optimal seeds there are for each such length. Our computer
experiments agree with theirs, but we also present on Table 1 the minimum
limiting autocorrelation demerit factors for all lengths ℓ ≤ 52, regardless of
whether or not the minimum is 1/3.
Now let us also consider crosscorrelation. In view of Proposition 2.8 and
Remark 2.9, we already know we can achieve limiting crosscorrelation demerit
factors as close to 0 as we like, but we have observed that pairs of stems with
very low limiting crosscorrelation demerit factor tend to have very high autocorrelation demerit factor, and are therefore of little practical value. This is not
surprising, given the bound (8) of Pursley and Sarwate. It is much more enlightening to ask how low one can make the limiting Pursley-Sarwate Criterion
(7), which combines both autocorrelation and crosscorrelation performance.
For each ℓ ≤ 28, we considered every possible pair of Littlewood polynomials (f0 , g0 ) of length ℓ, and used computers, including opportunistic use of
distributed resources through the Open Science Grid [14, 20], to calculate via
Theorem 1.2 the limiting crosscorrelation demerit factors of the pair of stems
(f0 , f1 , . . . ; g0 , g1 , . . .) constructed from our recursion (1). We also calculate the
limiting autocorrelation demerit factors for each of the two stems, and from
all three of these limits, we obtain the limiting Pursley-Sarwate Criterion. Table 2 records the lowest limiting Pursley-Sarwate Criterion achieved for each
ℓ ≤ 28, and records the seed pairs (f0 , g0 ) that give rise to the pairs of stems
that achieve this minimum. Seed pairs that are equivalent modulo the action of
the group Gℓ,ℓ described in Proposition 3.4 will always have the same limiting
Pursley-Sarwate Criterion by Corollary 3.5, and so we group seeds pairs into
orbits under the action of Gℓ,ℓ . We report one representative of each class on
Table 2 using our hexadecimal code (described above in the discussion of Table
1) and also report the size of the orbit. For some lengths there are multiple
24
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
Table 1. Lowest Limiting Autocorrelation Demerit Factor for
Seeds of Each Length
seed
limiting
number of number of
length
ADF(fn )
sequences
orbits
f0
= 0.3333 . . .
2
1
2
= 0.3333 . . .
4
1
0
= 0.6296 . . .
8
2
0
= 0.3333 . . .
8
1
1
= 0.5466 . . .
24
4
01
= 0.6296 . . .
56
8
01
= 0.4965 . . .
56
8
03
= 0.3333 . . .
32
4
06
= 0.4650 . . .
144
18
006
= 0.5466 . . .
504
64
006
= 0.4435 . . .
168
22
01C
= 0.4074 . . .
96
12
036
= 0.4280 . . .
344
44
0036
= 0.4965 . . .
2648
332
0036
= 0.4162 . . .
688
86
0163
= 0.3333 . . .
192
24
0359
= 0.4071 . . .
1472
184
001C9
= 0.4650 . . .
12992
1624
001C9
= 0.3998 . . .
784
98
00793
= 0.3333 . . .
128
16
05239
= 0.3938 . . .
1312
164
000F19
= 0.4435 . . .
35352
4420
000F19
= 0.3887 . . .
1696
212
0066B4
= 0.3518 . . .
320
40
00CD69
= 0.3845 . . .
2176
272
000CD29
= 0.4280 . . .
104920
13116
0007866
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
1
3
1
3
17
27
1
3
41
75
17
27
73
147
1
3
113
243
41
75
161
363
11
27
217
507
73
147
281
675
1
3
353
867
113
243
433
1083
1
3
521
1323
161
363
617
1587
19
54
721
1875
217
507
sample
equivalence classes that achieve the same minimum limiting Pursley-Sarwate
Criterion: each such class has its own line on the table.
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
25
Table 1. (continued) Lowest Limiting Autocorrelation Demerit Factor for Seeds of Each Length
seed
limiting
number of number of
length
ADF(fn )
sequences
orbits
f0
= 0.3808 . . .
1888
236
006B274
= 0.3605 . . .
512
64
00DB171
= 0.3777 . . .
3040
380
000E4B8D
= 0.4162 . . .
266688
33336
0006C729
= 0.3749 . . .
6368
796
001E2D33
= 0.3333 . . .
1536
192
003C5A66
= 0.3725 . . .
10400
1300
0003C5A66
= 0.4071 . . .
554752
69344
0003C5A66
= 0.3703 . . .
1216
152
001F1699C
= 0.3580 . . .
640
80
0034EC5A6
= 0.3683 . . .
1760
220
00035AC726
= 0.3998 . . .
840256
105032
00034E94E6
= 0.3666 . . .
4416
552
0019E2D2B3
= 0.3333 . . .
1088
136
0033C5A566
= 0.3650 . . .
7328
916
00033C5A566
= 0.3938 . . .
1589568
198696
0001E5A3599
= 0.3636 . . .
2592
324
0015B878CCB
= 0.3443 . . .
256
32
00178B4B326
= 0.3623 . . .
2272
284
0001E9663D33
= 0.3887 . . .
2690528
336316
0000FC31E199
= 0.3610 . . .
2752
344
0006E529E49C
= 0.3379 . . .
128
16
003C3315A9A6
= 0.3599 . . .
2720
340
0000F30F4A665
= 0.3845 . . .
3751392
468924
0000DD83C6696
= 0.3589 . . .
1536
192
0006F1C6D2372
= 0.3333 . . .
64
8
00C3CC459A96A
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
833
2187
53
147
953
2523
281
675
1081
2883
1
3
1217
3267
353
867
1361
3675
29
81
1513
4107
433
1083
1673
4563
1
3
1841
5043
521
1323
2017
5547
125
363
2201
6075
617
1587
2393
6627
73
216
2593
7203
721
1875
2801
7803
1
3
sample
In Table 3 we also present sequences with low limiting Pursley-Sarwate Criterion. For a given length ℓ ≤ 52, we first found all the seeds that produce
stems whose autocorrelation demerit factors reach the minimum value for that
26
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
Table 2. Lowest Limiting Pursley-Sarwate Criterion for Seeds
of Each Length
seed
limiting values as n → ∞
orbit
seeds
length
PSC(fn , gn )
ADF(fn )
ADF(gn )
CDF(fn , gn )
size
f0
g0
1
1.6666 . . .
0
8
0
1
3
1.3703 . . .
32
0
1
4
1.1666 . . .
16
1
2
5
1.3466 . . .
32
01
02
5
1.3466 . . .
32
01
08
5
1.3466 . . .
32
01
0D
6
1.2962 . . .
32
02
0D
6
1.2962 . . .
32
04
0B
7
1.2312 . . .
32
04
1A
8
1.1666 . . .
32
06
3A
8
1.1666 . . .
32
12
2E
9
1.2057 . . .
32
009
035
10
1.1733 . . .
32
04D
0A1
11
1.1818 . . .
32
032
251
12
1.1666 . . .
32
065
6A3
13
1.1734 . . .
32
00CA 03AD
13
1.1734 . . .
32
00CA 0907
14
1.1836 . . .
4
3
1
1
20
27
5
6
4
5
4
5
4
5
2
3
2
3
36
49
5
6
5
6
20
27
47
75
268
363
41
54
116
169
116
169
101
147
0
1.3333 . . .
1
3
1
3
17
27
1
3
41
75
41
75
41
75
17
27
17
27
73
147
1
3
1
3
113
243
41
75
161
363
11
27
281
507
281
507
73
147
4
2
1
3
1
3
17
27
1
3
41
75
41
75
41
75
17
27
17
27
73
147
1
3
1
3
113
243
41
75
161
363
11
27
217
507
217
507
73
147
32
0071 149B
length, as reported in Table 1. Then we compute which pairs of these seeds
(f0 , g0 ) produce pairs of stems with the lowest limiting crosscorrelation demerit
factor (and therefore the lowest limiting Pursley-Sarwate Criterion, since they
all have the same limiting autocorrelation demerit factors). Seed pairs that are
equivalent modulo the action of the group Gℓ,ℓ described in Proposition 3.4 will
always have the same limiting Pursley-Sarwate Criterion by Corollary 3.5, and
so we group seeds pairs into orbits under the action of Gℓ,ℓ . We report one
representative of each class on Table 3 using our hexadecimal code (described
above in the discussion of Table 1) and also report the size of the orbit. For
some lengths there are multiple equivalence classes that achieve the same minimum limiting Pursley-Sarwate Criterion: each such class has its own line on
the table.
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
27
Table 2. (continued) Lowest Limiting Pursley-Sarwate Criterion for Seeds of Each Length
seed
limiting values as n → ∞
orbit
seeds
length
PSC(fn , gn )
ADF(fn )
ADF(gn )
CDF(fn , gn )
size
f0
g0
15
1.1546 . . .
15C3
32
0A36
11D2
17
1.1407 . . .
32
0038D
0EE96
18
1.1481 . . .
32
0039A
0E8F6
19
1.1559 . . .
32
00E4D
38A16
19
1.1559 . . .
32
0C56D
0E013
20
1.1363 . . .
32
08FA6
5A230
21
1.1338 . . .
32
00F765
05DAF3
22
1.1515 . . .
32
0188B5
1341DE
22
1.1515 . . .
32
022D85
0C74FD
23
1.1203 . . .
32
0BA421
376A38
24
1.1292 . . .
32
02B25C
7A8C2C
25
1.1372 . . .
32
00A9273 0BFC9C7
26
1.1350 . . .
32
075D9AD 1ACFF83
27
1.1323 . . .
32
014E48A 03A3DE6
28
1.1258 . . .
52
75
37
48
212
289
166
243
788
1083
788
1083
113
150
316
441
257
363
257
363
1100
1587
13
18
92
125
343
507
172
243
71
98
024E
1.1041 . . .
23
45
1
3
353
867
113
243
497
1083
497
1083
1
3
65
147
161
363
161
363
617
1587
7
18
721
1875
83
169
299
729
59
147
32
16
281
675
1
3
353
867
113
243
433
1083
433
1083
11
25
521
1323
161
363
161
363
745
1587
23
54
157
375
217
507
961
2187
59
147
32
09467C5 60EA253
One can compare the results on Table 2 (which reports the minimum limiting
Pursley-Sarwate Criterion over all pairs of stems of a given length) with the
results of Table 3 (which reports the minimum limiting Pursley-Sarwate Criterion only over pairs of stems whose limiting autocorrelation demerit factors
equal the minimum value for that length). In some cases (lengths 13, 15, 19,
20, 21, and 23 through 28) the limiting Pursley-Sarwate Criterion reported on
Table 2 is lower. We were able to go to much greater lengths in Table 3 because
the computational burden is greatly reduced when we restrict the calculations
of crosscorrelation properties to only those pairs of seeds that produce minimum asymptotic autocorrelation demerit factors. The lowest asymptotic value
for the Pursley-Sarwate Criterion we discovered was 331/300 = 1.10333 . . . for
ℓ = 40 on Table 3. Our construction can be compared to randomly selected
long binary sequences, which typically have Pursley-Sarwate Criterion of about
2, and to high-performance sequence pairs constructed from finite field characters that have asymptotic Pursley-Sarwate Criterion of 7/6 (see [10, §II.E,
§IV.D] and [1, eq.(6)]).
28
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
Our computations of limiting crosscorrelation demerit factors in Tables 2
and 3 use a fast Fourier transform algorithm to speed up the convolutions
(Laurent polynomial multiplications) that appear in the formula for asymptotic
crosscorrelation demerit factor in Theorem 1.2. Because these calculations are
performed using floating point arithmetic, there are small rounding errors. We
checked that the approximate values of the norms and integrals in our formula
were always very close to integers: all discrepancies were less than 2 · 10−11 .
Table 3. Lowest Limiting Pursley-Sarwate Criterion among
Seed Pairs that Have the Lowest Limiting Autocorrelation Demerit Factor
seed
limiting values as n → ∞
orbit
seeds
length
PSC(fn , gn )
ADF(fn )
ADF(gn )
CDF(fn , gn )
size
f0
g0
1
1.6666 . . .
0
8
0
1
3
1.3703 . . .
32
0
1
4
1.1666 . . .
16
1
2
5
1.3466 . . .
32
01
02
5
1.3466 . . .
32
01
08
5
1.3466 . . .
32
01
0D
6
1.2962 . . .
32
02
0D
6
1.2962 . . .
32
04
0B
7
1.2312 . . .
32
04
1A
8
1.1666 . . .
32
06
3A
8
1.1666 . . .
32
12
2E
9
1.2057 . . .
32
009
035
10
1.1733 . . .
32
04D
0A1
11
1.1818 . . .
32
032
251
12
1.1666 . . .
32
065
6A3
13
1.1775 . . .
32
01DB
0D47
14
1.1836 . . .
32
0071
149B
15
1.1688 . . .
32
01AC
245C
16
1.1041 . . .
32
0A36
11D2
17
1.1407 . . .
32
0038D
0EE96
18
1.1481 . . .
4
3
1
1
20
27
5
6
4
5
4
5
4
5
2
3
2
3
36
49
5
6
5
6
20
27
47
75
268
363
41
54
380
507
101
147
508
675
37
48
212
289
166
243
0
1.3333 . . .
1
3
1
3
17
27
1
3
41
75
41
75
41
75
17
27
17
27
73
147
1
3
1
3
113
243
41
75
161
363
11
27
217
507
73
147
281
675
1
3
353
867
113
243
4
2
1
3
1
3
17
27
1
3
41
75
41
75
41
75
17
27
17
27
73
147
1
3
1
3
113
243
41
75
161
363
11
27
217
507
73
147
281
675
1
3
353
867
113
243
32
0039A
0E8F6
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
29
Table 3. (continued) Lowest Limiting Pursley-Sarwate Criterion among Seed Pairs that Have the Lowest Limiting Autocorrelation Demerit Factor
limiting values as n → ∞
seed
orbit
seeds
length
PSC(fn , gn )
ADF(fn )
ADF(gn )
CDF(fn , gn )
size
f0
g0
19
1.1643 . . .
0C549
16
05239
36E0A
21
1.1405 . . .
32
001C9A
063EAD
22
1.1515 . . .
32
0188B5
1341DE
22
1.1515 . . .
32
022D85
0C74FD
23
1.1424 . . .
16
00ECB6
1C312A
23
1.1424 . . .
16
08643E
14B9A2
24
1.1296 . . .
16
032695
03CE6A
25
1.1546 . . .
32
003B3CB
04E50A2
25
1.1546 . . .
32
01FAE32
0C42A69
26
1.1360 . . .
32
042347C
0A6B813
27
1.1545 . . .
32
0A109EC
3E2ACD6
28
1.1326 . . .
32
071DBB5
3B82B6D
29
1.1625 . . .
32
0089E14E
064BADE8
29
1.1625 . . .
32
013A6A2D
021C14EC
30
1.1288 . . .
32
02A6CF21
10164AC7
31
1.1394 . . .
32
067E2CAB
13AF5F63
32
1.1041 . . .
32
0A363905
11D21EDD
32
1.1041 . . .
32
1EDD11D2
39050A36
33
1.1303 . . .
32
00C03B652
0B5B9C517
34
1.1107 . . .
32
00598B0ED
0DEE9382B
35
1.1746 . . .
32
00963532E
1E280FB3B
36
1.1872 . . .
16
1D603A324
7190953ED
36
1.1872 . . .
276
361
121
150
988
1323
257
363
257
363
52
69
52
69
7
9
1444
1875
1444
1875
359
507
188
243
227
294
660
841
660
841
481
675
2204
2883
37
48
37
48
2476
3267
610
867
2956
3675
403
486
403
486
00F26
1.14
433
1083
1
3
521
1323
161
363
161
363
617
1587
617
1587
19
54
721
1875
721
1875
217
507
833
2187
53
147
953
2523
953
2523
281
675
1081
2883
1
3
1
3
1217
3267
353
867
1361
3675
29
81
29
81
32
20
433
1083
1
3
521
1323
161
363
161
363
617
1587
617
1587
19
54
721
1875
721
1875
217
507
833
2187
53
147
953
2523
953
2523
281
675
1081
2883
1
3
1
3
1217
3267
353
867
1361
3675
29
81
29
81
16
1DA30602B
7EACA6F12
30
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
Table 3. (continued) Lowest Limiting Pursley-Sarwate Criterion among Seed Pairs that Have the Lowest Limiting Autocorrelation Demerit Factor
seed
length
limiting values as n → ∞
orbit
seeds
PSC(fn , gn )
ADF(fn )
ADF(gn )
CDF(fn , gn )
size
f0
g0
37
1.1796 . . .
0636C1F2AA
16
0223D0E7AE
04164BD222
38
1.1209 . . .
32
0210F456C9
1A2E842E73
39
1.1477 . . .
32
0019E3352D
07819B4CAA
39
1.1477 . . .
32
166AAF0FCC
195A5FFCC3
40
1.1033 . . .
16
0033C66A5A
0F03369955
40
1.1033 . . .
16
33F0F55669
3CC005A566
41
1.1273 . . .
32
00033C66A5A
00F03369955
42
1.1171 . . .
32
01A5024EE35
0978ED8A2B1
43
1.1474 . . .
16
19685754CBC
3433FDFA1E6
43
1.1474 . . .
16
1BC85754E1C
3693FDFA346
44
1.1639 . . .
16
0BE7818814C
67D44D4B285
45
1.1675 . . .
32
01A1CE2AEC0D
097C919FF6BC
46
1.1109 . . .
32
00F34581196A
1EF59BD8C174
47
1.1475 . . .
16
0006E529E49C
363960F91AAA
47
1.1475 . . .
16
0886CF03E414
3EB94AD31A22
48
1.1643 . . .
16
235B4408706E
235B45778F91
48
1.1643 . . .
16
3185A800F4D9
3185AABF0B26
49
1.1690 . . .
32
007F570D131A6
0721128716E8A
49
1.1690 . . .
32
0220DA1E50EEE
0EEB06978B777
50
1.1093 . . .
32
092A07192BF8E
1F8C92BF8E6D4
51
1.1560 . . .
16
003CBF96B9CCE
133641E5434AA
51
1.1560 . . .
16
023615B4134EE
113CEBC7E9C8A
52
1.1863 . . .
16
00C3CC8A65695
03C0CFB9969AA
52
1.1863 . . .
3332
4107
3332
4107
781
1083
132
169
132
169
77
100
77
100
3844
5043
319
441
4348
5547
4348
5547
595
726
4892
6075
382
529
5212
6627
5212
6627
119
144
119
144
5828
7203
5828
7203
453
625
6220
7803
6220
7803
865
1014
865
1014
00035AC726
1.1796 . . .
1513
4107
1513
4107
433
1083
1673
4563
1673
4563
1
3
1
3
1841
5043
521
1323
2017
5547
2017
5547
125
363
2201
6075
617
1587
2393
6627
2393
6627
73
216
73
216
2593
7203
2593
7203
721
1875
2801
7803
2801
7803
1
3
1
3
16
37
1513
4107
1513
4107
433
1083
1673
4563
1673
4563
1
3
1
3
1841
5043
521
1323
2017
5547
2017
5547
125
363
2201
6075
617
1587
2393
6627
2393
6627
73
216
73
216
2593
7203
2593
7203
721
1875
2801
7803
2801
7803
1
3
1
3
16
30F3FC455A566
33F0FF76A9A59
CROSSCORRELATION OF RUDIN-SHAPIRO-LIKE POLYNOMIALS
31
Acknowledgment
This research was done using resources provided by the Open Science Grid
[14, 20], which is supported by the National Science Foundation award 1148698,
and the U.S. Department of Energy’s Office of Science. The authors thank Balamurugan Desinghu and Mats Rynge, who helped them set up the calculations
on the Open Science Grid.
References
[1] K. T. R. Boothby and D. J. Katz. Low correlation sequences from linear combinations
of characters. IEEE Transactions on Information Theory, 63(10):6158–6178, 2017.
[2] P. Borwein and M. Mossinghoff. Rudin-Shapiro-like polynomials in L4 . Math. Comp.,
69(231):1157–1166, 2000.
[3] M. J. E. Golay. Static multislit spectrometry and its application to the panoramic display
of infrared spectra. J. Opt. Soc. Am., 41(7):468–472, 1951.
[4] M. J. E. Golay. A class of finite binary sequences with alternate auto-correlation values
equal to zero. IEEE Trans. Inform. Theory, 18(3):449–450, 1972.
[5] S. W. Golomb and G. Gong. Signal design for good correlation. Cambridge University
Press, Cambridge, 2005.
[6] T. Høholdt. The merit factor problem for binary sequences. In Applied algebra, algebraic
algorithms and error-correcting codes, volume 3857 of Lecture Notes in Comput. Sci.,
pages 51–59. Springer, Berlin, 2006.
[7] T. Høholdt and H. E. Jensen. Determination of the merit factor of Legendre sequences.
IEEE Trans. Inform. Theory, 34(1):161–164, 1988.
[8] J. Jedwab. A survey of the merit factor problem for binary sequences. In T. Helleseth,
D. Sarwate, H.-Y. Song, and K. Yang, editors, Sequences and Their Applications - SETA
2004, volume 3486 of Lecture Notes in Computer Science, pages 19–21. Springer Berlin
/ Heidelberg, 2005.
[9] K. H. A. Kärkkäinen. Mean-square cross-correlation as a performance measure for department of spreading code families. In IEEE Second International Symposium on Spread
Spectrum Techniques and Applications, pages 147–150, 1992.
[10] D. J. Katz. Aperiodic crosscorrelation of sequences derived from characters. IEEE Transactions on Information Theory, 62(9):5237–5259, 2016.
[11] D. J. Katz and S. A. Trunov. Rudin-Shapiro-like polynomials with maximum asymptotic
merit factor. Preprint, arXiv:1711.02233,
2017.
P
P
[12] J. E. Littlewood. On polynomials n ±z m , n eαm i z m , z = eθi . J. London Math. Soc.,
41:367–376, 1966.
[13] J. E. Littlewood. Some Problems in Real and Complex Analysis. D. C. Heath and Co.
Raytheon Education Co., Lexington, MA, 1968.
[14] R. Pordes, D. Petravick, B. Kramer, D. Olson, M. Livny, A. Roy, P. Avery, K. Blackburn,
T. Wenaus, F. Wrthwein, I. Foster, R. Gardner, M. Wilde, A. Blatecky, J. McGee, and
R. Quick. The open science grid. Journal of Physics: Conference Series, 78(1):012057,
2007.
[15] M. B. Pursley and D. V. Sarwate. Bounds on aperiodic cross-correlation for binary sequences. Electronics Letters, 12(12):304–305, 1976.
[16] W. Rudin. Some theorems on Fourier coefficients. Proc. Amer. Math. Soc., 10:855–859,
1959.
[17] D. V. Sarwate and M. B. Pursley. Crosscorrelation properties of pseudorandom and
related sequences. IEEE Trans. Inform. Theory, 68(5):593–619, 1980. Correction in IEEE
Trans. Inform. Theory 68(12):1554, 1980.
[18] R. A. Scholtz and L. R. Welch. Group characters: sequences with good correlation properties. IEEE Trans. Inform. Theory, 24(5):537–545, 1978.
32
DANIEL J. KATZ, SANGMAN LEE, AND STANISLAV A. TRUNOV
[19] M. R. Schroeder. Number Theory in Science and Communication, volume 7 of Springer
Series in Information Sciences. Springer-Verlag, Berlin, fourth edition, 2006.
[20] I. Sfiligoi, D. C. Bradley, B. Holzman, P. Mhashilkar, S. Padhi, and F. Wurthwein. The
pilot way to grid resources using glideinwms. In 2009 WRI World Congress on Computer
Science and Information Engineering, volume 2, pages 428–432, March 2009.
[21] H. S. Shapiro. Extremal problems for polynomials and power series. Master’s thesis, Dept.
of Mathematics, Massachusetts Institute of Technology, Cambridge, 1951.
Department of Mathematics, California State University, Northridge, United
States
| 7cs.IT
|
MSDNN: Multi-Scale Deep Neural Network for Salient
Object Detection
Fen Xiao, Wenzheng Deng, Liangchan Peng, Chunhong Cao, Kai Hu, Xieping Gao*
Key Laboratory of Intelligent Computing & Information Processing of Ministry of
Education, Xiangtan University, Hunan 411105, China
*[email protected]
Abstract. Salient object detection is a fundamental problem and has been received a great deal of
attentions in computer vision. Recently deep learning model became a powerful tool for image
feature extraction. In this paper, we propose a multi-scale deep neural network (MSDNN) for salient
object detection. The proposed model first extracts global high-level features and context
information over the whole source image with recurrent convolutional neural network (RCNN).
Then several stacked deconvolutional layers are adopted to get the multi-scale feature
representation and obtain a series of saliency maps. Finally, we investigate a fusion convolution
module (FCM) to build a final pixel level saliency map. The proposed model is extensively
evaluated on four salient object detection benchmark datasets. Results show that our deep model
significantly outperforms other 12 state-of-the-art approaches.
Key words: salient object detection; recurrent convolution neural network; multi-scale feature
representation; fusion convolution module.
1 Introduction
Salient object detection, which aims to highlight the most important object regions or
distinctive parts in an image, has been a fundamental problem. As a key pre-processing, saliency
detection has been applied to benefit various computer vision applications, including image
segmentation [1], object detection [2, 3], image retrieval [4], image classification [5], etc. Many
computational models [6-12] have been proposed to capture different saliency cues.
Many conventional saliency models rely on various local or global low level features. Local
methods [7, 13] aim at evaluating the distinctiveness of each region or pixel with respect to local
context for color, texture, and edge orientation, which typically tend to highlight object boundaries
while often miss object interiors. Global methods [8, 14] take the entire image into consideration to
infer the salient objects. Unlike local methods which are sensitive to high frequency image
contents like edges, global methods are often less effective when the textured regions of salient
objects are similar to the background. Considering these problems, several researchers integrated
local with global contrasts [15, 16] to incorporate their complimentary interactions. However,
these methods are weak in capturing high-level semantic information and mainly rely on
bottom-up hand-crafted features, which may fail to describe complex image scenarios and object
structures, and it is difficult to find the optimal integration.
Convolutional neural networks (CNNs) have powerful capacities in learning complex
high-level semantic features of image by establishing deep architectures, which have been
successfully used in many applications of computer vision. The CNN-based salient object
detection can be divided into region-based models and pixel-based models. Region-based models
aim to extract features of each region, and then predict its saliency score. However, exiting
region-based methods predict the saliency label of a pixel only considering a limited size of local
image patch. They mostly fail to enforce spatial consistency and may inevitably make incorrect
predictions. On the other hand, pixel-based models lack the ability to get multi-scope of context
for each pixel. An appropriate scope of context is also very important to detect different sizes of
objects meanwhile keep those non-salient objects suppressed in the background.
In this paper, we propose a new computational model for saliency detection. The proposed
network is capable of extracting hierarchical high-level features and multi-scale contextual
simultaneous ly. First, we use the RCNN to extract global feature representation from the source
image. And then, we extend RCNN by introducing upsample layers and exploit FCM to enhance
the final saliency map accuracy. Our work has the following contributions:
(1) Multi-scale saliency maps are utilized to improve saliency detection performance. In this
framework, we can get different scopes of neighborhood information and robust features, which
are effective to detect the small object or multiple objects in the same input image.
(2) FCM has been proposed to fuse the multi-scale saliencies. We use FCM to predict the
final pixel level saliency map, which can effectively improve the model performance by integrating
different scale feature maps with a nonlinear manner.
2 Related works
2.1 Convolutional neural network for Salient Object Detection
Inspired by the fact that CNN has forceful ability to extract discriminative features
representation from raw pixels automatically, some deep CNN based models [17-20] have been
proposed for salient object detection and achieved the state-of-the-art performance. These models
can be divided into two categories according to their treatment of input images: region-based
methods and pixels-based methods. Region based methods formulate saliency detection as a
region classification task. While pixel-based methods directly predict saliency map with
end-to-end network. Here, we focus on discussing recent saliency detection models based on deep
learning architectures.
For the region-based CNN methods, Zhao et al. [17] developed a multi-context deep learning
framework which extracts features of given superpixels at global and local scale, and then predict
the saliency score of each region. Considering the fact that the saliency maps are greatly affected
by the number of regions, Li et al. [21] presented a CNN approach for salient object detection
based on multi-scale feature extraction. These region-based methods lack sufficient context to
accurately locate salient objects since they deal with each image regions independently. And the
pixel saliency value is given by the mean of pixel saliency values in its enclosing region.
To address the above issues, some works adopt diverse pixel-based deep architectures to
preserve details in saliency detection. In [22], Li et al. presented a fully convolutional neural
network for saliency detection. Liu et al. [23] proposed a hierarchically refine mechanism from
coarse to fine to progressively refine the global saliency map. Although the end-to-end approach
enables the model outperforms the region based model in some extent, the spatial information lost
in downsampling stage cannot be fully recovered. Recently, some kind of pixel-based information
was integrated with higher level region-based information to get the better salient object detection
performance. In [19], Yuan et al. proposed a dense and sparse labeling (DSL) framework for
saliency detection, both macro object contours and low-level image features were be extracted for
saliency estimations. In [20], Tang et al. proposed a novel saliency detection model named as
CRPSD, in which the pixel-level CNN and fusion CNN were jointly learned. These models tend
to extract local features from each patch and global features from source image. They are very
time-consuming, becoming the bottleneck of the computational efficiency of saliency detection.
In this paper, we propose to adopt multi-scale feature representation to enhance the capability
of the model to learn discriminative features and boost the details of pixels. The network does not
rely on image over-segmentation and post-processing, making the model effective and efficient.
2.2 Recurrent Convolution Neural Network
In [24], Liang et al. proposed a RCNN for object recognition, which extract the context
information by integrating the recurrent connections into each convolutional layer. As shown in
Fig.1, we unfold the recurrent connections for T time steps, the model becomes a deeper
feed-forward network with 4(T+1)+2 layers. For a unit located at (i,j) on the kth feature map in an
RCL, its net input at time step t is given by:
()
(
)
(
( )
)
( )(
)
(1)
Where ( ) and ( )(
) denote the feed-forward and recurrent input, respectively, which are
the vectorized patches centered at ( ) of the feature maps in the previous and current layer,
and
denote the vectorized feed-forward and recurrent weights, respectively, and
()
And its corresponding activity or state is a function of its net input:
(
(
is the bias.
( ) )).
In Eq.(1), the weight parameters are assumed to be time invariant, we can deepen the network
without any extra parameters and keep the computational effic ient. And depending on the
structure of the RCL, there may be multiple paths between the input layer and the output layer. As a
result the receptive filed of an RCL expands when the iteration number increases, which is
significant for the network to get more and more context information. In this paper, several
stacked RCLs will be used to extract high-level feature representation from the input image.
Feed-forward connection
Recurrent connection
t=3
t=2
t=1
x
t=0
u
Fig. 1 The architecture of RCL.
3 Proposed Method
The architecture of MSDNN is shown in Fig. 2. First, the RCNN is utilized to extract the
high-level context feature fc from global views. Then, hierarchical refined feature maps can be
produced based on . Next, we get multi-scale top-down saliency maps progressively by exploiting
upsampling layers and concatenation layers. At the end of the network, the multi-scale saliency
maps are fed into FCM to generate the final saliency map.
Fig. 2 The architecture of the proposed neural network.
3.1 RCNN for high-level feature extraction
In this subsection, we will use RCNN to extract global features. As shown in Fig 2, the RCNN
in our method is made up of six blocks: a convolutional block, four recurrent convolutional blocks
and a full connection block. The convolutional block (denoted as conv1 in Fig. 2) consists of two
convolution layers and two ReLU [25] layers, in which the number of feature maps is set to 64.
Each recurrent block is composed of a pooling layer and a RCL. The size of kernel is defined to 2*2
and the stride is set to 2 in the pooling layer. We set the number of feature maps to 96 and there are
3 recurrent connections in the RCL. A fully connected layer with 6272 nodes is followed by the last
recurrent block, which outputs a high-level feature vector
for the input image. Finally, we
reshape and upsample the
to obtain a coarse global feature map
(28*28*64), as shown in
Fig.3.
Fig. 3 Global feature maps obtained by RCNN.
The RCNN is effective in learning global features, while it is not sufficient for saliency
detection due to the details maybe lost in down-sample layers. In this paper, multi-scale saliency
prior maps will be used to capture local information of salient objects.
3.2 Multi-scale saliency maps
In this subsection, feature maps in different scales
(
) are generated and then
integrated with RCL to generate multi-scale saliency prior maps. As shown in Fig. 2, we combine
with the output of RCL4 (denoted as ) to generate the refiner feature map
can get the hierarchical feature maps of the input image:
(
(
)
. Similar, we
)
(2)
is the output of corresponding RCL. Conv
Where Concat represents concatenation operator and
denotes convolution operator using the kernel
(3*3). Thus, we can get the hierarchical feature
maps
, which will be utilized to predict the multi-scale saliency prior maps:
(
)
(3)
Where
represents upsample operator.
is the upsample kernel with stride of
, and the sizes
of
are set to 16*16, 8*8, 4*4 and 1*1, respectively. The upsample operator is exploited to ensure
the same spatial size of
with the original image.
In Fig. 4, we show the detailed framework of multi-scale saliency maps. The hierarchical
refined feature maps are exploited to produce multi-scale saliency maps. From
to
, the
receptive fields gradually shrink, and they are representing contextual information at 4 different
scales. For each pixel in the saliency map, the network can get multiple scopes of local
neighborhood information, which are complementary and can be fused to further improve the
performance to detect the small object or multiple objects in the input image. Even for the complex
image, the model also can get better effect.
Fig. 4 Saliency maps with different scales.
3.3 FCM for multi-scale saliency maps fusion
Given an input image, the proposed method efficiently produces four saliency prior maps.
Finally, FCM exploit the concatenated multi-scale saliency maps to enhance the final salient object
accuracy.
Fig. 5 The detailed framework of FCM.
The detailed structure of FCM is shown in Fig. 5. The four scale maps are first concatenated
into a 4-channel image. Then three convolutional and ReLU layers are used to fuse the feature maps
and predict the saliency map. For the former two convolutional layers, we set the number of feature
maps to 32 and 64 respectively. Last, we use only one convolutional kernel to convert the features to
the final salient map. During training stage, a loss layer is used to compute the errors between the
saliency and ground truth with the sigmoid cross-entropy. The FCM and former network are trained
jointly to gain the performance improvement and we don’t conduct any post-processing.
4 Experiments
4.1 Experimental Setup
4.1.1 Datasets
In this section, we give experimental results to evaluate the performance of MSDNN on four
standard benchmark datasets: ECSSD [9], iCoSeg [25], DUT-OMRON [26], MSRA10K [8].
ECSSD is a challenging dataset and includes 1000 semantical meaningful but structurally complex
images. iCoSeg was designed for co-segmentation and it contains 643 images and most of the
images may contain one or multiple salient objects. DUT-OMRON includes 5168 manually
selected high quality images with one or more salient objects and relatively complex backgrounds.
MSRA10K dataset is widely used for saliency detection and covers a large variety of image contents,
and most of the images include only salient object with high contrast to the background. All the
datasets is bundled with pixel-wise ground truth annotations.
4.1.2 Evaluation Criteria
We use standard tools for performance evaluation, precision-recall (P-R) curves, F-measure,
mean absolute error (MAE), and area under the ROC curve (AUC), which are most commonly used
evaluation metric for saliency detection.
For a saliency map, we generate a series of binary saliency masks by setting the threshold from
0 to 255 with an increment of 5. By comparing the binary masks against the ground truth, pairs of
precision and recall scores are obtained. Then we average scores of all images in the database to get
the P-R curve. The P-R curve reflects the mean precision and recall of different saliency maps at
different thresholds. For quantitative analysis, we binarize each saliency map with an adaptive
threshold, which is determined as twice the average value of saliency map. We compute average
precision and recall values based on the generated binary maps and the corresponding ground-truth.
F-measure score is harmonic mean of precision and recall:
(
)
(4)
where
is set to 0.3 as suggested in [9, 27].
And the
is the average per-pixel difference between the saliency map S and its
corresponding ground truth G:
∑
∑ | (
)
(
)|
(5)
where
and
are width and height of the saliency map, respectively.
is more
meaningful in evaluating the applicability of a model as it reveals the numerical distance between
the saliency map and the ground truth.
Area under the ROC curve (AUC) is one of the most widely used metrics for saliency methods
evaluation, which is calculated with Matlab code from Judd’s [28] homepage.
4.1.3 Implementation Details
We implement the proposed model on the popular open source framework of Caffe [29], it is
well known that training such a deep network from scratch is difficult. Therefore, we adapt
multi-steps method to train the network in our experiment. In the training state, we resize all the
images to 224*224 pixels. The momentums set to 0.9 and the weight decay is 0.0005. In our
experiment, we random selected 5500 images from MSRA10K and 2000 images from
DUT-OMRON as training set, and selected 500 images from MSRA10K and 500 images from
DUT-OMRON as validation, and we test the proposed model on the rest images and the other two
datasets. It costs about 11 hours to training the network.
4.2 Performance Comparison
We compared the proposed saliency model with 12 state-of-the-art models, including CA [30],
DSR [31], GMR [26], GR [32], HS [9], PRGL [15], RBD [33], BL [34], DRFI [35], DHSN [23],
MDF [17], DSL [19]. DHSN, MDF and DSL are CNN based methods. For fair comparison, the
source codes of these models are obtained from the project site of each algorithm or the benchmark
evaluation [27].
Fig. 6 Comparison of P-R curves of 12 salient object detection methods on 4 benchmark
datasets.
Fig. 7 Comparison of Precision, Recall and F-measure (using adaptive threshold) among
12different models on 4 datasets.
We compare the P-R curves in Fig. 6. As can be seen that proposed model significantly
outperformance the 12 algorithms in terms of the 4 public datasets. The precision achieves
maximum value when the recall from 0 to 1 in our model compared with the 12 methods. For the
complex and contain multiple objects dataset such as iCoSeg, the proposed method can detect all the
objects and obtain a better effect. From the P-R curves we can observe that the precision usually fell
down sharply only when the recall is very close to 1. On the contrary, most of other methods usually
dropt slowly within the scope of the recall, which shows that the proposed method can generates
high contrast saliency maps. Fig. 7 and Fig. 8 shows the precision, recall, F-measure, AUC and
MAE of the different models on the 4 datasets. The proposed method shows the highest F-measure
and the lowest MAE on all the datasets.
Fig. 8 Comparison of AUC and MAE (using adaptive threshold) among 12 different models on 4
datasets.
A visual comparison of saliency maps in Fig. 11. As can be seen, the CNN based methods is
generally superior to the conventional approaches and the proposed model generates more effect
and accurate saliency maps in various challenging cases, e.g., low contrast images, multiple small
objects and objects located the image boundary. It is obvious ly noticed that our saliency maps are
very similar to the ground truth and can highlight the saliency regions especially for the simple
images such as MSRA10K dataset.
4.3 Effectiveness of multi-scale saliency maps
,
As discussed in Section 3, our multi-scale saliency maps consist of four components,
,
and
. To demonstrate the effectiveness of different scale saliency maps, we trained
several models for comparison,
only,
denote as catenated
and
, similarity, the
proposed denote as the combination of these four components. The qualitative comparison on
ECSSD dataset is shown in Fig. 9, the proposed model consistently achieves the best performance,
and the model of
is better than
,
is better than
. And a visual comparison is
given in Fig. 10. The result demonstrates that the proposed multi-scale model not only can get the
details information but also can suppress the background noise.
Fig. 9 Component analysis on ECSSD dataset.
Fig. 10 Visualization of the multi-scale saliency maps. From left to right: original image, ground
truth,
,
,
, MSDNN.
Fig. 11 Visual comparison of saliency maps from state-of-the-art methods. From top to bottom:
ECSSD (first and second rows), iCoSeg (third and fourth rows), DUT -OMRON (fifth and sixth
rows), MSRA10K (last two rows). From left to right: input image, ground truth, CA, DSR, GMR,
GR, HS, PRGL, RBD, BL, MDF, DSL, DRFI, DHSN, MSDNN.
5 Conclusions
In this paper, we proposed a novel method of multi-scale deep supervised neural network for salient
object detection. The model exploits RCNN to learn discriminative information and multi-scale
image feature representation are adopted to hierarchically reconstruct final saliency map.
Experimental demonstrate that our deep model is effective and efficient and can significantly
improve the state-of-the-art.
Reference
1.
Jung, C. and C. Kim, A unified spectral-domain approach for salien cy detection and its
application to automatic object segmen tation. IEEE Transactions on Image Processing, 2012.
21(3): p. 1272-1283.
2.
Luo, P., et al. Switchable deep network for pedestrian detection. in Proceedings of th e IEEE
Conference on Computer Vision and Pattern Recognition. 2014.
3.
He, H., Saliency and depth-based unsupervised object seg menta tion. Iet Image Processing,
2017. 10(11): p. 893-899.
4.
Gao, Y., et al., Visual-textual joint relevance learning for tag-based social image sea rch. IEEE
Transactions on Image Processing, 2013. 22(1): p. 363-376.
5.
Wu, R., Y. Yu, and W. W ang. Scale: Supervised and cascad ed laplacian eig enmaps fo r visual
object recognition based on nearest neighbors. in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition. 2013.
6.
Xiao, F., et al., Salient object detection based on eye tracking data . Signal Processing, 2018.
144: p. 392-397.
7.
Itti, L., C. Koch, and E. Niebur, A model of saliency-based visual attention fo r rapid scene
analysis. IEEE Transactions on pattern analysis and machine intelligence, 1998. 20(11): p.
1254-1259.
8.
Cheng, M.-M., et al., Global contrast based salien t region detection. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 2015. 37(3): p. 569-582.
9.
Yan, Q., et al. Hierarchical salien cy detection. in Proceedings of the IEEE Conferen ce on
Computer Vision and Pattern Recognition. 2013.
10.
Sun, X., et al., Objectness to assist salient object detection . Iet Image Processing, 2016. 10(5):
p. 391-397.
11.
Liu, et al., Efficient saliency detection based on gaussian models. Image Processing Iet, 2011.
5(2): p. 122-131.
12.
Rabbani, N., et al., Efficient Bayesian approach to saliency detection based on Dirichlet
process mixture. Iet Image Processing, 2017. 11(11): p. 1103-1113.
13.
Xie, Y., H. Lu, and M.-H. Yang, Bayesian salien cy via lo w and mid level cues. IEEE Transactions
on Image Processing, 2013. 22(5): p. 1689-1698.
14.
Achanta, R., et al. Frequency-tuned salient region detection. in Computer Vision and Pattern
Recognition, 2009. CVPR 2009. IEEE Conference on. 2009.
15.
Tong, N., et al., Salient object detection via global and local cues. Pattern Recognition, 2015.
48(10): p. 3258-3267.
16.
Shi, K., et al. PISA: Pixelwise image saliency by aggregating co mplementa ry appearan ce
contrast measures with spatial priors . in Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition. 2013.
17.
Zhao, R., et al. Saliency detection by multi -context deep lea rning. in Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition. 2015.
18.
Wang, L., et al. Saliency detection with recurren t fully convolutional netwo rks. in European
Conference on Computer Vision. 2016. Springer.
19.
Yuan, Y., et al., Dense and Sparse Labeling with Multi-Dimensional Fea tures for Saliency
Detection. IEEE Transactions on Circuits & Systems for Video Technology, 2016. PP(99): p.
1-1.
20.
Tang, Y. and X. Wu. Saliency detection via co mbining region-level and pixel-level p redictions
with cnns. in European Conference on Computer Vision. 2016. Springer.
21.
Li, G. and Y. Yu. Visual saliency based on multiscale deep features. in Compu ter Vision and
Pattern Recognition. 2015.
22.
Li, X., et al., DeepSaliency: Multi-task d eep neu ral n etwork model for salien t object detection.
IEEE Transactions on Image Processing, 2016. 25(8): p. 3919-3930.
23.
Liu, N. and J. Han. DHSNet: Deep Hiera rchical Salien cy Network fo r Salient Object Detection.
in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . 2016.
24.
Liang, M. and X. Hu. Recurrent convolutional neural n etwork for object recognition. in
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2015.
25.
Batra, D., et al. icoseg: Interactive co-segmentation with in telligen t scribble guidance. in
Computer Vision and Pattern Recognition (CVPR), 2010 IEEE Conference on . 2010. IEEE.
26.
Yang, C., et al. Saliency detection via graph-based manifold ranking. in Proceedings of the
IEEE conference on computer vision and pattern recognition. 2013.
27.
Borji, A., et al., Salient object detection: A b enchma rk. IEEE Transactions on Image Processing,
2015. 24(12): p. 5706-5722.
28.
Judd, T., et al. Learning to predict where humans look. in Computer Vision, 2009 IEEE 12th
international conference on. 2009. IEEE.
29.
Jia, Y., et al. Caffe: Convolutional architecture for fast featu re embedding. in Proceedings of
the 22nd ACM international conference on Multimedia. 2014. ACM.
30.
Goferman, S., L. Zelnik-Manor, and A. Tal, Contex t-aware salien cy detection. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 2012. 34(10): p. 1915-1926.
31.
Li, X., et al. Saliency detection via dense and spa rse reconstru ction. in Pro ceedings of th e IEEE
International Conference on Computer Vision. 2013.
32.
Yang, C., L. Zhang, and H. Lu, Graph-Regularized Salien cy Detection With Convex-Hull-Based
Center Prior. IEEE Signal Processing Letters, 2013. 20(7): p. 637-640.
33.
Zhu, W., et al. Saliency optimization from robust background detection. in Pro ceedings of the
IEEE conference on computer vision and pattern recognition. 2014.
34.
Tong, N., et al. Salient object detection via bootstrap learning. in Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition. 2015.
35.
Jiang, H., et al. Salient object detection: A discriminative regional featu re integ ration
approach. in Proceedings of the IEEE conferen ce on computer vision and pattern recognition.
2013.
| 1cs.CV
|
THE APPLICATION OF S PECIAL MATRIX PRODUCT TO DIFFERENTIAL
QUADRAUTRE S OLUTION OF GEOMETRICALLY NONLINEAR BENDING
OF ORTHOTROPIC RECTANGULAR PLATES
W. Chen (Corresponding author) and
Permanent mail address: P. O. Box 2-19-201, Jiangshu University of Science &
Technology, Zhenjiang City, Jiangsu Province 212013, P. R. China
Present mail address (as a JSPS Postdoctoral Research Fellow): Apt.4, West 1st floor,
Himawari-so, 316-2, Wakasato-kitaichi, Nagano-city, Nagano-ken, 380-0926, JAPAN
Permanent email: [email protected]
C. Shu
Department of M echanical and Production Engineering, National University of
Singapore, Singapore
W. He
Department of Electrical Engineering, Jiangsu University of Science & Technology,
Zhenjiang, Jiangsu 212013, P. R. China.
ABSTRACT
The Hadamard and SJT product of matrices are two types of special matrix product.
The latter was first defined by Chen [1]. In this study, they are applied to the
differential quadrature (DQ) solution of geometrically nonlinear bending of isotropic
and orthotropic rectangular plates. By using the Hadamard product, the nonlinear
formulations are greatly simplified, while the SJT product approach minimizes the
1
effort to evaluate the Jacobian derivative matrix in the Newton-Raphson method for
solving the resultant nonlinear formulations. In addition, the coupled nonlinear
formulations for the present problems can easily be decoupled by means of the
Hadamard and SJT product. Therefore, the size of the simultaneous nonlinear
algebraic equations is reduced by two-thirds and the computing effort and storage
requirements are alleviated greatly. Two recent approaches applying the multiple
boundary conditions are employed in the present DQ nonlinear computations. The
solution accuracies are improved obviously in comparison to the previously given by
Bert et al. [9]. The numerical results and detailed solution procedures are provided to
demonstrate the superb efficiency, accuracy and simplicity of the new approaches in
applying DQ method for nonlinear computations.
1. INTRODUCTION
The differential quadrature (DQ) method was introduced by Bellman and his
associates [1, 2]. Since then, the method has been applied successfully to a broad
range of problems [2-16]. A striking merit of the method is of high efficiency in
computing nonlinear problems [2, 3, 5, 7-11, 14-16]. Compared with the standard
numerical techniques such as the finite element and finite difference methods, the DQ
method produces solution of reasonable accuracy with relatively small computational
effort. This method also does not require seeking the trial functions satisfying
boundary conditions as in the Rayleigh-Ritz and Galerkin methods [12]. Therefore, it
is easily used for practical purposes. It has been found that the DQ method is closely
related with the collocation (or pseudo-spectral) methods [12, 17, 18]. The principal
advantages of the DQ method over the latter are its simplicity and ease in using grid
spacing without restriction [8, 18].
The geometrically nonlinear behavior of thin plates is usually described by the von
Karman equations and has become a benchmark problem for testing numerical
solutions to nonlinear partial differential equations. Bert et al. [9] employed the DQ
2
method to solve the static von Karman equations in analyzing geometrically nonlinear
bending of isotropic and orthotropic rectangular plates. The work shows that the DQ
method is an efficient numerical technique in solving complex nonlinear problems by
comparing with the finite element and finite difference methods. However, the
application is not very successful in plate with all edge simply supported. The
solution procedure is also much more complex in comparison to the solution of linear
problems. Nonlinear problems presently are considered to be in the territory of high
performance workstation and supercomputer. With the development of personal
computers, it has become possible to develop a computationally inexpensive
approach for the nonlinear computations via PC computer. The DQ method should be
of choice in this respect. However, in solving complex multidimensional nonlinear
problems, for example, von Karman equations of plates, computational effort and
storage requirements are still rather high even in the DQ method. Therefore, if the DQ
method attempts to undertake more nonlinear computations via PC computer, its
computing cost need obviously be reduced. Recently, some remarkable advances are
achieved to simplify application and improve efficiency and accuracy of the DQ
method. In this study, values of some new techniques in the DQ computing are
verified through the solution of the nonlinear geometrical bending of plates. Especially,
the work validates the applicability, simplicity and high efficiency of special product
approach in the DQ nonlinear computations. It should be pointed out that due to a
great reduction in the computing effort and storage requirements by using the new
techniques, all results presented here were accomplished on a 386 personal computer
with only 4M B memory. In contrast, Bert et al. [9] used an IBM framework
computer for the same task. In the following, innovations in this work are briefly
introduced.
First, the Hadamard product of matrices and the DQ matrix approximate formulas are
used to express the formulation of nonlinear partial differential operator in explicit and
easily programmable matrix form. The DQ analog formulas in matrix form [20] can be
3
viewed as simple and compact version of the traditional polynomial approximation of
multidimensional problems [4]. By using these formulas, the formulation effort can be
simplified. The often-used ordinary matrix product rises from the concept of linear
transformation and is extended to handle the nonlinear problems. However, since
nonlinear problems are actually different from linear ones, the ordinary matrix product
seems not to undertake the task of nonlinear analysis and computations very well.
The Hadamard product is a kind of special matrix product and not well known to the
numerical computation community. It was found that the Hadamard product provides
an explicit, compact and convenient approach to formulate the nonlinear differential
operators in the DQ method [15] as well as the other numerical techniques [16].
Second, the SJT product was first introduced by Chen and Zhong [15] and Chen [16]
as an efficient and simple algorithm to compute the exact Jacobian derivative matrix in
the Newton-Raphson solution of the nonlinear formulations in the Hadamard product
form. The SJT product is also a kind of special matrix product. It is emphasized that
the SJT product approach may require minimal computing effort in all possible
approaches for the same task. M ore importantly, the Hadamard and SJT product
provides a decoupling technique in solving the coupled nonlinear DQ analog equations
of the von Karman equations. By comparing with Bert et al. [9], the resulting size of
the nonlinear simultaneous algebraic equations is reduced by two-thirds. Therefore,
the computational effort and storage requirements are alleviated significantly. Third, a
recent approach applying multiple boundary conditions in the DQ analysis of high
order boundary value problems, proposed by Wang and Bert [21], is employed for the
solution of the geometrically nonlinear simply-supported plates. It was verified that
in the linear cases, the accuracies of the DQ solutions using Wang and BertÕs approach
were evidently improved in comparison to using the conventional so-called δ
approach [12, 21, 22]. Wang and BertÕs approach is also effective for the other
boundary conditions except for the clamped-clamped (C-C) boundary condition [21,
22] and free edge of plate. Chen and Yu [23] proposed another different approach to
improve accuracy and eliminate instability caused by δ effect in the conventional δ
4
approach. Chen and YuÕs approach is applicable for problems with any boundary
constraint. In this study, the approach is applied to analyze the plate of clamped
edges. An improvement on this approach is also presented to avoid extra four
equations of corner points in the resulting formulation and reduce the size of the
resultant algebraic systems. It is believed that these two new approaches applying
multiple boundary conditions are first used in the DQ solution of nonlinear boundary
value problems. To avoid confusion, all three approaches are denoted as the DQδ
approach [6, 7, 9, 10], DQWB approach by Wang and Bert [21], and DQCY
approach by Chen and Yu [23]. Fourth, the Chebyshev grid points are used in the
present numerical study. As is expected, the DQ method using such grid points has
faster rate of convergence than using the equally spaced grid points in [9]. Finally,
some conclusions are drawn based on the results reported herein.
2. APPROACHES APPLYING BOUNDARY CONDITIONS AND MATRIX
APPROXIMATE FORMULAS IN THE DQ METHOD
The differential quadrature method approximates the partial derivative of a function at
a given discrete point as a weighted linear sum of function values at all of the discrete
points along the respective variable direction in the entire domain of the variable. The
m-th order derivative of a single function f(x) at a given discrete grid point i can be
approximated by the DQ method with N discrete grid points [2-6], namely,
∂m f
∂x m
N
x=x
i
= ∑ wij(m ) f j
j= 1
i = 1,2, K, N . ,
(2-1)
where fj=f(xj), wij(m) are the corresponding DQ weighting coefficients and can be
determined by requiring that Eq. (2-1) be exact for all polynomials less than or equal
to N-1. The explicit formulas [11, 17] were developed to conveniently obtain the
accurate DQ weighting coefficients. For the weighting coefficients of the 1st order
derivatives, we have
5
wij(1) =
N
1
xi − xk
,
∏
x j − xi k ≠i x j − x k
i=1,2,ÉN and j = 1,2,K , N .
(2-2a)
k≠ j
i≠ j
N
wii(1) = − ∑
k ≠i
1
,
xi − x k
i=1,2,ÉN
(2-2b)
The weighting coefficients for high order derivatives can be generated by recursion
formulas [17]:
(m+1)
wij
wii
(1) (m ) wij(m )
,
= m wij wii −
−
xi x j
( m+1)
N
= − ∑ wij
( m+1)
j ≠i
i≠ j
(2-3a)
,
(2-3b)
where the superscript (m) and (m+1) denote the order of the derivative.
In the present study, besides the uniform points, the zeros of Chebyshev polynomials
of the first kind are considered, namely,
ri = cos
(2i −1)π
,
2N
-1<ri<1,
i=1, 2, ..., N.
(2-4)
where ri is the i-th root of the N order Chebyshev polynomials. The zeros do not
included the end points of the normalized domains (0≤x≤1 or -1≤x≤1), while in the
present case, the end points (x=0,1 or x=±1) are needed to force boundary conditions.
Therefore, an algebraic transformation is required, namely,
ri − 1
rN − r1 ,
xi =
yj =
rj −1
rM − r1 ,
0≤ xi ≤1
i=1, 2, ..., N.
0≤ y j ≤1
(2-5a)
j=1, 2, ..., M .
(2-5b)
The above formulas produce the coordinates of sampling points. With such grid
points, a very simple formula can be obtained for computing the DQ weighting
coefficients of the 1st order derivatives.
(1)
wij
(i− j )
(
− 1) (rN
=
(
(
)
)
2
− r1 ) 1 − rj
ri − rj
1 − ri2 ,
i≠j
(2-6a)
6
wii(1) =
1 ri (rN − r1 )
,
2 1 − ri2
i=1,2,É, N
(2-
6b)
The weighting coefficients for high order derivatives can be easily computed by using
recurrence formulas (2-3a, b). M ore details about the DQ method see references [2, 4,
6, 11, 17]. It should be especially noted that in order to differ the variants from the
DQ weighting coefficients more clearly, the notations for the DQ weighting
coefficients of the 1st, 2nd, 3rd and 4th order derivatives are represented as Aij, Bij, Cij
and Dij, respectively, in all subsequent sections. A , B , C and D represent the
corresponding modified coefficient matrices with built-in boundary conditions by the
DQWB or DQCY approach.
2.1 Three approaches applying multiple boundary conditions
The governing equations in the structural mechanics usually involve the fourth-order
derivatives, and the two conditions need be satisfied at each boundary. Therefore, the
problems are more complex in comparison to the differential systems of no more than
second order such as the Poisson equation and convection-diffusion equation, which
do not involve more than one boundary condition at each boundary end of physical
domain. Some careful considerations are necessary how to properly implement the
double boundary conditions at each edge. The problem is actually the DQ solution of
high-order boundary value problems with multiple boundary conditions. In what
follows we discuss three existing approaches for such problems.
The earliest DQδ approach, proposed by Bert et al. [6] and widely used in literature
[7, 9, 10], enforces the geometry boundary condition at boundary points and
derivative boundary conditions at the δ points, which have very small distance δ
(δ≅10-5 in dimensionless value [22]) away from the boundary. Thus, the approach can
not satisfy derivative boundary conditions exactly at boundary points and the
accuracies of the solutions are affected. The solution accuracy depends on the proper
7
choice of δ. If the value of δ is small enough, the approach produce good results in
some situations such as clamped condition, however, failed to work well in the other
situations such as simply-supported and free edges [10, 21, 22]. On the other hand,
too small δ will deteriorate the computations. δ is usually determined by trial and
error for different cases [22], which is a rather tedious work. Arbitrariness in the
choice of the δ value may introduce the unexpected oscillation behavior of the
solutions. In addition, the number of grid points in the DQδ approach can not be large
due to the ill-conditioned matrices caused by the δ grid spacing.
To overcome the drawbacks in the DQδ approach, some new techniques were
presented in the application of the multiple boundary conditions. Wang and Bert [21]
developed a DQWB approach by incorporating the boundary conditions into the DQ
weighting coefficient matrices in advance, and then the weighting coefficients with
built-in boundary conditions are employed to analogize the governing equations of
problems of interests. The essence of the approach is that the boundary conditions are
applied during formulating the weighting coefficient matrices for inner grid points. The
technique resulted in an obvious improvement in the DQ solution of beams and plates
with free and SS boundary conditions [12, 21, 22]. However, it is regret that the
technique is not applicable for problems with the C-C boundary condition as well as
cross derivative and corner boundary conditions such as completely free plates. Thus,
a combination of the DQWB and DQδ approaches was used to handle the problems
with both the C-C boundary condition and the other boundary conditions in [22]. The
modification of the weighting coefficient matrices also causes some loss of use
flexibility and increases rather high additional computational effort, which require
some matrix product operations of O(N4) scalar multiplications. In this study, the
DQWB approach is used to handle the problems with simple supports.
8
Chen and Yu [23] proposed a different DQCY approach to cure the deficiencies of the
conventional DQδ approach. The fact was noted that that the rank of the DQ
weighting coefficient matrix for the i-th order derivative is N-i, where N is the number
of grid points. Naturally, the rank of the DQ coefficient matrix for the 4th order
derivative is N-4. Therefore, the DQ analog equations of the 4th order governing
equations at boundary points itself and the points immediate adjacent boundary need
be replaced by the analog equations of the boundary equations. DQCY approach
imposes all boundary conditions exactly at boundary points. Therefore, the solution
accuracy of the DQCY approach is improved evidently in comparison to the DQδ
approach. The approach is conceptually simple and effective for the C-C boundary
condition as well as any other boundary condition. Since the DQCY eliminates the δ
effect in the DQδ, the stability of the solution is improved and larger number of grid
points can be used. However, it has found that the DQCY approach is less efficiency
than the DQWB approach whenever the DQWB approach is applicable. Therefore, in
this paper the DQCY approach is applied to handle the problems with the C-C
boundary condition.
An improvement is also made here for the DQCY approach. In this application, the
first and last two rows of the original weighting coefficient matrices are removed, and
then the four boundary equations are applied to modify these coefficient matrices into
(N-4)×(N-4) matrix of full rank. The modified weighting coefficient matrices can be
used to formulate the differential equations of interest directly. Since the boundary
conditions are incorporated into the DQ weighting coefficient matrices in advance,
extra analog equations at four corner points of plate need no longer be considered in
the resulting formulation. The size of the resultant algebraic equations is also reduced.
For example, consider the C-C boundary:
w= 0,
∂w
=0 ,
∂x
at x=0, 1;
(2-7a)
at x=0, 1.
(2-7b)
9
The corresponding DQ approximate equations are given by
w1 = 0 ,
N
∑A
j =1
1j
wN = 0 ;
(2-8a)
N
∑A
wj = 0,
Nj
j =1
wj = 0,
(2-8b)
where wj is the displacement at the jth grid point. It is noted that all boundary
conditions are exactly satisfied at boundary points. By using equation (2-8a, b), the
desired displacements at the 2nd and (N-1)th grid points can be expressed in terms of
the unknown displacement values at interior points, namely,
w2 = −
1
A12
w N −1 = −
N −1
∑A
j= 3
1j
wj
(2-9a)
N−2
1
AN , N − 1
∑A
j =2
Nj
wj
(2-9b)
Substituting equations (2-8a) and (2-9a, b) into the DQ approximate formulas for the
1st, 2nd, 3rd and 4th derivatives, respectively, we have
v
2 v
3v
4 v
d w
d w
d w
dw
v
v
v
v
= Aw,
= Bw,
= Cw,
= Dw
2
3
4
dx
dx
dx
dx
v
where w = {w3 , w 4 , K , w N − 2 }. A ,
(2-10)
B , C and D are (N-4)×(N-4) modified
coefficient matrices different from those of (N-2)×(N-2) dimension in the DQWB
approach. For the other boundary conditions, the modified coefficient matrices can be
obtained in the similar way.
2.2 Matrix approximate formulas in the DQ method
One of the present authors [20] proposed the following DQ formulation in matrix
form for the partial derivatives of the function ψ(x,y) in two-dimensional domain:
∂ψ$
= A xψ$ ,
∂x
∂ 2ψ$
T
= ψ$ By ,
2
∂y
∂ψ$
∂ 2 ψ$
= ψ$A yT ,
= Axψ$ AyT ,
∂y
∂x ∂y
2 $
∂ ψ
∂ 2 ψ$
T
$
= Bx ψB y ,
= Dx ψ$ ,
∂x∂y
∂x 2
∂ 2 ψ$
= Bx ψ$ ,
∂x 2
∂ 2ψ$
T
= ψ$ Dy ,
2
∂y
(2-11)
where the unknown ψ$ is a rectangular unknown matrix rather than a vector as in [4, 6,
8, 11-14, 21, 22]. A , B and D with subscripts x and y stand for the DQ weighting
10
coefficient matrices for the 1st, 2nd and 4th order partial derivatives along x and y
directions, respectively. The superscript T means the transpose of the matrices. The
above analog equations are compact matrix version of the traditional polynomial
approximate formulas given by Civan and Sliepcevich [4]. In fact, both are equivalent.
3. Hadamard and SJT product
It is not an easy task to handle the problems involving nonlinearity. Chen and Zhong
[15] first found that the Hadamard product of matrices was a simple and effective
technique to formulate nonlinear differential equations in the DQ method. The SJT
product was also introduced there to efficiently compute the Jacobian derivative
matrix in the Newton-Raphson method for the solution of nonlinear formulation in the
Hadamard product form.
Definition 3.1 Let matrices A=[aij] and B=[bij]∈CN×M, the Hadamard product of
matrices is defined as A°B= [aij bij]∈CN×M, where CN×M denotes the set of N×M real
matrices.
Definition 3.2 If matrix A=[aij]∈CN×M, then A°q=[aijq]∈CN×M is defined as the
Hadamard power of matrix A, where q is a real number. Especially, if aij ≠0, A°(1)
=[1/aij]∈CN×M is defined as the Hadamard inverse of matrix A. A°0=11 is defined as
the Hadamard unit matrix in which all elements are equal to unity.
Definition 3.3 If matrix A=[aij]∈CN×M, then the Hadamard matrix function f
defined as f
o
( A) = [f (a
ij
o
(A )
is
)]∈ CN M.
×
Considering quadratic nonlinear differential operator W,xW,y, the DQ formulation can
be expressed in Hadamard product form as
v
v
W,xW, y = (A xW )o A yW ,
(
)
11
(3-1)
v
where W is the unknown vector. W with the comma subscript ,x and ,y denotes vector
of partial derivative of function W along x- and y- directions. Ax and Ay are the DQ
weighting coefficient matrices, modified by the boundary conditions, for the 1st order
derivative along the x- and y- directions, respectively. The above equation exposes an
essential idea in the latter analysis. For linear and nonlinear operators with varying
parameters c(x,y), the DQ analog can be obtained in the same way. For example,
v
c(x , y )U ,x = c x j , y j o (Ax U )
(3-2a)
{(
)}
and
{(
c (x , y )U, x U, y = c x j , y
j
)}o (A Uv ) o (A Uv ) .
x
y
(3-2b)
Theorem 3.1: If A, B and C∈CN×M, then
1> A°B=B°A
(3-3a)
2> k(A°B)=(kA)°B, where k is a scalar.
(3-3b)
3> (A+B)°C=A°C+B°C
(3-3c)
The Newton-Raphson method may be one of the most important techniques to
compute nonlinear algebraic equations. One of the major time-consuming calculations
in the method is to evaluate the Jacobian derivative matrix. In order to simplify the
computation of the Jacobian matrix, the postmultiplying SJT product of matrix and
vector was introduced by Chen and Zhong [15].
v
v
Definition 3.4 If matrix A=[aij]∈CN×M, vector V ={vj}∈CN×1, then A◊V =[aij
v
vj]∈CN×M is defined as the postmultiplying SJT product of matrix A and vector V ,
where ◊ represents the SJT product.
12
In what follows we give the operation rules applying the SJT product to compute the
Jacobian matrices, where
v
respect to vector U .
∂
v denotes the Jacobian derivative matrix operator with
∂U
v
{( )}o (A U ) , we have
1. For c(x , y )U ,x = c x j, y j
∂
v
∂U
x
{{c(x y )}o (A Uv )}= A ◊{c(x y )}.
j,
j
x
x
j,
(3-4a)
j
v oq
q
2. For (U ,x ) = (Ax U ) , where q is a real number, we have
∂
v
∂U
{(A Uv ) }= qA ◊(A Uv )
o( q −1)
oq
x
x
.
(3-4b)
(A U ) o (A U ), we have
v
3. For U, x U, y =
∂
v
∂U
x
v
x
v
y
v
v
v
{(A U )o (A U )}= A ◊(A U )+ A ◊(A U )
x
y
4. For sin U, x = sin
x
o
y
y
x
v
(A U ), we have
x
v
v
∂
v sin(Ax U ) = A x ◊ cos o (AxU )
∂U
v
o
5. For exp (U, x ) = exp (Ax U ) , we have
{
}
v
v
∂
v exp o (A xU ) = Ax ◊ exp o (Ax U )
∂U
{
(3-4c)
}
(3-4d)
(3-4e)
The above computing formulas give the analytic solution of the Jacobian derivative
matrix for the analog equations considered. The computational effort for one SJT
product is only n2 scalar multiplications, which may be the smallest computational
cost in all possible approaches. The premultiplying SJT product was also introduced
v
vm
∂W m
=
A
W
x
to compute the Jacobian matrix of the nonlinear formulation such as ∂x
(m≠1) [15, 16]. Such cases are seldom encountered in structure analysis. Therefore, it
is not presented here for the sake of brevity.
The finite difference method is a simple technique to obtain the approximate solution
of the Jacobian matrix in practical engineering and requires O(n2) scalar multiplications.
13
Both the SJT product approach and the finite difference method are essentially
comparable in computing effort. However, the approximate Jacobian matrix yielded
by the finite difference method affects the accuracy and convergence rate of the
Newton-Raphson method. In contrast, the SJT product produces the analytic solution
of the Jacobian matrix.
4. APPLICATIONS
The von Karman equations governing a thin, homogeneous, orthotropic rectangular
plate subject to a uniformly distributed transverse load are given by [9]
(
)
(4-1a)
(
)
(4-1b)
E1 u, xx + µG12 u, yy + Cv ,xy = − w,x E 1 w,xx + µG12 w,yy − Cw,y w,xy
E 2 v , yy + µG12 v ,xx + Cu,xy = − w,y E 2 w,yy + µG12 w,xx − Cw,x w,xy
D1 w,xxxx + 2 D3 w,xxyy + D2 w, yyyy = q +
h
1 2
u,x + w,x E 1 w,xx + υ 12 E 2 w,yy
µ
2
(
1
+ v ,y + w,2y E 2 w,yy + υ 21 E1 w, xx + 2 µG12 w, xy u,y + v ,x + w,x w,y
2
(
)
(
)
)]
(4-1c)
in terms of three displacement components, where ν12 and v21 are PoissonÕs ratio, E1
and E2 are the YoungÕs moduli. C is the shear modulus, D1, D2 and D4 are the
principal bending and twisting rigidities, µ=1-ν12v21, u, v and w are the desired inplane
and transverse displacements. a, b and h are width, length and thickness of plate,
respectively.
Applying the Hadamard product and the new DQ matrix approximate formulas (2-11),
the analog equations for this case are
(
$ )o (A wA
$ )
− C (wA
$ yT + CAx vA
$ yT = − (Ax w$ )o E 1 B x w
$ + µG12 wB
$ yT
E1 B x u$ + µG12 uB
T
y
(
x
)(
)
T
y
)
(4-2a)
$ yT + µG12 B x v$ + CA x uA
$ yT = − wA
$ yT o E 2 wB
$ yT + µG12 B x w$
E 2 vB
(
$ yT
− C (Ax w$ )o Ax wA
14
)
(4-2b)
$ yT + D2 wD
$ yT = q +
D1 Dx w$ + 2 D3 B x wB
h
µ
o2
1
Ax u$ + (Ax w$ )
2
(
T 1
$ + υ 12 E 2 wB
$ yT + vA
$ y + wA
$ yT
o E 1 Bx w
2
[
]
(
)[
$
) o [E wB
o2
2
(
$ yT o uA
$ yT + Ax v$ + (Ax w
$ )o wA
$ yT
+ 2 µG12 Ax wA
T
y
]
+ υ 21 E 1 B x w$
)]]
(4-2c)
where A , B and D with subscript x and y denote the modified weighting coefficient
matrix along x and y directions, respectively. The orders of these matrices are N-2 for
the DQWB approach and N-4 for the DQCY approach, where N is the number of grid
points. Note that the boundary conditions have been applied in the DQWB and
DQCY approach and thus are no longer considered. It is also noted that Ax and Ay
here are different form those in section 3. The latter are the stacked coefficient matrix,
v
v
the corresponding W and U are the desired vectors, while the former are in onedimensional sense and u$ , v$ and w$ in the above formulations are rectangular matrices.
v
v v
It is also pointed out that Ax , B x , Ay and B y for different u, v and w are the same,
respectively, in the cases of clamped or simply supported edges.
The variables are nondimensionalized as X ≡ x a, Y ≡ y b ,U ≡ u$ a ,V ≡ v$ b
and
W ≡ w$ h . Furthermore, by using the Kronecker product of matrices, we have
v
v
v
v
v
v
H1U + H2V = − (H 7W )o (H1W ) − (H8 W )o (H2 W )
(4-3a)
v
v
v
v
v
v
H2 U + H3V = − (H8W )o (H3W )− (H7 W )o (H2 W )
v qa 4
a 4 a 2
H4 W =
+
D1 h µD1 h h 2
(4-3b)
v o2
v b2 v 1
v o2
v 1
+
+
H
U
+
H
W
H
W
H
V
H
W
o
(
)
(
)
8
8
7
5
7
h 2
2
2
(4-3c)
v 2µG12
v
v
v
v
v
o H6 W +
(H 2W )o H8U + H7V + (H7W )o (H8W )
C
v v
v
W
U
V
where H1, H2, H3, H4, H5, H6, H7 and H8 see appendix.
,
and
are vectors
[
]
[
[
]
]
generated by stacking the rows of the corresponding rectangular matrix W, U, and V
into one long vector. By using the new matrix approximate formulas and Hadamard
product, the present nonlinear formulations are very easily accomplished. The matrix
15
form here is also simpler and more explicit than the conventional algebraic polynomial
form given by Bert et al. [9].
Equations (4-3a) and (4-3b) can be also restated as
v
v
v
H1U + H 2V = − L1 (W )
v
v
v
H2 U + H 3V = −L2 (W) ,
(4-4a)
(4-4b)
where
v
v
v
v
v
L1 (W ) = (H7W )o (H1W )+ (H8 W )o (H2 W )
(4-5a)
v
v
v
v
v
L2 (W ) = (H8W )o (H 3W )+ (H7 W )o (H2 W ) .
(4-5b)
v
v
v
The unknown vector U and V can be expressed in terms of W by
v
v
v
U = H9−1 H3−1 L2 (W ) − H9−1 H2−1 L1 (W )
(4-6a)
and
v
v
v
V = H10−1 H 2−1 L 2 (W ) − H 10−1 H 1−1 L1 (W ),
(4-6b)
−1
−1
where H9 = H2 H1 − H3 H 2
−1
−1
and H10 = H1 H2 − H2 H3 . By
product for the evaluation of the Jacobian matrix, we have
v
v
v
∂L2 (W)
∂L1 (W)
∂U
−1
−1
−1
−1
v = H 9 H3
v − H 9 H2
v
∂W
∂W
∂W
and
v
v
v
L
W
L
W
∂
∂
(
)
(
)
∂V
2
1
v = H10−1 H2−1
v − H10−1 H1−1
v ,
∂W
∂W
∂W
where
v
∂L1 (W )
v
v
v
v
v = H7 ◊(H1W )+ H1 ◊(H7 W )+ H8 ◊(H2 W )+ H2 ◊(H8 W )
∂W
applying the SJT
(4-7a)
(4-7b)
(4-8a)
and
v
∂L2 (W )
v
v
v
v
v = H8 ◊(H3W )+ H3 ◊(H8W )+ H7 ◊(H2 W )+ H2 ◊(H7W ) .
∂W
16
(4-8b)
v
v
v
∂U
∂V
v and v are relative Jacobian derivative matrices of dependent variable vector U
∂W
∂W
v
v
and V with respect to W . Applying formulas (4-6a, b) and (4-7a, b), the coupled
formulations (4-3a, b, c) are decoupled. The size of the resulting nonlinear
simultaneous algebraic equations are reduced from 3(N-2)×3(N-2) to (N-2)×(N-2) for
the DQWB approach or from 3(N-4)×3(N-4) to (N-4)×(N-4) for the DQCY approach.
It is known that each iteration step in the Newton-Raphson method has to solve a set
of linear simultaneous algebraic equations, which requires an order of n3 scalar
multiplications. n is the size of the solved equations. For example, the Gauss
elimination method requires n3/3 scalar multiplications, while the Gauss-Jordan
method requires n3/2 scalar multiplications. Therefore, the computational effort and
storage requirements here are only about one twenty-seventh and one-ninth,
v
respectively, as much as in reference [9]. W is a basic variable vector in the present
computations. Equation (4-3c) is chosen as the basic iteration equation, namely,
4
2
v
v
a a
ϕ W = H4 W −
µD1 h h 2
v o2
v b2 v 1
v o2
v 1
o
H
U
H
W
H
W
+
H
V
H
W
+
+
(
)
(
)
7
7
5
8
8
2
2
2
h
v
v
v
v
v
v qa 4
2µG12
o H6 W +
H 2W )o H8U + H7V + (H7W )o (H8W ) −
=0
(
C
D1h
[
{}
[
]
]
[
]
(4-9)
The Jacobian derivative matrix for the above iteration equation is given by
v
v
4
2
∂ϕ W
v
v
a a ∂U
v = H4 −
2 H7 v + H7 ◊(H 7W ) ◊ H5W
∂W
µD1 h h ∂W
v
2
v o 2 b 2 ∂V
v
v
a
v 1
+ 2 H5 ◊ H7 U + (H7 W ) + 2 H8 v + H8 ◊(H8 W ) ◊ H6 W
2
h
h ∂W
(4-10)
2
v o2 2µG12
v
v
v
v
b
v 1
+ 2 H6 ◊ H8 U + (H8 W ) +
H2 ◊ H8U + H7V + (H7 W )o (H8 W )
2
h
C
v
v
v
v
v
∂V
2 µG12 ∂U
+
H8 v + H7 v + H7 ◊(H8W )+ H8 ◊(H7 W ) ◊(H2 W )
∂W
C ∂W
{}
[
]
[
[
]
]
It is noted that the SJT product approach here yields the analytical solution of the
Jacobian matrix quite simply and efficiently. The Newton-Raphson iteration equation
for this case is
17
−1
v
∂ϕ (W ( k ) )
v ( k + 1)
v (k )
v
= W
−
W
v ϕ (W (k ) )
∂W
(4-11)
The DQ solutions for the corresponding linear isotropic and orthotropic plates are
chosen as the initial guess of the iteration procedures. It is noted that the NewtonRaphson method has rather big convergence domain in the present computations.
Even if the resulting nonlinear results are even eight times larger than the initial linear
solutions, the Newton-Raphson method still converges. M oreover, the solution
converges very rapidly, and the iterative times varies from 1 to 10 for various loadings
when the convergence criterion is the maximum residuals of equation (4-9) no more
than 10-5. In contrast, the IM SL subroutine NEONE used in the reference [9]
computed the Jacobian matrix approximately by a finite difference technique.
Therefore, as was shown in Fig. 8 in reference [9], the accuracy and converging rate of
the DQ solutions are affected, especially for simply supported plates.
References [8, 12, 17] pointed out that the accuracy and stability of the DQ method
can be improved significantly if the Chebyshev grid spacing is used. In the following
the DQWB and DQCY solutions are obtained by using Chebyshev grid spacing 7×7
for a simply supported plate and 9×9 for a clamped plate, respectively, unless where
specified. To avoid the effects of round-off errors on the accuracy of the solution,
double-precision arithmetic is used in all the results presented in this paper. Reference
[9] has pointed out the high efficiency and ease of use in the DQ method in
comparison to other numerical techniques such as the finite element, finite difference,
perturbation, Galerkin and Rayleigh-Ritz, etc., while this paper places its emphasis in
the simplification of the use and further improvement of the accuracy and efficiency in
the DQ method. Therefore, the comparisons with the other numerical techniques are
not discussed here.
The same simply-supported and clamped isotropic square plates subject to a
uniformly distributed loading as in example 1 of reference [9] are recalculated by the
18
present DQ method. The solutions by both the DQCY and conventional DQδ
approaches are nearly agreement and very accurate for the clamped plate, and, thus,
the DQCY solutions are not presented here for the sake of brevity. Nevertheless, it
should be emphasized that the DQδ approach can not use larger number of grid points
as in the DQCY approach due to instability caused by the δ effect. The results for
simply supported plate are shown in Fig. 1 and compared with the exact [24, 25] and
the conventional DQ [9] solutions. The present DQWB results show remarkable
agreement with those of Levy [24] and Yang [25]. It is also noted that the DQδ
approach using 7×7 grid points by Bert et al. [9] gave obviously better results in the
clamped cases than in the simply supported cases as shown in Figs 1 and 2 of
reference [9]. This is because the DQδ approach is not very suitable for the cases of
supported edges. As is expected, the DQWB approach gives more accurate results
than the DQδ approach for simply supported plate. Therefore, the former is a
competitive alternative to the latter for the nonlinear cases of simple supports.
In addition, we compute an isotropic simply supported square plate under a
uniformly distributed transverse load. The parameters of this case are a=16Ó, h=0.1Ó,
E=30E+6 and ν=0.316. Fig. 2 depicts the results obtained by the DQWB approach
using 5×5 Chebyshev grids and 7×7 equally spaced grids. All solutions agree very well
with those given by Levy [26]. As is expected, the DQWB method using the 5×5
Chebyshev grids yields more accurate results than using 7×7 equally spaced grids.
The center deflections of the clamped square plate (a=100, h=1.0, E=2.1E+6, ν=0.316,
q=3.0) and the simply-supported square plate (a=100, h=1.0, E=2.1E+6, ν=0.25,
q=1.0) subject to a uniformly distributed pressure are computed, respectively, by the
DQWB and DQCY methods. The DQ results as well as the analytical and FEM
solutions are listed in Table 1. The present DQ solutions show excellent agreement
with the analytical [27] and FEM solutions [27, 28]. However, the computational
19
effort in the present DQ method is much less than in the analytical method and FEM .
The DQ method is demonstrated again to be highly computationally efficient for
nonlinear structural analysis.
The numerical examples on the orthotropic rectangular plate provided by Bert et al.
[9] are recalculated by the present DQ methods. The specific parameters are
E1=18.7E+6psi; E2=1.3E+6psi, G12=0.6E+6psi; ν12=0.3; h=0.0624inch; a=9.4inch;
b=7.75inch. The center deflections in the clamped and simply supported cases are
displayed in Figs. 3 and 4, respectively. For the cases of clamped edges, the results by
the DQCY approach are very close to those by Bert and Cho. It is noted that the
DQCY approach using 15×15 grid spacing is still stable and give accurate results, but
computational effort also increases exponentially, while the DQδ approach can not
use so many grid points. The computational stability problem may be essential in
some cases when many grid points are required. Therefore, the DQCY approach may
have better promise in practical engineering than the DQδ approach. In this case, the
DQCY approach produces more accurate results by using more grid points (e.g., 9×9
grid spacing).
For the case of simple supports, it is noted that the results given by Bert, Striz and
Jang [9] using DQδ approach with 7×7 equally spaced grids are obviously larger than
those by Bert and Cho. In contrast, the present DQWB approach appears to give
results that are much closer to those by Bert and Cho as shown in Fig. 4. Bert and
ChoÕs solution values here are taken from graphs 9 and 10 in respective reference [9]
with appropriate scaling factors. Another example on orthotropic plates discussed in
reference [9] are also recalculated by the present DQ method and the same conclusions
are obtained.
20
Based on the centrosymmetric structures of the DQ weighting coefficient matrices, the
reduction technique in the DQ method was proposed in [16, 17, 20] and extended to
the nonlinear problems in [16]. For geometrically nonlinear bending of the isotropic
and orthotropic rectangular plates with symmetric boundary conditions, the
computational effort and storage requirements can be further reduced by 75% and
50% using such reduction technique, respectively.
5. CONCLUDING REMARKS
The DQ approach in using some new techniques is applied to analyze geometrically
nonlinear bending of isotropic and orthotropic plates with simply supported and
clamped edges. The new matrix approximate formulas offer a compact and convenient
DQ procedure to approximate partial derivatives. The DQWB approach is proved to
be a successful technique for geometrically nonlinear plate with SS-SS boundary
conditions. It is conceivable that the DQWB approach is also highly efficient for the
other boundary conditions whenever applicable as in the linear problems. The DQCY
approach is improved and shown to be a stable and accurate technique for handling the
cases with the C-C boundary conditions. Apparently, the results obtained by these
two approaches are more accurate than those by the traditional DQδ approach.
Although only simply-supported and clamped boundary conditions are involved in
the present study, it is straightforward that the same procedures can be easily
employed for handling problems with the other boundary conditions.
The publications in which the DQ method was applied to deal with nonlinear
problems are still few due to much more complex programming, storage requirements,
formulation and computing effort in comparison to linear problems. The Hadamard
and SJT product approach may provide a simple and efficient technique to greatly
reduce these impediments. The detailed solution procedures for the geometrically
nonlinear plate cases are provided here to show the simplicity and efficiency of the
Hadamard and SJT product approach. It is worth stressing that the Hadamard and
21
SJT products as well as the relative decoupling technique are applicable for the finite
difference, pseudo-spectral and collocation methods. Also, although only the
application of Hadamard and SJT product within one single interval is presented in
this paper, the extension of this procedure to solutions of problems with complex
geometries with the coordinate mappings and multidomain techniques should be
straightforward. The application of the present DQ solution procedure to the
nonlinear analysis of more complex plates with varying thickness, PoissonÕs ratio and
YoungÕs modulus is currently the subject of further investigation.
REFERENCES
1. R. E. Bellman and J. Casti, Differential quadrature and long-term integration. J.
Math. Anal. Appl. 34, 235-238 (1971).
2. R. E. Bellman, B. G. Kashef and J. Casti, Differential quadrature: A technique
for the rapid solution of nonlinear partial differential equations. J. Comput.
Phys.10, 40-52 (1972).
3. J. O. M ingle, The method of differential quadrature for transient nonlinear
diffusion. J. Math. Anal. Appl. 60, 559-569 (1977).
4. F. Civan and C. M . Sliepcevich, Differential quadrature for multidimensional
problems. J. Math. Anal. Appl. 101, 423-443 (1984).
5. G. Naadimuthu, R. E. Bellman, K. M . Wang and E. S. Lee, Differential quadrature
and partial differential equations: some numerical results. J. M ath. Anal. Appl. 98,
220-235 (1984).
6. C. W. Bert, S. K. Jang and A. G. Striz, Two new methods for analyzing free
vibration of structure components. AIAA J. 26, 612-618 (1988).
7. A. G. Striz, S. K. Jang and C. W. Bert, Nonlinear bending analysis of thin circular
plates by differential quadrature. Thin-Walled Struct. 6, 51-62 (1988).
8. J. R. Quan and C. T. Chang, New insights in solving distributed system equations
by the quadrature methods - 2: Numerical Experiments. Comput. Chem. Engrg. 13,
1017-1024 (1989).
22
9. C. W. Bert, A. G. Striz and S. K. Jang, Nonlinear bending analysis of orthotropic
rectangular plates by the method of differential quadrature. Comput. M ech. 5,
217-226 (1989).
10. Y. Feng and C. W. Bert, Application of the quadrature method to flexural vibration
analysis of a geometrically nonlinear beam, Nonlinear Dynamics. 3, 13-18 (1992).
11. C. Shu and B. E. Richards, Parallel simulation of incompressible viscous flows by
generalized differential quadrature. Computing Systems in Engrg. 3, 271-281
(1992).
12. C. W. Bert, X. Wang and A. G. Striz, Differential quadrature for static and free
vibrational analysis of anisotropic plates. Int. J. Solids Struct. 30, 1737-1744
(1993).
13. P. A. A. Laura and R. E. Rossi, ÒThe method of differential quadrature and its
application to the approximate solution of ocean engineering problems,Ó Ocean
Eng. 21, 57-66(1994).
14. M . M alik and F. Civan, A comparative study of differential quadrature and
cubature methods vis-ˆ-vis some conventional techniques in context of convectiondiffusion-reaction problems. Chem. Engrg. Sci. 50, 531-547 (1994).
15. Wen Chen and Tingxiu Zhong, The study on nonlinear computations of the DQ
and DC methods. Numer. M ethods for Partial Differential Equations 13, 57-75
(1997).
16. Wen Chen, Differential Quadrature method and its applications in Engineering,
Ph.D. dissertation, Shanghai Jiao Tong University, China, 1996.
17. J. R. Quan and C. T. Chang, New insights in solving distributed system equations
by the quadrature methods - 1. Comput. Chem. Engrg 13, 779-788 (1989).
18. G. M ansell, W. M erryfield, B. Shizgal and U. Weinerl, A comparison of
differential quadrature methods for the solution of partial differential equations,
Comput. Methods Appl. Mech. Engrg. 104, 295-316 (1993).
19. M . K. Burka, Solution of stiff ordinary differential equations by decomposition
and orthogonal collocation. AIChE J. 28, 11-20 (1982).
23
20. Wen Chen, X, Wang and Y. Yong, Reducing the computational effort of the
differential quadrature method,
Numerical M ethods for Partial Differential
Equations 12, 565-577 (1996).
21. X. Wang and C. W. Bert, A new approach in applying differential quadrature to
static and free vibrational analyses of beams and plates. J. Sound & Vibr. 162,
566-572 (1993).
22. X. Wang, C. W. Bert and A. G. Striz, Differential quadrature analysis of deflection,
buckling and free vibration of beams and rectangular plates. Comput. Struct. 48,
473-479 (1993).
23. Wen Chen and Y. Yu, Differential quadrature method for high order boundary
value problems. In: Proc. 1st Pan-Pacific Conf. on Comput. Engrg. (Kwak, B. M .,
Tanaka, M. eds.). Elsevier Sci. Publ. B. V. Netherlands, pp. 163-168 (1993).
24. T. Y. Yang, Finite displacement plate flexure by the use of matrix incremental
approach, Int. J. Numer. Meths. Engrg. 4, 415-432 (1972).
25. S. Levy, Square plate with clamped edges under normal pressure producing large
deflections. NACA Report No. 740 (1942).
26. S. Levy, Bending of rectangular plates with large deflections. NACA Report No.
737. 139-157 (1942).
27. G. Zhu and H. Wang, Quasi-conforming penalty FEM for large deflection of
composite laminated plate (in Chinese). ACTA M ATERIAE COM POSITE
SINICA, 6, 39-47 (1989).
28. G. P. Bazeley, Y. K. Cheung, B. M . Irens and O. C. Zienkiewicz, Triangular
elements in bending-conforming and non-conforming solutions, Proc. Conf. M atrix
M ethods in Structural M echanics, J. S. Przemieniecki et al. eds. Chio: Air Force
Ins. Tech. Wright-Patterson A F Base. 547-576 (1965).
29. P. Lancaster and M . Timenetsky, The Theory of M atrices with Applications, 2nd
edition, Academic Press. Orlando (1985).
24
APPENDIX
The desired unknowns in rectangular matrix form as in equations (4-2a, b, c) can be
converted into the conventional vector form by using the Kronecker product [29].
Lemma 1. If A∈ Cp×m, B∈ Cn×q and the unknown X∈ Cm×n, then
vec( AXB ) = ( A ⊗ B T ) vec( X )
(A1)
where vec( ) is the vector-function of a rectangular matrix formed by stacking the rows
of matrix into one long vector, ⊗ denotes the Kronecker product of matrices. Note
that vec( ) here is to stack rows rather than columns as in reference [29] and, thus, the
corresponding operation rules are also slightly different, but both are in fact equivalent.
v
v
In this paper, we define vec( AXB) = AXB and vec( X ) = X to simplify presentation.
Corollary:
v
v
1. AX = ( A ⊗ I n )X
v
v
2. XB = I m ⊗ B T X
v
v
v
T
3. AX + XB = A ⊗ I n + I m ⊗ B X
(
(
)
(A2)
)
where In and Im are the unit matrix. According to the above formulas, Eqs. (4-2a, b, c)
can be converted into Eqs. (4-3a, b, c), and the resulting coefficient matrices in Eqs.
(4-3a, b, c) are given by
H1 = E1 B x ⊗ I y
(
)
(
)
H2 = C Ax ⊗ A y
2
a
+ µG12
b
)
H4 = Dx ⊗ I y +
2
x
⊗ By
D2 a
D1 b
2
(B
x
2
(B
x
⊗ Iy
)
⊗ By +
h
h
H5 = E 1 B x ⊗ I y + υ 12 E 2
a
b
(A3)
(A5)
D3 a
D1 b
(I
x
⊗ Dy
)
(A6)
(A7)
h
h
H6 = E 2 I x ⊗ B y + υ 21 E 1 B x ⊗ I y
b
a
)
(A8)
)
2
(I
4
)
(
)
2
)
⊗ By
2
(
)
(A4)
b
H3 = E 2 I x ⊗ B y + µG12
a
(
(I
x
(
25
h
Ax ⊗ I y
a2
)
(A9)
h
H8 = 2 I x ⊗ A y
b
)
(A10)
H7 =
(
2
(
2
Table 1. The Center deflections of the clamped and simply-supported square
plates
Analytical[
Methods
FEM [28]
FEM
[27]
0.940
1.028(9.3%)
0.942(0.3%)
0.944(0.4%)
1.151
1.316(14.3%)
1.170(1.6%)
1.123(2.4%)
Present
27]
Simply
supported
Clamped
2
1.8
1.6
W/h
1.4
1.2
1
DQ (Bert et al. 1989)
0.8
Present DQ
0.6
Levy
0.4
Yang (1972)
0.2
0
0
50
100
150
200
qa4 /Eh4
Fig. 1. Central deflections for a simply supported square isotropic plate
26
2
1.8
1.6
W/h
1.4
1.2
Levy
1
0.8
DQ (Chebyshev grids 5×5)
0.6
0.4
DQ (equally spaced grids
7×7)
0.2
0
0
2
4
6
8
10
12
14
16
Normal pressure q (p.s.i)
Fig. 2. Comparison of the DQ accuracies of central deflections for a square simply
supported plate using the Chebyshev 5×5 and the equally spaced 7×7 grids.
27
0.16
Center deflection (in)
0.14
0.12
0.1
0.08
0.06
Bert and Cho
0.04
Present DQ
0.02
DQ (Bert et al.,
1989)
0
0
1
2
3
4
5
Normal pressure q (p.s.i)
Fig. 3. Central deflections for a clamped square orthotropic plate
0.2
Center deflection (in)
Bert and Cho
Present DQ
0.15
DQ (Bert et al., 1989)
0.1
0.05
0
0
0.5
1
1.5
2
2.5
3
Normal pressure q (p.s.i)
Fig. 4. Central deflections for a simply supported square orthotropic plate.
28
| 5cs.CE
|
Gradient-Based Meta-Learning with Learned Layerwise
Metric and Subspace
Yoonho Lee and Seungjin Choi
Department of Computer Science and Engineering
Pohang University of Science and Technology
arXiv:1801.05558v2 [stat.ML] 19 Feb 2018
77 Cheongam-ro, Nam-gu, Pohang 37673, Korea
{einet89, seungjin}@postech.ac.kr
February 20, 2018
Abstract
Gradient-based meta-learning has been shown to be expressive enough to approximate any learning
algorithm. While previous such methods have been successful in meta-learning tasks, they resort to simple
gradient descent during meta-testing. Our primary contribution is the MT-net, which enables the meta-learner
to learn on each layer’s activation space a subspace that the task-specific learner performs gradient descent on.
Additionally, a task-specific learner of an MT-net performs gradient descent with respect to a meta-learned
distance metric, which warps the activation space to be more sensitive to task identity. We demonstrate that
the dimension of this learned subspace reflects the complexity of the task-specific learner’s adaptation task,
and also that our model is less sensitive to the choice of initial learning rates than previous gradient-based
meta-learning methods. Our method achieves state-of-the-art or comparable performance on few-shot
classification and regression tasks.
1
Introduction
While recent deep learning methods achieve superhuman performance on various tasks including image
classification [16] or playing games [23], they can only do so using copious amounts of data and computational
resources. In many problems of interest, learners may not have such luxuries. Meta-learning [29, 30, 33]
methods are a potential solution to this problem; these methods leverage information gathered from prior
learning experience to learn more effectively in novel tasks. This line of research typically casts learning as
a two-level process, each with a different scope. The meta-learner operates on the level of tasks, gathering
information from several instances of task-specific learners. A task-specific learner, on the other hand, operates
on the level of datapoints, and incorporates the meta-learner’s knowledge in its learning process.
Model-agnostic meta-learning (MAML) [5] is a meta-learning method that directly optimizes the gradient
descent procedure of task-specific learners. All task-specific learners of MAML share initial parameters, and a
meta-learner optimizes these initial parameters such that gradient descent starting from such initial parameters
quickly yields good performance. An implicit assumption in having the meta-learner operate in the same space
as task-specific learners is that the two different scopes of learning require equal degrees of freedom.
Our primary contribution is the MT-net (Figure 1), a neural network architecture and task-specific learning
procedure. An MT-net differs from previous gradient-based meta-learning methods in that the meta-learner
determines a subspace and a corresponding metric that task-specific learners can learn in, thus setting the
degrees of freedom of task-specific learners to an appropriate amount. The activation space of the cell shown
1
(a)
(b)
(c)
Figure 1: Task-specific learning in an MT-net. (a) A cell (rounded rectangle) consists of two layers. In addition
to initial weights (black), the meta-learner specifies weights to be changed (dotted lines) by task-specific
learners (colored). (b) Activation of this cell has 3 dimensions, but activation of task-specific learners only
change within a subspace (white plane). (c) The value of T affects task-specific learning so that gradients of
W are sensitive to task identity. Best seen in color.
in Figure 1 is 3-dimensional, but because the task-specific learners can only change weights that affect two of
the three intermediate activations, task-specific learning only happens on a subspace with 2 degrees of freedom.
Additionally, meta-learned parameters T alter the geometry of the activation space of task-specific parameters
so that task-specific learners are more sensitive to change in task.
2
2.1
Background
Problem Setup
We briefly explain the meta-learning problem setup which is applied to few-shot tasks.
The problems of k-shot regression and classification are as follows. In the training phase for a meta-learner,
we are given a (possibly infinite) set of tasks {T1 , T2 , T3 , . . .}. Each task provides a training set and a test set
{DTi ,train , DTi ,test }. We assume here that the training set DTi ,train has k examples per class, hence the name
k-shot learning. A particular task T ∈ {T1 , T2 , T3 , . . .} is assumed to be drawn from the distribution of tasks
p(T ). Given a task T ∼ p(T ), the task-specific model fθT (a feedforward neural network is considered in this
paper) parameterized by θT is trained using the dataset DT ,train and its corresponding loss LT (θT , DT ,train ).
Denote by θeT parameters obtained by optimizing nLT (θT , DT ,train ).o Then, the meta-learner fθ is updated
using the feedback from the collection of losses LT (θeT , DT ,test )
, where the loss of each task
T ∼p(T )
is evaluated using the test data DT ,test . Given a new task Tnew (not considered during meta-training), the
meta-learner helps the model fθTnew to quickly adapt to the new task Tnew , by warm-starting the gradient
updates.
2.2
Model-Agnostic Meta-Learning
We briefly review model-agnostic meta-learning (MAML) [5], emphasizing commonalities and differences
between MAML and our method. MAML is a meta-learning method that applies to any model that learns
using gradient descent. This method is loosely inspired by fine-tuning, and it learns initial parameters of a
network such that the network’s loss after a few gradient steps is minimized.
2
Let us consider a model that is parameterized by θ. MAML alternates between the two updates (1) and (2)
to determine initial parameters θ for task-specific learners to warm-start the gradient descent updates, such that
new tasks can be solved using a small number of examples. Each task-specific learner updates its parameters
by the gradient update (1), using the loss evaluated with the data {DT ,train }. The meta-optimization across
tasks is performed such that the parameters θ are updated using the loss evaluated with {DT ,test }, which is
given in (2).
θeT
θ
← θ − α∇θ LT (θ, DT ,train )
X
← θ − β∇θ
LT θeT , DT ,test ,
(1)
(2)
T ∼p(T )
where α > 0 and β > 0 are learning rates and the summation in (2) is computed using minibatches of tasks
sampled from p(T ).
Intuitively, a well-learned initial parameter θ is close to some local optimum for every task T ∼ p(T ).
Furthermore, the update (1) is sensitive to task identity in the sense that θeT1 and θeT2 have different behaviors
for different tasks T1 , T2 ∼ p(T ).
Recent work has shown that gradient-based optimization is a universal learning algorithm [4], in the sense
that any learning algorithm can be approximated up to arbitrary accuracy using some parameterized model and
gradient descent. Thus, no generality is lost by only considering gradient-based learners as in (1).
Our method is similar to MAML in that our method also differentiates through gradient update steps to
optimize performance after fine-tuning. However, while MAML updates all parameters in θ to make θeT , our
method only alters a (meta-learned) subset of its weights. Furthermore, whereas MAML learns with standard
gradient descent, a subset of our method’s parameters effectively ’warp’ the parameter space of the parameters
to be learned during meta-testing to enable faster learning.
3
Meta-Learning Models
We present our two models in this section: Transformation Networks (T-net) and Mask Transformation
Networks (MT-net), both of which are trained by gradient-based meta-learning. A T-net learns a metric in its
activation space; this metric informs each task-specific learner’s update direction and step size. An MT-net
additionally learns which subset of its weights to update for task-specific learning. Therefore, an MT-net learns
to automatically assign one of two roles (task-specific or task-mutual) to each of its weights.
3.1
T-net
We consider a model fθ (·), parameterized by θ. This model consists of L cells, where each cell is parameterized∗ as TW:
fθ (x)
= TL WL σ TL−1 WL−1 . . . σ T1 W1 x
,
(3)
where x ∈ RD is an input, and σ(·) is a nonlinear activation function. T-nets get their name from transformation
matrices (T) because the linear transformation defined by a T plays a crucial role in meta-learning. Note that a
∗ For convolutional cells, W is a convolutional layer with some size and stride and and T is a 1 × 1 convolution that doesn’t change
the number of channels
3
Figure 2: A diagram of the adaptation process of a Transformation Network (T-net). Blue values are metalearned and shared across all tasks. Orange values are different for each task.
cell has the same expressive power as a linear layer. Model parameters θ are a collection of W’s and T’s, i.e.,
1
L
1
L
.
θ = W ,...,W ,T ,...,T
{z
} |
{z
}
|
θW
θT
Parameters θT , which are shared across task-specific models, are determined by the meta-learner. All taskspecific learners have the same initial θW but update to different values since each uses their corresponding
train set DT ,train . Thus we denote such (adjusted) parameters for task T as θeW,T . Though they may look
similar, T denotes tasks whereas T denotes a transformation matrix.
Given a task T sampled from p(T ), each W is adjusted with the gradient update
f T ← W − α∇W LT (θW , θT , DT ,train ) .
W
(4)
f1 , . . . , W
f L }. Using the task-specific learner θeW,T , the meta-learner improves
Again, θeW,T is defined as {W
T
T
itself with the gradient update
X
θ ← θ − β∇θ
(5)
LT θeW,T , θT , DT ,test .
T ∼p(T )
α > 0 and β > 0 are learning rate hyperparameters. We show our full algorithm in Algorithm 1.
Suppose that we are given a new task T∗ with the training set DT∗ ,train . The model parameters θeW,T∗
are computed as (4), where the gradient update starts from the initial value θW that was determined by the
meta-learner.
4
Algorithm 1 Transformation Networks (T-net)
Require: p(T )
Require: α, β
1: randomly initialize θ
2: while not done do
3:
Sample batch of tasks Ti ∼ p(T )
4:
for all Tj do
5:
for i = 1, · · · , L do
f T according to (4)
6:
Compute W
7:
end for
f1 , · · · W
fL }
8:
θeW,Tj = {W
Tj
Tj
9:
end for
P
10:
θ ← θ − β∇θ j LT (θeW,Tj , θT , DTj ,test )
11: end while
We now briefly examine a single cell:
y = TWx,
where x is the input to the cell and y its output. The squared length of a change in output ∆y = y∗ − y0 is
calculated as
>
k∆yk2 = ((∆W)x) T> T ((∆W)x) ,
(6)
where ∆W is similarly defined as W∗ − W0 . We see here that the magnitude of ∆y is determined by the
interaction between (∆W)x and T> T. Since a task-specific learner performs gradient descent only on W
and not T, the change in y resulting from (4) is guided by the meta-learned value T> T. We provide further
analysis of this behavior in Section 4.
3.2
MT-net
The MT-net is built on the same feedforward model (3) as the T-net:
fθ (x)
= TL WL σ TL−1 WL−1 . . . σ T1 W1 x
,
(7)
where the MT-net differs from the T-net is in the binary mask applied to the gradient update to determine
f T is given by
which parameters are to be updated. The update rule for task-specific parameters W
f T ← W − αM
W
∇W L(θW , θT , DT ,train ),
(8)
where is the Hadamard (elementwise) product between matrices of the same dimension. M is a binary
gradient mask which is sampled each time the task-specific learner encounters a new task. Each row of M is
either an all-ones vector 1 or an all-zeros vector 0. We parameterize the probability of row j in M being 1
with a scalar variable ζj :
[m1 , . . . , mn ]> ,
exp (ζj )
∼ Bern
1> ,
exp (ζj ) + 1
M =
m>
j
5
(9)
Figure 3: A diagram of the adaptation process of a Mask Transformation Network (MT-net). Blue values are
meta-learned and shared across all tasks. Orange values are different for each task.
where Bern(·) denotes the Bernoulli distribution. Each logit ζ acts on a row of a weight matrix W, so weights
that contribute to the same immediate activation are updated or not updated together.
We backpropagate through the Bernoulli sampling of masks using the Gumbel-Softmax estimator [12]:
g1 , g2
m>
j
∼
Gumbel(0, 1),
ζ +g
exp j T 1
←
ζ +g
exp j T 1 + exp
(10)
g2
T
>
1 ,
(11)
where T is a temperature hyperparameter. This reparameterization allows us to directly backpropagate through
the mask, which at the limit of zero temperature, follows the behavior of (9).
f1 , . . . , W
f L }. The meta-learner
As in T-nets, we denote the collection of altered weights as θeW,T = {W
T
T
learns all parameters θ:
1
L
1
L
1
L
θ = W ,...,W ,T ,...,T ,ζ ,...,ζ , .
(12)
|
{z
} |
{z
} | {z }
θW
θT
θ
ζ
As in a T-net, the meta-learner performs stochastic gradient descent on LT θeW,T , θT , θζ , DT ,test :
X
θ ← θ − β∇θ
LT θeW,T , θT , θζ , DT ,test .
(13)
T ∼p(T )
6
Algorithm 2 Mask Transformation Networks (MT-net)
Require: p(T )
Require: α, β
1: randomly initialize θ
2: while not done do
3:
Sample batch of tasks Ti ∼ p(T )
4:
for all Tj do
5:
for i = 1, · · · , L do
6:
Sample binary mask Mi according to (11)
f i according to (8)
7:
Compute W
Tj
8:
end for
f1 , · · · W
fL }
9:
θeW,Tj = {W
Tj
Tj
10:
end for
P
11:
θ ← θ − β∇θ j LT θeW,T , θT , θζ , DT ,test
12: end while
The full algorithm is shown in Algorithm 2.
We emphasize that the binary mask used for task-specific learning (M) depends on meta-learned parameter
weights (ζ). Since the meta-learner optimizes the loss in a task after a gradient step (8), the matrix M gets
assigned a high probability of having value 1 for weights that encode task-specific information. Furthermore, since we update M along with model parameters W and T, the meta-learner is incentivized to learn
configurations of W and T in which there exists a clear divide between task-specific and task-mutual neurons.
4
Analysis
In this section, we provide further analysis of the update schemes of T-nets and MT-nets.
Throughout this section, we focus on the space of y instead of A in a layer parameterized as y = Ax. This
is because when thinking about gradients with respect to a loss function, the two are equivalent. Note that the
influence of A on the loss function LT is bottlenecked by y. The chain rule shows that ∇A LT = (∇y LT )x> .
Assuming x is fixed, the space of possible ∇A LT under all loss functions is isomorphic to ∇y LT , which is in
turn isomorphic to Rn (n is the the dimension of y). We take advantage of this fact by learning a full-rank
(n × n) metric in the space of y; doing this in the space of A would require too many parameters for even a
small architecture.
4.1
T-nets Learn a Metric in Activation Space
We consider a cell in a T-net where the pre-activation value y is given by
y = TWx = Ax,
(14)
where A = TW and x is the input to the cell. We omit superscripts throughout this section.
A standard feedforward network resorts to the gradient of a loss function LT (which involves a particular
task T ∼ p(T )) with respect to the parameter matrix A, to update model parameters. In such a case, a single
gradient step yields
ynew
=
(A − α∇A LT )x
=
y − α∇A LT x.
7
(15)
The update of a T-net (4) results in the following new value of y:
ynew
=
=
T T−1 A − α∇T−1 A LT x
y − α TT> ∇A LT x,
(16)
where T is determined by the meta-learner.
Thus, in a T-net, the incremental change of y is proportional to
the negative of the gradient TT> ∇A LT , while the standard feedforward net resorts to a step proportional
to the negative of ∇A LT . Task-specific learning in the T-net is guided by a full rank metric in each cell’s
activation space, which is determined by each cell’s transformation matrix T. This metric (TT> )−1 warps
(scaling, rotation, etc.) the activation space of the model so that in this warped space, a single gradient step
with respect to the loss of a new task yields parameters that are well suited for that task.
4.2
MT-nets Learn a Subspace with a Metric
We now consider MT-nets and analyze what their update (8) means from the viewpoint of y = TWx = Ax.
MT-nets can restrict its task-specific learner to any subspace of its gradient space:
Proposition 1. Fix x and A. Let y = TWx be a cell in an MT-net and let ζ be its corresponding mask
parameters. Let U be a d-dimensional subspace of Rn (d ≤ n). There exist configurations of T, W, and ζ
such that the span of ynew − y is U while satisfying A = TW.
Proof. See Appendix B.
This proposition states that W, T, and ζ have sufficient expressive power to restrict updates of y to any
subspace. Note that this construction is only possible because of the transformation T; if we only had binary
masks M, we would only be able to restrict gradients to axis-aligned subspaces.
In addition to learning a subspace that we project gradients onto (U), we are also learning a metric in this
subspace. We first provide an intuitive exposition of this idea.
We unroll the update of an MT-net as we did with T-nets in (16):
ynew =T((T−1 A − αM
=y − αT(M
(T ∇A LT ))x
=y − αT(MT
=y − α(T
∇T−1 A LT )x)
>
T> )∇A LT x
M>
T )(MT
T> )∇A LT x.
(17)
Where MT is an m × m matrix which has the same columns as M. Let’s denote TM = MT T> . We
see that the update of a task-specific learner in an MT-net performs the update T>
M TM ∇A LT . Note that
T>
T
is
an
n
×
n
matrix
that
only
has
nonzero
elements
in
rows
and
columns
where
m is 1. By setting
M
M
appropriate ζ, we can view T>
T
as
a
full-rank
d
×
d
metric
tensor.
M
M
This observation can be formally stated as:
Proposition 2. Fix x, A, and a loss function LT . Let y = TWx be a cell in an MT-net and let ζ be its
corresponding mask parameters. Let U be a d-dimensional subspace of Rn , and g(·, ·) a metric tensor on U.
There exist configurations of T, W, and ζ such that the vector ynew − y is in the steepest direction of descent
on LT with respect to the metric g(·, ·).
Proof. See Appendix B.
8
Therefore, not only can MT-nets project gradients of task-specific learners onto a subspace of the preactivation (y) space, they can also learn a metric in that subspace and thereby learning a low-dimensional
linear embedding of the activation space. The MT-net update (8) is gradient descent in this low-dimensional
embedding, so the meta-objective shown in (13) is minimized when gradient descent in this embedding
requires few steps to converge and is sensitive to task identity.
5
Related Work
A successful line of research in few-shot learning uses feedforward neural networks as learners. These
approaches learn update rules [25, 19, 2] or directly generate weights [9]. A related research direction is to
learn initial parameters [5] while fixing the learning rule to gradient descent, or additionally learning learning
rates for each weight [20]. [8] interprets such gradient-based meta-learning as hierarchical bayesian inference,
and [4] states that such methods are expressive enough to approximate any learning algorithm.
Our work is closely related to this line of research. Unlike previous work in this line of research, MT-nets
learn how many degrees of freedom the task-specific learner should have at meta-test time. Additionally, while
MT-nets learn update rules, these update rules are directly embedded in the network itself instead of being
stored in a separate model.
Distance metric learning [38, 37] methods learn a distance function between datapoints. Like these
methods, MT-nets learn a full metric matrix. Whereas those methods required constrained optimization
techniques to enforce that the learned matrix represents a metric, our parameterization allows us to directly
learn such a metric using gradient descent. Recently, neural networks have been used to learn a metric between
images[15, 36, 31], achieving state-of-the-art performance on few-shot classification benchmarks. Our work is
similar to these recent methods, but we learn a metric in feature space instead of input space. Our method
applies to a larger class of problems including regression and reinforcement learning, since all MT-nets require
is a differentiable loss function.
Another line of research in few-shot learning is to use a recurrent neural network (RNN) as a learner
[28, 24]. Here, the meta-learning algorithm is gradient descent on an RNN, and the learning algorithm is the
update of hidden cells. The (meta-learned) weights of the RNN specify a learning strategy, which processes
training data and uses the resulting hidden state vector to make decisions about test data. A recent work that
uses temporal convolutions for meta-learning[21] is also closely related to this line of research.
6
Experiments
We performed experiments to answer:
• How do our novel components (TW, M etc) affect meta-learning performance?
• To what degree does T alleviate the need for careful tuning of step size α?
• In MT-nets, does learned subspace dimension reflect the difficulty of tasks?
• Can T-nets and MT-nets scale to large-scale meta-learning problems?
Most of our experiments were performed by modifying the code accompanying [5], and we follow their
experimental protocol and hyperparameters unless specified otherwise.
6.1
Toy Regression Problem
We start with a K-shot regression problem and compare results to previous meta-learning methods [5, 20].
The details of our regression task are the same as [20]. Each individual task is to regress from the input x to
9
Models
5-shot
10-shot
20-shot
MAML
Meta-SGD1
1.07 ± 0.11
0.88 ± 0.14
0.71 ± 0.07
0.53 ± 0.09
0.50 ± 0.05
0.35 ± 0.06
M-net-full
M-net
T-net
MT-net-full
MT-net
0.91 ± 0.09
0.88 ± 0.09
0.83 ± 0.08
0.81 ± 0.08
0.76 ± 0.09
0.63 ± 0.07
0.60 ± 0.06
0.56 ± 0.06
0.51 ± 0.05
0.49 ± 0.05
0.38 ± 0.04
0.41 ± 0.04
0.38 ± 0.04
0.35 ± 0.04
0.33 ± 0.04
1
Table 1: Loss on sine wave regression. Networks were meta-trained using 10-shot regression tasks. Reported
losses were calculated after adaptation using various numbers of examples.
1
Reported by [20].
the output y of a sine function
y(x) = A sin(wx + b)
(18)
For each task, A, w, b are sampled uniformly from [0.1, 5.0], [0.8, 1.2], and [0, π], respectively. Each task
consists of K ∈ {5, 10, 20} training examples and 10 testing examples. We sample x uniformly from
[−5.0, 5.0] for both train and test sets. Our regressor architecture has two hidden cells each with activation
size 40. After every T is a ReLU nonlinearity. The loss function is the mean squared error (MSE) between
the regressor’s prediction f (x) and the true value y(x). We used Adam [14] as our meta-optimizer with a
learning rate of β = 10−3 . Task-specifc learners used step size α = 10−2 . We initialize all ζ to 0, all T as
identity matrices, and all W as truncated normal matrices with standard deviation 10−2 . While we trained our
meta-learner with K = 10 examples, we tested using various numbers of examples (K ∈ {5, 10, 20}).
We show results in Table 1. To see if each of our novel components increased meta-learning performance,
we also performed the same experiments with variations of MT-nets. An M-net uses a mask M like an MT-net,
but each cell consists of a single matrix W instead of TW. A model with ”-full” at the end of its name learns
a separate mask for each weight of W instead of sharing a mask among weights that contribute to the same
activation. For example, if W has size 5 × 10, the corresponding ζ in an MT-net would be of dimension 5, but
in MT-net-full, the dimension of ζ would be 50. Note that MT-nets outperform MAML, meta-SGD, and all
variations of MT-nets.
6.2
Robustness to learning rate change
The transformation T of our method can change the effective step size α. To get the effect ofqimplicitly
p α0 i
i
changing the stepsize from α0 to αnew , one may replace each Ti with αnew
T and each W0i with αnew
α0 W0 .
We discuss this in more detail in Appendix A.
We performed experiments to see how robust our method is to variations in α. We perform the same
sinusoid experiment as in section 6.1, but with various step sizes (α ∈ {10−4 , 10−3 , 10−2 , 10−1 , 1, 10}). We
evaluate on K = 10 training examples, and all other settings are identical to the experiments in section 6.1.
We show losses after adaptation of both MAML and MT-nets in Table 2. We can see that MT-nets are
more robust to change in step size. This indicates that as shown in section 4.2, the matrix T is capable of
warping the parameter space to recover from suboptimal step size α.
10
α
MAML
T-net
MT-net
10
1
0.1
0.01
0.001
0.0001
171.92 ± 25.04
5.81 ± 0.49
1.05 ± 0.11
0.71 ± 0.07
0.82 ± 0.08
2.54 ± 0.19
4.08 ± 0.30
4.15 ± 0.30
0.68 ± 0.06
0.56 ± 0.06
0.59 ± 0.06
0.62 ± 0.06
4.18 ± 0.30
0.61 ± 0.07
0.54 ± 0.05
0.49 ± 0.05
0.59 ± 0.06
0.72 ± 0.07
Table 2: Loss on 10-shot sine wave regression. T-nets and MT-nets are bost robust to change in step size α.
This is due to the meta-learned matrix T inside each cell, which alters the effective step size.
Figure 4: 10-shot regression tasks to sets of polynomials of various degrees. MT-nets choose to update a larger
fraction of weights as the set of tasks gets more complex.
6.3
Task Complexity and Subspace Dimension
We performed this experiment to see whether the dimension of the learned subspace of MT-nets reflect the
underlying complexity of its given set of tasks.
We consider 10-shot regression tasks in which the target function is a polynomial. Just as we constructed
a sine wave regression meta-task in section 6.1 by varying amplitude, phase, and frequency, we construct
a polynomial regression
by varying coefficients of polynomials of the same order. To generate a
Pmeta-task
n
polynomial of order n ( i=0 ci xi ), we uniformly sampled c0 , . . . , cn from [−1, 1]. We used the same network
architecture and hyperparameters as in Section 6.1, and performed 10-shot regression for polynomial orders
n ∈ {0, 1, 2}. Since the number of free parameters is proportional to the order of the polynomial, we expect
higher-order polynomials to require more parameters to adapt to. The fraction of parameters that task-specific
e−ζ
learners change is calculated as the expected value of e−ζ
over all logits ζ.
+1
We show results in Figure 4, and additional results in Appendix C. The number of weights that the
meta-learner of an MT-net sets to be altered increases as the task gets more complex. We interpret this as
the meta-learner of MT-nets having an effect akin to Occam’s razor: it selects a task-specific model of just
enough complexity to learn in a set of tasks. This behavior emerges even though we do not introduce any
additional loss terms to encourage such behavior. We think this is caused by the noise inherent in stochastic
gradient descent. Since the meta-learner of an MT-net can choose whether or not to perform gradient descent
in a particular direction, it is incentivized not to do so in directions that are not model-specific, because doing
so would introduce more noise into the network parameters and thus (in expectation) increase loss.
11
Models
Matching Networks[36]
Prototypical Networks[31]
mAP-SSVM[34]
MAML[5]
Meta-SGD[20]
T-net (ours)
MT-net (ours)
5-way 1-shot acc. (%)
98.1
97.4
98.6
98.7 ± 0.4
99.53 ± 0.26
99.4 ± 0.3
99.5 ± 0.3
Models
Matching Networks[36]1
Prototypical Networks[31]2
mAP-SSVM[34]
Fine-tune baseline1
Nearest Neighbor baseline1
meta-learner LSTM[25]
MAML[5]
L-MAML[8]
Meta-SGD[20]
T-net (ours)
MT-net (ours)
20-way 1-shot acc. (%)
93.8
92.0
95.4
95.8 ± 0.3
95.93 ± 0.38
96.1 ± 0.3
96.2 ± 0.4
5-way 1-shot acc. (%)
43.56 ± 0.84
46.61 ± 0.78
50.32 ± 0.80
28.86 ± 0.54
41.08 ± 0.70
43.44 ± 0.77
48.70 ± 1.84
49.40 ± 1.83
50.47 ± 1.87
50.86 ± 1.82
51.70 ± 1.84
Table 3: Few-shot classification accuracy on (top) held-out Omniglot characters and (bottom) test split of
MiniImagenet. ± represents 95% confidence intervals.
1
2
6.4
Reported by [25].
Reported results for 5-way 1-shot.
Classification
To compare the performance of MT-nets to prior work in meta-learning, we evaluate our method on few-shot
classification on the Omniglot [17] and MiniImagenet [25] datasets. We used the miniImagenet splits proposed
by [25] in our experiments.
Our CNN model uses the same architecture as [5]. The model has 4 modules: each has 3 × 3 convolutions
and 64 filters, followed by batch normalization [11]. As in [5], we used 32 filters per layer in miniImagenet.
Convolutions have stride 2 × 2 on Omniglot, and 2 × 2 max-pooling is used after batch normalization instead
of strided convolutions on MiniImagenet. We evaluate with 3, 5, and 10 gradient steps for Omniglot 5-way,
Omniglot 20-way, and miniImagenet 5-way, respectively.
Results are shown in Table 3. MT-nets achieve state-of-the-art or comparable performance on both problems. Several works [22, 24, 32] have reported improved performance on miniImagenet using a significantly
more expressive architecture. We only report methods that have equal or comparable expressiveness to the
model first described in [36]. Not controlling for network expressivity, the highest reported accuracy so far on
5-way 1-shot miniImagenet classification is 57.02 [32].
12
7
Conclusion
We introduced T-nets and MT-nets. One can transform any feedforward neural network into an MT-net, so
any future architectural advances can take advantage of our method. Experiments showed that our method
alleviates the need for careful tuning of the learning rate in few-shot learning problems and that the mask
M reflects the complexity of the set of tasks it is learning to adapt in. MT-nets also showed state-of-the-art
performance in a challenging few-shot classification benchmark (miniImagenet). An interesting line of future
research is to theoretically analyze the optimization and generalization properties of MT-net cells, which we
have not explored in this work.
One of the biggest weaknesses of deep networks is that they are very data intensive. By learning what to
learn when a new task is encountered, we can train networks with high capacity using a small amount of data.
We believe that designing effective gradient-based meta-learners will be beneficial not just for the few-shot
learning setting, but also machine learning problems in general.
13
References
[1] Amari, S. Natural gradient works efficiently in learning. Neural Computation, 10(2):251–276, 1998.
[2] Andrychowicz, Marcin, Denil, Misha, Gómez Colmenarejo, Sergio, Hoffman, Matthew W, Pfau, David,
Schaul, Tom, Shillingford, Brendan, De Freitas, Nando, and Deepmind, Google. Learning to learn by
gradient descent by gradient descent. Advances in Neural Information Processing Systems (NIPS), 2016.
[3] Edwards, Harrison and Storkey, Amos. Towards a Neural Statistician. Proceedings of the International
Conference on Learning Representations (ICLR), 2017.
[4] Finn, C. and Levine, S. Meta-learning and universality: Deep representations and gradient descent can
approximate any learning algorithm, 2017. Preprint arXiv:1710.11622.
[5] Finn, C., Abbeel., P., and Levine, S. Model-agnostic meta-learning for fast adaptation of deep networks.
In Proceedings of the International Conference on Machine Learning (ICML), Sydney, Australia, 2017.
[6] Fort, S. Gaussian Prototypical Networks for Few-Shot Learning on Omniglot, 2017. Preprint
arXiv:1708.02735.
[7] Garcia, V. and Bruna, J.
arXiv:1711.04043.
Few-Shot Learning with Graph Neural Networks, 2017.
Preprint
[8] Grant, Erin, Finn, Chelsea, Levine, Sergey, Darrell, Trevor, and Griffiths, Thomas. Recasting gradientbased meta-learning as hierarchical bayes. Proceedings of the International Conference on Learning
Representations (ICLR), 2018.
[9] Ha, David, Dai, Andrew, and Le, Quoc V. Hypernetworks. Proceedings of the International Conference
on Learning Representations (ICLR), 2016.
[10] Hariharan, Bharath and Girshick, Ross B. Low-shot visual object recognition by shrinking and hallucinating features. Proceedings of the International Conference on Computer Vision (ICCV), 2016.
[11] Ioffe, S. and Szegedy, C. Batch Normalization: Accelerating Deep Network Training by Reducing
Internal Covariate Shift. Proceedings of the International Conference on Machine Learning (ICML),
2015.
[12] Jang, Eric, Gu, Shixiang, and Poole, Ben. Categorical Reparameterization with Gumbel-Softmax.
Proceedings of the International Conference on Learning Representations (ICLR), 2017.
[13] Kaiser, Lukasz, Nachum, Ofir, Roy, Aurko, and Bengio, Samy. Learning to Remember Rare Events.
ICLR, 2017.
[14] Kingma, D. P. and Ba, J. L. ADAM: A method for stochastic optimization. In Proceedings of the
International Conference on Learning Representations (ICLR), San Diego, CA, USA, 2015.
[15] Koch, G., Zemel, R., and Salakhutdinov, R. Siamese neural networks for one-shot image recognition. In
Proceedings of the International Conference on Machine Learning (ICML), Lille, France, 2015.
[16] Krizhevsky, Alex, Sutskever, Ilya, and Hinton, Geoffrey E. ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems (NIPS), 2012.
[17] Lake, Brenden M, Salakhutdinov, Ruslan, and Tenenbaum, Joshua B. Human-level concept learning
through probabilistic program induction. Science, 2015.
14
[18] Li, D., Yang, Y., Song, Y.-Z., and Hospedales, T. M. Learning to Generalize: Meta-Learning for Domain
Generalization. Proceedings of the AAAI National Conference on Artificial Intelligence (AAAI), 2017.
[19] Li, Ke and Malik, Jitendra. Learning to Optimize. Advances in Neural Information Processing Systems
(NIPS), 2016.
[20] Li, Zhenguo, Zhou, Fengwei, Chen, Fei, and Li, Hang. Meta-SGD: Learning to Learn Quickly for Few
Shot Learning, 2017. Preprint arXiv:1707.09835.
[21] Mishra, N., Rohaninejad, M., Chen, X., and Abbeel, P. Meta-Learning with Temporal Convolutions,
2017. Preprint arXiv:1707.03141.
[22] Mishra, Nikhil, Rohaninejad, Mostafa, Chen, Xi, and Abbeel, Pieter. A simple neural attentive metalearner. Proceedings of the International Conference on Learning Representations (ICLR), 2018.
[23] Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Riedmiller,
M., Fidjeland, A. K., Ostrovski, G., Petersen, S., Beattie, C., Sadik, A., Antonoglou, I., King, H.,
Kumaran, D., Wierstra, D., Legg, S., and Hassabis, D. Human-level control through deep reinforcement
learning. Nature, 518, 2015.
[24] Munkhdalai, T. and Yu, H. Meta networks. In Proceedings of the International Conference on Machine
Learning (ICML), Sydney, Australia, 2017.
[25] Ravi, S. and Larochelle, H. Optimization as a model for few-shot learning. In Proceedings of the
International Conference on Learning Representations (ICLR), Toulon, France, 2017.
[26] Ren, Mengye, Ravi, Sachin, Triantafillou, Eleni, Snell, Jake, Swersky, Kevin, Tenenbaum, Josh B.,
Larochelle, Hugo, and Zemel, Richard S. Meta-learning for semi-supervised few-shot classification.
Proceedings of the International Conference on Learning Representations (ICLR), 2018.
[27] Robbins, H. and Monro, S. A stochastic approximation method. The Annals of Mathematical Statistics,
22(3):400–407, 1951.
[28] Santoro, A., Bartunov, S., Botvinick, M., Wierstra, D., and Lilicrap, T. Meta-learning with memoryaugmented neural networks. In Proceedings of the International Conference on Machine Learning
(ICML), New York, NY, USA, 2016.
[29] Schmidhuber, J. Evolutionary Principles in Self-Referential Learning. PhD thesis, Technical University
of Munich, 1987.
[30] Schmidhuber, J., Zhao, J., and Wiering, M. Shifting inductive bias with success-story algorithm, adaptive
levin search, and incremental self-improvement. Machine Learning, 28:105–130, 1997.
[31] Snell, J., Swersky, K., and Zemel, R. Prototypical networks for few-shot learning. In Advances in Neural
Information Processing Systems (NIPS), volume 30, 2017.
[32] Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P. H. S., and Hospedales, T. M. Learning to Compare:
Relation Network for Few-Shot Learning, 2017. Preprint arXiv:1711.06025.
[33] Thrun, S. and Pratt, L. Learning to Learn. Kluwer Academic Publishers, 1998.
[34] Triantafillou, Eleni, Zemel, Richard, and Urtasun, Raquel. Few-shot learning through an information
retrieval lens. Advances in Neural Information Processing Systems (NIPS), 2017.
15
[35] Tsai, Y.-H. H. and Salakhutdinov, R. Improving One-Shot Learning through Fusing Side Information,
2017. Preprint arXiv:1710.08347.
[36] Vinyals, O., Blundell, C., Lillicrap, T., Kavukcuoglu, K., and Wierstra, D. Matching networks for one
shot learning. In Advances in Neural Information Processing Systems (NIPS), volume 29, 2016.
[37] Weinberger, K. Q., Blitzer, J., and Saul, L. K. Distance metric learning for large margin nearest neighbor
classification. In Advances in Neural Information Processing Systems (NIPS), volume 18, 2006.
[38] Xing, E. P., Ng, A. Y., Jordan, M. I., and Russel, S. Distance metric learning, with application to clustering
with side-information. In Advances in Neural Information Processing Systems (NIPS), volume 15. MIT
Press, 2003.
16
Appendices
A
Learning rates as a special case of T
In this subsection, we restrict our attention to the special case of T being a constant multiple of the identity
matrix I; we show that in this case, T corresponds to a learned learning rate.
Consider gradient descent with learning rate α on the following linear equation:
ŷ = Ax
(19)
and set the values to x = x0 and A = A0 . Naturally, the resulting ŷ is ŷ0 = A0 x0 . The gradient update with
respect to the loss function L(ŷ) = 21 (ŷ − y)2 is:
Anew =
A0 − α∇A 12 (ŷ − y)2
= A0 − α∇ŷ 21 (ŷ − y)2
=
A0
x>
ŷ0 0
A0 − α(ŷ0 − y)x>
0
(20)
and we see that the new value of y for the same input x0 is
ŷnew =Anew x0 = (A0 − α(ŷ0 − y)x>
0 )x0
=y0 − α(ŷ0 − y)x>
0 x0
(21)
Alternatively consider gradient descent on A in the following equation with learning rate 1:
√
ŷ = α(Alr x)
(22)
A0
and set the value of Alr to Alr0 = √
. We can easily see that given the input x0 , this equation produces the
α
√
√
A
lr
0
same output αA0 x0 = α √α x0 = A0 x0 .
The update here is:
1
Alrnew =Alr0 − ∇Alr (ŷ − y)2 √
A0
2
α
A0
1
= √ − ∇Alr x (ŷ − y)2 √y0 x>
0
2
α
α
1
A0 √
= √ − α∇y (ŷ − y)2 y x>
0
0
2
α
A0 √
= √ − α(ŷ − y)x>
0
α
(23)
The new value of y is
√
ŷnew = αAlrnew x0
√ A0 √
= α( √ − α(ŷ − y)x>
0 )x0
α
=ŷ0 − α(ŷ0 − y)x>
0 x0
(24)
We see that the resulting output from gradient descent on 22 with learning rate 1 is the same as gradient
descent on 19 with learning rate α.
Therefore,
√ the learning rate hyperparameter in gradient descent can be learned using our architecture
√ by
setting T to αI and updating with the equation θ ← W − ∇W L whilst meta-learning the value of α.
17
B
B.1
Proofs for Section 4
MT-nets Learn a Subspace
Proposition 1. Fix x and A. Let U be a d-dimensional subspace of Rn (d ≤ n). There exist configurations
of T, W, and ζ such that the span of ynew − y is U while satisfying A = TW.
Proof. We show by construction that Proposition 1 is true.
Suppose that {v1 , v2 , . . . , vn } is a basis of Rn such that {v1 , v2 , . . . , vd } is a basis of U. Let T be
the n × n matrix [v1 , v2 , . . . , vn ]. T is invertible since it consists of linearly independent columns. Let
W = T−1 A and let ζ 1 , ζ 2 , . . . , ζ d → ∞ and ζ d+1 , . . . , ζ n → −∞. The resulting mask M that ζ generates
is a matrix with only ones in the first d rows and zeroes elsewhere.
ynew − y = T(Wnew − W)x
∇W LT )x
= T(M
(25)
Since all but the first d rows of M are 0, (M ∇W LT )x is an n-dimensional vector in which nonzero
elements can only appear in the first d dimensions. Therefore, the vector T(M ∇W LT )x is a linear
combination of {v1 , v2 , . . . , vd }. Thus the span of ynew − y is U.
B.2
MT-nets Learn a Metric in their Subspace
Proposition 2. Fix x, A, and a loss function LT . Let U be a d-dimensional subspace of Rn , and g(·, ·) a
metric tensor on U. There exist configurations of T, W, and ζ such that the vector ynew − y is in the steepest
direction of descent on LT with respect to the metric du.
Proof. We show Proposition 2 is true by construction as well.
We begin by constructing a representation for the arbitrary metric tensor g(·, ·). Let {v1 , v2 , . . . , vn }
be a basis of Rn such that {v1 , v2 , . . . , vd } is a basis of U. Vectors u1 , u2 ∈ U can be expressed as
Pd
Pd
u1 = i=0 c1i vi and u2 = i=0 c2i vi . We can express any metric tensor g(·, ·) using such coefficients c:
g11 . . . g1d
c21
.
.
.
..
.. ...
g(u1 , u2 ) = c11 . . . c1d ..
(26)
,
|
{z
}
gd1 . . . gdd
c2d
c>
1
{z
} | {z }
|
G
c2
where G is a positive definite matrix. Because of this, there exists an invertible d × d matrix H such that
G = H> H. Note that g(u1 , u2 ) = (Hc1 )> (Hc2 ): the metric g(·, ·) is equal to the inner product after
multiplying H to given vectors c.
Using H, we can alternatively parameterize vectors in U as
u1 = v1 . . . vd c1
(27)
|
{z
}
V
= VH
−1
(Hc1 ) .
(28)
Here, we are using Hc1 as a d-dimensional parameterization and the columns of the n × d matrix VH−1 as
an alternative basis of U.
Let v1H , . . . , vdH be the columns of VH−1 , and set T = [v1H , . . . , vdH , vd+1 , . . . , vn ]. Since H is
invertible, {v1H , . . . , vdH } is a basis of U and thus T is an invertible matrix. As in Proposition 1, set W =
18
T−1 A, ζ 1 , ζ 2 , . . . , ζ d → ∞, and ζ d+1 , . . . , ζ n → −∞. Note that this configuration of ζ generates a mask
M that projects gradients onto the first d rows, which will later be multiplied by the vectors {v1H , . . . , vdH }.
We can express y as y = V cy = VH−1 (Hcy ), where cy is again a d-dimensional vector. Note that
VH−1 is constant in the network and change in W only affects Hcy . Since ∇W LT = (∇Wx LT )x> , the
task-specific update is in the direction of steepest descent of LT in the space of Hcy (with the Euclidean
metric). This is exactly the direction of steepest descent of LT in U with respect to the metric g(·, ·).
C
Additional Experiments
19
Figure 5: Additional qualitative results from the polynomial regression task
20
| 1cs.CV
|
MODELING AND SIMULATING RETAIL MANAGEMENT PRACTICES:
A FIRST APPROACH
ABSTRACT
Multi-agent systems offer a new and exciting way of understanding the world of
work. We apply agent-based modeling and simulation to investigate a set of problems
in a retail context. Specifically, we are working to understand the relationship
between people management practices on the shop-floor and retail performance.
Despite the fact we are working within a relatively novel and complex domain, it is
clear that using an agent-based approach offers great potential for improving
organizational capabilities in the future.
Our multi-disciplinary research team has worked closely with one of the UK’s top ten
retailers to collect data and build an understanding of shop-floor operations and the
key actors in a department (customers, staff, and managers). Based on this case study
we have built and tested our first version of a retail branch agent-based simulation
model where we have focused on how we can simulate the effects of people
management practices on customer satisfaction and sales. In our experiments we have
looked at employee development and cashier empowerment as two examples of shopfloor management practices.
In this paper we describe the underlying conceptual ideas and the features of our
simulation model. We present a selection of experiments we have conducted in order
to validate our simulation model and to show its potential for answering “what-if”
questions in a retail context. We also introduce a novel performance measure which
we have created to quantify customers’ satisfaction with service, based on their
individual shopping experiences.
KEYWORDS
Agent-Based Modeling and Simulation, Retail, Management Practices, Shopping
Behavior
1
BACKGROUND
1.1
Why is Retail Productivity Important?
The retail sector significantly contributes to the UK’s relatively low productivity
compared to France, Germany and the USA (Reynolds et al., 2005), popularly termed
the ‘productivity gap’. A large-scale literature review of management practices and
retail performance and productivity (Siebers et al., 2008) concluded that management
practices are multidimensional constructs which tend to demonstrate a complex
relationship with productivity. The authors concluded that it may be the contextspecific nature of management practices and productivity which precludes clear
patterns in the results of empirical studies (for a further review see Wall & Wood
2005). Many experts agree that the focus is shifting to looking inside organizations to
1
understand the source of the problem (Delbridge et al., 2006). Focusing on
management practices may offer an opportunity to further our understanding of the
UK’s relatively low levels of retail productivity (Porter & Ketels, 2003).
1.2
Customer Satisfaction and In-Store Experiences
Without customers a retailer is without a business; Hill and Alexander (2006, p.11)
advocate the only route to success is, “Do best what matters to customers.” Measuring
customer satisfaction is the key way in which a retailer can quantify and understand
their strengths and weaknesses. Empirical evidence suggests there is a need to
differentiate between the components of a global customer satisfaction measure (Rust
et al., 1995; Garbarino & Johnson, 1999). Consequently, instead of focusing on
overall satisfaction as a global evaluation, we will investigate customer satisfaction as
it is empirically driven by visitors’ in-store experiences and perceptions of service
they receive (Torres et al., 2001).
1.3
People Management Practices
Managers working in retail stores tend to be under a lot of pressure to allocate their
time effectively and to prioritize the competing tasks and impromptu issues that can
crop up throughout their shift. People management practices offer a way of enhancing
the overall operation of the store through its staff. People (or Human Resource)
management practices have been defined as the, “… organizational activities directed
at managing the pool of human capital and ensuring that the capital is employed
towards the fulfillment of organizational goals,” (p.304, Wright et al., 1994).
Examples of people management practices are empowerment, team-based working
and skill development.
2
INTRODUCTION
There exists a large body of work investigating the modeling and simulation of
operational management practices, whereas people management practices have often
been neglected. Yet research suggests that people management practices crucially
impact upon an organization's performance (for example, Birdi et al., 2008).
The overall aim of our project is to understand and predict the impact of different
people management practices on retail productivity and performance. One key
objective has been to apply simulation to devise a functional representation of the
retail shop-floor driven by a real system. To achieve this objective we have adopted a
case study approach and integrated applied research methods to collect
complementary qualitative and quantitative data. In summary, we have conducted
four weeks of informal participant observation, forty staff interviews supplemented by
a short questionnaire regarding the effectiveness of various management practices,
and drawn upon a range of internal company documentation. Early experimentation
with the model has led us to develop and advance its operation in order to facilitate
more comprehensive investigation of the impact of management practices. By
reducing the level of abstraction within the model we are able to evaluate simulation
runtime outcomes in terms more closely linked to those of the real system.
2
In this paper we describe the development of the first version of our simulation model
where we have focused mainly on how we can simulate the effects of people
management practices on customer satisfaction and sales. We have chosen employee
development and empowerment as examples of such people management practices. In
Sections 3, 4, and 5 we provide an overview of the research we have completed
leading up to the creation of our simulation model of two retail departments. Section 3
embeds our selection of modeling technique in the broader modeling literature, and
Section 4 describes the model design including data collection, model
conceptualization, and a description of how we incorporate the empirical data we
have gathered during our case study work. Section 5 explains the implementation of
these concepts and the data according to the first full version of our simulation model
(referred to as ManPraSim v1). In Section 6 we present two sets of validation
experiments, and then three sets of operational experiments to investigate to impact of
management practices on department performance measures including customers’
satisfaction with the service provided. We draw some conclusions in Section 7 and
identify priorities for future work.
3
WHY USE AGENT-BASED SIMULATION?
There are a number of competing approaches to modeling, and the decision to choose
Agent-Based Modeling and Simulation (ABMS) followed a careful review and
evaluation of different approaches.
3.1
Selection of modeling technique
Operations Research (OR) is applied to problems concerning the conduct and coordination of the operations within an organization (Hillier & Lieberman, 2005). An
OR study usually involves the development of a scientific model which attempts to
abstract the essence of the real problem. When investigating the behavior of a
complex system it is very important to select an appropriate modeling technique. In
order to be able to make a choice for our project, we reviewed the relevant literature
spanning the fields of Economics, Organizational Behavior, Psychology, Retail,
Marketing, OR, Artificial Intelligence, and Computer Science. Within these fields a
wide variety of modeling approaches are used which can be classified into three main
categories: analytical approaches, heuristic approaches, and simulation. In many cases
we found modelers had adopted an integrated approach and applied more than one
technique within a single model. Common combinations were ‘simulation /
analytical’ for comparing efficiency of alternative future scenarios (e.g. Greasley,
2005), and ‘simulation / analytical’ or ‘simulation / heuristic’ where analytical or
heuristic models were used to represent the behavior of the entities within the
simulation model (e.g. Schwaiger & Stahmer, 2003).
Simulation opens the door to a new way of thinking about social and economic
processes, based on ideas about the emergence of complex behavior from relatively
simple activities (Simon, 1996). Whereas analytical models tend to aim to explain
correlations between variables measured at one single point in time, simulation
models are concerned with the development of a system over time. Furthermore,
analytical models usually work on a much higher level of abstraction than simulation
models. For simulation models it is critical to define the right level of abstraction.
3
Csik (2003) states that on the one hand the number of free parameters must be kept as
small as possible. On the other hand, too much abstraction and simplification will
threaten the ability of the model to accurately represent the real system. OR usually
employs three different types of simulation modeling to help understand the behavior
of organizational systems, each of which has its distinct application area: Discrete
Event Simulation (DES), System Dynamics Simulation (SDS) and Agent Based
Simulation (ABS). The choice of the most suitable approach will always depend on
the focus of the model, which input data is available, the level of analysis and what
kind of answers are sought.
In our review we put particular emphasis on those publications that try to model the
link between management practices and productivity or performance in the retail
sector. We found a very limited number of papers that investigate management
practices in retail at the organizational level, with the majority of these papers
focusing on marketing practices (e.g. Keh et al., 2006). Agent-Based Modeling
(ABM), using simulation as the method of execution, was by far the most popular
technique. It seems to be accepted as the natural way of system representation for
organizations; active entities in the live environment are interpreted as actors in the
model.
3.2
Agent-Based Modeling and Simulation
Although computer simulation has been used widely since the 1960s, ABM only
became popular at the start of the 1990s (Epstein & Axtell, 1996). ABM can be used
to study how micro-level processes affect macro-level outcomes. A complex system is
represented by a collection of individual agents which are programmed to follow
simple behavioral rules. Agents can interact with one another and with their
environment, and these interactions can result in complex collective behavioral
patterns. Macro behavior is not explicitly simulated; it emerges from the microdecisions and actions of individual agents (Pourdehnad et al., 2002). The main
characteristics of agents are: autonomous operation, the ability to act flexibly in
response to the environment, and pro-activeness driven by internal motivations.
Agents are designed to mimic the attributes and behaviors of their real-world
counterparts. Simulation output can be used for explanatory, exploratory and
predictive purposes (Twomey & Cadman, 2002).
The way in which agents are modeled appears to be more suitable than DES for
modeling human-centric complex adaptive systems (Siebers, 2006). There is a
structural correspondence between the real system and the model representation,
which makes these models more intuitive and easier to understand than for example a
system of differential equations as used in SDS. Hood (1998) emphasized one of the
key strengths of this technique is that the system as a whole is not constrained to
exhibit any particular behavior because the system properties emerge from its
constituent agent interactions. Consequently assumptions of linearity, equilibrium and
so on, are not needed. Of course there are disadvantages; there is a general consensus
in the literature that it is difficult to empirically evaluate agent-based models, in
particular at the macro level, because the behavior of the system emerges from the
interactions between the individual entities (Moss & Edmonds, 2005). Furthermore
Twomey & Cadman (2002) state that problems often occur through the lack of
adequate empirical data, and that there is always a danger that people new to ABM
4
may expect too much from the models, in particular with respect to predictive ability,
though this last criticism applies to all the simulation approaches mentioned above.
Overall we can conclude that ABMS is the most appropriate technique to investigate
people management practices. This approach provides us with the opportunity to
model organizational characters and their interactions in realistic and valid ways.
4
MODEL DESIGN
We emphasize the central role of data collection and understanding of the real system
to inform the conceptualization and implementation of our model.
4.1
Knowledge gathering
The case studies were conducted in the same two departments across two branches of
a leading UK department store. We adopted an integrated approach using a
complementary set of data collection methods: participant observation; semistructured interviews; completion of a management practices questionnaire with team
members, managers and personnel managers; and the analysis of company data and
reports. Research findings were consolidated and fed back (via report and
presentation) to employees and managers with extensive experience and knowledge of
the case study departments in order to cross-validate our understanding and
conclusions.
Preliminary case study findings suggested that we needed to configure the model to
represent the different department types: Audio and Television (A&TV) and
Womenswear (WW). This approach also helps to ensure that the simulation results
remain as broadly applicable as possible. Case study work revealed substantial
differences between the two department types, a divergence which is generally driven
by fundamentally different product characteristics. For example, the average purchase
in A&TV is more expensive than in WW. The likelihood of a customer seeking
advice is higher in A&TV, and the average customer service time is longer than in
WW. Customers in WW are more likely to make a purchase than in A&TV.
Our empirical approach to understanding the real case studies has played a crucial
role in facilitating the conceptualization of how the real system is structured. This is
an important stage to any simulation project, revealing insights into the operation of
the system as well as the behavior of and interactions between the different characters
in the system.
4.2
Conceptual Modeling
To make the most of the empirical data and insights obtained through the case studies,
the core aspects of the model were conceptualized and mapped out prior to
implementation.
4.2.1 Modeling Approach
5
Building on the findings from our literature review we have used the agent paradigm
to conceptualize and model the actors within the system under investigation.
Our modeling approach has been iterative, firstly creating a relatively simple model
and progressively building in more and more complexity. We started by trying to
understand the particular problem domain and to generate the underlying rules
currently in place. We have since progressed to the process of building an ABS model
of the real system using the information gathered during our case study and to validate
our model by simulating the operation of the real system.
This approach allows us to assess the accuracy of the system representation. When the
simulation has provided a sufficiently good representation we have been able to move
to the next stage, and generate new scenarios for how the system could work using
new rules.
4.2.2 Concept for the Simulation Model
Our conceptual ideas for the simulation model are shown in Figure 1. Within our
simulation model we have three different types of agents (customers, shop-floor staff,
and managers) each with a different set of relevant attributes. Global parameters can
influence any aspect of the system, and define, for example, the number of agents in
the system. With regards to system outputs we hope to find some unforeseeable,
emergent behavior on the macro level. Maintaining a visual representation of the
simulated system and its actors will allow us to closely monitor and better understand
the interactions of entities within the system. In addition, by measuring the
performance of the system we will be able identify bottlenecks in the real system and
to subsequently to optimize it.
[INSERT FIGURE 1]
4.2.3 Concept for the Actors
The agents have been designed and represented using state charts. State charts display
the different states which an entity can be in and define the transitional events which
are the triggers driving an actor’s change from one behavioral state to another. This is
exactly the information we need in order to represent our agents later within the
simulation environment. Furthermore this form of graphical representation is helpful
for validating the agent design (micro-level face validation) because it is easy for
experts in the real system to understand.
The art of modeling relies on simplification and abstraction (Shannon, 1975). A
model is always a restricted copy of the real world and we have to identify the most
important components of a system to build effective models. In our case, instead of
looking for components we have identified the most important behaviors of each actor
and the triggers which initiate a move from one state to another. We have developed
state charts for all the relevant actors in our retail department model. Figure 2 shows
as an example the conceptual template of a customer agent. The transition rules have
been omitted here to keep the chart succinct (see Section 5.1 and 5.2 for a more
detailed explanation of the transition rules).
6
[INSERT FIGURE 2]
Once a customer enters the department he or she will be in the contemplating state.
This is a dummy state and represents the reality of an individual thinking through
their behavioral intentions prior to acting (Ajzen, 1985), regardless of whether the
department visit will result in a planned or unanticipated purchase (Kelly et al., 2000).
Even when a particular purchase is planned, the customer may change their mind and
go for a substitute product, if they buy at all. He or she will probably start browsing
and after a certain amount of time, he or she may require help, queue at the till or
leave the shop. If the customer requires help, he or she considers what to do and seeks
help by looking for a staff member and will either immediately receive help or wait
for attention. If no staff member is available, he or she has to join a queue and wait for
help. If the queue moves very slowly it could result in a customer becoming fed up of
waiting or running out of time and so he or she leaves the queue prematurely. This
does not mean necessarily that he or she will not make a purchase. Sometimes
customers would still make a purchase even without getting the advice they were
seeking. Another reason why a customer might come into the department is to ask for
a refund. We have added this activity to the conceptual model because we will later
experiment with different refund policies. From an organizational point of view the
refund process is very similar to the help process. The difference is that the refund
process will take place at the till. After the refund process is concluded the customer
will either continue shopping (i.e. start browsing) or leave the department.
It is important to observe that there is a sequential order to these events which is
incorporated into the customer state chart. Furthermore, there is a logical flow to these
states. Thus, for example, a customer is unlikely to be queuing at the till in WW to
buy something without having first picked up an item. Therefore, the condition for
queuing at the till to buy something would be that the customer has been browsing
before to pick up an item. These rules have been considered in the implementation
(see Section 5.1 and Figure 3 for more details).
During the process of conceiving the model we have questioned whether or not our
agents are intelligent. Wooldridge (2002) stated that in order to be intelligent, agents
need to be reactive, proactive and social. This is a widely accepted view. Being
reactive means responding to changes in the environment (in a timely manner), being
proactive means persistently pursuing goals and being social means interacting with
other agents (Padgham & Winikoff, 2004). Our agents perceive a goal in that they
intend to either make a purchase or return a previous purchase. The buying process
has a sub goal; the customer is trying to buy the right thing. If the customer is not sure
he or she may ask for help from a shop floor worker. Our agents are not only reactive
but also flexible, i.e. they are capable to recover from a failure of action. They have
alternatives inbuilt when they are unable to realize their goal in a timely manner. For
example if a customer wants to pay but the queue is not moving he or she will always
have the chance to leave a queue and pursue another action. This example illustrates
that customers can respond in a flexible way to certain changes in their environment,
in this case the length of the queue. Finally, as there is communication between agents
and staff, they can also be regarded as social entities interacting with others1.
1
An extensive discussion of the notion of ‘intelligence’, a topic which seems to split the simulation
community (encompassing the ABM community) into two halves, can be found in SIMSOC (2008).
7
4.2.4 Concept for a Customer Satisfaction Measure
Customer perceptions are crucial to measuring the impact of retail management
practices. Applied in conjunction with objective performance measures (e.g. sales
turnover), it becomes possible to obtain a rounded view of retail performance.
Customer service is by definition intangible, and an index of customer satisfaction
offers an invaluable way of quantifying customers’ perceptions of this. Measures of
customer satisfaction are important to provide an indicator of not only how the
business is performing at present, but also an idea of how many recent customers will
return to the retailer.
Global customer satisfaction is a multi-dimensional construct, an accumulation of
separate satisfaction evaluations of multiple facets (see for example Parasuraman et
al., 1988; Bolton & Drew, 1991; Mihelis et al., 2001). We conceptualize a measure
which draws upon a subset of these facets; focusing on those aspects of satisfaction
which we observed to be most pertinent to customer satisfaction in a retail
department. We will use customers’ perceptions of their in-store experiences, in
particular the service that is provided, as an indicator of customer satisfaction. The
aim is to go beyond the capabilities of existing measures to create a dynamic measure
that considers each step of the entire shopping experience as each individual customer
perceives it. The link between shop-floor management practices and this measure is
salient because achieving a high level of customer satisfaction is hinges on the
availability of suitably skilled staff when customers need them. Relating these to
Mihelis and colleagues’ (2001) model of global satisfaction, for example, these
components relate to two of five high-level components: service (e.g. waiting times,
service processes) and personnel (e.g. skills and knowledge).
Previous work examining service encounters in retail settings has shown that the
attitudes and behaviors of employees can positively influence customers’ perceptions
of quality, satisfaction, and hence purchase intentions (Babin et al., 1999; Baker et al.,
2002; Dabholkar et al., 1995; Parasuraman et al., 1994). Further to this, key aspects of
customer service quality have been shown to impact positively on customer
perceptions, and these include circumstances when employees have been perceived as
respectful, friendly, knowledgeable about products, responsive to the customer’s
needs and questions, able to give advice, and have not pursued a ‘hard sell’ (Darian et
al., 2001; Leo & Philippe, 2002). Some businesses continue to gain competitive
advantage through priding themselves on exceptional customer service. The
importance of providing a high quality service to customers is widely accepted as a
crucial topic for management success, as demonstrated by dedicated journals, such as
‘Managing Service Quality’. A recent large-scale consumer satisfaction study
(conducted by Which, cited by Fluke, 2008) surveyed more than 10,000 people and
found that, “shoppers are increasingly willing to spend extra for better service.” This
is convincing evidence that retailers who strive for high levels of customer
satisfaction through a favorable in-store experience are reaping the benefits.
Many methods of calculating customer satisfaction sample only those people who
visit the store and leave with a purchase. These methods ignore the store visitors who
could have made a purchase; in other words data is not collected from unrealized
customers. We would argue that the satisfaction of all store visitors is important and
8
valid for the long-term development and survival of the business. Everyone who visits
the store will remember their experiences, whether they have been positive or
negative, and this could influence his or her decision to come again or whether to
make a future purchase (e.g. Meyer, 2008). For this reason our customer satisfaction
measure draws on the perceptions of all department visitors, rather than restricting
this measure only to those individuals who make a purchase.
5
MODEL IMPLEMENTATION
Our simulation has been implemented in AnyLogic™ version 5.5 which is a Java™
based multi-paradigm simulation software (XJ Technologies, 2007). The simulation
model is initialized from an Excel™ spreadsheet. We have implemented the
knowledge, experience and data obtained from the case studies, resulting in a model
which supports the simulation of the two types of departments (A&TV and WW)
within which we conducted our case study work.
5.1
Implementing the Concept
The simulation model can represent the following actors: customers, service staff
(including cashiers, selling staff of two different training levels) and managers. Figure
3 shows a screenshot of the customer and staff agent logic as it has been implemented
in AnyLogic™. Boxes represent customer states, arrows transitions, circles with B
branches (decision nodes) and numbers denote satisfaction weights which as a whole
form the service level index.
[INSERT FIGURE 3]
There are two different types of customer goals implemented: making a purchase or
obtaining a refund. If a refund is granted, the customer’s goal may then change to
making a new purchase, or alternatively they will leave the shop straight away. The
customer agent template consists of three main blocks which all use a very similar
logic. These blocks are ‘Help’, ‘Pay’ and ‘Refund’. In each block, in the first instance,
customers will try to obtain service directly and if they cannot obtain it (no suitable
staff member available) they will have to queue. They will then either be served as
soon as the right staff member becomes available or they will leave the queue if they
do not want to wait any longer (an autonomous decision). A complex queuing system
has been implemented to support different queuing rules. In comparison to the
customer agent template, the staff agent template is relatively simple. Whenever a
customer requests a service and the staff member is available and has the right level
of expertise for the task requested, the staff member commences this activity until the
customer releases the staff member. While the customer is the active component of
the simulation model the staff member is currently passive, simply reacting to
requests from the customer. In future we planned to add a more pro-active role for the
staff members, for example offering services to browsing customers.
5.2
Input Parameters
We have used frequency distributions and probabilities to assign different values to
each individual agent. In this way a population is created that reflects the variations in
9
attitudes and behaviors of their real human counterparts. Often agents are based on
analytical models or heuristics and in the absences of adequate empirical data
theoretical models are employed; we have incorporated data from the real system
wherever possible.
We have used frequency distributions for modeling delays between state changes,
specifically triangular distributions supplying the time that an event lasts, using the
minimum, mode, and maximum duration. Our triangular distributions are based on
our own observation and expert estimates in the absence of numerical data. We have
collected this information from the two branches and calculated an average value for
each department type, building one set of data for A&TV and one set for WW. Table
1 lists some sample frequency distributions that we have used for modeling the
A&TV department (the values presented here have been slightly amended to comply
with confidentiality restrictions). The distributions have been used as exit rules for
most of the states. All remaining exit rules are based on queue development, i.e. the
availability of staff.
[INSERT TABLE 1]
We have used probabilities to model the decision making processes. The probabilities
are partly based on company data or published data (e.g. conversion rates, that is the
percentage of customers who buy something) and where empirical data has not been
available we have collected estimates from knowledgeable individuals working in the
case study departments (e.g. the patience of a customer before prematurely leaving a
queue). Some examples for the probabilities we have used to model the A&TV
department can be found in Table 2, and as before we have calculated average values
for each department type. The probabilities link to most of the transition rules at the
branches where decisions are made about what action to take (e.g. decision to seek
help). The remaining decisions are based on the state of the environment (e.g. leaving
the queue, if the queue does not get shorter quickly enough).
[INSERT TABLE 2]
5.3
Performance Measures
We have built a number of performance measures into the system to help us
understand the outcomes of a simulation run. In this paper we look at the number of
transactions, staff utilization indices, the number of satisfied customers, and overall
customer satisfaction. These measures are defined as follows. The number of
transactions acts as a proxy for departmental sales turnover, and allows us to draw
links between experimental results and the tangible financial outcomes of the real
system. The staff utilization indices are presented by staff type, either normal or
expert, and help us to understand whether the staff team’s composition is effectively
meeting the demands placed on it by customers. Satisfaction measures have been
introduced to allow the satisfaction of customers with their in-store experiences, and
any service with which they have been provided, to be recorded throughout the
simulated lifetime. The number of satisfied customers is the count of customers who
have left the department with a positive service level index (i.e. the count of satisfied
customers). Overall customer satisfaction is the sum of all customers’ service level
10
indices when they leave the department (i.e. the sum of all customers’ individual
satisfaction levels).
To implement customers’ perceptions in the model, we have decomposed each
customer’s shopping experience across different behavioral states and transitions
between them (see Figure 3). The satisfaction weights define the relative contribution
of each customer transition to the satisfaction score; actual figures are notional. The
weightings have been allocated in line with empirical data, or where this has not been
possible they have been allocated in an intuitive manner. For example, a large
evidence base supports the contention that waiting for service can lead to a customer
forming a negative impression of the service being received (Bitner et al., 1990; Katz
et al., 1991; Taylor, 1994). Figure 3 displays how this finding has been incorporated
into the model. If a customer seeks help and locates help immediately his or her
satisfaction score is increased by 2 + 2 = 4. If the customer has to wait for help but
gets the help in the end, the overall impact on his or her satisfaction cancels out: 2 – 2
= 0. If, however, the customer gets fed up with waiting for help and leaves the queue
prematurely, there is a strong adverse impact on his or her satisfaction score: -4. If
this customer then leaves the department without buying anything an additional -2 is
added to the satisfaction score, so that this customer at the end of his or her shopping
trip would contribute zero to the number of satisfied customers count and -6 to the
overall customer satisfaction measure.
Implementing satisfaction weights allows us to account for the differential impact of
different components of customers’ in-store experiences, and build a more realistic
measure of customers’ satisfaction with their visit. In line with the empirical findings
of Westbrook (1981), a simple linear additive model has been followed, whereby a
customer’s individual satisfaction weights collected at each relevant transition can be
summed up to calculate an overall level of satisfaction at the end of a department
visit. We measure customers’ service satisfaction in two different ways derived from
these weightings; number of satisfied customers and overall customer satisfaction.
Applied in conjunction with an ABMS approach, we expect to observe interactions
with individual customer differences; variations which have been empirically linked
to differences in customers’ service satisfaction. This helps the analyst to find out to
what extent customers underwent a positive or negative shopping experience. It also
allows the analyst to put emphasis on different operational aspects and try out the
impact of different management strategies.
Individual differences between customers have already been built into the model and
there is some potential to extend the modeled variability between customers by
introducing heterogeneous customer types (as discussed in Section 7), and so the
satisfaction weights remain static (unless the weights themselves are the experimental
variable – see Section 6.3). The rationale for modeling the weights in this way is
because although it is likely that any single situation will inevitably be perceived in
different ways across individuals, it can also be argued that multiple responses will
tend to a normal distribution, resulting in a single ‘most likely’ or mean response. It is
the estimated ‘most likely’ response which has been implemented in the simulation
model. Using static satisfaction weights in this way ensures that we can incorporate a
dynamic measure of customers’ service satisfaction without introducing unnecessary
variability.
11
At the end of the simulation run, there are a certain number of customers who are
satisfied, those who are neutral in opinion, and those who are dissatisfied. A higher
count of the number of satisfied customers means that more customers are satisfied
when they leave the store. A higher (or lower) level of overall customer satisfaction is
likely to be the result of a combination of a higher number of customers holding that
satisfied (or dissatisfied) point of view and also possessing a more extreme opinion.
6
EXPERIMENTING WITH THE MODEL
In this section we present and describe the results from a series of experiments. Firstly
we validate the model by varying the operational staffing configuration and
examining the impact on sales figures and customers’ service satisfaction. Then we
test the impact of management practices (task empowerment, empowerment to learn,
and employee development) on key department performance measures including
customers’ service satisfaction.
Despite our prior knowledge of how the real system operates, we have been unable to
hypothesize precise differences in variable relationships, instead predicting general
patterns of relationships. Indeed, ABMS is a decision-support tool and is only able to
inform us about directional changes between variables (actual figures are notional).
In the broader simulation literature there is some divergence about whether or not it is
appropriate to apply rigorous statistical tests in the analysis of simulation results (e.g.
Schmeiser, 2001). Law and Kelton (2000) advocate the application of T-tests and not
ANOVAs; nevertheless ANOVAs and T-tests are a similar type of statistical test (they
are both parametric tests), and so both rely on the same key assumptions (e.g. see
Howell, 2007, Pallant, 2001). Before applying any parametric test it is essential that
appropriate preliminary tests check these assumptions, and where these are not met
then appropriate corrections are applied. This systematic approach ensures that
appropriate statistical tests are applied in the correct way.
We conduct independent replications with our simulation model, resulting in
independent observations. Specifically, in line with Law & Kelton (2000) each run
uses: separate sets of different random numbers (i.e. not common random numbers);
the same initial conditions; and resets the statistical counters. For these reasons we are
confident that we can make an assumption fundamental to the application of rigorous
statistical tests, specifically the assumption of independence of observations.
6.1
Model Validation
To test the operation of our simulation model and ascertain confidence in the validity
of our model we have designed and run two sets of validation experiments for both
departments. Firstly we will look at the impact of varying the department staffing
configuration on performance measures, and secondly the impact of satisfaction
weights on overall customer satisfaction. All experiments hold the overall number of
staffing resources constant at 10 staff and we run the simulation for a period of 10
weeks. We have conducted 20 repetitions for every experimental condition.
6.1.1 Staffing Configuration
12
During our time in the case study departments, we observed that the number of
cashiers available to serve customers would fluctuate over time. In the first
investigation we vary the staffing arrangement (i.e. the number of cashiers) and
examine the impact on the volume of sales transactions and two levels of customer
satisfaction described earlier; number of satisfied customers and overall customer
satisfaction. In reality, we observed that allocating extra cashiers would reduce the
shop floor sales team numbers, and therefore the total number of customer-facing
staff in each department is kept constant at 10.
6.1.1.1 Hypotheses
Our case study work has helped us to identify the distinguishing characteristics of the
departments, for example higher customer arrival rates in WW compared to A&TV,
and longer service times in A&TV compared to WW. We expect these inherent
differences to impact on department performance, and we therefore predict that for
each of our dependent measures: number of sales transactions (1), number of satisfied
customers (2) and overall customer satisfaction (3):
• H1a, H2a, H3a: An increase in the number of cashiers will be linked to
increases in 1, 2 and 3 respectively to a peak level, beyond which 1, 2 and 3
will decrease.
• H1b, H2b, H3b: The peak level of 1, 2 and 3 respectively will occur with a
smaller number of cashiers in A&TV than in WW.
6.1.1.2 Results
Preliminary analyses were conducted for each department. The distributions of all 3
dependent variables are approximately normal (Kolmogorov-Smirnov statistics all
p>.05). For 1 and 2, Levene’s test of equality of variances was violated (p<.05). It has
been credibly established that this is not a problem; ANOVAs are robust to violations
of this assumption provided that the size of the groups are reasonably similar
(Stevens, 1996), and in our case the group sizes are identical (n=20). Therefore it is
appropriate to analyze each dependent variable using a two-way between-groups
analysis of variance (ANOVA). Where significant ANOVA results were found, posthoc tests have been applied where appropriate to investigate further the precise impact
on outcome variables under different experimental conditions. To address the
increased risk of a Type I error associated with multiple tests we have applied a
Bonferroni correction to create more conservative thresholds for significance
(corrected post-hoc p-value for 3 dependent variables = .0167).
Each ANOVA revealed statistically significant differences (see Table 3 for
descriptive statistics). For the number of sales transactions (1) there were significant
main effects for both department [F(1, 190) = 356441.1, p<.001] and staffing [F(4,
190) = 124919.5, p<.001], plus a significant moderating effect of department type
[F(4, 190) = 20496.37, p<.001]. Tukey’s post hoc tests were run to explore the impact
of staffing and revealed significant differences for every paired comparison (p<.001).
[INSERT TABLE 3]
13
There is clear support for Hla. We expected this to happen because the number of
normal staff available to provide customer advice will eventually reduce to the extent
where there will be a detrimental impact on the number of customers making a
purchase. Some customers will become impatient waiting increasingly long for
service, and will leave the department without making a purchase. Hlb is not
supported, the data presents an interesting contrast, in that 1 plateaus in A&TV
around 3 and 4 cashiers, whereas WW benefits greatly from the introduction of a
fourth cashier. Nonetheless this finding supports the thinking underlying this
hypothesis, in that we expected the longer average service times in A&TV to put a
greater ‘squeeze’ on customer advice with even a relatively small increase in the
number of cashiers.
For the number of satisfied customers (2), there were significant main effects for both
department [F(1, 190) = 391333.7, p<.001], and staffing [F(4, 190) = 38633.83,
p<.001], plus a significant moderating effect of department type [F(4, 190) = 9840.07,
p<.001]. Post hoc tests explored the impact of staffing, and revealed significant
differences for every single comparison (p<.001).
The results support both H2a and H2b. We interpret these findings in terms of
A&TV’s greater service requirement, combined with the reduced availability of
advisory sales staff. These factors result in a peak in the number of satisfied
customers with a smaller number of cashiers (4) than in WW (5).
For overall customer satisfaction (3), there were significant main effects for both
department [F(1, 190) = 117214.4, p<.001], and staffing [F(4, 190) = 29205.09,
p<.001], plus a significant moderation effect of department type [F(4, 190) = 6715.93,
p<.001]. Tukey’s post hoc comparisons indicated significant differences between all
staffing levels (p<.001).
Our results support H3a for A&TV, showing a clear peak in overall customer
satisfaction. H3a is only partially supported for WW, in that no decline in 3 is evident
with up to 5 cashiers, although increasing this figure may well expose a peak because
the overall customer satisfaction is starting to plateau out. The results offer firm
support in favor of H3b.
6.1.2 Sensitivity of the Service Level Index
ManPraSim v1 incorporates a novel way of measuring customers’ service satisfaction.
It is a new feature of the model, and before we progress to investigate management
practices it is important to conduct a sensitivity study with the satisfaction weights
and assess the impact on the overall customer satisfaction measure.
For this series of experiments we will focus on the two main customer activities
involving the interaction between customers and staff: buying and asking for help. We
have switched off the refund loop because it would not add any relevant information
to these results. Therefore we have two main customer blocks (pay block and help
block) where we will systematically change the satisfaction weights settings to
observe what effect these changes have on overall customer satisfaction. We will use
the same staffing configuration for both departments; 3 cashiers, 6 normal selling
staff, and 1 expert.
14
Our original service level index configuration allocates a satisfaction weight at each
relevant customer transition (see Figure 3). Depending on the impact of a transition,
the linked satisfaction weights will be either 1, 2 or 4, and can be positive or negative
depending on whether or not the customer perceives the interaction positively or
negatively. We will investigate 3 different scenarios each with 3 levels of satisfaction
weights.
In scenario 1 each satisfaction weight has been set to the same value. Different
experimental conditions have been created by systematically increasing this single
value. This has changed the relative relationship between the weights from the
standard model implemented described above (see Section 5.3). For scenario 2 the
satisfaction weights have all been multiplied by the 3 different values for each level
(1, 10, and 100). This keeps the inter-relationships between service level index values
constant. In scenario 3 we have increased satisfaction weights within the same
experiment in order to investigate more extreme inter-relationships between
satisfaction weights. The first set are the standard satisfaction weights (see Section
5.3, 1-2-4), the second set are the first set squared (see Section 5.3, 1-4-16), and the
third set are the second set squared for a further time (1-16-256). Logically we expect
that increasing the satisfaction weights will result in higher overall customer
satisfaction. Comparing the two departments, we expect increments in satisfaction
weights to be positively linked to higher overall customer satisfaction in WW than in
A&TV, given the higher visitor arrival rates and higher conversion rates in WW. The
A&TV department has a higher proportion of customers requiring advice, and when
the staffing levels are held constant we expect this will mean a relatively small growth
in overall customer satisfaction compared to WW. For the third scenario, in A&TV
customer demand for A&TV department’s expert is likely to negatively impact on
overall customer satisfaction. This occurs because a customer who leaves
prematurely, whilst waiting for expert advice, results in the highest satisfaction
penalty.
6.1.2.1 Hypotheses
Therefore we predict the following hypotheses:
• H4. For scenario 1, we predict a uniformly positive and linear relationship
between satisfaction weights and overall customer satisfaction in both
departments.
• H5. For scenario 1, we predict that increasing the satisfaction weights will
have a greater positive impact in WW than A&TV.
• H6. For scenario 2, we predict a uniformly positive and linear relationship
between satisfaction weights and overall customer satisfaction in both
departments.
• H7. For scenario 2, we predict that increasing the satisfaction weights will
have a greater positive impact in WW than A&TV.
• H8. For scenario 3, we predict a uniformly positive and non-linear relationship
between satisfaction weights and overall customer satisfaction in both
departments.
• H9. For scenario 3, we predict that increasing the satisfaction weights will
have a greater positive impact in WW than A&TV.
15
6.1.2.2 Results
Preliminary tests confirmed that the distributions of overall customer satisfaction, for
both departments, are approximately normal (Kolmogorov-Smirnov statistics all p
>.05). For Scenarios 1, 2 and 3 Levene’s test of equality of error variances was
violated (p<.01). Therefore the significance value has been set to a stricter level
(p<.01). Where significant ANOVA results were found, post-hoc tests have been
applied where appropriate to investigate further the precise relationship between
satisfaction weights and overall customer satisfaction.
Each ANOVA revealed statistically significant differences (see Table 4 for
descriptive statistics). For scenario 1 there were significant main effects for both
department [F(1, 114) = 30,363.42, p<.01] and satisfaction weights [F(2, 114) =
2,943.58, p<.01], plus a significant interaction effect [F(2, 114) = 2,439.80, p<.01].
The effect sizes are very large (partial eta-squared = .996, .998, and .977
respectively). Post-hoc comparisons confirmed that every single paired comparison
exhibited a significant difference (p<.01).
[INSERT NEW TABLE 4]
Results for Scenario 1 offer support for both H4 and H5. The pattern of relationships
can be clearly seen in Figure 4; overall customer satisfaction rises with the
satisfaction weights in both departments, but to a greater extent in WW as confirmed
by the significant interaction effect.
[INSERT FIGURE 4]
Investigating scenario 2, an ANOVA identified statistically significant main effects
for department [F(1, 114) = 32,726.63, p<.01] and satisfaction weights [F(2, 114) =
152,387.00, p<.01], with a significant interaction effect [F(2, 114) = 23,929.13,
p<.01]. The effect sizes are substantial (partial eta-squared = .997, 1.00, and .998
respectively). Tukey’s post-hoc tests revealed significant differences (p<.01) for every
paired comparison.
Findings support H6; the satisfaction weight is significantly related to overall
customer satisfaction. Results also support H7 whereby department type is linked to
significantly higher overall satisfaction values in WW than in A&TV. The pattern of
relationships is graphically displayed in Figure 5.
[INSERT FIGURE 5]
Finally, for scenario 3, an ANOVA revealed statistically significant main effects for
department [F(1, 114) = 13,280.03, p<.01] and satisfaction weight [F(2, 114) =
2,771.02, p<.01], with a significant interaction effect [F(2, 114) = 8,544.38, p<.01].
The effect sizes are substantial (partial eta-squared = .991, .980, and .993
respectively). Post-hoc comparisons revealed significant differences (p<.01) between
every single paired comparison.
For scenario 3 the use of multiple satisfaction weights made it impossible to fully
account for the variability of values in a single experiment. For this reason a proxy
16
has been used; the middle value (2, 4, 16). There is partial support for H8; for WW we
can see a steady increase in the overall customer satisfaction with higher squared
input values. Surprisingly however we observe the inverse for A&TV; with higher
squared values the overall customer satisfaction falls considerably. It appears that the
greater service requirement linked to A&TV customers interacts with more extreme
overall customer satisfaction scores, resulting in very negative overall customer
satisfaction. Figure 6 graphically displays these relationships and suggests non-linear
associations (NB the A&TV non-linear relationship is in the opposite direction).
Results support H9; WW is linked to higher overall customer satisfaction than
A&TV. We expected this pattern because the greater service component of A&TV
roles means that when customers’ demands increase, any surplus staff capacity for
dealing with requests is filled and further customer requirements cannot always be
satisfactorily met. In turn this triggers an increase in customer waiting times and
customers are five times more likely to leave prematurely than in WW.
[INSERT FIGURE 6]
6.1.3 Validation Summary
Investigation of the impact of staffing configuration has provided support for most of
our hypotheses. As these general patterns in the data were as we would expect, this
builds our confidence in the accuracy and validity of the model. Comparing different
customers’ service satisfaction scenarios has demonstrated that changing the relative
differences between satisfaction weightings, and not the absolute differences, has a
greater impact on customers’ service satisfaction measures. These validation
experiments have informed the standard configuration of ManPraSim v1 which has
been used in the following experiments of management practices.
The overall validation process permits us to conclude that our hypotheses have been
largely supported. We are satisfied that we can have sufficient confidence in the
ability of our model to provide valid results to progress and investigate more complex
phenomena, specifically the impact of people management practices on performance
measures.
6.2
Management Practice Experiments
We have designed 3 experiments to investigate the impact of task empowerment,
empowerment to learn, and employee development. Global model settings for A&TV
(the department under investigation in these experiments) are held constant across
these experiments; the staff group in every experiment consists of 3 cashiers, 7 normal
selling staff and 2 experts, with a customer arrival rate of 70 per hour and a runtime of
10 weeks. We have systematically manipulated only the independent variable of
interest in each experiment. We have conducted at least 20 replications for every
experimental condition enabling thorough analysis of the results.
6.2.1 Task Empowerment
During our case study work, we observed the implementation of a new refund policy.
This new policy allows any cashier to decide independently whether to make a
customer refund up to the value of £50, rather than being required to refer the
17
authorization to an expert employee. To simulate the impact of this practice on key
business outcomes, we have systematically varied the probability that employees are
empowered to make refund decisions autonomously. Cashiers were configured to
process a refund in 80% of cases, whereas experts were more critical and only accept
70% of refund claims.
6.2.1.1 Hypotheses
As we increase the level of empowerment, we expect to see more transactions as work
flows more effectively and cashiers can take more decisions autonomously and
quickly without requiring expert assistance. We also anticipate greater levels of
overall customer satisfaction (whether obtaining a refund or not), because staff time is
less consumed by the delays of locating expert assistance, resulting in more employee
time available to customers. As the level of empowerment increases, we predict:
• H10. higher numbers of transactions.
• H11. higher overall customer satisfaction (shopping).
• H12. higher overall customer satisfaction (refund).
6.2.1.2 Results
Preliminary analyses tested the assumptions of rigorous statistical tests. The
distributions of all 3 dependent variables approximate to the normal distribution
(Kolmogorov-Smirnov statistics all p>.05). Levene’s test of homogeneity of variances
was violated by the number of transactions only (p=.02), although this does not need
to be corrected for, given the equal group sizes (n=20, Stevens, 1996). Consequently
we can proceed to apply a series of one-way between groups ANOVAs. Where
significant ANOVA results were found, post-hoc tests have been applied to
investigate further the precise impact on outcome variables under different
experimental conditions. To address the increased risk of a Type I error we have
applied a Bonferroni correction (corrected post-hoc p-value for 3 dependent variables
= .0167).
ANOVAs revealed statistically significant differences across all three outcomes:
number of transactions [F(4, 95)=26.77, p<.01], overall customer satisfaction
(shopping) [F(4, 95)=12.35, p<.01], and overall customer satisfaction (refund) [F(4,
95)=2001.73, p<.01]. Consulting Table 5, we see that H10 has not been supported,
and the number of transactions actually decreases with empowerment, whereas H11
and H12 are confirmed. The effect size, calculated using eta squared, reveals
differences in the relative impact of empowerment on each outcome measure: 0.53 for
the number of transactions, 0.34 for overall customer satisfaction (shopping) and 1.00
for overall customer satisfaction (refund). Social scientists report 0.14 as indicative of
a large effect (Cohen, 1988) suggesting we are looking at substantial effect sizes.
[INSERT TABLE 5]
Post-hoc comparisons using Tukey’s test indicated a number of significant differences
between group means. Most notably the impact on overall customer satisfaction
(refund) was great, with every single increment in empowerment resulting in a
significant increase in overall customer satisfaction (refund). H10 has not been
supported. This unforeseen reduction in transactions has occurred because less
18
experienced employees take longer to make a decision on a refund, resulting in a
knock-on impact for customer waiting times. H11 holds, and this finding is intuitive
because customers prefer that one staff member can deal with their needs. H12 is
strongly supported, and makes sense because cashiers are also more likely to approve
a customer refund request.
6.2.2 Empowerment to Learn
Our case study work has revealed that a key way in which employees can develop
their product knowledge occurs when they are unable to fully meet a customer’s
request for advice. The employee calls over an expert colleague and the original
employee is empowered to choose whether or not to stay with them to learn from the
interaction. In this second set of experiments we are assuming that, given the
opportunity to choose to learn, an employee will usually decide to take up that
opportunity. We found that case study employees enjoyed providing excellent
customer service, and given the opportunity would do what they could to stay abreast
of product developments.
In our model, a normal staff member gains knowledge points on every occasion that
he or she stays with an expert to learn from a customer interaction. We have
systematically varied the probability that a normal staff member learns in this way.
Naturally we expect there to be a trade-off with short-term ability to meet customer
demand, and a customer may leave prematurely if they have to wait for too long. A
normal staff member will be occupied for longer when his or her will to learn is
stronger.
6.2.2.1 Hypotheses
By allowing employees to acquire new product knowledge from expert colleagues, we
anticipate performance improvements. We predict that increasingly empowering
employees to learn will result in:
• H13. an increase in the knowledge of normal staff.
• H14. an increase in the utilization of normal staff.
• H15. no change to the utilization of expert staff.
• H16. a short term reduction in the number of sales transactions.
• H17. a reduction in overall customer satisfaction.
6.2.2.2 Results
Preliminary tests confirm that the distribution of all2 5 dependent variables are
approximately normal (Kolmogorov-Smirnov statistics all p>.05). Levene’s test of
homogeneity of variances was violated by normal expertise3 only (p<.01); given the
near-equal group sizes (n=20 or 21) we can safely continue (Stevens, 1996) and it is
to appropriate to use a series of one-way between groups ANOVAs. Post-hoc tests
2
Normal staff member expertise was not tested in the zero empowerment condition because it does not
vary.
3
This largely related to the fact that normal expertise is constant (=0) in the condition where the
probability of an employee learning from a customer interaction equals zero.
19
were run to follow-up significant ANOVA results, and again a Bonferroni correction
was applied (corrected post-hoc p-value for 5 dependent variables = .01).
The ANOVAs (see Table 6 for descriptives) exposed a significant impact of
empowerment to learn on: normal staff expertise [F(4, 96)=2,794.12, p<.01],
utilization of normal staff [F(4, 96)=112.53, p<.01], and overall customer satisfaction
[F(4, 96)=29.16, p<.01]. Tests of expert staff utilization [F(4, 96)=1.28, p=.29] and
sales transactions [F(4, 96)=1.25, p=.30] were insignificant. Effect sizes of significant
relationships were all large (normal staff expertise = 0.99, normal staff utilization =
0.83, overall customer satisfaction = 0.55).
[INSERT TABLE 6]
Tukey’s post-hoc comparisons were run for the three significant findings. Both
normal staff expertise and utilization significantly increased with every single
increment in employee empowerment to learn. The largest significant differences in
overall customer satisfaction are observed at the polar ends of the scale. As predicted,
employees who are empowered to learn become more knowledgeable (H13), leading
to a more efficient utilization of employees as a whole (H14). H15 holds as expected,
meaning there is no significant impact on the utilization of expert staff in terms of the
time they spend engaged with customers. However we can see through the effects on
other outcome measures that higher levels of learning empowerment result in better
‘utilization’ of experts; in other words more efficient harnessing of their knowledge.
H16 has not been supported as the number of transactions does not significantly differ
between experimental conditions. The short-lived reduction that we had anticipated is
so negligible, it is inconsequential. Nonetheless, the associated increase in customer
waiting times has negatively influenced the customer service index, in support of
H17.
6.2.3 Employee development
Our final investigation of management practices goes one step further and explores
how time invested in learning impacts on medium-term system performance. Our
model mimics an evolutionary process whereby staff members can progressively
develop their product knowledge over a period of time. When a staff member has
accumulated a certain number of knowledge points from observing expert service
transactions, they are considered an expert. We have systematically varied the
number of knowledge points required to attain expert-level competence. All normal
staff members are programmed to take advantage of all learning opportunities.
6.2.3.1 Hypotheses
By investing time in developing and expanding employees’ specialist knowledge, we
anticipate even greater future savings in terms of key outcomes, beyond those already
observed in the previous experiment. The academic literature echoes the positive
business impact of employing individuals with greater expertise to provide better
customer service and advice (e.g. Crosby et al., 1990). We predict that increasing the
rate of employee development (by lowering the threshold for attaining expert status)
will result in more desirable outcome variables, specifically increases in:
• H18. normal staff member expertise.
20
•
•
•
•
H19. normal staff utilization.
H20. expert staff utilization.
H21. the number of transactions.
H22. overall customer satisfaction.
6.2.3.2 Results
Preliminary tests confirmed the distribution of all4 5 dependent variables are
approximately normal, with the exception of expert staff utilization for a promotion
criteria level of 0.8 (Kolmogorov-Smirnov statistic p<.01). This non-normal
distribution will have minimal impact on the significance or power of the test
(Stevens, 1996, p.240). Levene’s test of homogeneity of variances was violated by all
5 dependent variables5 (p<.05); again we rely on Steven’s (1996) empirical argument
that because we have equal group sizes (n=20) this is acceptable. We have applied a
series of one-way between groups ANOVAs with Tukey’s post-hoc tests to determine
the specific nature of significant ANOVA results, and again a Bonferroni correction
was applied (corrected post-hoc p-value for 5 dependent variables = .01).
The final ANOVAs revealed statistically significant differences in expert utilization
[F(5,114)=952.21, p<.01], volume of transactions [F(5,114)=193.14, p<.01] and
overall customer satisfaction [F(5,114)=959.01, p<.01]. The effect sizes of significant
relationships were again all very large (expert staff utilization = 0.98, volume of
transactions = 0.89, and overall customer satisfaction = 0.98). We were unable to
adequately test the impact of learning on normal staff expertise (H18) and utilization
(H19), because we do not have this data for all experimental conditions (see Table 7:
at the lower promotion thresholds, all normal staff have been promoted before the end
of the simulation run and so these values are recorded as zero).
[INSERT TABLE 7]
Tukey’s post-hoc comparisons revealed significant differences between every single
increment in the competence threshold for each variable, with the exception of the
two upper levels. However, only expert utilization was in the predicted direction. Our
evidence was strongly in favor of H20, whereas the exactly the contrary of H21 and
H22 have been supported. This is counter-intuitive because we had expected that the
greater the number of resulting experts, the greater the availability of top-quality
advice to customers. Indeed, it is possible that our simulation run is too short at just
ten weeks, and presents only a backward facing view of department performance; i.e.,
focusing on the time consumed in learning, and not on the time spent sharing their
new competence with customers. If we ran the simulation for longer we would still
expect our original hypotheses to hold true. In our model we are also assuming that
staff acquire expertise purely by learning from their colleagues, whereas in reality this
would be supported with other sources and forms of learning.
4
The distributions of normal staff expertise and normal staff utilization have not been tested for
promotion criteria levels 0, 0.2, 0.4, and 0.6 because these variables are constant (=0) under these
conditions.
5
The distributions of normal staff expertise and normal staff utilization have not been tested for
promotion criteria levels 0, 0.2, 0.4, and 0.6 because these variables are constant (=0) under these
conditions.
21
6.2.4 Summary and Implications of Management Practice Experiments
All in all we can conclude that through modeling and simulating the impact of
management practices we are able to evaluate hypotheses regarding department
performance. Although some empowerment effects may at first appear small or even
the inverse of what we had expected, over time these changes can have a meaningful
impact. For this reason it is important to incorporate management practices in a retail
department model. It is also worth noting that the full benefits of empowering staff
may also be in terms which our simulator does not currently measure, such as an
employee’s satisfaction with their job or their intention to stay with the business. It is
therefore likely that our performance figures are an under-estimate of the true impact
of the various ways of empowering shop-floor staff that we have looked at.
From our experimental investigations, it is apparent that some management practices
offer greater potential for performance improvements than others. Our results lead us
to make a number of suggestions for the A&TV department to which they pertain.
The first experiment investigated the impact of empowering staff to make refunds and
demonstrated improvements in overall customer satisfaction (shopping and refund),
but the number of transactions did not increase as we had expected. Employees who
are less experienced at processing refunds tend to take longer than someone who has
more experience. The implication for retailers is to ensure that each employee
receives sufficient training, and has the opportunity to fully familiarize himself or
herself with the refund process. Carefully implementing such a scheme would
minimize the amount of time these employees need to get fully ‘up to speed’ with
their new responsibilities on the shop-floor.
The second experiment examined the influence of employees being empowered to
learn new product knowledge from expert colleagues advising customers on the shopfloor. Our results suggest that there are performance improvements to be gained
through encouraging expert employees to share their knowledge with lessexperienced colleagues ‘on-the-job’. Retailers need to be in it for the long-term, in
that there is a short-term trade off with customers’ service satisfaction whilst less
experienced staff develop new capabilities. In reality an employer would benefit from
aligning this kind of scheme with a staff retention initiative to ensure that workers are
committed to staying with the organization and contributing to its long-term success.
The third and final experiment looked at the impact of employee development and
how time invested in learning impacts on medium-term system performance. Our
findings indicated that the time it takes to see performance benefits is longer than 10
weeks. We can, nevertheless, advise that to establish the optimal level of employee
training in a given context, the particulars of that situation would first need to be more
closely examined. For example, the cost of employee wages, availability of skilled
labor, and customers’ expectations of shop-floor staff can all vary to some extent by
geography and the specific nature of the work. Consequently this example does not
allow us to conclude more than illustrating the point that employee development takes
time to positively impact on performance.
7
CONCLUSIONS AND FUTURE DIRECTIONS
22
In this paper we have described our first approach to modeling people management
practices, in particular employee development and empowerment. We have also
described the development of a novel performance measure to gauge customers’
service satisfaction. In this version of the simulation model we use customer agents
that are fairly homogeneous and have no enduring memory (they only visit the shop
once in their lifetime). Still, with this simulation model we have been able to
demonstrate how one could implement different empowerment management practices
in a simulation model to investigate their effect on customers’ service satisfaction, the
development of staff knowledge and sales turnover.
As far as we are aware, this is the first time that researchers have tried to use an agentbased approach to simulate management practices such as employee development and
empowerment. Although our simulation model has been driven by case studies from
one retail organization, we believe that the general model could be adapted to other
businesses operating in different industries and countries, for businesses which use
management practices involving a high degree of human interaction.
From what we can conclude from our current analyses, some findings are as we
hypothesized whereas others are more mixed. Two sets of validation experiments
provide us with confidence in the model’s ability to produce reliable and valid results.
Early findings indicate that management practices tend to exert a subtle yet significant
effect on performance, consistent with our case study findings. Further
experimentation is required to explore the model’s operation, and more development
work would bring the model closer to the real system and allow us to incorporate
more complex and interesting features.
We have identified two major limitations that we want to address in our future work.
Firstly, the current version of the simulation model has a high level of abstraction
which represents the actors of the real system and their behaviors in a simplistic way,
cutting out some relevant features and behaviors. For example, the homogeneity of
our customers prevents us from representing the true variability of customer
perceptions of, and responses to, service. Currently the model represents customer
perceptions and responses to the same situation in the same way. Secondly, in the
current version we cannot measure any long-term performance effects because we
have an infinite number of homogenous customers who only visit the department
once. It is important to overcome this limitation because many organizations strive to
retain their customers over time.
In the second version of the simulation model we want to address the shortcomings
we have identified above. Our main focus will be to make the simulation a better
representation of the real system (i.e. the retail department) and the people in the real
system, including their characteristics and behavior. We will incorporate realistic
footfall data, heterogeneous customer types, a finite population in the form of a
customer pool, a prompt departure of customers at closing time, and an enhanced
version of our novel customer service satisfaction measure.
Adding customer types and thereby diversifying the population will allow us to
represent varied perceptions of and reactions to the shopping experience. Our case
study organization has identified the particular customer types which are important to
its business through market research, and we plan to find out how populations of
23
certain customer types influence sales. Through introducing a finite population each
customer agent will have a memory of his or her previous shopping experiences
which will impact on how the agent perceives the customer service on his or her next
visit. As we observed in Section 4.2.4 a bad experience might make him or her come
less frequently or might reduce subsequent patience level thereby reducing the
probability of that customer buying something.
Taking a step back, we believe that researchers need to become more involved in this
multi-disciplinary kind of work to gain new insights into the behavior of
organizations. In our view, the main benefit from adopting this approach is the
improved understanding of and debate about a problem domain, and the resulting
explicit convergence of understanding and agreement about a system’s functioning.
The very nature of the methods involved forces researchers to be explicit about the
rules underlying behavior and to think in new ways about them. As a result, we have
brought work psychology and ABM closer together to develop a new and exciting
research area.
24
REFERENCES
Ajzen, I. (1985). From intentions to actions: a theory of planned behavior. In: Kuhl, J.
& Beckmann, J. (eds). Action-control: from cognition to behavior. Springer:
Heidelberg, 11-39.
Babin, L.A., Babin, B.J., & Boles, J.S. (1999). The effects of consumer perceptions of
the sales person, product and dealer on purchase intentions. Journal of Retailing and
Consumer Services, 6 (2), 91-97.
Baker, J., Parasuraman, A., Grewal, D., & Voss, G.B. (2002). The influence of
multiple store environment cues on perceived merchandise value and patronage
intentions. Journal of Marketing, 66 (2), 120-142.
Birdi, K., Clegg, C.W., Patterson, M., Robinson, A., Stride, C.B., Wall, T.D., &
Wood, S.J. (2008). The impact of human resource and operational management
practices on company productivity: a longitudinal study. Personnel Psychology,
accepted for publication.
Bitner, M.J., Booms, B.H., & Tetreault, M.S. (1990). The service encounter:
diagnosing favorable and unfavorable incidents. Journal of Marketing, 54 (1), 71-84.
Bolton, R.N. and Drew, J.H. (1991). A multistage model of customers’ assessments of
service quality and value. Journal of Consumer Research, 17 (4), 375-384.
Cohen, J. (1988). Statistical power analysis for the behavioral sciences. Erlbaum:
Hillsdale, NJ.
Crosby, L.A., Evans, K.A., & Cowles, D. (1990). Relationship quality in services
selling: an interpersonal influence perspective. Journal of Marketing, 53 (3), 68-81.
Csik, B. (2003). Simulation of competitive market situations using intelligent agents.
Periodica Polytechnica Ser. Soc. Man. Sci., 11 (1), 83-93.
Dabholkar, P.A., Thorpe, D.I., & Rentz, J.O. (1995). A measure of service quality for
retail stores: scale development and validation. Journal of the Academy of Marketing
Science, 24 (1), 3-16.
Darian, J.C., Tucci, L.A., & Wiman, A.R. (2001). Perceived salesperson attributes
and retail patronage intentions. International Journal of Retail and Distribution
Management, 29 (5), 205-213.
Delbridge, R., Edwards, P., Forth, J., Miskell, P. & Payne, J. (2006). The organisation
of productivity: rethinking skills and work organisation. Advanced Institute of
Management Research. Available at:
http://www.aimresearch.org/publications/orgprod.pdf [Accessed 14/03/2007].
Epstein, J.M., & Axtell, R. (1996). Growing artificial societies: social science from
the bottom up. MIT Press: Cambridge, MA.
25
Fluke, K. (2008). The best and the worst shops on the high street. MSN Money,
Consumer
Centre.
Available
from:
http://money.uk.msn.com/consumer/article.aspx?cp-documentid=7349625 [Accessed
24/01/2008].
Garbarino, E. & Johnson, M.S. (1999). The different roles of satisfaction, trust, and
commitment in customer relationships. Journal of Marketing, 63 (2), 70-87.
Greasley, A. (2005). Using DEA and simulation in guiding operating units to
improved performance. Journal of the Operational Research Society, 56 (6), 727-731.
Hill, N. & Alexander, J. (2006). The handbook of customer satisfaction and loyalty
measurement. Gower Press: Aldershot, UK.
Hillier, F.S. & Lieberman, G.J. (2005). Introduction to Operations Research (8th
Ed.). McGraw-Hill Higher Education: Boston, MA.
Hood, L. (1998). Agent-based modeling. In: Conference Proceedings: Greenhouse
Beyond Kyoto, Issues, Opportunities and Challenges. 31 March - 1 April 1998,
Canberra,
Australia.
Available
from:
http://affashop.gov.au/product.asp?prodid=13231
Howell, D.C. (2007). Fundamental statistics for the behavioral sciences (6th ed.).
Thomson/Wadsworth: Belmont, CA.
Katz, K.L., Larson, B.M., & Larson, R.C. (1991). Prescription for the waiting-in-line
blues: entertain, enlighten, and engage. Sloan Management Review, 32 (Winter), 4453.
Keh, H.T., Chu, S. & Xu, J. (2006). Efficiency, effectiveness and productivity of
marketing in services. European Journal of Operational Research, 170 (1), 265-276.
Kelly, J.P., Smith, S.M., & Hunt, H.K. (2000). Fulfillment of planned and unplanned
purchases of sale- and regular-price items: a benchmark study. International Review
of Retail, Distribution and Consumer Research, 10 (3), 247-263.
Law, A.M. & Kelton, W.D. (2000). Simulation modeling and analysis (3rd ed.). New
York: McGraw-Hill.
Leo, P.Y. & Philippe, J. (2002). Retail centers: location and customer satisfaction,
Service Industries Journal, 22 (1), 122-146.
Meyer, T. (2008). Experience-based aspects of shopping attitudes: The roles of norms
and loyalty. Journal of Retailing & Consumer Services, 15 (4), 324-333.
Mihelis, G., Grigoroudis, E., Siskos, Y., Politis, Y., & Malandrakis, Y. (2001).
Customer satisfaction measurement in the private bank sector. Journal of Operational
Research, 130, 347-360.
26
Moss, S., & Edmonds B. (2005). Sociology and simulation: statistical and qualitative
cross-validation. American Journal of Sociology, 110(4), 1095-1131.
Padgham, L., & Winikoff, M. (2004). Developing intelligent agent systems - a
practical guide. Wiley: New York, NY.
Pallant, J. (2001). SPSS survival manual. Open University Press: Buckingham, UK.
Parasuraman, A., Zeithaml, V.A. & Berry, L.L. (1988). SERVQUAL: a multiple-item
scale for measuring consumer perceptions of service quality. Journal of Retailing, 64
(1), 12-40.
Parasuraman, A., Zeithaml, V.A., & Berry, L.L. (1994). Alternative scales for
measuring service quality: a comparative assessment based on psychometric and
diagnostic criteria, Journal of Retailing, 70 (3), 201-230.
Porter, M.E. & Ketels, C.H.M. (2003). UK competitiveness: moving to the next stage.
DTI Economics Paper No.3 (May).
Pourdehnad, J., Maani, K., & Sedehi., H. (2002). System dynamics and intelligent
agent-based simulation: where is the synergy? In: Proceedings of the 20th
International Conference of the System Dynamics Society. 28 July – 1 August 2002,
Palermo,
Italy.
Available
from:
http://www.systemdynamics.org/conferences/2002/proceed/PROCEED.pdf [Accessed
06/02/08].
Reynolds, J., Howard, E., Dragun, D., Rosewell, B., & Ormerod, P. (2005). Assessing
the productivity of the UK retail sector. International Review of Retail, Distribution
and Consumer Research, 15 (3), 237-280.
Rust, R. T., Zahorik, A. J., & Keiningham, T. L. (1995). Return On Quality (ROQ):
making service quality financially accountable. Journal of Marketing, 59 (April), 5870.
Schmeiser, B.W. (2001). Some myths and common errors in simulation experiments.
In: Peters B.A., Smith J.S., Medeiros D.J. and Rohrer M.W. (eds). Proceedings of the
2001 Winter Simulation Conference. IEEE: Piscataway,NJ, 39–46.
Schwaiger, A. & Stahmer, B. (2003). SimMarket: multi-agent based customer
simulation and decision support for category management. In: Schillo, M., Klusch,
M., Muller, J. and Tianfield, H. (eds). Lecture Notes in Artificial Intelligence (LNAI)
2831. Springer: Berlin, 74-84.
Shannon, R. E. (1975). Systems simulation: the art and science. Prentice-Hall:
Englewood Cliffs, NJ.
Siebers, P.-O. (2006). Worker performance modeling in manufacturing systems
simulation: proposal for an agent-based approach. In: Rennard J.P. (ed). Handbook of
research on nature inspired computing for economics and management. Hershey, PA:
Idea Group Publishing.
27
Siebers, P.-O., Aickelin, U., Battisti, G., Celia, H., Clegg, C. W., Fu, X., De Hoyos,
R., Iona, A., Petrescu, A., & Peixoto, A. (2008). The role of management practices in
closing the productivity gap. AIM Working Paper. Available from:
http://www.aimresearch.org/workingpaperspage.html [Accessed 06/02/08].
Simon, H. A. (1996). The sciences of the artificial (3rd ed.). MIT Press: Cambridge,
MA.
Stevens, J. (1996). Applied multivariate statistics for the social sciences (3rd ed.).
Lawrence Erlbaum: Mahway, NJ.
Taylor, S. (1994). Waiting for service: the relationship between delays and
evaluations of service, Journal of Marketing, 58 (2), 56-69.
Torres, I. M., Summers, T. A., & Belleau, B. D. (2001). Men’s shopping satisfaction
and store preferences. Journal of Retailing and Consumer Services, 8 (4), 205-212.
Twomey, P., & Cadman, R. (2002). Agent-based modeling of customer behavior in
the telecoms and media markets. Info - The Journal of Policy, Regulation and
Strategy for Telecommunications, 4 (1), 56-63.
Wall, T. D. & Wood, S. J. (2005). The romance of human resource management and
business performance, and the case for big science. Human Relations, 58 (5), 429462.
Westbrook, R. A. (1981) Sources of satisfaction with retail outlets. Journal of
Retailing, 57 (Autumn), 68-85.
Wright, P., McMahan, G., & McWilliams, A. (1994). Human resources and sustained
competitive advantage: a resourced-based perspective, International Journal of
Human Resource Management, 5 (2), 301-326.
Wooldridge, M. J. (2002). An introduction to multiagent systems. Wiley: New York,
NY.
XJ Technologies. XJ Technologies - Simulation Software and Services. Available via
<www.xjtek.com> [Accessed April 1, 2007].
28
FIGURES AND TABLES
Customer Agent
Visual Dynamic Stochastic Simulation Model
Customer Agent
Shopping need, attitudes,
Customer Agent
demographics etc.
Customer Agent
Emergent behaviour on
macro level
Understanding about
interactions of entities within
the system
Sales Staff Agent
Sales Agent
Attitudes, length of service,
competencies, training etc.
Identification of bottlenecks
Manager Agent
Leadership quality, length of
service, competencies,
training etc.
Global Parameters
Performance Measures
Number of customers, sales
staff, managers etc.
Staff utilisation, average
response time, customer
satisfaction etc.
Interface for User
Interaction during Runtime
Figure 1. Conceptual model of the simulation model.
Customer State-Chart
Queuing at till
Being served at till
(for refund)
(refund decision)
Seeking refund
Browsing
Contemplating
Leaving
Entering
(dummy state)
Queuing at till
Seeking help
Queuing for help
(to buy)
Being helped
Being served at till
(buying)
Figure 2. Conceptual model of a customer agent.
29
Figure 3. Customer (left) and staff (right) agent logic implementation in AnyLogic™.
Situation
leave browse state after …
leave help state after …
leave pay queue (no patience) after …
Min
1
3
5
Mode
7
15
12
Max
15
30
20
Table 1. Sample frequency distribution parameter values.
Event
someone makes a purchase after browsing
someone requires help
someone makes a purchase after getting help
Probability of event
0.37
0.38
0.56
Table 2. Sample probabilities.
30
Department Cashiers
Number of transactions
Mean
4853.50
9822.20
14279.90
14630.60
13771.85
8133.75
15810.10
25439.60
30300.70
28894.25
1
2
3
4
5
1
2
3
4
5
A&TV
WW
S.D.
26.38
57.89
96.34
86.19
97.06
22.16
56.16
113.66
249.30
195.75
Number of satisfied
customers
Mean
S.D.
12324.05
77.64
14762.45
81.04
17429.70
103.77
17185.00
99.09
16023.20
82.66
18508.20
88.68
22640.40
92.00
28833.10
115.65
32124.60
230.13
30475.20
176.41
Overall customer
satisfaction
Mean
S.D.
9366.40
563.88
19985.20
538.30
28994.80
552.60
32573.60
702.64
27916.05
574.56
17327.95
556.03
42339.10
736.61
58601.10
629.68
74233.30
570.79
76838.65
744.31
Table 3. Descriptive statistics for the first validation experiment (to 2 d.p.).
Satisfaction
Department weight value
level
1
A&TV
2
3
1
WW
2
3
Scenario 1
Mean
16,381.20
33,095.00
49,720.20
30,383.40
60,507.80
90,687.60
Scenario 2
S.D.
Mean
S.D.
242.42
27,624.55
797.73
568.22
274,347.00 4,061.43
881.91 2,743,685.00 59,264.25
286.14
63,219.15
812.36
699.92
632,188.00 8,101.63
1,655.07 6,344,035.00 78,434.26
Scenario 3
Mean
S.D.
27,624.55
797.73
27,541.75 1,973.94
-581,334.95 29,429.71
63,219.15
812.36
121,107.70 1,780.45
260,411.75 23,349.43
Table 4. Descriptive statistics for the second validation experiment (to 2 d.p.).
Mean overall customer satisfaction
100000
90000
80000
70000
60000
A&TV
50000
WW
40000
30000
20000
10000
0
1
2
3
Satisfaction weight level value
Figure 4. Scenario 1: Satisfaction weight level value by mean overall customer satisfaction.
31
Mean overall customer satisfaction
7000000
6000000
5000000
4000000
A&TV
WW
3000000
2000000
1000000
0
1
10
100
Satisfaction weight level value
Figure 5. Scenario 2: Satisfaction weight level value by mean overall customer satisfaction.
Satisfaction weights
1/2/4
1/4/16
1/16/256
Mean overall customer satisfaction
300,000
200,000
100,000
0
-100,000
A&TV
-200,000
WW
-300,000
-400,000
-500,000
-600,000
Figure 6. Scenario 3: Satisfaction weights by mean overall customer satisfaction.
Empowerment level
0
0.25
0.5
0.75
1
Number of
transactions
Mean
S.D.
15133.85 102.02
15114.75 60.04
15078.95 86.24
15008.45 52.53
14920.15 66.42
Overall customer Overall customer
satisfaction for
satisfaction for
shopping
refund
Mean
S.D.
Mean
S.D.
23554.30 892.55 -3951.40 288.84
24331.35 907.02 -2316.10 187.23
24476.95 907.48
-932.40 243.25
25213.10 898.61
613.70 182.03
25398.95 1092.50 1892.80 237.69
Table 5. Descriptive statistics for task empowerment experiment (to 2 d.p.).
32
Empowerment level
0
0.25
0.5
0.75
1
Normal
Utilization of Utilization of
expertise
normal staff
expert staff
Mean S.D. Mean S.D. Mean S.D.
0.00 0.00
0.82 0.01
0.93 0.00
18.36 2.10
0.83 0.01
0.94 0.00
35.66 2.54
0.84 0.01
0.94 0.00
53.44 2.98
0.85 0.01
0.94 0.01
69.35 2.85
0.85 0.01
0.94 0.00
Number of
transactions
Mean
S.D.
14830.00 99.82
14801.00 73.56
14782.00 79.90
14787.00 96.45
14823.00 80.42
Overall customer
satisfaction
Mean
S.D.
28004.00 823.19
26937.00 960.37
26310.00 916.38
25678.00 1269.68
24831.00 1043.79
Table 6. Descriptive statistics for empowerment to learn experiment (to 2 d.p.).
Promotional
threshold
0
0.2
0.4
0.6
0.8
1
Normal staff
expertise
Mean
0.00
0.00
0.00
0.00
67.68
68.83
S.D.
0.00
0.00
0.00
0.00
3.23
3.84
Normal staff
utilisation
Mean
0.00
0.00
0.00
0.00
0.86
0.86
S.D.
0.00
0.00
0.00
0.00
0.01
0.00
Expert staff
utilisation
Mean
0.77
0.80
0.82
0.86
0.93
0.94
S.D.
0.01
0.01
0.01
0.01
0.02
0.00
Number of
transactions
Mean
15482.35
15302.85
15125.15
14945.30
14801.95
14827.90
S.D.
97.66
75.00
52.03
118.41
92.79
76.14
Table 7. Descriptive statistics for employee development experiment (to 2 d.p.).
33
Overall customer
satisfaction
Mean
46125.25
40723.95
34992.75
28958.80
24661.75
24668.80
S.D.
1099.48
1209.39
1770.02
1460.78
1058.27
843.84
| 5cs.CE
|
1
Inter-Robot Interactions in
Multi-Robot Systems Using Braids
arXiv:1509.04826v1 [cs.MA] 16 Sep 2015
Yancy Diaz-Mercado and Magnus Egerstedt
Abstract—This paper describes a framework for multi-robot
coordination and motion planning with emphasis on inter-agent
interactions. We focus on the characterization of inter-agent
interactions with sufficient level of abstraction so as to allow
for the enforcement of desired interaction patterns in a provably
safe (i.e., collision-free) manner, e.g., for achieving rich movement
patterns in a shared space, or to exchange sensor information. We
propose to specify interaction patterns through elements of the
so-called braid group. This allows us to not focus on a particular
pattern per se, but rather on the problem of being able to execute
a whole class of patterns. The result from such a construction is
a hybrid system driven by symbolic inputs that must be mapped
onto actual paths that both realize the desired interaction levels
and remain safe in the sense that collisions are avoided.
I. I NTRODUCTION
M
ANY applications have been proposed for multirobot systems. For example, the multi-robot foraging
paradigm, in which agents wander around an environment
searching for items of interest [1], [2], can relate to many
real-world problems (e.g., waste or specimen collection in
hazardous environments, explosive detection and disposal, and
search and rescue operations after a crisis such as floods and
earthquakes). Another application of interest in the research
community and in industry is cooperative assembly [3], [4] and
self-assembly [5], [6]. In the cooperative assembly scenario,
multiple robots need to coordinate their motion in order to
cooperatively assemble a structure using a possibly heterogeneous set of tools, end effectors, skills, or parts. Similarly, in
the self-assembly scenario, multiple robots need to coordinate
their motion in order to collectively form a structure or achieve
a geometric configuration. There has been a recent push for
robotic farming and precision agriculture [7], [8], where a fleet
of robots is sent to gather data on the status of crops, tend
to and harvest them. In some communication architectures,
mobile agents called data MULEs or message ferries [9]–
[11] are used to transport data between sensors, access points,
or base stations, in situations where it is impractical to
establish physical communication. Multi-robot simultaneous
localization and mapping (SLAM) takes advantage of the robot
team’s size to attain a more complete (or more detailed) map
faster than with a single robot by combining information from
multiple sources and coordinating motion [12], [13]. Other
applications include transportation systems (e.g., intelligent
highways, air traffic control) [14], and the convoy protection
scenario (e.g., surveillance, coverage) [15].
The authors are is with the Department of Electrical and Computer
Engineering, Georgia Institute of Technology, Atlanta, GA, 30332 USA.
E-mail: {[email protected], [email protected]}.
In many of these applications, the overall objectives can
be stated in terms of making a team of robots follow a
physical path, such as a road or the movements of a ground
convoy, while ensuring that particular search patterns are
executed [16]–[18]. These patterns should be selected in such
a way that certain secondary geometric objectives are met,
including ensuring that an area along the path is covered, that
multiple views of the same objects are achieved, that an aerial
vehicle is always on top of the convoy, or that sufficiently
many vehicle-to-vehicle interactions take place for the purpose
of information sharing [19], [20]. In this work, we collect
all of these different secondary objectives under one unified
banner, namely multi-robot mixing. In particular, we specify
interaction patterns with certain desired levels of mixing, and
then proceed to generate the actual cooperative movements
that realize these mixing levels.
In this paper we study the problem of characterizing interrobot interactions for the sake of coordination and collision
avoidance. We specify the mixing patterns through elements
of the so-called braid group [21], [22], where each element
corresponds to a particular mixing pattern, i.e., we do not focus
on a particular pattern per se, but rather on the problem of
being able to execute a whole class of patterns. The result
from such a construction is a hybrid system driven by symbolic
inputs [23], i.e., the braids, that must be mapped onto actual
paths that both obtain the mixing level specified through the
braid, and remain safe in the sense that collisions are avoided.
The use of symbolic inputs allows us to abstract away
the geometry and physical constraints involved in the multirobot motion planning. As described in [23], this provides
the advantage of hierarchical abstractions that are typically
broken into three stages. At the top layer is the specification
level, which describes the motion tasks (such as robots A and
B should interact, and arrive at the goal simultaneously). The
second layer is the execution level, which describes how to
obtain the motion plans, e.g., by generating trajectories based
on optimality conditions. The bottom layer is the implementation level, which concerns itself with constructing the robot
controller, e.g., to track a reference trajectory. This work also
extends the notions presented in [24] for trajectory tracking of
virtual vehicles. The idea there is to have the physical vehicles
track a virtual vehicle, as opposed to a path (or trajectory)
itself. The virtual vehicle, which is being controlled directly
with simplified dynamics, can be controlled in order to track a
reference path while satisfying the constraints, compensating
for the physical vehicle’s dynamics and other disturbances.
The outline of the paper is as follows: in Section II we
start with a brief summary of the existing literature on braid-
2
ing for the sake of multi-robot motion planning; in Section
II-B, the braid group is introduced as a way of specifying
mixing levels and the corresponding symbolic braid objects
are given a geometric interpretation in terms of planar robot
paths; controllers are then proposed so that mixing strategies
satisfying a given specification can be executed by a class of
robots as described in Section III, together with a bound on the
highest achievable mixing level. In Section IV, implementation
of these controllers on actual robotic platforms is addressed,
and these ideas are deployed on a team of actual mobile robots.
II. B RAIDS AND ROBOTICS
The use of braids for robot motion planning has been
considered before. Using the notion of configuration space
[25] for robot motion planning, [26] studies the problem
of construction and classification of configuration spaces for
graphs, e.g., robots on a manufacturing floor constrained on
rails or paths. By studying the topological data associated with
these graphs, such as the braid groups, he is able to provide
a measure of the complexity of the control problem (e.g., the
construction of potential field controllers on homeomorphic
spaces).
In [27], the graph braid group is used as the fundamental
group of the configuration space of graphs that describe robot
motion. There, each graph in the configuration space represents discretized collision-free robotic motion plan (e.g. road
maps), where at each discrete time the graph vertices represent
the positions of robots and (possibly moving) obstacles, and
the edges represent fixed tracks connecting these vertices.
They present an algorithm to construct the presentation of the
graph braid group of N agents, where the group generators
(i.e., the braids) represent actual paths between configurations
of robots on the graph, i.e., the motion plan to transition from
one configuration to another. However, they only consider
zero-size robots where they rely on a “one edge separation”
between points at all times to avoid collisions, and as such,
their approach is purely academic and mainly focuses on the
combinatorics of ideal robots moving on a graph.
We will use the braid group’s generators as symbolic
inputs to multi-robot hybrid controllers which characterize
and enforce collision-free interactions, take into account kinematic and geometric constraints, and are executable on actual
robotic platforms. This is the first approach that uses braids
in multi-robot motion planning to symbolically characterize
and enforce rich interaction patterns, implementable on actual robotic platforms. In particular, we extend the notions
presented in [28], [29]. The definition on planar braids and
their geometric interpretations are generalized from their original form presented in [28]. A controller was designed for
nonholonomic vehicles to track a geometric path in [29].
In this paper, we modify this controller to instead produce
optimal trajectories that are provably safe in the collision-free
sense, satisfy a set of spatio-temporal constraints, and follow
desired geometric paths (Section III-B). This also allows us
to obtain tighter bounds on the amount of interaction patterns
that are achievable in a space than the one found in [29]. Two
new contributions to this work are found in Section III-C,
1
1
1
2
1
2
2
2
2
1
2
1
σ0
σ1
σ̂1
(a) Two robots, orig- (b) Three braids for two planar agents. Note that
inally at one side of “strand” 1 goes over “strand” 2 in σ1 and vice versa
a square, map to the in σ̂1 .
other.
Fig. 1. Two-robot interactions and corresponding symbolic braids.
where these trajectory generating controllers are extended to
non-rectangular regions, and Section IV-A, which includes
development of an optimal trajectory tracking controller with
formal guarantees on optimality and spatio-temporal constraint
satisfaction.
A. Planar Braids
Consider two agents on a square, initially located at the
two left vertices of the square as in Fig. 1a. The agents’ task
is to move to the two right vertices of the square. There are
two ways in which these target vertices can be assigned. The
first is to simply let the robots move along a straight line
while the second is to have them cross paths and move to
vertices diagonally across from their initial placement. If the
robots are not to collide with each other, one agent can cross
the intersection of the two paths first, and then the other (or
vice versa). In the braid group, these two options correspond
to different “braids,” and we have thus identified three planar
braids for two agents, as shown in Fig. 1b. Let us momentarily
denote these three braids, σ0 , σ1 , σ̂1 .
Now, given these three braids, we can concatenate them
together to form other braids, as seen in Fig. 2. The left braid is
given by σ1 ·σ1 and the right braid is σ1 ·σ̂1 . In the braid group,
what really matters is not the geometric layout of the paths, but
how the paths wrap around each other. As can be seen, if we
were to “pull” the right corners in σ1 · σ1 , the “strands” would
get “tangled” in the middle, while a “stretched-out” σ1 · σ̂1 is
simply σ0 . Thus we let σ0 be the identity braid, such that σ1
and σ̂1 are each others’ inverses in the sense that
σ1 · σ̂1 = σ̂1 · σ1 = σ0 .
In fact, every braid has an inverse and, as such, the set of braids
(together with the concatenation operation) is indeed a group.
And, as σ1−1 = σ̂1 (and σ1−1 = σ̂1 ), σ0 and σ1 (or σ0 and
σ̂1 for that matter) are the so-called generator braids for this
group in that all planar braids can be written as concatenations
of these two braids and their inverses [21], [22].
This notion of generator braids can be extended to the case
when there are N ≥ 2, with the only difference being that
we will have N generators rather than just two, i.e., let σ0
be the trivial generator with no interactions and σk denote the
interactions between agents k and k + 1, k = 1, . . . , N − 1.
If we let ΣN be the set of all planar generator braids over
N agents, this set will serve as the alphabet over which braid
strings (themselves braids) are produced from, and we let ΣM
N
denote the set of all braids of length M (i.e., braids composed
of M generators) over N agents.
3
σ(2)
σ(1)
yy
σ1 · σ1
σ1 · σ̂1
ξ (0, 1)
ξ (1, 1)
ξ (2, 3)
ξ (0, 2)
ξ (1, 3)
ξ (2, 1)
ξ (0, 3)
ξ (1, 2)
ξ (2, 2)
Fig. 2. Two concatenated braids σ1 · σ1 and σ1 · σ̂1 . The latter of these two
braids is the same as the identity braid σ0 .
B. A Geometric Interpretation
}
h
|
Although heavily inspired by geometry in [21], the braid
group is not concerned with the actual geometry of the braid
strands. For the sake of describing the robot motion plans,
we will associate geometric paths with the different braids.
First of all, we assume that the braid is geometrically located
in a rectangular area of height h and length ` no matter
how long the braid string is. Using the particular two-agent
braids discussed in the previous paragraphs, we assume that
the two agents are initially located at the points (0, 0) and
(0, h), while the final locations are at (`, 0) and (`, h). If
the total braid results from the use of one single generator
braid σ0 , or σ1 then no additional points are needed. However,
if the braid has length 2, then we also need to introduce
intermediary “half-way”
(`/2,
q
points
0) and (`/2, h). As such,
q
we let P2q =
`,
0
,
`,
h
be a set of uniformly
M
M
spaced1 positions for these intermediary points, where the
subscript 2 denotes the two-agent case, M is the length of
the braid to be executed, and q = 0, 1, . . . , M .
Using this notation, we can refer back to Fig. 1a and say that
each of the two generator braids correspond to an assignment,
i.e., a bijective map, between P20 and P21 , and we use the
following notation to denote this fact
{z
}
`
P30 = {(0, 0) , (0, 0.5h) , (0, h)}
P32 = {(`, 0) , (`, 0.5h) , (`, h)}
P31 = {(0.5`, 0) , (0.5`, 0.5h) , (0.5`, h)}
Fig. 3. The geometric interpretation of braid string σ = σ2 · σ1 for the
three-agent case. In this example, σ(1) = σ2 and maps P30 to P31 , while
σ(2) = σ1 and maps P31 to P32 .
where σ (k) is the k th braid2 in the string σ and
n
q
q
q
1
PN
=
`,
0
,
`,
h
,
M
M
(N −1)
o
q
q
2
,
M `, (N −1) h , . . . , M `, h
as shown for the three-agent case in Fig. 3.
We moreover use the notation ξ(i, j) ∈ R2 to denote the
point agent j should go to at step i, i = 1, . . . , M . We use
the convention that ξ(0, j) = (0, (j − 1)h/(N − 1)), j =
1, 2, . . . , N , to denote agent j’s initial position. In other words,
ξ(1, j) = σ(1)hξ(0, j)i,
ξ(2, j) = σ(2)hξ(1, j)i = σ(2) ◦ σ(1)hξ(0, j)i,
or more generally,
ξ(i, j) = σ(i)hξ(i − 1, j)i
= σ(i) ◦ σ(i − 1) ◦ · · · ◦ σ(1)hξ(0, j)i,
σi :
P20
→b P21 ,
i = 0, 1,
where →b denotes “bijection”. Note that this is a slight abuse
of notation in that σi now denotes both an element in the braid
group as well as a map – this distinction, however, should be
entirely clear from the context. Further, we will refer to these
points which agents are bijectively mapped to and from as
braid points.
If we generalize this to N ≥ 2 agents and let σ denote a
string of generators of length M ≥ 2, i.e., σ ∈ ΣM
N , we will
use the notation
where we use the h·i notation to denote the argument to
σ(i) and ◦ to denote composition. This construction is also
illustrated in Fig. 3 for the three-agent case.
The geometric interpretation we will make of the planar
braids is that the mobile agents that are to execute them must
traverse through these braid points. They must moreover do
so in an orderly and safe manner, which will be the topic of
the next section.
III. E XECUTING B RAIDS
Given a collection of N agents with dynamics
ẋj = f (xj , uj ),
(k−1)
σ(k) : PN
k
→ b PN
,
k = 1, . . . , M,
and planar output
yj = h(xj ) ∈ R2 , j = 1, . . . , N,
1 Unless otherwise stated, the rest of this document we will assume these
sets of intermediary points are uniformly spaced in h and `, but the notions
presented here extend to points which are not, as the application demands.
More on this in Section III-C.
2 Note that a “braid” here refers to a member of the braid group, e.g., a
single generator (viz., a single bijective map) or a concatenation of several
generators (viz., a composition of bijective maps).
4
then it is of interest to have these agents execute a braid
σ ∈ ΣM
N . We now define what it means for this braid to be
executed.
Given an input braid string σ, what each individual agent
should do is “hit” the intermediary braid points ξ(i, j) at
specified time instances. We let T denote the time it should
take for the entire string to be executed. As such, the first
condition for a multi-agent motion to be feasible with respect
to the braid is the following:
Definition 3.1 (Braid-Point Feasibility):
A multi-robot trajectory is braid-point feasible if
yj (ti ) = ξ(i, j), i = 0, . . . , M, j = 1, . . . , N.
where the ti ’s form a partition of a given time window [0, T ],
i.e.,
t0 = 0 < t1 < · · · < ti < · · · < tM = T.
On top of braid-point feasibility, we also insist on the robots
not colliding during the maneuvers. To a certain degree, this
condition is what restricts the level of mixing that is possible,
i.e., since the braid is constrained in a rectangle of fixed height
and width, what length strings the multi-robot system can
execute while maintaining a desired level of safety separation.
Definition 3.2 (Collision-Free):
A multi-robot trajectory is collision-free if
kyi (t) − yj (t)k ≥ δij ,
∀i 6= j,
A. Braid Controllers: Stop-Go-Stop Hybrid Strategy
Our first attempt at executing braids will be a hybrid control
strategy called the Stop-Go-Stop. We will assume that agent
dynamics are given by single integrator dynamics, i.e., ẋj =
uj ∈ R2 with yj = xj , j = 1, . . . , N . Further, for practical
considerations, assume that there is a cap on the maximum
velocity achievable by the agents, i.e., kuj k ∈ [0, vmax ]. The
idea is then at each braid step to “send” agents off straight
to their next braid point in order of the distance they’ll need
to travel, waiting just long enough to avoid collisions before
sending an agent off. To that end, we define si : {1, . . . , N } →
{1, . . . , N } to be a bijective mapping that denotes the farthest
distance ordering at step i, that is
si (p) < si (q)
where ties are arbitrarily broken such that si remains a
bijection.
heading angle be given by θi,j =
We let the agents
(ξ(i,j)−ξ(i−1,j))
tan−1 (ξ(i,j)−ξ(i−1,j))2 where the subscript indicates the
1
first or second component, and the unit heading vector be
T
ρ̂i,j = [cos(θi,j ), sin(θi,j )] . Lastly, the time the agents will
wait before entering GO mode will be given by their ordering
as (si (j) − 1)τ where
τ=
t ∈ [0, T ],
where δij > 0 is the desired level of safety separation between
agents i and j.
For convenience, we will refer to δ̄ = max δij as the
maximum safety separation such that no agent collides. This
will come up in Theorem 2. We are missing a notion to
describe what the multi-robot mixing problem is, that is, what
constitutes a braid controller.
Definition 3.3 (Braid Controller):
A multi-robot controller is a braid controller if the resulting
trajectories are both braid-point feasible and collision free, for
all collision-free initial conditions such that
yj (0) = ξ(0, j), j = 1, . . . , N.
=⇒
kξ(i, p) − ξ(i − 1, p)k ≥ kξ(i, q) − ξ(i − 1, q)k
As a final notion, we are interested in how much mixing a
particular system can support.
Definition 3.4 (Mixing Limit):
The mixing limit M ? is the largest integer M such that there
exists a braid controller for every string in ΣM
N.
Whenever two strands of the braid associated with a given
braid string cross, the two associated agents will have to interact. The mixing limit therefore serves as an input-independent
bound on how much mixing is achievable for a given team of
agents operating in a given environment. The mixing limit is
in general quite hard to compute, it needs to consider every
permutation of strings of varying lengths up to some number,
the geometry assigned to each string and is dependent on
the kinematical response of the multi-robot system. However,
under certain assumptions it is possible to find bounds on M ?
for a given braid controller.
δ
vmax cos(θ∗ )
is the time required to be δ apart horizontally3 , with cos(θ? ) =
√ 2 `/M2 2 being the maximum horizontal distance an agent
` /M +h
could travel given σ ∈ ΣM
N . To ensure that the agents do
not overtake the first agent (horizontally), the speed of the
agents should be scaled by the velocity of the first agent, i.e.,
uj = vmax cos(θi,s−1 (1) )/ cos(θi,si (j) ). The hybrid automaton
i
describing the stop-go-stop controller is given in Fig. 4.
Theorem 1: [28] The STOP-GO-STOP controller in Fig. 4 is
a braid controller if the braid points themselves are sufficiently
separated and
cos(θ? )vmax min (ti − ti−1 ) − (N − 1)τ
i
p
≥ `2 /M 2 + h2 .
Proof:
The STOP-GO-STOP controller ensures that the agents are
never within δ of each other by virtue of the fact that they
have to wait until they are indeed at least that far apart (horizontally) before entering GO mode. As such, the trajectories
are collision-free.
What remains to show is that they are also braid-point
feasible. Consider the agent that has to wait the longest before
it can move, i.e., the agent that has to wait a total of (N −1)τ ,
and at its worst has mini (ti − ti−1 ) − (N − 1)τ time left
to reach the next braid point. In other words, we need that
the distance traveled in that amount of time at the speed
3 Since we are interested in the mixing limit, the analysis is done with the
horizontal direction being the limiting factor for safe execution of the braids.
5
Mixing Bounds
t = (si (j) − 1) τ + ti−1
i := 1
i := i + 1
t = ti
STOP
ẋj = 0
GO
cos θi,s−1 (1)
i
ρ̂i,j
ẋj = vmax
cos θi,si (j)
xi = ξ (i, j)
Fig. 4. Hybrid STOP-GO-STOP braid controller.
vmax cos(θi,s−1 (1) )/ cos(θi,si (j) ) is greater than the distance
i
required. But, we note that
vmax cos(θi,s−1 (1) )/ cos(θi,si (j) ) ≥ vmax cos(θ? )
i
and, as we are only looking for a bound, we assume that we
use this lower speed and that the distance required to travel is
the largest distance possible (which it really is not). In other
words, we need
cos(θ? )vmax min (ti − ti−1 ) − (N − 1)τ
i
p
≥ `2 /M 2 + h2 ,
and the proof follows.
Note that Theorem 1 implicitly provides a lower bound on
the mixing limit as long as the agents’ paths are straight lines.
In Fig. 5 we can see the mixing limit as a function of the
number of agents in the team for parameters vmax = 5, T =
20, ` = 5, h = 10, and ∆ = 0.2. The problem with this
strategy is that it does not allow for more general geometric
paths (which could be interpreted as feasible trajectories under
a given robot dynamical model), nor does it ensure that agents
get within a specified distance from each other (which could be
necessary when collaboration is required). The next section we
present a new strategy called braid reparameterization which
will explicitly consider more general strand geometries, and
allow us to obtain analytical bounds on the mixing limit.
B. Braid Controller: Braid Reparameterization
We are now seeking a strategy that will allow us to follow
a given geometry while achieving a mixing strategy encoded
as a braid string σ ∈ ΣM
N . Further, we wish to enforce interagent interaction as dictated in σ by having agents get as close
as the safety separation δjk . Before moving forward, consider
the following lemma.
Lemma 1: If at any braid step, the generators in the braid
substring σ(k) = σ ⊆ Σm
N , m ≤ M , have indices that are two
or more apart, then any agent interacts with at most one other
agent at this step.
Proof:
k
Let BN
∈ RN ×2 be a matrix that contains the set of
0
braid points at time k such that BN
= [ξ(0, 1) · · · ξ(0, N )]T .
Consider the two-generator concatenation σi ·σj . As a bijective
map, if σi = σ0 , then the agents do not interact, and the agent
k−1
in braid point position [BN
]n gets mapped to braid point
Number of Generators M
STOP
ẋj = 0
10
9
8
7
6
5
4
3
2
1
0
10
20 30 40 50 60
Number of Agents N
70
Fig. 5. Lower bound on the mixing limit using the Stop-Go-Stop braid
controller.
k
]n , where [B]n corresponds to the nth row of B,
position [BN
n = 1, . . . , N . If σi 6= σ0 , then σi will swap the position of the
two agents occupying the braid point positions i and i + 1 at
k−1
k
]i+1
step k−1, i.e., it maps the agent occupying [BN
]i to [BN
k−1
k
and the agent occupying [BN ]i+1 to [BN ]i . Similarly, σj
swaps the position of the agents occupying the braid point
positions j and j + 1.
The two agents in positions i and i + 1 at k − 1 would only
interact with the two agents in positions j and j + 1 at k − 1
if i (or i + 1 for that matter) is equal to either j or j + 1. But
if we let |i − j| ≥ 2, then we get that
|i − j| ≥ 2 ⇔ 2 ≤ i − j or 2 ≤ j − i
⇒ j <j+1<j+2≤i<i+1
or i < i + 1 < i + 2 ≤ j < j + 1
i 6= j
i 6= j + 1
⇒
i + 1 6= j
i + 1 6= j + 1,
and as such these two-generator concatenation maps the agents
from one set of braid points to the next with at most two
interaction between at most two agents per interaction.
One of the two braid group relations [21], [22] tells us that
if |i − j| ≥ 2 then σi · σj = σj · σi . More generally, if we let
h : {1, . . . , m} → {1, . . . , N − 1}
be a surjective map such that |h(i) − h(j)| ≥ 2 for all
i, j ∈ {1, . . . , m} with i 6= j, then for any bijective map
g : {1, . . . , m} → {1, . . . , m} we have that
σ(k) = σh(1) · σh(2) · · · · · σh(m)
= σh(g(1)) · σh(g(2)) · · · · · σh(g(m)) .
As such, the braid generators can be rearranged to obtain any
permutation of two-generator concatenations from generators
in σ(k). Since for all permutations of two-generator concatenations we will have indices that are two or more apart, these
6
{σ1 · σ3 · σ2 }
= {σ3 · σ1 · σ2 }
(a)
{σ1 · σ3 } · σ2
= {σ3 · σ1 } · σ2
(b)
in order to safely execute braids, one need only consider the
case when two agents interact. We will devise a strategy for
reparameterizing the geometry such that should two agents
interact, the resulting parameters are at least δ from each other.
By tracking the parameterized paths, the agents’ trajectories
will be collision-free and braid point feasible. The resulting
controllers can be combined to satisfy a given braid string.
Consider the geometric path agent j must follow at step
i given by γij : [0, 1] → R2 with γij (0) = ξ(i − 1, j) and
γij (1) = ξ(i, j). Let ∆ be the arclength of this path, given by
Z 1 r
T
γ̇ij (p) γ̇ij (p) dp.
∆=
0
σ1 · σ3 · σ2
(c)
σ3 · σ1 · σ2
(d)
Fig. 6. A braid string σ = σ1 · σ3 · σ2 taking place in a varying number of
braid steps.
will map the agents at braid step k from the set of braid
k−1
k
with at most
points BN
to the next set of braid points BN
m interactions total, and at most one interaction per agent.
If the geometric interpretation of the braid string is restricted
to the case of only pairwise interactions at every braid step,
then it is possible to devise a hybrid strategy with the desired
properties, which we can then compose together to achieve
the desired interaction patterns. As such, we will restrict the
geometric interpretation of braid strings to those satisfying
pairwise interactions as in Lemma 1.
Restriction 1: As a bijective map from one set of braid
points to another, the braid σ(k) will only contain generators
whose indices are two or more apart, i.e., for some h : N →
{0, . . . , N − 1}
σ(k) = σh(1) · σh(2) · · · · , where |h(i) − h(j)| ≥ 2 ∀i 6= j
As an example, suppose that a braid string contains the
substring of three concatenated generators σ1 · σ3 · σ2 . The
first two generators, σ1 · σ3 , may take place simultaneously at
braid step 1 without incurring in more than one interaction per
agent with another agent, but we would require an additional
braid step for σ2 in order to avoid multiple interactions in the
same step. This is illustrated geometrically in Fig. 6. In (a),
an agent interacts with more than one other agent since all
the generators are not at least two indices apart and take place
at the same braid step. Note that in the remaining cases (b)(d), we introduce intermediate braid points while retaining the
desired level of interaction, final configuration of the agents,
and reducing to pairwise interactions.
Braid strings that satisfy Restriction 1 will have at most
interactions involving two agents at any given braid step. So
We wish to find a parameterization of the strand geometries
such that the parameters of two intersecting strands, thought of
as a virtual vehicles that live on their geometries, are collisionfree. The strategy to do so will be to impose constraints on
the agents separation from the path intersection at a specified
time. To ensure braid-point feasibility, the strategy will be to
impose constraints on the time in which the agents must get
from the beginning of one braid step to the end of that step. To
that end, we propose the following constrained optimization
problem
Z
1 ti 2
? ?
(vj , vk ) = arg min J(vj , vk ) = arg min
vj + vk2 dτ
(vj ,vk )
(vj ,vk ) 2 ti−1
(1)
subject to
ṗj = vj ,
∆+δ
ti−1 + ti
=
,
pj
2
2∆
pj (ti−1 ) = 0,
pj (ti ) = 1,
and
ṗk = vk ,
∆−δ
ti−1 + ti
pk
=
,
2
2∆
pk (ti−1 ) = 0,
pk (ti ) = 1.
The constraints at ti−1 and ti will ensure that the reparameterization is braid-point feasible. For collision avoidance we
will define a safety separation region, as depicted in Fig.
7, to be the region from the intersection point to the point
where it is possible for the
agents
two ∆+δ
to be within δjk of
,
for δ ∈ [0, ∆] such
each other, i.e., the set ∆−δ
2∆
2∆
that γij (pj ) − γik (pk ) ≥ δjk for all pj , pk ∈ [0, 1], where
we let δ be the distance from the intersection point to the
boundary of the safety separation region along the path. Note
that if the geometry of the braid strands are straight lines, the
distance δ can be easily computed by δ = δjk csc(θ), where
θ is the angle between the two intersecting lines, and this
would also imply the parameters get as close as δjk csc (θ/2).
Similarly, when the geometry is “city-block”-like, we can find
δ = δjk + 2(Nh−1) and the agent get as close as 2δjk . For
more general geometries, δ may be conservatively selected.
At t̄i := ti−12+ti , i.e., half-time along the braid step, we
will require one parameter to exit the safety separation region
as the other one is about to enter it. All these cases are
7
δjk
δjk
δ0
θ
δ
δ
Fig. 7. Safety separation region for three different-geometry, two-agent braids.
For straight lines (left) or “city-block”-like (right), the distance δ can be
computed exactly. For arbitrary curves (center), the distance δ 0 can be selected
conservatively.
illustrated in Fig. 7. This way we guarantee that the two
parameters are never inside the region simultaneously and thus
their separation will always be of at least δjk .
The optimality conditions for equation (1) are given by
ṗj = −λj ,
λ̇j = λ̇k = 0,
pj (t̄i ) =
∆+δ
2∆ ,
pk (t̄i ) =
∆−δ
2∆ ,
ṗk = −λk ,
pj (ti ) = pk (ti ) = 1.
Since λ̇j = λ̇k = 0, we will get that λj and λk are piecewise
constant. By using the midway condition, we have that for
t ∈ (ti−1 , t̄i ]
Z
t̄i
vj dτ =
pj (t̄i ) =
ti−1
⇒
∆+δ
2∆
vj (t̄i − ti−1 ) =
⇒
∆+δ
2∆
1 ∆+δ
vj =
∆ (ti − ti−1 )
and similarly for t ∈ (t̄i , ti ], the terminal condition tells us
that
Z ti
∆
∆+δ
pj (ti ) =
=
vj dτ +
2∆
∆
t̄i
1
(2∆ − ∆ + δ)
⇒
vj (ti − t̄i ) =
2∆
1 ∆−δ
⇒
vj =
∆ (ti − ti−1 )
Note that for agent k, the signs are reversed on the numerator. As such, the resulting braid parameterization will have
velocities given by
(
∆±δ
1
t ∈ (ti−1 , t̄i ]
∆ (ti −ti−1 )
ṗj (t) =
(2)
1
∆∓δ
t ∈ (t̄i , ti ]
∆ (ti −ti−1 )
where the sign in the numerator is determined by the interpretation given to the braid strand going “under” (e.g., the agent
crosses the intersection point first) or “over” (e.g., the agent
crosses the intersection point second), and in cases where there
are no intersections we set δ = 0 in the numerator.
It is possible to come up with an upper bound on the
length of the braid attainable under this mixing scheme. Under
Restriction 1, the following theorem provides an upper bound
on the mixing limit.
Theorem 2: Given the maximum safety separation δ̄ and
bounds on the agents’ velocities such that vj (t) ∈ [0, vmax ]
∀t, j, the mixing limit M ? for N -agent braids that can be
performed in a space of height h and length ` in time T is
bounded above by
q
` 4h2 − δ̄ 2 (N − 1)2
?
,
M ≤ min
δ̄h
)
(N − 1) vmax T − ` + δ̄
1
.
−
h
2
Proof:
Consider σ ∈ ΣM
N . Since at any step an agent can either interact with another agent or move straight ahead, the maximum
mixing will be achieved if an agent interacts with another
agent at every step. Thus, at every step it must be enforced
that there are no collisions. Assuming a uniform partition of
the time window, the arclength-normalized parameter velocity
will be given by
M ∆±M δ
if t ∈ i−1
T, 2i−1
T
T
∆
M
2M
vj =
i
M ∆∓M δ
if t ∈ 2i−1
T∆
2M T, M T
for i = 1, 2, . . . , M , where the sign on the numerator depends
on the interpretation of whether the “strand” goes over or
under, ∆ is the arclength of the strand geometry connecting
two braid point, assumed equal at each braid step and for both
agents due to symmetry, and δ is the safety separation distance
along the braid as described above. The total braid path length
∆ heavily depends on the geometry of path. However, for the
sake of obtaining bounds on the mixing limit, we may assume
that the braid points are uniformly distributed in height and
length, such that the length of sufficiently regular paths will
be bounded by
s
2 2
`
h
`
h
+
≤∆≤
+
(3)
N −1
M
N −1 M
where the lower bound assumes straight lines connecting the
braid points and the upper bound assumes “city block”-like
paths stepping midway between the two points. But after
normalizing the bounds on the parameter velocity, we see that
0≤
M (∆ − δ)
M (∆ + δ)
vmax
≤
≤
.
T∆
T∆
∆
(4)
The lower bound of (4) tells us that for the parameter to not
go backwards
δ ≤ ∆. To ensure this, we set δ =
r we need
2
h
` 2
δ̄ csc θ ≤
+ M
≤ ∆, and since we require
N −1
δ̄ ≤ Nh−1 for collision-free braid points, using the geometric
relationships to solve for M yields
q
2
` 4h2 − δ̄ 2 (N − 1)
M≤
.
δ̄h
Similarly, the right-hand side inequality of (4) tells us that
vmax T −( 21 ( Nh
−1 )+δ̄ )
∆ ≤ vmaxMT −δ =
, and to ensure this we
M
vmax T −( 12 ( Nh
−1 )+δ̄ )
h
`
set ∆ ≤ N −1 + M ≤
. Solving for M
M
yields
(N − 1) vmax T − ` + δ̄
1
− .
M≤
h
2
20
(a) Three agents braiding on a curved region.
60
0
40
10
Time Alloted T (s)
Mixing Limit Bound M
8
20
20
Number of Agents N
Fig. 8. Upper bound on the Mixing Limit presented in Theorem 2 for
parameters ` = 2 m, h = 4 m, δ = 0.13 m, vmax = 2 m/s
(b) Two agents need to move
from the left-most circles to
the right-most circles along
the curved region.
(c) The longer the braid length
in this curved road segment,
the closer the quadrilaterals
resemble the curved region.
Fig. 9. Agents braiding on a region that curves.
Thus
M ? ≤ min
q
` 4h2 − δ̄ 2 (N − 1)2
δ̄h
,
(N − 1) vmax T − ` + δ̄
h
1
−
2
)
.
Theorem 2 provides a compact expression to obtain an
upper bound on the mixing limit that abstracts away strand
geometry. It provides a notion of the whether or not desirable
mixing levels are achievable in the space, regardless of what
the actual movement patterns to achieve these mixing levels
are (encoded in the braid string of length M ≤ M ? ). Fig. 8
includes a plot of the bound on the mixing limit for varying
number of agents and time window size.
C. Non-Rectangular Regions
Up to this point, only the problem of braiding on a rectangular region of height h and length l has been considered.
On this region, the braid points were uniformly distributed
along both dimensions and bounds on the mixing limit were
provided through the use of the proposed braid controller. In
this section, the scheme is extended to more generally shaped
regions, e.g., the road on Fig. 9. As has been done previously,
discussion begins by first considering the two agent case.
Consider the two agents attempting to perform a braid
of length one on the arbitrarily curved region on Fig. 9b.
Connecting the braid points together results in the quadrilateral
depicted in the red dotted line. Let this quadrilateral be
considered as the space where the agent needs to perform a
braid of length one, rather than the curved region itself. Note
that as longer length braids are included in this road segment,
other quadrilaterals appended together will be obtained which
approximate the road slightly better. Since it is of interest to
obtain mixing strategies near the mixing limit, as longer length
braids are included in this road segment, better approximations
of the curved region will be obtained by these composition of
quadrilaterals. This is depicted in Fig. 9c.
The strategy for performing a mixing strategy in curved region will be to transform the curved region into a straightened
rectangular region of known height and length, as illustrated
in Fig. 10a. In this way the braid controller can be fashioned
as in previous sections and the resulting braid controller can
be transformed back into the actual curved region. After
distributing the braid points on both the curved region and
the rectangular region, as in Fig. 10a, the next step is to find
a transformation to map between these two regions.
q
Let agents j and k interact in at braid step i. Denote Si,j
to be the quadrilateral formed by connecting together the
braid points ξ(i − 1, j), ξ(i − 1, k), ξ(i, j), and ξ(i, k), like
r
the one depicted in Fig. 9b. Let Si,j
be a rectangular plane
of specified height and length whose corners are given by
ξr (i − 1, j), ξr (i − 1, k), ξr (i, j), and ξr (i, k).
With knowledge of these braid points and through the use of
a projective transform as in [30], it is possible to obtain a local
diffeomorphism that maps from a rectangle of specified height
and length to the convex arbitrarily shaped quadrilateral, i.e.,
q
r
Ti,j : Si,j
→ Si,j
.
Note that by selecting transforms based on the corners of
these quadrilaterals, a continuous curve that spans across the
r
r
boundary between Si,j
and Si+1,j
might be mapped to a
discontinuous curve that spans across the boundary between
q
q
Si,j
and Si+1,j
when transformed using Ti,j and Ti+1,j in their
respective spaces. However, there will certainly be continuity
in the mapping of curves passing through the braid points,
since these points are shared by the quadrilaterals and are
used to compute the transforms, i.e., Ti,j (ξr (i, j)) = ξ(i, j) =
Ti+1,j (ξr (i, j)).
Recall the braid controller proposed for the rectangular
9
point along the curve where the two curves intersect (resp.
where the agent exits the safety separation region).
In the special case where the geometry is given by straight
lines, then by setting
γin (p) = (1 − p)ξ(i − 1, n) + pξ(i, n), p ∈ [0, 1], n = j, k.
(a) The curved region will be mapped to the rectangular region of
known width and height where control design will take place.
(b) The bijective transformation T maps points in the rectangle to
points in the quadrilateral. Both shapes are defined by the braid points
which determine the corners.
as the path agent j and k must follow, the intersection point
s may be found by setting s = γij (πj ) = γik (πk ) where
πj
= A−1 (ξ(i − 1, k) − ξ(i − 1, j))
πk
with A = [(ξ(i, j) − ξ(i − 1, j)) , − (ξ(i, k) − ξ(i − 1, k))].
q
from the intersection point s,
Recall that the distance in Si,j
for the special case of the geometry being straight lines, was
given by δ = δjk csc(θ) with θ being
the angle between these
two lines, i.e., θ = cos−1 x̂>
j x̂k where x̂n is the unit vector
pointing towards the next point, i.e.,
Fig. 10. Rectangular and non-rectangular regions.
x̂n =
region. For agent j, at braid step i, this was given by
∆−δ
1
∆ ti −ti−1 if t ∈ (ti−1 , t̄i ] ,
vj (t) =
∆+δ
1
if t ∈ (t̄i , ti ] .
∆ ti −ti−1
−1
where it is expected that Ti,j
(x (ti−1 )) = ξr (i − 1, j) and
−1
Ti,j (x (ti )) = ξr (i, j). In the expression, ∆ corresponds to
the length of the geometric path agent j must follow to move
between ξr (i − 1, j) and ξr (i, j), while δ corresponds to the
distance along the path agent j switches velocities in order to
avoid collisions. Note that for a given parameterization of the
j
r
geometric path γi,r
(p) in Si,j
with parameter p ∈ [0, 1], the
arclength ∆ may be computed as follows
Z 1q
j>
j
γ̇i,r
(p)γ̇i,r
(p) dp.
∆=
0
q
If the geometric curve is directly given in Si,j
as γij (p) such
j
r
that the curve in Si,j may be parameterized as γi,r
(p) =
j
−1
Ti,j (γi (p)), then the arclength may be computed by
Z 1 r
>
γ̇ij (p) M(γij (p))γ̇ij (p) dp.
∆=
0
where M(γij (p))
=
>
−1
−1
γij (p)
DTi,j
γij (p)
DTi,j
−1
−1
and DTi,j
is the Jacobian of Ti,j
. In the special case
where the geometry is given by straight lines, then ∆ =
kξr (i, j) − ξr (i − 1, j)k.
Finally, it is of interest to find δ in order to avoid collisions.
q
If the parameterization γij (p) of the curve in Si,j
is known,
q
then the safety separation ball may be set in Si,j
as before
and the safety separation distance may be computed as
Z b r
>
δ=
γ̇ij (p) M(γij (p))γ̇ij (p) dp.
a
where [a, b] ⊂ [0, 1], and if the agent is meant to braid “over”
(resp. “under”) then γij (a) corresponds to the point along the
curve where the agent enters the safety separation region (resp.
where the two curves intersect) and γij (b) corresponds to the
(ξ(i, n) − ξ(i − 1, n))
,
kξ(i, n) − ξ(i − 1, n)k
n = j, k.
By setting γ̂(p) = (1 − p)s ± pδ x̂j , where sign depends on
whether the braid goes “over” or “under,” it is possible to
determine δ in the non-rectangular plane directly as
Z 1q
>
(±δ x̂j − s) M (γ̂(p)) (±δ x̂j − s) dp.
δr =
0
r
With this information, vj can be found in Si,j
for interactions
between agents j and k. Note that a reparameterization of the
path can now be set equivalent to the desired trajectory of
j
j
agent j by setting γi,d
(t) = γi,r
(pj (t)) (see (5) below). Thus,
the braid controller parameter velocity vjq (t) for agent j in
q
j
j
Si,j
will be given by vjq (t) = DTi,j (γi,d
(t))γ̇i,d
(t) .
This strategy was implemented in simulation over the
curved region illustrated in Fig. 11. In the figure, agents are
performing the mixing strategy given to them by a braid of
length 80 on the curved region (top) and simultaneously on the
straightened rectangular region (bottom). The parameters used
for this simulation were δjk = 7.7 cm and vmax = 1.5 m/s
∀j, k, and T = 30 s.
IV. I MPLEMENTING B RAIDS
Section II approached the problem of multi-robot mixing
from an execution level, where given a specification it is
possible to synthesize a braid controller that satisfies the
specification. In this section we consider the implementation
level, that is, we address how to find controllers that are
implementable on actual robotic systems that follow the braid
controllers found in the previous section. We then validate the
framework by implementing a mixing strategy on a team of
six robots.
A. Optimal Tracking Controller
We will now utilize the braid parameter velocity to find the
braid controller for the agents. By integrating (2) we obtain
the braid parameterization of the path γij
( t−t
∆±δ
i−1
t ∈ (ti−1 , t̄i ]
(ti −ti−1 ) ∆
pj (t) =
(5)
t−t̄i
∆±δ
∆∓δ
t ∈ (t̄i , ti ] .
(ti −ti−1 ) ∆ + 2∆
10
1.2
North (m)
1
Zoomed in Track (Real World)
3
0.6
2
North (m)
North (m)
Track (Real World)
Zoomed in Track (Real World)
0.8
0.4
0.2
2
North (m)
Track (Real World)
3
1
0.6
0
From Outer to Inner
Track Boundary (m)
0
1
2
3
East (m)
4
5
6
Straightened Track (Virtual World)
0.6
0.4
0.2
0
2
0
3.5
4
East (m)
4
6
8
10
12
Counter Clockwise along Path (m, along Center of Path)
14
0
From Outer to Inner
Track Boundary (m)
0
1
0.8
1
2
3
East (m)
4
5
6
0.4
5 5.2 5.4 5.6 5.8
East (m)
Straightened Track (Virtual World)
0.6
0.4
0.2
0
2
4
6
8
10
12
Counter Clockwise along Path (m, along Center of Path)
(a) Simulation Time 5.04 seconds.
14
(b) Simulation Time 9.57 seconds.
Fig. 11. Five agents performing a mixing strategy given by a braid of length 80. The top left plot represents the actual curved region the agents are mixing
in and the top right is a close-up of the agents. The lower plot is the virtual “straightened” rectangular region where the design of the braid took place.
Assume that agents have single integrator dynamics, i.e., ẋj =
uj with yj = xj ∈ R2 . The agent’s controller will be found
by optimally tracking the reparameterized path γij (pj (t)) to
minimize the cost
Z
T
1 ti
J(uj ) =
xj − γij Q xj − γij + uTj Ruj dτ (6)
2 ti−1
for Q = QT 0 and R = RT 0, with constraints
ẋj = uj , xj (ti−1 ) = γij (0), and xj (ti ) = γij (1). Using
the standard variational argument together with Pontryagin’s
minimum principle, the first order necessary conditions for
optimality tell us that the optimal tracking controller u∗j is
given by
u∗j = −R−1 λj
where λ is the so-called costate and satisfies
λ̇j = −Q xj − γij
with unknown terminal condition λj (ti ). Suppose that similarly to [31, Chapter 5.3] we can construct λj as an affine
combination of the unknown λj (ti ) and the state, i.e.,
λj (t) = H(t)xj (t) + K(t)λj (ti ) + E(t)
and similarly, the terminal state as
xj (ti ) = ξ(i, j) = F (t)xj (t) + G(t)λj (ti ) + D(t)
for some yet unknown functions H, K, E, F, G, and D. One
can differentiate these equations and manipulate the equations
to obtain
HR−1 H − Q − Ḣ xj + HR−1 K − K̇ λj (ti )
+ HR−1 E + Qγij − Ė = 0
and
F R−1 H − Ḟ xj + F R−1 K − Ġ λj (ti )
+ F R−1 E − Ḋ = 0.
In order to satisfy these equations for any value of xj (t) and
λj (ti ), the terminal conditions, and after noticing that F =
K T , we obtain that
Ḣ = HR−1 H − Q,
H(ti ) = 02×2
K̇ = HR−1 K,
K(ti ) = I2
T
Ġ = K R
Ė = HR
T
−1
−1
Ḋ = K R
K,
E+
−1
E,
G(ti ) = 02×2
Qγij ,
(7)
E(ti ) = 02×1
D(ti ) = 02×1 .
Note that G(ti ) = 0 and Ġ(t) 0 for all t, which suggests
that G(t) 0 for t < ti . If the problem is not abnormal, i.e.,
there exists a neighboring minimum solution, then G will be
invertible at some t < ti . In particular, by solving backwards
in time in the sequence H → K → E → G → D up to
t = ti−1 , we can find that
λj (ti ) =
G−1 (ti−1 ) ξ (i, j) − K T (ti−1 ) ξ (i − 1, j) − D (ti−1 )
resulting in the feedback optimal trajectory tracking control
law
h
u∗j (xj , t) = −R−1 H(t)xj (t) + K(t)G−1 (ti−1 ) ξ(i, j)
i
− K T (ti−1 )ξ(i − 1, j) − D(ti−1 ) + E(t) (8)
where H, K, E, G and D are the solutions to terminal value
problems in (7). which can be solved numerically backwards
from ti . As it turns out, these conditions are also sufficient for
optimality as presented in the following theorem.
Theorem 3: The tracking controller in (8) is a minimizer to
the cost functional (6) whose optimal value is given by
h
1
−1 T
∗
T
H − KG K
ξ(i − 1, j)
J(u ) = ξ(i − 1, j)
2
i
+ ξ(i − 1, j)T KG−1 (ξ(i, j) − D) + E + ϕ
t=ti−1
11
(a) The robots start at the beginning of the braid.
(b) Collision-free – robots get as close as δ.
(c) The controller is braid point feasible – braid points are reached simultaneously.
Fig. 12. Actual robots executing the mixing strategy in (10). The geometric paths and spatio-temporal constraints are being projected on the workspace with
an overhead projector for the sake of visualization.
where ϕ(t) is the solution to the terminal boundary problem
1
1 j T
T
(Λj (t)) R−1 Λj (t) −
γ (t) Qγij (t)
2
2 i
ϕ(ti ) = −ξ(i, j)T Λj (ti )ξ(i, j)
ϕ̇ =
with
Λj (t) = E(t)+
K(t)G−1 (ti−1 )(ξ(i, j) − K T (ti−1 )ξ(i − 1.j) − D(ti−1 )).
Proof:
As the control law was derived from the necessary conditions
for optimality, we only need to show that it is sufficient for
optimality. We will do so by leveraging the Hamilton-JacobiBellman theorem [32, Chapter 2].
Note that the optimization problem is regular as there exists
a uj that allows the Hamiltonian to achieve a minimum with
respect to it, i.e.,
H(xj , uj , λj )
T
1
j
j
T
Q xj − γi + uj Ruj + λTj uj
=
xj − γi
2
T
1
1
uj + R−1 λj R uj + R−1 λj − λTj R−1 λj
=
2
2
T
1
j
+
xj − γi
Q xj − γij
2
which attains a minimum with respect to uj when
u∗j = −R−1 λj .
12
0.5
1
Distance (m)
North (m)
0.4
0
−1
0.3
0.2
0.1
−2
−1
0
East (m)
1
0
2
(a) Robot trajectories in the plane.
min ≈ 0.132867m
0
10
20
30
40
Time (s)
50
60
(b) Instantaneous minimum inter-robot distance.
Fig. 13. Data associated with robotic implementation in Fig. 12.
Define V (z, t) as
1
V (z, t) = z T H(t)z + z T K(t)G−1 (ti−1 )ξ(i, j)
2
−z T K(t)G−1 (ti−1 ) K T (ti−1 )ξ(i−1, j)−D(ti−1 ) +E(t)
+ ϕ(t)
B. Robotic Implementation
where ϕ(t) is as defined above. It can be verified that
V (z, t) satisfies the terminal condition V (ξ(i, j), ti ) = 0, that
∂V (z,t)
∂z
T
z=xj
= λj , and that is satisfies the Hamilton-Jacobi-
Bellman equation, i.e.,
∂V (z, t)
∂V (z, t)
0 =
+ H xj , u∗j (xj , t) ,
∂t
∂z
z=xj
T
.
z=xj
As a consequence, u∗j is a minimizer to the cost functional
and
J(u∗ ) = V (ξ(i − 1, j), ti−1 ).
As a final note, the terminal costate value λj (ti ) was computed using the initial conditions for the problem. However,
as the gains involved are solved from terminal conditions,
the choice of initial conditions is arbitrary, and evaluating at
t = ti−1 results in control law (8) being open-loop in the
terminal costate value. This could yield undesired results under
the influence of disturbances and errors. To alleviate this, we
can rewrite the terminal costate as a function of the current
state value instead, i.e.,
λj (ti ) =
G−1 (ti−1 ) ξ (i, j) − K T (ti−1 ) ξ (i − 1, j) − D (ti−1 )
= G−1 (t) ξ (i, j) − K T (t) xj (t) − D (t)
in order to obtain the fully closed-loop optimal tracking
controller
u∗j (xj , t) = −R−1 H(t) − K(t)G−1 (t)K T (t) xj (t)
+K(t)G−1 (t) (ξ(i, j) − D(t)) + E(t) . (9)
In order to validate the above results in a practical setting,
the braid controllers were implemented on a team of Khepera
III differential-drive robots, which may be modelled using
unicycle dynamics, i.e.,
T
ẋj = νj cos θj νj sin θj ,
θ̇j = ωj .
where xj ∈ R2 is the robot’s position and θj its heading.
The single integrator control uj from (9) can be mapped to
unicycle dynamics as
νj = cos θj sin θj · uj ,
u
κ [ − sin θj cos θj ] · kujj k , if kuj k > 1
ωj =
κ [ − sin θj cos θj ] · uj ,
otherwise
for tuning gain κ > 0. Fig. 12 illustrates the team of 6 real
robots executing the braid string
σ = {σ1 ·σ3 ·σ5 }·σ2 ·σ3 ·σ4 ·{σ3 ·σ5 }·{σ2 ·σ4 }·σ1
(10)
where the braids grouped in braces are executed simultaneously. A visual representation of the braid string geometry
is being projected onto the robot workspace via an overhead
projector mounted on the ceiling of the lab space. Fig.
13b illustrates the minimum pairwise distance throughout the
execution of the braid. It can be seen that the minimum
distance achieved is approximately 0.132867m – greater than
the size of the robots which is approximately 0.13m. Fig. 13a
illustrates the trajectories taken by the robots and how they
follow the strand geometry, which were selected as straight
lines, while still accommodating for the agents’ unicycle
dynamics.
13
V. C ONCLUSION
The notion of multi-robot mixing is presented as a framework with which to characterize rich movement patterns in
a multi-robot system symbolically in terms of inter-robot
interactions. Two controllers are presented which allow to the
multi-robot system to achieve provably safe interactions, in the
sense that agents get close without colliding, while meeting
certain spatio-temporal constraints. Under these controllers, it
is possible to obtain theoretical bounds on the mixing limit, the
greatest level of mixing or inter-robot interactions achievable
in a given region. Further, tracking controllers that optimally
follow desired geometries are provided which can be mapped
to classes of nonholonomic dynamical systems. The concepts
presented here are validated on an actual multi-robot system
executing a interaction patterns with a desired mixing level.
Not found in this paper is the question of how to select the
interaction pattern that should be executed by these controllers.
This question addresses the specification level of our symbolic
approach. This discussion will be presented in a different paper
where motion tasks are encoded in linear temporal logic,
allowing for syntactically rich specifications at a high level,
such as “agents 1 and 3 need to collected data before arriving
at a goal location,” which are then translated into a mixing
strategy encoded as a braid.
ACKNOWLEDGMENT
This work is partially supported under a grant from the Air
Force Office of Scientific Research.
The authors would like to thank the members of the Georgia
Robotics and Intelligent Systems Laboratory (GRITS Lab) for
the many fruitful discussions.
R EFERENCES
[1] T. Balch, “The impact of diversity on performance in multi-robot
foraging,” in Proceedings of the third annual conference on Autonomous
Agents. ACM, 1999, pp. 92–99.
[2] K. Sugawara, T. Kazama, and T. Watanabe, “Foraging behavior of
interacting robots with virtual pheromone,” in Intelligent Robots and
Systems, 2004. (IROS 2004). Proceedings. 2004 IEEE/RSJ International
Conference on, vol. 3, Sept 2004, pp. 3074–3079 vol.3.
[3] M. Bonert, L. Shu, and B. Benhabib, “Motion planning for multirobot assembly systems,” International Journal of Computer Integrated
Manufacturing, vol. 13, no. 4, pp. 301–310, 2000.
[4] W. Nguyen and J. K. Mills, “Multi-robot control for flexible fixtureless
assembly of flexible sheet metal auto body parts,” in Robotics and
Automation, 1996. Proceedings., 1996 IEEE International Conference
on, vol. 3. IEEE, 1996, pp. 2340–2345.
[5] R. Groß, M. Bonani, F. Mondada, and M. Dorigo, “Autonomous selfassembly in swarm-bots,” Robotics, IEEE Transactions on, vol. 22, no. 6,
pp. 1115–1130, 2006.
[6] E. Klavins, “Programmable self-assembly,” Control Systems, IEEE,
vol. 27, no. 4, pp. 43–56, 2007.
[7] S. M. Pedersen, S. Fountas, H. Have, and B. Blackmore, “Agricultural
robotssystem analysis and economic feasibility,” Precision agriculture,
vol. 7, no. 4, pp. 295–308, 2006.
[8] N. Noguchi, J. Will, J. Reid, and Q. Zhang, “Development of a master–
slave robot system for farm operations,” Computers and Electronics in
agriculture, vol. 44, no. 1, pp. 1–19, 2004.
[9] R. C. Shah, S. Roy, S. Jain, and W. Brunette, “Data mules: Modeling
and analysis of a three-tier architecture for sparse sensor networks,” Ad
Hoc Networks, vol. 1, no. 2, pp. 215–233, 2003.
[10] I. Vasilescu, K. Kotay, D. Rus, M. Dunbabin, and P. Corke, “Data
collection, storage, and retrieval with an underwater sensor network,” in
Proceedings of the 3rd international conference on Embedded networked
sensor systems. ACM, 2005, pp. 154–165.
[11] W. Zhao, M. Ammar, and E. Zegura, “Controlling the mobility of
multiple data transport ferries in a delay-tolerant network,” in INFOCOM 2005. 24th annual joint conference of the IEEE computer and
communications societies. Proceedings IEEE, vol. 2. IEEE, 2005, pp.
1407–1418.
[12] A. Howard, “Multi-robot simultaneous localization and mapping using
particle filters,” The International Journal of Robotics Research, vol. 25,
no. 12, pp. 1243–1256, 2006.
[13] N. E. Özkucur and H. L. Akın, “Cooperative multi-robot map merging
using fast-slam,” in RoboCup 2009: Robot Soccer World Cup XIII.
Springer, 2010, pp. 449–460.
[14] R. M. Murray, “Recent research in cooperative control of multivehicle
systems,” Journal of Dynamic Systems, Measurement, and Control, vol.
129, no. 5, pp. 571–583, 2007.
[15] X. C. Ding, A. R. Rahmani, and M. Egerstedt, “Multi-uav convoy
protection: an optimal approach to path planning and coordination,”
Robotics, IEEE Transactions on, vol. 26, no. 2, pp. 256–268, 2010.
[16] R. Hall, “Efficient spiral search in bounded spaces,” IEEE Transactions
on Pattern Analysis and Machine Intelligence, vol. 4, no. 2, pp. 208–215,
March 1982.
[17] W. Huang, “Optimal line-sweep-based decompositions for coverage
algorithms,” in IEEE International Conference on Robotics and Automation, 2001.
[18] R. Hashemi, L. Jin, G. Anderson, E. Wilson, and M. Clark, “A
comparison of searchpatterns for cooperative robots operating in remote
environment,” in International Conference on Information Technology:
Coding and Computing, 2001.
[19] J. Kim, F. Zhang, and M. Egerstedt, “Simultaneous cooperative exploration and networking based on Voronoi diagrams,” in IFAC Workshop
on Networked Systems, October 2009.
[20] B. Tovar, L. Freda, and S. M. LaValle, “Using a robot to learn geometric
information from permutations of landmarks,” in Contemporary Mathematics. American Mathematical Society, 2007, vol. 438, pp. 33–45.
[21] E. Artin, “Theory of braids,” Annals of Mathematics, vol. 48, no. 1, pp.
101–126, 1947.
[22] J. S. Birman, Braids, links, and mapping class groups.
Princeton
University Press, 1975, no. 82.
[23] C. Belta, A. Bicchi, M. Egerstedt, E. Frazzoli, E. Klavins, and G. J. Pappas, “Symbolic planning and control of robot motion,” IEEE Robotics
and Automation Magazine, vol. 14, no. 1, pp. 61–70, March 2007.
[24] M. Egerstedt, X. Hu, and A. Stotsky, “Control of mobile platforms using
a virtual vehicle approach,” Automatic Control, IEEE Transactions on,
vol. 46, no. 11, pp. 1777–1782, 2001.
[25] T. Lozano-Perez, “Spatial planning: A configuration space approach,”
Computers, IEEE Transactions on, vol. C-32, no. 2, pp. 108–120, Feb
1983.
[26] R. Ghrist, “Configuration spaces and braid groups on graphs in robotics,”
AMS IP STUDIES IN ADVANCED MATHEMATICS, vol. 24, pp. 29–40,
2001.
[27] V. Kurlin et al., “Computing braid groups of graphs with applications to
robot motion planning,” Homology, Homotopy and Applications, vol. 14,
no. 1, pp. 159–180, 2012.
[28] Y. Diaz-Mercado and M. Egerstedt, “Multi-robot mixing using braids,”
in Decision and Control (CDC), 2013 IEEE 52nd Annual Conference
on, Dec 2013, pp. 2001–2005.
[29] ——, “Multi-robot mixing of nonholonomic mobile robots,” in Control
Applications (CCA), 2014 IEEE Conference on, Oct 2014, pp. 524–529.
[30] A. Criminisi, I. Reid, and A. Zisserman, “A plane measuring device,”
Image and Vision Computing, vol. 17, no. 8, pp. 625–634, 1999.
[31] A. E. Bryson, Applied optimal control: optimization, estimation and
control. CRC Press, 1975.
[32] A. Locatelli, Optimal control: An introduction. Springer Science &
Business Media, 2001.
| 3cs.SY
|
Submitted to ‘Chinese Physics C’
A user-friendly nano-CT image alignment and 3D reconstruction platform based
on LabVIEW
WANG Sheng-Hao(王圣浩)1, ZHANG Kai(张凯)2;1), WANG Zhi-Li(王志立)1, GAO Kun(高
昆)1, WU Zhao(吴朝)1, ZHU Pei-Ping(朱佩平)2, WU Zi-Yu(吴自玉)1,2;2)
1
National Synchrotron Radiation Laboratory, University of Science and Technology of China,
Hefei 230027, China
2
Institute of High Energy Physics, Chinese Academy of Sciences, Beijing 100049, China
Abstract: X-ray computed tomography at the nanometer scale (nano-CT) offers a wide range of
applications in scientific and industrial areas. Here we describe a reliable, user-friendly and fast
software package based on LabVIEW that may allow to perform all procedures after the
acquisition of raw projection images in order to obtain the inner structure of the investigated
sample. A suitable image alignment process to address misalignment problems among image
series due to mechanical manufacturing errors, thermal expansion and other external factors has
been considered together with a novel fast parallel beam 3D reconstruction procedure, developed
ad hoc to perform the tomographic reconstruction. Remarkably improved reconstruction results
obtained at the Beijing Synchrotron Radiation Facility after the image calibration confirmed the
fundamental role of this image alignment procedure that minimizes unwanted blurs and additional
streaking artifacts always present in reconstructed slices. Moreover, this nano-CT image alignment
and its associated 3D reconstruction procedure fully based on LabVIEW routines, significantly
reduce the data post-processing cycle, thus making faster and easier the activity of the users during
experimental runs.
Key words: nano-CT, image alignment, 3D reconstruction, LabVIEW
PACS:
1.
07.05.Hd, 87.57.nf, 87.57.Q-
Introduction
As a unique non-destructive high resolution visualization tool, X-ray computed tomography
at nanometer resolution (nano-CT) technique can be successfully applied in biomedical imaging,
materials science, inspection of integrated circuits and many other scientific researches and
industrial processes [1-3].
A full-field transmission X-ray microscope (TXM) with nanometer spatial resolution, named
nano-CT, has been designed and assembled at the Beijing Synchrotron Radiation Facility (BSRF),
a first-generation synchrotron radiation facility operating at 2.5 GeV. It operates continuously
from 5 to 12 keV with fluorescence mapping capability and an achieved spatial resolution better
than 30 nm [4]. In practical experiments with this nano-CT equipment, the sample stage rotates
discontinuously for well defined angular intervals to obtain projection images at different angles,
because it takes about 30 seconds for the CCD camera to capture a single projection image and 1
second is needed to transmit the data to the computer. During the rotation, errors in the mechanical
manufacture and in the assembly would cause jittering of the rotation axis of the sample stage.
* Supported by the Major State Basic Research Development Program (2012CB825800), the Science Fund
for Creative Research Groups (11321503), the Knowledge Innovation Program of the Chinese Academy of
Sciences (KJCX2-YW-N42), the National Natural Science Foundation of China (NSFC 11179004, 10979055,
11205189, 11205157) and the Fundamental Research Funds for the Central Universities (WK2310000021).
1) E-mail: [email protected]
2) E-mail: [email protected]
1/9
Also thermal expansion due to temperature variation during an experiment and external
environmental changes may impose a slight influence on the system. When dealing with micro-CT
and other larger scale CT equipment, thanks to high precision micro-mechanics, these
contributions are negligible. However, for the spatial resolution demands of nano-CT applications,
these contributions have to be taken into account seriously, as they provoke misalignments among
the projection image series, i.e. the transversal shift of the axis, axial vibrations of the sample
stage, and skew phenomena. If we directly perform 3D reconstruction without an accurate
calibration upon the raw dataset, all the misalignment contributions may generate blurring and
streaking artifacts which cause information loss and, probably, faked structures in the
reconstructed slices of the sample.
LabVIEW is a graphical programming language with a powerful function library and an
easy-to-use graphic user interface (GUI) design framework. The availability of drivers for a very
large number of hardware components, high-efficiency debugging functions and many other
remarkable features make it an ideal software development framework for instrument-oriented
programming. LabVIEW has been already used in X-ray imaging for large scale CT tomographic
reconstruction software development [5], X-ray flat-panel detector driver design [6] and CT
imaging system software building [7-9] by different teams. In addition, the LabVIEW’s Vision
Development Module (VDM) provides an excellent tool to perform digital image processing
[10,11], the utilization of the VDM reduces the complexity and shortens the development cycle.
The newly released LabVIEW Biomedical Workbench Toolkit (BWT) includes also a 3D image
reconstruction module, which could reconstruct a 3D model from a set of 2D image slices. It
provides users with a fast and reliable visualization technique for preliminary analysis and
qualitative evaluations. The technique is widely used in many areas from medicine to physics and
engineering, e.g., for MRI slices reconstruction. However, the BWT module cannot perform a 3D
reconstruction from projection images, which prevents its use in X-ray tomographic
reconstructions.
Next we will describe in detail a LabVIEW based user-friendly package designed for a
high-efficiency nano-CT image alignment and a flexible 3D tomographic reconstruction.
2.
Meterials and methods
Fig. 1.
The flow chart of the nano-CT image alignment and 3D reconstruction platform.
In order to fix the aforementioned image misalignment contributions affecting performance
of this nano-CT equipment, a spherical gold particle with a diameter of 0.5 um is mounted on the
sample stage. It is used as the reference point for the image alignment. Because of its strong
2/9
absorption in the 5-12 keV range (i.e., x-ray with 12 keV photon energy has an absorption of 83.1
% through a 0.5 um thick gold plate), the projection image of the gold particle on the detector is a
striking round dark area. Using this reference point present in each projection image, the software
may fix the major misalignments due to the drift of the rotation axis in the plane parallel to the
surface of the detector and of the sample stage’s undulate position in the axial direction, while the
rotation axis skew and other minor misalignments are considered negligible in our layout. The
program flow chart of the nano-CT image calibration and 3D reconstruction is outlined in Fig. 1.
2.1 Selection of the region of interest
Quickly browsing the projection image series with LabVIEW using the image display
indicator, it can be found that the gold particle jitters constantly within a small region. At this
stage users need to choose a region of interest (ROI) including the small region containing the
gold particle, where the latter image process takes place. The ROI will allow a faster and more
accurate automatic location of the reference point in different projection images. A convenient and
precision selection of the ROI proceeds as follows. At first, binary image corresponding to each
raw projection is generated in LabVIEW with the following scheme.
If A ij =A min , set A ij =0 .
If A ij >A min , set A ij =1 .
where A ij is gray value of the pixel i, j in the handled image and A min is the minimum gray
value of the entire image. Here gray value of the pixels at the center of the reference point is
assumed to be the minimum of the whole image. Actually, compared with the other region of the
image, this assumption is reasonable because the central part of the gold particle strongly absorbs
in the x-ray range from 5 to 12 keV. In the second step, the binary arrays of all the projection
image are multiplied one by one to generate an array product, where gray values of the pixels of
the gold particle’s jittering area would be one, while in other parts of the image they are often zero.
The method offers the possibility to easily choose an accurate ROI covering jittering area related
to the gold particle.
2.2 Automatic identification of the reference point
In the choice of the ROI, two automatic methods are used to locate the reference point in each
raw projection image. In the first one we position the reference point by calculating the gray value
barycenter (GVB) of the subdomain after a threshold segmentation [12]. Initially, a sub image
related to each projection image is extracted associated to the chosen ROI and then, by an iterative
algorithm for threshold segmentation, the threshold T of each sub image is obtained and used in
the next image processing step described as follows.
If A ij T , set A ij 255 A ij .
If A ij >T , set A ij 0 .
Assuming that the raw projection images are 8-bit unsigned and the procedure is valid for
each projection image, the GVB can be calculated according to Eq. 1. This set of coordinates is
considered to be the center of the reference point in the selected projection image.
Aij i
X= i
Aij
.
Aij j
Y= j
Aij
(1)
The second automatic identification procedure of the reference point is the circle fitting
method (CFM) based on the image morphology. It fully takes advantage of the LabVIEW’s VDM
module, a customized shape detection function is used to identify a geometrical shape in an image,
and to acquire the position of the reference point.
3/9
2.3 Image alignment
Based on the automatic procedure for the acquisition of the coordinates of the reference point
in each projection image, the image alignment process is carried out as follows. In the horizontal
direction, the symmetry axis of the image is considered the alignment target of the reference point
in all projection images, while in the vertical direction, the reference point in the first projection
image is regarded as the alignment target of all others. And then, via a small shift of each
projection image in both directions and with the aim of aligning the reference point with the target,
the image calibration procedure would be completed. However, if we still perform the image
alignment in the horizontal direction with the above mentioned strategy, when the gold particle is
mounted slightly away from the rotation axis of the sample stage, in the horizontal direction a very
narrow view is generated due to a large scale pixels movement. As a consequence we consider an
alternate method in which the pixel position X(k) in the horizontal direction, described in Eq. 2 is
set as the alignment target of the reference point in the kth projection image:
X(k)=
D
(k-1)
.
- R cos
2
N
(2)
Here D is the pixel width of the raw projection image and R is the pixel distance between the
reference point and the horizontal symmetry axis in the first projection image. The image
calibration in the horizontal direction can be performed in this way without a large shift of the
pixels.
2.4 Manual adjustment module
Because the characteristics of the reference point may change significantly among different
experiments, and in some circumstances neither GVB nor CFM generate a satisfactory alignment
upon the raw projection image series, although time-consuming, a manual correction module
(MCM) with a sub-pixel positioning accuracy has been introduced. It represents a relevant support
to the automatic protocols.
Finally, after image alignment of the raw nano-CT dataset, useless pixels in each projection
image resulting from the image shift are clipped, both to decrease the consumed memory and to
minimize the time required in the following 3D reconstruction.
2.5 3D reconstruction
The 3D tomographic reconstruction is performed here by the inverse radon transform function
iradon of the MATLAB software. LabVIEW’s MATLAB script node is used to invoke the
MATLAB software script server to execute scripts written in the MATLAB language syntax, and
the reconstructed slice is displayed with three basic orthographic views.
3.
Results
Fig. 2 is the main GUI of the nano-CT image alignment and the 3D reconstruction software
platform. It is made up of a 4 pages tab control representing the four key procedures, i.e. ROI
selection, automatic correction, manual correction and 3D reconstruction, respectively. Fig. 2(a)
shows the GUI of the ROI selection, after the program's execution, the projection images series of
the raw CT dataset is shown successively in the left image display indicator. During the process
users may quickly check the correction of the raw dataset and may also evaluate qualitatively the
jittering area of the gold particle. Then the cumprod (cumulative production) image will appear on
the right image display indicator, where reference points in all raw projection images form a dark
trail. A brown rectangle covering the dark trail could be drawn as showed in Fig. 2(a), then each
projection image with the over imposed rectangle, would be shown in the left image display
indicator, allowing user to check the accuracy of the selected ROI and perform further
adjustments.
The GUI of the automatic correction procedure is showed in Fig. 2(b). Users may choose one
automatic identification method between GVB and CFM. When dealing with each projection
4/9
image, a white circle with a radius matching the gold particle would overlay at the automatically
identified position, to assess the automatic identification performance. The radius of the white
circle can be customized to optimize the contrast respect to the reference point, while in some
cases a multiple threshold segmentation is required.
Fig. 2. (color online) GUI of the nano-CT image alignment and 3D reconstruction platform. Panels (a) (b) (c) (d)
show the procedures of ROI selection, automatic correction, manual correction and 3D reconstruction, respectively.
Fig. 2(c) represents the GUI of the manual calibration module introduced as an important
complementary tool of the automatic procedure. Actually, an image can be manually shifted in
four directions: up, down, left and right by a certain amount of pixels, and shortcut keys
corresponding to the four directions of shift are enabled. A white circle, appearing in the image at
the expected position, acts as the alignment target of the reference point. A sub-pixel positioning
accuracy can be achieved by an optimal matching between the reference point and the target as
shown in Fig. 2(c). The radius of the circle can be customized to achieve a better contrast and also
pixels per step can be changed to make possible and easier the adjustment when the reference
point is located far from the target.
Fig. 2(d) shows the GUI of the 3D reconstructor. In the left part, reconstruction parameters
such as start-stop projective angle, interpolation type (nearest, linear, spline, pchip, cubic, v5cubic)
and filter (Ram-Lak, Shepp-Logan, Cosine, Hamming, Hann, None) can be customized. After the
3D tomographic reconstruction, reconstructed slices appear in the three basic orthographic views,
while other cross-sections of the inspected material can be viewed by dragging the vernier in the
image display indicator or can be set to a definite three-dimensional coordinate.
4.
Performance
4.1 Image alignment efficiency
To evaluate the performance of the nano-CT image alignment package and the 3D
tomographic reconstructor, we present here some experimental results obtained at BSRF. The first
nano-CT raw dataset consists of 181 projections of 8 bit unsigned 1024 1024 images, and the
inspected sample is a standard spherical gold particle with a diameter of 4.5 um. During the image
alignment process, GVB and MCM image alignment methods are independently applied to correct
the nano-CT projection image series. Finally, the linear interpolation and Ram-Lak filter are
5/9
utilized for tomographic reconstruction. Fig. 3(a) shows the slice reconstructed from raw
projection images, while Fig. 3(b) and Fig. 3(c) show the slices reconstructed from corrected
datasets after GVB and MCM image calibrations, respectively. It is clear that a serious blurring
and evident additional fake artifacts exist in the slice reconstructed from the raw dataset as shown
in Fig. 3(a), which have been remarkably improved in the reconstruction image after an image
alignment procedure. And we can also see that almost no obvious visual discrepancies occur
between the slices reconstructed from projection images calibrated with GVB and MCM strategy.
Fig. 3. A cross section of the gold particle reconstructed with (a) the raw projection images, and projection
image series corrected independently with (b) the GVB and (c) the MCM calibration procedures.
Fig. 4. A transverse slice of an animal ovary, data are reconstructed from (a) raw projection images, and
projection images corrected independently with (b) the GVB, (c) the CFM and (d) the MCM calibration
procedures.
The second dataset is made up of 161 projections of 8-bit unsigned 512 512 images and the
material under investigation is the ovary of an insect. During the nano-CT image alignment
procedure, GVB, CFM and MCM image calibration strategy procedures are used independently.
Moreover, a linear interpolation and a Ram-Lak filter are utilized for a tomographic reconstruction.
The reconstructed transverse slices are shown in Fig. 4, while in Fig. 5 a coronal slice through the
vagina is showed. It is very clear as shown in Fig. 4(a) and Fig. 5(a) that evident ring artifacts and
6/9
blurring occur in the slices reconstructed from raw projection images, while almost ideal
reconstruction results are obtained after the image calibration with one of the three available
strategies. A comparison of the quality of the reconstructed slices from data corrected with GVB,
CFM and MCM points out that there are no obvious differences between them, except for the
existing slight blurring in the boundary of the coronal slice through the vagina (see Fig. 5(b)).
Fig. 5. A coronal slice through the vagina of an insect, data are reconstructed from (a) raw projection images, and
projection images corrected independently with (b) the GVB, (c) the CFM and (d) the MCM calibration
procedures.
4.2 Time consumed in image alignment and 3D reconstruction
Table 1.
Time required for nano-CT image alignment and 3D reconstruction of two datasets
Dataset
181
161
Image alignment with an automatic
/manual procedure (s)
3D reconstruction (s)
1024 1024
166.6/851.3
136.2
512 512
106.3/805.8
19.6
Reported in Table 1 are the values of the time consumed to perform the nano-CT image
alignment and the 3D reconstruction of the two raw datasets we discussed above. The evaluation
has been performed on a Core i7 3.40 GHz personal computer with 16 GB RAM and 64-bit
windows 7 operating system. The reconstruction results of these two datasets are 1024 slices of
724 724 pixels and 512 slices of 362 362 pixels, respectively. It takes 19.6 seconds to
7/9
reconstruct a 512 362 362 array, which could be considered a real breakthrough in increasing
the reconstruction speed, excluding other performance factors, compared with the existing
LabVIEW-based 3D Reconstructor [5].
5.
Discussion
Two automatic correction algorithms/methods and a manual adjustment module are described
in this manuscript to perform the nano-CT image calibration before a tomographic reconstruction.
Both automatic methods have advantages and drawbacks in different conditions, GVB is faster,
has less possibility to miss the target, and its positioning deviation is about 3-4 pixels. On the
contrary, CFM is slower, but it may identify more accurately the position, especially when the
reference point is not symmetric. However, CFM more frequently misses the reference point, a
condition that needs a supplementary manual adjustment. By combining GVB and CFM, the
system is much more powerful, in particular, in complex situations. The sub-pixel positioning
accuracy of the MCM calibration strategy makes it the most reliable image correction method if
not for time consideration, because about 15 minutes are necessary to handle a set of CT raw
dataset containing 181 projections. Fortunately, as shown in our experimental results, almost no
observable differences are detected in the slices reconstructed from datasets with an automatic
correction or a manual calibration method. Typically, about 2-4 pixels positioning deviation of the
reference point would exist after the two automatic alignment procedures, which normally is
considerately tolerant for the latter 3D tomographic reconstruction.
The automatic image alignment strategy we described allows labor-saving and fast data
pre-processing before the tomographic reconstruction. while the sub-pixel positioning accuracy
but time-consuming MCM module is supposed to only act as a complementary tool in challenging
situations where neither GVB nor CFM could effectively identify the reference point.
Several reasons are behind the choice of LabVIEW as the program development environment.
The first one is that LabVIEW is a graphical rather than a text based programming language such
as C++ or FORTRAN, which allow a much faster and easier software development. The second
one is that LabVIEW’s VDM module provides users with a unique and ideal digital image
processing tool, which remarkably reduces the project development cycle. Finally, LabVIEW
integrates perfectly with MATLAB, a choice that allows developers to fully take advantage of the
MATLAB's comprehensive library function and its superior computation ability. In our case, by
constructively invoking the inverse radon transform function iradon of MATLAB with the
LabVIEW’s MATLAB script node, complex codes of parallel beam tomographic reconstruction
are avoided. Moreover, the convenient multithreaded programming technique in LabVIEW makes
developers capable to fully exploit the multiple Central Processing Units of their computers.
6.
Conclusion
A LabVIEW based reliable and user-friendly nano-CT image alignment and 3D
reconstruction software package is presented and discussed. We have showed how instability
problems associated to rotation axis of the sample stage can be overcome, minimizing artifacts in
the 3D reconstruction. The nano-CT image alignment and 3D reconstruction software package
optimized for the nano-CT equipment existing at BSRF, significantly shortens the data
post-processing cycle, thus making easier the users activity during runs, and it also has great
potentials for the upgrade of other existing nano-CT instruments.
Acknowledgements
The authors acknowledge A. Marcelli for many fruitful helps and discussions.
References
[1]
[2]
[3]
[4]
V. Alt, D.V. Kögelmaier, K.S. Lips et al. Acta. Biomater, 2011, 7: 3773
P. Schneider, M. Stauber, R. Voide et al. J. Bone. Miner. Res, 2007, 22: 1557
P. Shearing, R. Bradley, J. Gelb et al. Electrochem. Solid. St, 2011, 14: B117
Q. Yuan, K. Zhang, Y. Hong et al. J. Synchrotron Radiat, 2012, 19: 1021
8/9
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
M. Dierick, B. Masschaele and L.Van Hoorebeke. Meas. Sci. Technol, 2004, 15: 1366
K. Bi, Q. Liu, X. Lv et al. J. X-Ray Sci. Technol, 2008, 16: 261
M. Dierick, D. Van Loo, B. Masschaele et al. J. X-ray Sci. Technol, 2010, 18: 451
C. Keleshis, C. Ionita, G. Yadava et al. Proc. SPIE, 2008, 6913: 69135A
J. Wang, P. Zhu, Q. Yuan et al. Radiat. Phys. Chem, 2006, 75: 1986
Y. Yuan, C.M. Li and B.L. Li. Adv. Mater. Res, 2012, 393: 471
H. Zhang, L. Gao, J. Liu et al. Appl. Mech. Mater, 2013, 263: 2515
A. Pérez and R.C.Gonzalez. IEEE. T. Pattern. Anal, 1987, 9: 742
9/9
| 5cs.CE
|
High-Dimensional Adaptive
Function-on-Scalar Regression
arXiv:1610.07507v1 [math.ST] 24 Oct 2016
Zhaohu Fan∗
Matthew Reimherr†‡
Pennsylvania State University
Abstract
Applications of functional data with large numbers of predictors have grown precipitously in
recent years, driven, in part, by rapid advances in genotyping technologies. Given the large numbers of genetic mutations encountered in genetic association studies, statistical methods which
more fully exploit the underlying structure of the data are imperative for maximizing statistical
power. However, there is currently very limited work in functional data with large numbers of
predictors. Tools are presented for simultaneous variable selection and parameter estimation in
a functional linear model with a functional outcome and a large number of scalar predictors;
the technique is called AFSL for Adaptive Function-on-Scalar Lasso. It is demonstrated how
techniques from convex analysis over Hilbert spaces can be used to establish a functional version
of the oracle property for AFSL over any real separable Hilbert space, even when the number
of predictors, I, is exponentially large compared to the sample size, N . AFSL is illustrated
via a simulation study and data from the Childhood Asthma Management Program, CAMP,
selecting those genetic mutations which are important for lung growth.
Keywords: Variable selection; Functional regression; Oracle property
1
Introduction
Many scientific areas are faced with the challenge of extracting information from increasingly large,
complex, and highly structured data sets. A great deal of modern statistical work focuses on
developing tools for handling such data. Networks, high dimensional data, images, functions,
∗
†
‡
Departments of Industrial Engineering and Statistics, Penn State University, University Park, PA, 16802
Department of Statistics, Penn State University, University Park, PA, 16802
Corresponding Author: [email protected]
1
surfaces, or shapes, all present data structures which are not well handled under a traditional
univariate or multivariate statistical paradigm. In this paper we present a new methodology, which
we call Adaptive Function-on-Scalar Lasso, AFSL, for analyzing highly complex functional outcomes
alongside large numbers of scalar predictors. Such data is becoming increasingly common due to
the prevalence of inexpensive genotyping technologies. Genome-wide association studies, GWAS,
examine hundreds of thousands or millions of genetic markers, attempting to find those mutations
which are associated with some outcome or phenotype. Many phenotypes of interest are now
complex outcomes, such as longitudinal measurements or biomedical images.
The functional linear model, FLM, is one of the primary modeling tools in FDA. There, one
assumes that the outcome is linearly related to some set of predictors. FLM are often categorized
by whether the outcome, the predictor, or both is functional (Reiss et al., 2010). While the
literature on FLM is now vast, Morris (2015) outlines most of the key work on low dimensional
FLM. However, to date, relatively little has been done when one has a large number of predictors
relative to the sample size, and of the work that does exist, nearly all of it is for scalar-on-function
FLM, the opposite setting we consider (Matsui and Konishi, 2011; Lian, 2013; Gertheiss et al., 2013;
Fan et al., 2015). For the function-on-scalar setting, we are aware of only two other works. Chen
et al. (2016) consider a basis expansion approach with an MCP style penalty and fixed number of
covariates. They also use a pre-whitening technique to exploit the within function dependence of
the outcomes. Asymptotic theory is developed for a fixed number of predictors and basis functions.
Barber et al. (2016) developed the Function-on-Scalar LASSO, FSL, which allows for an exponential
number of predictors relative to the sample size and establishes optimal convergence rates of the
estimates. In particular, it was shown that FSL achieves the same rates of convergence as in the
scalar case. However, as in the scalar case, this approach leads to estimates with a non–negligible
asymptotic bias due to the nature of the penalty.
To alleviate the bias problem inherent in FSL, we propose here an adaptive version, AFSL.
In addition to providing a novel statistical methodology, we develop a new theoretical framework
needed to establish its asymptotic properties. In particular, functional subgradients and tools from
convex analysis over Hilbert spaces are needed. In contrast, theory for FSL is built entirely on
basic inequalities and concentration inequalities, no theory for subgradients was required. The
contributions of this paper are thus as follows (1) we define a new variable selection and estimation
tool, AFSL, which alleviates the bias problems of FSL (2) we demonstrate how several tools and
techniques from convex analysis can be used for functional data problems and (3) we define a
2
functional version of the oracle property and show that AFSL achieves it. Additionally, we also go
a step beyond the traditional oracle property which states that the estimates recover the correct
support and are asymptotically normal, by showing that the oracle estimate and the AFSL estimate
are actually asymptotically equivalent.
The remainder of the paper is organized as follows. In Section 2, we provide the necessary
background material. In section 3 we outline the AFSL framework and in section 4 we establish
the oracle property. A simulation study and an application to the Childhood Asthma Management
Program, CAMP, is given in Section 5. Concluding remarks are given in Section 6. All theory is
provided in the Appendix.
2
Background
Let H denote a real separable Hilbert space, h·, ·i its inner product, and k · k the inner product
norm. In a function-on-scalar linear model we have that
(1)
Yn =
I
X
Xni βi? + εn = Xn> β ? + εn
1 ≤ n ≤ N,
i=1
where Yn ∈ H is a functional outcome, βi? ∈ H is a functional regression parameter, εn ∈ H is
a functional error, and Xni ∈ R is a scalar predictor. Throughout, we will use a ? to denote the
true data generating parameter so as to distinguish from β, which will usually represent a dummy
variable or the argument of a function. The most commonly encountered space for H is L2 [0, 1], i.e.,
the outcome is a function of time, though other spaces are used as well including spatial domains or
product spaces (for functional panels or multivariate functional data). If one wants to incorporate
smoothness assumptions of the data then one can work with Sobolev spaces or Reproducing Kernel
Hilbert Spaces, RKHS. We provide a few examples here to help emphasize the functional nature of
the data and highlight the wide impact of our theory and methods.
Example 1. Let H = L2 (D), where D is a compact subset of Rd . Then we write the model
Yn (s) =
I
X
Xni βi? (s) + εn (s)
s ∈ D,
i=1
and the norm is written as
2
Z
x(s)2 ds.
kxk =
D
3
Example 2. Let H = H 1,2 [0, 1], i.e., the Sobolev space of real valued functions over the unit
interval with one square integrable derivative. Then we have
Yn (t) =
I
X
Xni βi? (t) + εn (t)
t ∈ [0, 1],
i=1
with the norm given by
2
Z
1
Z
2
kxk =
x(t) dt +
0
1
x0 (t)2 dt.
0
Example 3. Let H, be an RKHS of functions over the unit interval with kernel operator K. Then
we have
Yn (t) =
I
X
Xni βi? (t) + εn (t),
i=1
but now the norm becomes
kxk2 = hK −1 x, xi,
where h·, ·i is the L2 [0, 1] inner product and K −1 is the inverse (operator) of K.
The most common method for estimating β is least squares, i.e. minimizing
N
L0 (β) =
1X
kYn − Xn> βk2 .
2
n=1
However, when I > N the problem is ill-posed and a unique minimizer does not exist. To address
this issue, Barber et al. (2016) introduced FSL, which is defined as the minimizer of
Lλ (β) =
The term
PI
i=1 kβi k
N
I
n=1
i=1
X
1X
kYn − Xn> βk2 + λ
kβi k.
2
can be viewed as a type of `1 norm on the product space HI which induces a
sparse estimate. As the name implies, this approach is an extension of the classic Lasso (Tibshirani,
1996) to functional outcomes. It was shown that under appropriate conditions, the FSL estimate
achieves optimal convergence rates and that these rates are the same regardless of the Hilbert space.
In other words, the converge rate for Y ∈ R is the same as for Y ∈ L2 [0, 1].
As in the scalar case, FSL does not achieve the oracle property. This is due to the nonnegligible bias of the estimates. In the next section we define an adaptive version of FSL and
show that it achieves a type of oracle property. Interestingly, we will use a different formulation of
the oracle property than the one commonly used in the scalar setting, namely, that estimates are
4
asymptotically normal. In the functional setting, some estimates are not asymptotically normal
even in the low dimensional setting (Cardot et al., 2007), and thus removing this connection is
useful.
3
Adaptive Function-on-Scalar Regression
We now define an adaptive version of FSL we call AFSL. As with FSL, this can be viewed as
an extension of the adaptive lasso (Zou, 2006) to functional outcomes. To produce an adaptive
estimate, we introduce data driven weights into the target function. Namely, we denote by β̂ the
AFSL estimate, which is the minimizer of
(2)
LA
λ (β) =
N
I
n=1
i=1
X
1X
kYn − Xn> βk2 + λ
w̃i kβi k.
2
These weights can be chosen in a variety of ways, but the approach we take here is to run FSL to
obtain preliminary estimates β̃i . We then set w̃i = kβ̃i k−1 so that small estimates are shrunk to
zero and large estimates are not shrunk as substantially. It is possible that w̃i = ∞ in which case
the predictor is dropped. This has the advantage that running AFSL after FSL is usually very fast
as most of the variables have already been screened. Another option, which is used in the scalar
case by (Huang et al., 2008), is to do marginal ordinary least squares regression to obtain each
weight. However, this does not reduce the dimension of the problem in anyway and requires an
additional theoretical justification, thus we do not pursue it further here.
Barber et al. (2016) developed optimal asymptotic convergence rates of FSL by combining
a basic inequality with concentration inequalities and a restricted eigenvalue condition. Here we
aim to show that AFSL achieves the oracle property which is a much stronger property. It is
therefore useful to switch to a different technique which involves functional subgradients. Given
that many readers may not be familiar with such concepts for general Hilbert spaces, we provide a
brief reminder of the core concepts. We refer interested readers to Boyd and Vandenberghe (2004);
Bauschke and Combettes (2011); Shor (2012) for more details on subgradients and convex analysis,
and Barbu and Precupanu (2012) specifically for convex analysis over Hilbert spaces.
Definition 1. Let f : H → R be a convex functional. An element h ∈ H is called a subgradient of
f at x if for all y ∈ H we have
f (y) − f (x) ≥ hh, y − xi.
5
We emphasize that in definition 1 the points x, y, and h are elements of an arbitrary Hilbert
space. If H = R, then they are scalars, if H = Rd then they are vectors, and if H = L2 [0, 1] or any
of the function spaces from 2, then they are functions. Note that subgradients are unique if the
functional is differentiable, but they also exist for nondifferentiable functions in which case they
need not be unique.
Definition 2. The set of subgradients of f at x is called the subdifferential and denoted ∂f (x)
Thus, one has a generalized notion of derivative for convex functionals. One can immediately
see from Defintion 1 that if 0 is in the subdifferential at a point x, then that point is a minimizer.
Both the FSL and AFSL are convex, we thus have the following property.
Theorem 1. The subdifferential of LA
λ (β) with respect to βi is given by
N
βi kβi k−1
X
∂ A
T
L (βi ) = −
Xni (Yn − Xn β) + λw̃i
∂βi λ
{h ∈ H : khk ≤ 1}
n=1
if kβi k =
6 0
.
if kβi k = 0
We again emphasize that βi is an element of H (i.e. is a function), thus the subdifferential in
Theorem 1 is a set of objects from H. Using Theorem 1 we can’t, in general, obtain an explicit
expression for β̂, but we can characterize the solution enough to establish the oracle property. In
particular, β̂ must satisfy the following corollary.
Corollary 1. If β̂ satisfies
−
N
X
Xni (Yn − Xn> β̂) + λw̃i β̂i kβ̂i k−1 = 0,
for β̂i 6= 0
n=1
and
N
X
Xni (Yn − Xn> β̂) < λw̃i
for β̂i = 0,
n=1
then β̂ is a minimizer of LA (β).
Using these properties combined with appropriate assumptions on the errors and the design
matrix, we establish the oracle property in the next section. However, to help better understand
the structure of AFSL, we consider the orthogonal design case as an illustrative example. In that
case, both FSL and AFSL produce explicit estimates.
Example 4 (Orthogonal Design). For this example only assume that I < N is fixed and that
N −1 X> X = II×I . From Corollary 1 if β̂i is not zero then it must satisfy
N
X
Xin Yn − N β̂i = λw̃i β̂i kβ̂i k−1 .
n=1
6
Since the design is orthogonal, β̂iLS =
1
N
PN
n=1 Xin Yn
is the least squares estimator. So we have
1
β̂i =
1+
λN −1 w̃
i kβ̂i
β̂iLS .
−1
k
We can get the norm of β̂i by considering
kβ̂iLS k = (1 + λN −1 w̃i kβ̂i k−1 )kβ̂i k = kβ̂i k + λN −1 w̃i ,
which implies that
kβ̂i k = kβ̂iLS k − λN −1 w̃i .
After a little algebra, the AFSL estimate can be expressed as
!
λw̃i
β̂iLS .
β̂i = 1 −
N kβ̂iLS k
Turning to the case where β̂i = 0 we have that
kXi> (Y − Xβ̂)k < λw̃i ,
or equivalently
kβ̂iLS k < λN −1 w̃i .
So, as in the scalar case, the FSL and AFSL procedures can be viewed as a soft thresholding applied
to the least squares estimator:
β̂i =
1−
λw̃i
N kβ̂iLS k
!+
β̂iLS .
At this point, we can now establish conditions for the oracle property to hold. Notice that for
βi = 0, we want λw̃i N −1 kβ̃i k−1 to grow quickly, while for βi 6= 0, we want it to shrink quickly (so
as to reduce the bias as much as possible). If βi = 0 then β̂iLS = OP (N −1/2 ) and we can assume
the same for w̃i−1 (e.g. this would be true if we used the least squares estimates to compute w̃i ).
We have
λw̃i
N kβ̂iLS k
= λOP (1),
so as long as λ → ∞, then β̂i will be shrunk to zero. Conversely, if βi 6= 0, then
λŵi
N kβ̂iLS k
= λN −1 OP (1),
so as long as λ/N → 0, then the estimate will be asymptotically unbiased. Under these conditions,
AFSL has variable selection consistency, but we need a bit more control of λ to ensure that the
estimate is asymptotically equivalent to the oracle. When βi 6= 0 we have
N 1/2 kβ̂i − β̂iLS k = N −1/2 λOP (1).
7
So, if λN −1/2 → 0 then the LS estimator for the nonzero coordinates (this is usually called the
oracle estimator) is asymptotically equivalent to the AFSL estimator and thus is asymptotically
normal. This is because, for βi 6= 0
N 1/2 (β̂i − βi? ) = N 1/2 (β̂iLS − βi? ) + N 1/2 (β̂i − β̂iLS ) = N 1/2 (β̂iLS − βi? ) + oP (1).
Therefore, we can conclude that for fixed I and an orthogonal design, AFSL achieves the oracle
property as long as 1 λ N 1/2 . Later on, we will see that this main dynamic still holds. With
a “looser” control of λ, we can get selection consistency, but to ensure AFSL is equivalent to the
oracle estimate, we need slightly tighter control.
4
Oracle Property
We begin by making the FLM assumption more explicit.
Assumption 1. Let Y1 , ..., YN be independent random elements of a real separable Hilbert space,
H, satisfying the functional linear model:
(3)
Yn =
I
X
Xni βi? + n .
i=1
Assume the N×I design matrix X = {Xni } is deterministic and has standardized columns, and that
n are i.i.d. Gaussian random elements of H with mean function 0 and covariance operator C.
As is common in high dimensional settings, the Gaussian assumption is not crucial and our arguments can be readily generalized to errors with thicker tails. We utilize the Gaussian assumption
here to make the rates in our subsequent assumptions simpler and more interpretable. We emphasize that the i.i.d. assumption means that response curves from different subjects are independent,
but observations from the same subject are allowed to be dependent, and no restriction is placed
on this dependence.
Define the true support as S = {i ∈ 1, . . . , I : βi? 6= 0} and I0 = |S|, i.e. the cardinality of S.
For notational simplicity, we will assume that X is ordered such that it can be partitioned into
X = (X1
X2 ),
where X1 ∈ RN ×I0 are covariates with nonzero coefficients and X2 ∈ RN ×(I−I0 ) are the covariates
with zero coefficients.
8
Next we define a functional version of the oracle property. For scalar outcomes, this is usually
phrased as having asymptotically normal estimates and estimating the true support with probability tending to one (Fan and Li, 2001). For functional data, we will divorce the normality from
the oracle property because there are examples of functional estimates which are not asymptotically normal, e.g. scalar-on-function regression (Cardot et al., 2007). Instead, we make the more
direct assumption that the oracle property means that an estimator asymptotically has the correct
support, and is asymptotically equivalent to the oracle estimator. The oracle estimate in this case
−1 >
is defined as β̂O = {(X1 X>
1 ) X1 Y, 0}, that is, all estimates from the second group are set to
zero and not used to form estimates for the first group. We now introduce the functional oracle
property.
Definition 3. We say that an estimate β̂ of β satisfies the functional oracle property if
s
1. P (β̂ = β) → 1 and
2.
√
N (β̂ − β̂O ) = oP (1).
s
Here = means that the two have the same support, i.e., same nonzero coordinates. We now
introduce the remaining more technical assumptions. All of these conditions are commonly found
in the literature and relate the orders of various terms.
Assumption 2. We assume that following four conditions hold:
1. Minimum Signal: Let bN = mini=1,...,I0 kβi k. Then we assume that
b2N
I02 log(I)
.
N
2. Tuning Parameter: We assume the tuning parameter satisfies
1/2
I0
log(I) λ
N b2N
1/2
.
I0
3. Design Matrix: Let Σ̂1 = N −1 X>
1 X1 , then we assume that the smallest and largest eigenvalues of Σ̂1 satisfy ,
1
≤ σmin Σ̂1 ≤ σmax Σ̂1 ≤ τ1
τ1
where τ1 > 0 is fixed scalar.
9
4. Irrepresentable Condition: Let Σ̂21 = N −1 X>
2 X1 , then we assume that
kΣ̂21 Σ̂−1
11 kop ≤ φ < 1,
where k · kop is the operator norm and φ is a fixed scalar.
Each of the above assumptions is common in the high dimensional regression literature. The
minimum signal condition allows the smallest βi to vary with the sample size, but it cannot get too
small. The tuning parameter assumption states a familiar rate, namely, the tuning parameter must
grow, but not too quickly. The eigenvalue assumption requires that the true predictors are well
behaved and not extremely highly correlated. The last assumption, the Irrepresentable condition,
essentially says that the true and null predictors cannot be too correlated. For the scalar oracle
property it has been shown that this is essentially necessary and cannot be weakened (Zhao and
Yu, 2006).
The above assumptions, as we will see, will guarantee that the AFSL estimate is variable
selection consistent. However, to control the asymptotic bias, and thus ensure the AFSL estimate
is asymptotically equivalent to the oracle, we also need the following.
Assumption 3. The tuning parameter satisfies
√
λ
N bN
1/2
.
I0
When we divide the right hand side of Assumption 2.2 by the left hand side , and assuming
that bN is bounded, we see the familiar assumption that
N
→ ∞.
I0 log(I)
However, to control the bias enough for the full oracle property to hold, we also need Assumption
3. If we take the right hand side of Assumption 3 and divide by the left hand side of Assumption
2.2 we arrive at another familiar form
√
N
→ ∞.
I0 log(I)
These rates are well known and necessary for scalar linear models (Bühlmann and van de Geer, 2011;
Cai and Guo, 2015). What is especially interesting here, is that we obtain the exact same rates
for any separable Hilbert space. In other words, the outcome can come from any separable Hilbert
space, and the oracle property holds with exactly the same rates as in the scalar case.
We are now in a position to state our main result.
10
Theorem 2. (Functional Oracle property) Let β̃ be the FSL estimator and let β̂ be a minimizer of
(2) with weights w̃i = kβ̃i k−1 . We then have the following.
1. If Assumptions 1 and 2 hold then
s
P (β̂ = β) → 1.
2. If in addition Assumption 3 holds then
√
n(β̂ − β̃O ) = oP (1).
As a corollary, since the oracle estimate for function-on-scalar regression is asymptotically normal, we have the following.
Corollary 2. If Theorem 2 holds, then for any nonzero sequence xN ∈ RI0 we have that
1
I0
X
q
−1
x>
N Σ̂1 xN
i=1
D
xN i (β̂i − βi? ) → N (0, C),
where N (0, C) is an H valued Gaussian process with zero mean and covariance operator C.
5
Numerical Studies
In this section we perform a simulation study to illustrate the performance gain of AFSL over
FSL. We also apply our methodology to data from Childhood Asthma Management project, a
longitudinal genetic association study. However, before exploring these examples, we discuss how
the methodology is implemented in both cases. The data are first preprocessed into functional
units via penalized (on the second derivative) bsplines basis expansions using the FDA package in
matlab. We use 100 cubic bsplines with equally spaced knots, and the smoothing parameter is
chosen via generalized cross-validation. This processing is now standard in the FDA literature
and we reference Horváth and Kokoszka (2012) for more details. After the functional units are
constructed, we rotate the data to the functional principal component, FPC, basis. This serves the
dual purpose of working with an orthonormal basis which is also tailored to the data. Practically,
one can work with a relatively small FPC basis compared to bsplines. We emphasize that we do not
use this basis for the purposes of serious dimension reduction, so we take the number of FPCs to
explain ≥ 99% of the variance so that the FPCA and bsplines approximations are nearly identical.
Next, we rephrase the problem as a group lasso problem, as in Barber et al. (2016), and use
11
alternating direction method of multipliers, ADMM, to find the solution (Boyd et al., 2011). This
approach can be used for both FSL and AFSL by using the weights to scale the predictors instead
of the β; one performs a change of variables with αi = w̃i βi , uses the discussed method for finding
the α̂i , and then changes variables back to the β̂i . We choose the tuning parameter via BIC, though
in the data application we also compare to the extended BIC as was used in Barber et al. (2016). A
complete matlab function and example are available for download from the corresponding author’s
website, www.personal.psu.edu/~mlr36, which can be readily used by other researchers. Access
to the discussed CAMP data is free, but one must submit an application through NIH’s dbGaP.
Those interested in the matlab code applied to the CAMP data should contact the corresponding
author directly.
5.1
Empirical Study
We generate random samples of size N = 100 and 500. For each sample, we generate the nonzero
βi from from a mean zero Gaussian process with covariance given in (4). Note that this is a
Matérn process with smoothness parameter ν = 2.5, which means the β are twice differentiable.
We also take the point-wise variance to be σ 2 = 1 and the range (i.e. how quickly the within
curve dependence decays) to be τ1 = 0.25. The errors, εn , are generated in nearly the same way,
but we change ν to 1.5, meaning that the errors are rougher than betas (as would be natural in
practice) possessing only one derivative. The corresponding covariance function is given in (5). We
also consider several different range parameters, τ2 = 0.01, 0.25, 1, 10, to examine the effect of the
within subject correlation.
√
(4)
C(t, s) = σ 2
5d 5d2
1+
+ 2
τ1
3τ1
√
(5)
C(t, s) = σ 2
3d
1+
τ2
√
!
exp
!
5d
−
,
τ1
d = |t − s|,
ν = 52 .
√
!
exp
!
3d
−
,
τ2
d = |t − s|,
ν = 32 .
Each curve is sampled at 50 evenly space points between 0 and 1. The total number of predictors
I is equal to 500 while the total number of true predictors, I0 , is set to 10. The predictors are Xni are
taken to be standard normal with a time series style correlation structure: Cov(Xni , Xnj ) = ρ|i−j| .
12
Figure 1: Plots of error terms, from left to right, with range parameters τ2 =0.01, 0.25, 1 and 10,
on the top row and the corresponding outcome Y on the bottom row.
As an illustration, in Figure 1 we plot four examples of error terms, εn , with four different range
parameters (τ2 =0.01, 0.25, 1 and 10) and the corresponding outcomes, Yn , when ρ = 0. We
consider three different values of ρ to examine the effect of predictor correlation on performance:
ρ = 0, 0.5, 0.99 corresponding to low, medium, and high correlation. We consider 1000 repetitions
of each scenario and in every setting, we choose the tuning parameter λ using BIC as in Barber et
al. (2016). The results are summarized in Table 1, where we report the average number of true
positives, average number of false positives, average root-mean-squared prediction error, and the
average computation time. Each of these values is reported for both FSL and AFSL.
13
True Positives
False Positives
RMSP
Time (seconds)
Parameters
FSL
AFSL
FSL
AFSL
FSL
AFSL
FSL
τ = 0.01; ρ=0; N=100
9.82
9.78
3.34
0.42
0.0129
0.0113
156.9379
165.5910
τ = 0.01; ρ=0; N=500
9.99
9.98
17.92
0.06
0.0051
0.0050
283.1586
311.3364
τ = 0.01; ρ=0.5; N=100
9.91
9.84
3.35
0.41
0.0126
0.0113
172.7202
180.0085
τ = 0.01; ρ=0.5; N=500
9.99
9.96
13.16
0.03
0.0049
0.0051
284.8761
316.0269
τ = 0.01; ρ=0.99; N=100
5.66
5.18
10.63
3.36
0.0182
0.0105
965.0788
859.6725
τ = 0.01; ρ=0.99; N=500
9.82
9.36
16.68
0.02
0.0037
0.0034
1167.7435
1191.7981
τ = 0.25; ρ=0; N=100
8.62
8.43
1.65
0.68
0.0193
0.0128
88.5702
91.2917
τ = 0.25; ρ=0; N=500
9.99
9.87
2.82
0.13
0.0054
0.0051
341.4661
369.8791
τ = 0.25; ρ=0.5; N=100
8.25
8.18
1.82
0.71
0.0202
0.0125
98.3453
98.1497
τ = 0.25; ρ=0.5; N=500
9.99
9.89
2.52
0.16
0.0054
0.0051
367.4788
395.9227
τ = 0.25; ρ=0.99; N=100
1.18
1.05
3.27
2.14
0.0328
0.0092
996.5986
476.5383
τ = 0.25; ρ=0.99; N=500
8.82
8.08
15.98
2.82
0.0045
0.0040
1291.9321
1310.9210
τ = 1; ρ=0; N=100
7.02
6.91
1.13
0.52
0.0239
0.0127
65.5368
62.4991
τ = 1; ρ=0; N=500
9.96
9.86
2.09
0.47
0.0056
0.0051
305.7731
332.5053
τ = 1; ρ=0.5; N=100
6.34
6.28
1.74
0.46
0.0262
0.0113
70.4651
58.5442
τ = 1; ρ=0.5; N=500
9.99
9.88
2.51
0.62
0.0057
0.0053
308.2203
333.3216
τ = 1; ρ=0.99; N=100
6.08
6.02
1.06
0.58
0.0263
0.0120
62.9913
54.8344
τ = 1; ρ=0.99; N=500
9.98
9.94
1.96
0.34
0.0057
0.0052
297.8822
322.7747
τ = 10; ρ=0; N=100
7.02
6.93
1.31
0.74
0.0239
0.0123
57.8824
55.3399
τ = 10; ρ=0; N=500
9.94
9.83
2.31
0.62
0.0057
0.0051
275.2240
301.7934
τ = 10; ρ=0.5; N=100
6.05
5.94
1.08
0.63
0.0262
0.0126
71.3873
63.0084
τ = 10; ρ=0.5; N=500
9.95
9.82
2.63
0.67
0.0056
0.0050
308.7230
334.8119
τ = 10; ρ=0.99; N=100
1.52
1.26
4.31
2.26
0.0301
0.0107
1031.6216
587.0757
τ = 10; ρ=0.99; N=500
8.38
7.56
17.11
3.84
0.0046
0.0041
1315.9980
1335.4355
AFSL
Table 1: Summary of the true positives, false positives, root-mean-square prediction errors (RMSP),
and computation time (seconds per repetition) for FSL and AFSL. Here N denotes the sample size,
the number of points per curve is 50, τ indicates the strength of the within curve correlation for
the errors, and ρ is the strength of the between predictor correlation.
14
True Positives
False Positives
RMSP
Time (seconds)
Parameters
FSL
AFSL
FSL
AFSL
FSL
AFSL
FSL
τ = 0.01; ρ=0; N=100
9.87
9.82
2.32
0.42
0.0242
0.0208
118.892
124.8117
τ = 0.01; ρ=0; N=500
9.99
9.99
6.98
0.02
0.0092
0.0091
289.4841
321.3352
τ = 0.01; ρ=0.5; N=100
9.85
9.66
2.84
0.55
0.0246
0.0218
129.4541
135.4117
τ = 0.01; ρ=0.5; N=500
9.99
9.96
6.34
0.01
0.0092
0.0092
306.2771
335.4572
τ = 0.01; ρ=0.99; N=100
4.05
3.57
7.71
3.39
0.0368
0.0173
856.7992
667.4319
τ = 0.01; ρ=0.99; N=500
9.73
9.34
17.21
0.33
0.0071
0.0068
1153.8001
1177.1012
τ = 0.25; ρ=0; N=100
8.35
8.11
1.37
0.52
0.0348
0.0235
79.9695
82.5338
τ = 0.25; ρ=0; N=500
9.99
9.93
2.45
0.13
0.0099
0.0094
317.6521
344.4366
τ = 0.25; ρ=0.5; N=100
7.72
7.55
1.59
0.72
0.0376
0.0211
83.9127
80.6943
τ = 0.25; ρ=0.5; N=500
9.99
9.95
2.41
0.13
0.0099
0.0093
284.5948
307.7951
τ = 0.25; ρ=0.99; N=100
1.38
1.21
3.19
1.84
0.0551
0.0149
921.1164
433.9953
τ = 0.25; ρ=0.99; N=500
8.83
8.25
15.83
2.86
0.0082
0.0073
1114.5110
1131.1231
τ = 1; ρ=0; N=100
6.68
6.57
1.13
0.62
0.0424
0.0195
61.2572
56.2244
τ = 1; ρ=0; N=500
9.99
9.91
2.42
0.45
0.0103
0.0094
285.0516
310.0837
τ = 1; ρ=0.5; N=100
6.83
6.68
1.55
0.87
0.0417
0.0199
79.4073
72.5188
τ = 1; ρ=0.5; N=500
9.98
9.91
2.25
0.42
0.0103
0.0095
274.4325
298.9810
τ = 1; ρ=0.99; N=100
1.08
0.92
2.49
1.52
0.0581
0.0145
926.1381
392.3812
τ = 1; ρ=0.99; N=500
8.80
8.06
16.25
3.39
0.0083
0.0074
1129.6421
1145.9041
τ = 10; ρ=0; N=100
6.15
6.04
0.72
0.47
0.0455
0.0196
57.8118
50.4278
τ = 10; ρ=0; N=500
9.94
9.91
2.57
0.62
0.0103
0.0094
248.1699
271.3013
τ = 10; ρ=0.5; N=100
5.61
5.54
0.97
0.57
0.0477
0.0190
64.7757
53.0382
τ = 10; ρ=0.5; N=500
9.97
9.89
2.69
0.58
0.0104
0.0096
257.8873
280.6499
τ = 10; ρ=0.99; N=100
1.12
1.02
2.54
1.41
0.0578
0.0141
988.6308
403.8979
τ = 10; ρ=0.99; N=500
9.01
8.18
15.92
3.42
0.0083
0.0074
1114.6311
1133.2537
AFSL
Table 2: Summary of the true positives, false positives, root-mean-square prediction errors (RMSP),
and computation time (seconds per repetition) for FSL and AFSL. Here N denotes the sample size,
the number of points per curve is 16, τ indicates the strength of the within curve correlation for
the errors, and ρ is the strength of the between predictor correlation.
15
1.5
1
0.5
0
-0.5
-1
-1.5
-2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Figure 2: One example of an estimated beta versus a true one for τ1 =0.25, τ2 = 0.25, N = 500 and
ρ = 0.The true β is plotted using a solid black line, the FSL estimate is plotted using blue squares,
and the AFSL estimate is plotted using a red dashed line.
Turning first to variable selection consistency in Table 1, we see that AFSL cannot beat FSL
in terms of selecting true predictors. This is because AFSL uses the FSL selection as a starting
point. The real gain is in terms of the false positives. For low and moderate correlation, we see
that the AFSL has about 50% lower false positive rate for N = 100, and this increases dramatically
for N = 500 resulting in an over 90% decrease in the false positive rate. This is accomplished
while maintaining nearly exactly the same true positive selection rate. Both methods, as we would
expect, have a noticeably harder time when the correlation between predictors is high, but AFSL
still substantially decreases the number of false positives. Turning to the prediction error (RMSP)
the differences are not quite as dramatic, but AFSL is still out performing FSL. For smaller samples,
the gain is around 10-20%, but this gain reduces for larger N . Lastly, adding AFSL on top of FSL
increases the computation time by less than 10%. We therefore conclude that there is very little
reason not to run AFSL after FSL. To help illustrate the resulting estimates, in Figure 2 we plot
one example of an estimated beta versus a true one for τ1 =0.25, τ2 = 0.25, N = 500 and ρ = 0.
There we can better visualize how AFSL compares to FSL. In general, FSL will pull the estimates
to zero more than AFSL. This results in too much bias for the nonzero betas, and thus worse
statistical performance.
As our last set of simulations, we rerun all of the scenarios in Table 1, but reduce the number of
time points sampled per curve to 16 to better emulate the CAMP data. The results are reported
in Table 2. As we can see, the performance of the two methods is nearly the same as before. This
16
helps justify applying our methodology to the CAMP data, which we will discuss more in the next
section.
5.2
Data Example
In this section we apply FSL and AFSL to data from the Childhood Asthma Management Program,
CAMP. CAMP was a multi–center, longitudinal clinical trial designed to better understand the long
term impact of two common daily asthma medications, Budesonide and Nedocromil, on children
(The Childhood Asthma Management Program Research Group, 1999). Phenotypic and genotypic
data are freely available at dbGaP, study accession phs000166.v2.p1. The data we consider here
consists of 540 caucasian subjects, ages 5–12, with asthma, who were randomly assigned a particular
treatment (one of the two drugs or placebo). Each subject made 16 clinical visits spread over a
four year period. The outcome is (log) forced expiratory volume in one second or FEV1, which is
a commonly used proxy for lung strength. For the domain, we use “study time” which represents
how long the subject has been on a particular treatment. In Figure 3 we plot log(FEV1) curves
for the 540 subjects. The left panel shows curves expanded using bsplines, with the same approach
as in Section 5.1. The right panel plots the curves after regressing out gender, age, and treatment,
using standard function-on-scalar regression techniques (Reimherr and Nicolae, 2014; Reimherr and
Nicolae, 2016). We use the standardized curves as response functions when applying our methods.
The data are observed on a grid, which is not quite evenly spaced as there are more observations
early on in the study. In particular, the first three observations are 1-2 weeks apart, the next two
are 2 months apart, and then remaining 11 visits are 4-5 months apart. As with the simulations,
the data are preprocessed using bsplines and FPCA. We mention that with 16 observed time points,
readers will rightly be concerned about the effects of such preprocessing on data which is not densely
observed. However, since the data is observed on a common grid and 16 is well beyond the typical
rule of thumb (5401/4 ≈ 5) we believe that this approach is justified. Furthermore, as we saw in
the simulations Section 5.1, our methods still work well with 16 time points.
We applied FSL and AFSL to a subset of 10,000 of single nucleotide polymorphisms, SNPs,
which were prescreened using methods from Chu et al. (2016). As in Barber et al. (2016), we
applied FSL and AFSL using both BIC and extended BIC (with a parameter of 0.2). It is well
known that the extended BIC helps control false positives better when the number of predictors
is large (Chen and Chen, 2008; Foygel and Drton, 2010). In Table 3 11 out of 10,000 SNPs were
selected by FSL, and 10 out of 11 SNPs selected by AFSL, regardless of whether BIC or EBIC was
17
Figure 3: The left panel is a plot of the log(FEV1) curves for the 540 subjects from CAMP. Each
trajectory is estimated using penalized bsplines. The right panel is the same, but the curves have
been standardized by age, gender, and treatment.
used. In Figure 4, we see that rs1875650, rs1368183, rs7751381, rs17372029, rs1540897, rs4734250,
rs4752250 and rs2019435 have a positive impact on lung development in growing children. We can
also conclude that rs953044 and rs2206980 have a negative impact on lung development in growing
children. Comparing FSL with AFSL, we see that FSL tends to over shrink the estimates and a
great deal of the curvature of the estimates is lost. In contrast, AFSL shrinks less and the changes
over time are much clearer.
As a final comparison we compare the prediction error between FSL and AFSL on the CAMP
data. To do this, we split the data set into a training set (80%) and test set (20%). We build our
model on the training set while comparing the predictive performance on the test set. We measure
performance in terms of the root mean squared error, RMSE:
v
u
N
u1 X
kYn − Ŷn k2 .
(6)
RMSE = t
N
n=1
Here Ŷn is the predicted functional outcome and Yn is the observed for the nth subject from the
test dataset. We repeat this split 10 times and average the results, which are given in Table 4. The
difference in terms of BIC vs EBIC seems to be negligible, while AFSL consistently produces an
RMSE which is about 15% lower than FSL. This further supports the idea that rs20141420 might
18
rs7751381
rs17372029
rs1540897
0.05
×10 -3
8
0.02
rs1875650
0.07
7
0.06
0.04
6
0.015
0.05
5
0.03
0.04
4
0.01
0.03
0.02
3
0.02
2
0.005
0.01
0.01
1
0
0
0
0
0.2
0.4
0.6
0.8
0
‘1
rs2019435
0.2
0.4
0.6
0.8
0
0
1
rs4734250
0.06
0.07
0.05
0.06
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
rs2206980
rs4752250
0.01
0.05
0
0.04
-0.01
0.05
0.04
0.03
0.04
-0.02
0.03
-0.03
0.03
0.02
0.02
0.02
-0.04
0.01
0.01
0.01
0
0
0.2
0.4
0.6
0.8
1
-0.05
0
0
0.2
0.4
0.6
0.8
1
0
-0.06
0
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
rs953044
rs1368183
12
0.2
×10 -3
0.5
10
×10 -4
0
8
-0.5
6
-1
4
-1.5
2
-2
0
-2.5
-2
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
Figure 4: Estimated β coefficients for top ten SNPs from CAMP. The FSL estimate is plotted using
blue squares and the AFSL is plotted using red dashes.
19
1
SNP
FSL
AFSL
Chromosome
Name
BIC
EBIC
BIC
EBIC
1
rs1875650
×
×
×
×
2
rs953044
×
×
×
×
5
rs1368183
×
×
×
×
6
rs7751381
×
×
×
×
6
rs2206980
×
×
×
×
7
rs17372029
×
×
×
×
8
rs1540897
×
×
×
×
8
rs4734250
×
×
×
×
10
rs4752250
×
×
×
×
15
rs2019435
×
×
×
×
20
rs2041420
×
×
Table 3: Top SNPs selected by FSL and by AFSL. A × indicated the SNP is selected when using
the corresponding variable selection criteria.
be a false positive, as this was the only one dropped by AFSL, as indicated in Table 3.
6
Conclusion and Future Work
In this paper we have presented an adaptive version of the function-on-scalar lasso introduced in
Barber et al. (2016). Using techniques from convex analysis over Hilbert spaces, we were showed
that AFSL estimates achieve the functional oracle property even when the number of predictors is
exponential relative to the sample size.
There are still a great deal of open problems in this area as relatively little has been done for
high dimensional functional data. We used AFSL to control the bias inherent in FSL, but other
approaches should work as well. For example, SCAD would also be an excellent candidate for
reducing the bias of the estimates, but serious theoretical developments would be required. In
another direction, alternating the penalty to more carefully control the level of smoothing could
prove beneficial as well. Additionally, these methods would greatly benefit from customized computational tools. In the present work and in Barber et al. (2016), an ADMM for scalar group lasso
20
BIC
EBIC
Test
FSL
AFSL
FSL
AFSL
1
0.1472
0.1310
0.1386
0.1194
2
0.1533
0.1275
0.1536
0.1272
3
0.1228
0.1267
0.1523
0.1062
4
0.1429
0.1098
0.1444
0.1315
5
0.1257
0.1299
0.1372
0.1248
6
0.1418
0.1178
0.1431
0.1167
7
0.1576
0.1219
0.1372
0.1248
8
0.1543
0.1094
0.1536
0.1272
9
0.1429
0.1098
0.1264
0.1338
10
0.1320
0.1502
0.1536
0.1248
Average
0.1421
0.1234
0.1440
0.1236
Table 4: RMSE for the CAMP data calculated by FSL and AFSL when using the corresponding
variable selection criteria.
was used to find the FSL and AFSL estimates. While this approach works well, it can be improved
by producing custom ADMM or coordinate descent methods which are specifically designed for
functional data.
Lastly, the tuning parameter was selected using BIC and EBIC, and in practice they work well,
but a stronger theoretical justification for their use is still needed in the functional context. It
would also be useful to explore other tuning parameter methods and see how they compare.
21
A
Proofs
In this section we collect the proofs for Theorems 1 and 2. The former is fairly straightforward as
the structure is nearly the same as for the scalar or multivariate setting. The latter is much more
involved and will be broken into several pieces.
A.1
Proof of Theorem 1
For a differentiable convex function, the subdifferential is equal to the (Fréchet) derivative. So we
have that the subdifferential of the first term of LA
λ with respect to βi is given by
N
1 X
−
Xni (Yn − Xn> β).
N
n=1
Turning to the second term, consider the functional g(x) = kxk. For x 6= 0, we need to verify that
subgradient is kxk−1 x, i.e., that for all y
g(y) − g(x) = kyk − kxk ≥ kxk−1 hx, y − xi.
This is equivalent to showing that
kykkxk − kxk2 ≥ hx, yi − kxk2 ,
but this follows from the Cauchy-Schwarz inequality. For x 6= 0 we need to show that for any
khk = 1 and all y we have
g(y) − g(x) = kyk ≥ hh, yi,
which again just follows from Cauchy-Schwarz. Thus the subdifferential of g(x) is verified. Combining this fact with the additivity and linearity of subdifferentials, we have that Theorem 1 holds.
A.2
Theorem 2: Preliminary
Here we state several key preliminary results which are useful for the main proof. We begin with
a reminder of the restricted eigenvalue condition.
Definition 4. (Restricted eigenvalue condition). A matrix M ∈ RN ×I satisfies the RE(I0 , α)
condition if, for all subsets S ⊂ {1, · · · , I} with |S| ≤ I0 ,
kM wk22 ≥ αN kwk22 for all w ∈ RI with kwsc kl1 ≤ 3kws |l1
22
In Barber et al. (2016) we showed that the scalar restricted eigenvalue condition implied a
functional one which was necessary for the convergence rates in FSL. We now restate a result
which follows from van de Geer and Bühlmann (2009), namely, that the irreprsentable condition
combined with a sparsity assumption implies that the restricted eigenvalue condition holds.
Lemma 1. If a matrix M satisfies Assumption 2 then it also satisfies RE(I0 , α).
Given the above result we can apply Theorem 3 from Barber et al. (2016) to obtain the
following.
Lemma 2. If Assumptions 1 and 2 hold then the FSL estimate β̃ satisfies
1/2
sup kβi − β̃i k = Op (rN )
where
rN =
i∈S
I0 log(I)
.
N
The following functional concentration inequality from Barber et al. (2016) will also be useful.
Lemma 3. Let X be an H valued Gaussian process with mean zero and covariance operator C.
Then we have the bound
√
P {kXk2 ≥ kCk1 + 2kCk2 t + 2kCk∞ t} ≤ exp(−t).
where kCk1 is the sum of the eigenvalues of C, kCk22 is the sum of the squared eigenvalues, and
kCk∞ is the largest eigenvalue.
Our last lemma is mainly for technical convenience as it shows that the operator norm of a
matrix is the same regardless of the Hilbert space it is acting on. This is especially useful as it
implies that the functional and scalar irrepresentable conditions are the same.
Lemma 4. Let A ∈ Rq×p be viewed as a linear operator from Hp → Hq . Then the operator norm
of A is the same for any Hilbert space H, and in particular when H = R.
Proof. Recall that when A is viewed as A : Rp → Rq . Then,
kAkop = sup |Ax| = σ1 ,
|x|=1
where σ1 is the largest singular value. When A is viewed as on operator on another Hilbert space
A : Hp → Hq we define
kAkF ;op =
sup
x∈Hp :kxk=1
23
kAxk.
Let {ej } be an orthonormal basis of H, then xi =
P
j
xij ej , where xij ∈ R. Define xj = {xij : i =
1 . . . p}. Then by Parseval’s identity
kAxk2 =
X
|Axj |2 ≤ σ12
j
X
|xj |2 = σ12 kxk2 .
j
Which means that kAkF,op ≤ kAkop . To see that they are equal, take x such that x1 is equal to
the largest right singular vector of A, and all other xj are zero. Then we have kAxk2 = |Ax1 |2 =
σ12 = kAk2op , thus the two operator norms are equal.
A.3
Theorem 2: Proof
We now have the necessary tools to establish the proof of Theorem 2. We will split this into multiple
pieces. We begin with the longer argument, namely, that AFSL is variable selection consistent. We
will then finish with the shorter argument, namely, that AFSL is also asymptotically equivalent to
the oracle estimate.
Proof of Theorem 2.2:
We will mimic the structure of Huang et al. (2008). Partition β̂ = {β̂1 , β̂2 }, where β̂1 is the estimate
of the true predictors, and β̂2 of the null predictors. Define
s̃1 = {w̃i kβ̂i k−1 β̂i , i ∈ S}.
Suppose that Ŝ = S, that is, the support of β̂ is S. Then from Corollary 1 we have that
X>
1 (Y − X1 β̂1 ) = λs̃1 .
Solving for β̂1 we get that
−1
β̂1 = (X>
1 X1 ) (X1 Y − λs̃1 )
or equivalently
−1
>
>
−1
>
β̂1 = (X>
1 X1 ) (X1 Y − λs̃1 ) = β1 + (X1 X1 ) (X1 ε − λs̃1 )
If we define β̂ := {β̂1 , 0}, as above, then we conclude the following
s
β̂1 =
β1
and
(7)
Ŝ = S
if
kX > (Y − X β̂ )k < λw̃ ∀ i ∈
/ S.
1 1
i
i
24
In the scalar case, the first condition above is replaced with the property that β̂i and βi point in
s
the same direction (as this will imply that β̂1 = β1 ). Since we are working in the functional setting,
we cannot use such a replacement. Instead we use the slightly stronger property kβ̂i − βi k < kβi k
for all i ∈ S. This says that the estimate lies in a ball centered at βi with radius less than βi . In
the scalar case this will imply that the two have the same sign.
Examining the second condition in (7) we have that
−1
>
Y − X1 β̂1 =X1 β1 + ε − X1 β1 + X1 (X>
1 X1 ) (X1 ε − λs̃1 )
−1 >
>
−1
=(I − X1 (X>
1 X1 ) X1 )ε − λX1 (X1 X1 ) s̃1 .
−1 >
Let H = (I − X1 (X>
1 X1 ) X1 ) then we get the following
ke> (N −1 Σ̂−1 (X>
1 ε − λs̃1 ))k < kβi k
i
11
Ŝ = S
if
(8)
kX > (Hε − λN −1 X Σ̂−1 s̃ )k < λw̃
1 11 1
i
i
∀i∈S
∀i∈
/ S,
where ei is a vector with 1 in the ith coordinate and zero everywhere else. Now define the following
four events:
B1
B2
B3
B4
1 > −1 >
kβi k
=
ke Σ X εk ≥
: for some i ∈ S
N i 11 1
2
λ > −1
kβi k
=
ke Σ s̃1 k ≥
: for some i ∈ S
N i 11
2
λw̃i
>
= kXi Hεk ≥
: for some i ∈
/S
2
1
w̃i
−1
>
=
kXi X1 Σ11 s̃1 k ≥
: for some i ∈
/S .
N
2
We have the inclusion
{Ŝ =
6 S} ⊂ B1 ∪ B2 ∪ B3 ∪ B4 .
So we can show that P (Ŝ 6= S) → 0 if we can show that P (Bi ) → 0 for j = 1, . . . , 4. Showing this
for B1 and B3 is very similar to the scalar case. For B2 and B4 this will require more work as our
s̃1 is a bit different than in the scalar case since we can’t assume that β̂i and βi have the same sign
(for i ∈ S).
Step 1 P (B1 ) → 0:
Denote by ei ∈ RI0 a vector which is 1 in the ith coordinate and zero
everywhere else. Then we can express B1 = ∪i∈S Ai where
kβi k
kβi k2
1 T −1 T
kei Σ11 X1 εk ≥
= N −2 kTi k2 ≥
,
Ai =
N
2
4
25
1 T −1
1
T
where Ti = eTi Σ−1
11 X1 ε. Since the ε are Gaussian we have that N Ti ∼ N (0, N ei Σ11 ei C) Trivially
P
we have P (B1 ) ≤
P (Ai ). Applying Lemma 3, we have that for any t > 0
√
1 T −1
T −1
−2
2
T −1
≤ exp (−t) .
P N kTi k ≥
e Σ ei kCk1 + 2ei Σ11 ei kCk2 t + 2ei Σ11 ei kCk∞ t
N i 11
We can apply this to our problem if we can find t such that
b2
√
1 T −1
T −1
t
+
2e
Σ
e
kCk
t
≤ N,
ei Σ11 ei kCk1 + 2eTi Σ−1
e
kCk
∞
2
i 11 i
11 i
N
4
(9)
where recall that bN = mini∈S kβi k. Applying Assumption 2.3 we have that
tKτ
√
1 T −1
1
ei Σ11 ei kCk1 + 2kCk2 t + 2kCk∞ t ≤
N
N
where K is some constant depending only on C, the covariance operator of the errors. So (9) holds
if
b2
tKτ1
≤ N.
N
4
So we take t =
b2N N
4Kτ1
, which gives
P (Ai ) = P
b2N
1
2
kT
k
≥
i
N2
4
2
b N
≤ exp − N
.
4Kτ1
And,
P (B1 ) ≤
X
2
2
b N
b N
P (Ai ) ≤ I0 exp − N
= exp − N + log(I0 ) → 0,
4Kτ1
4Kτ1
by Assumption 2.1, thus, P (B1 ) → 0, which gives the desired result.
Step 2 P (B2 ) → 0: Recall,
B2 =
kβi k
λ > −1
ke Σ s̃1 k ≥
: for some i ∈ S ,
N i 11
2
and s̃1 = {w̃i kβ̂i k−1 β̂i , i ∈ S}. Trivially we have
ks̃1 k2 =
X
i∈S
w̃i2
kβ̂i k2
kβ̂i
k2
=
X
w̃i2 =
i∈S
X
i∈S
wi2 +
X
(w̃i2 − wi2 ).
i∈S
Recall that w̃i = kβ̃i k−1 , where β̃i is computed from FSL. For a differentiable functional f : H → R,
a first order Taylor expansion is given by
0
f (x + h) = f (x) + hh, f (x)i + O(h2 ).
26
0
where f : H → H. In our case,
f (x) =
1
,
kxk2
0
and f (x) = −
2
x, x 6= 0.
kxk4
Applying a first-order Taylor expansion for w̃i2 , we obtain
w̃i2 − wi2 =
1
2
1
−
≈ hβ̃i − βi , −
βi i = −2kβi k−4 hβi − β̃i , βi i.
2
2
kβi k
kβi k4
kβ̃i k
Then, by the Cauchy–Schwarz inequality,
2
2
1
1
1
1/2
|hβi − β̃i , βi i| ≤
kβi − β̃i k ≤
(sup kβi − β̃i k)
=
Op (rN )wi2 ,
kβi k4
kβi k3
bN i∈S
kβi k2
bN
Therefore,
X
w̃i2 − wi2 ≤
X
1
1/2
Op (rN )
wi2 ,
bN
1/2
and since rN /bN → 0 by Assumption 2.1 we have that
ks̃1 k2 ≤
(10)
X
I0
wi2 (1 + op (1)) ≤ 2 (1 + op (1)).
bN
Returning to the original objective, we combine (10), an operator inequality, and Assumptions 2.2
and 2.3 to obtain
1/2
−1
−1
λke>
λke>
τ1 λI0
i Σ11 kop ks̃1 k λτ1 ks̃1 k
i Σ11 s̃1 k
≤
≤
≤
(1 + op (1)) → 0.
N kβi k
N kβi k
N bN
N b2N
Since this holds uniformly in i we have that P (B2 ) → 0, which gives the desired result.
Step 3 P (B3 ) → 0: Recall,
λw̃i
>
B3 = kXi Hεk ≥
: for some i ∈
/S
2
where
−1 >
H = (I − X1 (X>
1 X1 ) X1 ),
is the orthogonal projection. Let B3 = ∪i∈S c Ai , where
λw̃i
>
Ai = kXi Hεk ≥
.
2
Next define, Ti = XiT Hε. Then Ti ∈ H is Gaussian, has mean zero, and covariance CTi = XiT HXi C
since H is idempotent. Notice that we have
kXi> Hεk ≥
λ
λw̃i
⇐⇒ kβ̃i kkXi> Hεk ≥ .
2
2
27
1/2
By Lemma 2, supi∈S c kβ̃i k = Op (rN ). Then Ai can be expressed as
λ
1/2
Ai = Op (rN )kXi> Hεk ≥
,
2
where Op(1) is uniform with respect to i. Then, for any > 0 we can find K > 0 large such that
!
X
X
λ
>
P (B3 ) ≤
P (Ai ) ≤ /2 +
P kXi Hεk ≥
.
1/2
2KrN
i∈S c
So we need only to show that we can make the second term above small as well. Again, we will
apply Lemma 1, so we want to find t, such that
(11)
XiT HXi kCk1
+
2XiT HXi kCk2
√
t+
2XiT HXi kCk∞ t
!2
λ
≤
1/2
.
2KrN
Notice that because H is a projection and the Xi are standardized, we have that
√
XiT HXi (kCk1 + 2kCk2 t + 2kCk∞ t) ≤ N K2 t,
where K2 is a second constant depending on C only. So (11) holds if
!2
λ
tK2 N ≤
.
1/2
2KrN
2
Combining constants and labeling it K3 , we can take t= K3λN rN , which gives
!
λ
λ2
>
P kXi Hεk ≥
.
≤ exp −
1/2
K3 N rN
2KrN
We can then bound P (Bi ) as
P (Bi ) ≤
X
P (Ai ) ≤ I exp −
λ2
K3 N rN
λ2
+ /2 = exp −
+ log(I) + /2.
K3 I0 log(I)
By assumption 2.2, we can take N large to make the first term above smaller than /2 as well,
which gives the desired result.
Step 4 P (B4 ) → 0: Recall,
B4 =
w̃i
1
−1
>
kXi X1 Σ11 s̃1 k ≥
: for some i ∈
/S .
N
2
From (10) and Assumption 2.3 we have
1/2
I
ks̃1 k ≤ 0 (1 + oP (1))
bN
and
28
kΣ−1
11 kop ≤ τ1 .
From Step 3 we have
1/2
sup w̃i−1 = sup kβ̃i k = OP (rN ).
i∈S c
i∈S c
By Assumption 2.4 we have
kXi> X1 Σ−1
11 kop
≤ φ.
N
Combining these together we have
2kXi> X1 Σ−1
2kXi> X1 Σ−1
11 kop ks̃1 k
11 s̃1 k
≤
N w̃i
N w̃i
1/2 1/2
φrN I0
Op (1)
bN
I0 log1/2 (I)
=
Op (1) → 0,
bN N 1/2
≤
by Assumptions 2.1 and 2.2 uniformly in i, thus, P (B4 ) → 0, which gives the desired result.
Proof of Theorem 2.2:
As we have just shown, when Ŝ = S then the AFSL estimator takes the form β̂ = (β̂1 , 0) where
s
β̂ = β and
β̂1 = (XT1 X1 )−1 (XT1 Y − λs̃1 ) = β1 + (XT1 X1 )−1 (XT1 ε − λs̃1 ).
The oracle estimator is given by β̂O = (β̂1O , 0) where
−1
>
>
−1
>
β̂1O = (X>
1 X1 ) (X1 Y ) = β1 + (X1 X1 ) (X1 ε).
We then have, for > 0
P
n√
o
n√
o
N kβ̂O − β̂k ≥ ≤ P (Ŝ =
6 S) + P
N kβ̂1O − β̂1 k ≥ .
The first term can be made arbitrarily small by Theorem 2.1. Turning to the second term
√
λ
N kβ̂1 − β̂1O k = √ kΣ−1
11 s̃1 k.
N
From (10) and Assumption 3we have
1/2
λ
λτ1 I0
√ kΣ−1
Op (1) → 0,
11 s̃1 k ≤ √
N
N bN
which gives the desired result.
29
References
Barber, R.F., Reimherr, M., and Schill, T. (2016). Function-on-scalar lasso with applications to
longitudinal GWAS. Under Revision.
Barbu, V. and Precupanu, T. (2012). Convexity and optimization in banach spaces. Springer
Science & Business Media.
Bauschke, H. and Combettes, P. (2011). Convex analysis and monotone operator theory in hilbert
spaces. Springer Science & Business Media.
Boyd, S., Parikh, N., Chu, E., Peleato, B. and Eckstein, J. (2011). Distributed optimization and
statistical learning via the alternating direction method of multipliers. Foundations and Trends
in Machine Learning, 3, number 1, 1–122.
Boyd, S. and Vandenberghe, L. (2004). Convex optimization. Cambridge university press.
Bühlmann, P. and van de Geer, S. (2011). Statistics for high-dimensional data: methods, theory
and applications. Springer.
Cai, T. and Guo, Z. (2015). Confidence intervals for high-dimensional linear regression: Minimax
rates and adaptivity. arXiv preprint arXiv:1506.05539.
Cardot, H., Mas, A. and Sarda, P. (2007). CLT in functional linear regression models. Probability
Theory and Related Fields, 138, 325–361.
Chen, J. and Chen, Z. (2008). Extended Bayesian information criteria for model selection with
large model spaces. Biometrika, 95, 759–771.
Chen, Y., Goldsmith, J. and Ogden, T. (2016). Variable selection in function-on-scalar regression.
Stat, 5, 88–101.
Chu, W., Li, R. and Reimherr, M. (2016). Feature screening for time-varying coefficient models
with ultrahigh dimensional longitudinal data. Annals of Applied Statistics Forthcoming.
Fan, J. and Li, R. (2001). Variable selection via nonconcave penalized likelihood and its oracle
properties. Journal of the American Statistical Association, 96, 1348–1360.
Fan, Y., James, G. and Radchenko, P. (2015). Functional additive regression. Annals of Statistics,
43, number 5, 2296–2325.
30
Foygel, R. and Drton, M. (2010). Extended Bayesian information criteria for Gaussian graphical
models. In Advances in Neural Information Processing Systems, pp. 604–612.
Gertheiss, J., Maity, A. and Staicu, A.M. (2013). Variable selection in generalized functional linear
models. Stat, 2, 86–101.
Horváth, L. and Kokoszka, P. (2012). Inference for Functional Data with Applications. Springer.
Huang, J., Ma, S. and Zhang, C.H. (2008). Adaptive lasso for sparse high-dimensional regression
models. Statistica Sinica, 18, 1603.
Lian, H. (2013). Shrinkage estimation and selection for multiple functional regression. Statistica
Sinica, 23, 51–74.
Matsui, H. and Konishi, S. (2011). Variable selection for functional regression models via the L1
regularization. Computational Statistics & Data Analysis, 55, 3304–3310.
Morris, J. (2015). Functional regression. Annual Review of Statistics and Its Application, 2,
321–359.
Reimherr, M. and Nicolae, D. (2014). A functional data analysis approach for genetic association
studies. The Annals of Applied Statistics, 8, 406–429.
Reimherr, M. and Nicolae, D. (2016). Estimating variance components in functional linear models with applications to genetic heritability. Journal of the American Statistical Association,
Forthcoming.
Reiss, P.T., Mennes, M. and Huang, L. (2010). Fast function–on–scalar regression with penalized
basis expansions. International Journal of Biostatistics, 6, Article 28.
Shor, N. (2012). Minimization methods for non-differentiable functions, volume 3. Springer Science
& Business Media.
The Childhood Asthma Management Program Research Group (1999). The childhood asthma
management program (CAMP): design, rationale, and methods. Controlled Clinical Trials, 20,
91–120.
Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal
Statistical Society, B, 58, 267–288.
31
van de Geer, S. and Bühlmann, P. (2009). On the conditions used to prove oracle results for the
lasso. Electronic Journal of Statistics, 3, 1360–1392.
Zhao, P. and Yu, B. (2006). On model selection consistency of lasso. The Journal of Machine
Learning Research, 7, 2541–2563.
Zou, H. (2006). The adaptive lasso and its oracle properties. Journal of the American Statistical
Association, 101, 1418–1429.
32
| 10math.ST
|
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
1
Mapping Generative Models onto a
Network of Digital Spiking Neurons
arXiv:1509.07302v2 [cs.NE] 9 Oct 2015
Bruno U. Pedroni, Srinjoy Das, John V. Arthur, Paul A. Merolla, Bryan L. Jackson, Dharmendra S. Modha,
Kenneth Kreutz-Delgado, and Gert Cauwenberghs
Abstract—Stochastic neural networks such as Restricted Boltzmann Machines (RBMs) have been successfully used in applications ranging from speech recognition to image classification.
Inference and learning in these algorithms use a Markov Chain
Monte Carlo procedure called Gibbs sampling, where a logistic
function forms the kernel of this sampler. On the other side of the
spectrum, neuromorphic systems have shown great promise for
low-power and parallelized cognitive computing, but lack wellsuited applications and automation procedures. In this work, we
propose a systematic method for bridging the RBM algorithm
and digital neuromorphic systems, with a generative pattern
completion task as proof of concept. For this, we first propose
a method of producing the Gibbs sampler using bio-inspired
digital noisy integrate-and-fire neurons. Next, we describe the
process of mapping generative RBMs trained offline onto the
IBM TrueNorth neurosynaptic processor – a low-power digital
neuromorphic VLSI substrate. Mapping these algorithms onto
neuromorphic hardware presents unique challenges in network
connectivity and weight and bias quantization, which, in turn,
require architectural and design strategies for the physical realization. Generative performance metrics are analyzed to validate
the neuromorphic requirements and to best select the neuron parameters for the model. Lastly, we describe a design automation
procedure which achieves optimal resource usage, accounting for
the novel hardware adaptations. This work represents the first
implementation of generative RBM inference on a neuromorphic
VLSI substrate.
Index Terms—Generative model, neuromorphic VLSI, Restricted Boltzmann Machine, spiking digital neuron, Gibbs sampling.
(EEG) data feature learning and classification [5], [6]. RBMs
are generative learning algorithms and are particularly useful
in extracting features from unlabeled data (i.e. unsupervised
learning) [7]. Structurally, an RBM is a stochastic neural
network composed of 2 layers of neuron-like units: a layer
of visible units v which are driven by the real-world data of
interest and a layer of hidden units h which form connections
to these visible units. There are no interconnections within
a layer and the weights of connections between layers are
symmetric. Fig. 1a exemplifies an RBM with 4 visible and 3
hidden units.
The RBM defines a joint probability over the input data and
hidden variables specified by the Boltzmann distribution [8]:
e−E(v,h)
,
−E(v,h)
v,h e
p(v,h) = P
where E(v,h) = −vT Wh − bTv v − bTh h. Here p denotes
the Boltzmann probability distribution and E is a function
(also known as the “energy function”) of v and h, where v
denotes the binary state (0 or 1) of the visible units and h
represents the binary state of the hidden units. The weight
between visible and hidden units is represented by W, while
bv and bh represent the biases of v and h, respectively. The
denominator is the sum of all possible states of visible and
hidden units, also known as the partition function.
h
I. I NTRODUCTION
D
EEP Learning algorithms such as Restricted Boltzmann
Machines (RBMs) and Deep Belief Networks (DBNs)
have been successfully used in a wide range of cognitive
computing applications such as image classification [1], speech
recognition [2], [3], and motion synthesis [4]. Additionally,
these algorithms have been explored as possible solutions for
Brain-Computer Interfaces (BCI) and electroencephalography
Manuscript received October 9, 2015. This work was supported by the
DARPA SyNAPSE Program, the National Science Foundation Emerging Frontiers in Research Innovation ENG-1137279 and Expeditions in Computing
CCF-1317407, the Office of Naval Research (ONR MURI 14-13-1-0205),
and CNPq Brazil (CsF 201174/2012-0).
B.U. Pedroni and G. Cauwenberghs are with the Department of Bioengineering and the Institute for Neural Computation, UC San Diego, La Jolla,
CA 92093 USA (e-mails: [email protected], [email protected]).
S. Das and K. Kreutz-Delgado are with the Electrical and Computer
Engineering Department, UC San Diego, La Jolla, CA 92093 USA (e-mails:
[email protected], [email protected]).
J.V. Arthur, P.A. Merolla, B.L. Jackson, and D.S. Modha are with IBM
Research - Almaden, San Jose, CA 95120 USA (e-mails: arthurjo, pameroll,
bryanlj, [email protected]).
(1)
h3
W
v
h2
(a)
h1
h2
v1
data
h1
hk
v2
...
time
(b)
v
vk
data
(c)
Fig. 1: RBM and DBN representations. (a) An RBM formed
by 4 visible and 3 hidden units. (b) Gibbs sampling procedure
in an RBM. (c) A DBN formed by stacking RBMs.
Inference in an RBM can be performed using a Markov
Chain Monte Carlo (MCMC) procedure called Gibbs sam-
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
pling, where each unit in any given layer is sampled conditioning on its total input from units in the other layer. Fig. 1b
illustrates k steps of MCMC performed in an RBM. The
Gibbs sampling rule in binary RBMs is defined by the logistic
function,
σ(x) = 1/(1 + e−x ),
(2)
with the probability of activation of unit i as defined by [8]
X
p(xi = 1|xj ) = σ(
wij xj + bi ),
(3)
j
where wij is the weight from unit j to unit i for all j ∈
/
layer(i), and bi is the bias of unit i. DBNs are formed by
stacking layers of RBMs (Fig. 1c) and it has been shown
that inference in a DBN can be done in a successive layerby-layer manner on each RBM [9]. RBMs and DBNs can
be used with labeled data for classification tasks either as
feature extractors to an external classifier or as completely selfcontained discriminative machine learning frameworks [1].
However, most of the data in the real world is unlabeled and,
in such situations, RBMs and DBNs can be used to perform
generative inference tasks. Applications of inference in such
unsupervised frameworks include, for example, restoration
of incomplete or occluded images and prediction of motion
sequences.
Currently, inference tasks using RBMs and DBNs are overwhelmingly realized in software, which are typically run on
high performance CPUs (Central Processing Unit) and GPUs
(Graphical Processing Unit). For ultra low-power, real-time
realizations of these algorithms, such as in mobile devices,
the solution tends to be sending information to the cloud for
processing. However, this demands, in many cases, reliable
communication between client and server, along with large
amounts of transmitted data. In this context, the Neuromorphic
Computing paradigm is a more suitable solution in terms of
low-power client-side processing. Neuromorphic VLSI (Very
Large Scale Integrated Circuit) systems [10]–[17], inspired by
biological neural architectures and functions, have been realized with analog, digital, and mixed-signal circuit elements.
Such systems typically compute in a massively parallel fashion
and communicate asynchronously using spikes. The principal
benefit of this architecture, which stands in contrast to the
traditional von Neumann computing paradigm, is extremely
energy efficient computation in a highly concurrent fashion.
Algorithms which demand large matrix multiplications, such
as RBMs and DBNs, benefit greatly in terms of computation
(and, consequently, power) when implemented in spike-based
systems, mainly because multiplications by zero are avoided
(i.e. absence of spikes does not generate computation). Therefore, arrays of spiking neurons realized on neuromorphic VLSI
are ideal for classification, generation and other inference tasks
in the context of real-world high dimensional data.
The goal of our work is to develop a modular architecture in
a systematic fashion to form a foundation for building neural
networks, such as RBMs and DBNs, on substrates of digital
spiking neurons. As a proof of concept of our design approach,
we implement a pre-trained (i.e. trained offline) generative
2
RBM for pattern completion on the TrueNorth digital neuromorphic VLSI device using the MNIST handwritten digit
images dataset.
The remainder of this paper is divided in the following
manner: Section II describes the Markov chain analysis of
the digital neural sampler; Section III describes the TrueNorth
system and the challenges in implementing Deep Learning
algorithms, along with the necessary steps for mapping the
RBM algorithm onto digital spiking neuromorphic hardware;
Section IV discusses quality metrics and the impact on generative performance when using the digital neural sampler
and sparse network connectivity; Section V shows the developed 3-stage RBM architecture and the generative model on
TrueNorth; Section VI illustrates the spike processing flow in
the TrueNorth RBM; Section VII details the design automation procedure for optimal hardware utilization; Section VIII
presents the results of the physically-implemented TrueNorth
generative RBM; and the last section discusses conclusions
and future work.
II. M ARKOVIAN ANALYSIS OF THE DIGITAL NEURAL
LOGISTIC SAMPLER
The kernel of the MCMC procedure for inference in an
RBM is the Gibbs Sampler and involves sampling from a
logistic function (Eq. (3)) [18], [19]. More specifically, it
involves sampling from a Bernoulli distribution (defining the
state of the RBM unit, x in Eq. (3)) parameterized by a
logistic function (activation probability). Traditional methods
for realizing a logistic sampler in hardware demand a lookup table or functional approximation for the sigmoid [20]–
[22], which is then compared to the output of a pseudorandom number generator. On the other hand, in spiking neural
hardware, such as TrueNorth, the only computational primitives are neurons. Since sigmoidal activation functions are not
inherently present in TrueNorth, we therefore have to make use
of the deterministic and stochastic neurodynamical properties
of the system for efficient realization of the logistic sampler.
Below we describe the process of Gibbs sampling using digital
spiking neurons in a Markov chain framework, which is useful
for better understanding the sampler behavior and serves as
a means for producing the generative performance metrics
detailed in Section IV. The solution neatly combines producing
the logistic function and sampling the state of the RBM unit.
In [23] it was shown that a digital integrate-and-fire neuron with a uniformly-sampled threshold combined with a
Bernoulli-sampled leak can approximate a logistic spiking
probability for the corresponding RBM unit. Here we expand
on this by providing a Markov chain analysis of the discretetime neural sampler. The neural sampling procedure is initialized by setting the neural membrane potential (Vm ) to a value
equivalent to the argument of the logistic function (which is a
function of the weights, bias and unit states, as shown in Eq.
(3)). Afterward, the system uses three neural variables (two
stochastic and one deterministic) to produce an approximate
sigmoidal spiking probability. These variables are explained
next.
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
1) Stochastic leak. The stochastic leak is an integer
value added to the membrane potential, and is sampled from
Bernoulli trials with p = 0.5. In other words, at every time
step (“tick”), either the membrane potential remains the same
or it is incremented by the leak value (L). This type of leak
is inspired from the TrueNorth system, whose neurons can
be configured with stochastic non-voltage-dependent leak. For
our setup, we chose to use a positive leak, however TrueNorth
neurons can take on positive or negative leak values. The
TrueNorth system will be explained in detail in Section III.
2) Stochastic threshold. The stochastic threshold is an
integer sampled from a uniform distribution between Vth and
Vth + T R; the term T R stands for “threshold range”. At every
tick, the membrane potential of the neuron is compared with
the stochastic threshold and, in case the potential hits (i.e.
is equal to or exceeds) the threshold, a spike event will be
generated.
3) Sampling time window. The deterministic component
of the sampler is the sampling time window (TS ), which is
the number of ticks during which the neuron is observed. The
operation of the sampler during TS is the following:
a. If during TS the neuron hits the threshold at least once,
a spike event after TS is produced.
b. Even if the neuron hits the threshold more than once
during TS , the sampler must still produce a single spike
at the output after TS .
c. If no threshold events occur during TS , then no spike
event is produced at the output of the sampler after TS .
A. Sampling algorithm using digital neurons
The algorithm, using TrueNorth-based I&F neurons with
stochastic leak (L) and threshold (Vth rand ), for realizing the
sigmoidal sampling rule (Eq. (3)) to perform MCMC sampling
in RBMs is given below.
Vm = Vinit
spiked = 0
repeat
Vm = Vm + B(0.5)*L
Vth rand = U(Vth , Vth + T R)
if (Vm ≥ Vth rand ): spiked = 1
until TS steps;
will analyze the effect of the stochastic neural variables using
discrete-time Markov chains.
B. Adaptation of neural variables into Markov chains
Since we are dealing with a discrete-time digital system,
the stochastic neural variables can be modeled as coupling between two discrete-time Markov chains (DTMC): a stochastic
leak DTMC and a stochastic threshold DTMC. The sampling
time window determines how many steps should be taken in
these chains. Each state in a chain is the instantaneous value
of the membrane potential. Due to a limited number of bits
for data representation in the digital system, saturation levels
should be taken into account. For illustrative purposes, in our
examples we consider only positive leak values, implying that
only the positive saturation level will come into effect, as any
data point (i.e. membrane potential) beyond it will be clipped
to the saturation value.
The sampler operates by first initializing the stochastic leak
DTMC at the state which represents the initial membrane
potential value, and then taking alternate steps between the
stochastic leak and the stochastic threshold DTMCs. Different
initialization values of the stochastic leak DTMC yield different sigmoidal probabilities. Both DTMCs present the same
number of states, defined by the membrane potential range.
In terms of structure, the DTMCs will always present states
representing lower-valued membrane potentials to the left of
the chain, and consequently the rightmost state represents
membrane potential equal to Vsat . Next, we discuss the effect
of the three neural properties on the DTMCs.
1) Stochastic leak DTMC. Since the stochastic leak
chosen for our examples causes only non-negative change in
the membrane potential, the only possible transitions, at each
stochastic leak tick, from a state are: (1) to itself (in the event
of no leak occurrence) or (2) to the right (positive additive leak
occurrence). Figure 2 shows the general case of the DTMC
for the stochastic leak (L). The number next to each state
transition is the transition probability (set to 0.5 for all states).
Note how no value of membrane potential can surpass Vsat ,
which makes the state representing this specific membrane
potential an absorbing state [24], [25]. Since it is the only
absorbing state in the chain, it is called the terminating state
in a terminating DTMC. Also note this state will be reached by
more than one other state (not considering the self-connection)
when L > 1.
..
0.5
0.5
0.5
0.5
0.5
-Vsat
-Vsat
+1
+1
-Vsat
-Vsat
...
0.5
..
0.5
-Vsat
-Vsat
+1
+1
+L
+L
-Vsat
-Vsat
+L
+L
0.5
..
..
..
...
The term B(p) represents a Bernoulli sample (0 or 1)
with probability p and U(a, b) is an integer sampled from a
uniform distribution between a and b (both inclusive). The
membrane potential (Vm ) is initialized to Vinit and, during
the “repeat” cycle, if Vm crosses the threshold (equivalent to
Vm ≥ Vth rand ), the spiked variable will be set to 1, after
which it will remain in this state until the end of the TS
time steps. Therefore, the state of spiked after TS ticks will
produce a sample (given the initial membrane potential) from
an approximate sigmoidal spiking probability distribution. The
state of a sampled RBM unit using this algorithm is equal
to the state of spiked. How to generate the spiked variable
using TrueNorth neurons will be explained in Subsection III-B,
along with implementation details in Section V. Next, we
3
0.5
Vsat
Vsat
-1
-1
...
0.5
Vsat
Vsat
Vsat
1
Fig. 2: DTMC for stochastic leak in the neural sampler.
2) Stochastic threshold DTMC. The stochastic threshold
is sampled, at each stochastic threshold tick, from a uniform
distribution, which produces a linearly increasing transition
probability from states inside the range [Vth : Vth + T R]
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
Vsat - 2
Vsat - 1
Vsat
0 ...
1
2
0
0 ... 0
0
0
0
0
1
2
0 ... 0
1
2
0 ... 0
0
0
0
0
0
1
2
... 0
0
1
2
... 0
0
0
0
Pl =
0
0
0 ... 0
0
0 ...
1
2
0
0
0
0
0 ... 0
0
0 ... 0
1
2
0
0
0
0 ... 0
0
0 ... 0
0
1
2
1
2
1
2
1
2
0
0
0 ... 0
0
0 ... 0
0
0
1
A terminating Markov chain presents a single absorbing
state, also known as the terminating state; all the other states
are transient. The transition probability matrix – with rows
representing origin states and columns representing destination
states – of a terminating Markov chain can be defined in the
following manner:
P=
T T
0 1
0
(4)
In matrix P , the m × m transient-states transition matrix is
represented by T , the row-vector 0 represents the terminating
state’s non-transient transitions, and (Im −T )1 = T0 . Therefore,
the entire transition matrix P can be characterized by simply
knowing T .
Vsat
C. Matrix representation of Markov chains
Pth =
1
Vsat - 1
3) Sampling time window. The deterministic component
of the sampler, the sampling time window, defines the number
of steps taken in the Markov chains and is a two-phase
process. The first phase occurs in the leak DTMC, where
a new membrane potential value is assigned to the neuron.
The second phase is the evaluation of the newly-assigned
membrane potential in relation to the noisy threshold. This
entire process is considered one step in the coupled DTMCs.
In case the system resides in the terminating state, Vsat , of the
coupled DTMCs after TS , an ultimate single spike event will
be produced; if the system is in any other state, no spike event
will be produced. This results in spike events sampled from
the sigmoidal spiking probability, conditioning on the starting
state (Vinit ) of the procedure.
2) Stochastic threshold DTMC. The stochastic threshold
is characterized by the base threshold value (Vth ) and the
threshold range (T R). The threshold DTMC can be defined
by the transition matrix Pth in Eq. (6). The colors represent
the same individual components of Pth as in Eq. (4).
0 ... 0
0
0 ... 0
0 ... 0
0
0
1 ... 0
0
0 ... 0
0 ... 0
0
0
0 ... 1
0
0 ... 0
0 ... 0
0
0
0 ... 0 (1− Δ1 ) 0 ... 0
0 ... 0
0
0 ... 0
0 (1− Δ2 ) ... 0
0 ... 0
1
Δ
2
Δ
0
0 ... 0
0
...
To transform spikes into probabilities, we must produce a
single spike event after TS in case the neuron reached the
threshold during TS . This can be obtained by using the Vsat
state as the terminating state also for the stochastic threshold.
A two-fold effect is produced by this terminating state: (1) the
two DTMCs become coupled by using a common terminating
state; and (2) the sigmoidal firing probability can be extracted
directly from the terminating state in the stochastic threshold
DTMC after TS , as will be shown below.
(5)
...
Fig. 3: DTMC for stochastic threshold in the neural sampler.
-Vsat + L + 2
0
1
1/Δ
-Vsat + L + 1
Vsat - 3
Vsat
Vsat
-Vsat + L
1
2
Vth + TR
1
Vth + TR - 1
Vsat
Vsat
-1
-1
0 ...
1
Δ
0 ... 0
1
(1− )
Δ
0
0 ... 0
0
0 ... 0
0 ... 0
1
0
0 ... 0
0
0 ... 0
0 ... 0
1
0
0 ... 0
0
0 ... 0
0 ... 0
1
...
...
...
Vth
Vth
+TR
+TR
Vth + 1
1-2/Δ
1
Vth
Vth
+TR
+TR
-1
-1
Vth
1-1/Δ
...
Vth - 1
1
Vth
Vth
Vth
Vth
+1
+1
-Vsat + 1
1
1-1/Δ
2/Δ
-Vsat
-Vsat
-Vsat
...
...
1/Δ
-Vsat
-Vsat
+1
+1
-Vsat + 2
Δ = (TR + 1)
-Vsat + 1
1) Stochastic leak DTMC. The stochastic leak is characterized by the additive leak value (L). The leak DTMC can be
defined by the transition matrix Pl in Eq. (5). The colors of the
matrix components represent the same individual components
as in Eq. (4).
-Vsat
to the spiking state. Figure 3 shows the general case of the
DTMC for the stochastic threshold. Note how values outside
the range previously described are guaranteed not to hit the
threshold when V < Vth (realized by the self-connections)
and guaranteed to hit the threshold when V ≥ (Vth + T R)
(realized by the connections to Vsat ). For simplification, in
the figure the symbol ∆ = (T R + 1).
4
(6)
3) Spiking probability: coupled DTMCs and sampling
time window. To obtain the sigmoidal spiking probability,
the two transition matrices must first be coupled to produce
Pc = Pl Pth . The spiking probability can now be obtained by
computing Psample = PcTS , which represents TS steps taken
in the coupled DTMC. With this, the last column (terminating
state) of the final matrix will contain the spiking probability,
Pspike , of each initial membrane potential (rows in the matrix).
Therefore,
Pspike (si ) = Psample (si , 2Vsat + 1),
(7)
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
where si is the origin state corresponding to the initial membrane potential of the neuron prior to sampling.
D. Example
The example, shown in Fig. 4, illustrates the sampler
obtained using the previous calculations and compared with
actual stochastic neuron simulations. The x-axis represents the
membrane potential (Vm ) of the neuron at the start of the sampler operation. As can be seen, the neural sampler obtained via
the coupled DTMC computation (blue line) and the stochastic
simulation (averaged over 104 samples for each initial Vm )
of the neuron (blue circles) are overlapping. Besides this, the
results from the DTMC computation approximate the ideal
sampler scaled by a factor of 10 (red line) with considerable
precision.
Since the logistic function in Eq. (2) naturally presents a
dynamic range between -6 and +6, and because the TrueNorth
system deals only with integer-valued membrane potentials,
the scaling factor is a means of increasing the resolution of
the neural sampler. In other words, by “streching out” the
function, each integer increment in the initial value of the
membrane potential represents a smaller step in the function,
resulting in higher resolution. To realize this using the digital
neuron model described thus far, the appropriate values of
TS , Vth , T R, and L must be chosen. In Subsection III-B,
when physically implementing the neural sampler algorithm in
the TrueNorth system, the use of the scaling factor is further
discussed, and a quantitative analysis of the TrueNorth neural
sampler versus the ideal sampler for parameter selection is
detailed in Section IV.
1
Ts=5
Vsat=110
Vth=0
TR=100
L=10
P(spike | Vm)
0.8
0.6
Ideal sampler (scale=10)
DTMC computation
Simulated neuron
0.2
−100
main detail is that the last column of Pth , originally considered
the terminating state, would be able to be transitioned out of
due to the stochastic leak in the next step of the Markov chain.
Also, the terminating state in Pl would be the first column,
which does not “line up” with Pth .
On the other hand, if a negative leak is applied, though
not sufficient for a chain in the rightmost state to reach a
state below the maximum value of the threshold (i.e. below
Vth +T R), then the correct spiking probability can be obtained.
In this manner, even if the leak causes a transition to the left
in Pl , the following iteration of Pth will force the system to
return to the rightmost state. Interestingly, the coupled activity
of the two DTMCs can preserve the original terminating state,
even if it is not the terminating state in Pl .
Discrete phase-type distributions (DPTDs) [26] are very
similar in nature to the developed neural sigmoid sampler. The
main difference is that DPTDs result from a system of one or
more inter-related and sequentially occurring geometric distributions, while the neural sampler results from a combination
of geometric (leak as Bernoulli trials) and uniform (threshold)
distributions.
Lastly, the digital neural sampler is an elegant solution
for sampling from a logistic function by not only using bioinspired neural dynamics but also simultaneously realizing two
operations: computing the spiking probability and sampling
to obtain the new state of the unit. The DTMC presented
can be very useful when simulating the network dynamics:
the neuron’s transition operator can be extracted by simply
accessing the spiking probability curve obtained from the
DTMC. This removes the demand of having to simulate every
step of the neuron during the sampling time window (TS ),
which comes in handy during the analysis of the generative
performance of the sampler in Subsection IV-B.
III. A PPROACHES FOR D EEP L EARNING
ON T RUE N ORTH
0.4
0
5
−50
0
Vm
50
100
Fig. 4: Ideal sampler versus DTMC computation and neural
simulation.
The noise sources of the stochastic leak and threshold
are, respectively, Bernoulli and uniform. By applying these
noise sources in a single tick, it is not possible to obtain
the precise S-shape of Fig. 4; only straight lines could be
obtained. Therefore, an explanation for the “curved” part of
the sigmoid (around Vm equal to -35 and +35) is the non-linear
behavior produced by the temporal aspect of the sampler (TS ).
Throughout multiple ticks, the combination of these “linear”
noise sources results in a more non-linear curve by creating
shorter segments from the straight lines.
E. Considerations
The problem was analyzed for a positive additive stochastic
leak, yet the same would be possible with a negative leak. The
Neuromorphic substrates present unique challenges for creating spiking versions of machine learning algorithms due
to data precision and network connectivity constraints. In
this work, a step-by-step methodology for porting RBMs
and DBNs onto the IBM TrueNorth system is detailed. The
MNIST dataset, consisting of 28x28 pixel grayscale images of
handwritten digits 0 through 9, was chosen for the generative
inference task. For our experiments, the images were binarized
to zero-one values for adaptation to the neuromorphic scenario.
The following subsections present the TrueNorth system and
outline the approaches and quantitative analysis of the algorithm adaptations necessary for mapping the (offline-trained)
networks.
A. The TrueNorth digital neurosynaptic processor
IBM’s TrueNorth is a very low-power, brain-inspired digital neurosynaptic processor [16], with 4096 cores, totaling
1 million programmable spiking neurons and 256 million
configurable synapses (Fig. 5a). The core is the basic building
block of the system, each composed of 256 axons (inputs) and
256 neurons (outputs) (Fig. 5b), connected via a 256 x 256
crossbar of configurable synapses (Fig. 5c). Each neuron can
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
target its generated spikes to any axon on the chip, limited
to one axon per neuron, and presents over 20 individually
programmable features, including threshold, leak, reset, and
stochastic properties. From the user’s point-of-view, neurons
operate in 1 ms time steps, during which asynchronous spike
event transmission and processing occurs between and inside
the cores. Therefore, during each 1 ms interval, spikes are
delivered to and processed in their destination cores, after
which a global clock aligns the generation of the next set
of spikes.
6
configurable weights. Therefore, an axon can be configured to
be one of four types, and this defines which of the four possible
weight values – individually in each neuron it is connected to
– will be integrated if the axon is active.
The second line (Eq. (8b)) represents the leak integration,
where λj is a 9-bit signed integer. Depending on the value of
cj , the leak can be deterministic (cj = 0) or stochastic (cj =
1). When cj = 0, the value of λj is integrated in the membrane
potential. On the other hand, when cj = 1, the stochastic
function F (λj ) = |λj | ≥ ρ defines if a leak of value sgn(λj )
is integrated; the value of ρ is a sampled uniformly distributed
8-bit integer. In this manner, a stochastic leak can only take
on values of +1 or -1. However, the value of L in the digital
neural sampler (refer to the algorithm in Subsection II-A) can
take on much larger values. How to implement stochastic leaks
greater than 1 on TrueNorth will be explained in Section III-B.
The last line (Eq. (8c)) compares the integrated membrane
potential with the threshold, which has a base value of αj and
a uniformly sampled value of η(Mj ) ranging from 0 to 2M −1.
Therefore, if Vj (t) is equal to or surpasses the threshold, the
neuron spikes and its membrane potential is reset to Rj . Using
the TrueNorth system as a basis for digital neural processing,
the next section shows how an approximation to the Gibbs
Sampler can be obtained using these neural properties.
(a)
B. Gibbs sampling with TrueNorth neurons
active synapse
256
axons
Axons
(inputs)
Neurons
(outputs)
..
.
..
..
Crossbar
(b)
256 neurons
(c)
Fig. 5: The TrueNorth neurosynaptic processor: (a) chip layout, wafer, and chip package; (b) high-level view of the 256
axons (inputs) and 256 neurons (outputs); and (c) internal view
of the fully-configurable binary crossbar [16].
The digital integrate-and-fire (I&F) TrueNorth neurons
present stochastic and deterministic leak and threshold properties. A simplified representation of the dynamical behavior of
the membrane potential Vj (t) for neuron j at time t is defined
by the following set of (sequentially processed per neuron)
equations [27]:
Gi
255
(8a)
Vj (t) = Vj (t − 1) + Σi=0 Ai (t) wi,j sj
Vj (t) = Vj (t) + (1 − cj )λj + cj F (λj )sgn(λj )
(8b)
if (Vj (t) ≥ αj + η(Mj )), Spike and set Vj (t) = Rj (8c)
The first line (Eq. (8a)) represents the synaptic integration
of all active axons impinging on neuron j at time t. The term
Ai (t) is the binary-valued input spike arriving from the ith
axon at time t; wi,j is the binary-valued synaptic connection
i
between axon i and neuron j; and sG
j is the synaptic weight
between axon i and neuron j. This last term is particularly
interesting as each neuron presents four 9-bit signed integer
Neural sampling can be realized on the TrueNorth system by
means of the algorithm described in Subsection II-A. The first
step of the algorithm is to set the initial membrane potential
of the neuron (Vinit in the algorithm) to the equivalent value
of the argument of the logistic function. This is realized in
TrueNorth by appropriately activating the axons of neuron
j at time t = 1 to produce the desired membrane potential
(Vj (1) = Vinit in Eq. (8a)). The neuron is then free to run (no
axon activity) during a sampling time window, TS , defined in
number of 1 ms time steps (“ticks”), during which a stochastic
additive leak is applied and the updated membrane potential
is evaluated at every tick. If the neuron’s membrane potential
is greater than or equal to the stochastic threshold (i.e. the
neuron spikes) at least once during TS , the binary state of the
equivalent RBM unit is set to 1 (i.e. the RBM unit spikes).
For adapting the algorithm in Subsection II-A to TrueNorth,
the stochastic threshold can be directly modeled by setting the
appropriate values of αr and Mr for TrueNorth neuron r. The
stochastic leak, on the other hand, cannot be directly mapped
for absolute leak values greater than 1. An alternative to this
is to use an additional neuron l to act as the stochastic leak
for neuron r. For this, the parameters of neuron l are set to
cl = 1, λl = +128, αl = 1, Ml = 0, and Rl = 0. Therefore,
neuron nl naturally spikes with probability p = λl /255 ≈ 0.5,
because it leaks sgn(λl ) = +1 with this same probability and
the threshold is set to αl = 1. After spiking, it is reset to
Vl = 0 and will present the same behavior in the next tick. If
we then connect the output of neuron l to an input axon (of
i
type i) of neuron r and set the memory position sG
r equal to
the leak value L, we will obtain the desired spiking behavior.
In the algorithm, the state of the RBM unit is equivalent
to that of the spiked variable. However, in the TrueNorth
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
implementation multiple spikes may be produced by sampling
neuron r during TS . A solution for this is to create a so-called
“refractory effect” using an additional neuron k. What this
additional neuron essentially does is count how many spikes
are received from neuron r. For this, neuron k is configured
with threshold αk = 1 and its membrane potential set at the
start of the sampling phase (i.e. at the same moment neuron
r is set to Vinit ) to, for example, −TS . By incrementing the
membrane potential by 1 for every received spike from neuron
r, the membrane potential of neuron k will only be larger than
−TS if at least one spike was received. Therefore, after TS
has expired, we inject an axonal event of +TS into neuron k,
which, due to the unit-valued threshold, will cause it to spike
if at least one spike was produced by neuron r during the
sampling phase.
In sum, the outset of a single RBM unit can be analyzed as
capturing the dynamical behavior of two coupled DTMCs run
for a time interval of TS , which are used to set a threshold flag
(spiked). Two TrueNorth neurons (r and l) are used to form
the DTMCs, along with a third neuron (k) used for verifying
if the threshold flag has been triggered, after TS has expired
(thus, the “refractory effect”). The combination of all of this
comprises an RBM unit.
As a final note, it was shown that the argument of the
logistic function (x in Eq. (2)) is modeled as the membrane
potential of the neuron. Since TrueNorth neural membrane
potential takes on only signed integer values, and the logistic
function has a dynamic range between approximately -6 and
+6, it is necessary to apply a multiplicative scaling factor, s,
to the RBM weights and biases to increase the dynamic range
of the neural logistic sampler realization. As a result of this
scaling, the neural sampler must be realized with appropriate
values of TS , Vth , M , and L to enable the RBM to sample with
high precision from the logistic probability distribution. The
ideal sampler (with s = 50) is compared with the TrueNorth
realization (TS = 8, stochastic threshold ranging from 79 to
590, and stochastic leak of 49) in Fig. 6.
1
P(spike | Vm)
0.8
0.6
Ts=8
Vth=79
M=9
L=49
0.4
Ideal sampler (scale=50)
TrueNorth sampler
0.2
0
−400
−300
−200
−100
0
Vm
100
200
300
7
implementation, is connected to all 784 visible units. A viable
solution is to use a patching scheme over the original image
[28], thus reducing the area of the image “observed” by each
hidden unit. Reciprocally, since the generative RBM presents
feedback from the hidden layer to the visible layer, the quantity
of hidden units should also be selected in a way as to reduce
the number of units “observed” by the visible units. Fig. 7a
shows how a patch (yellow) is formed by an 8x8 pixel window
over a binarized MNIST image.
(a)
(b)
Fig. 7: Sparsity structure in RBM. (a) Illustration of an 8x8
pixel patch (in yellow). (b) Sparsity can be seen as applying a
mask over the network’s weight matrix during offline training.
In [28], square patches of size p × p were randomly placed
over the input image, with all the visible units belonging to a
patch connected to a single hidden unit. Though this resulted
in reduced network connectivity, for a physical implementation with fan-in constraints a systematic patching scheme is
necessary to produce a well established maximum number
of units observed in each layer. The systematic patching is
particularly important for the generative model due to the
feedback from hidden to visible units during inference (details
of generative RBM operation are given in Section V). If, for
example, patching were performed randomly, a visible unit
could possibly be captured by more than 256 patches, making
this fan-in unfeasible on TrueNorth. Therefore, we applied an
overlapping, yet deterministic, patching scheme developed for
the generative RBM realization. The method uses patches with
p2 pixels which are formed by “sliding” a square window over
the N 2 -pixel image and forming a new patch at every new
position. The total number of overlapping patches produced
using this method is defined by:
patches = hidden units = (N − p + 1)2 .
(9)
400
Fig. 6: Logistic sampler using TrueNorth neurons.
C. Sparse connectivity
The all-to-all connectivity between layers in the RBM
algorithm implementation has to be adapted to the available
connectivity in hardware. The 256-input cores in TrueNorth
present a constraint for the 784-pixel images used in the
hand-written digits pattern completion application described
in this paper, where each hidden unit, in a standard RBM
A patching scheme implemented in this manner can be interpreted as applying a mask, Wmask , over the RBM’s weight
matrix, where 0’s and 1’s in the mask represent, respectively,
no connection and presence of connection between visible and
hidden units. The mask is applied during the offline RBM
training and the resulting sparse weight matrix is then used
for mapping the RBM onto TrueNorth. In Fig. 7b, the sum of
column values in each row of Wmask represents the number of
hidden units observed by each visible unit, and the sum of row
values in each column represents the number of visible units
observed by each hidden unit. With this systematic patching,
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
IV. Q UALITY METRICS OF DIGITAL NEURAL SAMPLER AND
SPARSE NETWORK
The following subsections present quantitative analyses of
the impact of the adaptations demanded during the mapping of
the original RBM algorithm onto TrueNorth. First, the impact
of quantization due to digital hardware data representation is
verified. Second, the effect of approximate logistic sampling
using the digital neural sampler is analyzed. Lastly, due to
non-viability of all-to-all connections between neurons in
TrueNorth, we analyze the impact of sparsity in network
connectivity. For the neuromorphic adaptations, the generative
performances are verified using the Kullback-Leibler (KL)
divergence and Annealed Importance Sampling (AIS), which
are briefly explained next.
Kullback-Leibler divergence. KL divergence is a measure
of the difference between probability distributions. The probability distribution of an RBM is defined by Eq. (1), with
the denominator of this equation (i.e. the partition function)
demanding a countable normalizing sum of all state probabilities for computation. Therefore, since we want to compare
the performance of the samplers versus exact probability
distributions (computed by Eq. (1)), only small networks with
tractable partition functions can be analyzed. KL divergence
is particularly important for our analysis of the digital neural
sampler and, though we cannot directly extrapolate values of
this measure to larger networks, the results aid in identifying
expected performance for each sampler. KL divergence is
defined by the following equation [29]:
DKL (P ||Q) =
X
i
P (i) log
P (i)
,
Q(i)
(10)
where P and Q are two probability distributions, and DKL
is always non-negative. The state of the system is defined
by i. For our experiments, P was defined as the distribution
obtained in the experiment and Q as the true distribution.
Annealed Importance Sampling. AIS is a metric used
to estimate the log-probability of a generative model [30],
[31], where larger values indicate higher likelihood that the
model generated the data. For high-dimensional models, such
as RBMs, where calculation of the partition function is intractable, the AIS algorithm is very useful as it performs
a stochastic estimation of the partition function to compute
the log probability of the model with respect to the data.
Therefore, the AIS algorithm will be used for validating the
generative performance of the sparsely connected network by
verifying the patch size which produces the largest AIS value.
A. Quality of data quantization
The effect of data quantization can be verified by comparing
the quantized samplers to the ideal. The weights and biases can
be quantized by realizing the following: multiply their values
by a scaling factor (s), then round the result to the nearest
integer, and finally divide the second result by s. The KL
divergence of the network with quantized versus exact (high
precision) weights was computed over 1000 experiments, each
consisting of randomly generated weights and biases for a
network with 5 visible and 5 hidden units. For these, based
on experimental results of weights and biases from previously
trained RBMs, the values were sampled from the following
normal distributions: weights ∼ N (−0.05, 1.6e − 3), visible
biases ∼ N (−0.3, 1), and hidden biases ∼ N (0.5, 2.25). The
KL divergence results of quantized versus non-quantized data
are shown in Fig. 8, including a box plot of KL divergence for
s = 15−100. A saturation point can be seen around s = 50. It
is important to note that very large values of s are beneficial for
the algorithm, however they can be costly in terms of hardware
resources (cores, in the case of TrueNorth), since more neurons
and longer accumulation times will be required for mapping
larger values of weights and biases (explained in Section V).
0.014
−3
x 10
Average KL divergence
the bounds of the sums (both in rows and columns) are welldefined.
8
0.012
3
0.01
2
0.008
0.006
1
0.004
0
0.002
0
15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Scaling factor (s)
Fig. 8: Generative performance versus quantization.
B. Quality of the digital neural sampler
The scaling factor impacts the resource usage of the system
and it also impacts the latency – by increasing the accumulation time (see Section VI). The other parameter which also
affects latency is the sampling time window (TS ). Thus, when
neural samplers present the same generative performance, the
selected configuration will naturally be the one presenting the
lowest TS . Additionally, when selecting the neuron parameters (the “sampler configuration”), one important aspect of
TrueNorth neurons that should be taken into account is the
membrane potential range. From Eq. (8c), we can observe that
the upper bound (i.e. positive saturation) value of membrane
potential is defined by the sum αj +η(Mj ). Therefore, for our
analysis, we have chosen configurations with this sum close
to or surpassing the upper bound of the dynamic range of the
sigmoid function (≈ 6× scaling factor) while still presenting
adequate sigmoid fitting.
To begin the neuron parameter selection, we first fix the
scaling factor, then the quality of the digital neural sampler
can be verified by sweeping over values of sampling time
window (TS ) and neuron parameters (Vth , M , and L) to,
ideally, overlap with the sigmoid. The best fit was found for
each TS value by performing a parameter search to reduce the
mean squared error (MSE) between the ideal (scaled) logistic
function and the curve produced by the neural sampler. For
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
our experiments, five configurations were chosen, with the
TrueNorth neuron parameters of the configurations (G1-G5)
shown in Table I. The MSE of each versus the ideal logistic
function is presented in the rightmost column.
Config.
G1
G2
G3
G4
G5
Scaling factor
50
50
50
50
50
TS
1
2
4
8
16
V th
0
0
66
79
186
M
7
8
8
9
9
L
125
100
77
49
36
MSE
0.4878
0.1311
0.0741
0.0412
0.0415
9
measures were not identical – for example, the MSE for G4
and G5 were basically identical, yet the KL divergence for
G5 showed a slight improvement –, the figure clearly shows
similar trends for both measures. Thus, these results indicate
that using the DTMC analysis of the sampler combined with
the MSE measure can be a powerful tool for quick access to
estimating the generative performance of a sampler. For the
generative RBM implementation on TrueNorth, configuration
G5 was chosen due to slightly better KL divergence results.
1
TABLE I: Neuron configurations for neural sampler analysis.
Normalized MSE
Normalized KL div.
0.8
0.6
−1
Average KL divergence
10
−2
10
−3
10
KL divergence @ sample=100k
0.1
0.08
G1
G2
G3
G4
G5
Ideal sampler
0.06
−4
10
0.04
0.02
0
G1
G2
G3
G4
G5
Ideal
4
10
5
Samples
10
Fig. 9: Generative performance of neural Gibbs samplers.
As was mentioned in the end of Section II, the DTMC
computations of the neural sampler can be very useful when
simulating the network dynamics. Instead of having to simulate every step of the neuron during the sampling time window
(TS ), we can simply use the spiking probability curve obtained
from the DTMC as the neuron’s transition operator. In other
words, the probability of spiking after TS can be extracted
from the curve and this value is then compared to a uniformlysampled number between 0 and 1. Though this does not affect
in any sense the operation of the neural sampler algorithm
(and cannot be used in practice), it speeds up simulations
considerably.
A comparison of the normalized (i.e. all values divided by
the worst case = model G1) MSE and KL divergence (at
sample 105 ) is shown in Fig. 10. Though the results for both
0.4
0.2
0
G1
G2
G3
G4
G5
Neuron parameter configuration
Fig. 10: MSE and KL divergence of neural Gibbs samplers.
C. Quality of the sparse network
Since sparsity is difficult to evaluate in small networks, the
generative qualities of the sparse RBM were verified by means
of the AIS measure of a network with 784 visible and 500
hidden units, pre-trained using the MNIST dataset. In a related
vein, reference [28] shows how a sparsely connected RBM can
produce a more noise-tolerant model for classification. For our
application, sparsity is actually necessary for reducing the fanin of each neuron, which, in TrueNorth, is limited by the 256input cores. The patching scheme proposed for the generative
RBM, described in Subsection III-C, takes into account the
feedback from hidden to visible units. With this method, as
illustrated in Fig. 7b, the patch dimension (p) defines the
maximum number of connected units in both directions (i.e.,
visible → hidden and hidden → visible).
AIS measure versus patch dimension results are shown
in Fig. 11. For low p values, lower log-probabilities were
produced on account of less information captured by each
patch. For large p values, the log-probability is also lower
on account of less number of hidden units (refer to Eq. (9))
in the network. Given the performance results of the model,
for the generative RBM implementation an optimal patch size
of 8×8 was chosen, resulting in (N -p+1)2 = (28-8+1)2 = 441
hidden units.
−75
Log probability of data
The generative model performance for these configurations
was determined by means of average KL divergence of the
model, and also the ideal logistic sampler (using Eq. (3)),
versus the true distribution (computed by Eq. (1)), over 10
randomly sampled networks (5 visible and 5 hidden units),
with 15 experiments run for each network, and each experiment consisting of 105 samples. Fig. 9 shows the average KL
divergence results of the different parameter configurations.
The smaller plot in this figure is a boxplot of the 150 (10
networks × 15 experiments per network) KL divergence
values at sample 105 . Naturally, the configurations with lower
MSE also presented lower KL divergence, with G3 and G4
practically overlapping.
−80
−85
−90
−95
−100
2
4
6
8
10
12
14
Patch dimension (p)
Fig. 11: Generative performance versus sparsity.
16
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
Ref.-and-splitter
incoming
spikes
ref
+
Quantization
Acc.-and-sample
(x.W + b)
h
(h.W T + bv)
+
+
h
...
input
+
h
outgoing
spikes
(a)
ref
10
hidden
v
v
v
3rd sample
visible
2nd sample
input
ref
(v.W + bh)
+
+
visible
1st sample
hidden
“corruption”
output
(b)
original
(c)
Fig. 12: TrueNorth RBM. (a) The 3-stage architecture used to distribute spike events (splitter), produce the desired membrane
potential, and realize the sigmoid sampling. (b) The generative model structure, formed by combining two 3-stage blocks
and including the feedback between layers. (c) Example of a pattern completion task where the digit “6” is incrementally
reconstructed.
V. G ENERATIVE RBM ARCHITECTURE ON T RUE N ORTH
The generative RBM was mapped on TrueNorth by developing a modular 3-stage architecture, where each combination
of these three stages represents the transition between RBM
layers. The diversity of configurable parameters present in
TrueNorth is critical to the realization, with particular neuron
types, connectivity strategies, and reset modes in each stage.
The physical constraints of TrueNorth – particularly 256 axons
and neurons per core, only 1 destination axon per neuron, and
4 distinct weights per neuron – defined the design flow of the
RBM. The architecture, composed of stages (1) refractoryand-splitter, (2) quantization and (3) accumulate-and-sample,
is illustrated in Fig. 12a.
The generative application implemented was a pattern completion task of a corrupted MNIST image. The signal flow
in the TrueNorth RBM is illustrated in Fig. 12b, where each
row is a 3-stage module and the blue and red blocks represent
information related to visible and hidden units, respectively.
Note the second stage in each module contains both colors,
since this is the transition between visible and hidden layers,
i.e. where the arguments of p(v|h) and p(h|v) are computed.
Finally, the data flow for the application is represented in
Fig. 12c. For this task, part of an image of the digit “6” (not
used during training) was removed, and the figure shows the
first 3 reconstructions based on the partial data.
A. Stage 1a: Splitter
Stage 1 serves a dual role in the system: (1) a splitter for
input signals in the respective RBM layer and (2) a refractory
effect of the neuron. Since each RBM unit in the visible/hidden
layer is connected to multiple units in the hidden/visible layer,
along with the fact that TrueNorth neurons present only one-toone connections (i.e. each neuron can only target a single axon
on the entire chip), a signal splitter is necessary to create the
RBM’s one-to-many connections. Therefore, stage 1 generates
the required number of replicas of an RBM unit to be used
in the quantization stage. Fig. 13a illustrates a splitter core
used for generating the necessary number of replicas of each
of the visible units. The neurons are set to unit thresholds and
all synaptic connections are of weight equal to +1, which will
cause the neurons to spike whenever an axon event arrives.
The refractory effect function of this stage and the two control
signals (C+ and C− ) are discussed later in Subsection V-D.
B. Stage 2: Quantization
In TrueNorth, the weight of connections between axons and
neurons can be configured with two constraints: the weights
between axons connected to a given neuron are allowed to
have only 4 different values; and each axon is configured as
one of 4 types, reflecting on which of the 4 weights will be
used for the connection between the axon and the respective
neuron [27]. The first constraint limits the number of different
possible weights, while the second limits the “reutilization”
of axons between neurons. This is because an axon can be
used amongst two neurons only if the weight stored in each
neuron’s memory position – defined by the axon type – is
the desired synaptic weight for each of these connections.
Several methods proposing the usage of low-precision weights
and biases in artificial neural networks have been developed
[32], [33], however these methods target only discriminative
models. As was observed in Figure 8, a large scaling factor (i.e.
high precision) is critical for obtaining satisfactory generative
performance in RBMs. Since the precision and diversity of
weights and biases demanded by the generative RBM cannot
be directly represented by the TrueNorth memory structure,
a quantization stage is therefore necessary to realize the
connectivity between RBM units.
The representation of individual RBM weights and biases
was achieved by using a collection of neurons in stage 2, each
with its own weight, which together can produce the desired
membrane potential (i.e. the equivalent argument of σ(x)). For
this, linear-reset, unit-threshold neurons are used [27], and they
operate by decrementing their membrane potential by 1 every
time they spike, continuing to do so while the value is above
zero. In this manner, the collective activity of many stage 2
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
v1
v2
v3
C+
C-
+1
+1
...
+1
+1
+1
+1
C+ C+ C+ C+ C+ C+
C-
C-
v1 v2
C-
C-
C-
v2 v3 v3
C-
v3 ...
v1
v2
v3
bh1
11
h1+7
h1-8
h1-4
+7
+8
+4
+2
...
+8
+1
-1
-1
...
+2
L
L
L
...
h1+7 h1-8 h1-4 h1-2 h1+8 h1+2
(a)
...
h1 h2 h3 leak leak leak
(c)
(b)
Fig. 13: Example of TrueNorth RBM stages: (a) Refractory-and-Splitter, (b) Quantization, and (c) Accumulate-and-Sample.
neurons encodes the RBM weight/bias, while stage 3 will be
used to accumulate the spikes from these many neurons into
a single neuron.
The quantization of weights and biases is done by selecting
a maximum accumulation time (TA ), which will be the largest
value of membrane potential a stage 2 neuron can reach.
In other words, every input spike into stage 2 axons will
charge the membrane potential of each quantization neuron
up to at most TA , after which they will freely operate, with
spiking activity guaranteed to cease in a maximum of TA ticks.
Fig. 13b exemplifies a stage 2 core with TA =8 and visible units
v1 , v2 , and v3 connected to hidden unit h1 with weights +7,
-12, and -2, respectively. Since TA is a user-defined value,
intuitively we would select the lowest value possible as to
reduce the overall latency of the system. However, depending
on the number of weights to be mapped and their specific
values, attempting to use smaller values of TA will exceed the
number of available neurons in a core. Note that the sign for
the negative weights is actually positive, for stage 2 only takes
into account the intensity (absolute value) of the connection
between units, independent of being excitatory or inhibitory.
The actual sign of the connection is taken care of in stage 3.
Lastly, since bias values are independent of neuronal activity,
these are realized by sending an external spike event to the
bias axon (bh1 =10 in Fig. 13b) each time the sum of inputs
to a given RBM unit neuron is to be computed.
C. Stage 3: Accumulate-and-Sample
Stage 3 is used to accumulate the activity of the quantization
neurons into a single neuron, which will then be sampled
(as described in Subsection III-B). Prior to accumulation, the
membrane potentials of the stage 3 neurons are initialized
to zero. Then, during the first time window (TA ), stage 3
neurons accumulate spikes from stage 2 neurons to form a
membrane potential equivalent to the argument of the logistic
function. The neurons used in stage 3 have a non-resetting
property to prevent clearing the membrane potential during the
accumulation phase. This is followed by the time window TS ,
during which the stochastic threshold and leak properties of
the neuron are used for sampling from the logistic probability
distribution. During this second time window, if the neuron’s
membrane potential surpasses the threshold, the neuron may
spike multiple times since it is configured as non-resetting.
For the spikes to correctly represent a sample from the
logistic function, the refractory stage is necessary to register
a maximum of 1 spike event per sampling window, and is
described in the next subsection.
An example of a stage 3 core crossbar configuration is
shown in Fig. 13c, with the sign of the RBM weight/bias now
included in the synaptic weights. Note the use of recurrent
connections from additional neurons to realize the stochastic
leak. These additional neurons are necessary because an internally generated stochastic leak (for example, in neurons h1 ,
h2 , and h3 ) can only assume an absolute value of 1. Since our
digital neural sampler implementation usually demands larger
values of L, the “leak” neurons were created with threshold
of 1 and internal stochastic leak sampled from a Bernoulli
distribution with p ≈ 0.5 (refer to Subsection III-B). In this
manner, there is approximately 50% chance of these “leak”
neurons spiking at each tick, thus generating a spike event to
their respectively associated neuron (h1 , h2 , etc.), which can
be connected with a user-defined synaptic weight of L > 1.
D. Stage 1b: Refractory effect
As was discussed in Subsection III-B, for the multiple
spikes from the accumulate-and-sample stage to be converted
to a single spike event – which represents a sample from
the logistic probability distribution –, stage 1 neurons were
configured to produce a “refractory effect”. What essentially
occurs in stage 1 is a delayed propagation of the spiked
variable (in the digital neural sampler algorithm in Subsection
II-A), whose value is dependent on spikes from the previous
RBM layer’s stage 3. This delayed response after TS has
expired, therefore, results in a “frame alignment” (in the same
1 ms time step) of RBM unit samples to subsequent layers and
guarantees precise operation of the generative RBM algorithm.
The refractory effect is obtained in TrueNorth by configuring stage 1 splitter neurons with a negative saturating
membrane potential. The membrane potential of stage 1 neurons are initialized to the negative saturating value C− , with
|C− | ≥ TS , at the start of the sampling phase of stage 3
in the other RBM layer. Every incoming spike in stage 1
will cause the membrane potential of its associated neuron to
increase by 1. After TS , the membrane potential of the stage 1
neurons are incremented by C+ (= |C− |), causing the neurons
which received at least one spike to cross the threshold and
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
simultaneously generate (“frame alignment”) a spike to the
subsequent stage 2.
VI. S PIKE PROCESSING FLOW IN T RUE N ORTH RBM
An example of the spiking activity flow between RBM
layers is shown in Fig. 14, where the following parameters
were used: TS =10, TA =8, C− =-30, stage 3 stochastic threshold
ranging between 10 and 17 and stochastic leak of +3. In the
example, the x-axis denotes time (in 1 ms ticks) and the y-axis
denotes the value of the membrane potential (Vmem ). The blue
line is the neuron’s membrane potential, the solid red line is
the saturation level, the dashed red line is the threshold, and
the red circles represent spike events.
Vmem
Stage 1
c
a
Stage 2
Stage 3
d
b
c
e
f
g
time
Fig. 14: Example of spike processing flow. Stage 1 realizes two
functions: refractory effect in (a) and (b), and splitter (spike
distribution) in (c). Stage 2 quantizes the weights between
RBM layers in (c) and (d), producing the desired membrane
potential for stage 3 to sample from. Stage 3 accumulates the
spikes from stage 2 linear-reset neurons between (e) and (f),
and the sampling procedure is performed between (f) and (g).
The sequence of events (the letters) presented in Fig. 14 are
detailed below:
(a) Time=2. Stage 1 neurons are initialized to C− , after
which they begin accumulating spikes from stage 3 neurons
of the other RBM layer. (b) Time=10. After TS , the C+
signal is applied, and every neuron which captured at least
one spike from stage 3 neurons crosses the threshold=1.
(c) Time=11. As C+ is applied, spike events from stage 1
neurons are transmitted to stage 2 axons. In this example,
the stage 2 neuron is charged to a membrane potential of 6.
(d) Time=12-17. The linear-reset stage 2 neurons continuously
produce spike events to stage 3 axons until their membrane
potentials return to zero. (e) Time=13. At this moment, the
stage 3 neurons begin accumulation for TA ticks. (f) Time=21.
After the stage 3 neurons have accumulated their membrane
potentials to the desired values, the sampling phase begins.
The stage 1 neurons of the other RBM layer are initialized
to C− ; the stochastic threshold and leak come into effect at
stage 3. (g) Time=31. After TS ticks, the stage 3 neurons are
reinitialized.
The example shows the complete sampling procedure of
an RBM layer in the TrueNorth implementation. In stage 2,
weights and biases are converted from membrane potential
values to spikes, which are accumulated in stage 3 until
the appropriate membrane potential is formed (i.e. has been
grouped into a single neuron) at the start of the sampling
phase. The two TrueNorth neurons in stage 3 comprise the
coupled DTMCs used in the neural sampler. The stage 1 neurons produce the delayed spike response (“refractory effect”)
12
in the subsequent RBM layer, which constitutes a sample from
an RBM unit. Therefore, in this example, a new sample is
produced in an RBM layer at every (TA + TS + 2) = 20 ticks;
2 additional ticks are necessary for control signals. The entire
process of producing a new sample of the visible units – the
output of the generative RBM – would then take 2 × 20 = 40
ticks (= 0.04 seconds).
VII. D ESIGN AUTOMATION
TrueNorth system configuration can be realized using the
object-oriented Corelet Language, which is an abstraction for
representing the network of neurosynaptic cores [34]. The
developed design automation procedure consists of creating
systematic data structures, originating from the RBM weight
and mask matrices, RBM biases, and user-defined parameters,
which include: accumulation time (TA ); sampling time (TS );
data scaling factor (s); and sampler stochastic threshold
and leak. Once these have been defined, the automation
procedure produces an optimal configuration of cores which
minimizes the number of axons and neurons used for the
RBM realization. Three optimization strategies were created,
where strategies 1.1 and 1.2 are mutually exclusive, yet
they can be combined with strategies 2 and 3. Note that all
considerations for hidden units are also valid for visible units.
Strategy 1.1: The first strategy involves establishing the
number of neurons required for mapping each RBM weight
and bias. Without optimization in stage 2, the number of
neurons nj used when quantizing the weight between the
visibleP
units observed by hidden unit hj can be computed by
nj = i dwji /TA e. This direct method of mapping weights
and biases does not take into account the fact that possibly
many stage 2 neurons present low weights, which will cause
them to complete spiking (during the accumulation phase)
before neurons which represent higher values, such as weight
TA . Since the network must always go through TA ticks during
the accumulation phase, it would be more efficient to try to
connect a given neuron to as many possible axons, provided
the total synaptic weight is guaranteed not to exceed TA . In
the limiting case, neurons which map weights -1 and +1 can
have up to TA axons connected to them.
Though this first strategy benefits the core utilization
considerably, better optimizations are possible. This is
because the order in which the RBM weights are chosen to
be mapped in stage 2 is defined by the user, yet different
mapping sequences may utilize less cores. For example,
suppose TA =4 and the weights to be mapped are 1 through
6 for visible units v1 through v6 , respectively. If we were
to map them in this order, a total of 6 neurons would be
used (Fig. 15a). On the other hand, if we were to map in the
reverse order (6 through 1), a total of 7 neurons would be
necessary (Fig. 15b). Therefore, the order of weight mapping
affects the core utilization. Since the possible number of
weight orderings to be analyzed is intractable, better results
can be obtained by using strategy 1.2.
Strategy 1.2: In this strategy, the weights closest to a userdefined central weight value are mapped first. By sweeping
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
+4 +2 +3 +4 +3 +2 +1
+1 +1 +1 +1 +2 +2
v6
v5
v4
v3
v2
v1
v1
v2
v3
v4
v5
v6
3, the algorithm does the same as for stage 1, now selecting
the unit which uses the most number of axons (quantization
neurons from stage 2) without exceeding the axon limit in the
core.
VIII. R ESULTS
h1+1 h1+1 h1+1 h1+1 h1+2 h1+2
h1+4 h1+2 h1+3 h1+4 h1+3 h1+2 h1+1
(a)
(b)
Fig. 15: Strategy 1.1 examples of stage 2 quantization.
through all possible central weights, an optimal value can be
empirically obtained. Fig. 16 shows the number of neurons
used when mapping the weights -20 through 20 with TA = 4.
The red line is the number of neurons (120) obtained with no
optimization, while the black line is the number (110) when
using sequential mapping with weight neuron “reutilization‘’
(i.e. strategy 1.1). The blue line shows the results for the
central weight method (strategy 1.2). The reduction from
110 to 107 neurons when using a central weight of 5, for
example, is small (approximately 3%), though more significant
reductions are possible when this strategy is combined with
strategies 2 and 3.
118
No optimization
Strategy 1.1
Strategy 1.2
116
114
For realizing the generative model – the MNIST pattern
completion task detailed in Section V – on TrueNorth, an
RBM with 784 visible units and 441 hidden units (generated
by using 8×8 patches) was trained offline using the persistent Contrastive Divergence algorithm [35]. The generative
application demands a sampler with high fidelity with respect
to the ideal sampler. To achieve this, the parameters were
selected according to the criteria outlined in Section IV:
scaling factor = 50, TS =16, stochastic leak = 36, and stochastic
threshold ranging from 186 to 697. The choice of the scaling
factor directly impacts the RBM weight and bias magnitudes.
To map these weights in stage 2, a trade-off is necessary
between the accumulation time and the quantity of neurons and
cores demanded by the application. Therefore, a compromise
value of TA =32 was selected for the mapping. With these
parameters, a new RBM image is sampled at every 100 ticks
(= 0.1 seconds).
A. Resource utilization and power estimate
Using the automation strategies outlined in Section VII, the
generative RBM was realized with 865 cores, representing
21% of the total number of cores on TrueNorth. Table II
shows how applying the optimization strategies 1.2, 2 and 3
drastically reduced the core utilization.
120
Neurons used
13
112
110
108
106
−20
−15
−10
−5
0
5
10
15
20
Central weight value
Fig. 16: Optimization strategy comparison in terms of number
of neurons used to map the desired weights.
Strategy 2: One final optimization can be performed
in stage 2. When a hidden unit is mapped, the number
of remaining neurons in the core may be enough to map
additional hidden units. If this is the case, among the hidden
units to be mapped, we select the one which has the most
number of visible units in common with the hidden units
previously mapped in the given core. This is because the
patching scheme produces hidden units which may have some
visible units in common and, thus, specific units are capable
of sharing axons. This strategy is naturally also valid when
mapping visible units.
Strategy 3: Stages 1 and 3 can also be optimized via a
greedy optimization method. For stage 1, the algorithm selects
the unit which uses the most number of neurons (replicas for
axons in stage 2), yet does not exceed the neuron limit in
the core. If no unit can be mapped in the core, the algorithm
creates a new one, until all units have been mapped. For stage
Case
1
2
3
Strategies
none
1.1, 2, 3
1.2, 2, 3
# of cores
2956
906
865
Chip utilization
72.2%
22.1%
21.1%
TABLE II: Core utilization results based on optimizations.
In Figure 13, it was shown how each RBM unit is actually
formed by 3 TrueNorth neurons (stochastic leak, stochastic
threshold, and “refractory effect”). However, the final implementation of the 784 + 441 = 1225 RBM units consisted of
865 cores, with a total of 135k mapped TrueNorth neurons.
This number is mainly due to the splitters and to the stages
needed for weight and bias quantization, representing 82% of
the total neurons used. Virtually all of the remaining neurons
were used for the control signals for the system operation,
while the digital neural samplers – which represent the RBM
units per se – used up only 0.5% of the mapped TrueNorth
neurons. In practice, this results in a ratio of 110 TrueNorth
neurons required to implement each RBM unit, and shows how
generative models implemented on high dimensional datasets
incur a considerable overhead due to the aforementioned
hardware constraints. Nonetheless, given the network size (784
+ 441 RBM units), image patch size (p = 8), and accumulation
(TA = 32) and sampling (TS = 16) times, we conservatively
estimate a power consumption of 5 mW for the optimized
TrueNorth generative RBM (case 3 in Table II). This results
in an estimated 0.5 mJ of energy consumed to generate each
MNIST image sample.
B. Pattern completion outputs and metrics
Example outputs of the pattern completion task are shown
next. In Fig. 17a, one example output for each of the ten
digits is presented: the first column is the original data (“O”),
the middle column is the corrupted (“C”) image sent into the
TrueNorth RBM, and the third column is the reconstructed
(“R”) output after 50 RBM samples (= 5 seconds). These images were chosen to represent positive results, while Fig. 17b
shows images whose reconstruction was not ideal. Lastly, Fig.
17c illustrates a sequence of reconstructions for a corrupted
image of the digit “6”; the sample number is indicated above
each image. A decent reconstruction sample could be obtained
after about 4 seconds; however, with an earlier sample we
could possibly confuse the “6” with a “5”.
O
C
R
O
C
R
O
C
Normalized Hamming distance
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
14
0.2
Ideal sampler
Neural sampler
0.15
0.1
0.05
0
85% 75% 65% 55% 45% 35% 25% 15%
Image corruption (occlusion)
Fig. 18: Sampler generative performance analysis in terms of
incorrectly generated pixels (Hamming distance).
expenditure demanded for the generative task of MNIST image
reconstruction.
R
Normalized Hamming distance (with 35% occlusion)
mean @ sample 50
0.115
0.11
0.105
(a)
0
5
10
15
0.1
(b)
20
25
30
35
40
45
50
5
10
15
20
Reconstruction sample
25
30
Fig. 19: Normalized Hamming distance during reconstruction
of 35% occluded image on the TrueNorth RBM.
(c)
Fig. 17: TrueNorth pattern completion task outputs. Positive
(a) and negative (b) reconstruction results. Reconstruction of
the digit “6” in (c), with the sample number indicated above
each image.
Depending on the percentage of image occlusion (“corruption”), the RBM may or may not be able to reconstruct a
satisfactory representation of the original image. Therefore,
we performed experiments with different image occlusion
percentages and measured Hamming distance (HD) – identical
to the number of incorrectly reconstructed pixels in this case –
at the 50th RBM reconstruction sample for 1,000 test images.
The mean value of the HDs was normalized according to
the number of non-occluded pixels in the image. The results,
illustrated in Fig. 18, show that the reconstructive performance
of the neural sampler nearly matches that of the ideal sigmoid
sampler.
Lastly, the mean HD for the TrueNorth RBM can be
verified throughout the reconstruction process. In Fig. 19,
convergence to the mean HD value for 35% image occlusion
(dashed black line) occurs after about 10 RBM reconstruction
samples. This result is important to define the practical time
IX. C ONCLUSIONS AND FUTURE WORK
In this work, we have shown the first generative RBM implementation on neuromorphic hardware. For this, we followed
a step-by-step procedure for producing the Gibbs sampling
kernel – the sigmoidal spiking probability – using digital
spiking neurons and for mapping the generative RBM algorithm onto a digital neuromorphic VLSI substrate. The neural
sampler is an elegant solution as it uses bio-inspired dynamics
to simultaneously incorporate the logistic function look-up and
the comparison with a randomly generated number, which
together represent a Gibbs sample. A discrete-time Markov
chain (DTMC) analysis of the neural sampler was performed,
resulting in a simplified method of obtaining the spiking probability without the need for long neuron behavior simulations.
The generative performance of the neuromorphic adaptations
were then verified using the Kullback-Leibler (KL) divergence
and the Annealed Importance Sampling (AIS) algorithm. We
also showed how mean squared error (MSE), along with the
DTMC, can be used as an efficient method for obtaining
insight into the sampler quality.
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
In the TrueNorth system, we followed a systematic development and implementation process of a modular architecture,
which can be used for realizing generative RBMs and DBNs
on a substrate of digital neurosynaptic cores. The 3-stage
architecture and the design automation procedure provide a
path towards automated neural network applications on braininspired processors for more complex inference tasks, such
as natural image recognition and time series generation. The
modular characteristic of the architecture naturally lends itself
to implementations of deeper networks (DBNs). Also, the
architecture of stages 1 and 2 with the associated design
automation procedure can even be used to realize other neural
networks which are defined by sparse weight matrices. We are
currently working on new algorithms which incorporate more
of the hardware constraints during the training phase.
The developed architecture avails many of the features
present in neuromorphic systems. The spike flow of the 3-stage
architecture developed for the TrueNorth RBM uses spikes
for communication between cores, propagating information
between RBM layers. The number of computations is also
reduced in the neuromorphic scenario as only non-zero multiplications are performed (i.e. only when a spike occurs does
data processing take place), which is contrary to what occurs
traditionally for matrix multiplications in CPUs. Additionally,
the sampler makes use of stochastic neural properties to produce an approximate sigmoidal firing probability, necessary for
the RBM sampling procedure. Despite these positive features,
information processing in the network is somewhat sequential
(i.e. basically two stages are being used at each instant),
which is mainly a result of the limited weight values per
neuron in the present hardware. Inspired by the sampling
methods proposed in [36], [37], we are currently developing
paths towards algorithms on TrueNorth which incorporate the
hardware constraints yet present a more continuous flow of
spike processing for inference.
As a final note, research proposing RBMs and DBNs
as solutions to applications of BCIs and EEG classification
generally focuses on discriminative models [5], [6], [38].
However, BCIs could naturally benefit from generative models,
targeting applications such as time series EEG or neural signal
reconstruction for artificial limb control. An attractive feature
of spike-based neuromorphic processors for spike-based neural
interfaces would be the direct match between the event-driven
data formats of the artificial and biological neuronal networks
at the interface, potentially obviating the need for extra signal
processing to convert between spiking and mean-rate representations, and possibly allowing to exploit the inherent temporal
code of neuronal spike recordings or pulsed stimulation for
further improvements in BCI performance.
ACKNOWLEDGMENT
The authors would like to thank all the members of the
Brain-Inspired Computing Group at the IBM Almaden Research Center and also the participants in the NSF Telluride
Neuromorphic Cognition Engineering Workshop for their interaction and collaboration.
15
R EFERENCES
[1] H. Larochelle and Y. Bengio, “Classification using discriminative restricted Boltzmann machines,” in Proceedings of the 25th international
conference on Machine learning, pp. 536–543, ACM, 2008.
[2] H. Lee, P. Pham, Y. Largman, and A. Y. Ng, “Unsupervised feature learning for audio classification using convolutional deep belief networks,”
in Advances in neural information processing systems, pp. 1096–1104,
2009.
[3] G. E. Dahl, D. Yu, L. Deng, and A. Acero, “Context-dependent pretrained deep neural networks for large-vocabulary speech recognition,”
Audio, Speech, and Language Processing, IEEE Transactions on, vol. 20,
no. 1, pp. 30–42, 2012.
[4] G. W. Taylor, G. E. Hinton, and S. T. Roweis, “Modeling human
motion using binary latent variables,” Advances in neural information
processing systems, vol. 19, p. 1345, 2007.
[5] D. Wulsin, J. Gupta, R. Mani, J. Blanco, and B. Litt, “Modeling
electroencephalography waveforms with semi-supervised deep belief
nets: fast classification and anomaly measurement,” Journal of neural
engineering, vol. 8, no. 3, p. 036015, 2011.
[6] X. An, D. Kuang, X. Guo, Y. Zhao, and L. He, “A Deep Learning method
for classification of EEG data based on motor imagery,” in Intelligent
Computing in Bioinformatics, pp. 203–210, Springer, 2014.
[7] G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of
data with neural networks,” Science, vol. 313, no. 5786, pp. 504–507,
2006.
[8] S. Haykin, Neural Networks and Learning Machines (3rd Edition).
Prentice Hall, 2008.
[9] G. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning algorithm for
deep belief nets,” Neural computation, vol. 18, no. 7, pp. 1527–1554,
2006.
[10] G. Indiveri, E. Chicca, and R. Douglas, “A VLSI array of low-power
spiking neurons and bistable synapses with spike–timing dependent
plasticity,” IEEE Transactions on Neural Networks, vol. 17, pp. 211–
221, Jan 2006.
[11] M. M. Khan, D. R. Lester, L. A. Plana, A. Rast, X. Jin, E. Painkras,
and S. B. Furber, “Spinnaker: mapping neural networks onto a
massively-parallel chip multiprocessor,” in Neural Networks, 2008.
IJCNN 2008.(IEEE World Congress on Computational Intelligence).
IEEE International Joint Conference on, pp. 2849–2856, Ieee, 2008.
[12] S. Mitra, S. Fusi, and G. Indiveri, “Real-time classification of complex
patterns using spike-based learning in neuromorphic vlsi,” Biomedical
Circuits and Systems, IEEE Transactions on, vol. 3, no. 1, pp. 32–42,
2009.
[13] A. Basu, S. Ramakrishnan, C. Petre, S. Koziol, S. Brink, and P. E.
Hasler, “Neural dynamics in reconfigurable silicon,” Biomedical Circuits
and Systems, IEEE Transactions on, vol. 4, no. 5, pp. 311–319, 2010.
[14] C. Zamarreño-Ramos, A. Linares-Barranco, T. Serrano-Gotarredona,
and B. Linares-Barranco, “Multicasting mesh aer: a scalable assembly
approach for reconfigurable neuromorphic structured aer systems. application to convnets,” Biomedical Circuits and Systems, IEEE Transactions
on, vol. 7, no. 1, pp. 82–102, 2013.
[15] B. V. Benjamin, P. Gao, E. McQuinn, S. Choudhary, A. R. Chandrasekaran, J.-M. Bussat, R. Alvarez-Icaza, J. V. Arthur, P. Merolla,
K. Boahen, et al., “Neurogrid: A mixed-analog-digital multichip system
for large-scale neural simulations,” Proceedings of the IEEE, vol. 102,
no. 5, pp. 699–716, 2014.
[16] P. A. Merolla, J. V. Arthur, R. Alvarez-Icaza, A. S. Cassidy, J. Sawada,
F. Akopyan, B. L. Jackson, N. Imam, C. Guo, Y. Nakamura, et al., “A
million spiking-neuron integrated circuit with a scalable communication
network and interface,” Science, vol. 345, no. 6197, pp. 668–673, 2014.
[17] C. Mayr, J. Partzsch, M. Noack, S. Hanzsche, S. Scholze, S. Hoppner, G. Ellguth, and R. Schuffny, “A biological-realtime neuromorphic
system in 28 nm cmos using low-leakage switched capacitor circuits,”
Biomedical Circuits and Systems, IEEE Transactions on, vol. PP, 2015.
[18] S. Geman and D. Geman, “Stochastic relaxation, gibbs distributions,
and the bayesian restoration of images,” Pattern Analysis and Machine
Intelligence, IEEE Transactions on, no. 6, pp. 721–741, 1984.
[19] P. Smolensky, “Information processing in dynamical systems: Foundations of harmony theory,” D. E. Rumelhart, J. L. McClelland, and PDP
Research Group (Eds.), Parallel distributed processing, vol. 1, pp. 194–
281, 1986.
[20] M. Tommiska, “Efficient digital implementation of the sigmoid function
for reprogrammable logic,” in Computers and Digital Techniques, IEE
Proceedings-, vol. 150/6, pp. 403–411, IET, 2003.
THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE TBIOCAS FOR REVISION IN OCTOBER 2015
[21] A. Tisan, S. Oniga, D. MIC, and A. Buchman, “Digital implementation
of the sigmoid function for FPGA circuits,” ACTA TECHNICA NAPOCENSIS Electronics and Telecommunications, vol. 50, no. 2, p. 6, 2009.
[22] K. Lakshmi and M. Subadra, “A survey on FPGA based MLP realization
for on-chip learning,” Int. J. Sci. Eng. Res, pp. 1–9, 2013.
[23] S. Das, B. Pedroni, P. Merolla, J. Arthur, A. Cassidy, D. Modha,
G. Cauwenberghs, and K. Kreutz-Delgado, “Gibbs Sampling with LowPower Spiking Digital Neurons,” in Circuits and Systems (ISCAS), 2015
IEEE International Symposium on, IEEE, 2015. Manuscript submitted
for publication.
[24] J. G. Kemeny and J. L. Snell, Finite markov chains, vol. 356. van
Nostrand Princeton, NJ, 1960.
[25] A. Clark and S. Gilmore, “Terminating passage-time calculations on uniformised markov chains,” in Proceedings of the Twenty-Fourth annual
UK Performance Engineering Workshop, pp. 64–75, Citeseer, 2008.
[26] G. Latouche and V. Ramaswami, Introduction to matrix analytic methods
in stochastic modeling, vol. 5. Siam, 1999.
[27] A. S. Cassidy, P. Merolla, J. V. Arthur, S. K. Esser, B. Jackson,
R. Alvarez-icaza, P. Datta, J. Sawada, T. M. Wong, V. Feldman,
A. Amir, D. B. dayan Rubin, E. Mcquinn, W. P. Risk, and D. S.
Modha, “Cognitive computing building block: A versatile and efficient
digital neuron model for neurosynaptic cores,” in International Joint
Conference on Neural Networks (IJCNN). IEEE, 2013.
[28] Y. Tang and C. Eliasmith, “Deep networks for robust visual recognition,” in Proceedings of the 27th International Conference on Machine
Learning (ICML-10), pp. 1055–1062, 2010.
[29] T. M. Cover and J. A. Thomas, Elements of information theory. John
Wiley & Sons, 2012.
[30] R. M. Neal, “Annealed Importance Sampling,” Statistics and Computing,
vol. 11, no. 2, pp. 125–139, 2001.
[31] R. Salakhutdinov and I. Murray, “On the quantitative analysis of deep
belief networks,” in Proceedings of the 25th international conference on
Machine learning, pp. 872–879, ACM, 2008.
[32] L. K. Muller and G. Indiveri, “Rounding methods for neural networks
with low resolution synaptic weights,” arXiv preprint arXiv:1504.05767,
2015.
[33] E. Stromatias, D. Neil, M. Pfeiffer, F. Galluppi, S. B. Furber, and
S.-C. Liu, “Robustness of spiking deep belief networks to noise and
reduced bit precision of neuro-inspired hardware platforms,” Frontiers
in neuroscience, vol. 9, 2015.
[34] A. Amir, P. Datta, W. P. Risk, A. S. Cassidy, J. A. Kusnitz, S. K.
Esser, A. Andreopoulos, T. M. Wong, M. Flickner, R. Alvarez-Icaza,
et al., “Cognitive computing programming paradigm: a corelet language
for composing networks of neurosynaptic cores,” in Neural Networks
(IJCNN), The 2013 International Joint Conference on, pp. 1–10, IEEE,
2013.
[35] T. Tieleman, “Training restricted Boltzmann machines using approximations to the likelihood gradient,” in Proceedings of the 25th international
conference on Machine learning, pp. 1064–1071, ACM, 2008.
[36] P. O’Connor, D. Neil, S.-C. Liu, T. Delbruck, and M. Pfeiffer, “Realtime classification and sensor fusion with a spiking deep belief network,”
Frontiers in neuroscience, vol. 7, 2013.
[37] E. Neftci, S. Das, B. Pedroni, K. Kreutz-Delgado, and G. Cauwenberghs, “Event-driven Contrastive Divergence for spiking neuromorphic
systems,” Frontiers in Neuroscience, vol. 7, p. 272, 2013.
[38] D. Wulsin, J. Blanco, R. Mani, and B. Litt, “Semi-supervised anomaly
detection for EEG waveforms using deep belief nets,” in Machine Learning and Applications (ICMLA), 2010 Ninth International Conference on,
pp. 436–441, IEEE, 2010.
16
| 9cs.NE
|
Classifying local Artinian Gorenstein algebras
Joachim Jelisiejew∗
October 13, 2016
arXiv:1511.08007v3 [math.AC] 12 Oct 2016
Abstract
The classification of local Artinian Gorenstein algebras is equivalent to the study of orbits
of a certain non-reductive group action on a polynomial ring. We give an explicit formula
for the orbits and their tangent spaces. We apply our technique to analyse when an algebra
is isomorphic to its associated graded algebra. We classify algebras with Hilbert function
(1, 3, 3, 3, 1), obtaining finitely many isomorphism types, and those with Hilbert function
(1, 2, 2, 2, 1, 1, 1). We consider fields of arbitrary, large enough, characteristic.
MSc classes: 13H10 (Primary), 13N10, 14C05 (Secondary)
Contents
1 Introduction
1
2 Preliminaries and theoretical results
2.1 Power series ring S and its dual P . . . . . . . . . . . .
2.2 Automorphisms and derivations of the power series ring
2.3 Local Artinian Gorenstein algebras . . . . . . . . . . . .
2.4 Characteristic zero . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
4
4
6
8
11
3 Applications
3.1 Basic examples and tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Compressed algebras and generalisations . . . . . . . . . . . . . . . . . . . . . . .
3.3 Further examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
12
15
20
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Introduction
The problem of classifying Artinian Gorenstein algebras up to isomorphism is hard and long
studied, a comprehensive reference is [IK99]. The results usually rest on additional assumptions: small length [Poo08a, CN07, CN09, Cas10, CN11] or being a codimension two complete
intersection [Bri77, EV11, EH16]. See also [Iar81, ER93, EI14, FK12] for other approaches.
Let k be a field. So far we do not impose any conditions such as k being algebraically closed
or on its characteristic (see Section 3), however the reader may freely assume that k = C.
Let (A, mA , k) be a local Artinian Gorenstein k-algebra. As the notation suggests, throughout
the paper we assume that k → A/mA is an isomorphism. This assumption is automatic if k is
algebraically closed. It is standard and it is used to discard examples such as k = R ⊂ A = C.
An important numerical invariant of A is the Hilbert function HA (i) = dimk miA /mi+1
A . The
socle degree d is the maximal number L
such that HA (d) 6= 0. If HA (d − i) = HA (i) for all i, then
the associated graded algebra gr A = i>0 miA /mi+1
A is also Gorenstein, see [Iar94, Proposition
1.7].
We may now formulate our motivating problems:
∗
The author is a doctoral fellow at the Warsaw Center of Mathematics and Computer Science financed by the Polish program KNOW. This article is supported by Polish National Science Center, project
2014/13/N/ST1/02640. Contact information: Joachim Jelisiejew, Faculty of Mathematics, Informatics and Mechanics, University of Warsaw. Email: [email protected]
1
1. What is the classification up to isomorphism of local Artinian Gorenstein algebras with
Hilbert function HA ?
2. If HA is symmetric, what are the sufficient conditions for A ≃ gr A?
The first of these problems naturally leads to the second: suppose we are classifying A whose
Hilbert function HA is symmetric. Since gr A is in this case Gorenstein, one natural way is to
first determine possible gr A and then classify A having fixed gr A. Note that it is convenient to
classify graded Gorenstein algebras by using ideas from projective geometry: Waring, border and
smoothable ranks, and secant varieties, see [Ger96, IK99] for the overview of methods and [LO13,
BB14] for the state of the art on secant varieties. See [MZ15, Sta96] and references therein for
information about the possible Hilbert functions. See [Iar94, Appendix] for a list of possible
Hilbert functions for small values of dimk A.
Our main goal is to refine, simplify and make explicit the existing theory. Having done this,
we are gratified with classification results in several cases:
1. We reprove and extend the result of Elias and Rossi:
Proposition 1.1 ([ER12, ER15], Corollary 3.14). Let k be a field of characteristic other
than 2 or 3. A local Artin
Gorenstein k-algebra with Hilbert function (1, n, n, 1) or Hilbert
n+1
function 1, n, 2 , n, 1 is isomorphic to its associated graded algebras.
The original statement is proven in [ER12, ER15] for fields of characteristic zero. In
Example 3.15 we show that the statement is not true for fields of characteristic two or
three.
In Proposition 1.1 the condition on the Hilbert function can be rephrased by saying that
the algebra is compressed, see Section 3.2. The algebras appearing in Proposition 1.1 are
precisely compressed algebras of socle degree three and four. A analogous statement for
compressed algebras of higher socle degree is false. However, we show that compressed
algebras of all socle degrees are near to the graded algebras, see Corollary 3.13.
2. We prove that general algebras with Hilbert function (1, 2, 3, 3, 2, 1) and general algebras
with Hilbert function (1, 2, 2, 2, . . . , 2, 1) are isomorphic to their associated graded algebras
(see Example 3.17 or Example 3.25 for a precise meaning of the word general).
3. An algebra is canonically graded if it isomorphic to its associated graded algebra. We
investigate the set of (dual socle generators of) algebras which are canonically graded. In
Proposition 3.19 we show that this set is irreducible, but in general neither open nor closed
in the parameter space. This answers a question of Elias and Rossi [ER15, Remark 3.6].
We investigate density of the set canonically graded algebras and in Proposition 3.20 we
prove the only if part of the following conjecture.
Conjecture 1.2 (Conjecture 3.21). Assume that k is a field of characteristic not equal to
2, 3, 5. Say that an algebra A has type (n, d) if HA (1) = n and the socle degree of A is d.
Then a general algebra of type (n, d) is canonically graded if and only if (n, d) belongs to
the following list:
(a)
(b)
(c)
(d)
d 6 4 and n arbitrary,
d = 5 and n 6 6,
d = 6 and n = 2.
d arbitrary and n = 1.
4. We classify algebras with Hilbert function (1, 3, 3, 3, 1), obtaining eleven isomorphism types,
see Example 3.27. We also discuss the case of (1, 3, 4, 3, 1), show that there are infinitely
many isomorphism types and classify algebras which are not canonically graded, see Example 3.28. Non-canonically graded algebras with Hilbert function (1, 3, 4, 3, 1) are investigated independently in [MR16].
2
5. We classify algebras with Hilbert function (1, 2, 2, 2, 1, 1, 1) obtaining |k| + 1 isomorphism
types, see Example 3.29. This example was worked out by Elias and Valla [EV11].
We now explain our approach. In short, the classification problem may be reduced to representation theory, see [Ems78]. Assume for a moment for simplicity of presentation that the
characteristic of k is zero. Let n = HA (1). Then A may be seen as a quotient of a fixed power
series ring S = k[[α1 , . . . , αn ]]. The ring S acts on a polynomial ring P = k[x1 , . . . , xn ] by letting
αi act as the partial derivative with respect to xi . It turns out that A = S/ Ann(f ) for an
element f ∈ P , called the dual socle generator of A. Then the problem of classifying algebras
reduces to classifying quotients of S, which in turn boils down to classifying f ∈ P . There is a
certain group G acting on P such that S/ Ann(f ) ≃ S/ Ann(g) if and only if f and g lie in the
same G-orbit. We summarise this (for fields of arbitrary characteristic) in Proposition 2.14.
Our contribution to the above is in providing explicit formulas for all objects involved and
in applying Lie-theoretic ideas. In particular, we give formulas for the G-orbit and its tangent
space, thus we are able to compute dimensions of orbits. We also find the unipotent radical of
G and analyse its orbits, which are closed.
The explicit formulas make several known results, such as those by Elias and Rossi, corollaries
of the presented theory. This is, in our opinion, the most important contribution of this paper.
Accordingly, we have tried to keep the paper as elementary as possible.
The reader willing to get a quick idea of the contents of this paper should analyse Example 2.16, referring to previous results if necessary. Otherwise we recommend to read through
the paper linearly. The paper is organized as follows. First, we recall and develop the abstract
theory of inverse systems; in particular we give an explicit formula for the automorphism group
action in Proposition 2.9. Then we explain the link between classifying algebras and elements
of P in Proposition 2.14. We give some examples, in particular Example 2.16, and then discuss
the basic Lie theory in Section 2.4. Finally, in Section 3 we present the applications mentioned
above. Apart from the conjecture above, there are several natural questions that are in our
opinion worth considering:
1. What are H such that there are finitely many isomorphism types of Artinian Gorenstein
algebras with Hilbert function H? What is the classification in these cases?
2. An Artinian Gorenstein algebra is rigid if it cannot be deformed to a non-isomorphic
algebra, see [Sch73]. The G-orbit of a rigid algebra should be large. For k algebraically
closed, are there rigid k-algebras other than k?
3. Given the Hilbert function, are there non-trivial bounds on the dimensions of G-orbits of
algebras with this Hilbert function?
4. Can the above ideas of classification be generalised effectively to non-Gorenstein algebras?
Acknowledgements
I thank Pedro Macias Marques, Alessandra Bernardi and Kristian Ranestad for the joint work
which inspired the explicit formula in Proposition 2.9. I thank Jarosław Buczyński for helpful
suggestions and Aldo Conca for pointing some other possible applications. I thank Anthony
Iarrobino for comments on a near-final version of the paper and the referees for suggesting
many improvements. This paper has emerged from a desire to understand the surprising results
of Joan Elias and Maria Evelina Rossi [ER12, ER15]. I thank them for their hospitality in
Barcelona and Genova.
The computations of apolar algebras were done using Magma [BCP97] but could also be done
using Macaulay2 [GS] command fromDual or Joan Elias library [Eli15] for Singular [DGPS15].
3
2
Preliminaries and theoretical results
2.1
Power series ring S and its dual P
In this section we introduce our main objects of study: the power series ring S and its action on
the dual divided power (or polynomial) ring P .
By N0 we denote the set of non-negative integers. Let k be an algebraically closed field of
arbitrary characteristic. Let S be a power series ring over k of dimension dim S = n and let m
be its maximal ideal. By ord (σ) we denote the order of a non-zero σ ∈ S i.e. the largest i such
that σ ∈ mi . Then ord (σ) = 0 if and only if σ is invertible. Let S ∨ = Homk (S, k) be the space
of functionals on S. We denote the pairing between S and S ∨ by
h−, −i : S × S ∨ → k.
Definition 2.1. The dual space P ⊂ S ∨ is the linear subspace of functionals eventually equal
to zero:
P = f ∈ S ∨ | ∀D≫0 mD , f = 0 .
On P we have a structure of S-module via precomposition: for every σ ∈ S and f ∈ P the
element σ y f ∈ P is defined via the equation
hτ, σ y f i = hτ σ, f i
for every τ ∈ S.
(1)
This action is called contraction.
Existence of contraction is a special case of the following construction, which is basic and
foundational for our approach. Let L : S → S be a k-linear map. Assume that there exists
an integer s such that L(mi ) ⊂ mi+s for all i. Then the dual map L∨ : S ∨ → S ∨ restricts to
L∨ : P → P . Explicitly, L∨ is given by the equation
τ, L∨ (f ) = hL(τ ), f i
for every τ ∈ S, f ∈ P.
(2)
To obtain contraction with respect to σ we use L(τ ) = στ , the multiplication by σ. Later in this
paper we will also consider maps L which are automorphisms or derivations.
To get a down to earth description of P , choose α1 , . . . , αn ∈ S so that S = k[[α1 , . . . , αn ]].
Write αa to denote αa11 . . . αann . For every a ∈ Nn0 there is a unique element x[a] ∈ P dual to αa ,
given by
(
D
E
1 if a = b
αb , x[a] =
0 otherwise.
Additionally, we define xi as the functional dual to αi , so that xi = x[(0,...0,1,0,...,0)] with one on
i-th position.
Let us make a few remarks:
1. The functionals x[a] form a basis of P .
2. The contraction action is given by the formula
(
x[b−a] if b > a, that is, ∀i bi > ai
αa y x[b] =
0
otherwise.
Therefore our definition agrees with the standard, see [IK99, Definition 1.1, p. 4].
P
We say that x[a] has degree
ai . We will freely speak about constant forms, linear forms,
(divided) polynomials of bounded degree etc.
We endow P with a topology, which is the Zariski topology of an affine space. It will be used
when speaking about general polynomials and closed orbits, but for most of the time it is not
important.
Now we will give a ring structure Q
on P . It
2.9. For
in
Pwill be
P used crucially
QProposition
ai +bi
a+b
multi-indices a, b ∈ Nn0 we define a! = (ai !),
a=
ai and a+b
=
=
i
a
ai
b .
4
Definition 2.2. We define multiplication on P by
a + b [a+b]
[a]
[b]
x · x :=
x
.
a
(3)
In this way P is a divided power ring.
The multiplicative structure on P can be defined in a coordinate-free manned using a natural
comultiplication on S. We refer to [Eis95, §A2.4] for details in much greater generality.
Example 2.3. Suppose that
k is of characteristic 3. Then P is not isomorphic to a polynomial
[2]
[3]
[3]
ring. Indeed, x1 · x1 · x1 = 2x1 · x1 = 3x1 = 0. Moreover x1 is not in the subring generated
by x1 , . . . , xn .
Note that linear forms from S act on P as derivatives. Therefore we can interpret S as lying
inside the ring of differential operators on P . We will need the following related fact.
Lemma 2.4. Let σ ∈ S and denote by σ (i) its i-th partial derivative. For every f ∈ P we have
σ y (xi · f ) − xi · (σ y f ) = σ (i) y f.
(4)
Proof. Since the formula is linear in σ and f we may assume these are monomials. Let σ = αri τ ,
where αi does not appear in τ . Then σ (i) = rαir−1 τ . Moreover τ y (xi · f ) = xi · (τ y f ), thus we
may replace f by τ y f and reduce to the case τ = 1, σ = αri .
[s]
[s+1]
Write f = xi g where g is a monomial in variables other than xi . Then xi · f = (s + 1)xi
g
according to (3). If s + 1 < r then both sides of (4) are zero. Otherwise
[s+1−r]
σ y (xi ·f ) = (s+1)xi
[s−r]
g, xi ·(σ y f ) = xi ·xi
[s−r+1]
g = (s−r+1)xi
[s−(r−1)]
g, σ (i) y f = rxi
g,
so Equation (4) is valid in this case also.
Remark 2.5. Lemma 2.4 applied to σ = αi shows that αi y (xi · f ) − xi · (αi y f ) = f . This
can be rephrased more abstractly by saying that αi and xi interpreted as linear operators on P
generate a Weyl algebra.
Example 2.3 shows that P with its ring structure has certain properties distinguishing it from
the polynomial ring, for example it contains nilpotent elements. Similar phenomena do not occur
in degrees lower than the characteristic or in characteristic zero, as we show in Proposition 2.6
and Proposition 2.8 below.
P
Proposition 2.6. Let P>d be the linear span of {x[a] |
a > d}. Then P>d is an ideal of P ,
for all d. Let k be a field of characteristic p. The ring P/P>p is isomorphic to the truncated
polynomial ring. In fact
Ω : P/P>p → k[x1 , . . . , xn ]/(x1 , . . . , xn )p
defined by
is an isomorphism.
xa1 . . . xan
n
.
Ω x[a] = 1
a1 ! . . . an !
Proof. Since Ω maps a basis of P/Ip to a basis of k[x1 , . . . , xn ]/(x1 , . . . , xn )p , it is clearly well
defined
The fact that Ω is a k-algebra homomorphism reduces to the equality
and
Q (abijective.
a+b
i +bi )!
=
.
a
ai !bi !
5
Characteristic zero case. In this paragraph we assume that k is of characteristic zero. This
case is technically easier, but there are two competing conventions: contraction and partial
differentiation. These agree up to an isomorphism. The main aim of this section is clarify
this isomorphism and provide a dictionary between divided power rings used in the paper and
polynomial rings in characteristic zero. Contraction was already defined above, now we define
the action of S via partial differentiation.
Definition 2.7. Let k[x1 , . . . , xn ] be a polynomial ring. There is a (unique) action of S on
∂
. For f ∈ k[x1 , . . . , xn ] and σ ∈ S we denote
k[x1 , . . . , xn ] such that the element αi acts a ∂x
i
this action as σ ◦ f .
The following Proposition 2.8 shows that in characteristic zero the ring P is in fact polynomial
and the isomorphism identifies the S-module structure on P with that from Definition 2.7 above.
Proposition 2.8. Suppose that k is of characteristic zero. Let k[x1 , . . . , xn ] be a polynomial ring
with S-module structure as defined in 2.7. Let Ω : P → k[x1 , . . . , xn ] be defined via
xa1 . . . xan
n
.
Ω x[a] = 1
a1 ! . . . an !
Then Ω is an isomorphism of rings and an isomorphism of S-modules.
Proof. The map Ω is an isomorphism of k-algebras by the same argument as in Proposition 2.6.
We leave the check that Ω is a S-module homomorphism to the reader.
Summarising, we get the following corresponding notions.
Arbitrary characteristic
Characteristic zero
divided power series ring P
S-action by contraction (precomposition) denoted σ y f
x[a]
xi = x[(0,...0,1,0,...,0)]
polynomial ring k[x1 , . . . , xn ]
S action by derivations denoted σ ◦ f
xa /a!
xi
2.2
Automorphisms and derivations of the power series ring
Let as before S = k[[α1 , . . . , αn ]] be a power series ring with maximal ideal m. This ring has a
huge automorphism group: for every choice of elements σ1 , . . . , σn ∈ m whose images span m/m2
there is a unique automorphism ϕ : S → S such that ϕ(αi ) = σi . Note that ϕ preserves m and
its powers. Therefore the dual map ϕ∨ : S ∨ → S ∨ restricts to ϕ∨ : P → P . The map ϕ∨ is
defined (using the pairing of Definition 2.1) via the condition
hϕ(σ), f i = σ, ϕ∨ (f )
for all σ ∈ S, f ∈ P.
(5)
Now we will describe this action explicitly.
Proposition 2.9. Let ϕ : S → S be an automorphism. Let Di = ϕ(αi ) − αi . For a ∈ Nn0 denote
Da = D1a1 . . . Dnan . Let f ∈ P . Then
∨
ϕ (f ) =
X
[a]
x
a
· (D y f ) = f +
a∈Nn
0
n
X
xi · (Di y f ) + . . . .
i=1
Proof. We need to show that hσ, ϕ∨ (f )i = hϕ(σ), f i for all σ ∈ S. Since f ∈ P , it is enough to
check this for all σ ∈ k[α1 , . . . , αn ]. By linearity, we may assume that σ = αa .
For every g ∈ P let ε(g) = h1, gi ∈ k. We have
X a
X a
a−b
b
a−b
b
ε
α
y (D y f ) .
α
D y f =
hϕ(σ), f i = h1, ϕ(σ) y f i = ε (ϕ(σ) y f ) = ε
b
b
b6a
6
b6a
Consider a term of this sum. Observe that for every g ∈ P
a a−b
ε
α
y g = ε αa y x[b] · g .
b
(6)
Indeed it is enough to check the above equality for g = x[c] and both sides are zero unless
c = a − b, thus it is enough to check the case g = x[a−b] , which is straightforward. Moreover
note that if b 66 a, then the right hand side is zero for all g, because ε is zero for all x[c] with
non-zero c. We can use (6) and remove the restriction b 6 a, obtaining
! *
+
X
X
X
a
[b]
b
a
[b]
b
a
[b]
b
α y x · (D y f )
ε α y x · (D y f ) = ε
= α ,
x · (D y f ) =
b
b
b
αa , ϕ∨ (f ) = σ, ϕ∨ (f ) .
Consider now a derivation D : S → S. Here and elsewhere a derivation is a k-linear map
satisfying the Leibnitz rule. The derivation D gives rise to a dual map D ∨ : P → P . We wish
to describe it explicitly.
Proposition 2.10. Let D : S → S be a derivation and Di := D(αi ). Let f ∈ P . Then
∨
D (f ) =
n
X
xi · (Di y f ).
i=1
Proof. The proof is similar, though easier, to the proof of Proposition 2.9.
Remark 2.11. Suppose D : S → S is a derivation such that D(m) ⊆ m2 . Then deg(D ∨ (f )) <
deg(f ). We say that D lowers the degree.
Characteristic zero case. Let k be a field of characteristic zero. By xa we denote the monomial xa11 . . . xann in the polynomial ring k[x1 , . . . , xn ]. Then, in the notation of Proposition 2.8,
we have
1
Ω x[a] = xa .
a!
Clearly, an automorphism of S gives rise to an linear map k[x1 , . . . , xn ] → k[x1 , . . . , xn ]. We may
restate Proposition 2.9 and Proposition 2.10 as
Corollary 2.12. Let ϕ : S → S be an automorphism. Let Di = ϕ(αi ) − αi . For a ∈ Nn0 denote
Da = D1a1 . . . Dnan . Let f ∈ k[x1 , . . . , xn ]. Then
ϕ∨ (f ) =
n
X xa
X
xi (Di ◦ f ) + . . . .
(Da ◦ f ) = f +
a!
n
i=1
a∈N0
Let D : S → S be a derivation and Di := D(αi ). Then
∨
D (f ) =
n
X
xi (Di ◦ f ).
i=1
Example 2.13. Let n = 2, so that S = k[[α1 , α2 ]] and consider a linear map ϕ : S → S given by
ϕ(α1 ) = α1 and ϕ(α2 ) = α1 + α2 . Dually, ϕ∨ (x1 ) = x1 + x2 and ϕ∨ (x2 ) = x2 . Since ϕ is linear,
ϕ∨ is an automorphism of k[x1 , x2 ]. Therefore ϕ∨ (x31 ) = (x1 + x2 )3 . Let us check this equality
using Proposition 2.12. We have D1 = ϕ(α1 ) − α1 = 0 and D2 = ϕ(α2 ) − α2 = α1 . Therefore
D(a,b) = 0 whenever a > 0 and D(0,b) = αb1 .
7
We have
ϕ∨ x31 =
X
(a,b)∈N2
X xb
x2
x2
x3
xa1 xb2 (a,b) 3
2
(D
◦x1 ) =
(αb1 ◦x31 ) = x31 + ·(3x21 )+ 2 ·(6x1 )+ 2 ·(6) = (x1 +x2 )3 ,
a!b!
b!
1
2
6
b∈N
which indeed agrees with our previous computation.
When ϕ is not linear, ϕ∨ is not an endomorphism of k[x1 , x2 ] and computing it directly from
definition becomes harder. For example, if ϕ(α1 ) = α1 and ϕ(α2 ) = α2 + α21 , then
ϕ∨ (x1 ) = x1 ,
2.3
ϕ∨ (x41 ) = x41 + 12x21 x2 + 12x22 .
Local Artinian Gorenstein algebras
Let (A, mA , k)
be a local Artinian Gorenstein algebra. Recall that the socle degree of A is defined
as d = max i | miA 6= 0 . The embedding dimension of A is HA (1) = dim mA /m2A . Note that
A may be presented as a quotient of a power series ring S over k if and only dim S > HA (1).
Therefore it is most convenient to analyse the case when dim S = HA (1), which we will usually
do.
Recall that by Macaulay’s inverse systems, for every local Gorenstein quotient A = S/I there
is an f ∈ P such that I = Ann(f ), see [IK99, Iar94], [Eis95, Chapter 21] and [Mac94, Chapter IV].
Such an f is called a dual socle generator of A. Given I, the choice of f is non-unique, but any
two choices differ by a unit in S. Conversely, for f ∈ P we denote
Apolar (f ) = S/ Ann(f ),
which is called the apolar algebra of f and Ann(f ) is called the apolar ideal of f . It is important
that the apolar algebra of the top degree form of f is a quotient of gr A, see [Iar94, Chapter 1].
We may compare A and f as follows:
algebra A
module Sf
S-module A
dimk A
socle degree of A
certain quotient of gr A
If for all i we have
HA (d − i) = HA (i) then gr A
≃
=
=
≃
≃
Sf
dimk Sf
deg(f )
Apolar fdeg(f )
Apolar fdeg(f )
As mentioned in the introduction, our interest is in determining when two local Artinian
Gorenstein quotients of the power series ring are isomorphic. The following Proposition 2.14
connects this problem with the results of the previous section. It is well-known, see e.g. [Ems78].
Let S ∗ denote the group of invertible elements of S and let
G := Aut(S) ⋉ S ∗
be the group generated by Aut(S) and S ∗ in the group of linear operators on S. As the notation
suggests, the group G is a semidirect product of those groups: indeed ϕ ◦ µs ◦ ϕ−1 = µϕ(s) , where
ϕ is an automorphism, s ∈ S is invertible and µs denotes multiplication by s. By Equation (2)
we have an action of G on P . Here S ∗ acts by contraction and Aut(S) acts as described in
Proposition 2.9.
Proposition 2.14. Let A = S/I and B = S/J be two local Artinian Gorenstein algebras. Choose
f, g ∈ P so that I = Ann (f ) and J = Ann (g). The following conditions are equivalent:
1. A and B are isomorphic,
2. there exists an automorphism ϕ : S → S such that ϕ(I) = J,
8
3. there exists an automorphism ϕ : S → S such that ϕ∨ (f ) = σ y g, for an invertible element
σ ∈ S.
4. f and g lie in the same G-orbit of P .
Proof. Taking an isomorphism A ≃ B, one obtains ϕ′ : S → B = S/I, which can be lifted to an
automorphism of S by choosing lifts of linear forms. This proves 1. ⇐⇒ 2.
2. ⇐⇒ 3. Let ϕ be as in 2. Then Ann (ϕ∨ (f )) = ϕ(Ann (f )) = ϕ(I) = J. Therefore the
principal S-submodules of P generated by ϕ∨ (f ) and g are equal, so that there is an invertible
element σ ∈ S such that ϕ∨ (f ) = σ y g. The argument can be reversed.
Finally, 4. is just a reformulation of 3.
The invertible element in Point 3 cannot be discarded, see Example 2.16 and Example 3.27.
The outcome of this theorem is that we are interested in the orbits of elements of P under the
G action.
Example 2.15 (quadrics). Let f ∈ P2 be a quadric of maximal rank. Then G·f is the set of
(divided) polynomials of degree two, whose quadric part is of maximal rank.
Example 2.16 (compressed cubics). Assume that k has characteristic not equal to two.
Let f = f3 + f62 ∈ P63 be an element of degree three such HApolar(f ) = (1, n, n, 1), where
n = HS (1). Such a polynomial f is called compressed, see [Iar84, ER15] or Section 3.2 for a
definition.
We claim that there is an element ϕ ∈ G such that ϕ∨ (f3 ) = f . This implies that Apolar (f ) ≃
Apolar (f3 ) = gr Apolar (f ). We say that the apolar algebra of f is canonically graded.
Let A = Apolar (f3 ). Since HA (2) = n = HS (1), every linear form in P may be obtained as
δ y f forP
some operator δ ∈ S of order two, see Remark 3.11 below. We pick operators D1 , . . . , Dn
so that
xi · (Di y f3 ) = f2 + f1 . Explicitly, Di is such that Di y f3 = (αi y f2 )/2 + αi y f1 . Here
we use the assumption on the characteristic.
Let ϕ : S → S be an automorphism defined via ϕ(αi ) = αi + Di . Since (Di Dj ) y f = 0 by
degree reasons, the explicit formula in Proposition 2.9 takes the form
X
ϕ∨ (f3 ) = f3 +
xi · (Di y f3 ) = f3 + f2 + f1 .
The missing term f0 is a constant, so that we may pick an order three operator σ ∈ S with
σ y ϕ∨ (f3 ) = f0 . Then (1 + σ) y (ϕ∨ (f3 )) = f3 + f2 + f1 + f0 = f , as claimed.
The isomorphism Apolar (f ) ≃ Apolar (f3 ) was first proved by Elias and Rossi, see [ER12,
Thm 3.3]. See Example 2.22 for a more conceptual proof for k = C and Section 3.2 for generalisations. In particular Corollary 3.14 extends this example to cubic and quartic forms.
Remark 2.17 (Graded algebras). In the setup of Proposition 2.14 one could specialize to homogeneous ideals I, J and homogeneous polynomials f, g ∈ P . Then Condition 1. is equivalent
to the fact that f and g lie in the same GL(m/m2 )-orbit. The proof of Proposition 2.14 easily
restricts to this case, see [Ger96].
Now we turn our attention to derivations. The notation and motivation come from Lie theory,
see section on characteristic zero below. Let aut denote the space of derivations of S preserving
m, i.e. derivations such that D(m) ⊆ m. Let S ⊂ Homk (S, S) be given by sending σ ∈ S to
multiplication by σ. Let
g := aut + S,
where the sum is taken in the space of linear maps from S to S. Then g acts on P as defined
in Equation (2). The space g is naturally the tangent space to the group G. Similarly, aut · f is
naturally the tangent space of the orbit Aut(S) · f and gf is the tangent space to the orbit G·f
for every f ∈ P .
Sometimes it is more convenient to work with S than with P . For each subspace W ⊆ P
we may consider the orthogonal space W ⊥ ⊆ S. Below we describe the linear space (gf )⊥ . For
σ ∈ S by σ (i) we denote the i-th partial derivative of σ. We use the convention that deg(0) < 0.
9
Proposition 2.18 (tangent space description). Let f ∈ P . Then
aut · f = span hxi · (δ y f ) | δ ∈ m, i = 1, . . . , ni ,
gf = Sf +
n
X
m(xi · f ).
i=1
Moreover
n
o
(gf )⊥ = σ ∈ S | σ y f = 0, ∀i deg(σ (i) y f ) 6 0 .
Suppose further that f ∈ P is homogeneous of degree d. Then (gf )⊥ is spanned by homogeneous
operators and
n
o
(i)
(gf )⊥
=
σ
∈
S
|
σ
y
f
=
0,
∀
σ
y
f
=
0
.
i
6d
P
Proof. Let D ∈ aut and Di := D(αi ). By Proposition 2.10 we have D ∨ (f ) = ni=1 xi · (Di y f ).
For any δ ∈ m we may choose D so that Di = δ and all other Dj are zero. This proves the
description of aut · f . Now gf = Sf + span hxi · (σ y f ) | σ ∈ m, i = 1, . . . , ni. By Lemma 2.4 we
have xi (σ y f ) ≡ σ y (xi f ) mod Sf . Thus
X
gf = Sf + span hσ y (xi · f ) | σ ∈ m, i = 1, . . . , ni = Sf +
m(xi f ).
Now let σ ∈ S be an operator such that hσ, gf i = 0. This is equivalent to σ y (gf ) = 0, which
simplifies to σ y f = 0 and (σm) y (xi f ) = 0 for all i. We have σ y (xi f ) = xi (σ y f ) + σ (i) y f =
σ (i) y f , thus we get equivalent conditions:
σ y f = 0 and m y (σ (i) y f ) = 0,
and the claim follows. Finally, if f is homogeneous of degree d and σ ∈ S is homogeneous
of degree at most d then σ (i) y f has no constant term and so deg(σ (i) y f ) 6 0 implies that
σ (i) y f = 0.
Remark 2.19. Let f ∈ P be homogeneous of degree d. Let i 6 d and Ki := (gf )⊥
i . Proposition 2.18 gives a useful connection of Ki with the conormal sequence. Namely, let I = Ann (f )
and B = Apolar (f ) = S/I. We have (I 2 )i ⊆ Ki and the quotient space fits into the conormal
sequence of S → B, making that sequence exact:
0 → K/I 2 i → I/I 2 i → ΩS/k ⊗ B i → ΩB/k i → 0.
(7)
This is expected from the point of view of deformation theory. Recall that by [Har10, Theorem 5.1] the deformations of B over k[ε]/ε2 are in one-to-one correspondence with elements of
a k-linear space T 1 (B/k, B). On the other hand, this space fits [Har10, Prop 3.10] into the
sequence
0 → HomB (ΩB/k , B) → HomB (ΩS/k ⊗ B, B) → HomB (I/I 2 , B) → T 1 (B/k, B) → 0.
Since B is Gorenstein, HomB (−, B) is exact and we have T 1 (B/k, B)i ≃ Hom(K/I 2 , B)i for all
i > 0. The restriction i > 0 appears because Hom(K/I 2 , B) is the tangent space to deformations
of B inside S, whereas T 1 (B/k, B) parameterises all deformations.
Now we introduce a certain subgroup G+ of G. It plays an important part in characteristic
zero, because G+ -orbits are closed in P . This group is also very useful in applications, because
it preserves the top degree form, which allows induction on the degree type arguments.
Each automorphism of S induces a linear map on its cotangent space: we have a restriction
Aut(S) → GL(m/m2 ). Let us denote by Aut+ (S) the group of automorphisms which act as
identity on the tangent space: Aut+ (S) = ϕ ∈ S | ∀i ϕ(αi ) − αi ∈ m2 . We have the following
sequence of groups:
1 → Aut+ (S) → Aut(S) → GL(m/m2 ) → 1.
(8)
10
We define
G+ = Aut+ (S) ⋉ (1 + m) ⊆ G.
Note that we have the following exact sequence:
1 → G+ → G → GL(m/m2 ) × k∗ → 1.
(9)
Correspondingly, let aut denote the space of derivations preserving m, i.e. derivations such
that D(m)
⊆ m. Let aut+ denote the space of derivations such that D(m) ⊆ m2 . Denoting by
gl m/m2 the space of linear endomorphisms of m/m2 , we have we following sequence of linear
spaces:
0 → aut+ → aut → gl m/m2 → 0.
(10)
We define
g+ = aut+ + m.
Following the proof of Proposition 2.18 we get the following proposition.
P
Proposition 2.20. Let f ∈ P . Then g+ f = mf + m2 (xi f ) so that
o
⊥ n
g+ f = σ ∈ S | deg(σ y f ) 6 0, ∀i deg(σ (i) y f ) 6 1 .
If f is homogeneous of degree d then g+ f is spanned by homogeneous polynomials and
n
o
⊥
g+ f <d = σ ∈ S | σ y f = 0, ∀i σ (i) y f = 0 = (gf )⊥
<d .
2.4
Characteristic zero
For simplicity we restrict to k = C in this section, to freely speak about Lie groups and algebras.
As a general reference we suggest [FH91]. One technical problem is that the group Aut(S) is
infinite-dimensional. In this section we implicitly replace S with S/mD and Aut(S) with the
group Aut(S/mD ), where D ≫ 0 is larger than the degree of any element of P considered. Then
Aut(S) and G become Lie groups.
Note that the group G+ ⊆ GL(S) is unipotent: the map ϕ − id is nilpotent for every ϕ ∈ G+ .
Therefore, we get the following theorem.
Theorem 2.21. For every f ∈ P the orbit G+ ·f is closed in P , in both Euclidean and Zariski
topologies.
Classically, the Lie algebra corresponding to Aut(S) is aut. Moreover by the exact sequences (8) and (10), the Lie algebra of Aut+ (S) is aut+ . Finally, the Lie algebra of G is
aut + S = g and the Lie algebra of G+ is aut+ + m = g+ . In particular
X
dim G·f = dim (aut · f + Sf ) = dim Sf +
m(xi f ) .
We now give another proof of Elias-Rossi theorem on canonically graded algebras, in the following
Example 2.22.
Example 2.22 (compressed cubics, using Lie theoretic ideas). Assume that k = C.
Let f = f3 + f62 ∈ P63 be an element of degree three such HApolar(f ) = (1, n, n, 1), where
n = HS (1). Such a polynomial f is called compressed, see [Iar84, ER15] and Section 3.2 for
a definition. Then HApolar(f ) is symmetric, thus HApolar(f3 ) = (1, n, n, 1) as explained in Section 2.3.
We claim that there is an element ϕ of G such that ϕ∨ (f3 ) = f . This proves that Apolar (f ) ≃
Apolar (f3 ) = gr Apolar (f ). We say that the apolar algebra of f is canonically graded.
11
In fact, we claim that already G+ ·f3 is the whole space:
G+ ·f3 = f3 + P62 .
From the explicit formula in Proposition 2.9 we see that G+ ·f3 ⊆ f3 + P62 . Then it is a
Zariski closed subset. To prove equality it enough to prove that these spaces have the same
dimension, in other words that g+ f3 = P62 .
⊥
Let σ ∈ (g+ f3 )62 be non-zero. By Proposition 2.20 we get that σ y f3 = 0 and σ (i) y f3 = 0
for all i. Then there is a degree one operator annihilating f3 . But this contradicts the fact that
⊥
HApolar(f3 ) (1) = n = HS (1). Therefore (g+ f3 )62 = 0 and the claim follows.
3
Applications
In this part of the paper we give several new classification results using the above theory. The two
most important ones are the t-compressed algebras in Section 3.2, and algebras having Hilbert
function (1, 3, 3, 3, 1) in Example 3.27. We begin and end with some examples computable by
hand, which are nevertheless useful, see in particular the reference in Example 3.1.
We make the following assumption
Assumption. The field k is algebraically closed. It has characteristic zero or greater than the
degree of all considered elements of P .
For example, when we analyse cubics, we assume that k has characteristic different from 2 or
3. When analysing degree d polynomials, as in Example 3.1, we assume that the characteristic
is at least d + 1 or 0. The fact that k is algebraically closed is assumed because of various
geometrical arguments involved, such as in Proposition 3.3, and because of assumptions required
for the references: Example 3.27 refers to [LO13] and [CJN15] which both work over an algebraically closed field. It seems plausible that the classification is feasible over non-algebraically
closed fields, however then there will be more isomorphism types: already the classifications of
homogeneous quadrics over C and R differ. In contrast, the assumption that k has large enough
characteristic is hardly avoidable since key results are false without it, see Example 3.15 and compare with Example 2.16. This assumption is frequently used to guarantee that a non-constant
element of S has a non-zero derivative, which is needed to effectively apply Proposition 2.18, see
Proposition 3.12.
Under Assumption, we may always think of P as a polynomial ring, see Proposition 2.6.
Before we begin, we make the following observation. As shown by Equation (9), the group G
is build from G+ , GL(m/m2 ) and k∗ . We will be mostly concerned with G+ , in particular we
will analyse its orbits. This often suffices to prove important results as in the cubic case above
(Example 2.16) or below for t-compressed algebras. However sometimes the action of G+ is not
sufficient and the difference between this group and G becomes the substance of the result. This
is the case for algebras having Hilbert function (1, 3, 3, 3, 1). The group G+ is also convenient
because it allows degree-by-degree analysis which is our main technical tool, see Proposition 3.3.
When classifying Artinian Gorenstein quotients of S with Hilbert function H we will usually
assume that dim S = H(1). This does not reduce generality and simplifies the argument a bit.
3.1
Basic examples and tools
In this section we analyse the G-orbits of “easy” homogeneous polynomials: those with with small
Waring and border Waring rank.
Example 3.1 (Rank one). The simplest example seems to be the case f = ℓd for some linear
form ℓ. We may assume ℓ = x1 and apply Proposition 2.20 to see that
⊥
= (α2 , . . . , αn )2 ,
g+ xd1
<d
g+ xd1 = span hxr1 m | m ∈ P61 , r + 1 < di ,
12
or, in an invariant form, g+ ℓd = span hℓr · m | m ∈ P61 , r + 1 < di. Using Proposition 2.9 one
can even compute the orbit itself. For example, when d = 4, the orbit is equal to
G+ ·ℓ4 = ℓ4 + ℓ2 · m1 + ℓ · m2 + m3 + m21 | mi ∈ P61 .
This example together with Corollary 3.4 plays an important role in the paper [CJN15], compare
[CJN15, Lemma 4.2, Example 4.4, Proposition 5.13].
Example 3.2 (Border rank two). Consider f = xd−1 y. Assume S = k[[α, β]] and P = kdp [x, y].
As above, the apolar ideal of f is monomial, equal to (αd , β 2 ). Using Proposition 2.20 we easily
get that
⊥
g+ f <d = (β 3 )<d ,
so g+ f is spanned by monomials xa y b , where b 6 2 and a + b < d. Note that in contrast with
⊥
Example 3.1, the equality (g+ f )<d = Ann (f )2<d does not hold. This manifests the fact that
Apolar (f ) is a deformation of Apolar xd + y d .
Before we analyse the rank two polynomials i.e. xd + y d , we need a few more observations.
Let for every g ∈ P the symbol tdf(g) denote the top degree form of g, so that for example
tdf(x31 + x22 x3 + x24 ) = x31 + x22 x3 .
Proposition 3.3. Let f ∈ P . Then the top degree form of every element of G+ ·f is equal to
the top degree form of f . Moreover,
tdf(g − f ) | g ∈ G+ ·f = tdf(h) | h ∈ g+ f .
(11)
If f is homogeneous, then both sides of (11) are equal to the set of homogeneous elements of
g+ f .
Proof. Consider the S-action on P6d . This action descents to an S/md+1 action. Further in
the proof we implicitly replace S by S/md+1, thus also replacing Aut(S) and G by appropriate
truncations. Let ϕ ∈ G+ . Since (id −ϕ) mi ⊆ mi+1 for all i, we have (id −ϕ)d+1 = 0. By our
global assumption on the characteristic D := log(ϕ) is well-defined and ϕ = exp(D). We get an
injective map exp : g+ → G+ with left inverse log. Since exp is algebraic we see by dimension
count that its image is open in G+ . Since log is Zariski-continuous, we get that log(G+ ) ⊆ g+ ,
then exp : g+ → G+ is an isomorphism.
Therefore
!
d−1
d
∨ )i
∨ )i (f )
X
X
(D
(D
D ∨ (f ).
= f + D ∨ (f ) +
ϕ∨ (f ) = f +
i!
(i + 1)!
i=1
i=1
By Remark 2.11 the derivation D ∈ g+ lowers the degree, we see that tdf(ϕ∨ f ) = tdf(f ) and
tdf(ϕ∨ f − f ) = tdf(D ∨ (f )). This proves (11). Finally, if f is homogeneous then g+ f is equal to
the span htdf(h) | h ∈ g+ f i by Proposition 2.20, and the last claim follows.
For an elementary proof, at least for the subgroup Aut+ (S), see [Mat87, Proposition 1.2].
The following almost tautological Corollary 3.4 enables one to prove that a given apolar
algebra is canonically graded inductively, by lowering the degree of the remainder.
Corollary 3.4. Let F and f be polynomials. Suppose that the leading form of F − f lies in g+ F .
Then there is an element ϕ ∈ G+ such that deg(ϕ∨ f − F ) < deg(f − F ).
Proof. Let G be the leading form of f − F and e be its degree. By Proposition 3.3 we may find
ϕ ∈ Aut+ (S) such that tdf(ϕ∨ (F ) − F ) = −G, so that ϕ∨ (F ) ≡ F − G mod P6e−1 . By the
same proposition we have deg(ϕ∨ (f − F )− (f − F )) < deg(f − F ) = e, so that ϕ∨ (f − F ) ≡ f − F
mod P6e−1 . Therefore ϕ∨ (f ) − F = ϕ∨ (F ) + ϕ∨ (f − F ) − F ≡ f − G − F ≡ 0 mod P6e−1 , as
claimed.
13
Example 3.5 (Rank two). Let P = kdp [x, y], S = k[[α, β]] and F = xd +y d for some d > 2. Then
HApolar(F ) = (1, 2, 2, 2, . . . , 2, 1). We claim that the orbit G+ ·F consists precisely of polynomials
f having leading form F and such that HApolar(f ) = HApolar(F ) .
⊥
Let us first compute (g+ F ) . Since F is homogeneous, Proposition 2.20 shows that
n
o
⊥
g+ F <d = σ ∈ S<d | σ (x) y F = σ (y) y F = σ y F = 0 .
Since Ann (F ) = (αβ, αd − β d ), we see that
g+ F
⊥
<d
= (αβ)2<d .
Now we proceed to the description of G+ ·F . It is clear that every f ∈ G+ ·F satisfies the
conditions given. Conversely, suppose that f ∈ P satisfies these conditions: the leading form of
f is F and HApolar(f ) = HApolar(F ) . If d = 2, then the claim follows from Example 2.15, so we
may assume d > 3. By applying αd−2 and β d−2 to f , we see that x2 and y 2 are leading forms of
partials of f . Since HApolar(f ) (2) = 2, these are the only leading forms of partials of degree two.
Let G be the leading form of f −F . Suppose that G contains a monomial xa y b , where a, b > 2.
Then αa−1 β b−1 y f = xy + l, where l is linear, then we get a contradiction with the conclusion
of the previous paragraph.
Since G contains no monomials of the form xa y b with a, b > 2, we see that G is annihilated
⊥
by (g+ F ) , so it lies in g+ F . By Corollary 3.4 we may find u ∈ G+ such that deg(uf − F ) <
deg(f − F ). Thus, replacing f by uf we lower the degree of f − F . Repeating, we arrive to the
point where f − F = 0, so that f = F .
The analysis made in Example 3.5 may be generalized to obtain the following Proposition 3.6.
Proposition 3.6. Let f ∈ P be a polynomial with leading form F . Let I = Ann (F ). Fix an
integer t > 0 and assume that
1. dimk Apolar (f ) = dimk Apolar (F ).
⊥
2. we have (g+ F )i = Ii2 for all i satisfying t 6 i 6 d − 1.
Then there is an element F + g ∈ G+ · f such that deg(g) < t. Equivalently, Apolar (f ) ≃
Apolar (F + g) for some polynomial g of degree less than t.
Proof. We apply induction with respect to deg(f − F ). If deg(f − F ) < t then we are done.
Otherwise, it is enough to find u ∈ G+ such that deg(uf − F ) < deg(f − F ).
Since Apolar (F ) is a quotient of gr Apolar (f ), the first condition implies that
Apolar (F ) ≃ gr Apolar (f ) ,
Thus for every element i ∈ I = Ann (F ) we may find σ ∈ S such that (i + σ) y f = 0 and
ord (σ) > ord (i).
Let G be the leading form of f − F and t 6 r 6 d − 1 be its degree. We will now prove
that G ∈ g+ F . By assumption, it is enough to show that Ir2 annihilates G. The ideal I is
homogeneous so it is enough to show that for any elements i, j ∈ I such that deg(ij) = r we
have
(ij) y G = 0.
Take σ ∈ S such that (j − σ) y f = 0. Then ord (iσ) > ord (ij) = r = deg(f − F ), thus
(iσ) y f = (iσ) y F = 0. Therefore,
(ij) y G = (ij) y (F + G) = (ij) y f = i(j − σ) y f = 0.
and the claims follow.
14
Example 3.7 (Rank n). Let f ∈ P = kdp [x1 , . . . , xn ] be a polynomial of degree d > 4 with
leading form F = xd1 + . . . + xdn and Hilbert function HApolar(f ) = (1, n, n, n, . . . , n, 1). Then the
apolar algebra of f is isomorphic to an apolar algebra of F + g, where g is a polynomial of degree
at most three.
Indeed, let I = Ann (F ). Then the generators of I<d are αi αj for i 6= j, thus the ideals
⊥
I<d and (I 2 )<d are monomial. Also the ideal (g+ F ) is monomial by its description from
Proposition 2.20. The only monomials of degree at least 4 which do not lie in I 2 are of the form
⊥
⊥
αd−1
αj and these do not lie in (g+ F ) . Therefore Ii2 = (g+ F )i for all 4 6 i 6 d − 1. The
i
assumptions of Proposition 3.6 are satisfied with t = 4 and the claim follows.
It is worth noting that for every g of degree at most three the apolar algebra of F + g
has Hilbert function (1, n, n, . . . , n, 1), so the above considerations give a full classification of
polynomials in P with this Hilbert function and leading form F .
P
⊥
Since (g+ F )3 = span hαi αj αl | i < j < li, we may also assume that f = F + i<j<l λijl xi xj xl
for some λijl ∈ k, see also Example 3.27.
3.2
Compressed algebras and generalisations
In this section we use Proposition 2.18 to obtain a generalisation of wonderful results of Elias
and Rossi on being canonically graded (see Corollary 3.14) and answer their question stated in
[ER15, Remark 3.6].
Example 2.16 is concerned with a degree three polynomial f such that the Hilbert function
of Apolar (f ) is maximal i.e. equal to (1, n, n, 1) for n = HS (1). Below we generalise the results
obtained in this example to polynomials of arbitrary degree.
Recall that a local Artinian Gorenstein algebra A of socle degree d (see Section 2.3) is called
compressed if
i+n−1
d−i+n−1
HA (i) = min (HS (i), HS (d − i)) = min
,
for all i = 0, 1, . . . , d.
i
d−i
Here we introduce a slightly more general notation.
Definition 3.8 (t-compressed). Let A = S/I be a local Artinian Gorenstein algebra of socle
degree d. Let t > 1. Then A is called t-compressed if the following conditions are satisfied:
1. HA (i) = HS (i) = i+n−1
for all 0 6 i 6 t,
i
2. HA (d − 1) = HS (1).
Example 3.9. Let n = 2. Then HA = (1, 2, 2, 1, 1) is not t-compressed, for any t. The
function HA = (1, 2, 3, 2, 2, 2, 1) is 2-compressed. For any sequence ∗ the function (1, 2, ∗, 2, 1) is
1-compressed.
Note that it is always true that HA (d − 1) 6 HA (1) 6 HS (1), thus both conditions above
assert that the Hilbert function is maximal possible. Therefore they are open in P6d .
Remark 3.10. The maximal value of t, for which t-compressed algebras exists, is t = ⌊d/2⌋.
Every compressed algebra is t-compressed for t = ⌊d/2⌋ but not vice versa. If A is graded, then
HA (1) = HA (d − 1), so the condition HA (d − 1) = HS (1) is satisfied automatically.
The following technical Remark 3.11 will be useful later. Up to some extent, it explains the
importance of the second condition in the definition of t-compressed algebras.
Remark 3.11. Let A = Apolar (f ) be a t-compressed algebra with maximal ideal mA . We have
d−1
d−1
d
d
≃ md−1 f
dim P61 = HA (d − 1) + HA (d) = dim md−1
A /mA + dim mA = dim mA . Moreover mA
d−1
as linear spaces and m f ⊆ P61 . Thus
md−1 f = P61 .
15
The definition of t-compressed algebras explains itself in the following Proposition 3.12.
Proposition 3.12. Let f ∈ P be a polynomial of degree d > 3 and A be its apolar algebra.
Suppose that A is t-compressed. Then the G+ -orbit of f contains f + P6t+1 . In particular
f>t+2 ∈ G+ ·f , so that Apolar (f ) ≃ Apolar (f>t+2 ).
Proof. First we show that P6t+1 ⊆ g+ f , i.e. that no non-zero operator of order at most t + 1 lies
⊥
in (g+ f ) . Pick such an operator. By Proposition 2.20 it is not constant. Let σ ′ be any of its
non-zero partial derivatives. Proposition 2.20 asserts that deg(σ ′ y f ) 6 1. Let ℓ := σ ′ y f . By
Remark 3.11 every linear polynomial is contained in md−1 f . Thus we may choose a δ ∈ md−1
such that δ y f = ℓ. Then (σ ′ − δ) y f = 0. Since d > 3, we have d − 1 > ⌊d/2⌋ > t, so that σ − δ
is an operator of order at most t annihilating f . This contradicts the fact that HA (i) = HS (i)
for all i 6 t. Therefore P6t+1 ⊆ g+ f .
Second, pick a polynomial g ∈ f + P6t+1 . We prove that g ∈ G+ · f by induction on
deg(g − f ). The top degree form of g − f lies in g+ f . Using Corollary 3.4 we find ϕ ∈ G+ such
that deg(ϕ∨ (g) − f ) < deg(g − f ).
For completeness, we state the following consequence of the previous result.
Corollary 3.13. Let f ∈ P be a polynomial of degree
d > 3 and A be its apolar algebra. Suppose
that A is compressed. Then A ≃ Apolar f>⌊d/2⌋+2 .
Proof. The algebra A is ⌊d/2⌋-compressed and the claim follows from Proposition 3.12.
As a corollary we reobtain the result of Elias and Rossi, see [ER15, Thm 3.1].
Corollary 3.14. Suppose that A is a compressed Artinian Gorenstein local k-algebra of socle
degree d 6 4. Then A is canonically graded i.e. isomorphic to its associated graded algebra gr A.
Proof. The case d 6 2 is easy and left to the reader. We assume d > 3, so that 3 6 d 6 4.
Fix n = HA (1) and choose f ∈ P = kdp [x1 , . . . , xn ] such that A ≃ Apolar (f ). This is possible
by the existence of standard form, see [Iar94, Thm 5.3AB] or [IM15]. Let fd be the top degree
part of f . Since ⌊d/2⌋ + 2 = d, Corollary 3.13 implies that fd ∈ G+ ·f . Therefore the apolar
algebras of f and fd are isomorphic. The algebra Apolar (fd ) is a quotient of gr Apolar (f ). Since
dimk gr Apolar (f ) = dimk Apolar (f ) = dimk Apolar (fd ) it follows that
Apolar (f ) ≃ Apolar (fd ) ≃ gr Apolar (f ) ,
which was to be proved.
The above Corollary 3.14 holds under the assumptions that k is algebraically closed and of
characteristic not equal to 2 or 3. The assumption that k is algebraically closed is unnecessary as
proven for cubics in Example 2.16, the cases of quartics is similar. Surprisingly, the assumption
on the characteristic is necessary.
Example 3.15 (compressed cubics in characteristic two). Let k be a field of characteristic two.
Let f3 ∈ P3 be a cubic form such that HApolar(f3 ) = (1, n, n, 1) and α21 y f3 = 0. Then there is
a degree three polynomial f with leading form f3 , whose apolar algebra is compressed but not
canonically graded.
Indeed, take σ = α21 . Then all derivatives of σ are zero because the characteristic is two. By
Proposition 2.18 the element σ lies in (gf3 )⊥ . Thus gf3 does not contain P62 and so G·f3 does
not contain f3 + P62 . Taking any f ∈ f3 + P62 outside the orbit yields the desired polynomial.
[2]
In fact, one can explicitly check that f = f3 + xi is an example.
A similar example shows that over a field of characteristic three there are compressed quartics
which are not canonically graded.
16
Compressed algebras of socle degree d > 5 in two variables. As noted in [ER15, Example 3.4], the claim of Corollary 3.14 is false for d = 5. Below we explain this from the point of
view of our theory. First, we give an example of a compressed algebra of socle degree five, which
is not canonically graded.
Example 3.16 (HA = (1, 2, 3, 3, 2, 1), special). Let n = 2, P = kdp [x1 , x2 ] and S = k[[α1 , α2 ]].
Take F = x31 x22 ∈ P and A = Apolar (F ) = S/(α41 , α32 ). Since x21 , x1 x2 , x22 are all partials of F
we have HA (2) = 3 and the algebra A is compressed. It is crucial to note that
α42 ∈ (gF )⊥ .
(12)
Indeed, the only nontrivial derivative of this element is 4α32 ∈ Ann (F ), so (12) follows from
Proposition 2.18. Therefore G·F is strictly contained in F + P64 . Pick any element f ∈ F + P64
not lying in G·F . The associated graded of Apolar (f ) is equal to Apolar (F ) but Apolar (f ) is
not isomorphic to Apolar (F ), thus Apolar (f ) is not canonically graded. In fact one may pick
f = F + x42 , so that Apolar (f ) = S/(α41 , α32 − α31 α2 ).
On the positive side, Corollary 3.13 implies that every compressed algebra of socle degree 5 is
isomorphic to Apolar (f ) for some f = f5 + f4 . Thus we may always remove the cubic part of f .
But even more is true: a general compressed algebra of socle degree five is canonically graded,
as shown in Example 3.17 below.
Example 3.17 (HA = (1, 2, 3, 3, 2, 1), general). Let f ∈ P = kdp [x1 , x2 ] be a general polynomial
of degree five with respect to the natural affine space structure on kdp [x1 , x2 ]65 . Then Apolar (f )
and Apolar (f5 ) are compressed by [Iar84]. The ideal Ann (f5 ) is a complete intersection of a
cubic and a quartic. Since we assumed that f is general, we may also assume that the cubic
generator c of Ann (f5 ) is not a power of a linear form.
We claim that Apolar (f ) is canonically graded. Equivalently, we claim that f ∈ G·f5 . It is
enough to show that
G+ ·f5 = f5 + P64 .
(13)
⊥
We will show that (g+ f5 )64 = 0. This space is spanned by homogeneous elements. Suppose
that this space contains a non-zero homogeneous σ. Proposition 2.20 implies that σ is nonconstant and that all partial derivatives of σ annihilate f5 . These derivatives have degree at
most three, thus they are multiples of the cubic generator c of Ann (f5 ). This implies that all
partial derivatives of σ are proportional, so that σ is a power of a linear form. Then also c is
a power of a linear form, which contradicts earlier assumption. We conclude that no non-zero
⊥
σ ∈ (g+ f5 )64 exists. Now, the equation (13) follows from Corollary 3.4.
Example 3.18 (general polynomials in two variables of large degree). Let F ∈ P = kdp [x1 , x2 ]
be a homogeneous form of degree d > 9 and assume that no linear form annihilates F . The ideal
Ann (F ) = (q1 , q2 ) is a complete intersection. Let di := deg qi for i = 1, 2 then d1 + d2 = d + 2.
Since d1 , d2 > 2, we have d1 , d2 6 d.
We claim that the apolar algebra of a general (in the sense explained in Example 3.17)
polynomial f ∈ P with leading form F is not canonically graded. Indeed, we shall prove that
G·F ∩ (F + P<d ) is strictly contained in F + P<d which is the same as to show that K := (gF )⊥
<d
is non-zero.
The sequence (7) from Remark 2.19 becomes
0 → (K/I 2 )d−1 → (A[−d1 ] ⊕ A[−d2 ])d−1 → (A ⊕ A)d−2 ,
where the rightmost map has degree −1. Note that
dim (A[−d1 ] ⊕ A[−d2 ])d−1 = (d − 1 − d1 + 1) + (d − 1 − d2 + 1) = 2d − (d1 + d2 ) = d − 2.
By assumption d − 2 > 6 = dim (A ⊕ A)d−2 , so that (K/I 2 )d−1 6= 0 and our claim follows.
17
The set of canonically graded algebras is constructible but not necessarily open nor
closed. Again consider P6d as an affine space with Zariski topology. Suppose that there exists
a polynomial of degree d whose apolar algebra is not canonically graded. Elias and Rossi asked
in [ER15, Remark 3.6] whether in this case also the apolar algebra of a general polynomial of
this degree is not canonically graded. The following proposition answers this question negatively.
Proposition 3.19. Fix the degree d and the number of variables n = dim S. Consider the set
G ⊂ P6d of dual socle generators of canonically graded algebras of socle degree d. This set is
irreducible and constructible, but in general it is neither open nor closed.
Proof. The set G = G·Pd is the image of G×Pd , thus irreducible and constructible. Example 3.16
and Example 3.17 together show that for n = 2, d = 5 this set is not closed. Example 3.18 shows
that for n = 2, d > 9 this set is not open.
In view of Proposition 3.19 it is natural to ask for which degrees and numbers of variables
the answer to the question above is positive; the apolar algebra of a general polynomial in P6d is
a canonically graded algebra. Example 3.17 shows that this happens for n = 2, d = 5. However,
the list of cases where the answer may be positive is short, as we shall see now.
Proposition 3.20. Assume that k is a field of characteristic zero. Suppose that n = dim S and
d are such that the apolar algebra of a general polynomial in P6d is canonically graded. Then
(n, d) belongs to the following list:
1. d 6 4 and n arbitrary,
2. d = 5 and n 6 6,
3. d = 6 and n = 2,
4. d arbitrary and n = 1.
Proof. Fix n, d outside the above list and suppose that G · Pd is dense in P6d . Since we are
over a field of characteristic zero, the tangent map G × Pd → P6d at a general point (g, F ) is
surjective. Then by G-action, this map is surjective also at (1, F ). Its image is gF + Pd . Thus
gF + Pd = P6d .
(14)
Let us now analyse gF . By Proposition 2.18 the equations of gF are given by σ ∈ S satisfying
σ y F = 0 and
σ (i) y F = 0 for all i = 1, . . . , n.
(15)
Consider σ ∈ S homogeneous of degree d − 1. The space Ann (F )d−2 is of codimension
HApolar(F ) (d − 2) = HApolar(F ) (2) 6 n+1
. Thus for fixed i the condition σ (i) y F = 0 amounts
2
linear conditions on the coefficients of σ. Summing over i = 1, . . . , n we get
to at most n+1
2
n+1
n · 2 conditions. Now, if (n, d) is outside the list, then
n+1
n−1+d−1
n·
<
,
2
d−1
P
so that there exists a non-zero σ satisfying (15). Since σ = (d − 1)−1 αi σ (i) also σ y F = 0.
Thus σ ∈ (gF )⊥ is a non-zero element. Therefore gF does not contain the whole P6d−1 , which
contradicts (14).
We have some computational evidence that in all cases listed in Proposition 3.20 a general
polynomial indeed gives a canonically graded algebra. Of course for d 6 4 it is the result of Elias
and Rossi reproved in Corollary 3.14. We put forward the following conjecture.
18
Conjecture 3.21. Assume that k is a field of characteristic not equal to 2, 3, 5. The pairs (n, d)
such that the apolar algebra of a general polynomial in P6d is canonically graded are precisely the
pairs listed below:
1. d 6 4 and n arbitrary,
2. d = 5 and n 6 6,
3. d = 6 and n = 2.
4. d arbitrary and n = 1.
Here n = dim S is the number of variables.
We have some supporting evidence in all cases d = 5, n 6 6 and d = 6, n = 2. Namely, for a
pseudo-randomly chosen form F a machine computation of (gF )⊥ via Proposition 2.18 reveals
that
gF ⊃ P6d−1 .
Thus arguing by semicontinuity we get that the map G × Pd → P6d is dominating and the
conjecture follows. However we find the use of computer here far from satisfactory, since it dims
the reasons for this result. Hopefully someone would invent a computer-free proof.
Improvements using symmetric decomposition. The condition HA (d − 1) = n in definition of t-compressed algebras may be slightly weakened, which is sometimes useful. The price
one pays is a more technical assumption. We present the result below. An example of use of
Proposition 3.22 is given in Example 3.29. For information on the symmetric decomposition ∆r
of the Hilbert function, see [Iar94], [CJN15] or [BJMR14, Section 2]. In short, the Hilbert
Pd function H of an apolar algebra of socle degree d admits a canonical decomposition H = i=0 ∆i ,
where ∆i is a vector of length d − i which is symmetric: ∆i (j) = ∆i (d − i − j) for all j.
Proposition 3.22. Let f ∈ P be a polynomial of degree d > 3 and A be its apolar algebra. Let
ƥ be the symmetric decomposition of HA . Suppose that
for all 0 6 r 6 t.
1. HA (r) = HS (r) = r+n−1
r
2. ∆r (1) = 0 for all d − 1 − t 6 r.
Then the G+ -orbit of f contains f + P6t+1 . In particular f>t+2 ∈ G+ ·f .
Proof. In the notation of [BJMR14, Section 2] or [CJN15] we see that
n
o
d−1−a
∆a (1) = dim Ca,1 (1)/Ca−1,1 (1), where Ca,1 (1) = linear polynomials in m
f .
Moreover, the assumption HA (1) = n guarantees that Cd−2,1 (1) = P61 is the full space of linear
polynomials. The assumption ∆r (1) = 0 for all d − 1 − t 6 r 6 d − 2 shows that
P61 = Cd−2,1 (1) = Cd−3,1 (1) = . . . = Cd−2−t,1 (1) = mt+1 f.
(16)
Thus, for every ℓ ∈ P61 we have a δ of order greater than t such that δ y f = ℓ.
Now we repeat the proof of Proposition 3.12 with one difference: instead of referring to
Remark 3.11 we use Equation (16) to obtain, for a given linear form ℓ, an element δ ∈ S of order
greater than t and such that δ y f = ℓ.
19
3.3
Further examples
Below we present some more involved examples, which employ the tools developed in the previous
section.
Example 3.23 (Hilbert function (1, 2, 2, 1)). By the result of Elias and Rossi (Proposition 3.14)
every apolar algebra
(1, 2, 2, 1) is canonically graded. Then it is isomorphic
with Hilbert function
to Apolar x3 + y 3 or Apolar x2 y and these algebras are not isomorphic.
The example was treated in [CN07, p. 12] and [ER12, Prop 3.6].
Example 3.24 (Hilbert function (1, 2, 2, 2, 1)). Let f ∈ P64 be a polynomial such that HApolar(f ) =
(1, 2, 2, 2, 1). If f4 = x4 + y 4 then if fact Apolar (f ) ≃ Apolar (f4 ) by Example 3.5, so we may
assume f4 = x3 y. Since f is 1-compressed we may also assume f62 = 0, whereas by Corollary 3.4
together with Example 3.2 we may assume f3 = cy 3 for c ∈ k. Thus
Apolar (f ) ≃ Apolar x3 y + cy 3 .
By multiplying variables by suitable constants, we may assume c = 0 or c = 1 and obtain three
possibilities:
f = x4 + y 4 ,
f = x3 y,
f = x3 y + y 3 .
Note that the three types appearing above
are pairwise non-isomorphic.
Indeed, an isomorphism
may only occur between Apolar x3 y and Apolar x3 y + y 3 . Suppose that such exists, so that
there is ϕ ∈ G such that ϕ∨ (x3 y) = x3 y + y 3 . Write ϕ = gu, for g linear and u ∈ G+ . Then
g preserves x3 y and a direct check shows that g is diagonal. Then u∨ (x3 y) = x3 y + cy 3 for
some non-zero c ∈ k and y 3 = c−1 tdf(u(x3 y) − x3 y) ∈ g+ (x3 y), by Proposition 3.3. This is a
contradiction with Example 3.2.
This example was analysed, among others, in [Cas10], where Casnati classifies all Artinian
Gorenstein algebras of length at most 9. See especially [Cas10, Theorem 4.4].
Example 3.25 (Hilbert function (1, 2, 2, . . . , 2, 1), general). Consider the set of polynomials
f ∈ P = kdp [x, y]6d such that HApolar(f ) = (1, 2, . . . , 2, 1) with d − 1 twos occurring. This set
is irreducible in Zariski topology, by [CJN15, Proposition 4.8] which uses [Iar77, Theorem 3.13].
The leading form of a general element of this set has, up to a coordinate change, the form xd +y d .
Example 3.5 shows that the apolar algebra of such an element is canonically graded.
Remark 3.26. In general, there are d − 1 isomorphism types of almost-stretched algebras of
socle degree d and with Hilbert function (1, 2, 2, . . . , 2, 1) as proved in [Cas10, Theorem 4.4]; see
also [EV11, Remark 5., p. 447]. The claim is recently generalised by Elias and Homs, see [EH16].
Example 3.27 (Hilbert function (1, 3, 3, 3, 1)). Consider now a polynomial f ∈ P = kdp [x, y, z]
whose Hilbert function is (1, 3, 3, 3, 1). Let F denote the leading form of f . By [LO13] or [CJN15,
Prop 4.9] the form F is linearly equivalent to one of the following:
F1 = x4 + y 4 + z 4 ,
F2 = x3 y + z 4 ,
F3 = x2 (xy + z 2 ).
Since Apolar (f ) is 1-compressed, we have Apolar (f ) ≃ Apolar (f>3 ); we may assume that
the quadratic part is zero. In fact by the explicit description of top degree form in Proposition 3.3
we see that
G+ ·f = f + g+ F + P62 .
Recall that G/G+ is the product of the group of linear transformations and k∗ acting by multiplication.
⊥
The case F1 . Example 3.7 shows that (g+ F )63 is spanned by αβγ. Therefore we may assume
f = F1 + c · xyz for some c ∈ k. By multiplying variables by suitable constants and then
multiplying whole f by a constant, we may assume c = 0 or c = 1. As before, we get two
non-isomorphic algebras. Summarising, we got two isomorphism types:
f1,0 = x4 + y 4 + z 4 ,
f1,1 = x4 + y 4 + z 4 + xyz.
Note that f1,0 is canonically graded, whereas f1,1 is a complete intersection.
20
⊥
The case F2 . We have Ann (F2 )2 = (αγ, β 2 , βγ), so that (g+ F2 )63 = span β 3 , β 2 γ . Thus we
may assume f = F2 + c1 y 3 + c2 y 2 z. As before, multiplying x, y and z by suitable constants we
may assume c1 , c2 ∈ {0, 1}. We get four isomorphism types:
f2,00 = x3 y + z 4 ,
f2,10 = x3 y + z 4 + y 3 ,
f2,01 = x3 y + z 4 + y 2 z,
f2,11 = x3 y + z 4 + y 3 + y 2 z.
To prove that the apolar algebras are pairwise non-isomorphic one shows that the only linear
maps preserving F2 are diagonal and argues as an in Example 3.24 or as described in the case of
F3 below.
The case F3 . We have Ann (F3 )2 = (β 2 , βγ, αβ − γ 2 ) and
⊥
g+ F3 63 = span β 2 γ, β 3 , αβ 2 − 2βγ 2 .
We may choose span y 3 , y 2 z, yz 2 as the complement of g+ F3 in P3 . Therefore the apolar algebra
of each f with top degree form F3 is isomorphic to the apolar algebra of
f3,∗ = x3 y + x2 z 2 + c1 y 3 + c2 y 2 z + c3 yz 2
and two distinct such polynomials f3,∗1 and f3,∗2 lie in different G+ -orbits. We identify the set
of G+ -orbits with P3 /g+ F3 ≃ span y 3 , y 2 z, yz 2 . We wish to determine isomorphism classes,
that is, check which such f3,∗ lie in the same G-orbit. A little care should be taken here, since
G-orbits will be bigger than expected.
Recall that G/G+ ≃ GL(m/m2 ) × k∗ preserves the degree. Therefore, it is enough to look at
the operators stabilising F3 . These are c · g, where c ∈ k ∗ is a constant and g ∈ GL(m/m2 ) stabilises span hF3 i, i.e. g∨ (span hF3 i) = span hF3 i. Consider such a g. It is a linear automorphism
of P and maps Ann (F ) into itself. Since β(λ1 β + λ2 γ) for λi ∈ k are the only reducible quadrics
in Ann (F3 ) we see that g stabilises span hβ, γi, so that g∨ (x) = λx for a non-zero λ. Now it is
straightforward to check directly that the group of linear maps stabilising span hF3 i is generated
by the following elements
1. homotheties: for a fixed λ ∈ k and for all linear ℓ ∈ P we have g ∨ (ℓ) = λℓ.
2. for every a, b ∈ k with b 6= 0, the map ta,b given by
ta,b (x) = x, ta,b (y) = −a2 x + b2 y − 2abz, ta,b (z) = ax + bz.
which maps F3 to b2 F3 .
The action of ta,b on P3 /g+ F3 in the basis (y 3 , y 2 z, yz 2 ) is given by the matrix
6
b
0
0
−6ab5
b5
0
9 4
27 2 4
− 2 ab b4
2 a b
Suppose that f3,∗ = x3 y + x2 z 2 + c1 y 3 + c2 y 2 z + c3 yz 2 has c1 6= 0. The above matrix shows that
we may choose a and b and a homothety h so that
(h ◦ ta,b )(f3,∗ ) = c(x3 y + x2 z 2 + y 3 + c3 yz 2 ),
where c 6= 0, c3 ∈ {0, 1}.
Suppose c1 = 0. If c2 6= 0 then we may choose a, b and λ so that (h◦ta,b )(f3,∗ ) = x3 y +x2 z 2 +y 2 z.
Finally, if c1 = c2 = 0, then we may choose a = 0 and b, λ so that c3 = 0 or c3 = 1. We get at
most five isomorphism types:
f3,100 = x3 y + x2 z 2 + y 3 ,
f3,001 = x3 y + x2 z 2 + yz 2 ,
f3,101 = x3 y + x2 z 2 + y 3 + yz 2 ,
f3,010 = x3 y + x2 z 2 + y 2 z,
f3,000 = x3 y + x2 z 2 .
By using the explicit description of the G action on P3 /g+ F3 one checks that the apolar algebras
of the above polynomials are pairwise non-isomorphic.
21
Conclusion: There are 11 isomorphism types of algebras with Hilbert function (1, 3, 3, 3, 1).
We computed the tangent spaces to the corresponding orbits in characteristic zero, using a
computer implementation of the description in Proposition 2.18. The dimensions of the orbits
are as follows:
orbit
orbit
dimension
G·(x4 + y 4 + z 4 + xyz)
G·(x4 + y 4 + z 4 )
G· x3 y + z 4 + y 3 + y 2 z
G· x3 y + z 4 + y 3
G· x3 y + z 4 + y 2 z
G· x3 y + z 4
29
28
28
27
27
26
G·
G·
G·
G·
G·
x3 y + x2 z 2 + y 3 + yz
x3 y + x2 z 2 + y 3
x3 y + x2 z 2 + y 2 z
x3 y + x2 z 2 + yz 2
x3 y + x2 z 2
dimension
2
27
26
26
25
24
The closure of the orbit of f1,1 = x4 + y 4 + z 4 + xyz is contained in GL3 (x4 + y 4 + z 4 ) + P63 ,
which is irreducible of dimension 29. Since the orbit itself has dimension 29 it follows that it
is dense inside. Hence the orbit closure contains GL3 (x4 + y 4 + z 4 ) + P63 . Moreover, the set
GL3 (x4 + y 4 + z 4 ) is dense inside the set σ3 of forms F whose apolar algebra has Hilbert function
(1, 3, 3, 3, 1). Thus the orbit of f1,1 is dense inside the set of polynomials with Hilbert function
(1, 3, 3, 3, 1). Therefore, the latter set is irreducible and of dimension 29.
It would be interesting to see which specializations between different isomorphism types are
possible. There are some obstructions. For example, the GL3 -orbit of x3 y + x2 z 2 has smaller
dimension than the GL3 -orbit of x3 y + z 4 . Thus x3 y + x2 z 2 + y 3 + yz 2 does not specialize to
x3 y + z 4 even though its G-orbit has higher dimension.
Example 3.28 (Hilbert function (1, 3, 4, 3, 1)). Consider now a polynomial f ∈ kdp [x, y, z] whose
apolar algebra has Hilbert function H = (1, 3, 4, 3, 1). Let F be the leading form of f . Since H
is symmetric, the apolar algebra of F also has Hilbert function H. In particular it is annihilated
by 2-dimensional space of quadrics. Denote this space by Q. Suppose that the quadrics in Q
do not share a common factor, then they are a complete intersection. By looking at the Betti
numbers, we see that Ann (F ) itself is a complete intersection of Q and a cubic. From this point
of view, the classification of F using our ideas seems ineffective compared to classifying ideals
directly. Therefore we will not attempt a full classification. Instead we show that there are
infinitely many isomorphism types and discuss non-canonically graded algebras.
Let V = V (Q) ⊂ P2 be the zero set of Q inside the projective space with coordinates α, β, γ.
Infinite family of isomorphism types. Consider
F = Fλ = λ1 x4 + λ2 y 4 + λ3 z 4 + λ4 (x + y + z)4 .
(17)
for some non-zero numbers λi . Then the Hilbert function of Apolar (f ) is (1, 3, 4, 3, 1) and V (Q)
is a set of four points, no three of them collinear. Suppose that Fλ and Fλ′ are in the same
G-orbit. By Remark 2.17 there is an element of GL(m/m2 ) mapping Fλ to Fλ′ . Such element
stabilizes V (Q), which is a set of four points, no three of them collinear. But the only elements
of GL(m/m2 ) stabilizing such set of four points are the scalar matrices. Therefore we conclude
that the set of isomorphism classes of Fλ is the set of quadruples λ• up to homothety. This set is
in bijection with (k∗ )4 /k∗ ≃ (k∗ )3 , thus infinite. It fact the set of Fλ with λ4 = 1 is a threefold in
the moduli space of finite algebras with a fixed basis, see [Poo08b] for construction of this space.
Non-canonically graded algebras. We now classify forms F such that all polynomials f
with leading form F lie in G·F . As in Example 3.27 we see that
G+ ·F = F + g+ F + P62 .
⊥
Thus we investigate (g+ F )3 using Proposition 2.20. Let us suppose it is non-zero and pick a
⊥
non-zero element σ ∈ (g+ F )3 . Then σ (i) ∈ Ann (F )2 = Q for all i. Since Q is two-dimensional
22
we see that the derivatives of σ are linearly dependent. Thus up to coordinate change we may
assume that σ ∈ k[α, β]. If σ has one-dimensional space of derivatives, then σ = α3 up to
coordinate change and 31 σ (1) = α2 annihilates Q. If σ has two-dimensional space of derivatives
Q, then Q intersects the space of pure squares in an non-zero element α2 , thus α2 annihilates F
⊥
in this case also. Conversely, if α2 annihilates F , then α3 lies in (g+ F )3 .
Summarizing, F + P63 ⊂ G · F if and only if no square of a linear form annihilates F .
For example, apolar algebras of all polynomials with leading form Fλ from (17) are canonically
graded. On the other hand, for F = xz 3 + z 2 y 2 + y 4 and fλ = F + ax3 the apolar algebra of fλ
is canonically graded if and only if a = 0.
Non-canonically graded algebras with Hilbert function (1, 3, 4, 3, 1) are investigated independently in [MR16], where more generally Hilbert functions (1, n, m, n, 1) are considered.
Any discussion of isomorphism types would be incomplete without tackling the example of
Hilbert function (1, 2, 2, 2, 1, 1, 1), which is the smallest example where infinitely many isomorphism types appear. It is also instructive as a non-homogeneous example. Strangely enough, the
argument is similar to the previous examples. We will use some standard tools to deal with the
dual socle generator, see [CJN15, Chap 3].
Example 3.29 (Hilbert function (1, 2, 2, 2, 1, 1, 1)). Consider any f ∈ P such that
HApolar(f ) = (1, 2, 2, 2, 1, 1, 1).
Then f is of degree 6. Using the standard form, see [Iar94, Theorem 5.3] or [CJN15, Section 3],
we may assume, after a suitable change of f , that
f = x6 + f64 .
If y 4 appears with non-zero coefficient in f64 , we conclude that f ∈ G+·(x6 +y 4 ) arguing similarly
as in Example 3.5. Otherwise, we may assume that β 2 is a leading form of an element of an
annihilator of f , so that the only monomials in f4 are x4 , x3 y and x2 y 2 . By subtracting a suitable
element of g+ x6 and rescaling by homotheties, we may assume
f = x6 + cx2 y 2 + f63 ,
for c = 0 or c = 1. If c = 0, then the Hilbert function of f is (1, ∗, ∗, 1, 1, 1, 1), which is a
contradiction; see [Iar94, Lem 1.10] or [Jel13, Lem 4.34] for details. Thus c = 1.
⊥
with order at most three. Every non-zero partial
Consider elements of g+ (x6 + x2 y 2 )
derivative τ of such an element σ has order at most two and satisfies deg(τ y f ) 6 1. This is only
possible if τ2 = β 2 , so that σ3 = β 3 up to a non-zero constant. Therefore we may assume
f = x6 + x2 y 2 + λy 3 + f62
for some λ ∈ k.
The symmetric decomposition of Apolar (f ) is ∆0 = (1, 1, 1, 1, 1, 1, 1), ∆1 = 0, ∆2 = (0, 1, 1, 1, 0),
∆3 = 0, ∆4 = 0. In particular ∆4 (1) = 0, so that by Proposition 3.22 we may assume f62 = 0
and
f = fλ = x6 + x2 y 2 + λy 3 .
Similarly as in the final part of Example 3.24 we see that for every λ ∈ k we get a distinct algebra
Aλ = Apolar (fλ ) where Aλ ≃ Aλ′ if and only if λ = λ′ . Thus in total we get |k| + 1 isomorphism
types. Note that the |k| types fλ form a curve in the moduli space of finite algebras with a fixed
basis, see [Poo08b] for information on this space. The point corresponding to y 6 + x4 does not
seem to be related to this curve.
23
References
[BJMR14] Alessandra Bernardi, Joachim Jelisiejew, Pedro Macias Marques, and Kristian Ranestad. The
dimension of cactus varieties of cubic forms. ArXiv e-print arXiv:1211.7306v4, 2014.
[BCP97]
Wieb Bosma, John Cannon, and Catherine Playoust. The Magma algebra system. I. The user
language. J. Symbolic Comput., 24(3-4):235–265, 1997. Computational algebra and number
theory (London, 1993).
[Bri77]
Joël Briançon. Description de Hilbn C{x, y}. Invent. Math., 41(1):45–89, 1977.
[BB14]
Weronika Buczyńska and Jarosław Buczyński. Secant varieties to high degree Veronese reembeddings, catalecticant matrices and smoothable Gorenstein schemes. J. Algebraic Geom.,
23:63–90, 2014.
[Cas10]
Gianfranco Casnati. Isomorphism types of Artinian Gorenstein local algebras of multiplicity
at most 9. Comm. Algebra, 38(8):2738–2761, 2010.
[CJN15]
Gianfranco Casnati, Joachim Jelisiejew, and Roberto Notari. Irreducibility of the Gorenstein
loci of Hilbert schemes via ray families. Algebra Number Theory, 9(7):1525–1570, 2015.
[CN07]
Gianfranco Casnati and Roberto Notari. On some Gorenstein loci in Hilb6(P4k ). J. Algebra,
308(2):493–523, 2007.
[CN09]
Gianfranco Casnati and Roberto Notari. On the Gorenstein locus of some punctual Hilbert
schemes. J. Pure Appl. Algebra, 213(11):2055–2074, 2009.
[CN11]
Gianfranco Casnati and Roberto Notari. On the irreducibility and the singularities of the
Gorenstein locus of the punctual Hilbert scheme of degree 10. J. Pure Appl. Algebra,
215(6):1243–1254, 2011.
[DGPS15] Wolfram Decker, Gert-Martin Greuel, Gerhard Pfister, and Hans Schönemann.
Singular 4-0-2 — A computer algebra system for polynomial computations.
http://www.singular.uni-kl.de, 2015.
[EI14]
Michael Eastwood and Alexander Isaev. Invariants of Artinian Gorenstein algebras and isolated hypersurface singularities. In Developments and retrospectives in Lie theory, volume 38
of Dev. Math., pages 159–173. Springer, Cham, 2014.
[ER93]
Richard Ehrenborg and Gian-Carlo Rota. Apolarity and canonical forms for homogeneous
polynomials. European J. Combin., 14(3):157–181, 1993.
[Eis95]
David Eisenbud. Commutative algebra, volume 150 of Graduate Texts in Mathematics.
Springer-Verlag, New York, 1995. With a view toward algebraic geometry.
[Eli15]
Joan Elias. Inverse-syst.lib, Singular library for computing Macaulay’s inverse systems. ArXiv
e-print arXiv:1501.01786, 2015.
[EH16]
Joan Elias and Roser Homs. On the analytic type of Artin algebras. Comm. Algebra,
44(6):2277–2304, 2016.
[ER12]
Joan Elias and Maria Evelina Rossi. Isomorphism classes of short Gorenstein local rings via
Macaulay’s inverse system. Trans. Amer. Math. Soc., 364(9):4589–4604, 2012.
[ER15]
Joan Elias and Maria Evelina Rossi. Analytic isomorphisms of compressed local algebras.
Proc. Amer. Math. Soc., 143(3):973–987, 2015.
[EV11]
Joan Elias and Giuseppe Valla. Isomorphism classes of certain Artinian Gorenstein algebras.
Algebr. Represent. Theory, 14(3):429–448, 2011.
[Ems78]
Jacques Emsalem. Géométrie des points épais. Bull. Soc. Math. France, 106(4):399–416, 1978.
[FK12]
Gregor Fels and Wilhelm Kaup. Nilpotent algebras and affinely homogeneous surfaces. Math.
Ann., 353(4):1315–1350, 2012.
[FH91]
William Fulton and Joe Harris. Representation theory, volume 129 of Graduate Texts in
Mathematics. Springer-Verlag, New York, 1991. A first course, Readings in Mathematics.
[Ger96]
Anthony V. Geramita. Inverse systems of fat points: Waring’s problem, secant varieties of
Veronese varieties and parameter spaces for Gorenstein ideals. In The Curves Seminar at
Queen’s, Vol. X (Kingston, ON, 1995), volume 102 of Queen’s Papers in Pure and Appl.
Math., pages 2–114. Queen’s Univ., Kingston, ON, 1996.
24
[GS]
Daniel R. Grayson and Michael E. Stillman. Macaulay2, a software system for research in
algebraic geometry. Available at http://www.math.uiuc.edu/Macaulay2/.
[Har10]
Robin Hartshorne. Deformation theory, volume 257 of Graduate Texts in Mathematics.
Springer, New York, 2010.
[Iar77]
Anthony Iarrobino. Punctual Hilbert schemes. Mem. Amer. Math. Soc., 10(188):viii+112,
1977.
[Iar81]
Anthony Iarrobino. Deformations of zero-dimensional complete intersections after M. Granger,
T. Gaffney. In Proceedings of the Week of Algebraic Geometry (Bucharest, 1980), volume 40
of Teubner-Texte zur Math., pages 92–105. Teubner, Leipzig, 1981.
[Iar84]
Anthony Iarrobino. Compressed algebras: Artin algebras having given socle degrees and
maximal length. Trans. Amer. Math. Soc., 285(1):337–378, 1984.
[Iar94]
Anthony Iarrobino. Associated graded algebra of a Gorenstein Artin algebra. Mem. Amer.
Math. Soc., 107(514):viii+115, 1994.
[IK99]
Anthony Iarrobino and Vassil Kanev. Power sums, Gorenstein algebras, and determinantal
loci, volume 1721 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1999. Appendix
C by Anthony Iarrobino and Steven L. Kleiman.
[IM15]
Anthony Iarrobino and Pedro Macias Marques. Symmetric Decomposition of the Associated
Graded Algebra of an Artinian Gorenstein Algebra. Preprint, 2015.
[Jel13]
Joachim Jelisiejew. Deformations of zero-dimensional schemes and applications (Msc thesis).
arXiv:1307.8108, 2013.
[LO13]
Joseph M. Landsberg and Giorgio Ottaviani. Equations for secant varieties of Veronese and
other varieties. Ann. Mat. Pura Appl. (4), 192(4):569–606, 2013.
[Mac94]
Francis S. Macaulay. The algebraic theory of modular systems. Cambridge Mathematical
Library. Cambridge University Press, Cambridge, 1994. Revised reprint of the 1916 original,
With an introduction by Paul Roberts.
[MR16]
Shreedevi Masuti and Maria Evelina Rossi. Artinian Level algebras of socle degree 4. 2016.
In preparation.
[Mat87]
Jerzy Matczuk. Relations between algebraic derivations and algebraic automorphisms of
algebras. Bull. Soc. Math. Belg. Sér. B, 39(3):359–372, 1987.
[MZ15]
Juan Migliore and Fabrizio Zanello. Stanley’s nonunimodal Gorenstein h-vector is optimal.
arXiv:1512.01433., 2015.
[Poo08a]
Bjorn Poonen. Isomorphism types of commutative algebras of finite rank over an algebraically
closed field. In Computational arithmetic geometry, volume 463 of Contemp. Math., pages
111–120. Amer. Math. Soc., Providence, RI, 2008.
[Poo08b]
Bjorn Poonen. The moduli space of commutative algebras of finite rank. J. Eur. Math. Soc.
(JEMS), 10(3):817–836, 2008.
[Sch73]
Michael Schlessinger. On rigid singularities. Rice Univ. Studies, 59(1):147–162, 1973. Complex
analysis, 1972 (Proc. Conf., Rice Univ., Houston, Tex., 1972), Vol. I: Geometry of singularities.
[Sta96]
Richard P. Stanley. Combinatorics and commutative algebra, volume 41 of Progress in Mathematics. Birkhäuser Boston Inc., Boston, MA, second edition, 1996.
25
| 0math.AC
|
Simulated Tornado Optimization
S. Hossein Hosseini∗ , Tohid Nouri† , Afshin Ebrahimi∗ , and S. Ali Hosseini‡
arXiv:1701.00736v1 [math.OC] 31 Dec 2016
∗
ICT Research Center, Faculty of Electrical Engineering, Sahand University of Technology, Tabriz, Iran
†
Faculty of Civil Engineering, University of Mohaghegh Ardabili, Ardabil, Iran
‡
Faculty of Industrial Engineering, Urmia University of Technology, Urmia, Iran
Emails: {h hosseini, aebrahimi}@sut.ac.ir, [email protected], [email protected]
Abstract—We propose a swarm-based optimization algorithm
inspired by air currents of a tornado. Two main air currents spiral and updraft - are mimicked. Spiral motion is designed
for exploration of new search areas and updraft movements
is deployed for exploitation of a promising candidate solution.
Assignment of just one search direction to each particle at each
iteration, leads to low computational complexity of the proposed
algorithm respect to the conventional algorithms. Regardless of
the step size parameters, the only parameter of the proposed
algorithm, called tornado diameter, can be efficiently adjusted by
randomization. Numerical results over six different benchmark
cost functions indicate comparable and, in some cases, better
performance of the proposed algorithm respect to some other
metaheuristics.
Index Terms—air currents, spiral, updraft, tornado diameter
I. I NTRODUCTION
Metaheuristic or evolutionary algorithms are a family of
optimization algorithms inspired by nature, art, or social
developments. Despite of their lack in convergence guarantee,
these kind of algorithms have been popular in various engineering applications and scientific researches because of their
simplicity and efficiency in finding global optimum solution
for relatively low-dimension problems. Metaheuristics can be
divided to single-solution approaches and multi-solution methods. The well-known examples of the single-solution approach
are simulated annealing [1] and tabu search [2] algorithms
in which one solution vector is evolved through strategic
steps. On the other hand, in multi-solution(agent) methods
a number of solutions are appointed and their relations are
defined in an evolutionary or optimistic way such that they
approach hopefully to a globally optimal solution. Hence, the
main difference among the algorithms in the multi-solution
approaches is in the kind of relations (motions) defined among
the solutions.
Multi-solution approaches are recognized by two prior
works on genetic algorithm (GA) [3] and particle swarm optimization (PSO) [4]. Other algorithms, for example, differential
evolution [5], imperialist competition algorithm (ICA) [6], and
teaching-learning-based optimization (TLBO) [7] are evolved
versions of the mentioned two algorithms. They evolve to
handle an smart balance between exploration and exploitation.
The main differences are in the update rule and selection
criteria. Specifically, in the swarm-based approaches the main
difference is in the exploration strategy, since a motion toward
a globally best solution(s) is a fixed part of these algorithms.
The exploration item in PSO is accomplished by assigning
a memory to the particles to save the best result of their
search. In the ICA, the exploration was attended by appointing
several imperialists selected from best solutions. Finally, in
the TLBO, the exploration is a part of an educational system
which is accomplished by interaction between pairs of the
students chosen at random. Our proposed algorithm is another swarm-based optimization method in which the relation
and movements among dull air particles of the tornado are
mimicked. At the proposed algorithm, the exploration rule
is designed inspired by spiral motion of the tornado. Our
motivation behind modeling of the tornado motions was its
powerful capability in the reduction of ambient temperature
(up to 15◦ C) by its mysterious air currents.
In fact, in formulation of a metaheuristic algorithm, there
is a wide spectrum between random search and deterministic
search where the intensities of exploration and exploitation
of an new algorithm is determined. It gives an opportunity
to design different algorithms to handle various tradeoffs in
facing real-word problems. Two main drawbacks among the
metaheuristic algorithms are run-time and parameter adjustability. At this work, both of them, with an emphasis on the
former, is addressed. Two search directions is defined to boost
the exploration and exploitation capability of the proposed
algorithm. Each one is assigned to just one particle. That
is despite of conventional methods, e.g. TLBO and PSO
algorithms, which apply both main motions to each particle. Consequently, in the proposed algorithm, computational
complexity is considerably reduced. Regardless of the step
size parameters, the only parameter of the algorithm is the
number of particles that should be assigned to each motion.
The simulation results indicate that, in most of the problem,
this parameter can be floated without significant degradation
in the performance. This simple self-adjustment works, since
the assignment is done at random.
II. T ORNADO M ORPHOLOGY
Tornado is a kind of extreme weather. The strongest wind
that could be formed in the earth is related to this phenomenon.
When a layer of warm and moist air is located under the cool
and dry air, due to the lower density of the warmer one it
attempts to climb up to the top of the cold air. Conversely
cold air descent to replace the rising air. If this process is
occurred quickly, spiral airflow (like a funnel) is formed that
is called Tornado [8]. At the beginning of the Tornado the axis
of the rotating airflow is horizontal. At the present of wind,
the rising warm and moist air tilts the rotating air to vertical.
A schematic model of Tornado is shown in Fig. 1 [9]. During
the Tornado, the warm air blows towards the tornado and
rises through the spiral paths. This continues until combination
of warm and cold air together and establishment of thermal
equilibrium. Two main and specific air currents of the tornado,
i.e. updraft and spiral motions, form the basis of our proposed
algorithm. Moreover, atmospheric turbulence is an important
part of these motions that has significant influence in the
proposed simulated tornado optimization (STO) algorithm.
Fig. 2. Demonstration of two main movements in the proposed algorithm,
regardless of the uncertainty in directions governed by turbulence. (d1 < d2 )
a pure exploitation. The mentioned formulation is a common
rule in the attraction-based algorithms [10]. On the other side,
as an special motion - designed to promote the exploration
property of the proposed algorithm - a spiral particle at position xi (i = 2, ..., k1 ) moves toward the position of another
spiral particle or the coldest one, i.e. xj , j = 1, ..., k1 and j 6=
i. This index is determined according to the following criteria:
j = arg min kxj − xi k2
Fig. 1.
s.t.
f (xj ) < f (xi )
(2)
where k.k2 denotes Euclidian norm. Hence, each particle
moves toward another nearest-better spiral particle or the
coldest one (see Fig. 2). Corresponding update rule for a spiral
motion can be formulated as:
Illustration of air currents in a tornado [9]
III. P ROPOSED A LGORITHM
Let the columns of matrix X = [x1 x2 ... xk ] be consist
of k solution vectors xi = [x1 , x2 , ..., xN ]T (i = 1, ..., k).
Each solution includes N decision variables. The solutions are
interpreted as the position of k air particles in a tornado. Their
corresponding cost function f (xi ) determines the temperature
of ith particle at position xi . The particles are divided into
three types of coldest (x1 ), spiral (i = 2, ..., k1 ), and
updraft particles (i = k1 + 1, ..., k), where 1 < k1 < k.
Fig. 2 demonstrates this terminology and illustrates the idea
behind the direction of motion for each particle. The coldest
particle (x1 ), in top of all other particles, is best solution at
each iteration. For a minimization problem that is a position
vector with minimum cost function among the particles. Its
replacements, after each iteration, simulates the storm motion
(see Fig. 1, top of the left-hand picture). Other positions
(xi , i = 2, ..., k) are randomly sorted at the columns of X
matrix to participate as a spiral or updraft particle.
Updraft particles (the particles inside the funnel of a real
tornado) move directly toward the coldest particle. The following relation formulates this kind of motion:
xi := xi + µ
(xj − xi )
∀ i = 2, . . . , k1
(3)
After each iteration, the coldest position is refreshed and
spiral-updraft assignments are renewed, randomly.
A. On the Turbulence Model
The Turbulence vector µ imposes an uncertainty to both
spiral and updraft motion. It is modeled by i.i.d. normal distribution with zero mean and unit variance. This distribution for
both kind of updates leads to more generality and acceptable
performance respect to the uniform distribution. In addition,
that is more general in dealing with different problems respect
to the realistic atmospheric turbulence models like log-normal
distribution utilized in the communication channel modeling
[11]. Although, that is important to mention, according to
our observations, log-normal distribution had shown some
promisingly better performance than the normal distribution,
at the case of unimodal problems. Nevertheless, for the sake of
generality, our simulations are done using normal distribution.
B. Tornado Diameter
i
i
x := x + µ
1
i
(x − x )
∀ i = k1 + 1, . . . , k
(1)
where
indicates an entry-wise multiplication and µ is the
turbulence vector, consist of N step sizes for each decision
variable. That is responsible for slight divergences from direct
motion toward the coldest particle x1 , in order to avoid from
As it is illustrated at the simulations (Fig. 8), the number of
spiral particles k1 can considerably influence the performance
of the proposed algorithm in solving different problems. Its
ratio respect to the total number of particles k is called tornado
diameter α = kk1 . Each problem has its own optimal value for
tornado diameter. As a parameter-free version of the proposed
algorithm, we suggest to float this parameter at each iteration
to be selected uniformly among the feasible range of [1, k).
Simulation results confirm the efficiency of the randomized
tornado diameter. The finalized procedure of the proposed
algorithm is summarized as follows:
Algorithm 1: Simulated Tornado Optimization (STO)
Generate initial position of k particles by random.
2 Evaluate the temperature (cost function) of the particles
and appoint the particle with lowest temperature as the
coldest particle
3 Select a random integer number for the parameter k1
distributed uniformly between [1, k).
4 Randomly assign k1 particles for spiral current and rest
for updraft current, then update their positions.
5 Repeat 2 to 4 until tornado is vanished (position of the
solutions are same).
1
C. On the Convergence and Complexity
The movement toward best solution (coldest particle) in
the updraft current is inherently a converging motion [10].
However, that has a risk of falling into a local minimum.
The spiral current boosts the exploration capability of the
algorithm. Akin to the rings of a chain, the spiral particles are
partially connected together and conduce an implicit motion
toward the best solution. Hence, the spiral motion itself is
a kind of converging attraction. Intuitively, the algorithm still
converges after adding the spiral current. However, an analysis
of convergence would have insight in choosing the step sizes.
It is worth to mention that the randomized assignments of
the motions at each iteration is not an essential part of the
algorithm. Assignment according to the temperature can also
be effective in some problems.
On the other hand, STO has low computational complexity,
mainly because the motion of each particle is just in one
direction at each iteration. That is despite of the algorithms like
TLBO and PSO in which the final direction of movement for
each particle at each iteration is determined by the computation
of superposition of two or more direction vectors.
IV. S IMULATION R ESULTS
At this section, optimization of six different benchmark
cost functions are investigated. These functions are shown in
Fig. 3 and their formulations, along with domains, optimal
solutions, and minimum costs are postponed to appendix.
These functions were chosen from [12]. Simulations are conducted through five experiments. Firstly, the trajectory of the
particles in the proposed algorithm is demonstrated. Secondly,
the effect of tornado diameter on the optimization performance
is investigated. Thirdly, influence of the randomization of
tornado diameter was considered in a quantitative comparison
with a fixed diameter version. Next, fourth experiment includes
some qualitative comparisons with with GA, PSO, and TLBO.
Finally, the performance of the STO was evaluated on two
higher dimension problems.
Fig. 3. Benchmark cost functions: (a) Eggholder, (b) Ripple, (c) Beale, (d)
Modified Rosenbrock, (e) Styblinski-Tang, (f) Rastrigin
In all of the experiments the number of particles or
population size was fixed on 40 for∗ all of the algorithms.
2
Distortion function is defined as kxkx−x̂k
, where x∗ is the
∗k
2
real global minimum and x̂ is the result of optimization.
One run of the optimization algorithms was regarded as a
perfect optimization, if the distortion was less than 10−4 . For
Modified Rosenbrock and Rastrigin functions this definition
was considered as a cost value less than 36 and 10−4 ,
respectively. Parameters of the PSO were set according to the
constriction coefficients with Φ1 = Φ2 = 2.05 [13]. In the GA,
0.8 of the population were selected for crossover and the rest
for mutation. Crossover coefficient and mutation probability
were tuned on 0.05 and 0.08, respectively. The TLBO is a
parameter-free algorithm.
At the first experiment, trajectory of the air particles on
artificial landscape of EggHolder function was demonstrated
through Fig. 4 to Fig. 7. The number of spiral particles was
fixed to be equal with that of updraft particles (including
coldest one as a spiral particle). At the second experiment,
influence of the parameter of tornado diameter was investigated. For this aim, the algorithm was implemented in whole
range of the possible number of the spiral particles k1 . Fig.
8 illustrates the effect of this number in the probability of
perfect optimization of three test functions. The success rate
was computed over 1000 trials at each values of the k1 . As
shown, EggHolder and Ripple functions are two examples of
the functions that behave at extreme; the EggHolder function
was efficiently optimized by high diameter tornados while
the optimization of Ripple function was more efficient when
diameter was small. In between there is Beale function which
had low sensitivity to the tornado diameter. The proposed
algorithm, had best performance in the optimization of Beale
function when the ratio of the two spiral and updraft particles
was in middle points.
Fig. 4. Initial Position of the particles; Star indicator shows the coldest particle and circles and triangles indicate spiral and updraft particles, respectively.
Fig. 8. Probability of perfect optimization verse tornado diameter (in each
diameter, the number of particles in each air current is fixed)
Fig. 5. Tornado at iteration 10: Updraft movements toward coldest particle
and spiral motion among neighborhoods
assignment of the kind of motion for each particle position,
allows to float the number of particles for each air current
at each iteration. This randomization is not degrading and
in most functions leads to a comparable performance with a
tuned version of the STO. Table I compares the probability
of perfect optimization at the case of randomized tornado
diameter (indicated by STO2 ) with the best result obtained
by tuning of the number of particles assigned to spiral motion
(indicated by STO1 ). As shown, the performances are near
(regardless of Modified Rosenbrock function). Roughly speaking, possibility of the randomization of parameter k1 at each
iteration indicates a promising self-adjustment capability of the
proposed algorithm. In addition, the performance of other three
algorithms are included at this table. The results are obtained
over 1000 trials and all of the algorithms are stopped after 100
iterations. As inferred, the proposed algorithm has comparable
quantitative performance in most of the problems and a better
one at the case of Modified Rosenbrock function. In the rest
of experiments, STO in the randomized diameter version was
evaluated, except of Fig. 12.
TABLE I
P ROBABILITY OF SUCCESS IN PERFECT OPTIMIZATION
Fig. 6. Tornado at iteration 20: coldest particle converges to coldest position
(global minimum)
STO1
STO2
PSO
GA
TLBO
Fig. 7. Tornado at iteration 30: All particles approach coldest position and
tornado vanishes
As mentioned in the previous section, possibility of random
EggHolder
0.95
0.91
0.24
0.18
0.97
Ripple
0.94
0.93
0.26
0.88
0.93
Beale
0.99
0.98
0.89
0.47
1
Rosen. Mod.
0.54 (k1 = 35)
0.40
0.28
0.18
0.28
Rast.(5D)
0.99
0.09
0.46
0.97
Fig. 9 to Fig. 12 show a qualitative comparison of the
algorithms on optimization of EggHolder, Ripple, Beale, and
Modified Rosenbrock functions, respectively. The convergence
curves were obtained by averaging over 50 independent runs.
As shown in the first three figures, the proposed algorithm has
comparable performance with the TLBO as it was expected
from the quantitative comparisons. In addition, STO in both
tuned and randomized diameter versions, has best performance
in the optimization of Modified Rosenbrock (Fig. 12).
Fig. 13 shows the percentage of perfect optimizations (dis-
Fig. 9.
Convergence curves for EggHolder function
Fig. 10.
tortion less than 10−4 ) for the Styblinski-Tang function respect
to the dimension of problem. In two dimensions, this function
consists of 3 local minimums and 1 global optimum (see Fig.
3). As shown in Fig. 13, all of the algorithms are completely
successful in finding the global minimum at 2 dimensions.
However, as the dimension increases, the PSO, GA, and TLBO
were trapped in the local minimums such that their probability
of perfect optimization dramatically decreases, while, the
STO survives to explore new solutions in significantly higher
dimensions. Maximum number of the allowed iterations for
all of the algorithms was 5000. It was a limiting factor for the
proposed algorithm, since it rarely fall into the local minimums
of this problem and a better performance was possible by more
iterations.
Convergence curves for Ripple25 function
Fig. 13. probability of perfect optimization of Styblinski-Tang function in
high dimensions
Fig. 11.
Fig. 12.
Convergence curves for Beale Function
In the other experiment, convergence curve of the STO
in optimization of Rastrigin function with 5 dimensions was
evaluated in Fig. 14. This function has a global minimum in
the origin, with the cost function vlaue of zero. The results
indicate a comparable performance of the proposed algorithm
with the TLBO, both in quality and quantity (see Table I). In
order to have a sense of the low computational complexity
of the proposed algorithm, a normalized run-time of the
algorithms for last experiment was provided at Fig. 15.
Convergence curves for Modified Rosenbrock function
Fig. 14.
Convergence curves for Rastrigin function in 5 dimensions
(d) Modified Rosenbrock Function:
f = 74 + 100(x2 − x21 )2 + (1 − x1 ) − 400e
−(x1 +1)2 +(x2 +1)2
0.1
−2 ≤ xi ≤ 2, fmin (−0.9, −0.95) = 34.37
————————————————
(e) Styblinski-Tang Function with n decision variables:
n
Fig. 15. Comparison of relative run-time of the algorithms on last experiment
V. C ONCLUSION
Synchronized air currents in a tornado leads to a significant
reduction in the ambient temperature. It is a desired paradigm
in optimization. Simulated tornado optimization (STO)
mimicked two main air currents of a tornado in movement
toward colder positions (better solutions). Participation of each
air particle in only one current, or equivalently movement in
just one search direction, led to low computational complexity
in the proposed algorithm. Hence, it would be useful in the
applications with time constraint. A simple self-adjustment
is possible by randomization of the number of particles
appointed to each current (randomization of the tornado
diameter). Numerical results indicated the efficiency of the
proposed algorithm in facing with different problems. As
a future direction of research, more improvements can be
achieved by adaptive reduction in the range of variations of
the diameter. Also, a more realistic step size mechanism or
turbulence model would be of great interest.
A PPENDIX
Problems:
(a) EggHolder Function:
r
p
x1
+ 47|−x1 sin x1 − (x2 + 47)
f = −(x1 +47) sin |x2 +
2
−512 ≤ xi ≤ 512, fmin (512, 404.2319) ≈ −959.64
————————————————
(b) Ripple25 Function:
f=
2
X
−e(−2 ln 2(
xi −0.1 2
0.8 ) )
sin6 (5πxi )
i=1
0 ≤ xi ≤ 1, fmin (0.1, 0.1) = −2
————————————————
(c) Beale Function:
f = (1.5 − x1 + x1 x2 )2 + (2.25 − x1 + x1 x22 )2
+(2.625 − x1 + x1 x32 )2
−4.5 ≤ xi ≤ 4.5, fmin (3, 0.5) = 0
————————————————
f=
1X 4
(x − 16x2i + 5xi ), − 5 ≤ xi ≤ 5
2 i=1 i
fmin (−2.903534, . . . , −2.903534) = −39.1661657037n
————————————————
(f) Rastrigin Function with n decision variables:
n
f=
1X 2
(x − 10cos(2πxi ) + 10)
2 i=1 i
−5.12 ≤ xi ≤ 5.12, fmin (0, . . . , 0) = 0
————————————————
R EFERENCES
[1] S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, “Optimization by
simulated annealing,” Science, vol. 220, no. 4598, pp. 671–680, 1983.
[2] F. Glover, “Tabu search part I,” ORSA Journal on Computing, vol. 1,
no. 3, pp. 190–206, 1989.
[3] J. H. Holland, Adaptation in Natural and Artificial Systems. Cambridge,
MA, USA: MIT Press, 1992.
[4] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Neural
Networks, 1995. Proceedings., IEEE International Conference on, vol. 4,
pp. 1942–1948 vol.4, Nov 1995.
[5] R. Storn and K. Price, “Differential evolution – a simple and efficient
heuristic for global optimization over continuous spaces,” Journal of
Global Optimization, vol. 11, no. 4, pp. 341–359, 1997.
[6] E. Atashpaz-Gargari and C. Lucas, “Imperialist competitive algorithm:
An algorithm for optimization inspired by imperialistic competition,”
in IEEE Congress on Evolutionary Computation, pp. 4661–4667, Sept
2007.
[7] R. Rao, V. Savsani, and D. Vakharia, “Teaching learning-based optimization: A novel method for constrained mechanical design optimization
problems,” Computer-Aided Design, vol. 43, no. 3, pp. 303 – 315, 2011.
[8] J. Snow, “Tornado,” in Encyclopaedia Britannica (T. E. of Encyclopaedia Britannica, ed.), 2015 ed., 2015.
[9] E. P. Krider, “Thunderstorm,” in Encyclopaedia Britannica (T. E.
of Encyclopaedia Britannica, ed.), winter 2016 ed., 2016.
[10] X.-S. Yang, S. Deb, T. Hanne, and X. He, “Attraction and diffusion
in nature-inspired optimization algorithms,” Neural Computing and
Applications, pp. 1–8, 2015.
[11] X. Tang, Z. Ghassemlooy, S. Rajbhandari, W. Popoola, and C. Lee,
“Performance of the coherent optical binary polarization-shift-keying
heterodyne system in free space optical communications using a lognormal atmospheric turbulence model,” in Mosharaka international
conference on Communications, Propagation and Electronics, pp. 1–1,
Mar 2010.
[12] M. Jamil and X.-S. Yang, “A literature survey of benchmark functions
for global optimisation problems,” Int. J. of Mathematical Modelling
and Numerical Optimisation, vol. 4, no. 2, pp. 150–194, 2013.
[13] M. Clerc and J. Kennedy, “The particle swarm - explosion, stability, and
convergence in a multidimensional complex space,” IEEE Transactions
on Evolutionary Computation, vol. 6, pp. 58–73, Feb 2002.
| 9cs.NE
|
Development of a new model to predict indoor daylighting :
integration in CODYRUN software and validation
A. H. Fakra, F. Miranville, H. Boyer, S. Guichard
Physics and Mathematical Engineering Laboratory for Energy and Environment (P.I.M.E.N.T)
University of La Reunion
117 rue du Général Ailleret, 97430 Le Tampon (Reunion Island – French overseas Dpt)
[email protected]
Abstract
Many models exist in the scientific literature for determining indoor daylighting values. They are
classified in three categories: numerical, simplified and empirical models. Nevertheless, each of these categories
of models are not convenient for every application. Indeed, the numerical model requires high calculation time;
conditions of use of the simplified models are limited, and experimental models need not only important
financial resources but also a perfect control of experimental devices (e.g. scale model), as well as climatic
characteristics of the location (e.g. in situ experiment).
In this article, a new model based on a combination of multiple simplified models is established. The
objective is to improve this category of model. The originality of our paper relies on the coupling of several
simplified models of indoor daylighting calculations. The accuracy of the simulation code, introduced into
CODYRUN software to simulate correctly indoor illuminance, is then verified.
Besides, the software consists of a numerical building simulation code, developed in the Physics and
Mathematical Engineering Laboratory for Energy and Environment (P.I.M.E.N.T) at the University of Reunion.
Initially dedicated to the thermal, airflow and hydrous phenomena in the buildings, the software has
been completed for the calculation of indoor daylighting. New models and algorithms - which rely on a semidetailed approach - will be presented in this paper.
In order to validate the accuracy of the integrated models, many test cases have been considered as
analytical, inter-software comparisons and experimental comparisons. In order to prove the accuracy of the new
model – which can properly simulate the illuminance – a confrontation between the results obtained from the
software (developed in this research paper) and the major made at a given place is described in details. A new
statistical indicator to appreciate the margins of errors - named RSD (Reliability of Software Degrees) – is also
be defined.
1
The objective is not only to develop an efficient research tool to improve visual comfort and reduce
energy consumption, but also to transfer the knowledge through these decision-making aids tools to praticians
and decision makers.
Keywords: Indoor daylighting, building simulation software, experimental validation, CODYRUN.
2
Contents:
1.
Introduction ............................................................................................................................ 5
2.
Literature review ................................................................................................................... 5
2.1.
2.2.
3.
History ............................................................................................................................................ 5
State of the art – Daylighting, choice and selected models ........................................................ 6
Modelling of indoor daylighting .......................................................................................... 7
3.1. Brief description of CODYRUN software ................................................................................... 7
3.2. Implementation of a new model ................................................................................................... 8
3.2.1. Introduction ........................................................................................................................................................ 8
3.2.2. Improvement of the software ...................................................................................................................... 9
3.2.3. Description of the model ............................................................................................................................ 11
3.2.3.1.
3.2.3.2.
3.2.3.3.
3.2.3.4.
4.
Calculation of diffuse illuminance in the absence of sun patch: DF method .............................. 11
Calculation of diffuse illuminance part in the presence of the sun patch : new method .......... 12
Calculation of the sunlight ...................................................................................................................... 12
Recapitulation: new method of indoor illuminance calculation ..................................................... 12
Validation ............................................................................................................................. 13
4.1. Preliminary................................................................................................................................... 13
4.1.1. Limitations of models used in CODYRUN for indoor daylighting simulation .................. 13
4.1.2. New approach to determine the percentage of reliability ............................................................. 13
4.1.2.1.
4.1.2.2.
4.1.2.3.
Outdoor errors ............................................................................................................................................ 13
Indoor errors ............................................................................................................................................... 14
Reliability of the Software Degrees (RSD): new concept of quantification of the accuracy .. 15
4.1.3. Tests cases selection of the C.I.E............................................................................................................ 16
4.2. Summary and results of test cases submitted to CODYRUN ................................................... 16
4.3. Example of dynamic validation: experimentation on LGI cell ................................................ 17
4.3.1. Experimental test cell LGI......................................................................................................................... 18
4.3.2. Instrumentation and experimental procedure .................................................................................... 19
4.3.3. Measurement errors ...................................................................................................................................... 19
4.3.4. Simulation hypothesis in CODYRUN.................................................................................................. 20
4.3.5. Comparison of results .................................................................................................................................. 21
4.3.6. Conclusion of the transient study ........................................................................................................... 24
5.
Conclusions .......................................................................................................................... 25
6.
Future research ................................................................................................................... 26
3
Nomenclature
DF
Daylight Factor (%)
ERC
Externally Reflected Component (unitless)
Ein,dif, p
Indoor global Diffuse illuminance at point p (lux)
Ein,Dir, p
Indoor global Direct illuminance at point p (lux)
Emean
Illuminance Mean value of the dependent variable testing data set and N (number
of records of data in the testing set) or the Emeas,i mean values (lux)
Emeas,i
Measured value for horizontal illuminance (lux)
Emod,i
Illuminance predicted values (lux)
Eout, Dir, p
Direct illuminance from the sun at point p (lux)
Eout, Dir, S
Direct illuminance from the sun at point p (lux)
Eout, G, p
Outdoor Diffuse illuminance at point p (lux)
FC
Correction caused by the remoteness of a point illuminated by natural light from
an opening (unitless)
FR
Correction factor for window framing (unitless)
GL
Daylight transmission coefficient of the glass (unitless)
IRC
Internally Reflected Component (unitless)
MF
Correction factor for window dirt (unitless)
MG
Activity coefficient of the study site (unitless)
SC
Sky Component (unitless)
STS
Surface of the sun patch (m2)
ST
Floor surface (m2)
Vsim,i
Interior simulated illuminance value at a point of position index i (lux)
Vref,i
Interior reference illuminance value at a point of position index i (lux)
R2
Coefficient of correlation
Greek Symbol
sol
Indoor average slab reflectance (unitless)
vitre
Coefficient of light transmission of glazing (unitless)
Index
i
Positioning index of a point on a horizontal plan surface indoor the building. i
varies from 0 to N.
4
1. Introduction
CODYRUN software is a dynamic building simulation code developed at the Physics and Mathematical
Engineering Laboratory for Energy and Environment (P.I.M.E.N.T). The software uses a multiple-model
approach to simulate the energy behaviour of buildings, including thermal, aeraulic and hydrous aspects. It was
in its first release created in 1993 [1] for the needs of two types of users:
Building designers and operators
Building physics researchers
Thereafter, several works have been investigated to improve the computer tool. Many physical models were
tested and introduced into the software to make it more powerful and more diversified.
Recently, a new model has been developed and integrated to allow the simulation of indoor daylighting. The
objective is to simulate illuminance at any point of a work plane by using a new concept to simplify the
combination of photometric models. This model is original because it is based on the combination of three
simplified types of models to determinate indoor illuminance.
It is indeed important to provide a tool capable of simulating daylighting, simultaneously coupled with
thermal phenomena. This allows studying cases involving interactions between several energetic phenomena and
determining the global behaviour of buildings.
2. Literature review
The literature review will be only conducted on the quantity of indoor daylighting coming from sidelights
toward a point located on a specified useful work plane.
2.1. History
The CIE (International Commission of Illumination) was created in 1928 [2] in order to answer to the
problem associated with daylighting and lighting. The objective of the organization is to develop norms,
measurement methods and models to characterize luminous environment [3]. It is recognized by International
Standardisation Organisms due to its know-how and professionalism in illuminance analysis applications.
From 1960, Kittler [4] introduced artificial skies with the aim of simulating indoor daylighting of scale
models. Thus, more complex indoor daylighting configurations could be studied under normalised experimental
conditions.
In 1973 [5], the CIE published a new methodology to assess indoor daylighting from Daylight Factor (DF).
Some fifty other methods are presented in this same report.
5
Tregenza and Waters [6] introduced the notion of Daylight Coefficient (DC) in 1983.
Collins [7] marked the history and discoveries in the evaluation of indoor daylighting. In fact, in his article,
published in 1984, he reminded us that the very first research on quantitative study of daylighting was present
since 1865. The studies were made at that time from a graphical approach to be used with simple rectangular
sidelight (without considering complex cases).
In the same year, numerical calculation codes in daylighting and lighting have progressed in computers
technology, more precisely, with the development of computers and software: Superlite [8], Radiance [9],
Genelux [10], Lightscape 3.2 [11], Adeline [12], etc.
In 1988, Clarke and Janak [13] improved RADIANCE and ESP-r to determine indoor daylight from
Daylight Factor (or DF).
In 1999, Michel [14] introduced a new factor for the calculation of indoor daylighting, called the Partial DF.
In the same year, Paule [15] created a simulation code (DIAL) for study of indoor daylighting and lighting.
The characteristic of this software is that it allowed preliminary draft diagnosis of a building using simplified
and accurate models (minimal calculation time).
In 2004, Maamari [16] defined a series of test cases to verify the reliability of numerical simulation software
applicable to indoor natural and artificial lighting.
Currently, many laboratories and organizations across the world are working to improve numerical
simulation tools and modelling dedicated to indoor illuminance. This is the case, for example, of the models
developed by Enrique Ruiz in 2002 [17], Jenkins and al. in 2005 [18], and Danny in 2007 [19], or the task 31 of
IEA in 2007 [20].
2.2. State of the art – Daylighting, choice and selected models
Characterising indoor daylighting is very difficult because various parameters must be considered (view
factor, aperture size, depth of the local, etc.).
Generally, there are three types of methods to determine quantitative indoor illuminance:
Experimental (scale models, full-sized buildings, etc.);
Numerical (Radiosity, Ray tracing, etc.);
Simplified (Lumen, DF, etc.).
6
Experimental models require important financial means (scanning sky simulator, for example, in the case of
studies of daylighting from scaled models) and the perfect observation of experimental phenomena, through
measurements, on scale models or full-sized buildings. The second drawback in this method is that the time
delay to obtain exploitable data is relatively long and depends on the climatic constraints of the site. Concerning
scale models, errors are very high when compared to results obtained by full-scale measurements. This approach
is therefore to be avoided.
In the case of numerical detailed models, calculation time is considerable. In addition, these models are
generally used for visual rendering (quality of the illuminance) rather than for quantitative characteristics of
daylight in a room.
Finally, the major inconvenience of simplified models is that they are only applicable under certain
conditions (sky conditions, limited to a certain position, etc.). However, they have enormous advantages: easy to
understand and to implement, simple to use and less time consuming, give acceptable results, etc. These models
are more adaptable to the code source of CODYRUN. Therefore, this category has been selected for integration
of a daylighting model in CODYRUN. Our contribution in the field has been to combine three types of models
normally used in different conditions (overcast sky or not, only direct sunlight, etc.). This allowed not only to
circumvent the conditions calculations, but also to improve their applications range.
3. Modelling of indoor daylighting
In this paragraph, new models and algorithms for the numerical simulation of daylighting in CODYRUN are
presented.
3.1. Brief description of CODYRUN software
CODYRUN is a multi-zone software integrating natural ventilation, thermal and moisture transfers,
developed with a friendly interface.
The software is a tool dedicated to researchers and professionals. Physical models taken into account in the
software are:
Airflow and humidity transfer
Outdoor convection
Indoor convection
Outdoor long wave exchanges
Indoor long wave exchanges
7
Indoor short wave repartition
Heat conduction
Conductive exchanges with the ground
HVAC system
Outdoor illuminance
CO2 transport
Details of the multiple-model characteristics of CODYRUN are given in [21], the thermal model
constitution in [22], the pressure airflow model in [23], the outdoor illuminance model in [40], the datastructuration and the description of the tool in [24] and CO2 model in [25].
Many validation tests of CODYRUN code were successfully applied to the software. Most of them are part
of the BESTEST procedure [26] and led to the validation of the results.
3.2. Implementation of a new model
3.2.1.
Introduction
All these models allow the software to address not only the various issues related to building design but
also to make a detailed study of the physical variables of the surrounding.
However, the initial model does not study other aspects related to the interior environment, such as
visual comfort. An improvement of the software was necessary to overcome this.
Studies have thus been conducted to integrate indoor daylighting models in the software and ensure that
the software correctly simulates illuminance received at any point of a given work plane (see Figure 1).
This study is original because many research laboratories are only interested in the qualitative study of
the relationship between thermal transfer and luminous aspects (for example, the association of a temperature for
each luminous colour) as opposed to CODYRUN, which is able to examine quantitative values.
8
Figure 1. Objective of the study
3.2.2.
Improvement of the software
Many improvements have been made to the software CODYRUN to enable it to properly simulate
indoor daylighting. Figure 2 summarizes the various computer codes that have been developed in this study. The
objective is to develop a 3-D toolkit for building environments dedicated to the daylighting simulation. The tools
to represent this environment are algorithms and data structures in space vector (vector, polygon and projection).
These algorithms are often used in programming of any sort of 3D computer graphics.
For this, the computer code (program) of CODYRUN has been amended at four different levels:
In the simulation code: it was necessary to add indoor daylighting model, to create a work plane and to
make the mesh of the plane from an algorithm based on Euclidean geometry (polygons, vertex, vector,
etc.) and finally integrate a model capable of reproducing solar illuminance from a meteorological
database containing only irradiance (global, diffuse and direct beam).
9
In the result files: CODYRUN creates an Excel file where it usually stores all the information on the
results of simulations. This file is not suitable in the photometric study. So, two improvements have
been made at this level: first, the existing result file has been modified to take into account the value of
illumination on the working plane. On the other hand, new file storage information, much more detailed
in the field of lighting, was added. This file (output variable of the software) can be opened with a
notepad application.
In the weather file: the numerical simulation of a building cannot be done without prior knowledge of
the weather conditions. Indeed, CODYRUN needs to be informed about climatic conditions (input
variables of the software). Regarding the simulation of the interior illuminance, the software needs to
have information on the illumination normally provided by external data format files such as TMY2.
CODYRUN accepts such files. Sometimes, we do not have a similar database. Knowing that the
software has its own type of weather file but does not give any information on illuminance variable, an
improvement has therefore been made to provide information on outdoor illuminance. In the lack of
weather data, it is possible to simulate indoor illuminance. Indeed, at a given time it is sufficient to
indicate an overall value of outdoor illuminance for the calculations of the indoor illuminance at any
point on a work plane.
In the building description: another input variable should be given to the software: the description of the
studied building. However, the indoor illuminance strongly depends on the geometric position in space
of openings, building orientation, etc. Previously, the software could not take into account all these
parameters. This problem was solved by adding the cartesian coordinates in the declaration of each
component (wall, floor, ceiling, windows, etc.) of the building. Each component is constituted of two
sides and each side is characterized by one vertex.
10
Figure 2. Improvements to the software CODYRUN
3.2.3.
Description of the model
Two basic models are used in the algorithmic combination of the calculation code for indoor daylighting.
The Daylight Factor (or DF) is used to determine diffused illuminance whereas the classical method (taking into
account point position and aperture visible transmittance) is used to determine the direct illuminance through a
sidelight. The principle of calculation for each one of these two methods will be given in the following
paragraphs. In addition, a new specific mathematical relation for the determination of the diffuse illuminance on
a sun patch will be described.
3.2.3.1.
Calculation of diffuse illuminance in the absence of sun patch: DF method
The DF (Daylight Factor) method will be used to calculate diffuse illuminance. This method was elaborated
by BRE (Building Research Establishment) [27] and published by CIBSE [28]. We can consider the light falling
on a point in the rooms as being composed of three distinct components. Light that comes directly from the sky
and called the Sky Component (SC), light that comes from external surfaces such as buildings called the
Externally Reflected Component (ERC), and light that is reflected from internal surfaces called the Internally
Reflected Component (IRC). The sky component normally refers to the diffuse sky: it is not used to describe
direct sunlight. The DF is simply calculated as:
Equation.1
(1)
11
If a DF value (from the relation (1)) is known, it is possible to calculate the global inside illuminance at the
point of interest from the relation (2):
Equation.2
3.2.3.2.
(2)
Calculation of diffuse illuminance part in the presence of the sun patch : new method
It is considered that the diffuse illuminance at a point (which is inside the sun patch) is the sum of the
diffuse illuminance generated by the DF and that produced by the sun patch itself in the neighbourhood of the
point. A simple relation originally determines the diffuse part coming from the sun patch:
Equation.3
(3)
The general equation of diffuse illuminance at an inside point (of the room) on the sun patch becomes:
Equation.4
3.2.3.3.
(4)
Calculation of the sunlight
For this part, the Siret’s method is used [29]. The relation is written as followed:
Equation.5
3.2.3.4.
(5)
Recapitulation: new method of indoor illuminance calculation
Relations (2), (3) and (5) allow establishing a new combination model of simplified categories to calculate
indoor illuminance for every condition:
(6)
It is important to note that this simple new relation is original since (2), (3) and (5) are usually used
separately (for each according to different conditions of application). Whereas the new relation found (6) is the
combination of the three models. Thanks to this new combination, it is possible to calculate the illuminance
values for any sky conditions. Indeed, (2) is applied only in overcast sky condition, while (3) and (5) are applied
if and only if a light flux - coming from the sun - enters the room. The scientific literature does not give
information on the combination of relations (2), (3) and (5), which are combined together (i.e. relation (6)).
12
4. Validation
The validation step is very important to know the precision of the simulation results. Especially when
dealing with daylighting simulation programs, the achievement of the accuracy of the predictions is difficult. It is
thus necessary to refer to rigorous procedures, recognized worldwide.
There are few documents on the procedures to follow. In most cases, laboratories implement their own
experimental database to serve as a reference for comparisons between model predictions and measurements.
The study of Maamari [30] perfectly responds to our needs concerning the validation step. Indeed, his thesis
is related to the establishment of methods to check the reliability of simulation codes of indoor daylighting and
artificial lighting, from analytical and experimental tests cases. This work is used as reference in the task TC3-33
of the CIE [31].
Many other test cases found in the scientific literature (papers from CSTB, BRE, Task 21 of IEA and
experimental test case of CIBSE, etc.) have been applied to the simulation software CODYRUN. An intersoftware comparison has also been done and finally, database references for the local study in dynamic
conditions (from a cell called LGI) have been established. This last case will only be presented here.
4.1. Preliminary
4.1.1.
Limitations of models used in CODYRUN for indoor daylighting simulation
CODYRUN only provides daylight values for a given horizontal surface (i.e. no daylighting values are
calculated on vertical walls).
The software only considers sidelight. In addition, it does not take into account internal obstructions
(furnishing, occupants, etc.), bidirectional transmissions (BRDF) and lightwell. Therefore, simulation results will
be confronted to reference test cases, taking into account the above-said hypothesis.
4.1.2.
New approach to determine the percentage of reliability
It is important to consider a square percentage deviation to calculate the error between measurements
and used models, when studying outdoor or/and indoor daylighting.
4.1.2.1.
Outdoor errors
We have recently taken into account used models for the study of luminous efficacy [32]. It concerns
the references [33] and [34].
Generally, three statically indicators are considered to characterise the accuracy of the models associated to
outdoor and indoor daylight: the coefficient of determination R2, the Mean Bias Deviation (MBD) and the Root
13
Mean Square Deviation (RMSD). MBD demonstrates the model’s tendency to underestimation or to
overestimation. RMSD offers a deviation measure from the predicted values in relation to the measured values.
The use of the correlation coefficients R2, MBD, and RMSD determines the accuracy of the models. The
following relations give the definitions of these statistical estimators:
Equation.6
(6)
Equation.7
(7)
In order to increase accuracy, some statistical indicators also need to be defined as follows:
(8)
Equation.8
4.1.2.2.
Indoor errors
Concerning indoor daylighting study, the estimate is made from relative errors at the point i defined by:
Equation.9
(9)
Equation.10
(10)
Global relative () errors are thus written:
14
4.1.2.3.
Reliability of the Software Degrees (RSD): new concept of quantification of the
accuracy
In this study, a new percentage indicator that verifies the capacity of the software to correctly simulate
indoor daylighting is defined as RSD (or Reliability of Software Degree). It is the ability of the software to
simulate correctly and realistically a physical variable. In our case study, the variable is indoor daylighting.
Photometric phenomena are very complex, so we can consider that results of numerical simulation
software are correct compared to reality (or reference values) when its RSD is greater or equal to 50%. This
gives a precise idea of the limits, strengths and weaknesses of the software in its application field. As a
preliminary, this method requires the knowledge of reference values (or reference test case study).
Two situations are presented for the determination of RSD:
On the one hand, the reference values are given at each measurement. For example, positioning
reference point A1 at A4 (or reference calculation) in form of margin (upper and lower limits of the
curve below) of acceptable values for the results (results given by CODYRUN on the curve below) of
the simulation. These margins are given by measured values more or less the total estimated error with
and without errors linked to the simulation (see figure 3). In this case, the RSD is the percentage of
simulated values situated between references values.
Figure 3. Illustration of RSD evaluation when reference points are margins (upper and lower limits)
On the other hand, reference values are given at each measurement point (or reference calculation)
without margins. So, it is essential to consider relative mean errors between simulation and reference
values at each point. The RSD is given by subtracting 100% from the global mean relative error
obtained (refer to figure 4).
15
Figure 4. Illustration of RSD evaluation when reference points are given without margins
4.1.3.
Tests cases selection of the C.I.E
The reference scenarios of the C.I.E are classified into two categories: analytical and experimental. In each
category, we can distinguish the study of artificial lighting and/or daylighting. These references are defined in
the Task TC3-33 of the CIE. They are used for the validation of CODYRUN. Only 39% of these tests cases were
applied to CODYRUN and verified.
4.2. Summary and results of test cases submitted to CODYRUN
Table 1 summarises all test cases applicable to CODYRUN and associated errors for each case. CODYRUN
successfully passed selected test cases (the errors were acceptable).
Globally, results for CODYRUN are comparable to those of other codes for daylighting. This shows
that the model included in CODYRUN is well implemented and gives acceptable results.
Results of analytical test cases confirm the aptitude and the limitations of the software to simulate
daylighting. Unlike the approximate values of results concerning the calculation analytical test case, the
simulation of CODYRUN code gives exact values. Consequently, this difference of calculation increases the
errors in the comparison between the simulation in CODYRUN and the analytical test case.
Concerning the references of experimental test cases, it is important to know that they are very
complicated to implement, requiring a lot of logistics, technical and financial means. For example, accuracy for
spatial location of the sensors is very important; it is also important to ensure that the measureable limits of the
sensors corresponds to the range of measured values.
16
As a conclusion, this work allowed us to assess the strengths and weaknesses of the software. Thus, the
improvement of the software could be done on the illuminance calculation at the furthest and nearest point to the
apertures.
Test Type
Reference
Experimental
Relative
RSD
error (%)
(%)
Observation
[30]
II.10
3.87
91.44
Side opening
[30]
II.11
4.54
93.65
Different types of skies CIE
[30]
II.14.4
1.96
98.04
Diffuse (form factor)
[30]
II.14.9
15.61
84.34
Solar irradiance (outdoor)
[37]
CSTB
5.92
95
Daylight Factor (DF)
[30]
III.2.5
3.38
93.49
Artificial sky
[36]
I.E.A
9.31
74.46
Scale model E.P.F.L.
[38]
I.E.A
11.31
70.92
Real Building H.U.T.
[35]
LGI test
32
88.95
Transient (or dynamical
Analytical
Comparative
Procedure
aspect)
Minimum =
1.96
84.34
Maximum =
32
98.04
9.76
87.81
Average =
Table 1. Reference test cases applied to CODYRUN
4.3. Example of dynamic validation: experimentation on LGI cell
Table 1 shows the details of test cases submitted to the software CODYRUN to check its level of reliability
to simulate indoor daylighting.
The highest error between the reference values and those from the simulations were observed during the
dynamic study performed locally using the LGI cell (about 32%) of relative error [35]. The following paragraph
will present a detailed study to explain the results.
17
4.3.1.
Experimental test cell LGI
An existing experimental full-scale test cell has been instrumented in order to verify the reliability of
the software CODYRUN to correctly simulate daylighting in transient conditions during an entire day. The
database collected is used as a reference for the comparison between simulation and experimental values.
The experimental test cell LGI is located at Saint-Pierre (see figure 5). This test building was built for
experimental validation reference on physical models introduced in CODYRUN [10] and ISOLAB [39].
Figure 5. Experimental site for L.P.B.S. Laboratory
Two types of luxmeters were used for the global illuminance: one to characterize the outdoor
daylighting and another to characterize the penetration of indoor daylight.
The experimental platform includes a meteorological station and a technical room. The internal
and external walls were initially painted white. Absorption coefficient of the interior surfaces is about 60
%. The roof is constituted of corrugated iron and insulated with a radiant barrier. The vertical walls are
made of insulating board. The floor is composed of concrete slabs, of polystyrene on the top (4 and 5 cm
thick). The dimensions of the LGI cell are illustrated in the figure 8.
Figure 6. LGI cell test of L.P.B.S.
The LGI cell is representative of a typical room or office in Reunion Island. The door is made up of
aluminium, with a 6 mm glass window.
18
4.3.2.
Instrumentation and experimental procedure
The meteorological data are measured every minute. The station measures irradiation (global and diffuse)
and wind speed and direction (respectively at a height of 2 m and 10 m from the ground). We added an outdoor
daylight sensor that measures more than 14,000 values each day. The data logger is of CAMPBELL type. The
measurements of indoor and outdoor daylighting are made from instruments of AHLBORN type. All the
measuring instruments are synchronised. To simplify our study, we have positioned an exterior luxmeter (sensor)
on the roof of the LGI cell (see figure 8). Indoors luxmeters (sensors) were aligned on the slabs as shown in the
figure (see figure 9). These sensors are equidistant by 0.5 m. The horizontal distance between position of the
sensor A1 and aperture is 0.23 m.
Figure 7. Schematic position of the five interior luxmeters
Globally, three categories of days (refer to [40] for more information) are used as a reference experiment: it
consists of intermediate day, clear day and overcast day conditions. The reference Table 2 gives the subjective
and objective measurements errors. Total error associated with the experimental measurements (outdoor and
indoor measurements) is 13.22% and the global error is 15%.
4.3.3.
Measurement errors
Table 2 indicates the values of subjective and objective errors of our experiment. This approach to study the
measurement error is highly recommended by the C.I.E [31].
These values will allow us to give margins of lower and upper limits for each measurement. The mean error
(measurements errors) is 13.4% and global error (measurements and software hypothesis of simulation errors) is
15%. Table 3 gives the characteristics of the interior measurements sensors.
19
Error
Identifications
Error value (en %)
Sensor accuracy
1 (indoor) + 0.03 (datalogger) +
9.9 (outdoor)
Cosine correction
Objective
2.9 (indoor sensor) + 2,9 (outdoor
sensors)
Spectral sensitivity
1 (indoor sensor) + 7 (outdoor
sensor)
Subjective
Accuracy on positions and orientation of the
3
measuring points
Table 2. Subjective and objective errors of experimentation of the LGI cell test
Measuring range
0.05 – 12 500 lux
Spectral sensitivity
Class B, superior at 6%
Cosine error
<3%
Rated temperature
24°C +/- 2 K
Range of humidity
10 to 90% (without condensation)
Operating/storage temperature
0°C to +60°C / -10°C to +80°C
Table 3. Characteristics of the interior measurements sensors (FLA603VL2), of AHLBORN type
4.3.4.
Simulation hypothesis in CODYRUN
The simulation conditions of the LGI cell in CODYRUN are:
Albedo = 70%
Interior reflectance of the walls and ceiling (white colour) = 60%
Interior reflectance of the floor (grey-coloured slabs) = 20%
The height from the slabs to the workplane is at 0.01 m.
Dimension of the elementary surface grid = 0.1m x 0.1m
Correction factor of DF: MG = FR = 0.8 and MF = 0.9 (see (1) for the definition of these coefficients)
20
4.3.5.
Comparison of results
This paragraph presents the simulation results and the comparison with those of measured indoor
daylighting obtained for the three days conditions.
Clear day condition
Figure 8 indicates the comparison between measured values and simulated values at point A3. The trend at
points A1, A2, A4, and A5 are practically the same, but with different amplitudes.
Figure 8. Clear day: comparison between measured and simulated values at the point A3 (without direct
sunlight and for one-minute time step)
We obtained a RSD value of more than 50%. The simulation results obtained - without considering the
direct sunlight flux - better correspond to the values of measurements (see Figure 8). There is obviously no
inertia in photometric phenomena (i.e. abrupt variations of daylighting with time). So, during our measurements
we took a time step of one minute to take into consideration these abrupt variations. The inconvenience is that
this increases the errors between the simulation results and reference measurements. To verify this hypothesis,
we plotted the same curve with hourly averaged values and we got the following figure (refer to Figure 9, for the
same point A3). Reference values obtained hourly present more realistic results than minutely simulated results.
21
Figure 9. Clear day: comparison between measured and simulated values at the point A3 (for one-hour time
step)
Intermediate day
The comparison results of position of the reference point at A1 are given next. Figure 10 shows an example
of comparison between simulations and measured values. The results obtained for other positions are quite
identical and the observation is the same as for clear day. CODYRUN has a RSD equal to 71.3 %.
Figure 10. Intermediate day: comparison between measured and simulated values at the point A1 (for oneminute time step)
Figure 11 compares simulation results with results for an hourly time step at point A1. That day was
divided into two equal parts: a relatively clear morning and a cloudy evening. During some precise moments of
the morning, very high illuminance measurements have been observed, due to direct sun penetration. Indeed, the
interior sensors were not able to measure the direct illuminance flux of values which were superior to 12 500 lux
because of their measurements range, which does not go beyond this value (See Table 3). This explains why the
22
software seems to simulate incorrectly the global available illuminance in the morning. In fact, simulation results
in the evening better correspond to actual values.
Figure 11. Intermediate day: comparison between measured and simulated values at the point A1 (one-hour
time step)
Overcast day
In this study, we present simulation results obtained from reference positions points A5. Figure 12 illustrates
the results of simulation.
R2
R1
Figure 12. Indoor daylighting : Overcast day condition (LGI cell)
In most cases, simulated values in CODYRUN between 10:30 am and 12:45 pm are higher compared to
measured values (for example see the highlighted zone R2 in the Figure 12). The explanation of this result is that
the studied cell is equipped with an overhang on the northern façade. This prevents a major part of the sunlight
23
from penetrating into the room. At this time, this particular device is not taken into account by the software. We
noted that (for the five reference positions) a significant difference between simulation and measurement values
exists in the morning (highlighted region R1) compared to the evening (highlighted region R3):
In fact, from sunrise till 12:00 am (corresponding to highlighted R1 zone in the Figure 12)
CODYRUN better simulates the indoor daylighting. This is certainly due to the recognition
of obstacles by the software (buildings in the neighbourhood of the north-east façade
corresponding to the sun rising façade). This has been taken into account for indoor
daylighting by the software;
Concerning the evening, we noticed that (for example, the highlighted R3 zone in A5 - Figure
12) the simulated values are lower than the measured values. The fact is that South and West
parts of vertical façades of the test building are opaque. Then, the North-facing overhang
does not influence skylight at these hours of the day. We obtain more coherent values
between the simulation and the measurements.
As previously, we were interested in comparing these reference values with those simulated values for each
hour. The results (see the result comparison of figure 13 at reference point A1) show a good accuracy of the
software in the calculation of indoor daylighting, with the exception of close and far positions from the opening.
To conclude, in the case of an overcast day we can note that reliability accuracy is higher or equal to 68%.
Globally, CODYRUN has the tendency to underestimate indoor daylighting.
4.3.6.
Conclusion of the transient study
The case of a clear day study demonstrated the difficulty of simulating transient indoor daylighting. Indeed,
the photometric effect is very complex (it is difficult to try to draw the shape of the illuminance for a given day
and at a given point). For this reason, many other research laboratories prefer to conduct a time step study, but
not during an entire day (see [25] and [26]). However, we found that the simplified models introduced in the
software were able to simulate fairly well the dynamic photometric variable. Similar situations were presented in
an intermediate day study.
For an overcast day, we noted that the impact of the overhang in the simulation was important. Indeed, the
model introduced in CODYRUN does not consider, for the moment, the influence of overhangs. The fact that the
overhangs are not considered in the simulation increases the errors of comparisons in the experimental
validation. We also noted that the software made higher errors at reference points located close and far from the
opening.
24
During an entire day, the RMSD of the overall simulation is about 0.61. Considering the relative errors of
theses hourly simulations, we obtain an average error of 32.7%.
Figure 13. Overcast day: hourly comparison between measured and simulated values at the points A1
5. Conclusions
The sun is a renewable energy source often used in interior lighting (inside buildings).
This allows a decrease in the electrical energy consumption for artificial lighting in buildings. But, we do
not have enough information for the study concerning the needs of illuminance in tropical and humid climate
regions.
In the building, the integration of daylighting coming directly or indirectly from the sun is a variable, which is
more and more used in development tools to determine the illuminance availability. However, these tools must
satisfy the needs of architects and building design offices. According to them, the tool should not only quantify
daylight, but should also elaborate method tools, which are easily usable. CODYRUN tries to satisfy this
requirement.
As for any other software, the use of CODYRUN to simulate indoor daylighting presents advantages
and drawbacks.
Concerning the advantages, we can indicate:
Reduced calculation time. For example, the software takes 2 minutes to simulate daylighting on a
workplane of 39x35 position points inside a building and for a time step, it takes one minute all long
year);
25
The software is simple, user-friendly, and is perfectly adapted to design office (engineering consulting
firm) and to individuals;
This software is a tool used in research laboratories. We were able to show that CODYRUN was
reliable at 86.25%;
It is possible to reconstitute the outdoor illuminance from a meteorological file, which only has
irradiance information;
It is possible to simulate the effects of sunlight coming from a secondary light source;
We can monitor (or follow) the sun patch in the workplane;
CODYRUN can separately simulate the direct, diffuse and global daylighting for every sky condition,
through a new simplified model that is defined in this paper;
The software is capable of dynamical studies.
The limitations of the model introduced in the software are:
It does not consider overhangs and shadow masks in daylight simulation;
No daylight calculation on vertical or inclined walls are considered;
The cartesian coordinates of the building are manually inserted. This makes the description of the
building (in CODYRUN) a time-consuming task;
The are no graphical user interface, and no 3D visual representation of the building or visual rendering;
The used models are simplified categories, so they are not applicable to complex architectural
configurations;
It does not consider obstructions inside the room (furniture, occupants, etc.) when simulating
daylighting;
The opening should only be sidelights.
The conditions, which can be used in the software, are those of simple (the volumes are rectangular,
triangular, cubic, or L-shaped) closed systems (or sub-systems). We cannot simulate complex systems (for
example, inclined walls of buildings).
6. Future research
The future research in this field can be classified in three categories as follows:
The integration of an indoor artificial lighting model and the improvement of the new indoor
daylighting model in CODYRUN tool;
26
Complementary validation results, e.g. performed once again the experimental study for intermediate
and clear condition days in LGI test cell by using sensors capable of taking into account indoor direct
sunlight;
Using the tool to optimise the electrical energy consumption and to consider models taking into account
simultaneously the temperature and the daylight.
27
References
[1] H. Boyer, Ph D Thesis, Airflow and thermal design of multizone buildings : proposal of a tool are the
multiple choice models, University of Reunion, Saint-Denis, Reunion Island – French country, 1993. (H. Boyer,
Conception thermo-aéraulique de bâtiments multizones. Proposition d'un outil à choix multiple des modèles,
Université de La Réunion, France, thèse de Doctorat, 1993).
[2] CIE History 1913-1988, CIE 082-1990, ISBN 973 3 900734 19 0, 77 p. CIE publications.
[3] CIE, Standardization of luminious distribution on clear sky, Commission Internationale de l'Eclairage, Paris,
Technical Report, Publication CIE N°16, 1970.
[4] R. Kittler, An historical review of method and instrumentation for experimental daylight research by means
of models and artificial skies, Proceedings 14 Th CIE (Commission International de l'Eclairage) Session, vol. 8,
no. 59, pp. 319-334, 1960.
[5] CIE, Standardization of luminious distribution on clear sky, Commission Internationale de l'Eclairage, Paris,
Technical Report, Publication CIE N°22, 1973.
[6] P. R. Tregenza and I. M. Waters, Daylight coefficients, Lighting Research and Technology 1983;15(2):6571.
[7] J.B. Collin, The developpement of daylighting - a british view, Lighting Research and Technology 1984;16
(4):155-170.
[8] A. Grynberg. Comparison and validation of Radiance and Superlite. Rapport interne. Lawrence Berkeley
National Laboratory, 1988.
[9] G. J. Ward, Application of RADIANCE to architecture and lighting design, in Proceeding of the 1994 IES
National Conference, August 1994.
[10] R. Mitanchey, Ph D. Thesis, Synthesis imaging applied to indoor lighting, ENSM high school of SaintEtienne, Saint-Etienne, French, 1996. (R. Mitanchey. Synthèse d’images appliquée à l’éclairagisme intérieur des
bâtiments. Thèse en Informatique, spécialité images. Saint-Etienne: Ecole National Supérieure des Mines de
Saint-Etienne, 1996, 200p).
[11] A. B. Khodulev and E. A. Kopylov, "Physically accurate lighting simulation-computer graphics software",
in Sixth International Conference and Exhibition on Computer Graphics and Visualization in Russia, Saint
Petersburg, Russia, 1996.
[12] H. Erhorn, J. Stoffel, and M. Szermann, "Adeline 2.0 - Using computer tools to Evaluate Daylightong and
Electric lighting Applications in Building", Right-Light Four Conference, 1997.
28
[13] A. Clarke and M. Janak, Simulating the thermal effects of daylight-controlled lighting. Building
Performance (BEPAC) 1 (1998) Spring.
[14] L. Michel, Ph D. Thesis, Experimental method for evaluating illuminance performance of buildings, EPFL
High School, Lausanne, Switzerland, 1999. (L. Michel, Méthode experimentale d'évaluation des performances
lumineuses des bâtiments, Ecole Polytechnique Fédérale de la Lausanne, Suisse, Thèse de Doctorat, 1999).
[15] P. Bernard, Ph D Thesis, Application of Logic Fuzzy in indoor daylighting decision, EPFL high School,
Lausanne, Switzerland, 1999. (P. Bernard, Application de la logique floue à l'aide à la décision en éclairage
naturel, Ecole Polytechnique Fédérale de Lausanne, Suisse, Thèse de Doctorat, 1999).
[16] F. Maamari, Ph D. Thesis, The numerical simulation of lighting : advantages and potential, INSA of Lyon
High School, Lyon, French, 2004. (F. Maamari, La simulation numérique de l'éclairage : atouts et potentiels,
Institut National des Sciences Appliquées de Lyon, Lyon, France, Thèse de Doctorat, 2004).
[17] IEA, "Daylight for the 21St century : new opportunities with daylighting techniques", Solar Heating and
Cooling programmes for International Energy Agency, Washington, 2006 annual Report-Task 31, 2007.
[18] D. Jenkins, X. Zhang, T. Muneer, Formulation of semi-empirical models for predicting the illuminance of
light pipes, Energy Conversion and Management 2005;46(13-14):2288-2300.
[19] Danny H.W. Li. Daylight and energy implications for CIE standard skies, Energy Conversion and
Management 2007;48(3):745-755.
[20] Enrique Ruiz, Alfonso Soler, Luis Robledo. Statistical assessment of a model for global illuminance on
inclined surfaces from horizontal global illuminance. Energy Conversion and Management 2002;43(5):693-708.
[21] H. Boyer, F. Garde, J. C. Gatina, J. Brau, A multimodel approach to building thermal simulation for design
and research purposes, Energy and Buildings 1998;28(1):71-78.
[22] H. Boyer, J. P. Chabriat, B. Grondin-Perez, C. Tourrand, J. Brau, Thermal building simulation and
computer generation of nodal models, Building and Environment 1996;31(3):207-214.
[23] H. Boyer, A. P. Lauret, L. Adelard, T. A. Mara, Building ventilation : a pressure airflow model computer
generation and elements of validation, Energy and Buildings 1999;29:283-292.
[24] A. H. Fakra, F. Miranville, H. Boyer, D. Bigot, A simple evaluation of global and diffuse luminous efficacy
for all sky conditions in tropical and humid climate, Renewable Energy, 2011; 36: 298-306.
[25] H. Boyer, J. C. Gatina, J. Brau. Modelisation methods and data structuration induced for building thermal
simulation codes. In proceedings of the Conference of International Simulation Societies, Zurich, Switzerland,
pp. 729-733.
29
[26] D. Calogine, H. Boyer, S. Ndoumbe, C. Rivière, and F. Miranville, Identification of Parameters in Building
Concentration Dispersion Model , Indoor and Built Environment 2010;19:250-266.
[27] T. Soubdhan, T. A. Mara, H. Boyer, A. Younès, Use of BESTEST procedure to improve a building thermal
simulation program, World Renewable Energy Congress, Brighton, U. H., Part III, 1-7 July 2000, pp; 18001803.
[28] BRE, Estimating daylight in building : Part 1 and 2 / Diggest 309 and 310, Building Research Etablishment
Digest,England, 1986.
[29] CIBSE, Window design, application manual, Chartered Institution of Building Services Engineers, London,
1987.
[30] D. Siret, S. Houpert, A geometrical framework for solving sunlighting problems within CAD systems,
Energy and Buildings 2004;34 (4):343-351.
[31] F. Maamari, M. Fontoynont, and N. Adra, Application of the CIE test cases to assess the accuracy of
lighting computer programs, Energy and Buildings 2006;38:689-877.
[32] CIE, Test Cases to assess the acuracy of lighting computer programs, Commission Internationale de
l'Eclairage, Paris, France, CIE Technical Report (TC-3-33) 390 1 906 479, February 24, 2005.
[33] D. Dumortier, Ph D. Thesis, Measurement, analysis and modeling outdoor daylighting and application to
the assessment of illuminance performance into the building, University of Savoie, Savoie, French, 1995. (D.
Dumortier, Mesure, analyse et modélisation du gisement lumineux. Application à l'évaluation des performances
de l'éclairage naturel des bâtiments, Université de Savoie, Savoie, Thèse de Doctorat, 1995).
[34] L. Gabriel and A. G. Christian, Clear-sky solar luminous efficacy determination using artificial neural
networks, Solar Energy 2007;81:929-939.
[35] S. Pattanasethanon, C. Lertsatitthanakorn, S. Atthajariyakul, and S. Soponronnarit, "All sky modeling
daylight availability and illuminance / irradiance on horizontal plane for mahasarakhan", Energy Conversion and
management 2007;48(5):1601-1614.
[36] A. H. Fakra, Ph D. Thesis, Integration of daylighting and lighting models to a computer code in building
physics : modeling, validation and applications, University of Reunion, Tampon, Reunion Island – French
country, 2009. (A. H. Fakra, Intégration de modèles d’éclairage à un code de calcul en physique du bâtiment :
modélisation, validation et applications, Université de La Réunion, Thèse de Doctorat, 2009).
[37] L. Michel, IEA 21 Scale Model Daylighting Systems Evaluation, Internal Report of Solar Energy and
Physics Building Laboratory, Lausanne, Switzerland, September, 1998.
30
[38] M. Wetter, E. Polak, et V. P. Carey, BuildOpt 1.0.1 Validation, Lawrence Berkeley National Laboratory,
Californie, Technical Report LBNL-54658, March 25, 2004.
[39] IEA, Daylight Responsive lighting Control-Cases, International Energy Agency, IEA Task 21, Subtask B
(Daylight in Buildings), December, 2001.
[40] F. Miranville, H. Boyer, P. Lauret, F. Lucas, A combined approach for determining performance of radiant
barriers field conditions, Solar Energy 2006; 82: 399-410.
Vitae
Dr Ali Hamada FAKRA
Laboratoire Physique et Ingénierie Mathématique pour l’Energie et l’environneNT (P.I.M.E.N.T)
Faculté de Sciences de l’Homme et de l’Environnement
Université de La Réunion - Campus Sud
17, Rue du Général Ailleret - 97430 Tampon - Ile de la Réunion – France
Email: [email protected]
Tel : +262 692 56 23 95
Fax : +262 262 57 94 46
31
| 5cs.CE
|
arXiv:1604.00561v1 [math.ST] 2 Apr 2016
On the Conditional Distribution of the
Multivariate t Distribution
Peng Ding∗
Abstract
As alternatives to the normal distributions, t distributions are widely applied in robust analysis
for data with outliers or heavy tails. The properties of the multivariate t distribution are well
documented in Kotz and Nadarajah’s book, which, however, states a wrong conclusion about
the conditional distribution of the multivariate t distribution. Previous literature has recognized
that the conditional distribution of the multivariate t distribution also follows the multivariate t
distribution. We provide an intuitive proof without directly manipulating the complicated density
function of the multivariate t distribution.
Key Words: Bayes’ Theorem; Data augmentation; Mahalanobis distance; Normal mixture; Representation.
1 Introduction
The conventional version of the multivariate t (MVT) distribution X ∼ t p (µ, Σ, ν ), with location µ,
scale matrix Σ, and degrees of freedom ν , has the probability density function
f (x) =
n
o−(ν +p)/2
Γ{(ν + p)/2}
−1
⊤ −1
ν
1
+
.
(x
−
µ)
Σ
(x
−
µ)
Γ(ν /2)(νπ ) p/2 |Σ|1/2
(1)
We can see from the above that the tail probability of the MVT distribution decays at a polynomial
rate, resulting in heavier tails than the multivariate normal distribution. Because of this property, the
MVT distribution is widely applied in robust data analysis including linear and nonlinear regressions
(Lange et al. 1988; Liu 1994; Liu 2004), linear mixed effects models (Pinheiro et al. 2001), and
sample selection models (Marchenko and Genton 2012; Ding 2014).
∗ Peng Ding is Assistant Professor (Email: [email protected]), Department of Statistics, University of California, Berkeley, 425 Evans Hall, Berkeley, CA 94720, USA. The author thanks the Editor, Associate Editor and three
reviewers for their helpful comments.
1
In the following discussion, we use W ∼ χb2 /c to denote the scaled χ 2 distribution, with density
proportional to wb/2−1 e−cw/2 . We exploit the following representation of the MVT distribution:
√
X = µ + Σ1/2 Z/ q,
(2)
where Z follows a p dimensional standard normal distribution, q ∼ χν2 /ν , and Z is independent
of q (Kotz and Nadarajah 2004; Nadarajah and Kotz 2005). It differs from the multivariate normal
distribution N p (µ, Σ) only by the random scaling factor
√
q.
The above representation (2) implies X | q ∼ N p (µ, Σ/q), i.e., X follows a multivariate normal
distribution given the latent variable q. If we partition X into two parts, X1 and X2 , with dimensions
p1 and p2 , we obtain the following normal mixture representation conditional on q:
X1
µ1
Σ11 Σ12
X=
| q ∼ N p1 +p2
,
/q ,
X2
µ2
Σ21 Σ22
(3)
where the location and scale parameters are partitioned corresponding to the partition of X. Marginally,
we have X1 | q ∼ N p1 (µ1 , Σ11 /q), and therefore
X1 ∼ t p1 (µ1 , Σ11 , ν ),
which follows a p1 dimensional MVT distribution with degrees of freedom ν .
Although we can obtain the marginal distribution in an obvious way, the conditional distribution
of X2 given X1 is less transparent. In fact, the conditional distribution of the MVT distribution is also
a MVT distribution, with degrees of freedom different from the original distribution.
2 Conditional Distribution
Kotz and Nadarajah (2004, page 17) and Nadarajah and Kotz (2005) claimed that the conditional
distribution of the MVT distribution is not a MVT distribution except for some extreme cases. For
−1
the case with µ = 0, define x2|1 = x1 − Σ21 Σ−1
11 x1 , and define Σ22|1 = Σ22 − Σ21 Σ11 Σ12 as the
Schur complement of the block Σ11 in matrix Σ. They derived the conditional density of X2 given
2
X1 by calculating f (x)/ f1 (x1 ) using the probability density function in (1). Ignoring the normalizing
constant in formula (15) of Nadarajah and Kotz (2005), we present only the key term:
f2|1 (x2 | x1 ) ∝
1 + (ν + p1 )−1 x⊤
2|1
ν
−1
+ x⊤
1 Σ11 x1
ν + p1
Σ22|1
!−1
x2|1
−(ν +p1+p2 )/2
.
(4)
Kotz and Nadarajah (2004, page 17) and Nadarajah and Kotz (2005) obtained the correct conditional
density function, but due to the complex form of their formula (15), they did not recognize that the
key term in (4) is essentially the unnormalized probability density function of a p2 dimensional MVT
⊤ −1
distribution with location Σ21 Σ−1
11 x1 , scale matrix (ν + x1 Σ11 x1 )/(ν + p1 ) × Σ22|1 , and degrees of
freedom (ν + p1 ).
In the literature, some authors stated the right conclusion about the conditional distribution without
providing a proof (e.g., Liu 1994, page 5, DeGroot 2005, page 61), and some other authors re-derived
the correct conditional density function in (4) and pointed out that it is another MVT distribution (e.g.,
Roth 2013b). Instead of directly calculating the conditional density or using the general theory of
elliptically contoured distributions (Cambanis et al. 1981; Fang et al. 1990, Theorems 2.18 and 3.8;
Kibria and Joarder 2006), we provide an elementary and thus more transparent proof of the conditional
distribution based on the normal mixture representations in (3).
3 Proof via Representation
Our proof proceeds in two steps: we first condition on q, and then average over q.
First, we condition on both X1 and q. According to the property of the multivariate normal
distribution, X2 follows a multivariate normal distribution conditional on (X1 , q), i.e.,
−1
X2 | (X1 , q) ∼ N p2 µ2 + Σ21 Σ−1
11 (X1 − µ1 ), (Σ22 − Σ21 Σ11 Σ12 )/q = N p2 µ2|1 , Σ22|1 /q ,
where µ2|1 = µ2 + Σ21 Σ−1
11 (X1 − µ1 ) is the linear regression of X2 on X1 .
Second, we obtain the conditional distribution of q given X1 . According to Bayes’ Theorem, the
3
conditional probability density function of q given X1 is
fq|1 (q | X1 ) ∝
f1|q (X1 | q)π (q)
o
n q
(X
−
µ
)
qν /2−1 e−ν q/2
∝ |Σ11 /q|−1/2 exp − (X1 − µ1 )⊤ Σ−1
1
1
11
2
n q
o
∝ q(ν +p1 )/2−1 exp − (ν + d1 ) ,
2
where d1 = (X1 − µ1 )⊤ Σ−1
11 (X1 − µ1 ) is the squared Mahalanobis distance of X1 from µ1 with
scale matrix Σ11 . Therefore, the conditional distribution of q given X1 is q | X1 ∼ χν2+p1 /(ν + d1 ),
according to the density of the scaled χ 2 distribution.
Finally, we can represent the conditional distribution of X2 given X1 as
√
1/2
µ2|1 + Σ22|1 Z2 / q | X1
s
. χ2
ν +p1
1/2
∼ µ2|1 + Σ22|1 Z2
ν + d1
s
s
!
2
.
χ
ν + d1 1/2
ν +p1
Z2
Σ
,
∼ µ2|1 +
ν + p1 22|1
ν + p1
X2 | X1 ∼
(5)
where Z2 is a p2 dimensional standard multivariate normal vector, independent of χν2+p1 /(ν + p1 ).
From the normal mixture representation of the MVT distribution in (2), we can obtain the following
conditional distribution.
Conclusion One: The conditional distribution of X2 given X1 is
ν + d1
Σ22|1 , ν + p1 .
X2 | X1 ∼ t p2 µ2|1 ,
ν + p1
The conditional distribution of the multivariate t distribution is very similar to that of the multivariate normal distribution. The conditional location parameter is the linear regression of X2 on X1 .
The conditional scale matrix is Σ22|1 inflated or deflated by the factor (ν + d1 )/(ν + p1 ). Because
d1 ∼ p1 × F(p1 , ν ) with mean being p1 ν /(ν − 2) or infinity according to ν > 2 or ν ≤ 2 (Roth 2013a,
page 83), on average the factor (ν + d1 )/(ν + p1 ) is larger than one. With more extreme values of
X1 , the conditional distributions of X2 are more disperse. More interestingly, the conditional degrees
4
of freedom increase to (ν + p1 ). The more dimensions we condition on, the less heavy-tailedness
we have. When ν = 1, the MVT distribution is also called the multivariate Cauchy distribution. The
marginal distributions are still Cauchy, but the conditional distributions are MVT instead of Cauchy.
Although the marginal means of the multivariate Cauchy distribution do not exist for any dimensions,
all the conditional means exist because the conditional distributions of the multivariate Cauchy distribution follow MVT distributions with degrees of freedom at least as large as two.
As a byproduct, we can easily obtain from the representation (5) that
r
ν + p1
X2 − µ2|1 | X1 ∼ t p2 0, Σ22|1 , ν + p1 ,
ν + d1
which does not depend on X1 . The above result further implies the following conclusion.
q
1
Conclusion Two: X1 and νν +p
+d1 X2 − µ2|1 are independent.
It is straightforward to show that X1 and X2 − µ2|1 are uncorrelated for both multivariate nor-
mal and MVT distributions. If X follows the multivariate normal distribution, this also implies inde
pendence between X1 and X2 − µ2|1 . If X follows the MVT distribution, we need only to adjust
p
the linear regression residual X2 − µ2|1 by the factor (ν + p1 )/(ν + d1 ) and the independence
result still holds.
4 Discussion
Kotz and Nadarajah (2004, page 17) and Nadarajah and Kotz (2005) failed to recognize that the
conditional distribution of the MVT distribution is also a MVT distribution due to the complexity of
the conditional density function. Our proof, based on the normal mixture representation of the MVT
distribution, offers a more direct and transparent way to revealing the property of the conditional
distribution of the MVT distribution. Our representation in Section 3 implicitly exploits the data
augmentation formula f2|1 (x2 | x1 ) =
R
f2|1,q (x2 | x1 , q) fq|1 (q | x1 )dq (Tanner and Wong 1987). To
make our proof more intuitive, we used formula (5) to avoid the integral, which is essentially a Monte
5
Carlo version of the data augmentation formula.
Conclusion One could also follow from the general theory of elliptically contoured distributions,
including the MVT distribution as a special case (Cambanis et al. 1980; Fang et al. 1990). Conditional
distributions of elliptically contoured distributions are also elliptically contoured distributions. But this
does not immediately guarantee that conditional distributions of the MVT distributions are also MVT
distributions without some further algebra. The MVT distribution has the property of having MVT
marginal and conditional distributions. It will be interesting to find other sub-classes of elliptically
contoured distributions that have the same property.
References
Cambanis, S., Huang, S. and Simons, G. (1981). On the theory of elliptically contoured distributions.
Journal of Multivariate Analysis, 11, 368–385.
Ding, P. (2014). Bayesian robust inference of sample selection using selection-t models. Journal of
Multivariate Analysis, 124, 451–464.
DeGroot, M. H. (2005). Optimal Statistical Decisions. New York: John Wiley & Sons.
Fang, K. T., Kotz, S., and Ng, K. W. (1990). Symmetric Multivariate and Related Distributions.
London: Chapman & Hall.
Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D., Vehtari, A., and Rubin, D. B. (2014). Bayesian
Data Analysis, 2nd edition. London: Chapman & Hall/CRC.
Kibria, B. G., and Joarder, A. H. (2006). A short review of multivariate t-distribution. Journal of
Statistical Research, 40, 59–72.
Kotz, S. and Nadarajah, S. (2004). Multivariate t Distributions and Their Applications. New York:
Cambridge University Press.
Lange, K. L., Little, R. J., and Taylor, J. M. (1989). Robust statistical modeling using the t distribution.
6
Journal of the American Statistical Association, 84, 881–896.
Liu, C. (1994). Statistical Analysis Using the Multivariate t Distribution. Doctoral dissertation, Harvard University.
Liu, C. (2004). Robit regression: a simple robust alternative to logistic and probit regression. In
Gelman, A. and Meng, X.-L. (Eds.) Applied Bayesian Modeling and Causal Inference from
Incomplete-Data Perspectives (pp. 227–238). New York: John Wiley & Sons.
Marchenko, Y. V., and Genton, M. G. (2012). A Heckman selection-t model. Journal of the American
Statistical Association, 107, 304–317.
Nadarajah, S. and Kotz, S. (2005). Mathematical properties of the multivariate t distribution. Acta
Applicandae Mathematica, 89, 53–84.
Pinheiro, J. C., Liu, C., and Wu, Y. N. (2001). Efficient algorithms for robust estimation in linear mixed-effects models using the multivariate t distribution. Journal of Computational and
Graphical Statistics, 10, 249–276.
Roth, M. (2013a). Kalman Filters for Nonlinear Systems and Heavy-Tailed Noise, Licentiate of Engineering Thesis, Linköping University, Linköping.
Roth, M. (2013b). On the multivariate t distribution. Technical report from Automatic Control at
Linköpings Universitet. Report no.: LiTH-ISY-R-3059.
Tanner, M. A., and Wong, W. H. (1987). The calculation of posterior distributions by data augmentation. Journal of the American Statistical Association, 82, 528–540.
7
| 10math.ST
|
arXiv:1802.03554v1 [math.GR] 10 Feb 2018
Breaking points in centralizer lattices
Marius Tărnăuceanu
January 4, 2018
Abstract
In this note, we prove that the centralizer lattice C(G) of a group
G cannot be written as a union of two proper intervals. In particular,
it follows that C(G) has no breaking point. As an application, we
show that the generalized quaternion 2-groups are not capable.
MSC 2010 : Primary 20D30; Secondary 20D15, 20E15.
Key words : breaking point, centralizer lattice, interval, generalized quaternion
2-group, capable group.
1
Introduction
Let G be a finite group and L(G) be the subgroup lattice of G. The starting
point for our discussion is given by [2], where the proper nontrivial subgroups
H of G satisfying the condition
(1)
for every X ∈ L(G) we have either X ≤ H or H ≤ X
have been studied. Such a subgroup is called a breaking point for the lattice
L(G), and a group G whose subgroup lattice possesses breaking points is
called a BP-group. Clearly, all cyclic p-groups of order at least p2 are BPgroups. Note that a complete classification of BP-groups can be found in [2].
Also, we observe that the condition (1) is equivalent to
(2)
L(G) = [1, H] ∪ [H, G],
1
where for X, Y ∈ L(G) with X ⊆ Y we denote by [X, Y ] the interval in L(G)
between X and Y . A natural generalization of (2) has been suggested by
Roland Schmidt, namely
(3)
L(G) = [1, M] ∪ [N, G] with 1 < M, N < G,
and the abelian groups G satisfying (3) have been determined in [1].
The above concepts can be naturally extended to other remarkable posets
of subgroups of G, and also to arbitrary posets. We recall here that the
generalized quaternion 2-groups
n−2
Q2n = ha, b | a2
n−1
= b2 , a2
= 1, b−1 ab = a−1 i, n ≥ 3
can be characterized as being the unique finite non-cyclic groups whose posets
of cyclic subgroups and of conjugacy classes of cyclic subgroups have breaking
points (see [7] and [3], respectively).
In the current note, we will focus on the centralizer lattice
C(G) = {CG (H) | H ∈ L(G)}
of G. Note that this is a complete meet-sublattice of L(G) with the least
element Z(G) = CG (G) and the greatest element G = CG (1). We will
prove that there are no proper centralizers M and N such that C(G) =
[Z(G), M] ∪ [N, G]. This implies that C(G) does not have breaking points.
As an application, we show that Q2n is not a capable group, i.e. there is no
group G with G/Z(G) ∼
= Q2n (see e.g. the main theorem of [6]).
Most of our notation is standard and will usually not be repeated here.
Elementary concepts and results on group theory can be found in [4]. For
subgroup lattice notions we refer the reader to [5] .
2
Main results
Our main theorem is the following.
Theorem 1. Let G be a group and C(G) be the centralizer lattice of G. Then
C(G) cannot be written as C(G) = [Z(G), M] ∪ [N, G] with M, N 6= Z(G), G.
Proof. Assume that there are two proper centralizers M and N such that
C(G) = [Z(G), M] ∪[N, G]. Then for every x ∈ G we have either CG (x) ≤ M
or N ≤ CG (x). In the first case we infer that x ∈ M, while in the second one
we get x ∈ CG (CG (x)) ≤ CG (N), that is x ∈ CG (N). Thus, the group G is
the union of its proper subgroups M and CG (N), a contradiction.
2
Clearly, by taking M = N in Theorem 1, we obtain the following corollary.
Corollary 2. The centralizer lattice C(G) of a group G has no breaking point.
Next we remark that for an abelian group G we have C(G) = {G}, and
also that there is no non-abelian group G with C(G) = {Z(G), G} (i.e. C(G)
is not a chain of length 1). Since chains of length at least 2 have breaking
points, Corollary 2 implies that:
Corollary 3. The centralizer lattice C(G) of a group G cannot be a chain of
length ≥ 1. Moreover, C(G) is a chain if and only if G is abelian.
Another consequence of Corollary 2 is:
Corollary 4. The generalized quaternion 2-groups Q2n , n ≥ 3, are not capable groups.
Proof. Assume that there is a group G such that G/Z(G) ∼
= Q2n . Obviously,
G is not abelian. Since Q2n has a unique subgroup of order 2, it follows that
the lattice interval [Z(G), G] contains a unique minimal element, say H. If
H ∈ C(G), then it is a breaking point of C(G), contradicting Corollary 2. If
H∈
/ C(G), then it is (properly)
T contained in all minimal centralizers M1 , M2 ,
..., Mk of G, and so H ⊆ ki=1 Mi . Note that a intersection of centralizers
T
is also a centralizer, that is ki=1 Mi ∈ C(G). On the other hand, we have
T
k ≥ 3 because G is non-abelian. Then ki=1 Mi < Mj , for any j = 1, 2, ..., k,
T
and therefore ki=1 Mi = Z(G) by the minimality of Mj ’s. Consequently,
H ⊆ Z(G), a contradiction.
Finally, we formulate an open problem concerning the above study.
Open problem. Let G be a group. Then C′ (G) = {CG (H) | H E G)} is also
a complete meet-sublattice of L(G) with the least element Z(G) = CG (G)
and the greatest element G = CG (1). Which are the groups G such that
C′ (G) has breaking points? (note that this can happen, as for G = S3 )
References
[1] A. Breaz and G. Calugareanu, Abelian groups whose subgroup lattice is
the union of two intervals, J. Aust. Math. Soc., vol. 78 (2005), no. 1,
pp. 27-36.
3
[2] G. Călugăreanu and M. Deaconescu, Breaking points in subgroup lattices,
Proceedings of Groups St. Andrews 2001 in Oxford, vol. 1, Cambridge
University Press, 2003, pp. 59-62.
[3] Y. Chen and G. Chen, A note on a characterization of generalized
quaternion 2-groups, C. R. Math. Acad. Sci. Paris, vol. 352 (2014), no.
6, pp. 459-461.
[4] I.M. Isaacs, Finite group theory, Amer. Math. Soc., Providence, R.I.,
2008.
[5] R. Schmidt, Subgroup lattices of groups, de Gruyter Expositions in Mathematics 14, de Gruyter, Berlin, 1994.
[6] S. Shariari, On normal subgroups of capable groups, Arch. Math., vol.
48 (1987), no. 3, pp. 193-198.
[7] M. Tărnăuceanu, A characterization of generalized quaternion 2-groups,
C. R. Math. Acad. Sci. Paris, vol. 348 (2010), no. 13-14, pp. 731-733.
Marius Tărnăuceanu
Faculty of Mathematics
“Al.I. Cuza” University
Iaşi, Romania
e-mail: [email protected]
4
| 4math.GR
|
Cognitive Database: A Step towards Endowing Relational
Databases with Artificial Intelligence Capabilities
Rajesh Bordawekar
IBM T. J. Watson Research
Center
Yorktown Heights, NY 10598
arXiv:1712.07199v1 [cs.DB] 19 Dec 2017
[email protected]
∗
Bortik Bandyopadhyay
†
The Ohio State University
Columbus, OH 43210
CS Department, Technion
Haifa 32000, Israel
bandyopadhyay.14
@osu.edu
[email protected]
ABSTRACT
We propose Cognitive Databases, an approach for transparently enabling Artificial Intelligence (AI) capabilities in
relational databases. A novel aspect of our design is to first
view the structured data source as meaningful unstructured
text, and then use the text to build an unsupervised neural
network model using a Natural Language Processing (NLP)
technique called word embedding. This model captures the
hidden inter-/intra-column relationships between database
tokens of different types. For each database token, the model
includes a vector that encodes contextual semantic relationships. We seamlessly integrate the word embedding model
into existing SQL query infrastructure and use it to enable
a new class of SQL-based analytics queries called cognitive
intelligence (CI) queries. CI queries use the model vectors to
enable complex queries such as semantic matching, inductive
reasoning queries such as analogies, predictive queries using
entities not present in a database, and, more generally, using
knowledge from external sources. We demonstrate unique
capabilities of Cognitive Databases using an Apache Spark
based prototype to execute inductive reasoning CI queries
over a multi-modal database containing text and images.
We believe our first-of-a-kind system exemplifies using AI
functionality to endow relational databases with capabilities that were previously very hard to realize in practice.
1.
Oded Shmueli
INTRODUCTION
Artificial Intelligence: Systems that perform actions that, if performed by humans, would be
considered intelligent –Marvin Minsky
Wikipedia defines cognition as the mental action or process of acquiring knowledge and understanding through thought,
∗Work done while the author was visiting the IBM Watson
Research Center.
†Work done while the author was visiting the IBM Watson
Research Center.
1
experience, and the senses. In broad terms, cognition refers
to the process of building knowledge capabilities using innate resources (i.e. intelligence), enriching it with external inputs such as experiences or interactions, and applying
the knowledge to solve problems which feeds back towards
knowledge building. While these definitions are more relevant to animate objects, they can be also applicable to
scenarios in which inanimate entities simulate cognitive processes.
We focus on a particular cognitive process of reading comprehension of text via contexts and apply it to relational
databases. In the relational model, some relationships between database values and entities are defined at the schema
level: data types, keys, and functional (and other) dependencies. Relationships at the instance level (i.e. actual data
tables) are left to be explored by queries. In a strong sense,
the actual semantics of the data mostly lies in users’ minds
and is expressed via queries. We take a significant diversion
from this point of view. We postulate that there is significant latent knowledge in a database instance irrespective of
querying. To capture this latent knowledge we propose to
use Artificial Intelligence (AI) techniques that take advantage of contexts.
Specifically, in the relational database model, the main
sources of latent information include the structure of database
(e.g., column names in a relation) as well as the types of associated data values that include unstructured natural language text, strings, numerical values, images, SQL Dates
etc. Together, these factors lead to inter- and intra-column
semantic relationships. Current systems have limited support to exploit this information, namely via SQL and extensions such as text extenders [14] or RDF-based ontologies [38]. However, SQL queries rely mainly on value-based
predicates to detect patterns. In addition, the relational
data model ignores many inter- or intra-column relationships. Thus, traditional SQL queries lack a holistic view of
the underlying relations and thus are unable to extract and
exploit semantic relationships that are collectively generated
by the various entities in a database relation.
A few examples may serve to clarify what we mean by
latent knowledge. The first example considers a Human Resources (HR) database. This database contains relations
with information about employees, their work history, pay
grade, addresses, family members and more. Lately there
have been some issues with an employee, John Dolittle. As
a HR professional, you are interested in names of employees
who know John well. Sure, you can get on the phone (or any
other media) and start making calls, collecting information
until you have a few names of people with whom you would
like to consult. Much of the information you will obtain is
already hidden in the legacy database, but is diffused and
hard to get at. This may include people who worked with
John, managed him, complimented him, complained about
him, provided technical services to him, were members of
a small team with him and so on. Wouldn’t it be nice if
you could write a SQL query that would use this hidden
knowledge and essentially ask provide the names of the 10
employees most related to John Dolittle.
In the previous example, we relied on the content of the
database in isolation. The next example involves entities
external to the database. Suppose you have database of active vacations, featuring diving, hiking, skiing, desert driving and more. You are a bit worried as to which vacation package is the most dangerous one. Naturally, the official descriptions in the database will not always provide
the information. It is likely that the words accident, danger, wounded, and death will not even be present in the
database. These words are present in other data sources,
from Wikipedia to news articles. Suppose you have access
to such external sources. Wouldn’t it be nice if you could
utilize these external sources and pose a SQL query expressing which are the most dangerous vacation packages. In this
example, as well as the previous ones, you may want also
to have a degree of certainty associated with each potential
answer.
It is worth pointing out what distinguishes the level of
intelligence we are looking for from known extensions to relational systems. In current systems one needs to pose a
query based on some knowledge of the relational schema.
The query may be assisted by text-aware features such as
DB2 Text Extender [14], WordNet or using RDF-based ontologies [38]. These may be used to identify synonyms and
related terms and relax the query by allowing it to explore more possibilities than those explicitly specified by the
user [11]. Of course, such relaxation may result in obtaining a larger result set. But, all these useful features assume
that the user knows how to specify a backbone query. The
example problems we listed above are such that formulating
an effective SQL query is a daunting task. In fact, these
examples resemble research projects rather than standard
queries. One can also allow the user to specify the query
in natural language [37], but this pushes the problem of expressing a query to an automated tool; again, it is unclear
how a tool will approach these problem if the tool’s writer
does not have a ready recipe. This highlights the need of a
new set of tools to enable far richer querying.
In this paper, we explore the potential of using Natural
Language Processing (NLP) approaches to endow databases
with query expression capabilities that were very hard, or
perhaps impossible, to realize in practice, and at a reasonable cost in terms of storage overhead as well as processing
time. The unique aspect of our proposal is to first represent
the data and optionally, schema, of a (structured) relational
database as an unstructured text document and then use a
NLP technique, Vector Space Models (VSM) [55], to extract
latent semantic relationships via associations in the generated text. The trained VSM model represents the semantic
meaning of the words as vectors and enables operations on
these vectors to mimic cognitive operations on natural language words. As these words represent relational entities
and values, the VSM model, in fact, captures intra-/intercolumn relationships in the relational database. We then integrate the VSM model into an existing standard SQL query
processing system and expose the novel vector-based cognitive operations via a new class of SQL analytics queries,
called Cognitive Intelligence (CI) queries [8]. We believe
this is one of the first examples of AI transperently augmenting a relational database system. Clearly, this is only
one of the many possible ways of integrating AI capabilities in database systems, e.g., for enhancing their querying
capabilities or improving their operational capabilities [50,
59, 37]. While our current focus is on enhancing relational
databases, we believe this approach can be applied to other
database domains such as XML/RDF or JSON databases,
document databases, graph databases, and key-value stores.
We are currently developing an Apache Spark-based prototype to implement our vision of an AI-enhanced cognitive
relational database. The rest of the paper provides more details on the design and implementation of such a system. In
Section 2, we introduce the vector space modeling process
and detail the execution flow of the system we envisage.
In Section 3, we provide specifics of the data preparation
and building a specialized vector space model. In Section 4
we discuss three significant classes of CI queries: Similarity
Queries, Inductive Reasoning Queries and Cognitive OLAP
Queries; we also present cognitive extensions to the Relational Data Model. We also describe the design of cognitive
User Defined Functions (UDFs). Section 5 describes a practical scenario which demonstrates unique aspects of a cognitive database system: ability to invoke inductive reasoning (e.g., analogies, semantic clustering, etc.) queries over
multi-modal data (e.g., images and text). We also discuss
CI query performance issues, with focus on an important
building block: Nearest-Neighbor Computations. Related
work is discussed in Section 6 and we conclude by outlining
extensions, future work, and success criteria in Section 7.
2.
DESIGNING A COGNITIVE DATABASE
Our goal is to build a cognitive relational database system
that not only extracts latent semantic information, but can
also enrich it by using external input (e.g., external knowledge bases, new data being inserted, or types of invoked
queries) and use it transparently to enhance its query capabilities. To achieve these goals, we rely on the VSM approach that infers word meanings using the distributional
hypothesis, which states that words in a neighborhood (or
context) contribute to each other’s meanings [24, 26]. Specifically, we use a predictive [4] implementation of the VSM
approach, commonly referred to as word embedding, which
assumes a probabilistic language model to capture relationships between neighborhood words [5]. The word embedding
approach fixes a d-dimensional vector space and associates
a vector of continuous-valued real numbers to a word to
encode the meaning of that word. Thus, for a given text
corpus, the meaning of a word reflects collective contributions of neighborhood words for different appearances of the
word in the corpus. Two words are closely related or have
similar meaning if they appear often within close proximity
of the same, or similar meaning, words. If two words have
similar meaning, their meaning vectors point in very similar
directions, i.e., the cosine similarity between their vectors is
high (vector similarity is measured as a cosine of the angle
2
between two vectors and can vary from 1.0 to -1.0). One surprising application of word-embedding vectors is their usage
in solving inductive reasoning problems such as computing
analogies by using vector algebra calculations [58, 54, 35].
Over a past few years, a number of methods have been
developed to implement the word embedding. Recently, an
unsupervised neural network based approach,
Word2Vec [42, 45], has gained popularity due to its performance and ability to capture syntactic as well semantic
properties of words. We use Word2Vec as it is easy to adapt,
can be trained incrementally, and can be used for building
models from both structured and unstructured data sources
(alternatively, we could use approaches such as GloVe [51]).
In the database context, vectors may be produced by either learning on text transformed and extracted from the
database itself and/or using external text sources. For learning from a database, a natural way of generating vectors is
to apply the word embedding method to a string of tokens
generated from the database: each row (tuple) would correspond to a sentence and a relation would correspond to a
document. Thus, vectors enable a dual view of the data: relational and meaningful text. To illustrate this process, consider Figures 1 and 2 that present a simple customer sales
table. Figure 1 shows an English sentence-like representation of the fourth row in the table (note that the numeric
value 25.00 is represented by the string cluster 10. We
discuss the reasons in the next section.). Using the scope of
the generated sentence as the context, the word embedding
approach infers latent semantic information in terms of token associations and co-occurrences and encode it in vectors.
Thus, the vectors capture first inter- and intra-column relationships within a row (sentence) and then aggregate these
relationships across the relation (document) to compute the
collective semantic relationships. At the end of training,
each unique token in the database would be associated with
a d-dimensional meaning vector, which can be then used to
query the source database. As a simple example, the relational entity custD is semantically similar to custB due
to many common semantic contributors (e.g., Merchant B,
Stationery, and Crayons). Equivalently, custA is similar
to custC due to similar reasons.
custID
Date
Merchant
Address
Category
custA
9/16/17
Merchant_A
NY
Fresh Produce
custB
10/16/17
Merchant_B
NJ
Stationery
custC
10/16/17
Merchant_A
NJ
Fresh Produce
custD
9/16/17
Merchant_B
NY
custD
9/16/17
Merchant_B
Stationery
NY
Stationery
Items
60.80
Bananas, Oranges
80.10
Crayons, Folders
25.00
Crayons Folders
Merchant
Address
Category
9/16/17
Merchant_A
NY
Fresh Produce
custB 10/16/17
Merchant_B
NJ
Stationery
custC 10/16/17
Merchant_A
NJ
Fresh Produce
custD
Merchant_B
NY
9/16/17
Stationery
NY
Amount
Items
200.50
Bananas, Apples
Crayons, Pens, Notebooks
60.80
Bananas, Oranges
80.10
Crayons, Folders
25.00
cluster_10
Figure 2: Example of customer analytics with a different relational view
This examples illustrates a key design feature of our cognitive database: the neighborhood context used for building
the word embedding model is determined by the relational
view being used. Hence, the inferred semantic meaning of
the relational entities reflect the collective relationships defined by the associated relational view.
Vector
Domain
Learned Vectors
External
Learned Vectors
UDFs
External Text sources
Text
Domain
Relational
Domain
Tokenized Relations
Relational
Tables
(1) Optional Training Phase
Relational
System Tables
(2) Vector Storage Phase
CI Queries
Relations
(3) Query Execution Phase
Figure 3: End-to-end execution flow of a cognitive
relational database
The cognitive relational database has been designed as an
extension to the underlying relational database, and thus
supports all existing relational features. The cognitive relational database supports a new class of business intelligence
(BI) queries called Cognitive Intelligence (CI) queries. The
CI queries extract information from a relational database
based, in part, on the contextual semantic relationships among
database entities, encoded as meaning vectors. Figure 3
presents key phases in the end-to-end execution flow of a cognitive relational database system. The first, optional, phase
involves (1) Generating token sequences from the database
tables (textification), and then applying a word embedding
model training method on the unstructured text corpus created from these token sequences. Following model training,
the resultant vectors are stored in a relational system table (phase 2). At runtime, the SQL query execution engine
uses various user-defined functions (UDFs) that fetch the
trained vectors from the system table as needed and answer
CI queries (phase 3). The CI queries take relations as input and return a relation as output. CI queries augment
the capabilities of the traditional relational BI queries and
can be used in conjuction with existing SQL operators (e.g.,
OLAP [22]).
200.50
Crayons, Pens, Notebooks
Date
custA
custD 9/16/17
Amount
Bananas, Apples
custID
cluster_10
Figure 1: Example of customer analytics
We may use a relational view of a table, rather than the
original table, to generate text representing the database
content. This may be useful for a supporting a particular class of applications. Consider a scenario in which
a view of the table is defined (Figure 2), the view only
projects data from the bold columns (Cust, Date, Address
and Amount). In this case, the generated sentence-like representation would be different than the first case. Hence, it
will generate a different word embedding model.
3.
BUILDING THE SEMANTIC MODEL
The key to artificial intelligence has always been
the representation. –Jeff Hawkins
In this section, we discuss how we train a word embedding
model using data from a relational database. Our training
3
approach is characterized by two unique aspects: (1) Using
unstructured text representation of the structured relational
data as input to the training process (i.e. irrespective of the
associated SQL types, all entries from a relational database
are converted to text tokens representing them), and (2) Using the unsupervised word embedding technique to generate
meaning vectors from the input text corpus. Every unique
token from the input corpus is associated with a meaning
vector. We now elaborate on these two aspects.
3.1
specify rules to define ranges for the numeric values and
use them to generate string tokens for the numeric values.
For example, consider values for a column name, Cocoa
Contents. The value 80%, can be replaced by the string
token choc dark, while the value 35%, can be replaced by
the string token choc med, etc., and (3) user-directed clustering: an user can choose values of one or more numerical columns and cluster them using traditional clustering
algorithms such as K-Means. Each numeric value is then
replaced by a string representing the cluster in which that
value lies (e.g., cluster 10 for value 25 in Figure 1).
For image data, we use approaches similar to ones used for
numerical values. The first approach represents an image by
its string token, e.g., a string representing the image path or
a unique identifier. The second approach uses pre-existing
classifers to cluster images into groups and then uses the
cluster information as the string representation of the image. For example, one can use a domain-specific deep neural
network (DNN) based classifier to cluster input images into
classes [31] and then use the corresponding class information to create the string identifiers for the images. The final
approach applies of-the-shelf image to tag generators, e.g.,
IBM Watson Visual Recognition System (VRS) [30], to extract image features and uses them as string identifiers for
an image. For example, a Lion image can be represented by
the following string features, Animal, Mammal, Carnivore,
BigCat, Yellow, etc.
Once text, numeric values and images are replaced by
their text representations, a relational table can be viewed as
unstructured meaningful text corpus to be used for building
an word embedding model. For Null values of these types,
we replace them by the string column name Null. The methods outlined here can be applied to other data types such
as SQL Date and spatial data types such as lattitude and
longitude.
Data Preparation
The data preparation stage takes a relational table with
different SQL types as input and returns an unstructured
but meaningful text corpus consisting of a set of sentences.
This transformation allows us to generate a uniform semantic representation of different SQL types. This process of
textification requires two stages: data pre-processing and
text conversion (Figure 4).
(A)
Relational Data
Text, Numeric Values, Images, ...
(B)
Preprocessing
(C)
Meaningful Text
(D)
Word
Embedding
NN
(E)
Trained Model
Externally
Pre−trained Model
Figure 4: Multiple stages in creating the word embedding model
The textification phase processes each relational row seperately and for each row, converts data of different SQL data
types to text. In some scenarios, one may want to build a
model that also captures relational column names. For such
cases, the pre-processing stage first processes the column
names before processing the corresponding data.
For SQL variables of VARCHAR type, preprocessing involves
one or more of the following actions: (1) prepend the column
attribute string to a SQL variable, (2) creating a single concept token from a group of VARCHAR tokens, e.g., JPMorgan
Chase is represented as JPMorgan Chase, (3) creating a single
token for semantically similar sequences of VARCHAR tokens,
e.g., two sequences of tokens, bank of america and BANK
OF AMERICA, can be represented by a single compound token BANK OF AMERICA, and (4) Using an external mapping
or domain-specific ontologies to create a common representative token for a group of different input tokens. This approach is useful for enabling transfer learning via reusing
the same training model for a group of related tokens. After
pre-processing, all input text tokens have uniform representations.
In addition to text tokens, our current implementation
supports numeric values and images (we assume that the
database being queried contains a VARCHAR column storing
links to the images). These techniques can be applied to
other SQL datatypes such as SQL Date as well. For numeric values, we use three different approaches to generate
equivalent text representations: (1) creating a string version of the numerical value, e.g., value 100.0 for the column name price can be represented by either PRICE 100.0
or ‘‘100.0’’, (2) User-managed categorization: a user can
3.2
Model Training
We use an unsupervised approach, based on the Word2Vec
(W2V) implementation [42], to build the word embedding
model from the relational database data. Our training approach operates on the unstructured text corpus, organized
as a collection of English-like sentences, separated by stop
words (e.g., newline). There is no need of labelling the training data as we use unsupervised training. Another advantage of unsupervised training is that users do not need to
do any feature engineering [20]; features of the training set
are extracted automatically by the training process.
During model training, the classical W2V implementation
uses a simplified 3-layer shallow neural network that views
the input text corpus as a sequence of sentences. For each
word in a sentence, the W2V code defines a neighborhood
window to compute the contributions of nearby words. Unlike deep learning based classifiers, the output of W2V is
a set of vectors of real values of dimension d, one for each
unique token in the training set (the vector space dimension
d is independent of the token vocabulary size). In our scenario, a text token in a training set can represent either text,
numeric, or image data. Thus, the model builds a joint latent representation that integrates information across different modalities using untyped uniform feature (or meaning)
vectors.
Our training implementation builds on the classical W2V
implementation, but it varies from the classical approach in
4
empl Relation
(Primary Key)
a number of ways (Figure 4):
empNum
• A sentence generated from a relational row is generally
not in any natural language such as English.1 Therefore, W2V’s assumption that the influence of any word
on a nearby word decreases as the word distances increases, is not applicable. In our implementation, every token in the training set has the same influence on
the nearby tokens; i.e. we view the generated sentence
as a bag of words, rather than an ordered sequence.
firstName
259
John
lastName
salary
Smith
95
dept
jobDescr
Multimedia
manager
multimedia
entertainment
eval
address
good people skills
not punctual
need improvement
119
(Foreign Key)
address Relation
id
stNum
119
100
street
10th
city
state
Newark
NJ
zip
07105
remarks
alternate 19 Chatsworth Ave Larchmont NY
(Primary Key)
259 John Smith 95 Multimedia manager multimedia entertainment good people skills
not punctual need improvement 119 100 10th Newark NJ 07105
alternate 19 Chatsworth Ave Larchmont NY ..............................................................
• Another consequence is that unlike an English sentence, the last word is equally related to the first word
as to its other neighbors. To enable such relationships, we use a circular neighborhood window that
wraps around a sentence (i.e. for the last word, the
first word can be viewed as its immediate neighbor).
Figure 5: Text view of two tables joined using primary and foreign keys
• Build models each based on linked tables where, usually, the linking is based on foreign keys appearing in
say table A pointing to tuples into another, say table
B. When a foreign key is present, during tokenization
of table A, we can follow the foreign key to a row in
table B. We can then tokenize fields of interest in the
row of table B and insert the resulting sequences into
the sequence generated for table A. Figure 5 presents
another example of a database table, address, and a
resulting token sequence that utilizes a relationship between the empl table and the address table; namely
the address table provides the addresses for the employees of database table empl. Technically, the resulting token sequence is based on foreign key 119 in
the address column of the table emp which provides
a value for key column id of the address table. The
straight forward way to tokenize with foreign keys is
to insert the subsequence generated out of the B row
immediately after the one generated for the A row as
depicted in Figure 5; another possibility is to intermix
the subsequence from the B row within the A row sequence following the tokenization of the foreign keys
values (again, other options may apply).
• For relational data, we provide special consideration
to primary keys. First, the classical W2V discards
less frequent words from computations. In our implementation, every token, irrespective of its frequency,
is assigned a vector. Second, irrespective of the distance, a primary key is considered a neighbor of every
other word in a sentence and included in the neighborhood window for each word. Also, the neighborhood
extends via foreign key occurrences of a key value to
the row in which that value is key.
• In some cases, one may want to build a model in which
values of particular columns are given higher weightage for their contributions towards meanings of neighborhood words. Our implementation enables users to
specify different weights (or attention [3]) for different
columns during model training (in this scenario, one
needs to use a training set in which column names are
embedded).
• Finally, our implementation is designed to enable incremental training, i.e. the training system takes as
input a pre-trained model and a new set of generated
sentences, and returns an updated model. This capability is critical as a database can be updated regularly
and one can not rebuild the model from scratch every
time. The pre-trained model can be built from the
database being queried, or from an external source.
Such sources may be publicly available general sources
(e.g., Wikipedia), text from a specific domain (e.g.,
from the FDA regarding medical drugs), text textified from other databases or text formed from a different subset of tables of the same database. The use of
pre-trained models is an example of transfer learning,
where a model trained on an external knowledge base
can be used either for querying purposes or as a basis
of a new model [20].
• A collection of tables may be identified and textified
into a collection of texts. These texts may be concatenated to form a single text which may be used in
training. The tables in this collection should form a
coherent informative subset of the database.
• In all the options above, the training text may be augmented with text from external sources.
4.
BUILDING A COGNITIVE DATABASE
SYSTEM
Cognitive intelligence (CI) queries are standard SQL queries
and can be implemented using the existing SQL query execution infrastructure. The distinguishing aspect of cognitive intelligence queries, contextual semantic comparison between relational variables, is implemented using user-defined
functions (UDFs). These UDFs, termed cognitive UDFs,
take typed relational values as input and compute semantic
relationships between them using uniformly untyped meaning vectors. This enables the relational database system to
seamlessly analyze data of different types (e.g., text, numeric
values, and images) using the same SQL CI query.
Our current implementation is built on the Apache Spark
2.2.0 infrastructure. Our system follows the cognitive database
In practice, enterprise database systems, as well as data
warehouses, are built using many inter-related database tables. Forming a training corpus from multiple tables is nontrivial. There are numerous options, including:
• Build separate models (i.e. a set of word vectors), each
based on an individual, informative, table.
1
Currently, we assume that database tokens are specified
using the English language.
5
execution flow as presented in Figure 3. The system first
initializes in-memory Spark Dataframes from external data
sources (e.g., relational database or CSV files), loads the associated word embedding model into another Spark Dataframe
(which can be created offline from either the database being
queried or external knowledge bases such as Wikipedia), and
then invokes CI queries using Spark SQL. The SQL queries
invoke Scala-based cognitive UDFs to enable computations
on the meaning vectors (we also provide a Python based
implementation).
4.1
(1) Similarity/Dissimilarity Queries: The basic UDF
that compares two sets of relational variables can be integrated into an existing SQL query to form a similarity CI
query. Figure 6 illustrates a SQL CI query that identifies
similar customers by comparing their purchases. Assume
that sales is a table that contains all customer transactions for a credit card company and whose sales.Items
column contains all items purchased in a transaction (Figure 1). The current query uses a UDF, similarityUDF(),
that computes similarity match between two sets of vectors,
that correspond to the items purchased by the corresponding customers. Unlike the food item scenario, the purchased
item list can be viewed as an unordered bag of items; and
individual pair-wise distances contribute equally to the final result. The query shown in Figure 6 uses the similarity score to select rows with related customers and returns
an ordered set of similar customer IDs sorted in descending order of their similarity score. This query can be easily
tweaked to identify dissimilar customers based on their purchases. The modified CI query will first choose rows whose
purchases have lower similarity (e.g., < 0.3) and if the results are ordered in an ascending form using the SQL ASC
keyword, returns customers that are highly dissimilar to a
given customer (i.e., purchasing very different items). If the
results are ordered in the descending order using the SQL
DESC keyword, the CI query will return customers that are
somewhat dissimilar to a given customer.
Design of Cognitive UDFs
A cognitive UDF takes as input either relational query
variables or constant tokens, and returns a numeric similarity value that measures the semantic relationships between
the input parameters. A user can then control the result
of the CI query by using a numerical bound for similarity
result value as a predicate for selecting eligible rows. A user
can also use SQL ordering clauses, DESC or ASC, to order
results based on the similarity value that captures the semantic closeness between the relational variables: higher is
the similarity value, closer are these two relational variables.
The UDFs perform three key tasks: (1) processing input relational variables to generate tokens used for training. This
involves potentially repeating the steps executed during the
data preparation stage, such as creating compound tokens.
For numeric values, one can use the centroid information to
identify the corresponding clusters. For images, the UDF
uses the image name to obtain corresponding text tokens,
(2) Once the training tokens are extracted, the UDF uses
them to fetch corresponding meaning vectors, and (3) Finally, the UDF uses the fetched vectors to execute similarity computations to generate the final semantic relationship
score.
The basic cognitive UDF operates on a pair of sets (or
sequences) of tokens associated with the input relational parameters (note: value of a relational parameter can be a set,
e.g., {Bananas, Apples}, see Figure 1). The core computational operation of a cognitive UDF is to calculate similarity
between a pair of tokens by computing the cosine distance
between the corresponding vectors. For two vectors v1 and
·v2
.
v2 , the cosine distance is computed as cos(v1 , v2 ) = kvv11kkv
2k
The cosine distance value varies from 1.0 (very similar) to
-1.0 (very dissimilar). For sets and sequences, the individual
pair-wise similarity values are then aggragated to generate
the result. In case of sequences, computation of the final
similarity value takes into account the ordering of tokens:
different pair-wise distances contribute differently to the final value based on their relative ordering. For example, two
food items, a chicken item consisting of (chicken, salt), is
not very similar to a corn item consisting of (corn, salt), although both contain salt (however, the corn item is closer
to chicken item than a wheat item that contains (wheat,
sugar)).
4.2
SELECT X.custID, Y.custID, similarityUDF(X.Items,
Y.Items) AS similarity
FROM sales X, sales Y
WHERE similarityUDF(X.Items, Y.Items) > 0.5
ORDER BY similarity DESC
Figure 6: Example of an CI similarity query: find
similar customers based on their purchased items
A modified version of the query (not shown) can identify
similar customers based on their overall purchasing pattern
as evidenced in a number of rows. The word embedding
model creates a vector for each customer name that captures the overall purchases made by that customer. Then,
the customers with similar purchase patterns would have
vectors that are close using the cosine distance metric. The
pattern observed in this query can be applied to other domains as well, e.g., identifying patients that are taking similar drugs, but with different brand names, or identifying
food items with similar ingredients, or recommending mutual funds with similar investment strategies. As we will see
in the next section, the similarity query can be applied other
data types, such as images.
SELECT X.custID, Y.custID, Y.Merchant,
valueSimUDF(X.Amount, Y.Amount) AS similarity
FROM sales X, sales Y
WHERE X.custID=’custA’ AND valueSimUDF(X.Amount,
Y.Amount) > 0.5 AND X.custID != Y.custID AND X.amount >
150.0 AND Y.amount < 100.0
ORDER BY similarity DESC
Cognitive Intelligence Queries
The basic UDF and its extensions are invoked by the SQL
CI queries to enable semantic operations on relational variables. Each CI query uses the UDFs to execute nearest
neighbor computations using the vectors from the current
word-embedding model. Thus, CI queries provide approximate answers that reflect a given model. The CI queries
can be broadly classified into four categories as follows:
Figure 7: Example of an CI value similarity query:
find similar transactions using purchased amount for
comparison
6
SELECT X.custID, similarityUDF(X.Items, ‘listeria’) AS
similarity
FROM sales X
WHERE similarityUDF(X.Items, ‘listeria’) > 0.3
ORDER BY similarity DESC
LIMIT 10
SELECT X.custID, analogyUDF(‘Frozen Goods’,
‘custF’,‘Fresh Produce’,X.custID) AS similarity
FROM sales X
WHERE analogyUDF(‘Frozen Goods’, ‘C3423567’,‘Fresh
Produce’,X.custID) > 0.5
ORDER BY similarity DESC
Figure 8: Example of a prediction query: find customers that have purchased items affected by a listeria recall
Figure 9: Example of an analogy query
• Analogies: Wikipedia defines analogy as a process of
transferring information or meaning from one subject
to another. A common way of expressing an analogy is to use relationship between a pair of entities,
source 1 and target 1, to reason about a possible target entity, target 2, associated with another known
source entity, source 2. An example of an analogy
query is Lawyer : Client :: Doctor :?, whose answer
is P atient. To solve an analogy problem of the form
(X : Y :: Q :?), one needs to find a token W whose
meaning vector, Vw , is closest to the ideal response
vector VR , where VR = (VQ + VY − Vx ) [54]. Recently,
several solutions have been proposed to solve this formulation of the analogy query [35, 39, 45]. We have
implemented the 3COSMUL approach [35] which uses
both the absolute distance and direction for identifying
the vector VW as
Figure 7 presents a CI query executing similarity operations using a numeric variable. For the sake of example, assume that no two transactions have the same amount value
and each unique numerical value is associated with its own
string token. In this scenario, one wants to identify transactions from the sales table (Figure 1): using similarity based
on the purchase amount (in this case, 200.50 for customer
custA). The UDF, valueSimUDF(), takes two numeric values
as input parameters, and compares them using their overall
context (which is captured in their meaning vectors), not
their numerical values. The most similar amount to 200.50,
would be 80.10 (for customer custC), as it shares the most
context (e.g., category, address, merchant, and items). The
least similar amount would be 60.80 as it has completely different context than amount 200.50. This example also illustrates how one can combine the value-based and semanticbased comparisons in the same SQL query.
The third use case provides an illustration of a prediction
CI query which uses a model that is externally trained using
an unstructured data source or another database (Figure 8).
Consider a scenario of a recall of various fresh fruit types due
to possible listeria infection. This example assumes that we
have built a word embedding model using the recall notices
as an external source. Assume that the recall document lists
all fruits impacted by the possible listeria infection, e.g., Apples, Peaches, Plums, Nectarines,... The model will create
vectors for all these words and the vector for the word listeria will be closer to the vectors of Apples, Peaches, Plums,
etc. Now, we can import this model and use it to query the
sales database to find out which customers have bought
items that may be affected by this recall, as defined by the
external source. As Figure 8 shows, the similarityUDF()
UDF is used to identify those purchases that contain items
similar to listeria, such as Apples. This example demonstrates a very powerful ability of CI queries that enables
users to query a database using a token not present in the
database (e.g., listeria). This capability can be applied to
different scenarios in which recent, updatable information,
can be used to query historical data. For example, a model
built using a FDA recall notices could be used to identify
those customers who have purchased medicines similar to
the recalled medicines.
(2) Inductive Reasoning Queries: An unique feature
of word-embedding vectors is their capability to answer inductive reasoning queries that enable an individual to reason
from part to whole, or from particular to general [54, 58]. Solutions to inductive reasoning queries exploit latent semantic structure in the trained model via algebraic operations
on the corresponding vectors. We encapsulate these operations in UDFs to support following five types of inductive
reasoning queries: analogies, semantic clustering, analogy
sequences, clustered analogies, and odd-man-out [54].
arg max
W ∈C
cos(VW , VQ )cos(VW , VY )
cos(VW , VX ) +
(1)
where = 0.001 is used to avoid the denominator becoming 0. Also, 3COSMUL converts the cosine similarity value of c to (c+1)
to ensure that the value being
2
maximized is non-negative.
Figure 9 illustrates a CI query that performs an analogy computation on the relational variables using the
UDF analogyUDF(). This query aims to find a customer from the sales table (Figure 1), whose relationship to the category, Fresh Produce, is similar to what
C3423567 has with the category, Frozen Goods (i.e. if
C3423567 is the most prolific shopper of frozen goods,
find other customers who are the most prolific shoppers
of fresh produce). The analogyUDF() UDF fetches vectors for the input variables, and using the 3COSMUL
approach, returns the analogy score between a vector
corresponding to the input token and the computed
response vector. Those rows, whose variables (e.g.,
custID) have analogy score greater than a specified
bound (0.5), are selected, and returned in descending
order of the score. Since analogy operation is implemented using untyped vectors, the analogyUDF() UDF
can be used to capture relationships between variables
of different types, e.g., images and text.
• Semantic Clustering: Given a set of input entities,
{X, Y, Z, ..}, the semantic clustering process identifies
a set of entities, {W, ..}, that share the most dominant
trait with the input data. The semantic clustering operation has a wide set of applications, including customer segmentation, recommendation, etc. Figure 10
presents a CI query which uses a semantic clustering UDF, semclusterUDF(), to identify customers that
7
SELECT X.custID, semclusterUDF(‘custX’, ‘custY’,
‘custZ’, X.custID) AS similarity
FROM sales X
WHERE semclusterUDF(‘custX’,‘custY’, ‘custZ’, X.custID)
> 0.5
ORDER BY similarity DESC
SELECT X.Category, MAX(X.Amount)
FROM sales X
WHERE similarityUDF(‘Merchant Y’,X.Merchant) > 0.5
GROUP BY X.Category
Figure 11: Example of a cognitive OLAP (aggregation) query
Figure 10: Example of a semantic clustering query
execution requires context-specific semantic clustering
over the meaning vectors. Specifically, the clustering
aims to partition the data into two clusters, one with
only one member, and the other containing the remaining data. One obvious application of the odd-man-out
CI query would anomaly detection, e.g., for identifying
a fraudalant transaction for a customer.
have the most common attributes with the input set of
customers, e.g., custF, custM, and custR. For solving
a semantic clustering query of the form, (X, Y, Z ::?),
one needs to find a set of tokens Sw = {W1 , W2 , .., Wi }
whose meaning vectors Vwi are most similar to the centroid of vectors VX , VY , and VZ (the centroid vectors
captures the dominant features of the input entities).
(3) Cognitive OLAP Queries: Figure 11 presents a simple example of using semantic similarities in the context of
a traditional SQL aggregation query. This CI query aims to
extract the maximum sale amount for each product category
in the sales table for each merchant that is similar to a specified merchant, Merchant Y. The result is collated using the
values of the product category. As illustrated earlier, the
UDF similarityUDF can also be used for identifying customers that are different than the specified merchant. The
UDF can use either an externally trained or locally trained
model. This query can be easily adapted to support other
SQL aggregation functions such as MAX(), MIN(), and AVG().
This query can be further extended to support ROLLUP operations over the aggregated values [27].
We are also exploring integration of cognitive capabilities
into additional SQL operators, e.g., IN and BETWEEN. For
example, one or both of the value ranges for the BETWEEN
operator can be computed using a similarity CI query. For
an IN query, the associated set of choices can be generated by
a similarity or inductive reasoning queries. Another intriguing extension involves using contextual similarities to choose
members of the schema dimension hierarchy for aggregation
operations like ROLLUP or CUBE. For example, instead of aggregating over all quarters for all years, one can use only
those quarters that are semantically similar to a specified
quarter.
(4) Cognitive Extensions to the Relational Data Model:
There are powerful extensions to SQL that are enabled by
word vectors. For this we need the ability to refer to constituent tokens (extracted during textification) in columns
of rows, in whole rows and in whole relations. The extension
is via a declaration, in the FROM clause, of the form Token
e1 that states that variable e1 refers to a token. To locate
a token we use, in the WHERE clause, predicates of the form
contains(E, e1) where E can be a column in a row (e.g.,
EMP.Address), a whole row (e.g., EMP.*) or a whole relation
(e.g., EMP). With this extension we can easily express queries
such as asking for an employee whose Address contains a token which is very close to a token in a row in the DEPT relation (Figure 12). Furthermore, we can also extend SQL with
relational variables, say of the form $R and column variables,
say X, whose names are not specified at query writing time;
they are bound at runtime. We can then use these variables
in queries, in conjunction with Token variables. This enables database querying without explicit schema knowledge
which is useful for exploring a database. Interestingly, the
• Analogy Sequences and Clustered Analogies: These
two types of inductive reasoning queries, analogy sequences and clustered analogies, can be implemented
by combining strategies for semantic clustering and
analogies. The analogy sequence query takes as input a
sequence of analogy pairs, with the same source entity
and aims to identify a set of target entities that exhibit
the same relationships with the source entity as the
set of input target entities. To answer this query, one
needs to first compute the centroid vector of the input
target vectors and then use it to answer the following
analogy problem: source : input centroid :: source :?
using the 3COSMUL approach to return a set of target
entities.
Unlike analogy sequences, the clustered analogy operation takes as input a set of analogy pairs, each with different (sourcei , targeti ) entity pairs and aims to predict a set of (sourceo , targeto ) pairs that shares the
following relationships with the input sequence: the
result set of source entities, sourceo , share the dominant trait with the input set of source entities, sourcei ,
and each resultant target entity, targeto , is related to
the corresponding source entity via the analogy relationship. Therefore, to solve clustered analogy queries,
we first perform semantic clustering to compute the result source entities, sourceo , and then for each result
source entity, we compute a set of target entities using
the analogy sequence approach. Unlike the analogy
sequences query, the result of the clustered analogy
query is a set of sets: a set of source entities, each
associated with a set of target entities.
• Odd-man-out2 : As the name suggests, given a set of
items, the odd-man-out query identifies an item that
is semantically different from the remaining items [18].
The odd-man-out query can be viewed as a complementary query to semantic clustering. For example,
given a set of animals, {Hippopotamus, Giraffe, Elephant, and Lion}, one answer can be Lion, as it is
only carnivorous animal in the collection. However, if
the word-embedding model of these animals captures
their locations, Elephant may be the answer if it is
not present in that location. Thus, the odd-man-out
2
Odd-man-out is often used in practice to quantify human
intelligence [18].
8
notation $R.X is basically syntactic sugar. A software translation tool can substitute for $R.X an actual table name and
an actual column. Then, perform the query for each such
substitution and return the union of the results [8].
(A) Original Table
Picture ID
National Park
Country
Image Path
PK_01
Corbett
India
Img_01.JPEG
PK_05
Kruger
South Africa
Img_05.JPEG
PK_09
Sunderbans
India
Img_09.JPEG
PK_11
Serengeti
Tanzania
Img_11.JPEG
(B) Training Table
SELECT EMP.Name, EMP.Salary, DEPT.Name
FROM EMP, DEPT, Token e1, e2
WHERE contains(EMP.Address, e1) AND
contains(DEPT.*, e2) AND
cosineDistance(e1, e2) > 0.75
Picture ID
Image Path
National Park
Country
Animal Name
Class
Dietary Habit
color
PK_01
Img_01.JPEG
Corbett
India
Tiger
Mammal
Carnivorous
Yellow
PK_05
Img_05.JPEG
Kruger
South Africa
Rhinoceros
Mammal
Herbivores
Gray
PK_09
Img_08.JPEG
Sunderbans
India
Indian Gharial
Reptile
Carnivorous
Gray
PK_11
Img_11.JPEG
Serengeti
Tanzania
Crocodile
Reptile
Carnivorous
Gray
(C) Sentence for training
“PK_01 Img_01.JPEG Corbett India Tiger Mammal Carnivorous Yellow”
Figure 12: Example of an SQL query with entities
Training a shared model across features from multi-modal data
Lastly, one may wonder how numeric bounds on UDFs
(e.g., cosineDistance(e1, e2) > 0.75 in Figure 12) are
determined. The short answer is that these bounds are
application dependent, much like hyperparameters in machine learning. One learns these by exploring the underlying
database and running experiments. In the future, one can
envision a tool that can guide users to select an appropriate
numeric bound for a particular CI query.
5.
Figure 13: Steps in training a multi-modal database
with text and image fields
either hidden or exposed to the user. The training table is
then converted to the textual representation (Figure 13(C))
to build the text corpus for training the word embedding
model. Each sentence in the text corpus includes both original non-image (e.g., Corbett) and extracted image features
(e.g., Tiger, or Carnivorous). In the resulting multi-modal
word embedding model, the non-image features will contribute to the meaning of image features and vice versa,
and all meaning vectors will be uniformly represented using
vectors of dimension d = 200.
COGNITIVE QUERIES IN PRACTICE
We now illustrate some unique capabilities of our cognitive database system by discussing a scenario in which CI
queries are used to gain novel insights from a multi-modal relational database.3 In this scenario, we consider a database
of national parks across multiple countries, with links to images of animals in the associated national parks (Figure 13).
We use images from the open source Image database, ImageNet [16], to populate our database. We use this database
to present results from inductive reasoning CI queries using
our Spark 2.2.0 based prototype on an Intel Xeon E5-2680
system. Our prototype is implemented in Scala and supports
queries written in either Scala or Python using Spark SQL,
Spark Dataframes, or Python pandas SQLite interfaces. Although the database under evaluation is fairly simple, its architecture is similar to many other real life databases, e.g., a
multi-modal patient database with text fields describing patient characteristics and image fields referring to associated
images (e.g., radiology or FMRI images), or an insurance
claims database with text fields containing the claim information and image fields storing supporting pictures (e.g.,
car collision photos).
Figure 13(A) presents the original relational table as created by an user. It contains only text fields that list paths to
the images and provide additional information on every image. This table, other than the image path, does not provide
any details on the referred images. To create a shared word
embedding model from the text and images, we employ the
automatic tag generator approach outlined in Section 3.1.
Figure 14 presents the workflow of extracting image features from the referred images. Each image in the database
(e.g., a Lion) is first uploaded to the IBM Watson Visual
Recognition System (VRS) [30] for classification and text
description. The Watson system’s JSON response is then
parsed, and a set of text attributes for the input image is
extracted. These attributes form the features of the images
and are added to the original table to create a training version of the table (Figure 13(B)). This training table can be
List of Images
(JPEG format,
size < 2 MB)
[n00015388_39388.JPEG]
API call
IBM
Watson
Python
SDK
IBM Watson Visual
Recognition Service
[Default Classifier]
Raw JSON
Response
Offline
Parsing
"{\n\"images\": [\n {\n \"image\": \"./n00015388_39388.JPEG\", \n
\"classifiers\": [\n {\n \"classes\": [\n {\n \"score\": 0.969, \n
\"class\": \"lion\", \n\"type_hierarchy\":
\"/animal/mammal/carnivore/feline/big cat/lion\"\n}, \n {\n
\"score\": 0.978, \n \"class\": \"big cat\"\n}, \n {\n \"score\": 0.979,
\n \"class\": \"feline\"\n}, \n {\n \"score\": 0.98, \n \"class\":
\"carnivore\"\n }, \n {\n \"score\": 0.98, \n \"class\":
\"mammal\"\n}, \n {\n \"score\": 0.98, \n \"class\": \"animal\"\n}, \n
{\n \"score\": 0.942, \n \"class\": \"yellow color\"\n}, \n {\n
\"score\": 0.887, \n \"class\": \"pale yellow color\"\n}\n], \n
\"classifier_id\": \"default\", \n \"name\": \"default\"\n}\n]\n }\n],
\n \"custom_classes\": 0, \n \"images_processed\": 1\n}"
Text
Attribute
Database
of Images
[“n00015388_39388.JPEG”,
“animal”,
“mammal”,
“carnivore feline big_cat”,
“lion”,
“yellow pale_yellow”]
Figure 14: Illustration of workflow to generate a
Text Attribute Database of images using the IBM
Watson Visual Recognition Service (VRS) using a
sample image from Imagenet dataset
Once the meaning vectors are computed they can be used
to evaluate semantic relationships between values in the
original relational table. For example, if one were to compare National Parks, Serengati and Sunderbans would be
the most similar as they both share multiple image features.
It should be noted that one can not get this insight by using standard SQL queries as the original table does not have
any image information. Further, as many values in the training database are syntactically different (e.g., Crocodile and
Indian Gharial), existing SQL systems will fail to extract
any semantic similarities.
The first two examples of SQL CI queries over multimodal data assume that the users have access to the training database (Figure 13(B)). Figure 15 illustrates an analogy
3
Due to space limitations, we focus only on certain types of
queries.
9
Find all images whose classD satisfies the
analogy query [reptile : monitor_lizard ::
aquatic_vertebrate : ?] using analogyQuery
UDFhaving similarity score greater than 0.5.
Sort the results in descending order of their
similarity score.
Output Table
SELECT X.imagename,
X.classB, X.classC, X.classD,
analogyQuery(’reptile’,’monitor_lizard’,’
aquatic_vertebrate’,X.classD)
AS Score
FROM ImageDataTable X
WHERE
(analogyQuery(’reptile’,’monitor_lizard’,
’aquatic_vertebrate’, X.classD) > 0.5)
ORDER BY Score DESC
n02512053_1493
n02512053_6867
n02512830_2279
Figure 15: Analogy queries over images
n00015388_18458
query over images, while Figure 16 illustrates an analogy sequence query. In both cases, the CI queries are formulated
using UDFs, analogyQuery() and analogySequence(), that
take values from the training database as input. In case of
the analogy query, the goal is to find all images whose classD
feature (i.e. extracted name) has the same relationship to its
classC feature (i.e. class aquatic vertebrate) as the specified relationship, reptile::monitor lizard. For each row
in the table, the UDF first fetches meaning vectors for the
input parameters and uses the 3COSMUL approach to find
a relational value whose vector maximizes the analogy similarity score as defined in Equation 1. The SQL query returns
the corresponding images, whose similarity score is higher
than 0.5 and reports them in a descending order of similarity score. Figure 15 presents an output fragment of the CI
query and the corresponding images of spiny finned fish.
The analogy sequence query (Figure 16) uses a UDF that
operates on a sequence of analogy pairs that share the source
entity (e.g., mammal), but has different target entities, (e.g.,
jackal and mongoose). The UDF converts the analogy sequence problem into a traditional analogy problem by first
computing the average vector of the target entities and then
using it in the 3COSMULT approach. Figure 16 presents the
SQL CI query, the analogy sequence UDF, and its output
(images of grey fox).
Output
n01314663_1129
[grey_fox, kit_fox]
n01315805_3451
[grey_fox]
n02075612_7257
[grey_fox]
n01324431_4949
[silver_fox,kit_fox,grey_fox]
Input
n01316422_255
n01315581_997
n01604330_12473
andean_condor, condor
sloth_bear
n01316422_1684
glutton_wolverine
Find all images whose similarity to user
chosen images of [lion, vulture, shark]
using combinedAvgSim UDF is greater
than 0.75. Exclude the input images and
sort the result in descending order of
their similarity score.
SELECT X.imageName,
combinedAvgSim(X.imagename,
’n00015388_18458.jpeg’,
’n01316422_255.jpeg’,
’n01315581_997.jpeg’) AS SimScore
FROM ImageDataTable X WHERE
(X.imagename <> ’n00015388_18458.jpeg’) AND
(X.imagename <> ’n00015388_19237.jpeg’) AND
(X.imagename <> ’n00015388_18797.jpeg’) AND
(combinedAvgSim(X.imagename,
’n00015388_18458.jpeg’,
’n01316422_255.jpeg’,
’n01315581_997.jpeg’) > 0.75)
ORDER BY SimScore DESC
Output
n01324431_7056
andean_condor, tayra
Figure 17: Semantic clustering of images
Find all images whose classD satisfies
the analogy query [ (mammal : jackal),
(mammal : mongoose) :: mammal : ?]
having similarity score greater than
0.5. Filter out images of common
carnivorous animals and sort the
results in descending order of their
similarity score.
n02075612_340
[grey_fox]
are similar to every image in the set of user chosen images.
Such images share one or more features with the input set of
images. For this query, we select images of a lion, a vulture,
and a shark as the input set and use the combinedAvgSim()
UDF to identify images that are similar to all these three
images. Although the input images display animals from
three different classes, they share one common feature: all
three animals are carnivorous. The UDF computes the average vector from the three input images and then selects
those images whose vectors are similar to the computed average vector with similarity score higher than 0.75. Figure 17
shows the top three image results: andean condor, glutton
wolverine, and tyra. Although these animals are from different classes, they all are carnivores, a feature that is shared
with the animals from the input set.
SELECT X.imagename, analogySequence
(’mammal’, ’jackal’,’mammal’,
’mongoose’,’mammal’,X.ClassD)
AS AnalogyScore
FROM ImageDataTable X WHERE
(stringPresent(X.classD, ’tigress’) == 0) AND
….
stringPresent(X.classD, ’jackal’) == 0) AND
(stringPresent(X.classD, ’mongoose’) == 0) AND
(analogySequence(’mammal’, ’jackal’,
’mammal’,’mongoose’,’mammal’,X.classD,) 0.5)
ORDER BY AnalogyScore DESC
Figure 16: Analogy sequence queries over images
The next example (Figure 17) illustrates execution of a semantic clustering query on the original multi-modal database
table. The goal of this query is to identify all images that
The final example demonstrates the use of an external
semantic model for querying a multi-modal database. In
this scenario, we first train a word embedding model from
an external knowledge base derived from Wikipedia. Similar to the model trained from the database, the external
model assigns d dimensional meaning vectors to unique tokens (for the external model, we use d = 200). From the
wikipedia model, we select a token associated with a concept
Hypercarnivore, which refers to a class of animals whose
diet has more than 70% meat. Examples of hypercarnivores include lions, sharks, polar bears, crocodiles, hyenas,
etc. Therefore, in our model, the Hypercarnivore meaning vector is related to meaning vectors of tokens shark,
crocodiles, etc. For this query, we employ this externally
trained model to extract images that are similar to the concept Hypercarnivore. The UDF proximityAvgForExtKB()
uses the external model, finds images from the database
whose classD features (i.e., names) are related to
Hypercarnivore, and returns those images whose similarity score is higher than 0.5. Figure 18 shows the CI query
and its result: pictures of hyenas, who are members of the
hypercarnivore class4 . This example also demonstrates the
unique capability of cognitive databases that allows querying a database using a token not present in the database.
In our case, both the original and training databases do not
contain the token Hypercarnivore.
4
one of the images is a picture of a big dog which has been
mislabelled as an hyena by Watson VRS.
10
Output
n01316422_10446
[hyena, spotted_hyena]
n01317541_6440
[hyena, water_dog_dog,
spotted_hyena]
n01321579_5386
n01316422_10406
[hyena, spotted_hyena] [hyena, spotted_hyena]
n02075612_2316
[hyena, pouched_mammal,
spotted_hyena, bear_cub]
can be converted into a single matrix-vector multiplication operation. This can be generalized to a matrixmatrix multiplication operation to enable distance computations between two sets of vectors.
Find all images of animals whose
classD similarity score to the Concept
of ‘‘Hypercarnivore" of Wikipedia
using proximityAvgExtKB UDF is
greater than 0.5. Exclude images that
are already tagged as carnivore,
herbivore, omnivore or scavenger. Sort
the results in descending order of their
similarity score.
• Reducing redundant computations: For a given vector,
we can first identify a candidate set of vectors that
are spatially closer to it in the d dimensional vector
space using either locality sensitive hashing (LSH) [9],
or clustering via the Spherical K-Means algorithm [17].
We can then invoke distance calculations on the candidate set to compute precise distances. In the LSH
approach, the d dimension vector locations are mapped
to bit-vector signatures of length d via projecting them
on random planes. For a given vector, spatially closer
vectors can be identified by choosing those with small
hamming distance (1 or 2) between their corresponding signatures. In the K-Means approach, one can use
the centroid information to identify a candidate set of
spatially close vectors. Both approaches can also be
accelerated using either SIMD functions or GPUs [7].
SELECT X.imagename, proximityAvgExtKB
(’CONCEPT_Hypercarnivore’, X.classD)
AS SimScore
FROM ImageDataTable X WHERE
…….
…….
…….
( (stringPresent(X.classD, ’scavenger’) == 0) AND
(proximityAvgForExtKB
(’CONCEPT_Hypercarnivore’,X.classD) > 0.5)
ORDER BY SimScore DESC
Figure 18: Similarity query over images using
a model trained from external knowledge base
(Wikipedia)
Although we used hypothetical scenarios to demonstrate
our ideas, CI queries are applicable to a broad class of domains. These include finance, insurance, retail, customer
care, log analytics, healthcare, genomics, semantic search
over documents (patent, legal, or financial), healthcare informatics, and human resource management.
These examples demonstrate several unique capabilities
of cognitive database systems, namely: (1) ability to build
joint cross-modal semantic model from multi-modal data,
(2) transperent integration of novel cognitive queries into
existing SQL query infrastructure, (3) using untyped vector representations to support contextual semantic (i.e. not
value based) queries across multiple SQL data types, and
(4) ability to import an externally trained semantic model
and apply it to enable querying a database using a token not
present in the database. To the best of our knowledge, none
of the current industrial, academic, or open source database
systems support these capabilities.
5.1
• Using relational query optimization approaches: One
can also reduce redundant computations by using relational view that pre-selects rows from a table based
on certain criteria. In addition, once the candidate
set of SQL variables is known, the SQL query engine
can pre-compute the pair-wise distances and use the
cached results during execution of UDFs later.
• Using hardware acceleration and parallelization: The
core nearest-neighbor distance computations, namely,
dotproduct, matrix-vector, and matrix-matrix computations can be accelerated via hardware accelerators
such as on-chip SIMD or using GPUs [6]. Most numerical libraries such as MKL, ESSL, or OpenBLAS provide hardware accelerated matrix computation kernels.
Further, the nearest neighbor computations can be
also parallelized either using CPU-based multithreading (e.g., using pthreads) or distributing it over a cluster of machines using a distributed infrastructure such
as the Apache Spark.
Optimizing Cognitive Intelligence Queries
Cognitive Intelligence queries are standard SQL analytics
queries which invoke UDFs to enable contextual semantic
operations between relational variables. Irrespective of the
kind of CI Query, core UDF computation involves computing pair-wise similarities between vectors, which can be then
used to identify nearest or furthest neighbors of a vector. In
the worst case, a CI query can invoke a UDF for every row
combination being evaluated and the UDF, in turn, can operate on a large number of vectors. Since in practice, the
number of row combinations can very high, it is critical to
optimize the performance of distance computations in the
nearest-neighbor calculations for CI queries.
In a d dimensional vector space, pair-wise distance between vectors v1 and v2 is calculated by computing the co·v2
. As we use normalized
sine distance, cos(v1 , v2 ) = kvv11kkv
2k
forms of vectors (i.e. kvk = 1), the pair-wise distance calculation gets simplified to a vector dotproduct, v1 · v2 . In
general, there are four basic ways of optimizing distance
computations:
Optimizations of the cognitive intelligence queries is an
open problem and is the current focus of our activities.
6.
RELATED WORK
Language Embedding: Over the past few years, a number of methods have been introduced for obtaining a vector
representation of words in a language [5], called language
embedding. The methods range from brute force learning
by various types of neural networks (NNs) [5], to log-linear
classifiers [43] and to various matrix formulations, such as
matrix factorization techniques [36]. Lately, Word2Vec [42,
45, 44, 35] has gained prominence as the vectors it produces appear to capture syntactic as well semantic properties of words. The exact mechanism employed by Word2Vec
and suggestions for alternatives are the subject of much research [46, 19, 35]. Although Word2Vec has gained much
prominence it is one of many possible methods for generating word representing vectors. For example, GloVe [51]
also builds word embeddings by a function optimization approach over the word co-occurrence matrix. Vectors may be
associated with larger bodies of text such as paragraphs and
• Increasing computation granularity: In cases where
one needs to compute the distance of a vector from
many vectors, many individual dotproduct operations
11
even documents. Applications to the paragraph and document embedding appear in [34, 15]. Recent work has also
been exploring applying word embeddings to capture image
semantics [57].
Applications of Word Embedding: The word embedding model is being used for a wide variety of applications
beyond NLP. Wu et al [64] provide a general neural framework for building vector embaddings of entities of different types into a vectorial embedding space. The common
representation can be then used for different tasks such as
text classification, link prediction, document recommendation, etc. The YouTube recommendation system uses embedding approaches to capture user behavior [13]. Similar embedding-based approaches have been used for recommending news articles [48], for discovering topics [65], or for
personalized fashion shopping [47, 2]. Hope et. al. have proposed using word embedding for supporting analogy queries
over knowledge bases such as the US Patent database [29].
DeepWalk [52] and Node2Vec [23] have proposed using word
embedding approaches for learning neighborhood features
of nodes in a network graph. Word embedding approaches
are also being used for a variety of semantic web applications, e.g., embedding RDF triples [53, 12] and encoding
geo-spatial proximity [33]. Using latent feature vectors for
data integration in knowledge databases has been explored
in [61, 60].
Relational Databases: In the context of SQL, text capabilities, e.g., the use of synonyms, have been in practice for
a while [14]. In the literature, techniques for detecting similarity between records and fields have also been explored.
Semantic similarity between database records is explored
in [32]. Phrase-based ranking by applying an IR approach
to relational data appears in [40]. Indexing and searching relational data by modeling tuples as virtual documents
appear in [41]. Effective keyword-based selection of relational databases is explored in [66]. A system for detecting
XML similarity, in content and structure, using a relational
database is described in [62]. Related work on similarity
Join appears in [10]. Semantic Queries are described in [49].
Most recently, Shin et al. have described DeepDive [56] that
uses machine learning techniques, e.g., Markov Logic based
rules, to convert input unstructured documents into a structured knowledge base.
The proposed cognitive database system can be distinguished by the following unique features: (1) Encoding relational data using word embedding techniques, (2) Using semantic vectors to enable a new class of SQL analytics queries
(CI queries), (3) Ability to make contextual semantic matching, unlike the traditional value (syntactical) matching supported by current SQL queries, (4) Capturing relationalships
across multiple data types, including images, and (5) Ability of using external knowledge bases. Further, semantic
vectors are primarily based on the database itself (with external text or vectors as an option). This means that we assume no reliance on dictionaries, thesauri, word nets and the
like. Once these vectors are generated they may be used in
vastly enriching the querying expressiveness of virtually any
query language. These capabilities go far beyond analytical
capabilities present in current relational systems. All wellknown commercial and open source (e.g., Apache Spark [1],
MADlib [25]) database systems have built-in analytics capabilities, e.g., Spark MLLib. Apache Spark can also create
a deep-learning pipeline in which it can invoke an external
deep-learning infrastructure e.g., TensorFlow [21] to train a
model, and then load the trained model to perform inferencing operations [28]. However, such systems view databases
as repositories for storing input features and results for the
analytics or deep-learning frameworks. On the other hand,
cognitive databases use the word embedding model to extract features from the database entities and use them to enhance its querying capabilities. Systems based on statistical
relational learning models combine probabilistic graphical
models and first-order logic to encode uncertain first-order
logic rules based on known information [63]. In contrast, a
cognitive database learns information about the relational
data which is not known apriori.
7.
CONCLUSIONS AND SUCCESS CRITERIA
In this paper we presented Cognitive Database, an innovative relational database system that uses the power
of word embedding models to enable novel AI capabilities
in database systems. The word embedding approach uses
unsupervised learning to generate meaning vectors using
database-derived text. These vectors capture syntactic as
well as semantic characteristics of every database token. We
use these vectors to enhance database querying capabilities.
In essence, these vectors provide another way to look at
the database, almost orthogonal to the structured relational
regime, as vectors enable a dual view of the data: relational
and meaningful text. We thereby introduce and explore a
new class of queries called cognitive intelligence (CI) queries
that extract information from the database based, in part,
on the relationships encoded by these vectors.
We are implementing a prototype system on top of Apache
Spark [1] to exhibit the power of CI queries. Our current
infrastructure enables complex SQL-based semantic queries
over multi-modal databases (e.g., inductive reasoning queries
over a text and image database). We are now working on
accelerating model training and nearest neighbor computations using a variety of approaches (e.g., using GPUs),
and developing new techniques for incremental vector training. We believe CI queries are applicable to a broad class of
application domains including healthcare, bio-informatics,
document searching, retail analysis, and data integration.
7.1
Success Criteria
We believe Cognitive Databases are truly a new technology for incorporating AI capabilities into relational databases.
As such it holds a great promise for innovative applications
with a very different view of data as compared to today’s
database systems. Since it is based on a new concept, there
are no easy comparisons. For example, there are no relevant benchmarks except for ones used in NLP for testing
language features such as analogies [45]. So, success of the
cognitive databases will be mainly evaluated based on new
applications within known domains (such as Retail), new
domains of applications (such as medical drugs selection)
that are currently in the sphere of AI-based systems, and
the adaptation of CI capabilities as a standard feature by
leading vendor as well as open source database systems. Finally, we hope this work spurs new research initiatives in this
exciting emerging area in database management systems.
12
8.
REFERENCES
1994.
[19] Y. Goldberg and O. Levy. word2vec explained:
deriving mikolov et al.’s negative-sampling
word-embedding method. CoRR, abs/1402.3722, 2014.
[20] I. Goodfellow, Y. Bengio, and A. Courville. Deep
Learning. The MIT Press, 2016.
[21] Google Inc. TensorFlow: An open-source software
library for Machine Intelligence.
[22] J. Gray, S. Chaudhuri, A. Bosworth, A. Layman,
D. Reichart, M. Venkatrao, F. Pellow, and
H. Pirahesh. Data cube: A relational aggregation
operator generalizing group-by, cross-tab, and
sub-totals. Data Mining and Knowledge Discovery,
1(1):29–53, 1997.
[23] A. Grover and J. Leskovec. Node2vec: Scalable feature
learning for networks. In Proceedings of the 22nd ACM
SIGKDD International Conference on Knowledge
Discovery and Data Mining, KDD ’16, pages 855–864,
2016.
[24] Z. S. Harris. Distributional structure. Word,
10(2-3):146–162, 1954.
[25] J. M. Hellerstein, C. Re, F. Schoppmann, D. Z. Wang,
E. Fratkin, A. Gorajek, K. S. Ng, C. Welton, X. Feng,
K. Li, and A. Kumar. The MADlib analytics library:
or MAD skills, the SQL. Proc. VLDB Endow., 5(12),
August 2012.
[26] G. E. Hinton. Learning distributed representations of
concepts. In Proceedings of the Eighth Annual
Conference of Cognitive Science Society, pages 1–12,
1986.
[27] C.-T. Ho, R. Agrawal, N. Megiddo, and R. Srikant.
Range queries in olap data cubes. In Proceedings of
the 1997 ACM SIGMOD International Conference on
Management of Data, pages 73–88, 1997.
[28] S. A. Hong, T. Hunter, and R. Xin. A vision for
making deep learning simple, June 2017. DataBricks
Engineering Blog.
[29] T. Hope, J. Chan, A. Kittur, and D. Shahaf.
Accelerating innovation through analogy mining. In
Proceedings of the 23rd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining,
KDD ’17, pages 235–243, 2017.
[30] IBM Watson. Watson visual recognition service.
www.ibm.com/watson/services/visual-recognition/.
[31] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev,
J. Long, R. Girshick, S. Guadarrama, and T. Darrell.
Caffe: Convolutional architecture for fast feature
embedding. arXiv preprint arXiv:1408.5093, 2014.
[32] V. Kashyap and A. P. Sheth. Semantic and schematic
similarities between database objects: A context-based
approach. VLDB J., 5(4):276–304, 1996.
[33] M. Kejriwal and P. Szekely. Neural embeddings for
populated geonames locations. In Proceedings of the
16th International Semantic Web Conference, 2017.
[34] Q. V. Le and T. Mikolov. Distributed representations
of sentences and documents. CoRR, abs/1405.4053,
2014.
[35] O. Levy and Y. Goldberg. Linguistic regularities in
sparse and explicit word representations. In
Proceedings of the Eighteenth Conference on
Computational Natural Language Learning, CoNLL
[1] Apache Foundation. Apache spark: A fast and general
engine for large-scale data processing, 2017. Release
2.2.
[2] S. Arora and D. Warrier. Decoding fashion contexts
using word embeddings. In KDD Workshop on
Machine learning meets fashion, 2016.
[3] D. Bahdanau, K. Cho, and Y. Bengio. Neural machine
translation by jointly learning to align and translate.
In Proceedings of the ICLR 2015, 2015.
[4] M. Baroni, G. Dinu, and G. Kruszewski. Dont count,
predict! a systematic comparison of context-counting
vs. context-predicting semantic vectors. In Proceedings
of The 2014 Conference of the Association for
Computational Linguistics, 2014.
[5] Y. Bengio, R. Ducharme, P. Vincent, and C. Janvin.
A neural probabilistic language model. Journal of
Machine Learning Research, 3:1137–1155, 2003.
[6] R. Bordawekar, B. Blainey, and R. Puri. Analyzing
Analytics. Morgan & Claypool Publishers, November
2015. Synthesis Lectures on Computer Architecture.
[7] R. Bordawekar and P. D’Souza. Optimizing
Out-of-core Nearest Neighbor Problems on
Multi-GPU Systems using NVLink. Nvidia Global
Technical Conference (GTC), May 2017.
[8] R. Bordawekar and O. Shmueli. Enabling cognitive
intelligence queries in relational databases using
low-dimensional word embeddings. CoRR,
abs/1603.07185, March 2016.
[9] M. S. Charikar. Similarity estimation techniques from
rounding algorithms. In Proceedings of the
Thiry-fourth Annual ACM Symposium on Theory of
Computing, pages 380–388, 2002.
[10] S. Chaudhuri, V. Ganti, and R. Kaushik. A primitive
operator for similarity joins in data cleaning. In
Proceedings of the 22nd International Conference on
Data Engineering, Washington, DC, USA, 2006. IEEE
Computer Society.
[11] W. W. Chu, H. Yang, and G. Chow. A cooperative
database system (cobase) for query relaxation. In
Proceedings of ARPI 1996, 1996.
[12] M. Cochez, P. Ristoski, S. P. Ponzetto, and
H. Paulheim. Global rdf vector space embeddings. In
Proceedings of the 16th International Semantic Web
Conference, 2017.
[13] P. Covington, J. Adams, and E. Sargin. Deep neural
networks for youtube recommendations. In
Proceedings of the 10th ACM Conference on
Recommender Systems, New York, NY, USA, 2016.
[14] R. Cutlip and J. Medicke. Integrated Solutions with
DB2. IBM Press, 2003.
[15] A. M. Dai, C. Olah, and Q. V. Le. Document
embedding with paragraph vectors. CoRR,
abs/1507.07998, 2015.
[16] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and
L. Fei-Fei. ImageNet: A Large-Scale Hierarchical
Image Database. In CVPR09, 2009.
[17] I. S. Dhillon and D. S. Modha. Concept
decompositions for large sparse text data using
clustering. Mach. Learn., 42(1-2):143–175, Jan. 2001.
[18] M. N. Diascro and N. Brody. Odd-man-out and
intelligence. Intelligence, 19(1):79–92, July-August
13
2014, pages 171–180, 2014.
[36] O. Levy and Y. Goldberg. Neural word embedding as
implicit matrix factorization. In Annual Conference on
Neural Information Processing Systems 2014, pages
2177–2185, 2014.
[37] F. Li and H. V. Jagadish. Constructing an interactive
natural language interface for relational databases.
Proc. VLDB Endow., 8(1):73–84, Sept. 2014.
[38] L. Lim, H. Wang, and M. Wang. Semantic queries by
example. In Proceedings of the 16th International
Conference on Extending Database Technology (EDBT
2013), 2013.
[39] T. Linzen. Issues in evaluating semantic spaces using
word analogies. arXiv preprint arXiv:1606.07736,
2016.
[40] F. Liu, C. T. Yu, W. Meng, and A. Chowdhury.
Effective keyword search in relational databases. In
Proceedings of the ACM SIGMOD International
Conference on Management of Data, Chicago, Illinois,
USA, June 27-29, 2006, pages 563–574, 2006.
[41] Y. Luo, X. Lin, W. Wang, and X. Zhou. Spark: top-k
keyword query in relational databases. In Proceedings
of the ACM SIGMOD International Conference on
Management of Data, Beijing, China, June 12-14,
2007, pages 115–126, 2007.
[42] T. Mikolov. word2vec: Tool for computing continuous
distributed representations of words, 2013.
github.com/tmikolov/word2vec.
[43] T. Mikolov, K. Chen, G. Corrado, and J. Dean.
Efficient estimation of word representations in vector
space. CoRR, abs/1301.3781, 2013.
[44] T. Mikolov, Q. V. Le, and I. Sutskever. Exploiting
similarities among languages for machine translation.
CoRR, abs/1309.4168, 2013.
[45] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and
J. Dean. Distributed representations of words and
phrases and their compositionality. In 27th Annual
Conference on Neural Information Processing Systems
2013., pages 3111–3119, 2013.
[46] A. Mnih and K. Kavukcuoglu. Learning word
embeddings efficiently with noise-contrastive
estimation. In 27th Annual Conference on Neural
Information Processing Systems 2013., pages
2265–2273, 2013.
[47] C. Moody. A word is worth a thousand vectors. Stitch
Fix Multithreaded Blog.
[48] S. Okura, Y. Tagami, S. Ono, and A. Tajima.
Embedding-based news recommendation for millions
of users. In Proceedings of the 23rd ACM SIGKDD
International Conference on Knowledge Discovery and
Data Mining, KDD ’17, pages 1933–1942, 2017.
[49] Z. Pan and J. Heflin. DLDB: extending relational
databases to support semantic web queries. In PSSS1
- Practical and Scalable Semantic Systems, 2003.
[50] Y. Park, A. S. Tajik, M. Cafarella, and B. Mozafari.
Database learning: Toward a database that becomes
smarter every time. In Proceedings of the 2017 ACM
International Conference on Management of Data,
pages 587–602, 2017.
[51] J. Pennington, R. Socher, and C. D. Manning. GloVe:
Global vectors for word representation. In Proceedings
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
[60]
[61]
[62]
[63]
[64]
[65]
[66]
14
of the 2014 Conference on Empirical Methods in
Natural Language Processing, pages 1532–1543, 2014.
B. Perozzi, R. Al-Rfou, and S. Skiena. Deepwalk:
Online learning of social representations. In
Proceedings of the 20th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining,
KDD ’14, pages 701–710, 2014.
P. Ristoski and H. Paulheim. Rdf2vec: Rdf graph
embedding for data mining. In Proceedings of the 15th
International Semantic Web Conference, pages
498–514, 2016.
D. E. Rumelhart and A. A. Abrahamson. A model for
analogical reasoning. Cognitive Psychology, 5(1):1 –
28, 1973.
G. Salton, A. Wong, and C. S. Yang. A vector space
model for automatic indexing. Communications of the
ACM, 18(11):613–620, 1975.
J. Shin, S. Wu, F. Wang, C. De Sa, C. Zhang, and
C. Ré. Incremental knowledge base construction using
deepdive. Proc. VLDB Endow., 8(11), July 2015.
R. Socher, M. Ganjoo, H. Sridhar, O. Bastani, C. D.
Manning, and A. Y. Ng. Zero-shot learning through
cross-modal transfer. CoRR, abs/1301.3666, 2013.
R. J. Sternberg and M. K. Gardner. Unities in
inductive reasoning. Technical Report Technical rept.
no. 18, 1 Jul-30 Sep 79, Yale University, 1979.
D. Van Aken, A. Pavlo, G. J. Gordon, and B. Zhang.
Automatic database management system tuning
through large-scale machine learning. In Proceedings
of the 2017 ACM International Conference on
Management of Data, pages 1009–1024, 2017.
P. Verga, D. Belanger, E. Strubell, B. Roth, and
A. McCallum. Multilingual relation extraction using
compositional universal schema. CoRR,
abs/1511.06396, 2015.
L. Vilnis and A. McCallum. Word representations via
gaussian embedding. CoRR, abs/1412.6623, 2014.
W. Viyanon and S. K. Madria. A system for detecting
xml similarity in content and structure using
relational database. In Proceedings of the 18th ACM
Conference on Information and Knowledge
Management, CIKM, pages 1197–1206, 2009.
D. Z. Wang, Y. Chen, C. Grant, and K. Li. Efficient
in-database analytics with graphical models. IEEE
Data Engineering Bulletin, 2014.
L. Wu, A. Fisch, S. Chopra, K. Adams, A. Bordes,
and J. Weston. Starspace: Embed all the things!
CoRR, abs/1709.03856, 2017.
G. Xun, Y. Li, J. Gao, and A. Zhang. Collaboratively
improving topic discovery and word embeddings by
coordinating global and local contexts. In Proceedings
of the 23rd ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining, KDD ’17,
pages 535–543, 2017.
B. Yu, G. Li, K. R. Sollins, and A. K. H. Tung.
Effective keyword-based selection of relational
databases. In Proceedings of the ACM SIGMOD
International Conference on Management of Data,
Beijing, China, June 12-14, 2007, pages 139–150,
2007.
| 2cs.AI
|
1
GPI-based Secrecy Rate Maximization
Beamforming Scheme for Wireless Transmission
with AN-aided Directional Modulation
arXiv:1711.06463v2 [cs.IT] 29 Jan 2018
Hai Yu, Simin Wan, Wenlong Cai, Ling Xu, Xiaobo Zhou, Jin Wang, Yongpeng Wu,
Feng Shu, Jiangzhou Wang, and Jianxin Wang
Abstract—In a directional modulation network, a general
power iterative (GPI) based beamforming scheme is proposed to
maximize the secrecy rate (SR), where there are two optimization
variables required to be optimized. The first one is the useful
precoding vector of transmitting confidential messages to the
desired user while the second one is the artificial noise (AN)
projection matrix of forcing more AN to eavesdroppers. In such
a secure network, the paramount problem is how to design or
optimize the two optimization variables by different criteria. To
maximize the SR (Max-SR), an alternatively iterative structure
(AIS) is established between the AN projection matrix and the
precoding vector for confidential messages. To choose a good
initial value of iteration process of GPI, the proposed Max-SR
method can readily double its convergence speed compared to the
random choice of initial value. With only four iterations, it may
rapidly converge to its rate ceil. From simulation results, it follows
that the SR performance of the proposed AIS of GPI-based MaxSR is much better than those of conventional leakage-based and
null-space projection methods in the medium and large signalto-noise ratio (SNR) regions, and its achievable SR performance
gain gradually increases as SNR increases.
Index Terms—Secrecy rate, artificial noise, directional modulation, general power iterative, alternatively iterative structure.
I. I NTRODUCTION
In the recent decade, physical-layer security in wireless
networks, as a new tool to provide an incremental safeguard
of confidential message over conventional cryptography, has
drawn tremendous research attention and interests from both
academia and industry [1]–[11]. In [1], the author’s seminal
research work has established the channel model and found
This work was supported in part by the National Natural Science Foundation of China (Grant Nos. 61771244, 61702258, 61472190, 61501238 and
61602245), the China Postdoctoral Science Foundation (2016M591852), the
Postdoctoral research funding program of Jiangsu Province(1601257C), the
Natural Science Foundation of Jiangsu Province (Grants No. BK20150791),
and the open research fund of National Mobile Communications Research
Laboratory, Southeast University, China (No.2013D02).
Hai Yu, Simin Wang, Ling Xu, Feng Shu, Xiaobo Zhou, Jin Wang, and
Jianxin Wang are with School of Electronic and Optical Engineering, Nanjing
University of Science and Technology, Nanjing, 210094, China.
Wenlong Cai is with the faculty of National Key Laboratory of Science
and Technology on Aerospace Intelligence Control and Beijing Aerospace
Automatic Control Institute, Beijing, China. E-mail: [email protected]
Yongpeng Wu is with the Shanghai Key Laboratory of Navigation and
Location-Based Services, Shanghai Jiao Tong University, Minhang 200240,
China. E-mail: [email protected]
Feng Shu is also with the College of Computer and Information Sciences,
Fujian Agriculture and Forestry University, Fuzhou 350002, China.
Jiangzhou Wang is with the School of Engineering and Digital Arts,
University of Kent, Canterbury CT2 7NT, U.K. E-mail: [email protected]
the tradeoff curve between the transmission rate and the data
equivocation seen by the wire-tapper. More importantly, the
author also proved that reliable transmission at rates up to
C, is possible in approximately perfect secrecy. As a physical
layer secure transmit technique suitable for line-of-propagation
(LoP) scenario, directional modulation (DM) has made great
progresses in many aspects with the aid of artificial noise
(AN) and antenna array beamforming [12]–[19]. To enhance
security, the symbol-level precoder in [13] was presented by
using the concept of constructive interference in directional
modulation with the goal of reducing the energy consumption
at transmitter. In the presence of direction measurement error,
the authors in [17], [18] designed three new robust DM
synthesis methods for three different application scenarios:
single-desired user, multi-desired user broadcasting, multidesired user MIMO by fully making use of the property of direction measurement error. Compared with conventional nonrobust methods, like null-space projection (NSP) [16], [20], the
proposed robust methods actually harvests an appealing rate
gain and almost an order-of-magnitude bit error rate (BER)
performance improvement along desired directions. The works
above only present an investigation on conventional DM with
only direction-dependent property. However, if eavesdropper
lies on the same direction from as the desired user, and its
distance from the DM transmitter is different from the distance
from the desired user to the DM transmitter, then it can still
intercept the confidential messages successfully. This is the
existing intrinsic secure problem for the conventional DM
networks. To address the serious secure issue, recently, the
authors in [21], [22] proposed an original concept of secure
and precise wireless transmission. In their works, the confidential messages are transmitted to the desired position precisely
and securely. If eavesdropper is outside the small area around
the desired position and locates in the same direction as the
desired direction, it can not intercept the confidential messages
successfully due to AN corruption and frequency random
property. However, in this paper, we focus still on how to
maximize the SR for conventional DM networks via GPI
algorithm.
However, in a three-node directional modulation network,
what is the maximum achievable secrecy rate in the absence of
direction estimation error? This is an open NP-hard problem.
In the following, we will solve the problem via a combination
of general power iterative (GPI) method [23] and alternatively
iterative structure (AIS). From simulation results, we find that
2
the proposed GPI-based AIS can achieve an obvious secrecy
rate (SR) performance gain compared to conventional NSP and
leakage-based methods in the medium and large SNR regions.
Our main contributions are summarized as follows:
1) In line-of-sight (LoS) scenarios, such as mmWave communications (massive MIMO), internet of things (IoT),
unmanned aerial vehicle (UAV) and satellite communications, we propose a general power iterative (GPI)
scheme in DM system to maximize secrecy rate (SR),
which is shown to be much better than conventional NSP
and leakage-based methods in terms of SR.
2) To accelerate the convergence rate of our proposed
method, the leakage-based precoding vector and AN
projection matrix are chosen to be the initial values
instead of random generating. From simulation results, it
follows that the new initial values can greatly reduce the
number of iterations, and doubles the convergence rate
compared to the random ones. This means that a good
choice of initial values leads to a dramatic reduction in
computational amount of the proposed method.
The remainder is organized as follows: Section II describes
the system model. An AIS of maximizing secrecy rate (MaxSR) based on GPI with aided AN is proposed in Section
III. Section IV presents the simulation results and complexity
analysis. Finally, we draw conclusions in Section V.
Notations: throughout the paper, matrices, vectors, and
scalars are denoted by letters of bold upper case, bold lower
case, and lower case, respectively. Signs (·)T , (·)∗ and (·)H
denote transpose, conjugate, and conjugate transpose, respectively. Notation E{·} stands for the expectation operation.
Matrices IN denotes the N × N identity matrix and 0M×N
denotes M × N matrix of all zeros. tr(·) denotes matrix trace.
Operation ⊗ denotes the Kronecker product of two matrices
[24].
II. S YSTEM M ODEL
Fig. 1 sketches a diagram block of directional modulation
(DM) network consisting of one base station (BS, Alice)
equipped with N antennas, one desired node (Bob) and one
eavesdropping node (Eve). Here, it is assumed that both
desired node and eavesdropping node are employed with single
antenna and BS employs an N -element uniformly spaced
linear array. Due to its directional property, DM usually
works in LoP channel. By introducing the precoding vector of
confidential messages and AN projection matrix at transmitter,
the transmit baseband signal from antenna array of BS is of
the form
p
p
s = β1 Ps vd x + αβ2 Ps PAN z,
(1)
H
where x is the confidential message with E x x = 1
and z ∈ CN ×1 denotes the AN vector of obeying complex
Gaussian distribution CN (0, IN −1 ). Ps denotes the total transmit power, β1 and β2 stand for the power allocation (PA)
factors of confidential message and AN with β12 + β22 = 1.
vd ∈ CN ×1 denotes the transmit beamforming vector to align
the confidential message to desired direction, vdH vd = 1.
Bob
Eve
h(q d )
h(q e )
Alice
Fig. 1. Directional modulation network
PAN ∈ CN ×(N −1) is the projection matrix, α normalizes
PAN z such that α2 E{Tr[PAN zzH PH
AN ]} = 1.
After experiencing the LoP channel, the received signal
along direction θ is given by
y(θ) = hH (θ)s + nr
p
p
= β1 Ps hH (θ)vd x + αβ2 Ps hH (θ)PAN z + nr ,
(2)
where nr is additive white Gaussian noise (AWGN) with
nr ∼ CN (0, σr2 ), h(θ) ∈ CN ×1 is
√ the normalized steerN −1 [exp (j2πΨθ (1)) ,
ing vector defined by h(θ) =
T
· · · , exp (j2πΨθ (N ))] , and the phase function Ψθ (n)
along direction θ is denoted as Ψθ (n) , −[n − (N +
1)/2]d cos θλ−1 , (n = 1, 2, · · · , N ), where n indexes the
elements of transmit antenna array, d denotes the spacing of
two adjacent antennas, and λ is the wavelength of the carrier.
The receive signals at Bob and Eve are written as
p
p
y(θd ) = β1 Ps hH (θd )vd x + αβ2 Ps hH (θd )PAN z + nd ,
(3)
and
y(θe ) = β1
p
p
Ps hH (θe )vd x + αβ2 Ps hH (θe )PAN z + ne ,
(4)
where nd ∼ CN (0, σd2 ) and ne ∼ CN (0, σe2 ) represent the
noise at Bob and Eve. With the help of some classic estimation
algorithms such as Capon’s method, MUSIC and ESPRIT, the
DM transmitter in Fig. 1 can obtain the direction knowledge
of eavesdroppers and desired users. Without loss of generality,
it is assumed that the variances of σr2 , σd2 , and σre are equal,
i.e., σr2 = σd2 = σe2 = σ 2 .
To derive the SR, the achievable rates of Bob and Eve,
R(θd ) and R(θe ) are defined as follows
β12 Ps hH (θd )vd vdH h(θd )
R(θd ) = log2 1 + 2
,
σd + α2 β22 Ps hH (θd )PAN PH
AN h(θd )
(5)
and
R(θe ) = log2 1 +
β12 Ps hH (θe )vd vdH h(θe )
σe2 + α2 β22 Ps hH (θe )PAN PH
AN h(θe )
,
(6)
Then, the SR Rs (vd , PAN ) is defined as follows
Rs (vd , PAN ) = max{0, R(θd) − R(θe )}
H
vd (Hd + Ad IN )vd
= max{0, log2
×
B
},
vdH (He + Ae IN )vd
(7)
3
H
where Hd , hd hH
d , H e , he he ,
2
2
−1
Ad = α2 β22 β1−2 hH (θd )PAN PH
, (8)
AN h(θd ) + σ (β1 Ps )
2 2
−1
Ae = α2 β22 β1−2 hH (θe )PAN PH
, (9)
AN h(θe ) + σ (β1 Ps )
and
B=
2
2 2
−1
hH (θe )PAN PH
AN h(θe ) + σ (α β2 Ps )
.
H
2
H
2
2
h (θd )PAN PAN h(θd ) + σ (α β2 Ps )−1
(10)
The RS defined in (7) should be larger than or equal to zero.
If it is less than zero, then the eavesdropper will get more
mutual information than the desired user. This is not desired
in a secure system. However, the second term on right hand
side of (7) can be less than zero sometimes, i.e.,
log2
vdH (Hd + Ad IN )vd
×B
vdH (He + Ae IN )vd
<0
(11)
if the eavesdropper is closer to the DM transmitter than the
desired user. Thus, the operation max in RS given by (7)
should be kept such that the SR is larger than or equal to
zero. In what follows, we will maximize the SR by optimizing
the beamforming vector vd of confidential messages and the
projection matrix PAN of AN, which can be casted as
(P1) : max
vd ,PAN
Rs (vd , PAN )
(12)
−2
subject to vdH vd = 1, Tr[PAN PH
.
AN ] = α
Obviously, the above optimization problem is NP-hard, and
it is hard to be solved directly or find a closed-from solution
to the above problem.The objective function in (12) as shown
in (7) is divided into a sum of two non-convex functions of
variables vd , and PAN . This means that it is still a non-convex
function. Thus, we make a conclusion that the optimization
problem in (12) is a nonlinear non-convex optimization.
III. P ROPOSED I TERATIVE ALGORITHM TO DESIGN
vd AND PROJECTION MATRIX PAN
BEAMFORMING VECTOR
Since the joint optimization vd and PAN in (12) is too complicated, we divided it into two mutual coupling subproblems
and construct an iterative structure between them. By fixing
the beamforming vector vd , the optimal PAN is solved by
utilizing the GPI algorithm [23]. For given projection matrix
PAN , the optimal vd is derived according to generalized
Rayleigh-Ritz ratio in [24]. This forms an alternatively iterative structure. The Max-SR method outputs the beamforming
vector vd and projection matrix PAN by repeatedly applying
the GPI in [23].
Tr(PH
Tr(PH
AN Ce PAN )
AN Bd PAN )
×
Tr(PH Be PAN ) Tr(PH
AN Cd PAN )
H AN
H
w (IN −1 ⊗ Ce )w
w (IN −1 ⊗ Bd )w
,
×
= log2
wH (IN −1 ⊗ Be )w wH (IN −1 ⊗ Cd )w
(14)
Rs = log2
where w , vec(PAN ) ∈ CN (N −1)×1 ,
Bd =
β22
σ2
H
+
(
+ hH (θd )vd vdH h(θd ))IN ,
d
β12
β12 Ps
(15)
Be =
β22
σ2
He + ( 2 + hH (θe )vd vdH h(θe ))IN ,
2
β1
β1 Ps
(16)
Cd = Hd + σ 2 (β22 Ps )−1 IN , and Ce = He + σ 2 (β22 Ps )−1 IN .
Consider that shrinking or stretching PAN does not change
the ratio value as shown in (14), the problem (P1.1) is
equivalent to
(P1.2) : max
w
wH (IN −1 ⊗ Ce )w
wH (IN −1 ⊗ Bd )w
×
.
wH (IN −1 ⊗ Be )w wH (IN −1 ⊗ Cd )w
(17)
Since (17) is a non-convex quadratic fractional function,
(IN −1 ⊗ Bd ), (IN −1 ⊗ Be ), (IN −1 ⊗ Cd ) and (IN −1 ⊗ Cd )
are N (N − 1) × N (N − 1) positive semi-definite matrices. w
can be solved by utilizing GPI algorithm in [23]. Then, PAN
is able to be reconstructed from w.
B. Optimize vd for fixed PAN
If the AN projection matrix PAN in (12) is fixed , the
optimization problem (12) is simplified to
(P1.3) : max
vd
Rs (vd , fixed PAN )
(18)
vdH vd = 1
subject to
Observing (7), we find that above optimization problem is
equivalent to
(P1.4) : max
vd
subject to
vdH (Hd + Ad IN )vd
vdH (He + Ae IN )vd
vdH vd = 1
(19)
Actually, this is a generalized Rayleigh-Ritz ratio problem
and the optimal vd is the eigenvector corresponding to the
largest eigenvalue of the matrix
(He + Ae IN )−1 (Hd + Ad IN ).
(20)
C. Initialization of PAN and vd
A. Optimize PAN for fixed vd
If the beamforming vector vd in (12) is fixed, then the
optimization problem (12) can be reduced to
(P1.1) : max
PAN
subject to
Rs (fixed vd , PAN )
−2
Tr[PAN PH
.
AN ] = α
The cost function Rs (fixed vd , PAN ) is rewritten as
(13)
Based on the two previous subsections, we propose an
AIS by alternatively solving PAN and vd to further improve
secrecy rate Rs as shown in Fig. 2. The basic idea is to
apply the two steps in Subsections A and B individually and
repeatedly until the SR converges. The detailed process is as
follows. Firstly, we take the initial values of PAN and vd
being the associated eigen-vectors of the largest egienvalues
of (23) and (25), respectively. Then, we compute PAN by
4
utilizing GPI algorithm in [23] for a given fixed vd . Subsequently, vd is chosen to be the associated eigen-vectors of the
largest egienvalues of matrix(20) under the condition PAN is
fixed. We repeat the process until the termination condition is
satisfied.
Begin
Input : h d , h e , b1 , b 2 , a , d , and i = 1
i
PAN
Projection
Matrix PAN
1
Compute the initial value of PAN
and
Beamforming
Vector v d
v1d according to (23) and (25)
v id
i = i +1
Fig. 2. Schematic diagram of AIS
Generally, the convergence rate of the iterative algorithm
in Fig. 2 depends intimately on the the initial value. In what
flows, we focus on the initialization of PAN and vd using the
leakage concept in accordance with (3) and (4).
From the aspect of interfering the eavesdropper, the AN
along direction θe should be viewed a useful component for
eavesdropper, and the leakage of AN to the desired direction
θd is regarded as interference. Hence, the AN-to-leakage-plusnoise ratio (ANLNR) corresponding to PAN is defined as
AN LN R(PAN ) =
=
α2 β22 Ps hH (θe )PAN PH
AN h(θe )
2
2
α β22 Ps hH (θd )PAN PH
AN h(θd ) + σ
Tr[PH
AN He PAN ]
.
(21)
2
H
Tr[PAN (Hd + βσ2 Ps IN )PAN ]
2
When maximizing the above cost function AN LN R(PAN ),
the optimization variable PAN is a matrix. To solve this
problem, we convert matrix PAN into a column vector w
using vec operation, and the associated cost function becomes
AN LN R(w) =
wH (IN −1 ⊗ He )w
.
2
wH [IN −1 ⊗ (Hd + βσ2 Ps IN )]w
(22)
2
Therefore, by maximizing the above objective function
AN LN R(w), we have the optimum w being the eigenvector
corresponding to the largest eigenvalues of matrix
−1
(IN −1 ⊗ He ). (23)
IN −1 ⊗ (Hd + σ 2 (β22 Ps )−1 IN )
This completes the initialization of PAN .
Similarly, as the desired user hopes that the confidential
message x should be leaked to the eavesdropper along the
eavesdropper direction θe as little as possible, we define
the confidential signal-to-leakage-plus-noise ratio (CSLNR)
corresponding to vd as
CSLN R(vd ) =
=
β12 Ps vdH h(θd )h(θd )H vd
2
β1 Ps vdH h(θe )h(θe )H vd + σ 2
vdH Hd vd
.
2
vdH (He + βσ2 Ps IN )vd
(24)
1
Maximizing (24) yields the initial value of vd being the
eigenvector corresponding to the largest eigenvalues of matrix
(He + σ 2 (β12 Ps )−1 IN )−1 Hd .
(25)
i
Solve PAN
by GPI algorithm in [23]
Update v id based on (20)
i
Compute Rsi in (7) using PAN
and vid
No
| Rsi - Rsi -1 |< d
Yes
Output : PAN , v d , Rs
End
Fig. 3. Flow graph of our proposed algorithm
Finally, the detailed implementation process of our proposed
AIS scheme is summarized in Algorithm 1.
In Algorithm 1, parameter δ is the tolerance factor. To make
the above algorithm more clear, the corresponding detailed
flow graph is also presented in Fig. 3.
IV. S IMULATION , D ISCUSSION ,
A NALYSIS
AND
C OMPLEXITY
In this section, simulation results are presented to evaluate
the performance of the proposed algorithm. The leakagebased method in [25] and NSP method in [20] are used For
comparison, . The system
parameters are
√ chosen as: N = 8,
√
PA factors β1 = 0.9 and β2 = 0.1, θd = 45◦ and
θe = 70◦ , and QPSK.
A. Simulation results and analysis
Fig. 4 shows the achievable SR versus the number of
iterations between PAN and vd for leakage-based and random
5
Algorithm 1 Proposed iterative algorithm to solve PAN and
vd
Input: hd , he , β1 , β2 , α and δ
Output: PAN , vd , Rs
Initialization: i = 1, and compute the initial value of P1AN
and vd1 by using (23) and (25).
repeat
1. i = i + 1.
2. Update PiAN utilizing GPI algorithm in [23];
3. Update vdi based on (20);
4. Compute Rsi in (7) using updated PiAN and vdi .
until |Rsi − Rsi−1 | < δ.
return PAN , vd and Rs .
9
Proposed method
Leakage-based method in [25]
NSP in [20]
8
Secrecy Rate(bits/s/Hz)
7
6
5
4
3
2
1
0
0
5
10
15
20
25
SNR(dB)
3.45
Fig. 5. Achievable secrecy rate versus SNR of three different methods
100
3.35
10-1
10-2
3.3
BER
Secrecy Rate(bits/s/Hz)
3.4
10-2
3.25
35 40 45 50 55
10-3
Leakage-based initialization
Random initialization
Proposed method
Leakage-based method in [25]
NSP in [20]
3.2
1
2
3
4
5
6
7
8
9
10
Number of iterations
10-4
0
20
40
60
80
100
120
140
160
180
Directional angle (degree)
Fig. 4. Curves of achievable secrecy rate versus the number of iterations
(SNR=10dB)
initial values, respectively. Here, SNR is chosen to be 10dB. It
is seen that the iterative algorithm with leakage-based solution
as initialization value converges more rapidly than that with
random initialization. As shown in Fig. 4, the former converges
to a constant rate after 4 iterations, while the latter requires
7 iterations to converge. And the two initialization methods
finally converge to the same SR limit value. This implies that
the SLNR initialization can provide a faster convergence speed
than the random initialization .
Fig. 5 illustrates the achievable SR versus SNR of the
proposed method, leakage-based method in [25] and NSP
method in [20]. It can be seen that our proposed method
performs better than the remaining two methods in almost
all SNR regions. With the increase of SNR, the secrecy rate
gain over them achieved by the proposed method show a
gradual growth trend. For example, at SNR=15dB, the proposed method attains an approximate ten-percent and twentypercent rate improvements over leakage and NSP methods,
respectively.
Fig. 6 plots the bit error rate (BER) versus SNR of the
proposed method, leakage-based method and NSP. All the
three methods achieve their best BER performance along the
Fig. 6. Curves of BER versus directional angles of three different methods
desired direction θd = 45◦ , and a sharp BER performance
degradation appears once the desired receiver deviates from the
main beam of the desired direction 45◦ . Both NSP and our proposed method have approximately the same BER performance
around the desired direction 45◦ . The main reason is that the
proposed method reduces the effect of artificial noise on the
desired direction by maximizing SR. The NSP even makes
AN vanish in the desired direction. However, the conventional
NSP one shows a better BER performance than the proposed
method along the eavesdropper direction θe = 70◦ , which
means that the confidential messages can be easily intercepted
along the direction.
B. Complexity comparison and convergence analysis
The complexities of the proposed Max-SR, NSP, and
leakage-based methods are O(I(N 6 )), O(N 3 ), and O(N 3 )
floating-point operations (FLOPs), respectively, where I stands
for the number of iterations. Our proposed method is due to the
fact that the optimization variable w , vec(PAN ) in (13) is an
N (N − 1)-D column vector. Clearly, the proposed method has
much higher complexity than NSP and leakage-based methods.
6
Both NSP and leakage-based methods have the same order of
complexity.
For the aspect of convergence, we propose two different
choices of initial values in Section III: random and leakagebased. From Fig. 4, the former requires 7 iterations to converge
the limit and the latter needs only 4 iterations. Clearly, the
leakage-based initialization approximately doubles the convergence rate compared to the random initialization. This
means a good initial choice accelerates the convergence of
our proposed method.
V. C ONCLUSION
In this paper, we have investigated three beamforming
schemes including the proposed GPI-based Max-SR, leakage,
and NSP. Compared to the last two methods based on NSP and
leakage, the proposed GPI-based Max-SR method achieved a
substantial SR improvement in the medium and large SNR
regions. In particular, as SNR increases, the SR performance
gain increases. Additionally, by an appropriate choice of initialization value, for example, with the leakage-based solution
as initial value of GPI, the proposed method required only 4
iterations to converge to the limit value of SR. The random
initialization requires 7 iterations. Thus, the leakage-based
initialization saves about 40-percent computational amount
over random one. In the coming future, the proposed scheme
will be potentially applied to the following diverse applications
such as mmWave communications (massive MIMO), IoT,
UAV, satellite communications, and flying ad-hoc networks.
R EFERENCES
[1] A. D. Wyner, “The wire-tap channel,” Bell. Syst. Tech. J., vol. 54, no. 8,
pp. 1355–1387, Oct. 1975.
[2] N. Zhao, F. R. Yu, M. Li, Q. Yan, and V. C. Leung, “Physical layer
security issues in interference-alignment-based wireless networks,” IEEE
Commun. Mag., vol. 54, no. 8, pp. 162–168, Aug. 2016.
[3] S. Yan, X. Zhou, N. Yang, B. He, and T. D. Abhayapala, “Artificialnoise-aided secure transmission in wiretap channels with transmitter-side
correlation,” IEEE Trans. Wireless Commun., vol. 15, no. 12, pp. 8286–
8297, Dec. 2016.
[4] Y. L. Zou, B. Champagne, W. P. Zhu, and L. Hanzo, “Relay-selection
improves the security-reliability trade-off in cognitive radio systems,”
IEEE Trans. Commun., vol. 63, no. 1, pp. 215–228, Jan. 2015.
[5] B. Li, Z. Fei, Z. Chu, and Y. Zhang, “Secure transmission for heterogeneous cellular networks with wireless information and power transfer,”
IEEE Syst. J., vol. PP, no. 99, pp. 1–12, June 2017.
[6] H. M. Wang, Q. Yin, and X. Xia, “Distributed beamforming for physicallayer security of two-way relay networks,” IEEE Trans. Sig. Proc.,
vol. 60, no. 7, pp. 3532–3545, Jul. 2012.
[7] T.Lv, H. Gao, and S. Yang, “Secrecy transmit beamforming for heterogeneous network,” IEEE Journal on Selected Areas in Communications,
vol. 33, no. 6, pp. 1154–1170, June 2015.
[8] M.F.Hanif, L.N.Tran, M.Juntti, and S.Glisic, “On linear precoding strategies for secrecy rate maximization in multiuser multiantenna wireless
networks,” IEEE Transactions on Signal Processing, vol. 62, no. 14, pp.
3536–3551, July 2014.
[9] G. Han, L. Liu, and W. Zhang, “A hierarchical jammed-area mapping
service for ubiquitous communication in smart communities,” IEEE
Communications Magazine, vol. 56, pp. 92–98, 2018.
[10] G. Han, X. Yang, L. Liu, and W. Zhang, “A joint energy replenishing
and data collection algorithm in wireless rechargeable sensor networks,”
IEEE Internet of Thing Journal, vol. PP, pp. 1–1, December 2017.
[11] G. Han, X. Yang, L. Liu, W. Zhang, and M. Guizani, “A disaster
management-oriented path planning for mobile anchor-based localization in wireless sensor networks,” IEEE Transactions on Emerging
Topics in Computing, vol. PP, pp. 1–1, 2017.
[12] A. Maha, D. Spano, A. Kalantari, C. Tsinos, D. Christopoulos,
S. Chatzinotas, and B. Ottersten, “Symbol-level and multicast precoding for multiuser multiantenna downlink: A survey, classification
and challenges,” arXiv preprint, vol. 3, no. 1, pp. 1–21, Mar. 2017,
arXiv:1703.03617.
[13] A. Kalantari, M. Soltanalian, S. Maleki, S. Chatzinotas, and B. Ottersten,
“Directional modulation via symbol-level precoding: A way to enhance
security,” IEEE Jour. on Sel. Topics in Signal Process., vol. 10, no. 8,
pp. 1478–1493, Dec. 2016.
[14] A. Babakhani, D. Rutledge, and A. Hajimiri, “Transmitter architectures
based on nearfield direct antenna modulation,” IEEE J. Solid-State
Circuits, vol. 43, no. 12, pp. 2674–2692, Dec. 2008.
[15] M. P. Daly and J. T. Bernhard, “Directional modulation technique for
phased arrays,” IEEE Trans. Antennas Propag., vol. 57, no. 9, pp. 2633–
2640, Sep. 2009.
[16] Y. Ding and V. Fusco, “A vector approach for the analysis and synthesis
of directional modulation transmitters,” IEEE Trans. Antennas Propag.,
vol. 62, no. 1, pp. 361–370, Jan. 2014.
[17] J. S. Hu, F. Shu, and J. Li, “Robust synthesis method for secure
directional modulation with imperfect direction angle,” IEEE Commun.
Lett., vol. 20, no. 6, pp. 1084–1087, Jun. 2016.
[18] F. Shu, X. M. Wu, J. Li, R. Q. Chen, and B. Vucetic, “Robust
beamforming scheme for secure multi-beam directional modulation in
broadcasting systems,” IEEE Access, vol. 4, pp. 6614–6623, Oct. 2016.
[19] F. Shu, W. Zhu, X. Zhou, J. Li, and J. Lu, “Robust secure transmission
of using main-lobe-integration based leakage beaforming in directional
modulation MU-MIMO systems,” IEEE Syst. J., vol. PP, no. 99, pp. 1–1,
2017.
[20] Y. Ding and V. Fusco, “Orthogonal vector approach for synthesis of
multi-beam directional modulation transmitters,” IEEE Antennas Wireless Propag. Lett., vol. 14, pp. 1330–1333, Feb. 2015.
[21] J. Hu, S. Yan, F. Shu, J. Wang, J. Li, and Y. Zhang, “Artificial-noiseaided secure transmission with directional modulation based on random
frequency diverse arrays,” IEEE Access, vol. 5, pp. 1658–1667, 2017.
[22] F. Shu, X. Wu, J. Hu, R. Chen, and J. Wang, “Secure precise wireless
transmission with random-subcarrier-selection-based directional modulation transmit antenna array,” arXiv preprint, pp. 1–14, Apr 2017,
arXiv:1704.07996.
[23] N. Lee, H. J. Yang, and J. Chun, “Achievable sum-rate maximizing
af relay beamforming scheme in two-way relay channels,” in ICC
Workshops - 2008 IEEE International Conference on Communications
Workshops, May 2008, pp. 300–305.
[24] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge, U.K.:
Cambridge University Press, 2013.
[25] M. Sadek, A. Tarighat, and A. H. Sayed, “A leakage-based precoding
scheme for downlink multi-user MIMO channels,” IEEE Trans. Wireless
Commun., vol. 6, no. 5, pp. 1711–1721, May 2007.
| 7cs.IT
|
1
Enhanced Uplink Resource Allocation in
Non-Orthogonal Multiple Access Systems
arXiv:1711.08900v1 [cs.IT] 24 Nov 2017
Rukhsana Ruby, Member, IEEE, Shuxin Zhong, Hailiang Yang, and Kaishun Wu, Member, IEEE
Abstract—Non-orthogonal multiple access (NOMA) is envisioned to be one of the most beneficial technologies for next generation wireless networks due to its enhanced performance compared to other conventional radio access techniques. Although
the principle of NOMA allows multiple users to use the same
frequency resource, due to decoding complication, information of
users in practical systems cannot be decoded successfully if many
of them use the same channel. Consequently, assigned spectrum
of a system needs to be split into multiple subchannels in order to
multiplex that among many users. Uplink resource allocation for
such systems is more complicated compared to the downlink ones
due to the individual users’ power constraints and discrete nature
of subchannel assignment. In this paper, we propose an uplink
subchannel and power allocation scheme for such systems. Due to
the NP-hard and non-convex nature of the problem, the complete
solution, that optimizes both subchannel assignment and power
allocation jointly, is intractable. Consequently, we solve the
problem in two steps. First, based on the assumption that the
maximal power level of a user is subdivided equally among its
allocated subchannels, we apply many-to-many matching model
to solve the subchannel-user mapping problem. Then, in order to
enhance the performance of the system further, we apply iterative
water-filling and geometric programming two power allocation
techniques to allocate power in each allocated subchannel-user
slot optimally. Extensive simulation has been conducted to verify
the effectiveness of the proposed scheme. The results demonstrate
that the proposed scheme always outperforms all existing works
in this context under all possible scenarios.
Index Terms—NOMA Systems; Optimal Resource Allocation;
Many-to-Many Matching Model; Geometric Programming; Iterative Water-Filling Algorithm
I. I NTRODUCTION
T
He data traffic over cellular networks is projected to grow
explosively in the coming years due to the proliferation of smartphones, tablets, smart terminals and emerging
applications (e.g., machine-type-communications (MTC)) [1]–
[4]. Consequently, future radio access networks [5], [6] are
expected to have the capability of supporting massive connectivity, diverse sets of users and applications with radically
different requirements in terms of delay, bandwidth and so on.
In order to obtain fruitful outcome in this context, designing an
effective and efficient radio access technology [7] is one of the
possible solutions. Through experimentation and theoretical
This research was supported in part by the China NSFC
Grant 61472259, Guangdong Natural Science Foundation (No.
2017A030312008, 2016A030313036), Shenzhen Science and Technology
Foundation (No. JCYJ20170302140946299, JCYJ20170412110753954,
JCYJ20150324140036842).Guangdong Talent Project 2014TQ01X238,
2015TX01X111, GDUPS(2015). Kaishun Wu is the corresponding author. R.
Ruby, S. Zhong, H. Yang and K. Wu are with the college of Computer Science
and Software Engineering, Shenzhen University, Shenzhen, Guangdong,
518060 China, e-mail: [email protected]
analysis, it is proved that non-orthogonal multiple access
(NOMA) technique is able to provide enhanced performance
comparing with other orthogonal multiple access (OMA)
techniques, such as time division multiple access (TDMA)
and frequency division multiple access (FDMA) [8]–[11].
Consequently, NOMA is considered as the future dominating
radio access technique, and is expected to satisfy the everincreasing demands of future cellular networks.
Conceptually, power-domain NOMA allows multiple users
to occupy the same frequency channel. By applying successive interference cancellation (SIC) [12] in NOMA systems,
superposition coded signal can be correctly decoded and demodulated at the receiver. Although NOMA technique allows
multiple users to be superimposed on the same frequency
channel, due to the error propagation in the SIC technique,
it is not an optimal design to assign large number of users
on the same channel. Consequently, dedicated spectrum of a
system needs to be subdivided into multiple subchannels in
order to support increased number of users. At the same time,
how to allocate these subchannels among users in a multiplex
manner given the allowable maximum number of users that
can use a subchannel simultaneously, is an important problem. Extensive research has been conducted on the downlink
subchannel and power allocation for such NOMA systems.
Based on some assumption of having constant power on the
subchannels, typically, existing works provide some heuristics for subchannel-user mapping task. Once the subchanneluser mapping information is known, in order to enhance the
performance of the system further, different existing works
have provided different schemes for power allocation across
the allocated subchannel-user slots. For example, in [13],
[14], the authors use fractional transmit power allocation
technique among users and equal power allocation concept
across subchannels. [15] uses water-filling-based approach for
power allocation, and in [16], the authors use difference of
convex (DC) programming-based [17] approach for the power
allocation in both user and subchannel levels. Energy-efficient
downlink resource allocation has also been studied in some
papers, such as [18]–[20].
Unlike the downlink one, uplink resource allocation even in
conventional OMA systems is considered challenging [21]–
[23] because of the individual users’ power constraints and
discrete nature of subchannel assignment. On the other hand,
decoding technique in NOMA systems, SIC, is a multiuser detection technique that uses the structured nature of
interference to decode multiple concurrent transmissions. Each
individual signal from the composite signal is retrieved one
by one following some order. If any of the signals is failed
2
to be decoded, it is unlikely that the rest of the signals can
be decoded. Therefore, decoding order plays the significant
role on the success of decoding operation and the throughput
achieved by each individual signal. All these complications
bring further difficulties in the uplink resource allocation of
NOMA systems compared to OMA systems. Although NOMA
principle does not enforce the decoding order of received
superposition coded signals, it is proved in [24] that decoding
of stronger signals ahead of weaker signals is beneficial for
the system in terms of throughput and proportional fairness.
Compared to the downlink resource allocation in NOMA
systems, uplink resource allocation is not that much studied.
Still, there are some works in this context. Unlike the system in our work (in which NOMA technique is employed
in the frequency domain), for a system in which NOMA
technique is employed in the time domain, a set of uplink
resource allocation schemes is provided in [24], [25] with
the objective of throughput maximization and fairness of the
system. Although the scheduling scheme in [25] assumes that
the system has one time slot and a set of users with their
power constraints to schedule, the scheduling scheme in [24]
optimizes the total throughput and fairness of the system over
a set of time slots and users. On the other hand, the main
drawback of the work in [24] is, the time slots are resource
elements and are invariant over time, which is very impractical
for wireless networks. Moreover, given the power constraint
of each individual user, each user can get only one time slot
(i.e., one resource block), the concept of which fails to exploit
multi-user-channel diversity of wireless systems. However, in
practice, if multiple time slots are allocated to a user, the
performance of that user may be enhanced. Another very close
work compared to our work is [26]. In this work, the authors
have proposed an uplink subchannel and power allocation
scheme based on iterative water-filling technique [27]. With
the expectation of utilizing the multi-user-channel diversity,
this resource allocation scheme overcomes the drawback of
the solution in [24] by assigning multiple subchannels (i.e.,
multiple resource blocks) to each user. Moreover, this scheme
assigns exactly the maximum allowable number of users to
each subchannel and gives more preference to the users with
better channel while solving the subchannel-user mapping
problem. However, even in uplink OMA systems [22], we
previously observed that not necessarily the more the users
allocated to each subchannel, the better the throughput is,
especially in worse channel condition. This is because each
user requires to subdivide its limited power level among
its allocated subchannels. Furthermore, in NOMA systems,
giving less privilege to the users with worse channel not
necessarily enhances the throughput. Since the power level
of other users in such systems is considered as interference
for some particular user, pairing users with highly different
channel condition is sometimes conducive to the performance
of the system. In this paper, our objective is to overcome
the drawbacks of existing aforementioned uplink resource
allocation schemes, and to take NOMA-specific all useful
scheduling insights into account.
The contribution of this paper is an elegant uplink subchannel and power allocation scheme in a NOMA system with en-
hanced performance. Since this problem considers subchannels
assignment which are associated with discrete variables in the
formulated problem, the problem is NP-hard. Moreover, even
if the subchannel assignment information is known, because
of the interference power resultant from the superposition
coded signals of other users on a specific subchannel, the
power allocation of the problem is non-convex [28]. As a
result, joint subchannel assignment and power allocation of
this problem can be considered as a mixed integer non-linear
programming (MINLP) problem. Overall, joint optimization
of both subchannel assignment and power allocation is not
tractable for this case. Consequently, we solve this problem
in two steps. Based on the assumption that the maximal
power level of each user is subdivided equally among its
allocated users, we apply many-to-many matching model [29],
[30] to solve the subchannel-user mapping problem. Then, we
apply iterative water-filling [27] and geometric programming
(GP) [31] techniques to allocate power across the subchannels for different users. Iterative water-filling is a multi-userchannel power allocation technique, which is developed based
on the insights of single-user water-filling solution. On the
other hand, GP technique can solve special-form of nonconvex problems using convex optimization solvers through
variable transformation. Given the subchannel-user mapping
information, our uplink power allocation problem is amenable
to GP after applying some transformation on the objective
function. Extensive simulation has been conducted to verify
the effectiveness of our proposed uplink resource allocation
scheme comparing with two very similar existing works [24],
[26]. The results demonstrate that the proposed scheme always
outperforms the existing works in terms of computational
complexity, the usage of resource and overall performance.
The rest of the paper is organized as follows. Along with the
background information and the description of the system, we
formulate our uplink resource allocation problem in Section II.
The detailed solution approach is provided in Section III.
Followed by the simulation methodology, we evaluate the
performance of our uplink resource allocation scheme in
Section IV. Finally, Section V concludes the paper with some
direction on future research.
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
In this paper, we consider an uplink scenario of a cellular
network, which has one base station. Time is divided into
frames, and in each frame, the entire pre-assigned spectrum
for the system is divided into N subchannels with equal
bandwidth. The resultant subchannels are the elements of a
set, denoted by N. There are M number of users in the system,
and the corresponding set holding these users is denoted by
M. Using the subchannels in set N as the transmission media,
the users in set M transmit data to the base station. Each
user m in set M has the maximal power level, denoted by
pmax
. Both the base station and the users in the system
m
are equipped with NOMA technologies. The users transmit
their data using superposition coding (SC) technique over
a set of subchannels. Whereas, the receiver, i.e., the base
station applies SIC technique on each subchannel to decode the
3
We consider the block fading channel model [32]. It implies
that the CSI of the subchannels in the system remains constant
over a time frame, however vary independently across different time frames. Although NOMA techniques have various
classification, we plan to exploit power-domain NOMA. We
assume that the base station assigns Mn number of users on
the nth subchannel, and the corresponding set holding these
users is denoted by Mn . If each individual user m transmits
√ n
pm sm symbol on subchannel n, the symbol received by the
base station on this subchannel can be expressed as
Subchannel n
user 1
user 2
user 1
user 3
user 2
Fig. 1: A sample example for the organization of decoding
order in subchannel n ((b) is the organized version of (a)),
where pn1 g1n = 20, pn2 g2n = 10 and pn3 g3n = 30 (1st case), or
g1n = 0.2, g2n = 0.1 and g3n = 0.3 (2nd case).
m=1
n
ym
= hnm xn + zn
X p
√ n n
= p m hm sm +
pni hnm si + zn ,
(2)
i=1,i6=m
where hnm is the channel gain of user m on the nth subchannel.
zn is the noise power over subchannel n, which follows
Additive White Gaussian Noise (AWGN) [33] distribution
with mean zero and variance σn2 , i.e., zn ≈ CN (0, σn2 ). The
noise power of subchannel n is statistically same for all users.
In NOMA systems, each subchannel is shared by multiple
users. Consequently, each user on subchannel n receives its
signal as well as the interference signals from other users
on the same subchannel. Therefore, without SIC at the base
station, the received SINR of the mth user on subchannel n
is given by
SINRnm =
n
pnm gm
pnm |hnm |2
=
, (3)
Mn
Mn
X
X
σn2 +
1+
pni |hni |2
pni gin
i=1,i6=m
i=1,i6=m
where σn2 = E[|zn |2 ] is the noise power on subchannel n, and
gin = |hni |2 /σn2 is the normalized channel gain of user m on
subchannel n. Based on Shannon’s capacity formula [34], the
sum-rate of subchannel n is given by
Mn
X
log2 (1 + SINRnm )
m=1
(1)
where sm is the modulated symbol of the mth user on
subchannel n, and pnm is the power level assigned to user m
on subchannel n. Consequently, the signal of user m, received
by the base station on subchannel n, can be represented as
(b)
(a)
Rn =
Mn
X
p
pnm sm ,
xn =
user 3
Interference
for user 3
superimposed signals, and extracts the corresponding signal of
each individual user. However, before the uplink transmission
operation, it is required to schedule subchannels and power
across the users optimally so that the capacity of the system
is maximized. We assume that the scheduling scheme in the
system is centralized, and the base station is appointed to
conduct this operation. To develop this scheduling scheme,
the entire channel state information (CSI) of the system is
required, and hence the base station is aware of all these
information. At the beginning of each time frame, all users
send their CSI to the base station via some reliable control
channels.
n
pnm gm
=
log2 1 +
.
M
n
X
m=1
n n
1+
pi g i
Mn
X
(4)
i=1,i6=m
In NOMA systems, the SIC process is implemented at the
receiver to reduce the interference from other users on the
same subchannel. According to [24], it has been proved that
the optimal decoding order is equivalent to the decreasing
order of received power. In this way, the interference imposed
on each user by other users of the same subchannel is
reduced, and consequently the sum-capacity and proportional
fairness of the system are enhanced. Based on this truth, we
adopt the following rule while determining the decoding order
of the users transmitting on the same subchannel. As we
mentioned in the introduction that our solution approach of
this problem consists of two steps. In the first step, based
on the assumption that the maximal power level of each user
is equally subdivided among its allocated subchannels. In this
case, we decode the users superimposed on a subchannel in the
decreasing order of their received power. A sample example
of this idea has been provided in Fig. 1.
For the second case, when the user and subchannel assignment information are known, the power allocation of the superimposed users should follow some order. From the insights
of the decoding order idea of [24] as well as intuitively, it
is obvious that the user with better gain in any subchannel
should be assigned with larger transmit power. In this way, the
interference imposed on any user (assigned to any subchannel)
caused by other users is reduced. A sample example of this
idea is provided in Fig. 1 as well. Due to the decoding order
concept and the principle of SIC technique, not necessarily
4
other all users assigned to a particular subchannel impose
interference on a specific user. On subchannel n, denote
that the users in set Mm
n produce interference for user m.
Consequently, the sum-rate of subchannel n (in (4)) can be
rewritten as
!
Mn
n
X
pnm gm
P
.
(5)
Rn =
log2 1 +
1 + i∈Mm pni gin
m=1
n
In this work, while preserving the power constraints of
all users, our objective is to allocate the subchannels in
set N across all users in set M, and assign power level to
each subchannel-user slot so that the capacity of the system
is maximized. Clearly, this is an optimization problem. To
formulate this problem, we define a binary variable αnm .
αnm = 1 implies that subchannel n is allocated to user m,
and αnm = 0 means the other case. It is proved in [26] that the
more users are assigned to a subchannel, the better the system
capacity is. However, due to the varying nature of wireless
channels and the decoding complication of SIC technique, not
necessarily more users assigned to a subchannel will enhance
the system throughput. While giving weight to this observation
and insight, we assume that maximum K number of users can
be assigned to a subchannel. Therefore, the uplink subchannel
and power allocation problem in this context can be formulated
as follows.
n
pnm gm
X
αnm log 1 +
,
n n
{pm ,αm }
1+
pi g i
n∈N m∈Mn
max
n
n
X X
(6)
i∈Mm
n
subject to,
X
αnm ≤ K, ∀n ∈ N,
(7)
αnm ≤ ∞, ∀m ∈ M,
(8)
m∈Mn
X
n∈N
αn ∈ {0, 1}, ∀m ∈ Mn , ∀n ∈ N,
Xm
αnm pnm ≤ pmax
, pnm ≥ 0, ∀m ∈ M.
m
(9)
(10)
n∈N
In the above formulation, there are two types of variables,
i.e., {αnm } and {pnm }. {αnm } are the set of discrete variables,
and the problem is NP-hard because of these variables. On
the other hand, even if the informationPof set {αnm } is known,
pni gin inside the
because of the interference term 1 + i∈Mm
n
log term of (6), the problem is non-convex. While considering
the overall structure, we can say that the problem is jointly
NP-hard and non-convex.
III. S OLUTION A PPROACH
In this section, we explore the solution approach of the uplink resource allocation problem of a NOMA system described
in the previous section. The entire problem is formulated
in (6)-(10). Apparently, due to the discrete nature of subchannel assignment (i.e., variables αnm ) and the continuous nature
of power assignment (i.e., variables pnm ), this is a MINLP
problem. This type of problem even in conventional OMA
systems is intractable. When it comes to the case of NOMA
systems, due to the superimposition of multiple users on the
same subchannel, the solution of this problem brings further
complication. Therefore, we have decomposed the problem in
two parts. In the first part, based on the assumption that the
maximal power level of a user is subdivided equally among its
allocated subchannels, we solve the subchannel-user mapping
problem. In this case, we find that two-sided matching model
is appropriate to capture the structure of this problem. Since
one user can be assigned with multiple subchannels and one
subchannel can have multiple users, many-to-many matching
scheme is expected to solve the first problem. For the solution
of the second part of the problem, we assume that we have
the subchannel-user mapping information. Even though the
subchannel-user mapping information is known, the power
allocation across the subchannels and users, given the power
constraints of the users, is a non-convex problem. Based on
the structure of the problem, we find that iterative water-filling
and GP are very appropriate techniques to solve this problem.
A. Subchannel and User Mapping Scheme
In this system, intuitively, assignment of many users to a
subchannel and allocating multiple users to a subchannel (to
follow the guidelines of NOMA technique) is envisioned to
enhance the overall throughput. Given the power constraint of
each user, this problem is NP-hard. However, the nature of
the problem implies that many-to-many matching model [29],
[30] is appropriate to solve this problem. Given that maximum
K number of users can be multiplexed on a subchannel, M
users in set M and N subchannels in set N are two sets of
players of this many-to-many matching relation. Note that each
user m can have infinite (N in practice) number of users if
possible. However, since user m has maximal power constraint
pmax
, this should be subdivided equally among its allocated
m
subchannels.
Definition 1: A many-to-many matching µ is a mapping from
set M to set N such that every m ∈ M and n ∈ N satisfy the
following properties:
• µ(m) ⊆ N and µ(n) ⊆ M
• |µ(m)| ≤ ∞, ∀m ∈ M
• |µ(n)| ≤ K, ∀n ∈ N
• n ∈ µ(m) if and only if m ∈ µ(n)
where µ(m) is the set of partners for user m and µ(n) is the set
of partners for subchannel n under the matching model µ. The
definition states that each user in set M is matched to a subset
of subchannels in set N, and vice versa. In other words, each
user may choose a set of subchannels as the communication
media, whereas each subchannel may choose a set of users to
be assigned with in order to maximize the overall benefit of
the system. However, before accomplishing these assignment
operations, each user needs to have preference list based on
some criteria. The criterion of constructing preference list for
users is based on their received power from the subchannels.
n
For example, if the gain of subchannel n for user m is gm
n
and assigned power level of this subchannel is pm , then the
n
. We use the
received power from this subchannel is pnm gm
′
′′
notation Ω m ≻Ω m to imply that user m wants to have the
5
subchannels in subset Ω′ m than the subchannels in subset
Ω′′ m , where Ω′ m ⊆ N and Ω′′ m ⊆ N. Similar analogy can
be made for any subchannel n in set N. Preference of each
subchannel n is based on the overall benefit (i.e., throughput)
of the system. For example, if user m chooses subchannel n,
this subchannel only accepts this user if and only the system
performance is enhanced by this allocation.
To solve our subchannel-user mapping problem, we are
interested to look at a stable solution, in which there are no
players that are not matched to one another but they all prefer
to be partners. Since subchannel players give preference to the
overall throughput of the system while choosing partners from
set M, stable solution is envisioned to be the optimal solution
for this problem. In many-to-many matching models [30],
many stability concepts can be considered depending on the
number of players that can improve their utility by forming
new partners among one another. However, due to the large
number of players (M∪N) in our problem, identifying optimal
subset of partners for a player is intractable. Consequently, we
choose to solve the matching problem by identifying partner
one by one from the opposite set. This way of choosing partner
in the matching model brings pair-wise stability.
In Definition 2 and Definition 3, we highlight some properties of pair-wise stable matching relation. For the sake of
these definitions, we define some notations as follows. Faced
with a set N̂ ⊆ N of possible partners, a player m ∈ M can
determine which subset of set N̂, it wishes to match to. We
denote this choice set as Cm (N̂).
Definition 2: A matching relation µ is pairwise stable if there
does not exist a pair (m, n) with m 6∈ µ(n) and n 6∈ µ(m)
such that φ ∈ Cm (µ(m) ∪ {n}) and ϕ ∈ Cn (µ(n) ∪ {m}),
and at the same time both {φ}≻m µ(m) and {ϕ}≻n µ(n) are
satisfied.
Definition 3: Let M̂ is the subset of users in set M. The
preference of subchannel n is called substitutable if there exist
users such that m, m′ ∈ Cn (M̂), then m ∈ Cn (M̂/ {m′ }) is
satisfied.
While satisfying the properties of stable many-to-many
matching relation, we have proposed an algorithm in Algorithm 1. Note that in this algorithm, we are interested
in pair-wise stability, and hence the players (i.e., users and
subchannels) choose their partners one-by-one instead of a
subset. We have adopted a few paradigms or strategies in
order to bring stability in this matching relation or enhance the
overall system performance. The description of the algorithm
is as follows. First, Ωm , m ∈ M are initialized with ∅, which
basically contains the allocated subchannels of user m ∈ M.
At the same time, Mn , n ∈ N are initialized with ∅ as
well. Over the iterations, these sets are filled by the allocated
subchannels and users, respectively. At the initialization phase,
each user m ∈ M also constructs its subchannel preference list
based on the descending order of their received power level.
n
If the gain of subchannel n for user m is gm
and the assigned
n
power level is pm , the received power level of this subchannel
n
for this user is pnm gm
. Since we have an assumption that
the maximal power level of user m is subdivided equally
among its allocated subchannels, the preference list of user m
pmax
is constructed based on the assumption that |Ωmm |+1 amount
of power is reserved for subchannel n ∈ {N/ Ωm }. Then,
inside the outer-most loop (between step 3 and step 34), if no
assignment is possible inside the second outer loop (between
line 4 and line 33), the algorithm terminates1 . Inside the innermost loop (between step 5 and step 32), each user m chooses
its most preferred unallocated subchannel n. At this point,
two conditions are possible. The first condition is that the
number of allocated users on subchannel n can be less than
K (maximum allowable number of users per subchannel), and
the second condition is the other case. If the first condition
is true, we can apply two strategies for this subchannel-user
assignment: either user m is substituted by one of the existing
users (e.g., m′ ∈ Mn ) on subchannel n, or user m can be
added to this subchannel. Each of these strategies is inserted
to strategy set S (which was initialized before initiating the
loop). Whereas, for the second case, only addition strategy is
possible. After filling the strategy set S no matter the number
of allocated users on subchannel n is less than or equal to
K, the elements of S are filtered based on some criterion,
which is as follows. If strategy s is a substitution policy, let
m′ is to-be-replaced user, and hence N′ = Ωm ∪ Ωm′ is the
set of affected subchannels. Moreover, let thr′ be the total
computed throughput (following (5)) of the subchannels in set
N′ before applying strategy s. Then, after applying strategy
s and adjusting the power level of user m and m′ in set
N′ , in the similar manner, the total throughput is computed
(denoted by thr). Strategy s is only added to set CS if and
only if this throughput (due to applying this strategy) is bigger
than thr′ . For the addition strategy, the affected subchannels
are the ones in set Ωm , and hence N′ = Ωm . For this
strategy, in the similar manner, set S is filtered and set CS
is updated. Finally, sBest strategy is chosen based on the total
throughput each strategy incurs. If sBest is empty, the innermost loop continues, and the next user is chosen from set M
for building its possible strategy set. If sBest is not empty,
the corresponding strategy is executed. As a result, set Ωm ,
set Ωm′ (only for the substitution strategy), and set Mn are
updated. The power level of user m for the subchannels in
set Ωm are adjusted, and its preference list is updated as well
due to the updated power level. For only substitution strategy,
the power level of user m′ in its affected subchannels and
its preference list is updated. By analyzing the algorithm, we
conclude Proposition 1, Lemma 1 and Theorem 1.
Proposition 1: Rejected users by the subchannels are not final.
For example, if a user m ∈ M is rejected by subchannel n ∈ N
at some iteration i, at the i′ th (i′ ≥ i) iteration, it is possible
that both m ∈ µ(n) and n ∈ µ(m) will appear true.
1 At this point, it is assumed that the system has reached a stable situation
or the performance improvement is no longer possible.
6
Algorithm 1 The uplink subchannel-user mapping algorithm
using many-to-many matching model.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
Ωm ← ∅, ∀m ∈ M; Mn ← ∅, n ∈ N.
Each user m ∈ M produces its preference list.
repeat
for i ← 1 to N do
for m ∈ M do
n ← The best unallocated subchannel from the
preference list of user m.
S ← ∅, CS ← ∅, and Thrput ← ∅.
if The number of assigned users on subchannel n
is less than K then
Construct each strategy s that is supposed to
replace each user m′ ∈ Mn by user m, and
insert the corresponding s to set S.
Construct another strategy s that is supposed to
add user m to set Mn , and insert the corresponding s to set S.
else if The number of assigned users on subchannel
n is equal to K then
Construct each strategy s that is supposed to
replace each user m′ ∈ Mn by user m, and
insert the corresponding s to set S.
end if
for s ∈ S do
Determine the affected subchannel list N′ for
strategy s.
Adjust the power level of user m′ ∈
Mn′ , ∀n′ ∈ N′ .
thr ← The sum-throughput of all subchannels
in set N′ due to strategy s.
Insert strategy s to CS and insert thr to set
Thrput if thr is larger than the sum-throughput
of affected subchannel list N′ before applying
strategy s.
end for
sBest ← arg maxs∈CS Thrput(s).
if sBest is the replacement strategy then
Ωm ← Ωm ∪ {n}, Ωm′ ← Ωm′ / {n}, and
Mn ← (Mn / {m′ }) ∪ {m}. {m′ is the tobe-replaced user and m is the new user on
subchannel n}
Adjust the power level of the subchannels in Ωm
and Ω′m , and update the preference lists of user
m and user m′ .
else if sBest is the addition strategy then
m is the new user on subchannel n.
Ωm ← Ωm ∪ {n}, and Mn ← Mn ∪ {m}. {m
is the new user on subchannel n}
Adjust the power level of the subchannels in
Ωm , and update the preference list of user m.
end if
if sBest is not empty then
Terminate this loop.
end if
end for
end for
until The performance enhancement of the system is not
possible
Proof: Let assume K = 2, which implies that maximum
allowable number of users per subchannel is 2. Consider that
user 1 and user 2 are already matched with subchannel n in
some iteration i, and pn2 > pn1 holds. At iteration i′ (i′ > i),
user 3 has come to obtain subchannel n with the power level
pn3 , and pn2 > pn3 > pn1 holds. For the sake of simplicity, we
further assume that g1n = g2n = g3n = 1. We know that if
pn2 > pn3 > pn1 holds, log2 (1 + pn2 + pn3 ) > log2 (1 + pn2 + pn1 )
always satisfies. Therefore, at this stage, based on the replacement strategy our algorithm provides, user 3 is replaced by
user 1. Since user 1 is unallocated from subchannel n, in
other subchannels (of set Ω1 ) to which user 1 is belonged
to, the power level of this user will be increased. This is
because the algorithm ensures that the maximal power level
pmax
is equally subdivided among the allocated subchannels
1
of user 1. In the similar manner and for the same reason, user
1 may further be replaced by some other user on its some other
assigned subchannel (in set Ω1 ). At this stage, at iteration i′′
(i′′ > i′ ), user 1 is able to compete (although rejected already
at iteration i′ ) for obtaining the nth subchannel again with the
increased level of power, denoted by p̃n1 (p̃n1 > pn3 ). Thus,
we prove that user 1 may come again to choose subchannel
n, and can be replaced by user 3. This is due to the fact
log2 (1+pn2 + p̃n1 ) > log2 (1+pn2 +pn3 ) because of p̃n1 > pn3 . By
adopting the strategy in this proposition, the algorithm ensures
as better performance as possible for the system.
Lemma 1: The subchannel-user mapping algorithm (i.e.,
Algorithm 1) is guaranteed to converge to a pair-wise stable
matching relation.
Proof: We prove this lemma by contradiction. Suppose that
there exist a user m and a subchannel n with m 6∈ µ(n)
and n 6∈ µ(m) such that φ ∈ Cm (µ(m) ∪ {n}), ϕ ∈
Cn (µ(n) ∪ {m}), and at the same time both {φ}≻m µ(m)
and {ϕ}≻n µ(n) are satisfied. Since {n}≻m µ(m) is true, user
m must propose subchannel n in some earlier iteration to be
paired with. However, at the same time, both m 6∈ µ(n) and
n 6∈ µ(m) are true. Consequently, at the proposal time of user
m, either subchannel n had some better preference compared
with user m and rejected this user, or accepted this user and
then made a replacement with some other user in the latter
iteration. Therefore, m 6∈ Cn (µ(n) ∪ {m}) cannot be a false
statement, and hence matching relation µ cannot be unstable.
Theorem 1: Algorithm 1 terminates after finite number of
iterations.
Proof: Algorithm 1 is proposed to solve the subchannel-user
mapping problem, and this is an optimization problem. No
matter the problem is convex or non-convex, we know that
every optimization problem has a unique global solution. If
the problem is non-convex, there might be some local optimal
solutions. However, for the convex problem, the solution is
unique, which can be assumed as both local and global. As we
verify before that the problem discussed herein is non-convex
and NP-hard, the global optimal solution requires to search
all possible feasible solution spaces, which is computationally
intensive and is not feasible to implement in a wireless system,
where channel turnaround time is in millisecond/microsecond
level. Consequently, we proposed Algorithm 1 to solve this
problem. We do not claim that the algorithm always finds
7
122.5
69
System Throughput
System Throughput
122
68
67
66
K=6
K=8
121.5
121
120.5
K=6
K=8
120
119.5
119
65
118.5
118
64
117.5
1
2
3
4
5
6
7
8
9
1
2
3
4
5
6
7
8
9
Outer-Most Loop Iteration Index
Outer-Most Loop Iteration Index
(a) M = 40.
(b) M = 60.
10
11
Fig. 2: Increasing system throughput as the outer-most loop
iteration of Algorithm 1 advances.
13
14
12
13
Average Number of Iterations
Average Number of Iterations
the global optimal solution as this is based on the many-tomany matching scheme. In the many-to-many matching model,
since the number of players (i.e., users and subchannels) on
the both sides is large, in terms of their associations, many
combinations are possible as mentioned above. Therefore, we
mostly focused on pair-wise stability, and Algorithm 1 is developed based on this concept. The outer-most loop continues if
and only if at least one allocation (subchannel-user mapping)
is executed. Interestingly, in our algorithm, every allocation
enhances the system throughput compared to the throughput
before that particular allocation. Step 18 of the algorithm is the
evidence of this statement. Through addition and substitution
operations, for a user, we define possible strategies for the
tagged user and its preferred subchannel. This implies that the
tagged user will be added to its most preferred subchannel or
will replace an existing user of that subchannel if only if the
system capacity is enhanced by this allocation. Therefore, in
every allocation, if the system capacity is enhanced, eventually,
the process approaches the convergence state as the solution of
the problem exists in the finite domain. Moreover, no matter
the problem is convex or non-convex, the global or local
optimum point has convex nature. Consequently, after finite
number of iterations, the algorithm converges and terminates.
In order to provide practical evidence of the convergence
event, we have plot Fig. 2. In each subfigure of this figure,
we show the system throughput with the increasing outermost loop iterations for different values of K. As observed,
in each iteration, the system throughput is increased little by
little before reaching the convergence state.
In Fig. 3a and Fig. 3b, we present the average number of
iterations the outer-most loop runs before reaching the optimal
point with the increasing number of users and the increasing
value of K, respectively. The total number of subchannels
in both figures is constant and fixed. Intuitively, given the
number of users and the number of subchannels to be matched,
the outer-most loop should run the times surrounding the
value of K. However, when the number of users is less,
not necessarily, each subchannel obtains exactly K number
of users. Moreover, when the number of users in the system
is less, the number of strategies with substitution operation
is relatively less compared to the case when the number
of users in the system is higher. In this case, the outermost loop terminates in less number of iterations, which is
obvious in Fig. 3a. On the other hand, when the value of K
is lower, for the given number of users and subchannels in
the system, each subchannel obtains relatively less number of
users compared to the case with larger value of K. Therefore,
due to the restriction on the less number of users allocated to
each subchannel, the number of strategies accompanied with
substitution and addition operations is less in this case as
well similar to the other case. Consequently, the outer-most
loop requires less number of iterations to run before reaching
the optimal possible point of the system. Fig. 3b presents the
corresponding observation in this context.
Computational Complexity of Algorithm 1: The joint worst
case complexity of the inner two loops of the algorithm is
O(M N ). Inside these loops, all operations occur in constant
time, and so we can ignore the complexity of these operations.
11
10
9
8
7
6
12
11
10
9
8
7
6
K=6
K=8
5
M=40
M=60
5
10
20
30
40
50
60
2
4
6
8
10
12
Number of Users (M)
Maximum Allowable Users on a Subchannel (K)
(a) Results while varying M.
(b) Results while varying K.
Fig. 3: The number of outer-most loop iterations required
before achieving stability.
Mainly, the running time of the outer-most loop dominates
the computation time of the entire algorithm. As shown in
Fig. 3a and Fig. 3b, the iterations of this outer-most loop is
proportional to the number of users (M ) and the maximum
allowable number of user per subchannel (K), but this is not
a very large number. Therefore, we conclude that the algorithm
has polynomial time complexity. Whereas, the complexity
of the brute-force searching operation to obtain the optimal
N
N
×···
) ≈ O(M N ), which is an order
solution is O( M 1×(M−1)
N ×2N ×···
of exponential series.
B. Power Allocation Schemes
From Algorithm 1, we know the subchannel-user mapping
information, i.e., Mn , n ∈ N and Ωm , m ∈ M. This information is derived based on the assumption that the maximal
power level pmax
of user m is equally subdivided among
m
its allocated subchannels, i.e., pnm = pmax
/|Ωm |, m ∈ M.
m
However, in (6), we see that the instantaneous rate of user m
on subchannel n is positively proportional to pnm and inversely
proportion to the interference power level of other users, i.e.,
pni , i ∈ Mm
n . Consequently, even if the information about
Mn , n ∈ N and Ωm , m ∈ M is known, the power allocation
across all subchannel-user slots, i.e., pnm , m ∈ M, n ∈ N, is
an optimization problem. Consequently, the next objective of
this resource allocation scheme is to allocate power across
all subchannel-user slots optimally. We have adopted two approaches to solve this power allocation problem, the individual
description of which is provided in the following discussions.
1) Iterative Water-Filling Algorithm:
Given that
subchannel-user mapping information, Mn , n ∈ N and
Ωm , m ∈ M are known, the power allocation problem can be
written as
8
max
{pn
m }m∈Mn ,n∈N
s.t.,
X
log2 1 +
X
n
pnm gm
m∈Mn
n∈N
!
,
n∈Ωm
pnm ≤ pmax
, ∀m ∈ M,
m
X
n∈Ωm
(11)
+
X
=
λm
m∈M
X
log2
1+
n
pnm gm
m∈Mn
n∈N
pmax −
m
X
X
n∈Ωm
pnm
!
.
!
pnm =
1
1
− n 1 +
λm gm
1
1
− n 1 +
λm
gm
λm =
X
i∈Mn ,i6=m
1:
2:
3:
,
5:
pni gin , ∀m ∈ Mn , ∀n ∈ N.
(13)
If we compare the solution in (13) with the single-user
n
water-filling solution pnm = 1/λm −1/gm
, it is obvious that the
optimal power level of one user considers the received power
of other users as noise. Based on this intuition, we develop
an iterative algorithm in Algorithm 2 to solve this power
allocation problem. The algorithm works as follows. First,
power level of all users over all subchannels are initialized.
Then, for each user, water-filling power allocation problem
is solved assuming the power level of other users as noise.
Once the single-user water-filling solutions are obtained for all
users, the resultant solutions of all users are replaced by the
previously initialized power levels. This operation is continued
until the performance of the system appears to be saturated.
While solving the single-user water-filling problem in Algorithm 2, typically, bisection search is applied to obtain the
optimal value of λm , m ∈ M. If the proper interval of the
bisection search is not chosen, running time of the bisection
search is huge. Moreover, the accuracy of the solution obtained from the bisection search is greatly dependent on the
precision level of λm as this is a variable with continuous
nature. Therefore, to obtain the optimal value of λm , we have
developed a low complexity procedure in Algorithm 3. In this
algorithm,
user m, Am is a vector,
the elements
of which
i
h for P
1
n n
are gn 1 + i∈Mn ,i6=m pi gi , n ∈ Ωm . The insights of
m
this procedure is developed based on the following relation for
individual user m
s.t.,
X
|Ωm |
(14)
pni gin = pmax
m
X
1+
i∈Mn ,i6=m
.
(15)
pni gin
n∈Ωm
pnm
n∈Ωm
6:
7:
i∈Mn ,i6=m
pnm ← 0, ∀m ∈ Mn , ∀n ∈ N.
repeat
for m ∈ M do
X
n
arg max
log2 1 + pnm gm
+
{pn
m }n∈Ωm
(12)
+ g1n
m
X
Algorithm 2 The iterative water-filling algorithm to calculate
optimal {pnm }m∈Mm ,n∈N .
4:
Taking the derivative of the Lagrangian in (12) with respect
to pnm , m ∈ Mn , n ∈ N, we obtain
pmax
m
where the objective function is the simplified version of the
objective function in (6). This is actually a multi-user waterfilling problem. Using the dual decomposition method [28],
the solution of this problem is described as follows. Taking
the dual variables λm , m ∈ M, the Lagrangian of the problem
in (11) can be written as
L({pnm }m∈Mn ,n∈N )
X
8:
X
i∈Mn ,i6=m
≤ pmax .
m
pni gin ,
end for
Update the values of previous {pnm }m∈Mn ,n∈N by optimal {pnm }m∈Mn ,n∈N obtained from this iteration.
until The performance improvement is not possible
Algorithm 3 The iterative process to calculate optimal λm for
user m in Algorithm 2.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
The elements of Am is sorted in the ascending order.
j ← 1, Gn ← 0, Gd ← 0.
repeat
Gn ← Gn + 1.
Gd ← Gd + Am (j).
Gn
.
λm (j) ← pmax
+Gd
m
if λm (j) ≥ 1/Am (j + 1) then
λ∗m ← λm (j).
Break the loop.
end if
j ← j + 1.
until j ≤ |Ωm |
2) Geometric Programming: Another way to write the
problem in (11) is as follows
min
{pn
m }m∈Mn ,n∈N
log2 Y
(1 +
n∈N
≈
min
{pn
m }m∈Mn ,n∈N
Y
n∈N
s.t.,
X
n∈Ωm
pnm
(1 +
1
X
n
pnm gm
)
m∈Mn
1
X
n
pnm gm
)
,
m∈Mn
≤ pmax
, ∀m ∈ M.
m
(16)
GP [31], [35] is an optimization technique that can solve
some non-convex problem by adopting some transformation
on the optimization variables. The objective and constraint
functions with which GP deals are posynomials and monomials. The objective function in (16) is the ratio of two posyn-
9
omials. The ratio of two posynomials is not a posynomial2,
and hence this problem is still not amenable to GP. However,
there are some heuristics, such as single condensation method,
double condensation method [31] that can be used to make the
problem amenable to GP. We have adopted single condensation
method to solve this problem. According to this method, the
denominator (which is a posynomial) has to be approximated
by some monomial. If we denote the denominator of the optimization problem by G({pnm }m∈Mn ,n∈N ), the approximated
monomial of this function is
G({pnm }m∈Mn ,n∈N ) =
≈λ
Y
Y
(1 +
n∈N
X
n
pnm gm
)
m∈Mn
n
(pnm )am .
(17)
m∈Mn ,n∈N
where anm , m ∈ Mn , n ∈ N and λ are auxiliary variables.
Given the values of {pnm }m∈Mn ,n∈N , the values of the auxiliary
variables can be obtained as follows.
anm =
pnm
n
G({pm }m∈Mn ,n∈N )
λ=
∂G({pnm }m∈Mn ,n∈N )
,
∂pnm
(18)
G({pnm }m∈Mn ,n∈N )
Y
n ,
(pnm )am
(19)
m∈Mn ,n∈N
and
∂G({pnm }m∈Mn ,n∈N )
n
= gm
∂pnm
Y
(1 +
n′ ∈N,n′ 6=n
X
′
′
n
pnm′ gm
′ ).
m′ ∈Mn′
(20)
Given some initial values of {pnm }m∈Mn ,n∈N , we require
an iterative process in order to obtain the optimal values of
these variables step by step. The steps of this iterative process
are provided in Section III of [36]. The final values of the
variables pnm , m ∈ Mn , n ∈ N, obtained in the last iteration of
the iterative process, is the solution of our defined optimization
problem.
IV. P ERFORMANCE E VALUATION
In this section, we evaluate the performance of our proposed
uplink resource allocation scheme via extensive simulation.
The detailed system setup and simulation settings are provided
in the following subsection. Then, we exhibit the results
obtained from the conducted simulation to verify the effectiveness of our scheme.
A. Simulation Setup
The cellular network, that we consider, has a circle-like
shape. Since our proposed resource allocation scheme did not
capture the interference from neighboring cells, we assume
that the cellular network is isolated. The base station is located
at the center of the cell, and the users are uniformly distributed
in a circular range with 500 m radius. We set the minimum
2 The
ratio of a posynomial and a monomial is a posynomial.
distance between users to 40 m, and the minimum distance
from the users to the base station to 50 m. As mentioned
previously, time is divided into frames. Each time frame is
equivalent to 1 s, and during this frame, spectrum is subdivided
among 20 subchannels and these are available to be allocated
among M users.
Each subchannel is assumed to have 200 KHz bandwidth.
According to [37], the theoretical limit of the channel capacity
, where Pb denotes the Bit Error Rate
is given by ln−1.5
(5Pb )
(BER). BER for the channel is configured as 106 . The channel
between the base station and a user is affected by shadow
and Rayleigh fading. Shadowing effect follows log-normal
distribution with variance 3.76. In order to calculate lognormal shadowing effect of a subchannel, we assume the
reference distance as 1 km and the SNR for this reference
distance is 28 dB. Reference shadowing effect has also the log
normal distribution with variance 3.76. Rayleigh fading effect
for each user over a subchannel follows Rayleigh distribution
with zero mean and 10 scale factor. Using all these parameters,
the gain of each subchannel for a user towards the base station
is computed following (22) in [23]. The maximal power level
of all users is set to 30 W. SC-coded signal on each subchannel
at the base station is decoded following the SIC technique
in [12], [38].
In addition to implement our proposed resource allocation
scheme, we have implemented relevant other algorithms [24],
[26] proposed in the literature already. For example, in [24],
the authors proposed two heuristics in order to maximize the
overall capacity and proportional fairness across the system.
In the figures demonstrated in the following subsection, these
are referred by Alg. 1 [24] and Alg. 2 [24]. The algorithm
proposed in [26] is referred by Alg. [26]. Since we have
adopted two techniques in order to allocate power to all
subchannels across all users, while referring our algorithm,
we use IWF and GP for iterative water-filling and GP power
allocation schemes, respectively. Furthermore, obtaining the
optimal subchannel-user mapping information is computationally intensive for a large-scale system, and hence we apply
brute-force search for a system with 10 and 20 users. In
the following subsection, for each data point, we conduct the
simulation over 10000 time frames.
B. Simulation Results
In Fig. 4, we show the increasing system capacity with
the increasing number of users. This is a natural trend. The
more the number of users in the system the more the overall
capacity. In this figure, we set K to 6. No matter the number of
users in the system, our proposed scheme always outperforms
existing other algorithms. Actually the algorithms proposed
in [24] have an assumption that each user can get only
one subchannel, which is the main reason of such degraded
performance compared to other algorithms. If one user obtains
only one subchannel, that assumption fails to exploit the multiuser-channel diversity of wireless systems, and consequently
the overall capacity of the system is much lower. On the
other hand, the total capacity obtained by the algorithm that
is designed to maximize the overall capacity should be larger
10
System Throughput (Mbps)
120
100
80
60
Alg. 1 [24]
Alg. 2 [24]
Alg. [26]
Our Alg. w. IWF
Our Alg. w. GP
Brute-Force w. IWF
40
20
10
20
40
60
80
100
Number of Users
Fig. 4: Comparison of system throughput with the increasing
number of users.
85.5
85
System Throughput (Mbps)
84.5
Alg. [26]
Our Alg. w. IWF
Our Alg. w. GP
84
83.5
83
82.5
82
81.5
81
80.5
80
2
4
6
8
10
12
Maximum Allowable Number of Users on a Subchannel (K)
Fig. 5: Comparison of system throughput with the increasing
value of K.
than that obtained by the algorithm which is designed to
maximize the proportional fairness. Fairness of the system is
always achieved by sacrificing the capacity of the system.
The algorithm proposed in [26] has very close performance
compared to ours. This algorithm is designed based on iterative water-filling algorithm. The idea of the algorithm is
as follows. First, it is assumed that all users are allocated
to all subchannels. On this setup, the iterative water-filling
algorithm is applied for the power allocation. Obviously, the
subchannel-user slot which has the worst gain, obtains the
least amount of power. Consequently, the corresponding user
is unallocated from the corresponding subchannel. Then, again
iterative water-filling algorithm is applied until the constraint
that, each subchannel obtains exactly K number of users,
is met. At this point, this algorithm is terminated. From the
nature and design, it is obvious that the algorithm only keeps
the users to a subchannel which have relatively better gain
compared to other users. Better gain of a subchannel for a
user implies, that user obtains more power on that subchannel.
However, if a subchannel has a number of allocated users with
relatively better level of power, it does not necessarily enhance
the capacity of that subchannel. This is because SIC technique
considers the power level of other users as interference level
while decoding the signal of one user. As a result, such
mechanism of the algorithm does not enhance the system
capacity. There must be some other algorithm that should solve
the subchannel-user mapping problem in such a way that the
drawback of this algorithm can be overcome.
Consequently, we have proposed an efficient algorithm to
solve the subchannel-user mapping problem based on the
many-to-many matching model. The algorithm is designed
in such a way that in every iteration, the system capacity
is enhanced little by little. The algorithm terminates only
when the performance of the system cannot be enhanced
anymore. Based on the first requirement of many-to-many
matching model, each user constructs its preference list in
the descending order of the received power achieved from the
subchannels. Then, each user only wants to obtain its most
preferred subchannel, and the corresponding subchannel either
adds this user or substitutes the existing user if and only if
the sum-capacity of that subchannel and other affected subchannels is enhanced. Note that while solving the subchanneluser mapping problem, it is assumed that each user subdivides
its maximal power equally among its allocated subchannels.
Furthermore, even if the value of K is large, exactly K number
of users allocated to a subchannel (especially the one with
worse channel) not necessarily enhances the system throughput. The algorithm adds additional user to a subchannel if
that user improves the sum-throughput of that subchannel and
other affected subchannels. All these design mechanisms allow
the algorithm to overcome the drawbacks of the algorithm
in [26], and outperforms it. Although for the power allocation
of our scheme, we have adopted two techniques, iterative
water-filling algorithm outperforms GP technique. Moreover,
the way we have implemented iterative water-filling algorithm
such that it has much less computational complexity compared
to the other one. Therefore, we recommend iterative waterfilling as the power allocation technique for our proposed
resource allocation scheme.
For M = 40 in the system, Fig. 5 depicts the increasing
system capacity with the increasing value of K. It is apparent
that the more the number of users in a subchannel, the more
enhanced the system capacity is. However, not necessarily
the same K number of users on every subchannel enhances
the system capacity. This phenomenon particularly happens
for the subchannel which has worse condition due to the
subdivision of limited power level of each user among its
allocated subchannels. Therefore, our algorithm assigns less
number of users to some subchannels whenever necessary.
However, the algorithm in [26] blindly assigns exactly K
number of users to each subchannel without giving attention
to the performance of the system. This statement has been
well-proved in this figure.
The evidence that our algorithm not necessarily assigns
exactly K number of users to all subchannels is strengthened
in Fig. 6. In this figure, K is set to 8. The subchannel
which has relatively better gain can have larger number of
users compared to other subchannels. On the other hand,
the algorithm in [26] always assigns K number of users to
all subchannels. The algorithms in [24] assign the user to
a subchannel which has the lowest interference level and
11
12
Algs. [24]
Alg. [26]
Our Alg.
Brute-Force
8
10
6
Alg. 1 [24]
Alg. 2 [24]
Alg. [26]
Our Alg.
5
4
Resource Efficiency
Number of Allocated Users
7
8
6
4
3
2
2
1
10
20
40
60
80
100
Number of Users
0
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16
Subchanel Index
Fig. 6: Comparison of allocated users to each individual
subchannel when M = 40.
Fig. 9: Comparison of resource efficiency with the increasing
number of users.
Algs. [24]
Alg. [26]
Our Alg.
6
30
Alg. 1 [24]
Alg. 2 [24]
Alg. [26]
Our Alg. w. IWF
Our Alg. w. GP
User Throughput (Mbps)
25
20
Resource Efficiency
5
4
3
2
15
1
2
10
4
6
8
10
12
Maximum Allowable Number of Users on a Subchannel (K)
Fig. 10: Comparison of resource efficiency with the increasing
value of K.
5
0
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15
Users with the Increasing Distance
Fig. 7: Comparison of throughput for each individual user
when M = 40.
18
Alg. 1 [24]
Alg. 2 [24]
Alg. [26]
Our Alg.
Number of Allocated Subchannels
16
14
12
10
8
6
4
2
0
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16
User Index
Fig. 8: Comparison of allocated subchannels to each individual
user when M = 40.
each user only obtains one subchannel. Therefore, these two
algorithms also assign less number of users compared to the
value of K.
In Fig. 7 and Fig. 8, we provide some detailed information
about our algorithm to justify its superiority. In these figures,
K is set to 8 as well. Since our algorithm is based on
the consideration to maximize the system capacity, the user
who has the best gain on all subchannels, should have the
highest throughput. At the same time, since Rayleigh fading
effect is statistically similar for all users over all subchannels,
the user closest to the base station should have the highest
throughput, and so thus observed in Fig. 7. There are some
exceptions as well due to the random nature of Rayleigh
fading effect, such as users 9, 10, 11 and 12. Other than some
exceptions, the throughput of the users have the decreasing
trend with the incresing distance from the base station. Due to
the aforementioned justifications, other algorithms incur less
throughput for all users except the one in [26]. Previously,
we claimed that our proposed algorithm is not necessarrily
globally optimal, and therefore the algorithm in [26] incurs
larger throughput for one or two users. However, it is obvious
in Fig. 7 that the throughput of more number of users incurred
by our algorithm are better compared to the one in [26]. One of
12
the design insights of our proposed resource allocation scheme
is to exploit multi-user-channel diversity of wireless systems.
Consequently, by assigning more number of subchannels to
each user, it is possible to enhance the system throughput, and
so thus our algorithm does (as depicted in Fig. 8). However, the
algorithms in [24] assign only one subchannel to a user, and
hence their performance is much worse compared to others.
In Fig. 9 and Fig. 10, we further justify that our algorithm is superior in terms of resource efficiency as well. In
these two figures, we plot the ratio of total subchannel-user
slots and the number of used subchannel-user slots with the
increasing number of users and with the increasing value
of K, respectively. If the number
of used subchannel-user
P
slots is denoted by D (i.e., n∈N |Mn |), the metric on the
Y-axis of these two figures is KN
D . It is obvious that the
more the number of users and the larger the value of K,
the more subchannels are used to support more users and
to enhance the system throughput. Previously, we observed
that the overall throughput achieved by the algorithm in [26]
has close performance to ours, however using more resource
slots in the system as depicted in Fig. 9 and Fig. 10. On the
other hand, although the algorithms in [24] have the highest
resource efficiency, they have the worst overall performance
as proved in the previous results. These algorithms incur the
highest performance in this case because of assigning only
one subchannel to each user. Whereas, to exploit multi-usersubchannel diversity of wireless systems, our scheme assigns
more subchannels to the users, and consequently achieves the
highest overall performance.
V. C ONCLUSION
AND
F UTURE W ORK
In this paper, we proposed an uplink resource allocation
scheme of a NOMA system, in which the spectrum is divided
into multiple subchannels. The objective of our resource
allocation scheme is to allocate power and subchannels across
the users of the system. Due to the discrete nature of subchannels and the characteristics of NOMA systems, the problem
is NP-hard and non-convex. Since the optimal solution of
the problem is intractable, we solved the problem in two
steps. First, the subchannel-user mapping problem was solved
using many-to-many matching model. Then, iterative waterfilling algorithm and GP technique were applied to allocate
power optimally across all assigned subchannels and users
of the system. We conducted extensive simulation to verify
the effectiveness of our proposed resource allocation scheme
comparing with other existing works in the literature.
One of our previous experience [39] while dealing with
interference is, reducing the allocated power level on the
subchannels may bring better performance for the system
compared to the case when the maximal power of the users
are used. This is what exactly observed while allocating
subchannels and power among the users in an OFDMAbased network surrounded by many neighboring cells. In such
systems, transmission of users in one cell causes interference
for the transmission of users in other cells. Since in NOMA
systems, multiple users are superimposed on the same subchannel, power level of one user may cause interference for
other users. Both water-filling algorithm and GP technique use
full power of a user to its allocated subchannels. However,
we believe that using less power, it might be possible to
enhance the system performance in terms of both capacity and
resource efficiency. Another assumption of our work is, the
network for which we proposed the uplink resource allocation
scheme, is isolated and does not have any neighboring cell.
Such assumption of the network is equivalent to ignoring
the interference from neighboring cells. However, interference
from the neighboring cells is a crucial factor especially for the
uplink case in this context. As of our future work, we would
like to continue the research in this direction in order to obtain
more promising and useful results.
R EFERENCES
[1] S. Tekinay, “Next generation wireless networks,” Springer US, 2002.
[2] S. C. Borst, A. Buvaneswari, L. M. Drabeck, M. J. Flanagan, J. M.
Graybeal, G. K. Hampel, M. Haner, W. M. MacDonald, P. A. Polakos,
G. Rittenhouse, I. Saniee, A. Weiss, and P. A. Whiting, “Dynamic
optimization in future cellular networks,” Bell Labs Technical Journal,
vol. 10, pp. 99–119, 2005.
[3] A. Agarwal and K. Agarwal, “The next generation mobile wireless
cellular networks: 4G and beyond,” American Journal of Electrical and
Electronic Engineering, vol. 2, no. 3, pp. 92–97, 2014.
[4] A. Ali, W. Hamouda, and M. Uysal, “Next generation
M2M cellular networks: Challenges and practical considerations,” CoRR, vol. abs/1506.06216, 2015. [Online]. Available:
http://arxiv.org/abs/1506.06216
[5] S. Ghosh, K. Basu, and S. K. Das, “An architecture for next-generation
radio access networks,” IEEE Netw., vol. 19, no. 5, pp. 35–42, Sept
2005.
[6] V. Suryaprakash and I. Malanchini, “Reliability in future radio access
networks: From linguistic to quantitative definitions,” in Proc. IEEE
IWQoS, June 2016, pp. 1–2.
[7] T. Edler and S. Lundberg, “Energy efficiency enhancements in radio
access networks,” Ericsson Review, vol. 21, no. 1, 2014.
[8] K. Higuchi and Y. Kishiyama, “Non-orthogonal access with random
beamforming and intra-beam SIC for cellular MIMO downlink,” in Proc.
IEEE VTC Fall, 2013, pp. 1–5.
[9] Y. Endo, Y. Kishiyama, and K. Higuchi, “Uplink non-orthogonal access
with MMSE-SIC in the presence of inter-cell interference,” in Proc.
ISWCS, Aug 2012, pp. 261–265.
[10] J. Umehara, Y. Kishiyama, and K. Higuchi, “Enhancing user fairness
in non-orthogonal access with successive interference cancellation for
cellular downlink,” in Proc. IEEE ICCS, Nov 2012, pp. 324–328.
[11] N. Otao, Y. Kishiyama, and K. Higuchi, “Performance of non-orthogonal
access with SIC in cellular downlink using proportional fair-based
resource allocation,” in Proc. ISWCS, Aug 2012, pp. 476–480.
[12] N. I. Miridakis and D. D. Vergados, “A survey on the successive
interference cancellation performance for single-antenna and multipleantenna OFDM systems,” IEEE Commun. Surveys Tutorials, vol. 15,
no. 1, pp. 312–335, Jan 2013.
[13] Y. Saito, A. Benjebbour, Y. Kishiyama, and T. Nakamura, “System-level
performance evaluation of downlink non-orthogonal multiple access
(NOMA),” in Proc. IEEE PIMRC, June 2013, pp. 611–615.
[14] A. Benjebbour, A. Li, Y. Saito, Y. Kishiyama, A. Harada, and T. Nakamura, “System-level performance of downlink NOMA for future LTE
enhancements,” in Proc. IEEE GLOBECOM Workshops, June 2013, pp.
611–615.
[15] M. R. Hojeij, J. Farah, C. A. Nour, and C. Douillard, “Resource
allocation in downlink non-orthogonal multiple access (NOMA) for
future radio access,” in Proc. IEEE VTC Spring, May 2015, pp. 1–6.
[16] P. Parida and S. S. Das, “Power allocation in OFDM based NOMA
systems: A DC programming approach,” in Proc. IEEE Globecom
Workshops, Dec 2014, pp. 1026–1031.
[17] N. Vucic, S. Shi, and M. Schubert, “DC programming approach for
resource allocation in wireless networks,” in Proc. International Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless
Networks, May 2010, pp. 380–386.
[18] S. Han, C. L. I, Z. Xu, and Q. Sun, “Energy efficiency and spectrum
efficiency co-design: From NOMA to network NOMA,” IEEE Commun.
Society MMTC E-Lett., vol. 9, no. 5, pp. 21–24, Sept 2013.
13
[19] Q. Sun, S. Han, C. L. I, and Z. Pan, “Energy efficiency optimization for
fading MIMO non-orthogonal multiple access systems,” in Proc. IEEE
ICC, June 2015, pp. 2668–2673.
[20] F. Fang, H. Zhang, J. Cheng, and V. C. M. Leung, “Energy-efficient resource allocation for downlink non-orthogonal multiple access network,”
IEEE Trans. Commun., vol. 64, no. 9, pp. 3722–3732, Sept 2016.
[21] J. Huang, V. G. Subramanian, R. Agrawal, and R. Berry, “Joint scheduling and resource allocation in uplink OFDM systems for broadband
wireless access networks,” IEEE J. Sel. A. Commun., vol. 27, no. 2, pp.
226–234, Feb 2009.
[22] R. Ruby and V. Leung, “Uplink scheduling solution for enhancing
throughput and fairness in relayed long-term evolution networks,” IET
Commun., vol. 6, no. 8, pp. 813–825, 2014.
[23] R. Ruby, V. C. M. Leung, and D. G. Michelson, “Uplink scheduler for
SC-FDMA-based heterogeneous traffic networks with QoS assurance
and guaranteed resource utilization,” IEEE Trans. Veh. Technol., vol. 64,
no. 10, pp. 4780–4796, Oct 2015.
[24] M. Mollanoori and M. Ghaderi, “Uplink scheduling in wireless networks
with successive interference cancellation,” IEEE Trans. Mobile Comput.,
vol. 13, no. 5, pp. 1132–1144, May 2014.
[25] K. Kumaran and L. Qian, “Scheduling on uplink of CDMA packet
data network with successive interference cancellation,” in Proc. IEEE
WCNC, vol. 3, March 2003, pp. 1645–1650 vol.3.
[26] M. Al-Imari, P. Xiao, and M. A. Imran, “Receiver and resource allocation optimization for uplink NOMA in 5G wireless networks,” in Proc.
ISWCS, Aug 2015, pp. 151–155.
[27] W. Yu, W. Rhee, S. Boyd, and J. M. Cioffi, “Iterative water-filling for
gaussian vector multiple-access channels,” IEEE Trans. Inform. Theory,
vol. 50, no. 1, pp. 145–152, Jan 2004.
[28] D. P. Bertsekas, “Nonlinear programming,” Athena Scientific, 1999.
[29] K. Hamidouche, W. Saad, and M. Debbah, “Many-to-Many matching
games for proactive social-caching in wireless small cell networks,” in
Proc. IEEE WiOpt, May 2014, pp. 569–574.
[30] A. Roth, “Stability and polarization of interests in job matching,”
Econometrica, vol. 52, no. 1, pp. 47–57, 1984.
[31] M. Chiang, “Geometric programming for communication systems,”
Commun. Inf. Theory, vol. 2, no. 1/2, pp. 1–154, Jul. 2005.
[32] D. Tse and P. Viswanath, “Fundamentals of wireless communication,”
Cambridge University Press, 2005.
[33] K. McClaning, “Radio receiver design,” Noble Publishing Corporation,
2000.
[34] M. Hazewinkel, “Transmission rate of a channel,” Encyclopedia of
Mathematics, Springer, 2001.
[35] S. Boyd, S.-J. Kim, L. Vandenberghe, and A. Hassibi, “A tutorial on
geometric programming,” Optimization and Engineering, vol. 8, no. 1,
p. 67, 2007.
[36] R. Ruby, V. C. M. Leung, and D. G. Michelson, “Centralized and game
theoretical solutions of joint source and relay power allocation for AF
relay based network,” IEEE Trans. Commun., vol. 63, no. 8, pp. 2848–
2863, Aug 2015.
[37] X. Qiu and K. Chawla, “On the performance of adaptive modulation in
cellular systems,” IEEE Trans. Commun., vol. 47, no. 6, pp. 884–895,
Jun 1999.
[38] S. Vanka, S. Srinivasa, Z. Gong, P. Vizi, K. Stamatiou, and M. Haenggi,
“Superposition coding strategies: Design and experimental evaluation,”
IEEE Trans. Wirel. Commun., vol. 11, no. 7, pp. 2628–2639, July 2012.
[39] L. Zhou, R. Ruby, H. Zhao, X. Ji, J. Wei, and V. Leung, “A graph-based
resource allocation scheme with interference coordination in small cell
networks,” in Proc. IEEE Globecom Workshops, 2014, pp. 211–218.
| 7cs.IT
|
A semiparametric single-index estimator for a class of
estimating equation models
arXiv:1608.04244v2 [math.ST] 26 Apr 2017
Marian Hristache∗ Weiyu Li† Valentin Patilea‡
Abstract
We propose a two-step pseudo-maximum likelihood procedure for semiparametric
single-index regression models where the conditional variance is a known function of
the regression and an additional parameter. The Poisson single-index regression with
multiplicative unobserved heterogeneity is an example of such models. Our procedure is
based on linear exponential densities with nuisance parameter. The pseudo-likelihood
criterion we use contains a nonparametric estimate of the index regression and therefore a rule for choosing the smoothing parameter is needed. We propose an automatic
and natural rule based on the joint maximization of the pseudo-likelihood with respect to the index parameter and the smoothing parameter. We derive the asymptotic
properties of the semiparametric estimator of the index parameter and the asymptotic
behavior of our ‘optimal’ smoothing parameter. The finite sample performances of our
methodology are analyzed using simulated and real data.
Keywords: semiparametric pseudo-maximum likelihood, single-index model, linear exponential densities, bandwidth selection.
∗
CREST (Ensai), email: [email protected]
Corresponding author. CREST (Ensai), email: [email protected]
‡
CREST (Ensai), email: [email protected]. Valentin Patilea gratefully acknowledges support from
the research program New Challenges for New Data of Genes, LCL and Fondation de Risque.
†
1
1
Introduction
In this paper we consider semiparametric models defined by conditional mean and conditional variance estimating equations. Models defined by estimating equations for the first
and second order conditional moments are widely used in applications. See, for instance,
Ziegler (2011) for a recent reference. Here we consider a model that extends the framework
considered by Cui, Härdle and Zhu (2011).
To provide some insight on the type of models we study, consider the following semiparametric extension of the classical Poisson regression model with unobserved heterogeneity:
T
the observed variables are Y, Z T where Y denotes the count
variable and Z is the vector
of d explanatory variables. Let r (t; θ) = E Y | Z T θ = t . We assume that there exists
θ0 ∈ Rd such that
E (Y | Z) = E Y | Z T θ0 = r Z T θ0 ; θ0 .
The parameter θ0 and the function r are unknown. Given Z and an unobserved error term ε,
the variable Y has a Poisson law of mean r Z T θ0 ; θ0 ε. If E (ε | Z) = 1 and V ar (ε | Z) = σ 2 ,
then
V ar (Y | Z) = V ar (E (Y | Z, ε) | Z) + E (V ar (Y | Z, ε) | Z)
= r Z T θ0 ; θ0 1 + σ 2 r Z T θ0 ; θ0 .
(1.1)
This model is a semiparametric single-index regression model (e.g., Powell, Stock and Stoker
(1989), Ichimura (1993), Härdle, Hall and Ichimura (1993), Sherman (1994b)) where a second
order conditional moment is specified as a nonlinear function of the conditional mean and an
additional unknown parameter. This extends the framework of Cui, Härdle and Zhu (2011)
where the conditional variance of the response is proportional to a given function of the
conditional mean.
Our first contribution is to propose a new semiparametric estimation procedure for singleindex regression which incorporates the additional information on the conditional variance
of Y . For this we extend the quasi-generalized pseudo maximum likelihood method introduced by Gouriéroux, Monfort and Trognon (1984a, 1984b) to a semiparametric framework.
More precisely, we propose to estimate θ0 and the function r(·) through a two-step pseudomaximum likelihood (PML) procedure based on linear exponential families with nuisance
parameter densities. Such densities are parameterized by the mean r and a nuisance parameter that can be recovered from the variance. Although we use a likelihood type criterion,
no conditional distribution assumption on Y given Z is required for deriving the asymptotic
results.
As an example of application of our procedure consider the case where Y is a count
variable. First, write the Poisson likelihood where the function r(·) is replaced by a kernel
estimator and maximize this likelihood with respect to θ to obtain a semiparametric PML
estimator of θ0 . Use this estimate and the variance formula (1.1) to deduce a consistent
moment estimator of σ 2 . In a second step, estimate θ0 through a semiparametric Negative
Binomial PML where r is again replaced by a kernel estimator and the variance parameter
of the Negative Binomial is set equal to the estimate of σ 2 . Finally, given the second step
estimate of θ0 , build a kernel estimator for the regression r(·). For simplicity, we use a
2
Nadaraya-Watson estimator to estimate r(·). Other smoothers like local polynomials could
be used at the expense of more intricate technical arguments.
The occurrence of a nonparametric estimator in a pseudo-likelihood criterion requires a
rule for the smoothing parameter. While the semiparametric index regression literature contains a large amount of contributions on how to estimate an index, there are much less results
and practical solutions on the choice of the smoothing parameter. Even if the smoothing
parameter does not influence the asymptotic variance of a semiparametric estimator of θ0 ,
in practice the estimate of θ0 and of the regression function may be sensitive to the choice
of the smoothing parameter.
Another contribution of this paper is to propose an automatic and natural choice of the
smoothing parameter used to define the semiparametric estimator. For this, we extend the
approach introduced by Härdle, Hall and Ichimura (1993) (see also Xia and Li (1999), Xia,
Tong and Li (1999) and Delecroix, Hristache and Patilea (2006)). The idea is to maximize the
pseudo-likelihood simultaneously in θ and the smoothing parameter, that is the bandwidth
of the kernel estimator. The bandwidth is allowed to belong to a large range between n−1/4
and n−1/8 . In some sense, this approach considers the bandwidth an auxiliary parameter for
which the pseudo-likelihood may provide an estimate. Using a suitable decomposition of the
pseudo-log-likelihood we show that such a joint maximization is asymptotically equivalent
to separate maximization of a purely parametric (nonlinear) term with respect to θ and
minimization of a weighted (mean-squared) cross-validation function with respect to the
bandwidth. The weights of this cross-validation function are given by the second order
derivatives of the pseudo-log-likelihood with respect to r. We show that the rate of our
‘optimal’ bandwidth is n−1/5 , as expected for twice differentiable regression functions.
The paper is organized as follows. In section 2 we introduce a class of semiparametric
PML estimators based on linear exponential densities with nuisance parameter and we provide a natural bandwidth choice. Moreover, we present the general methodology used for
the asymptotics. Section 3 contains the asymptotic results. A bound for the variance of
our semiparametric PML estimators is also derived. In section 4 we use the semiparametric
PML estimators to define a two-step procedure that can be applied in single-index regression
models where an additional variance condition like (1.1) is specified. Section 5.1 examines
the finite-sample properties of our procedure via Monte Carlo simulations. We compare
the performances of a two-step generalized least-squares with those of a Negative Binomial
PML in a Poisson single-index regression model with multiplicative unobserved heterogeneity. Even if the two procedures considered lead to asymptotically equivalent estimates, the
latter procedure seems preferable in finite samples. An application to real data on the frequency of recreational trips (see Cameron and Trivedi (2013), page 246) is also provided.
Section 6 concludes the paper. The technical proofs are postponed to the Appendix.
2
Semiparametric PML with nuisance parameter
T
T
Consider that the observations Y1 , Z1T , ..., Yn , ZnT
are independent copies of the ran
T T
d
dom vector Y, Z
∈ R × R . Assume that there exists θ0 ∈ Rd , unique up to a scale
3
normalization factor, such that the single-index model (SIM) condition
E (Y | Z) = E Y | Z T θ0 = r Z T θ0 ; θ0
(2.1)
holds. In this paper, we focus on single-index models where the conditional second order
moment of Y given Z is a known function of E [Y | Z] and of a nuisance parameter. To be
more precise, in the model we consider,
V ar (Y | Z) = g (E (Y | Z) , α0 ) = g r Z T θ0 ; θ0 , α0 ,
(2.2)
for some real value α0 . The function g (·, ·) is known and, for each r, the map α → g (r, α) is
one-to-one. Our framework is slightly more general that the one considered by Cui, Härdle
and Zhu (2011) where the conditional variance of Y given Z is a given function of the
conditional mean of Y given Z multiplied by an unknown constant.
To estimate the parameter of interest θ0 in a model like (2.1)-(2.2), we propose a semiparametric PML procedure based on linear exponential families with nuisance parameter.
The density used to build the pseudo-likelihood is taken with mean and variance equal to
r and g(r, α), respectively. In this section we suppose that an estimator of the nuisance
parameter is given. In section 4 we show how to build such an estimator using a preliminary
estimate of θ0 and condition (2.2).
2.1
Linear exponential families with nuisance parameter
Gouriéroux, Monfort and Trognon (1984a) introduced a class of densities, with respect to
a given measure µ, called linear exponential family with nuisance parameter (LEFN) and
defined as
l (y | r, α) = exp [B (r, α) + C (r, α) y + D (y, α)] ,
where α is the nuisance parameter. Since the dominating measure µ need not be Lebesgue
measure, the law defined by l is not necessarily continuous. The functions B (·, ·) and C (·, ·)
are such that the expectation of the corresponding law is r while the variance is [∂r C (r, α)]−1 .
(∂r denotes the derivative with respect to the argument r.) Recall that for any given α, the
following identity holds:
∂r B (r, α) + ∂r C (r, α) r ≡ 0.
If α is fixed, a LEFN becomes a linear exponential family (LEF) of densities. Gouriéroux,
Monfort and Trognon (1984a, 1984b) used LEFN densities to define a two-step PML procedure in nonlinear regression models where a specification of the conditional variance is given.
Herein, we extend their approach to a semiparametric framework.
In the case of the SIM defined by equation (1.1), the conditional variance is given by
g (r, α) = r (1 + αr) with r and α > 0. In this case take
1
B (r, α) = − ln (1 + αr)
α
and
C (r, α) = ln
r
,
1 + αr
which define a Negative Binomial distribution of mean r and variance r (1 + αr). Note that
the limit case α = 0 corresponds to a Poisson distribution. As another example, consider
g (r, α) = r2 /α with r and α > 0. Now, take the LEFN density given by B (r, α) = −α ln r
and C (r, α) = −α/r, which is the density of a gamma law of mean r and variance r2 /α.
4
2.2
The semiparametric estimator
In order to define our semiparametric PML estimator in the presence of a nuisance parameter
let us introduce some notation: given {cn } , a sequence of numbers growing slowly to infinity
(e.g., cn = ln n), let
−1/8
Hn = h : cn n−1/4 ≤ h ≤ c−1
n n
be the range from which the ‘optimal’ bandwidth will be chosen. Define the set Θn =
{θ : kθ − θ0 k ≤ dn }, n ≥ 1, with {dn } some sequence decreasing to zero.
Let α∗ be some real value of the nuisance parameter. Typically, α∗ = α0 if the conditional
variance formula (2.2) is correctly specified. Otherwise, α∗ is some pseudo-true value of the
nuisance parameter. Suppose that a sequence {e
αn } such that α
en → α∗ , in probability, is
given. Set1
ψ (y, r; α) = ln l (y | r, α)
with l (y | r, α) the LEFN density of expectation r and nuisance parameter α. Define the
semiparametric PML estimator in the presence of a nuisance parameter and the optimal
bandwidth as
n
1X
b
b
θ, h = arg max
ψ Yi , r̂hi ZiT θ; θ ; α
en τn (Zi ),
θ∈Θn , h∈Hn n i=1
where
1
n−1
r̂hi (t; θ) =
Yj Kh t − ZjT θ
j6=i
P
1
Kh t − ZjT θ
n−1
P
j6=i
(2.3)
=:
γ
bhi (t; θ)
fbi (t; θ)
h
denotes the leave-one-out version of the Nadaraya-Watson estimator of the regression function
γ (t; θ)
,
r (t; θ) = E Y |Z T θ = t =:
f (t; θ)
with f (·; θ) the density of Z T θ. The function K (·) is a second order kernel function and Kh (·)
stands for K (·/h) /h, where h is the bandwidth. τn (·) denotes a trimming function. If the
sequence α
en is constant or ψ does not depend on α, equation (2.3) defines a semiparametric
PML based on a LEF density.
A trimming is designed to keep the density estimator fbhi away from zero in computations
and it is usually required for analyzing the asymptotic properties of the nonparametric
regression estimator and of the ‘optimal’ bandwidth. The practical purpose of a trimming
recommends a data-driven device like I{z: fbi (zT θ;θ)≥c} (·), with some fixed c > 0 . (Herein,
h
IA (·) denotes the indicator function of the set A.) However, to ensure consistency with such
a trimming, one should require in addition that
θ0 = arg max E ψ Y, rθ Z T θ I{z: f (zT θ;θ)≥c} (Z) .
θ
1
Herein, we focus on ψ (y, r; α) = ln l (y | r, α) where l (y | r, α) = exp [B (r, α) + C (r, α) y + D (y, α)] is
a LEFN density. However, other functions ψ (y, r; α) having the required properties can be considered (see
Appendix A).
5
Meanwhile, a trimming like I{z: f (zT θ0 ;θ0 )≥c} (·) is easier to handle in theory. Here, we consider
τn (·) = I{z: fbi
hn (z
T θ ; θ )≥c
n n
} (·)
(2.4)
with θn ∈ Θn , n ≥ 1, a sequence with limit θ0 and hn , n ≥ 1, a sequence of preliminary bandwidths such that nε hn → 0 and n1/2−ε hn → ∞ for some 0 < ε < 1/2. The
trimming procedure we propose represents an appealing compromise between the theory
and the applications. On one hand, it is easy to implement. On the other hand, we show
below that, in a certain sense, our trimming is asymptotically equivalent to the fixed trimming I{z: f (zT θ0 ;θ0 )≥c} (·) and this fact greatly simplifies the proofs. We prove this equivalence under two types of assumptions: either i) Z is bounded and θn − θ0 = o (1), or ii)
E [exp(λ kZk)] < ∞, for some λ > 0, and θn −
ln n)
θ0 = o (1/
. To be more precise, define
T
d
δ
T
A = z : f z θ0 ; θ0 ≥ c ⊂ R and A = z : f z θ0 ; θ0 − c ≤ δ , δ > 0. By little
algebra, for all θ ∈ Θn , h and i,
I{z: fbi (zT θ;θ)≥c} (Zi ) − IA (Zi ) ≤ IAδ (Zi ) + I(δ,∞) (Gn ),
h
where
Gn = max
sup
1≤i≤n θ∈Θn , h
Let
fbhi ZiT θ; θ − f ZiT θ0 ; θ0 .
n
1X
b
ψ Yi , r̂hi ZiT θ; θ ; α
en IA (Zi )
S θ, h; α
en , A =
n i=1
with A = A or Aδ . Without loss of generality, consider that ψ (·, · ; ·) ≤ 0. (Since ψ is the
logarithm of a LEFN density, for any given y and α, the map r → ψ (y, r ; α) attains its
maximum at r = y; thus, up to a translation with a function depending only on y and α, we
may consider ψ ≤ 0.) In this case we have
n
1X
ψ Yi , r̂hi ZiT θ; θ ; α
en I{z: fbi (zT θn ;θn )≥c} (Zi ) − Sb (θ, h; α
en , A)
hn
n i=1
(2.5)
n
I(δ,∞) (Gn ) X
ψ Yi , r̂hi ZiT θ; θ ; α
en .
≤ −Sb θ, h; α
en , Aδ −
n
i=1
We show that Sb θ, h; α, Aδ = oP (Sb (θ, h; α,A)), uniformly over Θn × Hn and uniformly in
α, provided that δ → 0 and P f Z T θ0 ; θ0 = c = 0. On the other hand, we prove that
P (Gn > δ) → 0, provided that δ → 0 slowly enough and h → 0 faster than nε and slower
than n1/2−ε , for some 0 < ε < 1/2. (See Lemma B.2 in the appendix; in that lemma we
distinguish two types
of assumptions depending on whether Z is bounded or not.)
bb
Deduce that θ,
h is asymptotically equivalent to the maximizer of Sb (θ, h; α
en , A) over
Θn × Hn . Therefore, hereafter, we simply write Sb (θ, h; α
en ) instead of Sb (θ, h; α
en , A) and we
consider
bb
θ,
h = arg max Sb (θ, h; α
en ) .
(2.6)
θ∈Θn , h∈Hn
6
2.3
Methodology
The semiparametric pseudo-log-likelihood Sb (θ, h; α
en ) can be split into a purely paramete
ric (nonlinear) part S (θ; α
en ), a purely nonparametric one T (h; α∗ ) and a reminder term
R(θ, h; α
en ), where
n
1 X
en − ψ Yi , r ZiT θ0 ; θ0 ; α∗ IA (Zi ) ,
ψ Yi , r ZiT θ; θ ; α
Se (θ; α
en ) =
n i=1
(2.7)
n
1X
T (h; α ) =
ψ Yi , r̂hi ZiT θ0 ; θ0 ; α∗ IA (Zi ) ,
n i=1
∗
n
1 X
R (θ, h; α
en ) =
ψ Yi , r̂hi ZiT θ; θ ; α
en − ψ Yi , r ZiT θ; θ ; α
en IA (Zi )
n i=1
n
1 X
−
ψ Yi , r̂hi ZiT θ0 ; θ0 ; α∗ − ψ Yi , r ZiT θ0 ; θ0 ; α∗ IA (Zi )
n i=1
(see Härdle, Hall and Ichimura (1993) for a slightly different splitting). Given this decomposition, the simultaneous optimization of Sb (θ, h; α
en ) is asymptotically equivalent to
e
separately maximizing S (θ; α
en ) with respect to θ and T (h; α∗ ) with respect to h, provided
that R (θ, h; α
en ) is sufficiently small.
A key ingredient for proving that R (θ, h; α
en ) is negligible with respect to Se (θ; α
en ) and
∗
T (h; α ) , uniformly in (θ, h) ∈ Θn ×Hn and for any {e
αn } , is represented by the orthogonality
conditions
E ∂2 ψ Y, r Z T θ0 ; θ0 ; α | Z = 0
(2.8)
and
E ∂θ ∂2 ψ Y, r Z T θ0 ; θ0 ; α | Z T θ0 = 0,
(2.9)
that must hold for any α, where ∂2 denotes the derivative with respect to the second argument
of ψ (·, ·; ·) and ∂θ is the derivative with respect to all occurrences of θ, that is given y, z and
α,
∂
∂2 ψ y, r z T θ; θ ; α θ=θ0
∂θ ∂2 ψ y, r z T θ0 ; θ0 ; α =
∂θ
(see also Sherman (1994b) and Delecroix, Hristache and Patilea (2006) for similar conditions). If
ψ (y, r; α) = ln l (y | r, α) = B (r, α) + C (r, α) y + D (y, α) ,
with ∂r B (r, α) + ∂r C (r, α) r ≡ 0, then ∂2 ψ (y, r; α) = ∂r C (r, α) (y − r) and thus (2.8) is a
consequence of the SIM condition (2.1). To check the second orthogonality condition note
that
2
2
E ∂22
ψ Y, r Z T θ0 ; θ0 ; α | Z = E ∂22
ψ Y, r Z T θ0 ; θ0 ; α | Z T θ0
and
E ∂θ r Z T θ0 ; θ0 | Z T θ0 = E r0 Z T θ0 ; θ0 Z − E Z | Z T θ0 | Z T θ0 ,
7
where r0 (·; θ0 ) is the derivative of r(·; θ0 ). The last identity is always true under the SIM
condition (e.g., Newey (1994), page 1358). Let us point out that conditions (2.8)-(2.9) hold
even if the variance condition (2.2) is misspecified.
Since R (θ, h; α
en ) is negligible with respect to Se (θ; α
en ) and T (h; α∗ ) does not contain
the parameter of interest, the asymptotic distribution of θb will be obtained by standard
arguments used for M −estimators in the presence of nuisance parameters applied to the
objective function Se (θ; α
en ). We deduce that θb behaves as follows: i) if the SIM condition
(2.1) holds and α
en − α∗ = OP (1) , for some α∗ , then θb is asymptotically normal; ii) if SIM
condition holds, the conditional variance (2.2) is correctly specified and α
en − α0 = OP (1) ,
then θb is asymptotically normal and it has the lowest variance among the semiparametric
√
PML estimators based on LEF densities. In any case, the asymptotic distribution of n(θb−
θ0 ) does not depend on the choice of α
en . Let us point out that in our framework
√ we only
impose α
en convergent in probability without asking a rate of convergence OP (1/ n) , as it is
usually supposed for M −estimation
in the presence of nuisance
parameters. This because the
usual orthogonality condition E ∂α ∂θ ψ Y, r Z T θ0 ; θ0 ; α = 0 is true for any α, provided
that ψ (y, r; α) = ln l (y | r, α) with l (y | r, α) a LEFN density. Indeed, we have
E ∂α ∂θ ψ Y, r Z T θ0 ; θ0 ; α
= E ∂α ∂r ψ Y, r Z T θ0 ; θ0 ; α ∂θ r Z T θ0 ; θ0
= E E ∂α ∂r B r Z T θ0 ; θ0 ; α + ∂α ∂r C r Z T θ0 ; θ0 ; α Y | Z ∂θ r Z T θ0 ; θ0
=0
because E (Y | Z) = r Z T θ0 ; θ0 and ∂α ∂r B (r, α) + ∂α ∂r C (r, α) r ≡ 0, for any α.
For the bandwidth b
h we obtain an asymptotic equivalence with a theoretical ‘optimal’
bandwidth minimizing −T (h; α∗ ) , that is we prove that the ratio of the two bandwidths converges to one, in probability. Remark that −T (h; α∗ ) is a kind of ψ−CV (cross validation)
function. It can be shown that, up to constant additive terms, −T (h; α∗ ) is asymptotically
equivalent to a weighted (mean-squared) CV function. When ψ (y, r; α) = − (y − r)2 , the
function −T (h; α∗ ) is the usual CV function that one
would use for choosing the bandwidth
for the Nadaraya-Watson estimator of E Y | Z T θ0 . By extension of classical results for
nonparametric regression, it can be proved that the rate of the theoretical ‘optimal’ bandwidth minimizing −T (h; α∗ ) is n−1/5 (see Lemma B.3 in Appendix B; see also Härdle, Hall
and Ichimura (1993) for the case ψ (y, r; α) = − (y − r)2 ). Deduce that b
h is also of order
−1/5
n
.
2.4
Extensions
Given the model conditions (2.1)-(2.2), the idea is to choose a LEFN density with mean r
and variance g(r, α) and to construct a semiparametric PML estimator given a preliminary
estimate of the nuisance parameter α0 . However, it may happen that no such LEFN density
exists or that one prefers another type of LEFN densities. Then, the idea is to reparametrize
8
the conditional variance of Y given Z. More precisely, we may consider
l (y | r, η) = exp [B (r, η) + C (r, η) y + D (y, η)] ,
where η stands for the nuisance parameter. Let Σ = Σ(r, η) denote the variance of the law
given by this density. Assume that for any given r, the map η → Σ(r, η) is one-to-one. In
this case, in order to provide a LEFN density with variance g(r, α) it suffices to consider
l (y | r, η) with η = Σ−1 (r, g(r, α)). For instance, if g(r, α) = r(1 + αr2 ), one may use a
Negative Binomial density of mean r and nuisance parameter αr. Another solution is to
consider a normal density of mean r where the variance equal to r(1 + αr2 ) plays the role
of the nuisance parameter. In this case, given an estimate of r(1 + αr2 ), our semiparametric
PML becomes a semiparametric generalized least-squares (GLS) procedure. Note that this
example of function g(r, α) leads us to the situation where the nuisance parameter is replaced
by a ‘nuisance’ function of r and some additional parameters.
At the expense of more complicated writings, our methodology can be extended to take
into account the case of a ‘nuisance’ function. More precisely, consider a more general pseudolog-likelihood function ψ (y, r; Ψ(r, g(r, α))) where Ψ(·, ·) is a given real-valued function and
α is the nuisance parameter. See also Gouriéroux, Monfort and Trognon (1984a). To define
bb
en ) in equation (2.3), where (θn , α
en ) → (θ0 , α∗ ),
( θ,
h ), one replaces α
en by Ψ(b
rhn (ZiT θn ; θn ); α
in probability, for some α∗ , and rbhn (·; θn ) is a Nadaraya-Watson estimator of the regression
r(·; θn ). The same type of decomposition of the pseudo-log-likelihood criterion into a purely
parametric part function of θ
n
1 X
ψ Yi , r ZiT θ; θ ; Ψ r ZiT θn ; θn , g r(ZiT θn , θn ), α
en
n i=1
−ψ Yi , r ZiT θ0 ; θ0 ; Ψ r ZiT θ0 ; θ0 , g r(ZiT θ0 , θ0 ), α∗
IA (Zi ) ,
a purely nonparametric part function of h
n
T (h; α∗ ) =
1X
ψ Yi , r̂hi ZiT θ0 ; θ0 ; Ψ r ZiT θ0 ; θ0 , g r(ZiT θ0 , θ0 ), α∗
IA (Zi )
n i=1
and a negligible reminder function of θ and h can be used. For brevity, the details of this
more general case are omitted. However, we sketch a quick argument that applies for the
semiparametric GLS.2 Consider the semiparametric GLS criterion
n
2
−1
1X
g rbhn (ZiT θn ; θn ); α
en
Yi − r̂hi ZiT θ; θ IA (Zi )
Sb (θ, h; θn , α
en , hn ) = −
n i=1
with (θn , α
en ) → (θ0 , α∗ ), in probability, and hn , n ≥ 1, a sequence of bandwidths. Assume
that
max g rbhn (ZiT θn ; θn ); α
en − g r(ZiT θ0 ; θ0 ); α∗ IA (Zi ) = oP (1)
(2.10)
1≤i≤n
2
This semiparametric generalized least-squares procedure is a particular case for Picone and Butler (2000).
However, they do not provide a bandwidth rule.
9
and g r(z T θ0 ; θ0 ); α∗ IA (z) stays away from zero. Then the GLS criterion Sb (θ, h; θn , α
en , hn )
is asymptotically equivalent to the infeasible GLS criterion
n
2
−1
1 X
−
Yi − r̂hi ZiT θ; θ g r(ZiT θ0 ; θ0 ); α∗
IA (Zi ) ,
n i=1
that is we can decompose the two criteria in such way that, up to negligible reminders, they
have exactly the same purely parametric and purely nonparametric parts. Finally, we apply
the methodology3 described in the previous subsection with ψ (y, r; α) = − (y − r)2 and the
−1
. In order to ensure condition (2.10), it
trimming IA (Zi ) multiplied by g r(ZiT θ0 ; θ0 ); α∗
suffices to suppose that the map (r, α) → g (r; α) satisfies a Lipschitz condition and that hn
is such that
max rbhn (ZiT θ0 ; θ0 ) − r(ZiT θ0 ; θ0 ) IA (Zi ) = oP (1)
1≤i≤n
and max1≤i≤n |∂θ rbhn ZiT θ; θ |IA (Zi ) is bounded in probability, uniformly with respect to
θ in oP (1) neighborhoods of θ0 . For instance, a bandwidth of order n−1/5 satisfies these
conditions (see Andrews (1995); see also Delecroix, Hristache and Patilea (2006)).
Other possible extensions of the framework we consider is to allow a multi-index regression
and/or multivariate dependent variables. For instance, the SIM condition can be replaced
by the multi-index condition
E (Y | Z) = E Y | Z T θ01 , ..., Z T θ0p
with p smaller than the dimension of Z, while the second order moment condition remains
V ar (Y | Z) = g (E (Y | Z) , α0 ) . On the other hand, for multivariate dependent variables
one may consider PML estimation based on the multivariate normal or multivariate generalizations of Poisson, Negative Binomial distributions (Johnson, Kotz and Balakrishnan
e T and R as above can still be
(1997)). The decomposition of the pseudo-log-likelihood in S,
used for these cases but the detailed analysis of these extensions will be considered elsewhere.
3
Asymptotic results
b
In this section we obtain the asymptotic
distribution
for θ and the corresponding estimator
T
of the regression function r (t; θ) = E Y | Z θ = t as well as the asymptotic behavior of b
h,
bb
with (θ,
h) defined in (2.3). A consistent estimator for the asymptotic variance matrix of θb
is proposed. Moreover, a lower bound for the asymptotic variance matrix of θb is derived.
For the identifiability of the parameter of interest θ0 , hereafter fix its first component,
that is θ0 = (1, θe0T )T , θe0 ∈ Rd−1 . Therefore, we shall implicitly identify a vector θ = (1, θeT )T
with its last d − 1 components and redefine the symbol ∂θ as being the vector of the first
order partial derivatives with respect to the last d − 1 components of θ.
Notice that the trimming function z → IA (z) with A = z : f z T θ0 ; θ0 ≥ c can be written as a
function of z T θ0 . In view of our proofs, it becomes obvious that the methodology described in the previous
subsection remains valid if IA (Zi ) is multiplied by a function depending only on ZiT θ0 .
3
10
Let v (t; θ) =V ar (Y | Xθ = t). If the
SIM assumption and variance condition (2.2) hold,
then v Z T θ0 ; θ0 = g r Z T θ0 ; θ0 , α0 . For a given θ, let r0 (·; θ) and r00 (·; θ) denote the
first and second order derivatives of the function r (·; θ) . Similarly, f 0 (·; θ) is the derivative
of f 0 (·; θ) . Define4
2
K1 1
E
∂r C r Z T θ0 ; θ0 ; α∗
(3.1)
C1 = −
2
4
"
0 T
#2
0
T
2 r Z θ0 ; θ0 f Z θ0 ; θ0
× r00 Z T θ0 ; θ0 +
I
(Z)
A
f (Z T θ0 ; θ0 )
∗
1
1
T
T
v Z θ0 ; θ0 IA (Z) ,
∂r C r Z θ0 ; θ0 ; α
C2 = − K2 E
2
f (Z T θ0 ; θ0 )
R
R
with K1 = u2 K (u) du, K2 = K 2 (u) du, and consider
4
−1 −1
hopt
=
arg
max
C
h
+
C
n
h
= (C2 /4C1 )1/5 n−1/5 .
1
2
n
h
Define the (d − 1) × (d − 1) matrices
n
o
∗ 2
T
T
T
T
T
I = E ∂r C r Z θ0 ; θ0 ; α
v Z θ0 ; θ0 ∂θ r Z θ0 ; θ0 ∂θ r Z θ0 ; θ0 IA (Z)
h
T
J = E ∂r C r Z θ0 ; θ0 ; α
∗
T
T
∂θ r Z θ0 ; θ0 ∂θ r Z θ0 ; θ0
T
i
IA (Z) .
Note that I = J if the variance condition (2.2) holds and α∗ = α0 .
Now, we deduce the asymptotic normality of the semiparametric PML θb estimator in
the presence of a nuisance parameter. Moreover, we obtain the rate of decay to zero of the
‘optimal’ bandwidth b
h. The proof of the following result is given in Appendix refproof.
Theorem 3.1 Suppose that the assumptions in Appendix A hold. Define the set Θn =
{θ : kθ − θ0 k ≤ dn }, n ≥ 1, with dn ln n → 0 and α
en , n ≥ 1, such that α
en − α∗ = oP (1). Fix
bb
c > 0. If (θ,
h) is defined as in (2.3)-(2.4), then b
h/hopt
n → 1, in probability, and
√
D
n θb − θ0 −→ N 0, J −1 IJ −1 .
If Z is bounded, the same conclusion remains true for any sequence dn → 0.
4
2
2
2
Note that ∂22
ψ(y, r) = ∂rr
C(r, α) (y − r) − ∂r C(r, α). Thus, −∂r C can be replaced by ∂22
ψ in the
definition of the constants C1 and C2 .
11
In applications J −1 IJ −1 is unknown and therefore it has to be consistently estimated. To
this end, we propose an usual sandwich estimator of the asymptotic variance J −1 IJ −1 (e.g.,
Ichimura (1993)). Let fbh (·; θ) denote the kernel estimator for the density of Z T θ. Define
n
i2
i2 h
1 Xh
b θb
b θb ; α
en
Yi − rbbh ZiT θ;
∂r C rbbh ZiT θ;
n i=1
T
b θb ∂θ rbb Z T θ;
b θb I b T b b
× ∂θ rbbh ZiT θ;
i
h
{z: fhb (z θ; θ)≥c} (Zi )
n
T
1X
b θb ; α
b θb ∂θ rbb Z T θ;
b θb I b T b b
Jn =
en ∂θ rbbh ZiT θ;
∂r C rbbh ZiT θ;
i
h
{z: fhb (z θ; θ)≥c} (Zi ).
n i=1
In =
Proposition 3.2 Suppose that the conditions of Theorem 3.1 hold. Then, Jn−1 In Jn−1 →
J −1 IJ −1 , in probability.
Proof. The arguments are quite standard (e.g., Ichimura (1993), section 7). On one
hand, the convergence in probability of θb and
en and, on the other hand, the convergence in
α
probability of rbbh z T θ; θ and ∂θ rbbh z T θ; θ , uniformly over θ in neighborhoods shrinking to
θ0 and uniformly over z ∈ A (e.g., Andrews (1995), Delecroix, Hristache and Patilea (2006))
imply In → I and Jn → J, in probability.
Theorem 3.1 shows, in particular, that θb is asymptotically equivalent to the semiparametric PML based on the LEF pseudo-log-likelihood ψ (y, r; α∗ ) = ln f (y, r | α∗ ) . As in the
parametric case, we can deduce a lower bound for the asymptotic variance J −1 IJ −1 with
respect to semiparametric PML based on LEF densities. This bound is achieved by θb if the
SIM assumption and the variance condition (2.2) hold and α∗ = α0 . The proof of the following proposition is identical to the proof of Property 5 of Gouriéroux, Monfort and Trognon
(1984a, page 687) and thus it will be skipped.
Proposition 3.3 The set of asymptotic variance matrices of the semiparametric PML estimators based on linear exponential families has a lower bound equal to K, where
n
o
−1
T
K−1 = E v Z T θ0 ; θ0
∂θ r Z T θ0 ; θ0 ∂θ r Z T θ0 ; θ0 IA (Z) .
Concerning the nonparametric part, we have the following result on theasymptotic distribution of the nonparametric estimator of the regression. The proof is omitted (see Härdle
and Stoker (1989)).
Proposition 3.4 Assume that the conditions of Theorem 3.1 are fulfilled. Then, for any t
such that f (t; θ0 ) > 0,
p
D
nb
h rbbh t; θb − r (t; θ0 ) − b
h2 β (t) −→ N 0, K2 v(t; θ0 )f (t; θ0 )−1
where β (t) = (K1 /2) r00 (t; θ0 ) + 2r0 (t; θ0 ) f 0 (t; θ0 ) f (t; θ0 )−1 .
12
Note that, for any z such that f z T θ0 ; θ0 > 0,
p
D
−1
b θb − r z T θ0 ; θ0 − b
h2 β z T θ0 → N 0, K2 v(z T θ0 ; θ0 )f z T θ0 ; θ0
.
nb
h rbbh z T θ;
Indeed, use the results of Andrews (1995) to deduce that ∂θ rbbh zT θ; θ → ∂θ r z T θ; θ ,
in probability, uniformly over neighborhoods of θ0 where f z T θ; θ stays away from zero.
Therefore, we can write
Tb b
T
Tb b
rbbh z θ; θ − r z θ0 ; θ0 = rbbh z θ; θ − rbbh z T θ0 ; θ0 + rbbh z T θ0 ; θ0 − r z T θ0 ; θ0
= ∂θ rbbh z T θ0 ; θ0 θb − θ0 + oP θb − θ0 + rbbh z T θ0 ; θ0 − r z T θ0 ; θ0
= OP θb − θ0 + rbbh z T θ0 ; θ0 − r z T θ0 ; θ0
b θ)
b as a consequence of the √n−consistency
and obtain the asymptotic normality of rbbh (z T θ;
of θb and the asymptotic behavior of the Nadaraya-Watson estimator.
4
Two-step semiparametric PML
Here, we consider a two-step semiparametric PML procedure that can be applied in semiparametric single-index regression models when a conditional variance condition like
V ar (Y | Z) = g (E (Y | Z) , α0 ) = g r Z T θ0 ; θ0 , α0 ,
(4.2)
is specified. Assume that this conditional variance condition is correctly specified. At the
end of this section we also discuss the misspecification case.
First, we have to build a sequence {θn } with limit θ0 . Moreover, in the case of unbounded
covariates, θn should approach θ0 faster than 1/ ln n. For this purpose, we maximize with
respect to θ a pseudo-likelihood based on a LEF density l (y | r). We use a fixedtrimming
IB (·) with B a subset of Rd such that, for any θ and any z ∈ B, we have f z T θ; θ ≥ c > 0.
To ensure consistency for such a PML estimator, we have to check that
θ0 = arg max E ln l Y | r Z T θ; θ IB (Z) ,
(4.3)
θ
and θ0 is unique with this property.
Recall
that the SIM condition specifies θ0 as the unique
T
vector satisfying E [Y | Z] = E Y | Z θ0 . On the other hand, if ln l (y | r) = B (r) +
C (r) y + D (y) , then B (m) + C (m) r ≤ B (r) + C (r) r (cf. Property 4, Gouriéroux, Monfort
and Trognon (1984a, page 684)). Deduce that for any z,
θ0 = arg max E ln l Y | r z T θ; θ
θ
and θ0 is the unique maximizer. Hence, condition (4.3) holds for any set B. This leads us to
the following definition of a preliminary estimator.
13
STEP 1 (preliminary step). Consider a sequence of bandwidths hn , n ≥ 1, such that
n hn → 0 and n1/2−ε hn → ∞ for some 0 < ε < 1/2. Moreover, let l (y | r) be a LEF density.
Define
n
1X
θn = arg max
ln l Yi | r̂hn ZiT θ; θ IB (Zi ).
n i=1
θ
ε
Delecroix, Hristache and Patilea (2006) showed that, under the regularity conditions
required by Theorem 3.1, we have θn − θ0 = oP (1/ ln n) . Using the preliminary estimate θn
and the variance condition (4.2) we can build α
en , n ≥ 1, such that α
en → α0 , in probability
(see the end of this section). Let l (y | r, α) denote a LEFN density
with mean r and variance
−1/8
g (r, α) . Consider cn → ∞ (e.g., cn = ln n), define Hn = h : cn n−1/4 ≤ h ≤ c−1
.
n n
Moreover, consider Θn = {θ : kθ − θ0 k ≤ dn }, n ≥ 1 with {dn } as in Theorem 3.1. Fix some
small c > 0.
STEP 2. Define
n
1X
b
b
ln l Yi | r̂hi ZiT θ; θ ; α
en I{z: fbi (zT θn ; θn )≥c} (Zi ),
θ, h = arg max
hn
θ∈Θn , h∈Hn n i=1
with θn and hn from Step 1.
The following result is a direct consequence of Theorem 3.1.
Corollary 4.1 Suppose that the assumptions of Theorem 3.1 hold. If θb and b
h are obtained
as in Step 2 above, then
√
D
n θb − θ0 −→ N (0, K) ,
with
K
−1
=E
n
T
v Z θ0 ; θ0
−1
T
T
∂θ r Z θ0 ; θ0 ∂θ r Z θ0 ; θ0
T
o
IA (Z) .
Moreover,
b
h
(C2 /4C1 )1/5 n−1/5
→ 1,
in probability, where C1 and C2 are defined as in (3.1) with α∗ = α0 .
Remark 1. Let us point out that simultaneous optimization of the semiparametric criterion in Step 1 with respect to θ, α and h (or with respect to θ and α for a given h) is not
recommended, even if the conditional variance V ar (Y | Z) is correctly specified. Indeed, if
the true conditional distribution of Y given Z is not the one given by the LEFN density
l (y | r, α) = exp ψ (y, r; α) , joint optimization with respect to θ and α leads, in general, to
an inconsistent estimate of α0 . (This failure is well-known in the parametric case where r
is a known function; see comments of Cameron and Trivedi (2013), pages 84-85. In view of
decomposition (2.7) we deduce that this fact also happens in the semiparametric framework
14
where r has to be estimated.) In this case the matrices I and J defined in section 3 are
no longer equal and thus the asymptotic variance of the one-step semiparametric estimator
of θ obtained by simultaneous maximization of the criterion in Step 1 with respect to θ,
α does not achieve the bound K. However, when the SIM condition holds and the true
conditional law of Y is given by the LEFN density l = exp ψ, our two-step estimator θb and
the semiparametric MLE of θ0 obtained by simultaneous optimization with respect to (θ, α)
are asymptotically equivalent.
Remark 2. Note that if we ignore the efficiency loss due to trimming, K is equal to
the efficiency bound in the
semiparametric model defined only by the single-index condition
T
E (Y | Z) = E Y | Z θ0 when the variance condition (4.2) holds. To see this, apply the
bound of Newey and Stoker (1993) with the true variance given by (4.2). Our two-stage
estimator achieves this SIM efficiency bound (if the variance is well-specified). However,
this SIM bound is not necessarily the two moment conditions model bound. The latter
should take into account the variance condition (see Newey (1993), section 3.2, for a similar
discussion in the parametric nonlinear regression framework). In other words our two-stage
estimator has some optimality properties but it may not achieve the semiparametric efficiency
bound of the two moment conditions model. The same remark applies for the two-stage
semiparametric generalized least squares (GLS) procedure of Härdle, Hall and Ichimura
(1993) [see also Picone and Butler (2000)]. Achieving semiparametric efficiency when the
first two moments are specified would be possible, for instance, by estimating higher orders
conditional moments nonparametrically. However, in this case we face again the problem of
the curse of dimensionality that we tried to avoid by assuming the SIM condition.
To complete the definition of the two-step procedure above, we have to indicate how
to build a consistent sequence {e
αn }. Such
a sequence can be obtained from the moment
T
condition (4.2) after replacing r z θ0 ; θ0 by a suitable estimator. This kind of procedure
is commonly used in the semiparametric literature (e.g., Newey and McFadden (1994)). For
simplicity, let us only consider the Negative Binomial case where, for any z, we have
E (Y − E (Y | Z))2 | Z = z = r z T θ0 ; θ0 1 + α0 r z T θ0 ; θ0 .
(4.4)
Consider a set B ⊂ Rd such that, for any θ and any z ∈ B, we have f z T θ; θ ≥ c > 0. We
can write
n h
i
o
n
o
2
2
E E Y − r Z T θ0 ; θ0
− r Z T θ0 ; θ0 | Z IB (Z) = α0 E r Z T θ0 ; θ0 IB (Z) .
Consequently, we may estimate5 α0 by
2
i
Pn h
1
T
T
Y
−
r
b
Z
θ
;
θ
−
r
b
Z
θ
;
θ
IB (Zi )
i
h
n
n
h
n
n
n
n
i
i
i=1
n
α
en =
P
2
n
1
bhn (ZiT θn ; θn ) IB (Zi )
i=1 r
n
(4.5)
5
One can expect little influence of the choice of the bandwidth used to construct the α
en . This is indeed
confirmed by the simulation experiments we report in section 5.1.
15
with θn and hn from Step 1 and rbhn the Nadaraya-Watson estimator with bandwidth hn .
Since θn → θ0 , deduce that α
en → α0 , in probability (see also the arguments we used in
subsection 2.4).
Now, let us comment on what happens with our two-step procedure if the second order moment condition is misspecified, while the SIM condition still holds. In general, the
sequence α
en one may derive from the conditional variance condition and the preliminary
estimate of θ0 is still convergent
to some pseudo-true value α∗ of the nuisance parameter.6
bb
Then, the behavior of θ,
h yielded by Step 2 is described by Theorem 3.1, that is θb is still
√
n−asymptotically normal and b
h is still of order n−1/5 .
Finally, if the SIM condition does not hold, then θb estimates a kind of first projectionpursuit direction. In this case, our procedure provides an alternative to minimum average
(conditional) variance estimation (MAVE) procedure of Xia et al. (2002). The novelty would
be that the first projection direction is defined through a more flexible PML function than
the usual least-squares criterion. This case will be analyzed elsewhere.
5
Empirical evidence
In our empirical section we consider the case of a count response variable Y . A benchmark
model for studying event counts is the Poisson regression model. Different variants of the
Poisson regression have been used in applications on the number of patents applied for and
received by firms, bank failures, worker absenteeism, airline or car accidents, doctor visits,
etc. Cameron and Trivedi (2013) provide an overview of the applications of Poisson regression. In the basic setup, the regression function is log-linear. An additional unobserved
multiplicative random error term in the conditional mean function is usually used to account for unobserved heterogeneity. In this section we consider semiparametric single-index
extensions of such models.
5.1
Monte Carlo simulations
To evaluate the finite sample performances of our estimator θb and of the optimal bandwidth
b
h, we conduct a simulation experiment with 500 replications.
We consider three explanatory variables Z = (Z1 , Z2 , Z3 )> ∼ N (0, Σ) with Σ = [σij ]3×3
and σij = 0.5|i−j| . The regression function is
E(Y | Z) = (Z > θ0 )2 + 0.5
(1)
(2)
(3)
and θ0 = (θ0 , θ0 , θ0 )T = (1, 3, −2)T . The conditional distribution of Y given Z and
ε is Poisson of mean r(Zθ0 ; θ0 ) · ε with ε independent of Z and distributed according to
6
For instance, α
en defined in (4.5) is convergent in probability to
2
IB (Z)] − E[r Z T θ0 ; θ0 IB (Z)]
E[ Y − r Z T θ0 ; θ0
α =
.
E[r(Z T θ0 ; θ0 )2 IB (Z)]
∗
To ensure that the limit of α
en is positive, one may replace α
en by max (e
αn , ρ) for some small but positive ρ.
16
Gamma(0.5, 2) or Uniform(0, 2). Thus, the conditional variance of Y given Z is given by
the function g (r, α) = r (1 + αr) with α0 = 2 for ε ∼ Gamma(0.5, 2) and α = 1/3 for
ε ∼ Uniform(0, 2).
For this simulation experiment we generate samples of size n = 200 and 300. For the
2
nonparametric part we use a quartic kernel K (u) = (15/16) (1 − u2 ) I[−1,1] (u) . To estimate
the parameter θ0 and the regression r(·; θ0 ) we use two semiparametric two-step estimation
procedures as defined in section 4: i) A procedure with a Poisson PML in the first step and
(2)
(3)
a Negative Binomial PML in the second step; let θbN B−SP = (1, θbN B−SP , θbN B−SP )> denote
the two-step estimator. ii) a procedure with a least-squares method in the first step and
(2)
(3)
a GLS method in the second step; let θbGLS−SP = (1, θbGLS−SP , θbGLS−SP )> be the two-step
estimator. Note that θbN B−SP and θbGLS−SP have the same asymptotic variance. In both
two-step procedures considered, we estimate α0 using the estimator defined in (4.5). The
bandwidth hn is equal to 3n−1/5 . We also consider the parametric two-step GLS method as a
benchmark. In this case the link function and the variance parameter are considered given;
(3)
(2)
let θbGLS−P = (1, θbGLS−P , θbGLS−P )> denote the corresponding estimator.
Table 1. Poisson regression with unobserved heterogeneity ε ∼ Gamma(0.5, 2). The true conditional
variance of Y given Z is r(Zθ0 ; θ0 )(1 + 2r(Zθ0 ; θ0 )) with r (t; θ0 ) = t2 + 0.5. The true vector θ0 is
T
(1, 3, −2) . Let θbN B−SP and θbGLS−SP denote the two-step estimators obtained from the Negative
Binomial pseudo-likelihood and GLS criterion, respectively. The first step Poisson PML estimator is
denoted by θbP OI−SP . The superscripts indicate the components of the vectors.
n
(2)
θbGLS−P
(2)
θbGLS−SP
(2)
θbP OI−SP
(2)
θbN B−SP
(3)
θbGLS−P
(3)
θbGLS−SP
(3)
θbP OI−SP
(3)
θbN B−SP
200
mean
std.
MSE
2.8977
0.8097
0.3822
2.8019
0.8986
0.8467
3.0177
0.9937
0.9879
3.1249
0.9481
0.9145
-1.9501
0.6268
0.3929
-1.6954
0.5170
0.3600
-1.7955
0.6435
0.4559
-2.0520
0.5580
0.3167
300
mean
std.
MSE
2.9422
0.4600
0.2150
2.8261
0.7741
0.6295
2.9982
0.9288
0.8628
3.0758
0.8297
0.6941
-1.9594
0.5002
0.2519
-1.7215
0.4568
0.2862
-1.8028
0.5705
0.3643
-1.9569
0.4670
0.2199
Table 2. The same setup as in Table 1 but with ε ∼ Uniform(0, 2) and the true conditional variance of Y
given Z equal to r(Zθ0 ; θ0 )(1 + (1/3)r(Zθ0 ; θ0 )).
n
(2)
θbGLS−P
(2)
θbGLS−SP
(2)
θbP OI−SP
(2)
θbN B−SP
(3)
θbGLS−P
(3)
θbGLS−SP
(3)
θbP OI−SP
(3)
θbN B−SP
200
mean
std.
MSE
2.9842
0.2505
0.0630
2.8460
0.4537
0.2295
2.9755
0.6619
0.4387
3.0613
0.4917
0.2456
-1.9961
0.2551
0.0651
-1.8702
0.2874
0.0994
-1.9094
0.4117
0.1777
-2.0127
0.2921
0.0855
300
mean
std.
MSE
2.9919
0.2213
0.0497
2.8956
0.4279
0.1940
2.9422
0.5753
0.3343
3.0618
0.3658
0.1376
-1.9946
0.2443
0.0597
-1.8999
0.2647
0.0800
-1.8953
0.3639
0.1433
-2.0052
0.2237
0.0500
17
The results on the estimates of the components of θ0 are provided in Table 1 and Table 2.
We report the mean, the standard deviation and the estimated mean squared error (MSE)
for each component. The two semiparametric estimators that incorporate the information on
the conditional variance clearly outperform the semiparametric single-index estimator that
ignores that information. Moreover, they behave reasonably well compared to the parametric
benchmark.
5.2
A real data example
In order to further illustrate our methodology, we consider a real dataset on recreational trips
as presented by Cameron and Trivedi (2013). This data initially collected by Sellar, Stoll
and Chavas (1985) is built from a survey that includes the number of recreational boating
trips to Lake Sommerville, Texas. We reproduce below the tables that describe the observed
frequencies and the explanatory variables. We do not use all the explanatory variables for
estimation since the variables C1, C3 and C4 are almost perfectly correlated in the sample.
(Indeed, Corr(C1, C3) = 0.977, Corr(C1, C4) = 0.987 and Corr(C3, C4) = 0.964.) To
avoid collinearity problems, we drop C3 and C4. We standardize the variables IN C and
C1.
Table 3. The recreational trips data set: actual frequency distribution.
Number of Trips
Frequency
0
417
1
68
2
38
3
34
4
17
5
13
6
11
7
2
8
8
9
1
10
13
Number of Trips
Frequency
11
2
12
5
15
14
16
1
20
3
25
3
26
1
30
3
40
3
50
1
88
1
Table 4. Explanatory variables for the recreational trips counts.
Variable
Definition
Mean
Std
T RIP S
Number of recreational boating trips in 1980
by a sample group
Facility’s subjective quality ranking on a scale of 1 to 5
Equal 1 if engaged in water-skiing at the lake
Household income of the head of the group ($10,000/year)
Equal 1 if user’s fee paid at Lake Sommerville
Hundreds of dollar expenditure when visiting Lake Conroe
Hundreds of dollar expenditure when visiting Lake Somerville
Hundreds of dollar expenditure when visiting Lake Houston
2.244
6.292
1.419
0.367
0.385
0.019
0.554
0.599
0.560
1.812
0.482
0.185
0.139
0.467
0.488
0.461
SO
SKI
IN C
F C3
C1
C3
C4
The model we consider is the one given by equations (2.1)-(2.2) with g(r, α) = r(1 + αr).
First, we assume that the regression function is log-linear, that is we consider the standard
Negative Binomial Parametric model (NB-P). Next, we no longer assume that the regression function is known and we apply our semiparametric methodology, the semiparametric
18
5
4
3
hat r(t)
2
1
−2
0
2
4
t
Figure 1: The link function
Negative Binomial pseudo-likelihood procedure. In the semi-parametric procedures the coefficient of the variable SO is set to 1. For the nonparametric part we use the quartic kernel
K(u) = (15/16)(1 − u2 )2 I[−1,1] (u). The parameter estimates and estimated standard errors
are gathered in Table 5, the plot of the estimated link function is provided in Figure 1.
Table 5. Estimation results: parametric (NB-P) versus semiparametric model ( NB-SP).
Parameters
NB-P
NB-SP
Intercept
SO
SKI
IN C
F C3
C1
α
h
-1.7452 (0.1441)
0.9017 (0.0430)
0.4420 (0.1707)
-0.2245 (0.0906)
1.5813 (0.4404)
-0.3258 (0.1018)
2.2983 (0.2210)
.
.
1
-0.2489 (0.0405)
0.1963 (0.0690)
-0.1399 (0.0702)
-0.2987 (0.0995)
5.5764
5.6530
Note that the estimate of the coefficient of SO in the parametric model is close to one,
while in the semiparametric approach we fixed it to one. Thus the estimated values of the
remaining parameters in the parametric and semiparametric cases are almost directly comparable. The results obtained with the semiparametric approach seem more realistic. For
instance, the coefficient of IN C covariate is positive with NB-SP and the link function is
strictly monotone. This suggests that a higher income more likely induces a larger number
of recreational trips. The NB-P model leads to the opposite conclusion. The reported parametric and semiparametric standard errors cannot be directly compared on the same basis
since we can only compute the standard error of a ratio of parameters in the semiparametric cases. The large bandwidth could be explained by the large conditional variance of the
19
response and a link function with a second derivative close to zero. This leads to a large
constant (C2 /4C1 )1/5 in the expression of hopt
n , see equation (3.1) above.
In order to evaluate the overall performance of the parametric and semiparametric models
and of the estimation methods, we consider various goodness-of-fit measures such as the
Pearson statistic, the deviance statistic and the deviance pseudo R-squared statistic. The
Pearson statistics is given by
n
X
(Yi − rbi )2
,
P =
ω
b
i
i=1
where rbi is the estimated conditional mean for individual i and ω
bi is the estimated conditional
variance computed according to equation (2.2). The deviance statistic is given by
n
X
Yi
Yi + 1/b
α
D=2
Yi ln
− (Yi + 1/b
α) ln
,
r
b
r
b
+
1/b
α
i
i
i=1
with α
b the estimated value of the nuisance parameter with the values given in the Table
5. Finally, if Y denotes the sample mean of the variable Y , the deviance pseudo R-squared
statistic is
i
n h
P
+1/b
α
Yi ln (Yi /b
ri ) − (Yi + 1/b
α) ln Yrbii+1/b
α
i=1
2
RDEV
=1− P
i .
n h
Yi +1/b
α
Yi ln Yi / Y − (Yi + 1/b
α) ln Y +1/bα
i=1
Another model diagnostic is obtained when comparing fitted probabilities and actual
probabilities by the mean of a chi-square type statistic. The statistic we consider is
2
J
X
pj − pbj
,
ξ=n
p
j
j=1
where the possible values of Y are aggregated in J non overlapping cells. 7 The actual
frequency for cell j is denoted pj while pbj is the corresponding predicted probability by the
model under study. For both methods GLS-SP and NB-SP we used the probabilities of
a negative binomial distribution to compute pbj . We consider seven cells corresponding to
the values T RIP = 0, ..., 5 and T RIP > 5. All the results are summarized in Table 6.
The semiparametric model performs better than the parametric model. We also give the
estimators of the probability in Table 7. We can see that our estimators are close to the
empirical probability of T RIP . The semiparametric approach greatly improves the standard
parametric modeling.
2
Table 6. Goodness-of-fit statistics: P −Pearson statistic, D−deviance statistic, RDEV
−deviance pseudo
R-squared statistic and ξ−chi-square statistic.
7
The chi-square statistic we consider is not necessarily chi-square distributed under the null hypotheses
of a well specified model. This is because it does not correctly take into account the estimation error in pbj .
See Andrews (1988) for the general definition of the chi-square goodness-of-fit test statistic in nondynamic
regression models. Here, we only use ξ as a crude diagnostic for the three types of fitted probabilities pbj .
20
P
D
2
RDEV
ξ
NB-P
NB-SP
5296.506
1158.41
0.4780
968.9416
608.7212
405.1771
0.1886
3.1922
Table 7. Empirical probability and
T RIP S
0
1
2
Empirical probability 0.6327 0.1031 0.0576
NB–P
0.1111 0.1572 0.1596
NB–SP
0.6314 0.1045 0.0568
6
estimate
3
0.0515
0.1407
0.0381
probability
4
5
0.0257
0.0197
0.1147
0.0889
0.02797 0.0215
>5
0.1092
0.2273
0.1194
Conclusion
We consider a semiparametric single-index model (SIM) where an additional second order
moment condition is specified. To estimate the parameter of interest θ we introduce a
two-step semiparametric pseudo-maximum likelihood (PML) estimation procedure based on
linear exponential families with nuisance parameter densities. This procedure extends the
quasi-generalized pseudo-maximum likelihood method proposed by Gouriéroux, Monfort and
Trognon (1984a, 1984b). We also provide a natural rule for choosing the bandwidth of the
nonparametric smoother appearing in the estimation procedure. The idea is to maximize
the pseudo-likelihood of the second step simultaneously in θ and the smoothing parameter
h. The rate of the bandwidth is allowed to lie in a range between n−1/4 and n−1/8 . We derive
b
the asymptotic behavior of
√ θ, the two-step semiparametric PML we propose. If the SIM
b
condition holds, then θ is n−asymptotically normal. We also provide a consistent estimator of its variance. When the SIM condition holds and the conditional variance is correctly
specified, then θb has the best variance amongst the semiparametric PML estimators. The
‘optimal’ bandwidth b
h obtained by joint maximization of the pseudo-likelihood function in
the second step is shown to be equivalent to the minimizer of a weighted cross-validation
function. From this we deduce that n1/5b
h converges to a positive constant, in probability. In
b
particular, our optimal bandwidth h has the rate expected when estimating a twice differentiable regression function nonparametrically. We conduct a simulation experiment in which
the data were generated using a Poisson single-index regression model with multiplicative
unobserved heterogeneity. The simulation confirms the significant advantage of estimators
that incorporate the information on the conditional variance. We also applied our semiparametric approach to a benchmark real count data set and we obtain a much better fit than
the standard parametric regression models for count data.
21
A
Appendix: Assumptions
e with Θ
e a compact subset of Rd−1 with nonvoid interior. Depending on the
Let Θ = {1} × Θ
context, Θ is considered a subset of Rd−1 or a subset of Rd .
Assumption A.1 The observations Y1 , Z1T
T
random vector Y, Z T ∈ Rd+1 .
T
, . . . , Yn , ZnT
T
are independent copies of a
Assumption A.2 Let r (t; θ) = E Y | Z T θ = t . There
exists a unique θ0 interior point of
T
T
Θ such that E (Y | Z) = E Y | Z θ0 = r Z θ0 ; θ0 .
Assumption A.3 For every θ ∈ Θ, the random variable Z T θ admits a density f (·; θ) with
respect to the Lebesgue measure on R.
Assumption A.4 E [exp (λ kZk)] < ∞, for some λ > 0. Moreover, E(Y 4+ε ) < ∞, for
some ε > 0.
Assumption A.5 With probability one, the matrix (1, Z T )T (1, Z T ) is positive definite.
Assumption A.6 There exists c0 > 0 and a positive integer k0 such that, for any θ ∈ Θ
and 0 < c ≤ c0 , the set {t : f (t; θ) = c} has at most k0 elements.
The last two assumptions ensure that P f (Z T θ0 ; θ0 ) = c = 0, for any 0 < c ≤ c0 .
CONDITION L A function g : Θ × R → R is said to satisfy Condition L if, for any Λ
a compact set on the real line, there exists B > 0 and b ∈ (0, 1] such that
b
|g (θ, t) − g (θ0 , t0 )| ≤ B k(θ, t) − (θ0 , t0 )k ,
θ, θ0 ∈ Θ,
t, t0 ∈ Λ.
Assumption A.7 a) The function (θ, t) → f (t; θ) ≥ 0, θ ∈ Θ, t ∈ R, satisfies a Lipschitz
condition, that is there exists a ∈ (0, 1] and C > 0 such that
|f (t; θ) − f (t0 ; θ0 )| ≤ C k(θ, t) − (θ0 , t0 )k
a
f or
θ, θ0 ∈ Θ
and
t, t0 ∈ R.
b) The function (θ, t) → r (t; θ) , θ ∈ Θ, t ∈ R, satisfies Condition L.
c) For any θ ∈ Θ, the functions t → γ (t; θ) and t → f (t; θ) are twice differentiable. Let
γ 00 (t; θ) and f 00 (t; θ) denote the second order derivatives. The functions (θ, t) → γ 00 (t; θ) and
(θ, t) → f 00 (t; θ) , θ ∈ Θ, t ∈ R, satisfy Condition L with b = 1.
d) For any θ ∈ Θ and any component Z (j) of Z, the functions t → E Z (j) | Z T θ = t and
t → E Y Z (j) | Z T θ = t are twice differentiable and their second order derivatives satisfy
Condition L with b = 1.
e) For any t ∈ R, the function θ → r (t; θ) is twice continuously differentiable and, for
2
any θ ∈ Θ, the functions t → ∂θ r (t; θ) and t → ∂θθ
r (t; θ) are continuous. Moreover, the
function (θ, t) → ∂θ r (t; θ) satisfy Condition L with b = 1.
Let v (t; θ) = V ar Y | Z T θ = t be the conditional variance of Y given Z T θ = t.
22
Assumption A.8 The function (θ, t) → v (t; θ) satisfies Condition L.
Consider the functions B, C : R × N → R, with Y, R, N ⊂ R. Define Λ =
f (t; θ) ≥ c}, with c, δ > 0, and
S
θ∈Θ {t
:
D(c, δ) = {r : ∃ (θ, t) ∈ Θ × Λ such that |r − r(t; θ)| ≤ δ}.
Assumption A.9 If c > 0, there exists δ > 0 such that D(c, δ) is strictly included in R.
Assumption A.10 The kernel function K (·) is differentiable, symmetric, positive and compactly supported. Moreover, K (·) and the derivative K 0 (·) are of bounded variation.
Up to a term depending only on y and α, the three arguments function ψ (·, ·; ·) involved
in equation (2.3) is defined as
ψ (y, r; α) = B(r, α) + C(r, α)y
where l(y | r, α) = exp [B(r, α) + C(r, α)y + D(y, α)] is a LEFN density with mean r and
variance [∂r C(r, α)]−1 .
Assumption A.11 The functions B (r, α) and C (r, α) are twice differentiable in the first
argument. Moreover, for any c and δ > 0 for which D(c, δ) is strictly included in R, there
exists a constant M such that
2
sup
|∂rr
G(r, α)| + |∂r G(r, α)| ≤ M,
r∈D(c,δ), α∈N
sup
r,r0 ∈D(c,δ), α,α0 ∈N
2
2
∂rr
G(r, α) − ∂rr
G(r0 , α0 ) ≤ M (|r − r0 | + |α − α0 |) ,
where G stands for B or C. The functions ∂r B(r; α) and ∂r C(r; α) are continuously differentiable in α.
Assumption A.12 For any c and δ > 0 for which D(c, δ) is strictly included in R, we have
∂r C(r, α) > 0, ∀r ∈ D(c, δ), ∀α ∈ N.
Assumption A.12 ensures that the (d − 1) × (d − 1) matrix
2
T
∗
J = −E ∂θθ
IA (Z)
T ψ Y, r Z θ0 ; θ0 ; α
h
i
T
= E ∂r C r Z T θ0 ; θ0 ; α∗ ∂θ r Z T θ0 ; θ0 ∂θ r Z T θ0 ; θ0 IA (Z)
is positive definite.
Let us notice that the asymptotic results remain valid even if the function ψ (y, r; α) is
not the logarithm of a LEFN. It suffices to adapt Assumption A.11, to suppose that there
exists F (·; ·) such that ψ (y, r; α) ≤ F (y; α) , ∀r ∈ R, to ensure
that J is positive definite
T
and to assume that, for any α, E ∂2 ψ Y, r Z θ0 ; θ0 ; α | Z = 0 and
E ∂θ ∂2 ψ Y, r Z T θ0 ; θ0 ; α | Z T θ0 = 0.
23
B
Appendix: Technical lemmas
Let Hn = n−(1/2−ε) , n−ε , with 0 < ε < 1/2, and Θn = {θ : kθ − θ0 k ≤ dn }, with dn → 0.
We use C to denote a positive constant, not necessarily the same at each occurrence.
Lemma B.1 Assume that the kernel K is a symmetric, positive, compactly supported function of bounded variation. Suppose that the map (θ, t) → f (t; θ) ≥ 0, θ ∈ Θ, t ∈ R, satisfies
a Lipschitz condition, that is there exists a ∈ (0, 1] and C > 0 such that
|f (t1 ; θ1 ) − f (t2 ; θ2 )| ≤ C k(θ1 , t1 ) − (θ2 , t2 )ka
f or
θ1 , θ2 ∈ Θ
and
t1 , t2 ∈ R.
(B.1)
Then
max
sup
1≤i≤n θ, z, h∈Hn
fbhi z T θ; θ − f z T θ; θ = OP h−1 n−1/2 + O (ha ) .
The proof of Lemma B.1 can be distilled from many existing results (e.g., Andrews (1995),
Sherman (1994b), Delecroix, Hristache and Patilea (2006)) and therefore it will be omitted.
Lemma B.2 a) If δ > 0, then
sup
θ∈Θn , h∈Hn
I{z: fbi (zT θ;θ)≥c} (Zi ) − IA (Zi ) ≤ IAδ (Zi ) + I(δ,∞) (Gn ),
h
1 ≤ i ≤ n,
where Aδ = z : f z T θ0 ; θ0 − c ≤ δ and
Gn = max
sup
1≤i≤n θ∈Θn , h∈Hn
fbhi ZiT θ; θ − f ZiT θ0 ; θ0 .
b) Suppose that K (·) and f (·; ·) satisfy the assumptions of Lemma B.1 for some a, C > 0.
Moreover, assume that either i) Z is bounded and dn → 0 or, ii) E [exp (λ kZk)] < ∞ for
some λ > 0 and dn = o (1/ ln n) (with dn from the definition of Θn ). Let δn → 0 such that
−a
δn /n−aε → ∞ and either i) δn d−a
→ ∞. Then
n → ∞ if Z is bounded or, ii) δn [dn ln n]
−α
I(δn ,∞) (Gn ) = oP (n ) , ∀α > 0.
Proof. a) We have
I{z: fbi (zT θ;θ)≥c} (Zi ) − IA (Zi ) ≤ I{z: fbi (zT θ;θ)≥c}\A (Zi ) + IA\{z: fbi (zT θ;θ)≥c} (Zi ).
h
h
h
For any θ, h and δ, we can write
o
n
o
n
T
i
T
i
T
b
b
fh Zi θ; θ ≥ c \A ⊂ fh Zi θ; θ ≥ c, f Zi θ0 ; θ0 < c−δ ∪ c−δ ≤ f ZiT θ0 ; θ0 < c
and
o
n
o n
A\ fbhi ZiT θ; θ ≥ c ⊂ fbhi ZiT θ; θ < c, f ZiT θ0 ; θ0 ≥ c+δ ∪ c ≤ f ZiT θ0 ; θ0 < c+δ
which proves the inequality.
24
b) It suffices to prove that P (Gn > δn ) → 0. First consider the case of unbounded Z.
Note that, for any z and θ,
a/2
2
f z T θ1 ; θ1 − f z T θ2 ; θ2 ≤ C z T θ1 − z T θ2 +kθ1 − θ2 k2
≤ C (1+kzk)a kθ1 − θ2 ka .
Combine this inequality and Lemma B.1 and write
fbhi ZiT θ; θ − f ZiT θ; θ
1≤i≤n θ∈Θ, h∈Hn
+ max sup f ZiT θ; θ − f ZiT θ0 ; θ0
Gn ≤ max
sup
1≤i≤n θ∈Θn
fbhi z T θ; θ − f z T θ; θ + C kθ − θ0 ka max (1 + kZi k)a
sup
1≤i≤n
1≤i≤n θ∈Θ, h∈Hn , z
a
−aε
−ε
a
= O n
+ OP n
+ O (dn ) max (1 + kZi k) .
≤ max
1≤i≤n
On the other hand, we can write
n
X
a
a
P dn max (1 + kZi k) > δn
≤
P ((1 + kZi k)a > δn /dan )
1≤i≤n
i=1
= nP exp (λ(1 + kZi k)) > exp λδn1/a /dn
≤ n
eλ E [exp (λ kZi k)]
.
1/a
exp λδn /dn
1/a
Since δn / (dn ln n) and δn /n−aε → ∞, deduce that P (Gn > δn ) → 0.
If Z lies in a compact, condition (B.1) implies that for any z in the support of Z,
f z T θ1 ; θ1 − f z T θ2 ; θ2 ≤ C kθ1 − θ2 ka ,
θ1 , θ2 ∈ Θ,
with C > 0 some constant independent of z. In this case
Gn = O n−aε + OP n−ε + O (dan ) .
Thus P (Gn > δn ) → 0 provided that δn → 0 such that δn /n−aε and δn /dan → ∞
The proofs of the following three lemmas are lengthy and technical. These proofs are
provided in Delecroix, Hristache and Patilea (2006) and therefore it will be omitted herein.
The key ingredients for the three proofs are the results on uniform rates of convergence
for U −processes indexed by Euclidean families; see Sherman (1994a). See also Pakes and
Pollard (1989) for the definition and the properties of Euclidean families of functions.
The first of the three lemmas is a refined version of a standard result for cross-validation
in nonparametric regression
(e.g., Härdle
−(1/2−ε)
and Marron (1985)). The result holds uniformly
−ε
in θ,α and for h in Hn = n
,n
, with 0 < ε < 1/2.
25
Lemma
S B.3 Suppose that Assumptions A.1 to A.8 hold. Fix some small c > 0 and let
Λ = θ∈Θ {t : f (t; θ) ≥ c}. Consider a family of functions (y, t) → wθ,α (y, t), θ ∈ Θ, α ∈ N
for which there exist a real-valued function B(·) with E[B(Y )4+ε ] < ∞, for some ε > 0, and
b0 ∈ (0, 1] such that, for each y
|wθ,α (y, t) − wθ0 ,α0 (y, t0 )| ≤ B(y) (θT , α, t)T − (θ0T , α0 , t0 )T
b0
e such that
for any θ, θ0 ∈ Θ, t, t0 ∈ Λ, and α, α0 ∈ N . Moreover, there exist θ, α and B(·)
e and E[B(Y
e )4+ε ] < ∞.
supt∈Λ |wθ,α (·, t)| ≤ B(·)
For (θ, h) ∈ Θ×Hn and α ∈ N, define
n
2
1X
U (θ, h; α) =
wθ,α Yi ,ZiT θ r̂hi ZiT θ; θ −r ZiT θ; θ I{z: f (zT θ;θ)≥c} (Zi )
n i=1
;
the kernel is a continuous probability density function K with the support in [−1, 1]. Moreover, K is of bounded variation and symmetric. Then,
U (θ, h; α) = −h4 C1 (θ; α) −
1
C2 (θ; α) + ρ (θ, h; α) ,
nh
where
"
0 T
#2
0
T
2r Z θ; θ f Z θ; θ
T
00
T
C1 (θ; α) =
E −wθ,α Y, Z θ r Z θ; θ +
I{z:f (zT θ;θ)≥c} (Z) ,
4
f (Z T θ; θ)
(
)
wθ,α Y, Z T θ
T
C2 (θ; α) = K2 E −
v
Z
θ;
θ
I{z: f (zθ;θ)≥c} (Z) ,
f (Z T θ; θ)
R
R
with K1 = u2 K (u) du, K2 = K 2 (u) du and
sup
ρ (θ, h; α) = oP h4 + (nh)−1 .
K12
θ∈Θ, h∈Hn ,α∈N
Lemma B.4 Assume that the conditions of Lemma B.3 hold. Let
n
1X
Te(θ, h; α) =
π Yi , ZiT ; α r̂hi ZiT θ; θ − r ZiT θ; θ I{z: f (zT θ0 ;θ0 )≥c} (Zi ) ,
n i=1
where, for some ε > 0, E [π (Y, Z; α) | Z] = 0 and E |π (Y, Z; α)|4+ε I{z: f (zT θ0 ;θ0 )≥c} (Z) <
∞, for any α. Then,
Te(θ, h; α) = oP h4 + n−1 h−1 ,
uniformly in h ∈ Hn , in θ ∈ Θn and in α ∈ N.
The following lemma will provide the order of the reminder term in the decomposition
(2.7). The proof relies on orthogonality conditions like (2.8) and (2.9).
26
Lemma B.5 Let A = z : f z T θ0 ; θ0 ≥ c ⊂ Rd , for some c > 0. Let φ (y, r; α) =
B (r; α) + C (r; α) y such that such that B 0 (r; α) + C 0 (r; α) r ≡ 0. Suppose that the Assumptions A.1, A.3, A.4 and A.7 to A.11 hold. Then
n
1 X
φ Yi , r̂hi ZiT θ; θ ; α − φ Yi , r ZiT θ; θ ; α IA (Zi )
R (θ, h; α) =
n i=1
n
1 X
−
φ Yi , r̂hi ZiT θ0 ; θ0 ; α − φ Yi , r ZiT θ0 ; θ0 ; α IA (Zi )
n i=1
2
1
1
h
4
√
= OP h + OP
+ OP
+ OP √
× OP (kθ − θ0 k)
nh2
n
n nh4
1
2
+ O h + OP √ 2
× OP kθ − θ0 k2
nh
when n → ∞, uniformly in α, uniformly in h ∈ n−(1/2−ε) , n−ε , with 0 < ε < 1/2, and
uniformly in θ ∈ Θn .
C
Appendix: proofs of the main results
b
b
Proof of Theorem 3.1. Assume for the moment that θ, h are defined by maximization
of Sb (θ, h; α
en ) which is defined with the fixed trimming IA (·)
(see
equation (2.6)). At the
b
b
end of the proof we show that the same conclusions hold for θ, h defined in (2.3) with the
data-driven √
trimming.
b By the decomposition (2.7) we have
Part I : n−asymptotic normality of θ.
Sb (θ, h; α
en ) = Se (θ; α
en ) + T (h; α∗ ) + R (θ, h; α
en ) .
Our objective is to show that R (θ, h; α
en ) is negligible when compared to Sb (θ; α
en ) from which
b
e
we deduce that θ behaves as the maximizer of S (θ; α
en ). Define
n
1 X
R1 (θ, h; α) =
ψ Yi , r̂hi ZiT θ; θ ; α − ψ Yi , r ZiT θ; θ ; α IA (Zi )
n i=1
and use Taylor expansion to write
R (θ, h; α
en ) = R1 (θ, h; α
en ) − R1 (θ0 , h; α∗ )
= [R1 (θ, h; α
en ) − R1 (θ0 , h; α
en )] + [R1 (θ0 , h; α
en ) − R1 (θ0 , h; α∗ )] .
Apply Lemma B.5 to obtain the order of R1 (θ, h; α
en ) − R1 (θ0 , h; α
en ) . Next, note that
27
R1 (θ0 , h; α
en ) − R1 (θ0 , h; α∗ ) does not depend on θ. Deduce that
2
1
1
h
4
√
+ OP
× OP (kθ − θ0 k)
R (θ, h; α
en ) = OP h + OP
+ OP √
nh2
n
n nh4
1
2
+ O h + OP √ 2 × OP kθ − θ0 k2
nh
+ {terms not depending on θ} ,
uniformly in h ∈ n−(1/2−ε) , n−ε , with 0 < ε < 1/2, uniformly in θ ∈ Θn and uniformly
with respect to {e
αn } . It follows that, up to terms not containing θ,
√
R (θ, h; α
en ) = oP kθ − θ0 k / n + oP kθ − θ0 k2 ,
uniformly in h ∈ Hn , θ ∈ Θn and uniformly with respect to {e
αn } .
∗
e
e
e
Now, write S (θ; α
en ) = S1 (θ; α
en ) − S1 (θ0 ; α ) with
n
1X
Se1 (θ, α) =
ψ Yi , r ZiT θ; θ ; α IA (Zi ) .
n i=1
Notice that
h
i h
i
Se (θ, α
en ) = Se1 (θ, α
en ) − Se1 (θ0 , α
en ) + Se1 (θ0 , α
en ) − Se1 (θ0 , α∗ ) ,
where the last difference does not contain θ, so that
arg max Se (θ, α
en ) = arg max[Se1 (θ, α
en ) − Se1 (θ0 , α
en )].
θ
θ
2 e
2 e
∗
Furthermore, for any α
en → α∗ , in probability, we have ∂θθ
en ) − ∂θθ
T S1 (θ0 , α
T S1 (θ0 , α ) =
oP (1) . Therefore, using the Taylor expansion we can write
Se1 (θ, α
en ) − Se1 (θ0 , α
en )
T
2 e
= (θ − θ0 ) ∂θ Se1 (θ0 , α
en ) + (θ − θ0 )T ∂θθ
en ) (θ − θ0 ) + oP kθ − θ0 k2
T S1 (θ0 , α
h
i
= (θ − θ0 )T ∂θ Se1 (θ0 , α∗ ) + (θ − θ0 )T ∂θ Se1 (θ0 , α
en ) − ∂θ Se1 (θ0 , α∗ )
∗
2 e
+ (θ − θ0 )T ∂θθ
T S1 (θ0 , α ) (θ − θ0 )
h
i
∗
2 e
2 e
+ (θ − θ0 )T ∂θθ
S
(θ
,
α
e
)
−
∂
S
(θ
,
α
)
(θ − θ0 )
T
0
1
0
n
θθT 1
+oP kθ − θ0 k2
h
i
1
= √ (θ − θ0 )T Vn + (θ − θ0 )T ∂θ Se1 (θ0 , α
en ) − ∂θ Se1 (θ0 , α∗ ) − (θ − θ0 )T Wn (θ − θ0 )
n
+oP kθ − θ0 k2 ,
28
uniformly in θ in oP (1) neighborhoods of θ0 , where
n
1 X
Vn = √
∂θ ψ Yi , r ZiT θ0 ; θ0 ; α∗ IA (Zi ) ,
n i=1
n
Wn
1X 2
= −
∂θθT ψ Yi , r ZiT θ0 ; θ0 ; α∗ IA (Zi )
n i=1
2 e
2
(here, ∂θ Se1 and ∂θ ψ are vectors in Rd−1 , while ∂θθ
T S1 and ∂θθ T ψ are (d−1)×(d−1) matrices).
Next, write
∂θ Se1 (θ0 , α
en ) − ∂θ Se1 (θ0 , α∗ ) = ∂α ∂θ Se1 (θ0 , α) [e
αn − α ∗ ]
(∂α denote the partial derivative with respect to α) with α between α
en and α∗ , and notice
that by the definition of ψ as the logarithm of a LEFN density, for any α,
h
i
E ∂α ∂θ Se1 (θ0 , α) = E ∂θ ∂α ψ Y, r Z T θ0 ; θ0 ; α IA (Z) = 0.
√
en → α ∗ ,
Consequently, ∂α ∂θ Se1 (θ0 , α) = OP (1/ n) , uniformly in α. Deduce that for all α
√
in probability, ∂θ Se1 (θ0 , α
en ) − ∂θ Se1 (θ0 , α∗ ) = oP (1/ n) . Use this fact and the order of
R (θ, h; α
en ) to write
1
1
Sb (θ, h; α
en ) = √ (θ − θ0 )T Vn − (θ − θ0 )T Wn (θ − θ0 )
(C.2)
2
n
√
+oP kθ − θ0 k / n + oP kθ − θ0 k2 + {terms not depending on θ} ,
uniformly in h ∈ Hn , θ ∈ Θn and α
en in oP (1) neighborhoods of α∗ . By little algebra,
h
i
∗
∗ T
T
T
I = E ∂θ ψ Y, rθ0 Z θ0 ; α ∂θ ψ Y, rθ0 Z θ0 ; α
IA (Z)
2
J = −E ∂θθ
ψ Y, rθ0 Z T θ0 ; α∗ IA (Z) .
Deduce from the assumptions that Vn converges in distribution to N (0, I) and Wn → J, in
probability. Finally, deduce that θb has the same asymptotic distributions as the maximizer
of the quadratic form (C.2). More precisely, apply Theorems 1 and 2 of Sherman (1994a) to
√
obtain first, the n−consistency of θb and next, the asymptotic normality
√
D
b
n θ − θ0 −→ N 0, J −1 IJ −1 .
Part II : the behavior of b
h. By Taylor expansion we can write
T (h; α∗ ) = T0 + T1 (h; α∗ ) + T2 (h; α∗ ) + {negligible terms} ,
where T0 is independent of h,
n
1X
T1 (h; α ) =
∂2 ψ Yi , r ZiT θ0 ; θ0 ; α∗ r̂hi ZiT θ0 ; θ0 − r ZiT θ0 ; θ0 IA (Zi ) ,
n i=1
∗
29
n
2
1X1 2
∂22 ψ Yi , r ZiT θ0 ; θ0 ; α∗ r̂hi ZiT θ0 ; θ0 − r ZiT θ0 ; θ0
IA (Zi ) .
T2 (h; α ) =
n i=1 2
∗
By Lemma B.3, T2 (h; α∗ ) = −C1 h4 − C2 /nh + oP (h4 + 1/nh), uniformly over Hn , with C1 ,
C2 defined in (3.1). Moreover, by Lemma B.4, T1 (h; α∗ ) = oP (T2 (h; α∗ )), uniformly over Hn .
Finally, recall that
R(θ, h; α
en ) = [R1 (θ, h; α
en ) − R1 (θ0 , h; α
en )] + [R1 (θ0 , h; α
en ) − R1 (θ0 , h; α∗ )] .
The order of R1 (θ0 , h; α
en ) and R1 (θ0 , h; α∗ ) can be obtained in the same way as the order of
∗
T (h; α ) . Taking the differences R1 (θ0 , h; α
en ) − R1 (θ0 , h; α∗ ) vanishes the constants of the
dominating terms containing h. Thus, up to terms independent of h, the second bracket is
negligible compared with T (h; α∗ ), uniformly in {e
αn } . On the other hand, by Lemma B.5,
the first bracket is of order oP (T2 (h; ; α∗ )), uniformly in θ in OP (n−1/2 ) neighborhoods of θ0
and h ∈ Hn and uniformly in {e
αn } in oP (1) neighborhood of α∗ . Since θb was shown to be
√
n−consistent, deduce that b
h is asymptotically equivalent to the maximizer of T2 (h; α∗ ).
1/5 −1/5
opt
n
.
More precisely, b
h/hopt
n → 1, in probability, where hn = (C2 /4C1 )
b
b
To close the proof it remains to show that (θ, h) defined in (2.3) is asymptotically equivalent to the maximizer of the objective function (θ, h) → Sb (θ, h; α
en ) in equation (2.6). For
this we use inequality (2.5) and Lemma B.2. Moreover, we can decompose Sb θ, h; α
en , Aδ
in the same way as Sb (θ, h; α
en ) and obtain the same orders, uniformly over Θn × Hn , uniformly with respect to {e
αn } and uniformly in δ ∈ [0, δ0 ], for some small δ0 . Note that Aδ
shrinks to the set z : f z T θ0 ; θ0 = c as δ → 0. Therefore, the constants appearing in
δ
b θ, h; α
the dominating
terms
of
the
decomposition
of
S
e
,
A
vanishes as δ → 0, provided
n
that P f Z T θ0 ; θ0 = c = 0. Consequently, the OP (·) orders are transformed in oP (·)
en , A)), uniformly in θ ∈ Θn , h ∈ Hn and α
en a
orders and thus Sb θ, h; α
en , Aδ = oP (Sb (θ, h; α
∗
sequence convergent to α , in probability, provided that δ → 0. The proof is complete.
REFERENCES
Andrews, D.W.K. (1988). Chi-Square diagnostic tests for econometric models: theory. Econometrica,
56, 1419-1453.
Andrews, D.W.K. (1995). Nonparametric kernel estimation for semiparametric models. Econometric
Theory, 11, 560-596.
Cameron, A.C., and Trivedi, P.K. (2013). Regression Analysis of Count Data. Second Edition. Econometric Society Monographs, Cambridge University Press.
Cui, X., Härdle, W. and Zhu, L. (2011). The EFM approach for single-index models. Ann. Statist.,
39, 1658–1688.
Delecroix, M., Hristache, M. and Patilea, V. (2006). On semiparametric M-estimation in singleindex regression. J. Statist. Plann. Inference, 136, 730–769
30
Gouriéroux, C. Monfort, A. and Trognon, A. (1984a). Pseudo maximum likelihood methods:
theory. Econometrica, 52, 681-700.
Gouriéroux, C. Monfort, A. and Trognon, A. (1984b). Pseudo maximum likelihood methods:
applications to Poisson models. Econometrica, 52, 701-720.
Härdle, W., Hall, P. and Ichimura, H. (1993). Optimal smoothing in single-index models. Ann.
Statist., 21, 157-178.
Härdle, W. and Marron, J.S. (1985). Optimal bandwidth selection in nonparametric regression function estimation. Ann. Statist., 13, 1465-1481.
Härdle, W. and Stoker, T.M. (1989). Investigating smooth multiple regression by the method of
average derivatives. J. Amer. Statist. Assoc., 84, 986-995.
Ichimura, H. (1993). Semiparametric least squares (SLS) and weighted SLS estimation of single-index
models. J. Econometrics, 58, 71-120.
Johnson, N.L., Kotz, S. and Balakrishnan N. (1997). Discrete Multivariate Distributions. New York,
John Wiley.
Newey, W.K. (1993). Efficient estimation of models with conditional moment restrictions, in G.S. Maddala, C.R. Rao and H.D. Vinod (eds.) Handbook of Statistics, vol. 11, pp. 419- 454, New-York:
North-Holland.
Newey, W.K. (1994). The asymptotic variance of semiparametric estimators. Econometrica, 62, 13491382.
Newey, W.K. and Stoker, T.M. (1993). Efficiency of weighted average derivative estimators and index
models. Econometrica, 61, 1199-1223.
Newey, W.K. and McFadden, D. (1994). Large sample estimation and hypothesis testing, in R.F.
Engle and D.L. McFadden (eds.) Handbook of Econometrics, vol. IV, pp. 2111- 2245, New-York:
North-Holland.
Pakes, A. and Pollard, D. (1989). Simulation and the asymptotics of optimization estimators. Econometrica, 57, 1027-1057.
Powell, J.L., Stock, J.M. and Stoker, T.M. (1989). Semiparametric estimation of index coefficients.
Econometrica, 57, 1403-1430.
Picone, G.A. and Butler, J.S. (2000). Semiparametric estimation of multiple equation models. Econometric Theory, 16, 551-575.
Sellar, C., Stoll, J.R. and Chavas, J.P. (1985). Validation of empirical measures of welfare change:
a comparison of nonmarket techniques. Land Economics, 61, 156-175.
Sherman, R.P. (1994a). Maximal inequalities for degenerate U −processes with applications to optimization estimators. Ann. Statist., 22, 439-459.
31
Sherman, R.P. (1994b). U -processes in the analysis of a generalized semiparametric regression estimator.
Econometric Theory, 10, 372-395.
Xia, Y. and Li, W.K. (1999). On single-index coefficient regression models. J. Amer. Statist. Assoc.,
94, 1275-1285.
Xia, Y., Tong, H. and Li, W.K. (1999). On extended partially linear single-index models. Biometrika,
86, 831-842.
Xia, Y., Tong, H., Li, W.K. and Zhu, L.-X. (2002) An adaptive estimation of dimension reduction
space (with discussions). J. R. Statist. Soc. B, 64, 363–410.
Ziegler, A. (2011). Generalized Estimating Equations. Lecture Notes in Statistics, Volume 204. NewYork: Springer.
32
| 10math.ST
|
Continuum contact models for coupled adhesion and friction
Janine C. Mergel1a , Riad Sahlib , Julien Scheibertb , Roger A. Sauer2,a
a
Graduate School AICES, RWTH Aachen University, Templergraben 55, 52056 Aachen, Germany
Laboratoire de Tribologie et Dynamique des Systèmes LTDS UMR5513, Ecole Centrale de Lyon,
ENISE, ENTPE, CNRS, F-69134, Ecully, France
arXiv:1803.00046v1 [cs.CE] 28 Feb 2018
b
Abstract: We develop two new continuum contact models for coupled adhesion and friction,
and discuss them in the context of existing models proposed in the literature. Our new models
are able to describe sliding friction even under tensile normal forces, which seems reasonable
for certain adhesion mechanisms. In contrast, existing continuum models for combined adhesion and friction typically include sliding friction only if local contact stresses are compressive.
Although such models work well for structures with sufficiently strong local compression, they
fail to capture sliding friction for soft and compliant systems (like adhesive pads), for which the
resistance to bending is low. This can be overcome with our new models. For further motivation, we additionally present experimental results for the onset of sliding of an elastomer cap on
a smooth glass plate under low normal loads. As shown, the findings from these experiments
agree well with the results from our models. In this paper we focus on the motivation and
derivation of our continuum contact models. Their implementation in a nonlinear finite element framework as well as the algorithmic treatment of adhesion and friction will be discussed
in future work.
Keywords: continuum contact mechanics, van der Waals adhesion, sliding friction, Amontons’
law, Coulomb’s law, adhesion-controlled friction
1
Introduction
This work is concerned with general contact models for adhesion mechanisms that are characterized by coupled adhesion and friction. Such mechanisms appear in both natural and technical
applications; they also include the adhesive microstructures of many insects and lizards, which
may resemble either soft adhesive pads (e.g. for flies and beetles), or rather stiff fibrils (e.g. for
the gecko). In particular the gecko mechanism has several other remarkable properties: It
enables fast detachment, self-cleaning, and non-clumping [6]. To replicate the characteristics
of such bio-adhesives, within the last one and a half decades various synthetic materials have
been developed [46, 135, 47, 45, 81, 11]. Like their natural counterparts, bio-inspired adhesives are often reusable and removable without leaving any residue. This makes them relevant
for different medical applications; other possible applications include climbing robots, microelectro-mechanical systems, and pick-and-place methods.
Initiated by the study of Autumn et al. [9], it is generally believed that bio-adhesion is mainly
caused by van der Waals forces. Nevertheless, also other effects may have an influence, such as
capillary [54] or electrostatic forces [56]. Within this paper, we mainly focus on dry adhesion
due to van der Waals interactions. As experiments reveal, the function of bio- and bio-inspired
adhesives further depends on friction (or “shear”) forces [7, 138]. Thus, the applications discussed so far do not only involve dry adhesion but also dry friction. Combined, adhesion and
1
2
Email: [email protected]
Corresponding author, email: [email protected]
1
friction play an important role in various other systems; these include small frictional hairs at
the legs of insects [68], the peeling of adhesive tapes [2], or elastomer spheres sliding on glass
plates [107, 127, 27, 97].
Regarding the adhesion and friction of bio-adhesive systems in general, experiments show that
friction forces can occur even if the normal force is negative, i.e., tensile [7, 37, 138]. A similar
observation can be made in experiments with chemically modified graphite surfaces [34]. Note
that these findings usually refer to the total force in the system, while locally the contact stresses
may vary between tension and compression within the same macroscopic contact area. This
was demonstrated by Eason et al. [40], who used a tactile sensor to investigate the distribution
of the normal stress underneath an attached gecko toe.
If an adhesive system has a large stiffness (or low compliance, respectively), its contact surface
tends to be characterized by regions of tensile and regions of compressive contact forces. Then,
the fraction of compressive and tensile areas is large compared to a more compliant system,
which may justify the application of continuum models that capture friction only at the locally
compressed parts of the contact area. However, for compliant, tape-like structures or soft
adhesive pads (e.g. of insects, see [37]), the assumption of a high stiffness is not realistic.
The continuum models that have been considered for combined adhesion and friction so far
include sliding friction only if the local contact stress is positive, i.e., compressive. Although
they work well for systems with sufficiently strong (local) compression, they fail to capture
sliding resistance for in the cases described above.
For this reason, we propose two general contact models for the analysis of physical problems
that are characterized by coupled adhesion and friction. The first contact model is based on a
constant sliding threshold within the contact area, which means the sliding shear is independent
of the local distance of the contacting surfaces. This model turns out to be favorable for low
normal loads, when friction is purely adhesion-controlled. The second model can be considered
as a local version of Amontons’ law extended to adhesion. Both models are capable of capturing
the frictional behavior for zero or negative (i.e. tensile) contact pressure. This approach is
motivated by findings for soft and compliant bio-adhesive pads, which are able to generate
friction forces under tensile normal loads. We additionally present new experimental results for
the onset of sliding of a smooth elastomer sphere on a glass plate. As shown, the findings from
these experiments agree well with the results for the proposed models. Our contact models are
formulated in the framework of large deformations, and apply to arbitrary geometries. For many
applications this requires applying computational methods for analysis. An implementation
of the proposed models in the framework of nonlinear finite element methods is addressed
separately [75, 77].
Note that in the following we will also refer to the terms sticking for static friction, and sliding
for kinetic (or dynamic) friction, respectively.
The remainder of this paper is structured as follows. In Sect. 2 we discuss various mechanisms
that may contribute to dry adhesion and friction, and outline existing modeling approaches.
We then present new experimental results for adhesive friction of a smooth elastomer sphere
on glass; see Sect. 3. In Sect. 4 we present the two proposed models for coupled adhesion and
friction, and discuss their general behavior. Sect. 5 then provides a comparison of the model
with the experiments, and Sect. 6 concludes this paper.
2
2
Origin and modeling of adhesive friction
This section provides an overview of relevant experimental findings, theoretical models, and
computational approaches that have been proposed in the past.
2.1
Adhesive and non-adhesive friction
Like adhesion, dry friction may stem from (combinations of) different mechanisms at smaller
length scales. This depends on the involved materials, surface properties, and the considered
length scale. Mechanisms that contribute to friction are classified in [80] as follows: 1) adhesion;
2) deformation of asperities; 3) plastic deformation at the interface; 4) fracture; 5) interlocking;
and 6) wear or contamination particles between the surfaces. Another possible mechanism is
viscous dissipation in the material, e.g. for rubber or various tissues. Although sliding friction
may be caused by a mixture of these mechanisms, in engineering applications it is often described
by the well-known formula
Ft = µ Fn ,
Fn > 0,
(1)
where µ is the coefficient of sliding friction, which relates the normal and tangential forces, Fn
and Ft , to each other. Eq. (1) results from the following three laws, which were named after
Guillaume Amontons and Charles de Coulomb:
1) The friction force is proportional to the applied normal load.
2) The friction force is independent of the apparent contact area.
3) The friction force is independent of the sliding
velocity.4
(Amontons’ 1st law)3
(Amontons’ 2nd law)
(Coulomb’s law)
We will refer to relation (1) also as Coulomb-Amontons friction. The three laws stated above can
be interpreted by introducing both an apparent (nominal) and a real (true) contact area, Areal ,
and assuming that compressive contact occurs only between small microasperities. Such an
approach goes back to Refs. [35, 18, 5], and explains the apparent contradiction in several
experiments showing the friction force to be proportional to the contact area [108, 41]. In many
applications, Eq. (1) describes frictional sliding of the system appropriately. Nevertheless, for
considerably smooth surfaces or soft materials (such as polymers), Eq. (1) may not be valid
anymore [87]. This was also shown for smooth elastomer contacts, see e.g. [27, 97].
In his study of adhesion and friction of crystal layers, Derjaguin [35] used statistical analysis to
derive a generalization of Amontons’ first law,
Ft = µ (F0 + Fn ) = µ Areal (p0 + pn ).
(2)
Here, pn := Fn /Areal is an averaged normal pressure, and p0 := F0 /Areal is an adhesion parameter. After introducing a material-specific, critical shear stress τ0 := µ p0 , Eq. (2) becomes
Ft = τ0 Areal + µ Fn = τ Areal ,
τ := τ0 + µ pn .
(3)
This will be referred to as extended Amontons’ law. According to Ref. [94], Eq. (2) describes
the friction force between dry surfaces that slide smoothly over each other in the presence of
adhesion. For low normal loads, the first term in Eq. (3) predominates, which is called adhesioncontrolled friction. For large normal loads, the second term (equivalent to Eq. (1)) predominates,
3
4
Note that in engineering mechanics, this is often referred to “Coulomb’s law” or “Coulomb friction” instead.
According to Ref. [80], Coulomb stated this relation for intermediate velocities.
3
which is denoted pressure-controlled friction.5 The influence of both terms strongly depends on
the specific application, and in particular on the smoothness, compliance, and adhesion of the
contacting surfaces. This is demonstrated by the experimental results [53] shown in Fig. 1. For
Removed for license reasons from arXiv
preprint. See Fig. 4(a) in Ref. [53].
Removed for license reasons from arXiv
preprint. See Fig. 4(b) in Ref. [53].
(a) Uncoated surfaces.
(b) Surfaces coated with calcium stearate.
Figure 1 Friction force and contact area measured for two mica cylinders sliding over each other
in dry air. Dashed arrows indicate the transition between adhesion- and pressure-dependent
friction.
two molecularly smooth (and initially undamaged) mica cylinders sliding over each other, the
friction force is directly proportional to the contact area; Ft = τ0 Areal . This means that with
increasing normal load, the friction force follows the triangles in the figure. At a sufficiently high
normal load, damage occurs and forms wear particles at the interface. Afterwards, the normal
load is decreased again. From that moment on, the friction force is proportional to the normal
load according to Eq. (1), and not to the contact area anymore; see the filled circles in Fig. 1. As
a comparison of Fig. 1(a) and Fig. 1(b) reveals, it depends on the specific parameters whether
either adhesion- or pressure-controlled friction provide larger friction forces. The transition
between adhesion- and pressure-controlled friction is further discussed in Refs. [13, 44, 57].
Eq. (3) has been considered in various research fields, sometimes with a different terminology. It
had already been discussed in a similar form before Derjaguin [35], also by Coulomb himself [44].
What he proposed was a fracture criterion for granular and brittle materials, known as the
Mohr- or Navier-Coulomb criterion, respectively [31, 88]. Besides, Tabor [115] used Eq. (3)
to derive a pressure-dependent friction model with an effective coefficient of sliding friction,
µ̂ := Ft /Fn = τ0 /pn + µ. Extended Amontons’ law was further discussed in Refs. [21, 117, 134].
It also appeared in the context of microtribology [20, 14, 94], molecular dynamics simulations
[44, 113, 78], and adhesive friction of bio- and bio-inspired systems [136, 52].
2.2
Existing modeling approaches
In the following, we give an overview of the most relevant models for dry adhesion and friction,
with a special focus on continuum methods that are suitable for computational formulations. For
a discussion of friction and tribology from a physical point of view, we refer to monographs [84,
16]. In addition, different aspects of nanotribology are addressed in book [15]. A recent review
on modeling and simulation in tribology is given in Ref. [124]. For an overview of computational
models for adhesion, see also [100].
In order to model adhesion (and repulsion) due to van der Waals interactions, various continuum
5
In [94] the expression “load-controlled” is used.
4
models have been developed based on the Lennard-Jones (LJ) potential
r 12
r 6
0
0
φ(r) = ε
− 2ε
.
r
r
(4)
Here, φ is the potential between two molecules separated by the distance r; it depends on
the parameters ε and r0 , which denote the depth of the potential well and the molecular
equilibrium distance, respectively. In order to increase the efficiency of computational models based on Eq. (4), the interactions between two (in general arbitrarily
shaped) bodies are
3
often approximated byeffective volumetric forces with unit N/m or effective surface tractions with unit N/m2 . Such an approach is pursued in various computational formulations
[104, 38, 106, 137, 58, 59, 42].
In general, models based on the Lennard-Jones potential do not contain any contribution due
to friction. For this reason, Deng et al. [34] additionally consider a constant tangential contact
stress during sliding wherever the normal contact stress is compressive. This corresponds to the
first term in Eq. (3), and thus shows similarities to one of our proposed models (Sect. 4.2.1).
Apart from that, Jiang and Park [60] propose to extend the original Lennard-Jones potential by
an additional “friction potential” in order to describe the friction properties of layered materials
such as graphene. However, this approach is restricted to static friction.
Van der Waals forces can also be modeled by fitting traction-separation laws within a cohesive
zone model; see e.g. Refs. [131, 83, 103, 105]. Although these papers use models that do not
include tangential sliding, there also exist numerous cohesive zone models that also include
sliding friction. In early works [122, 70], sliding friction according to Coulomb and Amontons
was modeled subsequently after full tangential debonding. More recent cohesive zone models
account for sliding friction also during partial separation [23, 93, 33, 114]. Besides, there exist
several cohesive zone models including friction, which are based on representative interface or
volume elements [1, 96, 95, 51].
In all of the cohesion models discussed so far, the damage caused by debonding is considered
to be irreversible. This, however, is not realistic for van der Waals interactions. Schryve [112]
and Cocou et al. [26] thus modified model [92] to capture partial recovering (“healing”) of the
adhesive strength when the gap between the interfaces is closed again. Schryve applied this
approach to investigate adhesion and friction between elastomer and glass.
Macroscopic adhesion and friction can also be modeled by statistical homogenization of the
microscopic contact behavior [128]. Tworzydlo et al. [123], for instance, combined finite element
simulations of small surface asperities with statistical analysis, mainly focusing on metallic
surfaces. In that study adhesion was modeled with an integrated Lennard-Jones potential.
Besides, there exists a recursive multiscale approach [129] to investigate contact of rubber and
rough, rigid surfaces (like tire-road contact). This model captures van der Waals interactions by
means of a fitted cohesive zone model, and restores adhesion fully when contact is re-established.
For completeness we also outline the most relevant models for non-adhesive friction. In a general
continuum framework, sliding law (1) is usually stated in terms of the normal and tangential
contact tractions, tn and tt ,
ktt k = µ tn ,
tn > 0.
(5)
In many engineering applications it is sufficient to consider the coefficient of sliding friction, µ, as
a constant parameter for the material pair at the interface; see e.g. Refs. [25, 119, 3, 4, 133].
Depending on the application, however, µ may be affected by the sliding velocity [50, 130],
contact pressure [110], temperature [50], or microscopic time scales [109, 116]. For an overview
of approaches suitable for different kinds of materials see Ref. [128]. In addition, there exists
also the class of rate-and-state models (see e.g. [12]), which further depend on contact duration
5
and contact wear. Such models are usually applied to investigate earthquakes and seismic
activities [84, 64].
Note that both the friction and the cohesive zone models discussed above include sliding friction
only under local compression. This means that after tangential debonding, frictionless sliding
occurs if the normal contact stresses are tensile.
Next, we outline the most relevant small-deformation models for adhesion and friction. Regarding the frictionless adhesion of spheres, these include the well-known JKR [63] and DMT
[36] models, which apply to compliant and stiff materials, respectively. Both models were later
combined in Ref. [73]. Although these models are restricted to small deformations and special
contact geometries, they have been successfully applied to fit experimental data. In the experimental results shown in Fig. 1, for instance, the contact area agrees well with that predicted by
the JKR theory. For adhesion-controlled friction, it is then possible to determine the frictional
shear stress τ0 (a material-pair parameter) from sphere-plane contact as follows: 1) Predict the
real contact area, Areal , with the JKR theory by inserting Young’s modulus, the sphere radius,
the applied normal load, and the interfacial adhesion energy; 2) measure the friction force, Ft ;
and 3) combine both values using Ft = τ0 Areal ; see also Sect. 3. In Ref. [107], the JKR theory
was extended further to incorporate the influence of tangential shearing. According to that
model, for a fixed normal force, the contact area decreases with increasing tangential force.
This agrees with the experimental findings presented therein. Related studies and important
extensions include Refs. [61, 62, 127].
As mentioned in Sect. 2.1, friction may be explained by small asperities sliding over each other.
Such models have been applied to study friction not only between macroscopic rough surfaces
[19, 120, 121], but also between molecularly smooth surfaces, and go back to models [89, 118].
Asperity-based friction models generally differ in the surface shape and the number of involved
length scales; examples include periodic and wavy substrate, fractal surfaces, or stochastic
profiles. Note that such approaches are also used in many of the papers outlined in Sect. 2.1, in
particular to verify a linear relation between the real contact area and the normal force (see also
[72, 85, 10]). Asperity-based models are not only used to investigate friction, but also adhesion.
Multiasperity contact was combined both with the JKR [43] and DMT theory [24, 74]. Further
extensions include Refs. [49, 91]. Like asperity-based friction models, models for rough surface
adhesion can be based e.g. on a single layer of wavy asperities or on single- and multiscale fractal
surfaces, respectively [82, 79]. For an overview see also Ref. [86].
Finally, there exists the analytical tape model [65] by Kendall, which was also extended numerous times to investigate directional peeling of thin strips or tapes. Some of these models
also incorporate sliding friction. As experiments show, approaches of this type are suitable to
accurately capture the peeling force of structures with small bending stiffness; these include
elastomer films [28], adhesive tape [30], and adhesive beetle pads [67]. These models are thus
often applied to discuss the properties of bio-adhesive systems. For adhesive fibrils with finite
bending stiffness, however, such models may underestimate the actual peel-off force considerably [99, 76].
3
Experimental results for sliding of a smooth elastomer sphere
We now present new experimental results for the onset of sliding of a smooth polydimethylsiloxane (PDMS) sphere on a smooth glass plate under low normal loads. Our experimental setup,
inspired by Ref. [90] and fully described in Ref. [97], is illustrated in Fig. 2. As shown, the
glass plate is attached to a double cantilever, which can deflect in the vertical direction (vertical
6
stiffness 410±20 N/m). The other end of the cantilever is mounted on a vertical translation
stage (not shown), which allows the glass plate to be brought into and out of contact with the
elastomer sphere. First contact is found by very slowly moving the double cantilever down and
stopping as soon as the sphere snaps into adhesive contact with the glass. Going further down
allows to obtain interfaces with increasing initial contact areas. The vertical displacement is
then kept constant during the experiments. Interestingly, once first contact has been formed,
it is possible to move the double cantilever back up some distance without losing contact. This
allows us to obtain interfaces in a global tensile state.
sphere: •
•
•
•
•
•
material: PDMS Sylgard 184
Young’s modulus: 1.6 ± 0.1 MPa
Poisson’s ratio: 0.5
radius of curvature: 9.42 mm
radius of cap: 12 mm
work of adhesion (PDMS – glass):
27 ± 1 mJ/m2
Figure 2 Monitoring of the incipient sliding of an elastomer sphere on a glass plate: Sketch of
the experimental setup.
The vertical translation stage itself is mounted on a motorized horizontal translation stage,
enabling motion of the glass plate at a constant velocity v = 0.1 mm/s. We measure the
time evolution of the tangential load, Ft , by means of a load sensor (resolution ca. 1 mN). In
synchronization, we monitor the time evolution of the contact area, Areal , by analyzing the
reflection images of the contact interface (see the method described in [97]). A typical evolution
of the contact area is shown in the snapshots of Fig. 3. As one can see, with increasing tangential
loading, the area shrinks along the loading direction, both at its left and right edges, while the
peeling front on the left moves faster than the one on the right.
0.00 s
0.45 s
0.62 s
0.73 s
0.88 s
1.25 s
1.75 s
Figure 3 Incipient sliding of a glass plate on an elastomer sphere: Snapshots of the contact
area for an initial area of 0.27 mm2 , which is the upper yellow curve in Fig. 4(b); the sphere is
fixed with respect to the frame of the camera, while the glass plate moves to the right.
Fig. 4(a) shows the tangential force, Ft , as a function of time, for different initial contact areas.
Here, the origin of time is taken when the motor starts moving. For each curve, the typical
behavior is the following: Ft increases, first almost linearly and then with a weakening slope,
reaches its maximum (the static friction peak), and rapidly drops afterwards before entering
a slow decay during macroscopic sliding. This slow decay, arising from a small residual angle
between the glass plate and the horizontal, was negligible in Ref. [97]. Here, it is detectable
because of the (about 20 times) smaller initial areas and the (about 8 times) stiffer cantilever.
7
The time evolution of the contact area, Areal , is shown in Fig. 4(b). As shown in Refs. [107,
127, 97], Areal decreases as the interface is progressively sheared. The rate of area decrease
(i.e., the slope of the curve) significantly drops when the contact enters macroscopic sliding.
The subsequent slow decrease of the area is the counterpart of the slow decrease of the sliding
friction force. Consistently with Ref. [127], contacts with the smallest initial areas (below about
0.1 mm2 ) abruptly vanish upon shearing, without entering a macroscopic sliding regime. In
contrast, the contacts with the highest initial areas (above about 0.19 mm2 ) do not vanish
during the time window shown in Figs. 4(a) and 4(b). For each experiment, this time window
covers the time that is necessary to reach the static friction peak (squares in Figs. 4(a) and 4(b)),
and to slide further by a distance that is equal to half of the width of the contact area when
the force becomes largest.
150
0.5
point of maximum Ft
0.4
$
#
Areal mm2
Ft [mN]
100
0.3
50
0.2
0
-50
0.1
0
1
2
3
0
4
0
1
t ! tstart [s]
2
3
4
t ! tstart [s]
(a) Tangential force over time.
(b) Contact area over time.
0.5
$
#
Areal mm2
0.4
0.3
0.2
1==0
0.1
0
0
20
40
60
80
100
120
140
Ft [mN]
(c) Contact area as a function of the tangential force.
Figure 4 Sliding of an elastomer sphere on a glass plate: Friction force and contact area
measured under low normal loads; the squares in Figs. (b) and (c) indicate the states with
maximum Ft (see Fig. (a)); the driving velocity is v = 0.1 mm/s.
Fig. 4(c) shows the dependency between the contact area and the tangential force in the incipient
loading phase, before the friction peak is reached. As can be seen, at the onset of sliding (see
the squares) the friction force and the contact area are, to a good approximation, proportional
to each other. This is consistent with experimental results on identical smooth sphere/plane
contacts [97], albeit for much larger contacts. It is also consistent with experiments on rough
contacts involving soft materials (see e.g. [130, 32, 132, 97]).
As our results demonstrate, considering the shear stress to be constant during sliding is a
8
reasonable assumption when modeling friction of soft and smooth surfaces. It further agrees
well with observations for adhesion-controlled friction, when Ft ≈ τ0 Areal according to Eq. (3).
For the present experiments, we measure a value of τ0 = 0.43 ± 0.01 MPa (see Fig. 4(c)).
4
Models for adhesive friction
Before formulating adhesive friction between two arbitrary objects mathematically, we must
first quantify the separation of their surfaces appropriately. To this end, we conceptually introduce variables for the distances in normal and tangential directions following classical contact
formulations [69, 128]. Using a common notation in continuum mechanics (see e.g. [17]), we
use uppercase letters for variables defined in the reference configuration of a body, and small
letters for variables in the current configuration.
For a given point xk on the contact surface of one of the two bodies Bk , k = 1, 2, we first
introduce its closest neighbor (or projection point), xp` , on the contact surface of the opposing
body B` , ` = 2, 1. At xp` , the outward unit normal vector of the surface is denoted np . Once
these quantities are determined, we define a normal gap vector, g n , as well as a (scalar) signed
normal gap, gn , as
(6)
gn = g n · np .
g n = xk − xp` ,
In addition, we conceptually introduce a vector for the tangential slip, g t , which reduces to
the scalar gt for 2D problems. This quantity contains the magnitude and direction of the
tangential displacement between the two surfaces. The normal and tangential gaps, gn and g t ,
as well as the normal vector, np , will be used to derive and illustrate our new contact models.
Their computation is addressed e.g. by Sauer and De Lorenzis [101, 102], while the algorithmic
treatment of combined adhesion and friction is discussed by Mergel et al. [75, 77].
4.1
Adhesive and repulsive contact
In order to describe general adhesion and repulsion between the two bodies under large deformations, we consider the coarse-grained contact model (CGCM) [98, 104, 106]. This model
is derived by first integrating the Lennard-Jones (LJ) potential from Eq. (4) over one of the
two bodies (index `), assuming that
its surface appears flat in the range of φ. This yields a
volumetric force with unit N/m3 , which acts at each point within the other body (index k).
In a second step, this body
force is projected onto the surface of body Bk to obtain the contact
traction with unit N/m2
θk
AH
1 r0 9 1 r0 3
T n,k =
−
.
(7)
Tn (gn ) np ,
Tn (gn ) =
J`
3 gn
2πr03 45 gn
Note that this contact law is stated in the undeformed reference configuration. As can be seen
in Fig. 5, Tn can be split into a repulsive (power 9) and an attractive (power 3) term. In
Eq. (7), the parameter AH = 2π 2 β01 β02 ε r06 is the Hamaker constant [55]; it contains the initial
molecular densities β0k of the two bodies and the parameters ε and r0 of the Lennard-Jones
potential. The scalar J` ≥ 0 denotes the volume change of the surrounding material during
deformation. It corresponds to the determinant of the deformation gradient of the body B`
[106]. If this body is rigid, incompressible, or considerably stiffer than body k, one can assume
that J` is either equal or close to one. The volume change can be related to the local surface
stretch Jc` by using J` = Jc` · λ` , where λ` is the stretch along the thickness (perpendicular to
9
3
integrated LJ potential
repulsive term
attractive term
regularization with greg = geq
Tn = T0
2
1
repulsion
geq =r0
0
attraction
!Tmax =T0
-1
0
0.5
Wadh =(T0 r0 )
1
1.5
2
2.5
3
gn = r0
Figure 5 Contact traction in the model of Sauer and Wriggers [106] (“surface force formulation”) for frictionless adhesion and repulsion; T0 = AH / (2πr03 ).
the surface). Since the potential φ(r) rapidly decays to zero at large separations, it is reasonable
to assume that within its effective range λ` ≈ 1, see also Sect. 2.3.2 of Ref. [101].
Eq. (7) contains a scalar, θk , that includes the current alignment of the two interacting surfaces
as well as the volume change of body Bk . Like with J` , θk requires the computation of the
deformation in the vicinity of surface point xk . If both surfaces are parallel, and if the influence
of the surface stretch of body Bk is negligible, one can set θk ≈ 1, as is shown in Ref. [106]. In
the following, we also assume that θk ≈ 1. An alternative approximation (similar to J` ≈ Jc` )
is proposed by Mergel in [75].
Before we proceed with the frictional part, let us define some characteristic parameters shown
in Fig. 5:
1. The equilibrium distance geq = r0 /
√
6
15 at which Tn (geq ) = 0 [99];
2. The work of adhesion per unit area for full separation, which is obtained by integrating
|Tn (gn )| from gn = geq to gn = ∞ [99],
√
3
15 AH
Wadh =
;
(8)
16πr02
3. The location gmax of the maximum adhesive traction Tmax ,
gmax
r0
,
= √
6
5
√
Tmax = |Tn (gmax )| =
5 AH
.
9πr03
(9)
Since Tn (gmax ) is negative (i.e., attractive), gmax is the global minimum of Tn .
Furthermore, in a computational implementation, one may increase robustness by regularizing
the normal traction for small normal gaps; see App. A and Fig. 5. This approach prevents
ill-conditioning caused by the slope of Tn approaching minus infinity for decreasing gn .
4.2
Frictional contact
We now propose two new phenomenological models that combine adhesion and repulsion with
sliding friction. To this end, we assume here that the sliding resistance is equal to the threshold
10
for static friction. This means that the tangential traction required to initiate the sliding
process agrees with the traction in the final sliding state. This assumption is physically justified
by experimental observations for both biological adhesives and rough elastomers, indicating
that for such systems static friction is comparable to kinetic friction [8, 7, 138, 48, 90]. For the
validity and restrictions of this assumption see also Sect. 4.3.
To shorten the notation, we now omit index k. Let tt (gn , g t ) denote the tangential traction
vector due to frictional sticking or sliding, which is a force per current area. tt depends on both
the normal gap from Eq. (6) and the tangential slip g t . We now assume that tt satisfies
(
< tslide (gn ) during sticking,
ktt (gn , g t )k
(10)
= tslide (gn ) during sliding,
where tslide (gn ) ≥ 0 is a function that defines the sliding threshold. For tslide we propose two
different approaches in the next two subsections. Note that the classical Coulomb-Amontons
law in Eq. (5) corresponds to tslide := µ tn for tn > 0.
4.2.1
Model DI: Distance-independent sliding friction in the contact area
The first proposed friction law is motivated by the experimental results discussed in Sect. 3. We
simply assume that the sliding threshold is independent of the distance gn , i.e., constant within
the current contact area. As a consequence, the resulting friction force is proportional to this
area if the entire body is sliding. Let us first define some cutoff distance, gcut , up to which the
surfaces are sufficiently close to each other in order to experience friction. After introducing a
constant parameter τDI > 0 for the frictional shear strength, we define
(
τDI , gn ≤ gcut ,
tslide (gn ) =
(11)
0,
gn > gcut .
Since this function is discontinuous at gn = gcut , we regularize it with the logistic function,
tslide (gn ) =
1+
τDI
,
k
DI
e (gn −gcut )
(12)
where kDI > 0 (with unit 1/m) is a sufficiently large parameter. Both the original and the
regularized models are illustrated in Fig. 6(a). Fig. 6(b) depicts the resulting friction law, i.e.,
the tangential traction-separation relation for arbitrary but fixed normal distances.
Note that for the regularized model tslide (gcut ) = τDI /2. As seen in Fig. 6, for model DI both
the value τDI and the cutoff distance gcut can be chosen independently of the normal traction.
Nevertheless, for the sake of comparison with the other model we introduce a coefficient that
relates the constant τDI to the maximum adhesive traction Tmax , defined in Eq. (9), as
µDI :=
τDI
.
Tmax
(13)
The sliding threshold τDI is related to the actual contact surface (in the current configuration).
This is motivated by the experiments from Sect. 3, for which the force is proportional to the
current size of the contact area. If we defined the parameter τDI in the reference configuration
instead, we would miss the change in contact area. The differences between those two approaches
are discussed in Ref. [75].
11
0.6
0.4
0.4
gn = geq
=DI =T0
0.2
0.2
tt = T0
Tn = T0 and tslide = T0
0.6
0
gmax =r0
gcut =r0
-0.2
-0.6
0.5
1
1.5
0
gn = 1:2 gcut
-0.2
Tn
Tn (regularized)
tslide
tslide (regularized)
-0.4
gn = gmax
-0.4
-0.6
-1
2
-0.5
gn = r0
0
0.5
1
gt = r0
(a) Sliding threshold for varying gn .
(b) Friction law in 2D for different but fixed gn .
Figure 6 Model DI: Constant sliding traction within the contact area defined by the cutoff
distance gcut ; τDI = 0.5 Tmax , gcut = 1.1 r0 , and T0 = AH / (2πr03 ); the colored asterisks mark
the normal gap values geq (where Tn = 0, light red), gmax (yellow), and 1.2 gcut (purple).
When recapitulating the unregularized version of model DI, one may recognize that Eq. (11) has
some similarities to one of the earliest cohesive zone models, the Dugdale model [39]. In contrast
to that model, however, we here define the sliding threshold for dynamic friction (instead of the
normal stress during pure debonding). As mentioned also in Sect. 2.2, a similar approach is used
in [34] to model sliding of graphene sheets. Therein, a constant stress during sliding is considered
for those parts of the surfaces that are in compressive contact. This would correspond to the
unregularized friction law (11) with gcut = geq . The current model is more general, because it
can also be used to describe sliding resistance for tensile contact.
4.2.2
Model EA: Extended Amontons’ law in local form
The second proposed traction-separation law is inspired by extended Amontons’ law (3). Importantly, that law was originally formulated in terms of force resultants and average tractions. In
our continuum formulation, however, the normal contact stress (7) can vary within the contact
area between attraction and repulsion. To include a tangential resistance against sliding even
for zero or negative normal pressure, we now propose a model that can be regarded as extended
Amontons’ law in local form.
Let us first shift the traction Tn in Eq. (7) by a value smaller than or equal to the maximum
adhesive traction, Tmax ; see also Eq. (9) and Fig. 5. To this end, we introduce a distance, gcut ,
which lies somewhere between the equilibrium distance, geq , and the location gmax of Tmax :
gcut = scut gmax + (1 − scut ) geq ,
scut ∈ [0, 1].
(14)
As can be seen in Figs. 7(a) and 7(e), in this range Tn (gcut ) is smaller than or equal to zero. We
then consider a sliding threshold that is proportional to the shifted curve Tn (gn ) + |Tn (gcut )|:
µEA Tn (gn ) − Tn (gcut ) , gn < gcut ,
Jc`
Tslide (gn ) =
(15)
0,
gn ≥ gcut .
Note that in this case, the sliding threshold, Tslide , directly depends on the normal traction, Tn ,
which is defined in the reference configuration (Sect. 4.1).
12
0.6
Tn
Tn (regularized)
Tslide
Tslide (regularized)
0.4
0.2
0.4
gn = 0:5 r0
0.2
(gcut = geq )
Tt = T0
Tn = T0 and Tslide = T0
0.6
0
gmax =r0
-0.2
gn = geq
0
gn = gmax
-0.2
-0.4
-0.4
scut = 0
-0.6
0.5
1
1.5
-0.6
-1
2
-0.5
gn = r0
(a) Sliding threshold for varying gn ; scut = 0.
0.5
1
(b) Friction law for different but fixed gn ; scut = 0.
0.6
0.6
Tn
Tn (regularized)
Tslide
Tslide (regularized)
0.4
0.2
gn = 0:5 r0
0.4
0.2
(gcut = (geq + gmax )=2)
Tt = T0
Tn = T0 and Tslide = T0
0
gt = r0
0
gmax =r0
-0.2
gn = geq
0
gn = gmax
-0.2
-0.4
-0.4
scut = 0:5
-0.6
0.5
1
1.5
-0.6
-1
2
-0.5
gn = r0
(c) Sliding threshold for varying gn ; scut = 0.5.
0.5
1
(d) Friction law for different but fixed gn ; scut = 0.5.
0.6
0.6
Tn
Tn (regularized)
Tslide
Tslide (regularized)
0.4
0.2
gn = 0:5 r0
0.4
0.2
(gcut = gmax )
Tt = T0
Tn = T0 and Tslide = T0
0
gt = r0
0
gmax =r0
-0.2
gn = geq
0
gn = gmax
-0.2
-0.4
-0.4
scut = 1
-0.6
0.5
1
1.5
-0.6
-1
2
gn = r0
-0.5
0
0.5
1
gt = r0
(e) Sliding threshold for varying gn ; scut = 1.
(f) Friction law for different but fixed gn ; scut = 1.
Figure 7 Model EA: Extended Amontons’ law in local form illustrated for µEA = 0.2, Jc` ≡ 1,
and three different values of scut ; T0 = AH / (2πr03 ); the colored asterisks mark the normal gap
values 0.5 r0 (green), geq (where Tn = 0, light red), and gmax (yellow).
13
Fig. 7 illustrates model EA for three different values of the parameter scut . The left-hand side of
the figure shows the dependence of the sliding traction, Tslide , on the normal gap, gn ; dashed lines
indicate a regularized version according to App. A. If scut = 0 (Figs. 7(a) and 7(b)), tangential
sliding occurs only for positive — i.e. compressive — normal tractions. This corresponds to
classical Coulomb-Amontons friction (see Eq. (5)) for non-adhesive contact. It is further used
in many cohesive zone models (see Sect. 2.2) to include frictional sliding. If scut > 0, a tangential
sliding resistance is present even for tensile normal tractions. Note that the curve for Tslide is
smooth (C 1 -continuous) only for scut = 1 (Fig. 7(e)); otherwise, a kink occurs at gn = gcut ,
which requires special treatment in a computational implementation [75]. For the particular
case scut = 0, this kink is exactly located at the equilibrium position, geq .
4
gn = geq
Tt = (7EA T0 )
gn = geq
0
A
!
gn = gmax
-2
-4
-1
gn = geq
gn = 0:5 r0
A
A
A
model EA, scut = 0
sliding law (5) with 7 = 7EA
gn ! 1
0
1
gn = geq
A
3
gn = gmax
A
model EA, scut = 0:5
sliding law (5) with 7 = 7EA
gn ! 1
-4
-1
4
A
A
!
A
0
-2
A
2
A
2
Tt = (7EA T0 )
gn = 0:5 r0
2
4
A
0
1
Tn = T0
A
2
3
4
Tn = T0
(a) Model EA with gcut = geq .
(b) Model EA with gcut = (geq + gmax )/2.
4
gn = geq
gn = gmax
1
A
gn = 0:5 r0
A
!
A
gn = gmax
A
-2
-4
-1
0
1
2
A
-1
3
4
A
!
-0.5
Tn = T0
gn = geq
gn ! 1
!
0
-0.5
model EA, scut = 1
sliding law (5) with 7 = 7EA
gn ! 1
A
!
0
A
tt = =DI
gn = geq
0.5
A
!
Tt = (7EA T0 )
2
gn = geq
A
!
model DI
DI (regularized)
A
gn = gcut
A
0
0.5
1
Tn = T0
(c) Model EA with gcut = gmax .
(d) Model DI.
Figure 8 Tangential traction during sliding vs. normal traction for (a) – (c) model EA with
Jc` ≡ 1 and (d) model DI; T0 = AH / (2πr03 ); the blue arrows indicate the sequence from gn = 0
to gn → ∞; for the colored asterisks see Figs. 6 and 7, respectively.
Fig. 8 depicts the (directed) tangential traction during sliding versus the normal traction, either
for model EA and different values of scut , or for model DI. As shown, for model EA this relation
resembles a shifted version of the cone describing the classical friction law (5). In fact, up to
gn ≤ gcut , it corresponds to the Mohr-Coulomb model mentioned in Sect. 2.2. In model DI, the
tangential traction, tt , is not a function of the normal traction, tn , anymore.
14
4.3
General comments, validity, and restrictions
Like any other model, our models have limitations. We address those in the following.
In Eq. (10) we assume that the static friction threshold coincides with the resistance for kinetic
friction. As mentioned in the beginning of Sect. 4.2, this assumption agrees well with experimental observations for bio-adhesive systems. Nevertheless, these observations refer to the globally
measured force resultant instead of the local traction at the contact zone. It remains to be
discussed further whether static and kinetic friction coincide because of the bio-material itself,
or whether this is caused by a split of the effective contact area into a large number of small
areas. The influence of such a split on both static friction and stick-slip motion is addressed
in Refs. [125, 126, 71]. Regarding bio-inspired adhesives, it may be arguable whether the assumption of equal static and kinetic friction is still valid. See, for instance, the experimental
results [125] obtained for a microstructured polyvinyl siloxane (PVS) surface sliding on glass. If
required, our model might be extended to account for differing parameters for static and kinetic
friction.
In our model we consider dry (i.e., non-lubricated) adhesion and friction. We hence omit the
influence of any secretion that may cover the adhesive device, as observed for many beetles
and other insects. In fact, experimental studies on insect pads show that the depletion of
such secretion (caused e.g. by sliding or by repeating detachment) affects both their frictional
resistance [37, 22] and (although less strongly) their adhesion [66]. Nevertheless, if the amount
of sliding is sufficiently small, one can assume that the frictional resistance does not change
considerably. In order to include the influence of a finite amount of fluid, it would be possible
to adjust the friction parameter with the distance of sliding.
As shown in Sects. 1 and 3, there exist several applications for which van der Waals forces
may affect the macroscopic behavior even at larger scales (micrometers or millimeters). For
increasing length scales, a direct computational implementation of adhesion models based on
the Lennard-Jones potential, such as the CGCM model (Sect. 4.1) and friction model EA, can
become very inefficient, because they require nanoscale finite element resolution. For this reason,
it would be very promising to develop an effective adhesion model that is regularized by the
compliance of the surrounding material. An approach in this direction has been developed in the
context of adhesive joints under small deformations [111]. Another possibility, which is pursued
in Ref. [75], is to calibrate the parameters in Sect. 4.1 (like AH , Tmax , or Wadh ) such that they
match with experimental data. The curve (7) is then regularized by an automatic increase of the
length parameter r0 . As a consequence, the model does not require nanoscale resolution. When
adjusting these parameters it is important to distinguish between the nominal (or apparent)
contact area and the true contact area due to very small asperities, which interact at the contact
surface; see also the discussion in Sect. 2.1. For the CGCM of Sect. 4.1 the normal contact stress
is integrated over an apparent, nominally flat contact area. Thus, simply inserting the material
constant AH for the considered pair of materials would overestimate the real strength of adhesion
by several orders of magnitude. This can be overcome by directly inserting effectively measured
values as described before. Note that in general, the ratio between the true and the nominal
contact areas may depend on the contact pressure, as discussed in Sect. 5.2.4 of Ref. [128] from
a computational point of view.
15
5
Qualitative comparison between model and experiment
We now apply the proposed two models to study adhesive friction between a soft cylindrical
cap and a rigid substrate, see Fig. 9(a).
(a) Problem setup.
(b) Sliding under zero normal load.
Figure 9 Adhesion of a soft cap: (a) Setup; (b) stress distribution during sliding for zero
preload and model DI (gcut = gmax , µDI = 1); the colors show the first invariant of the Cauchy
stress in units of E.
We consider this example as a preliminary study to investigate the qualitative behavior of our
models, and show that it agrees with the experimental behavior presented in Sect. 3. For length
scale reasons (addressed in Sect. 4.3), a quantitative comparison is not possible at this stage.
Therefore, the cap in the example is 2D and smaller than that in Sect. 3. Also slightly different
material parameters (Tab. 1) and smaller friction values are used. A 2D plane strain, nonlinear
finite element formulation based on a Neo-Hooke material model is used to simulate the example.
For further details see Ref. [75]. All results shown here are normalized by Young’s modulus, E,
an arbitrary out-of-plane thickness (or width), W , and a unit length, L0 (see Tab. 1).
E
ν
Tmax
Wadh
R
H
L0
2 MPa
0.4
0.33 MPa
0.027 J/m2
47.1 L0
10 L0
1 µm
Table 1 Adhesion of a soft cap: Parameters.
As illustrated in Fig. 9(a), we slide the cap horizontally after prescribing a fixed normal force. As
a special case we investigate the sliding behavior also under zero load, for which the attractive
and repulsive stresses in the contact area equilibrate each other. The finite element mesh consists
of 42,300 Q1N2.1 elements [29]. Fig. 9(b) shows the cap during sliding under zero load. Fig. 10
shows both the friction force and the contact area for different loads.
Here we use models DI with gcut = gmax , EA with scut = 1, and EA with scut = 0 (CoulombAmontons friction, see also Eq. (5)). Before full sliding, the qualitative behavior of all models
is close to our experimental results in Figs. 4(a) and 4(b). For classical Coulomb-Amontons
friction, however, the sliding force is considerably lower than for the other models, because the
compressed area is much smaller than the total contact area. When the entire contact surface is
sliding, both the force and the area remain constant for the numerical models. In contrast, in the
experiments, the tangential force drops over a finite time scale before entering a rather steady
sliding regime. These differences are most likely caused by two phenomena, partly addressed in
Sect. 4.3 already: The overshoot followed by a drop-off most probably results from the static
friction threshold being larger than kinetic friction for the elastomeric contact. The gradual
decrease over a finite time is likely caused by the viscosity in the bulk material. It thus makes
sense to extend our model in future work to account for different static and kinetic friction
parameters in the contact law (see Sect. 4.3), and to include a viscoelastic material model.
16
2
12
area for max. force
11.5
Areal = [L0 W ]
Fx = [E L0 W ]
1.5
1
0.5
0
DI, gcut = gmax
EA, scut = 1
EA, scut = 0 (Coulomb-Amontons)
0
1
2
3
4
5
11
10.5
10
9.5
9
6
0
1
2
u
7 = L0
3
4
5
6
u
7 = L0
(a) Friction force under zero preload, Fy∗ = 0.
(b) Contact area under zero preload, Fy∗ = 0.
1
9
area for max. force
8
Areal = [L0 W ]
Fx = [E L0 W ]
0.8
0.6
0.4
0
6
DI, gcut = gmax
EA, scut = 1
EA, scut = 0 (Coulomb-Amontons)
0.2
0
1
2
3
4
5
7
5
6
u
7 = L0
0
1
2
3
4
5
6
u
7 = L0
(c) Friction force under tensile preload, Fy∗ = 0.288.
(d) Contact area under tensile preload, Fy∗ = 0.288.
Figure 10 Adhesion of a soft cap: Friction force and contact area for sliding under a constant
preload (µEA = µDI = 1); the squares
indicate
the onset of full sliding; the dots mark the
∗
configurations of Fig. 12; Fy = Fy / E L0 W denotes the normalized preload.
Fig. 11 (like Fig. 4(c)) shows the contact area versus the tangential force up to the point of full
sliding. The classical law by Coulomb and Amontons (model EA for scut = 0) obviously fails
to capture the qualitative behavior observed in the experiments. In contrast, the differences to
the experimental results are smaller for model EA with scut = 1, but they are still considerable.
Model DI on the other hand, agrees very well with the experiments.
Finally, Fig. 12 shows (a vertically stretched view of) the contact interface for model DI. As a
comparison with Fig. 3 reveals, model DI appropriately captures also the qualitative behavior
of the two peeling fronts at the right (slowest front) and left (fastest front) edges. In summary,
combining model DI with a viscous material model and initial stiction (i.e., different parameters
for static and kinetic friction) seems a very promising approach to model sliding of smooth
rubber spheres on glass.
6
Conclusion
In this paper, we present two new continuum contact models for coupled adhesion and friction
that are suitable for soft systems such as biological and bio-inspired adhesives. The two models
are based on the 3D continuum adhesion formulation [106], which is suitable to describe large
17
14
Areal = [L0 W ]
12
zero load
10
8
1==DI
tensile load
6
4
DI, gcut = gmax
EA, scut = 1
EA, scut = 0 (Coulomb-Amontons)
2
0
0
0.5
1
1.5
Fx = [E L0 W ]
Figure 11 Adhesion of a soft cap: Contact area as a function of the tangential force for the
onset of sliding under different preloads (µEA = µDI = 1); τDI = µDI Tmax .
ū = 0.5 L0
ū = 1.0 L0
ū = 1.5 L0
ū = 2.0 L0
ū = 2.5 L0
ū = 3.0 L0
ū = 3.5 L0
ū = 4.0 L0
Figure 12 Adhesion of a soft cap: Evolution of the contact interface during sliding for model DI
(see the dots in Fig. 10(b)); the vertical axis is stretched by the factor ≈ 3.93; the configuration
at zero tangential displacement (ū = 0) is shown in red; the colors show the first invariant of
the Cauchy stress in units of E.
deformations and large sliding motions. We motivate the proposed models in the context of
existing theoretical and experimental findings. We also present new experimental results for a
PDMS cap sliding on a smooth glass plate under low normal loads. Our findings indicate that
the classical sliding law by Coulomb and Amontons (with linear dependence between normal and
tangential loads) is not applicable for strong adhesion or small roughnesses. Instead, the friction
force is also affected by an additional adhesion term. For some applications, this additional term
dominates, and the sliding resistance can thus be described by a constant, material-dependent
frictional shear strength multiplied with the real contact area. As a consequence, the proposed
models can generate sliding friction even when the normal pressure is zero or negative (i.e.,
tensile). This is useful because soft bio-adhesive pads are observed to generate non-negligible
sliding forces under zero normal load. The proposed models contain the law of Coulomb and
Amontons as a special case. We investigate their behavior during adhesive sliding of a PDMS
cap and glass, and show that the model behavior is in qualitative agreement to experiments.
In addition, it is shown in [75] that model DI is suitable to describe also the characteristics of
friction devices in stick insects. Also, for some applications models EA and DI show a very
similar behavior as investigated in Refs. [75, 77].
18
This paper focuses on the motivation and derivation of our new contact models. Both their
implementation in terms of a nonlinear finite element framework and the algorithmic treatment
of adhesive friction are discussed separately [75, 77].
In future work, several extensions of the model should be developed in order to overcome the
restrictions mentioned in Sect. 4.3. The most crucial one, is the restriction to small length
scales that is inherent to van der Waals based contact formulations. A contact homogenization
technique should be able to overcome this. Other important model extensions include the
incorporation of differing static and kinetic friction, and the incorporation of viscosity in the
bulk material. The current comparison between model and experiment is qualitative. With a
refined model, it will be possible to conduct quantitative comparisons. For this purpose, further
experiments are called for. In particular, it would be interesting to apply model EA to problems
in which both adhesion- and pressure-controlled friction play a major role.
A
Regularization for small normal distances
To avoid ill-conditioning, it is possible to regularize the contact stress for small normal gaps,
gn → 0, e.g. by means of linear extrapolation. To this end, we introduce Tn (gn ) ≈ Tnreg (gn ),
(
Tn (gn ),
gn ≥ greg ,
reg
Tn (gn ) =
(16)
0
Tn (greg ) + Tn (greg ) (gn − greg ), gn < greg ;
see also Fig. 5. Tn0 denotes the derivative of Tn with respect to gn . It is reasonable to choose
a regularization distance of greg ≈ geq , where geq is the equilibrium distance. For greg ≤ geq ,
the attractive part of the curve in Fig. 5 remains unaffected; if greg > geq , one also modifies the
work of adhesion, Wadh , in Eq. (8), and shifts geq to another position.
Acknowledgements
The authors are grateful to the German Research Foundation (DFG) for supporting this research
under grants SA1822/5-1 and GSC111. They also thank Dr. David Labonte (University of Cambridge) for helpful comments, and Udit Pillai (RWTH Aachen University) for his help with the
FE mesh of the cylindrical cap. This work was further supported by LABEX MANUTECHSISE (ANR-10-LABX-0075) of Université de Lyon, within the program Investissements d’Avenir
(ANR-11-IDEX-0007) operated by the French National Research Agency (ANR). In addition,
it received funding from the People Program (Marie Curie Actions) of the European Union’s
Seventh Framework Program (FP7/2007–2013) under Research Executive Agency Grant Agreement PCIG-GA-2011-303871.
References
[1] G. Alfano and E. Sacco. Combining interface damage and friction in a cohesive-zone model. Int. J. Numer.
Methods Eng., 68(5):542–582, 2006.
[2] N. Amouroux, J. Petit, and L. Léger. Role of interfacial resistance to shear stress on adhesive peel strength.
Langmuir, 17:6510–6517, 2001.
[3] D. S. Amundsen, J. Scheibert, K. Thøgersen, J. Trømborg, and A. Malthe-Sørenssen. 1D model of precursors
to frictional stick-slip motion allowing for robust comparison with experiments. Tribol. Lett., 45:357–369,
2012.
19
[4] D. S. Amundsen, J. K. Trømborg, K. Thøgersen, E. Katzav, A. Malthe-Sørenssen, and J. Scheibert.
Steady-state propagation speed of rupture fronts along one-dimensional frictional interfaces. Phys. Rev. E,
92:032406, 2015.
[5] J. F. Archard. Elastic deformation and the laws of friction. Proc. R. Soc. London A, 243:190–205, 1957.
[6] K. Autumn. Properties, principles, and parameters of the gecko adhesive system. In A. M. Smith and J. A.
Callow, editors, Biological Adhesives, pages 225–256. Springer, Berlin Heidelberg, 2006.
[7] K. Autumn, A. Dittmore, D. Santos, M. Spenko, and M. Cutkosky. Frictional adhesion: A new angle on
gecko attachment. J. Exp. Biol., 209:3569–3579, 2006.
[8] K. Autumn, Y. A. Liang, S. T. Hsieh, W. Zesch, W. P. Chan, T. W. Kenny, R. Fearing, and R. J. Full.
Adhesive force of a single gecko foot-hair. Nature, 405:681–685, 2000.
[9] K. Autumn, M. Sitti, Y. A. Liang, A. M. Peattie, W. R. Hansen, S. Sponberg, T. W. Kenny, R. Fearing,
J. N. Israelachvili, and R. J. Full. Evidence for van der Waals adhesion in gecko setae. Proc. Natl. Acad.
Sci. U.S.A., 99(19):12252–12256, 2002.
[10] J. R. Barber. Multiscale surfaces and Amontons’ law of friction. Tribol. Lett., 49:539–543, 2013.
[11] M. D. Bartlett, A. B. Croll, D. R. King, B. M. Paret, D. J. Irschick, and A. J. Crosby. Looking beyond
fibrillar features to scale gecko-like adhesion. Adv. Mater., 24:1078–1083, 2012.
[12] T. Baumberger and C. Caroli. Solid friction from stick-slip down to pinning and aging. Adv. Phys.,
55(3–4):279–348, 2006.
[13] A. Berman, C. Drummond, and J. Israelachvili. Amontons’ law at the molecular level. Tribol. Lett.,
4:95–101, 1998.
[14] A. Berman, S. Steinberg, S. Campbell, A. Ulman, and J. Israelachvili. Controlled microtribology of a metal
oxide surface. Tribol. Lett., 4:43–48, 1998.
[15] B. Bhushan, editor. Nanotribology and Nanomechanics. Springer, Berlin Heidelberg, 2005.
[16] B. Bhushan. Introduction to Tribology. Wiley, New York, 2nd edition, 2013.
[17] J. Bonet and R. D. Wood. Nonlinear Continuum Mechanics for Finite Element Analysis. Cambridge
University Press, 1997.
[18] F. P. Bowden and D. Tabor. The area of contact between stationary and between moving surfaces. Proc.
R. Soc. London A, 169:391–413, 1939.
[19] O. M. Braun and M. Peyrard. Modeling friction on a mesoscale: Master equation for the earthquakelike
model. Phys. Rev. Lett., 100:125501, 2008.
[20] B. J. Briscoe and S. L. Kremnitzer. A study of the friction and adhesion of polyethylene-terephthalate
monofilaments. J. Phys. D: Appl. Phys., 12:505–517, 1979.
[21] B. J. Briscoe and D. Tabor. Friction and wear of polymers: the role of mechanical properties. Br. Polym.
J., 10:74–78, 1978.
[22] J. M. R. Bullock, P. Drechsler, and W. Federle. Comparison of smooth and hairy attachment pads in
insects: friction, adhesion and mechanisms for direction-dependence. J. Exp. Biol., 211:3333–3343, 2008.
[23] J. L. Chaboche, R. Girard, and A. Schaff. Numerical analysis of composite systems by using interphase/interface models. Comput. Mech., 20:3–11, 1997.
[24] W. R. Chang, I. Etsion, and D. B. Bogy. Adhesion model for metallic rough surfaces. J. Tribol., 110:50–55,
1988.
[25] M. Ciavarella. The generalized Cattaneo partial slip plane contact problem. II — Examples. Int. J. Solids
Struct., 35(18):2363–2378, 1998.
[26] M. Cocou, M. Schryve, and M. Raous. A dynamic unilateral contact problem with adhesion and friction
in viscoelasticity. Z. angew. Math. Phys. (ZAMP), 61:721–743, 2010.
[27] C. Cohen, F. Restagno, C. Poulard, and L. Léger. Incidence of the molecular organization on friction at
soft polymer interfaces. Soft Matter, 7:8535–8541, 2011.
[28] R. R. Collino, N. R. Philips, M. N. Rossol, R. M. McMeeking, and M. R. Begley. Detachment of compliant
films adhered to stiff substrates via van der Waals interactions: role of frictional sliding during peeling. J.
R. Soc. Interface, 11:20140453, 2014.
[29] C. J. Corbett and R. A. Sauer. NURBS-enriched contact finite elements. Comput. Methods Appl. Mech.
Eng., 275:55–75, 2014.
[30] M.-J. Dalbe, R. Villey, M. Ciccotti, S. Santucci, P.-P. Cortet, and L. Vanel. Inertial and stick-slip regimes
of unstable adhesive tape peeling. Soft Matter, 12:4537–4548, 2016.
20
[31] F. V. de Blasio. Introduction to the Physics of Landslides. Lecture Notes on the Dynamics of Mass Wasting.
Springer, 2011.
[32] E. Degrandi-Contraires, C. Poulard, F. Restagno, and L. Léger. Sliding friction at soft micropatterned
elastomer interfaces. Faraday Discuss., 156:255–265, 2012.
[33] G. Del Piero and M. Raous. A unified model for adhesive interfaces with damage, viscosity, and friction.
Eur. J. Mech. A/Solids, 29:496–507, 2010.
[34] Z. Deng, A. Smolyanitsky, Q. Li, X.-Q. Feng, and R. J. Cannara. Adhesion-dependent negative friction
coefficient on chemically modified graphite at the nanoscale. Nat. Mater., 11:1032–1037, 2012.
[35] B. Derjaguin. Molekulartheorie der äußeren Reibung. Z. Phys., 88:661–675, 1934.
[36] B. V. Derjaguin, V. M. Muller, and Y. P. Toporov. Effect of contact deformations on the adhesion of
particles. J. Colloid Interface Sci., 53(2):314–325, 1975.
[37] P. Drechsler and W. Federle. Biomechanics of smooth adhesive pads in insects: influence of tarsal secretion
on attachment performance. J. Comp. Physiol. A, 192:1213–1222, 2006.
[38] Y. Du, L. Chen, N. E. McGruer, G. G. Adams, and I. Etsion. A finite element model of loading and
unloading of an asperity contact with adhesion and plasticity. J. Colloid Interface Sci., 312:522–528, 2007.
[39] D. S. Dugdale. Yielding of steel sheets containing slits. J. Mech. Phys. Solids, 8:100–104, 1960.
[40] E. V. Eason, E. W. Hawkes, M. Windheim, D. L. Christensen, T. Libby, and M. R. Cutkosky. Stress
distribution and contact area measurements of a gecko toe using a high-resolution tactile sensor. Bioinspir.
Biomim., 10:016013, 2015.
[41] M. Enachescu, R. J. A. van den Oetelaar, R. W. Carpick, D. F. Ogletree, C. F. J. Flipse, and M. Salmeron.
Observation of proportionality between friction and contact area at the nanometer scale. Tribol. Lett.,
7:73–78, 1999.
[42] H. Fan and S. Li. A three-dimensional surface stress tensor formulation for simulation of adhesive contact
in finite deformation. Int. J. Numer. Methods Eng., 107(3):252–270, 2015.
[43] K. N. G. Fuller and D. Tabor. The effect of surface roughness on the adhesion of elastic solids. Proc. R.
Soc. London A, 345:327–342, 1975.
[44] J. Gao, W. D. Luedtke, D. Gourdon, M. Ruths, J. N. Israelachvili, and U. Landman. Frictional forces and
Amontons’ law: From the molecular to the macroscopic scale. J. Phys. Chem. B, 108:3410–3425, 2004.
[45] L. Ge, S. Sethi, L. Ci, P. M. Ajayan, and A. Dhinojwala. Carbon nanotube-based synthetic gecko tapes.
Proc. Natl. Acad. Sci. U.S.A., 104(26):10792–10795, 2007.
[46] A. K. Geim, S. V. Dubonos, I. V. Grigorieva, K. S. Novoselov, A. A. Zhukov, and S. Y. Shapoval. Microfabricated adhesive mimicking gecko foot-hair. Nat. Mater., 2:461–463, 2003.
[47] S. Gorb, M. Varenberg, A. Peressadko, and J. Tuma. Biomimetic mushroom-shaped fibrillar adhesive
microstructure. J. R. Soc. Interface, 4:271–275, 2007.
[48] N. Gravish, M. Wilkinson, S. Sponberg, A. Parness, N. Esparza, D. Soto, T. Yamaguchi, M. Broide,
M. Cutkosky, C. Creton, and K. Autumn. Rate-dependent frictional adhesion in natural and synthetic
gecko setae. J. R. Soc. Interface, 7:259–269, 2010.
[49] J. A. Greenwood and K. L. Johnson. The mechanics of adhesion of viscoelastic solids. Phil. Mag. A,
43(3):697–711, 1981.
[50] K. A. Grosch. The relation between the friction and visco-elastic properties of rubber. Proc. R. Soc. A,
274:21–39, 1963.
[51] I. Guiamatsia and G. D. Nguyen. A thermodynamics-based cohesive model for interface debonding and
friction. Int. J. Solids Struct., 51:647–659, 2014.
[52] G. C. Hill, D. R. Soto, A. M. Peattie, R. J. Full, and T. W. Kenny. Orientation angle and the adhesion of
single gecko setae. J. R. Soc. Interface, 8:926–933, 2011.
[53] A. M. Homola, J. N. Israelachvili, P. M. McGuiggan, and M. L. Gee. Fundamental experimental studies in
tribology: The transition from “interfacial” friction of undamaged molecularly smooth surfaces to “normal”
friction with wear. Wear, 136:65–83, 1990.
[54] G. Huber, H. Mantz, R. Spolenak, K. Mecke, K. Jacobs, S. N. Gorb, and E. Arzt. Evidence for capillarity
contributions to gecko adhesion from single spatula nanomechanical measurements. Proc. Natl. Acad. Sci.
U.S.A., 102(45):16293–16296, 2005.
[55] J. N. Israelachvili. Intermolecular and Surface Forces. Academic Press, 3rd edition, 2011.
[56] H. Izadi, K. M. E. Stewart, and A. Penlidis. Role of contact electrification and electrostatic interactions in
gecko adhesion. J. R. Soc. Interface, 11:20140371, 2014.
21
[57] A. Jagota and C.-Y. Hui. Adhesion, friction, and compliance of bio-mimetic and bio-inspired structured
interfaces. Mater. Sci. Eng. R, 72:253–292, 2011.
[58] U. B. Jayadeep, M. S. Bobji, and C. S. Jog. Energy loss due to adhesion in longitudinal impact of elastic
cylinders. Eur. J. Mech. A/Solids, 45:20–31, 2014.
[59] U. B. Jayadeep, M. S. Bobji, and C. S. Jog. A body-force formulation for analyzing adhesive interactions
with special considerations for handling symmetry. Finite Elem. Anal. Des., 117–118:1–10, 2016.
[60] J.-W. Jiang and H. S. Park. A Gaussian treatment for the friction issue of Lennard-Jones potential in
layered materials: Application to friction between graphene, MoS2 , and black phosphorus. J. Appl. Phys.,
117:124304, 2015.
[61] K. L. Johnson. Continuum mechanics modeling of adhesion and friction. Langmuir, 12:4510–4513, 1996.
[62] K. L. Johnson. Adhesion and friction between a smooth elastic spherical asperity and a plane surface.
Proc. R. Soc. London A, 453:163–179, 1997.
[63] K. L. Johnson, K. Kendall, and A. D. Roberts. Surface energy and the contact of elastic solids. Proc. R.
Soc. London A, 324:301–313, 1971.
[64] H. Kawamura, T. Hatano, N. Kato, S. Biswas, and B. K. Chakrabarti. Statistical physics of fracture,
friction, and earthquakes. Rev. Mod. Phys., 84:839–884, 2012.
[65] K. Kendall. Thin-film peeling – the elastic term. J. Phys. D: Appl. Phys., 8:1449–1452, 1975.
[66] D. Labonte and W. Federle. Rate-dependence of ‘wet’ biological adhesives and the function of the pad
secretion in insects. Soft Matter, 11:8661–8673, 2015.
[67] D. Labonte and W. Federle. Biomechanics of shear-sensitive adhesion in climbing animals: peeling, pretension and sliding-induced changes in interface strength. J. R. Soc. Interface, 13:20160373, 2016.
[68] D. Labonte, J. A. Williams, and W. Federle. Surface contact and design of fibrillar ‘friction pads’ in stick
insects (Carausius morosus): mechanisms for large friction coefficients and negligible adhesion. J. R. Soc.
Interface, 11:20140034, 2014.
[69] T. A. Laursen. Computational Contact and Impact Mechanics. Springer, 2002.
[70] C. J. Lissenden and C. T. Herakovich. Numerical modelling of damage development and viscoplasticity in
metal matrix composites. Comput. Methods Appl. Mech. Eng., 126:289–303, 1995.
[71] B. Lorenz and B. N. J. Persson. On the origin of why static or breakloose friction is larger than kinetic
friction, and how to reduce it: the role of aging, elasticity and sequential interfacial slip. J. Phys.: Condens.
Matter, 24:225008, 2012.
[72] A. Majumdar and B. Bhushan. Fractal model of elastic-plastic contact between rough surfaces. J. Tribol.,
113:1–11, 1991.
[73] D. Maugis. Adhesion of spheres: The JKR-DMT transition using a Dugdale model. J. Colloid Interface
Sci., 150(1):243–269, 1992.
[74] D. Maugis. On the contact and adhesion of rough surfaces. J. Adhes. Sci. Technol., 10(2):161–175, 1996.
[75] J. C. Mergel. Advanced Computational Models for the Analysis of Adhesive Friction. Doctoral thesis,
RWTH Aachen University, Germany, 2017. Submitted.
[76] J. C. Mergel and R. A. Sauer. On the optimum shape of thin adhesive strips for various peeling directions.
J. Adhesion, 90(5-6):526–544, 2014.
[77] J. C. Mergel, J. Scheibert, and R. A. Sauer. A computational contact formulation for coupled adhesion
and friction. 2018. Under preparation.
[78] Y. Mo, K. T. Turner, and I. Szlufarska. Friction laws at the nanoscale. Nature, 457:1116–1119, 2009.
[79] M. H. Müser, W. B. Dapp, R. Bugnicourt, P. Sainsot, N. Lesaffre, T. A. Lubrecht, B. N. J. Persson, K. Harris, A. Bennett, K. Schulze, S. Rohde, P. Ifju, W. G. Sawyer, T. Angelini, H. A. Esfahani, M. Kadkhodaei,
S. Akbarzadeh, J.-J. Wu, G. Vorlaufer, A. Vernes, S. Solhjoo, A. I. Vakis, R. L. Jackson, Y. Xu, J. Streator,
A. Rostami, D. Dini, S. Medina, G. Carbone, F. Bottiglione, L. Afferrante, J. Monti, L. Pastewka, M. O.
Robbins, and J. A. Greenwood. Meeting the contact-mechanics challenge. Tribol. Lett., 65:118, 2017.
[80] M. Nosonovsky and B. Bhushan. Multiscale friction mechanisms and hierarchical surfaces in nano- and
bio-tribology. Mater. Sci. Eng. R, 58:162–193, 2007.
[81] A. Parness, D. Soto, N. Esparza, N. Gravish, M. Wilkinson, K. Autumn, and M. Cutkosky. A microfabricated wedge-shaped adhesive array displaying gecko-like dynamic adhesion, directionality and long lifetime.
J. R. Soc. Interface, 6:1223–1232, 2009.
[82] L. Pastewka and M. O. Robbins. Contact between rough surfaces and a criterion for macroscopic adhesion.
Proc. Natl. Acad. Sci. U.S.A., 111(9):3298–3303, 2014.
22
[83] Z. L. Peng, S. H. Chen, and A. K. Soh. Peeling behavior of a bio-inspired nano-film on a substrate. Int. J.
Solids Struct., 47:1952–1960, 2010.
[84] B. N. J. Persson. Sliding Friction. Springer, Berlin Heidelberg New York, 2nd edition, 2000.
[85] B. N. J. Persson. Theory of rubber friction and contact mechanics. J. Chem. Phys., 115(8):3840–3861,
2001.
[86] B. N. J. Persson, O. Albohr, U. Tartaglino, A. I. Volokitin, and E. Tosatti. On the nature of surface
roughness with application to contact mechanics, sealing, rubber friction and adhesion. J. Phys.: Condens.
Matter, 17:R1–R62, 2005.
[87] B. N. J. Persson, I. M. Sivebæk, V. N. Samoilov, K. Zhao, A. I. Volokitin, and Z. Zhang. On the origin of
Amonton’s friction law. J. Phys.: Condens. Matter, 20:395006, 2008.
[88] E. Popova and V. L. Popov. The research works of Coulomb and Amontons and generalized laws of friction.
Friction, 3(2):183–190, 2015.
[89] L. Prandtl. Ein Gedankenmodell zur kinetischen Theorie der festen Körper. Z. angew. Math. Mech.
(ZAMM), 8(2):85–106, 1928.
[90] A. Prevost, J. Scheibert, and G. Debrégeas. Probing the micromechanics of a multi-contact interface at
the onset of frictional sliding. Eur. Phys. J. E, 36:17, 2013.
[91] Y. I. Rabinovich, J. J. Adler, A. Ata, R. K. Singh, and B. M. Moudgil. Adhesion between nanoscale rough
surfaces. I. Role of asperity geometry. J. Colloid Interface Sci., 232:10–16, 2000.
[92] M. Raous, L. Cangémi, and M. Cocu. A consistent model coupling adhesion, friction, and unilateral contact.
Comput. Methods Appl. Mech. Eng., 177:383–399, 1999.
[93] M. Raous and Y. Monerie. Unilateral contact, friction and adhesion: 3D cracks in composite materials.
In J. A. C. Martins and M. D. P. Monteiro Marques, editors, Contact Mechanics, pages 333–346. Kluwer,
2002.
[94] M. Ruths, A. D. Berman, and J. N. Israelachvili. Surface forces and nanorheology of molecularly thin films.
In B. Bhushan, editor, Nanotribology and Nanomechanics, pages 389–481. Springer, Berlin Heidelberg, 2005.
[95] E. Sacco and F. Lebon. A damage-friction interface model derived from micromechanical approach. Int.
J. Solids Struct., 49:3666–3680, 2012.
[96] E. Sacco and J. Toti. Interface elements for the analysis of masonry structures. Int. J. Comput. Methods
Eng. Sci. Mech., 11:354–373, 2010.
[97] R. Sahli, G. Pallares, C. Ducottet, I. E. Ben Ali, S. Al Akhrass, M. Guibert, and J. Scheibert. Evolution
of real contact area under shear and the value of static friction of soft materials. Proc. Natl. Acad. Sci.
U.S.A., 115(3):471–476, 2018.
[98] R. A. Sauer. An Atomic Interaction based Continuum Model for Computational Multiscale Contact Mechanics. PhD thesis, University of California, Berkeley, USA, 2006.
[99] R. A. Sauer. The peeling behavior of thin films with finite bending stiffness and the implications on gecko
adhesion. J. Adhesion, 87(7–8):624–643, 2011.
[100] R. A. Sauer. A survey of computational models for adhesion. J. Adhesion, 92(2):81–120, 2016.
[101] R. A. Sauer and L. De Lorenzis. A computational contact formulation based on surface potentials. Comput.
Methods Appl. Mech. Eng., 253:369–395, 2013.
[102] R. A. Sauer and L. De Lorenzis. An unbiased computational contact formulation for 3D friction. Int. J.
Numer. Methods Eng., 101(4):251–280, 2015.
[103] R. A. Sauer and M. Holl. A detailed 3D finite element analysis of the peeling behaviour of a gecko spatula.
Comput. Methods Biomech. Biomed. Eng., 16(6):577–591, 2013.
[104] R. A. Sauer and S. Li. A contact mechanics model for quasi-continua. Int. J. Numer. Methods Eng.,
71:931–962, 2007.
[105] R. A. Sauer and J. C. Mergel. A geometrically exact finite beam element formulation for thin film adhesion
and debonding. Finite Elem. Anal. Des., 86:120–135, 2014.
[106] R. A. Sauer and P. Wriggers. Formulation and analysis of a three-dimensional finite element implementation
for adhesive contact at the nanoscale. Comput. Methods Appl. Mech. Eng., 198:3871–3883, 2009.
[107] A. R. Savkoor and G. A. D. Briggs. The effect of tangential force on the contact of elastic solids in adhesion.
Proc. R. Soc. London A, 356:103–114, 1977.
[108] A. Schallamach. The load dependence of rubber friction. Proc. Phys. Soc. B, 65(9):657–661, 1952.
[109] A. Schallamach. A theory of dynamic rubber friction. Wear, 6:375–382, 1963.
23
[110] J. Scheibert, A. Prevost, J. Frelat, P. Rey, and G. Debrégeas. Experimental evidence of non-Amontons
behaviour at a multicontact interface. Lett. J. Explor. Front. Phys. (EPL), 83:34003, 2008.
[111] P. Schmidt and U. Edlund. A finite element method for failure analysis of adhesively bonded structures.
Int. J. Adhesion Adhesives, 30:665–681, 2010.
[112] M. Schryve. Modèle d’adhésion cicatrisante et applications au contact verre/élastomère. PhD thesis,
Université de Provence Aix-Marseille I, France, November 2008.
[113] I. M. Sivebæk, V. N. Samoilov, and B. N. J. Persson. Frictional properties of confined polymers. Eur.
Phys. J. E, 27:37–46, 2008.
[114] L. Snozzi and J.-F. Molinari. A cohesive element model for mixed mode loading with frictional contact
capability. Int. J. Numer. Methods Eng., 93:510–526, 2013.
[115] D. Tabor. Friction — the present state of our understanding. J. Lubr. Technol., 103:169–179, 1981.
[116] K. Thøgersen, J. K. Trømborg, H. A. Sveinsson, A. Malthe-Sørenssen, and J. Scheibert. History-dependent
friction and slow slip from time-dependent microscopic junction laws studied in a statistical framework.
Phys. Rev. E, 89:052401, 2014.
[117] C. Thornton. Interparticle sliding in the presence of adhesion. J. Phys. D: Appl. Phys., 24:1942–1946,
1991.
[118] G. A. Tomlinson. CVI. A molecular theory of friction. Phil. Mag. Ser. 7, 7(46):905–939, 1929.
[119] J. Trømborg, J. Scheibert, D. S. Amundsen, K. Thøgersen, and A. Malthe-Sørenssen. Transition from
static to kinetic friction: Insights from a 2D model. Phys. Rev. Lett., 107:074301, 2011.
[120] J. K. Trømborg, H. A. Sveinsson, J. Scheibert, K. Thøgersen, D. S. Amundsen, and A. Malthe-Sørenssen.
Slow slip and the transition from fast to slow fronts in the rupture of frictional interfaces. Proc. Natl. Acad.
Sci. U.S.A., 111(24):8764–8769, 2014.
[121] J. K. Trømborg, H. A. Sveinsson, K. Thøgersen, J. Scheibert, and A. Malthe-Sørenssen. Speed of fast and
slow rupture fronts along frictional interfaces. Phys. Rev. E, 92:012408, 2015.
[122] V. Tvergaard. Effect of fibre debonding in a whisker-reinforced metal. Mater. Sci. Eng., 125:203–213,
1990.
[123] W. W. Tworzydlo, W. Cecot, J. T. Oden, and C. H. Yew. Computational micro- and macroscopic models
of contact and friction: formulation, approach and applications. Wear, 220:113–140, 1998.
[124] A. I. Vakis, V. A. Yastrebov, J. Scheibert, C. Minfray, L. Nicola, D. Dini, A. Almqvist, M. Paggi, S. Lee,
G. Limbert, J. F. Molinari, G. Anciaux, R. Aghababaei, S. Echeverri Restrepo, A. Papangelo, A. Cammarata, P. Nicolini, C. Putignano, G. Carbone, M. Ciavarella, S. Stupkiewicz, J. Lengiewicz, G. Costagliola,
F. Bosia, R. Guarino, N. M. Pugno, and M. H. Müser. Modeling and simulation in tribology across scales:
An overview, 2018. DOI: 10.1016/j.triboint.2018.02.005.
[125] M. Varenberg and S. Gorb. Shearing of fibrillar adhesive microstructure: friction and shear-related changes
in pull-off force. J. R. Soc. Interface, 4:721–725, 2007.
[126] M. Varenberg and S. N. Gorb. Hexagonal surface micropattern for dry and wet friction. Adv. Mater.,
21:483–486, 2009.
[127] J. F. Waters and P. R. Guduru. Mode-mixity-dependent adhesive contact of a sphere on a plane surface.
Proc. R. Soc. A, 466:1303–1325, 2010.
[128] P. Wriggers. Computational Contact Mechanics. Springer, Berlin Heidelberg, 2nd edition, 2006.
[129] P. Wriggers and J. Reinelt. Multi-scale approach for frictional contact of elastomers on rough rigid surfaces.
Comput. Methods Appl. Mech. Eng., 198:1996–2008, 2009.
[130] F. Wu-Bavouzet, J. Cayer-Barrioz, A. Le Bot, F. Brochard-Wyart, and A. Buguin. Effect of surface pattern
on the adhesive friction of elastomers. Phys. Rev. E, 82:031806, 2010.
[131] H. Yao and H. Gao. Mechanics of robust and releasable adhesion in biology: Bottom-up designed hierarchical structures of gecko. J. Mech. Phys. Solids, 54:1120–1146, 2006.
[132] S. Yashima, V. Romero, E. Wandersman, C. Frétigny, M. K. Chaudhury, A. Chateauminois, and A. M.
Prevost. Normal contact and friction of rubber with model randomly rough surfaces. Soft Matter, 11:871–
881, 2015.
[133] V. A. Yastrebov. Sliding without slipping under Coulomb friction: opening waves and inversion of frictional
force. Tribol. Lett., 62:1, 2016.
[134] H. Yoshizawa, Y.-L. Chen, and J. Israelachvili. Fundamental mechanisms of interfacial friction. 1. Relation
between adhesion and friction. J. Phys. Chem., 97:4128–4140, 1993.
24
[135] B. Yurdumakan, N. R. Raravikar, P. M. Ajayan, and A. Dhinojwala. Synthetic gecko foot-hairs from
multiwalled carbon nanotubes. Chem. Commun., 30:3799–3801, 2005.
[136] H. Zeng, N. Pesika, Y. Tian, B. Zhao, Y. Chen, M. Tirrell, K. L. Turner, and J. N. Israelachvili. Frictional
adhesion of patterned surfaces and implications for gecko and biomimetic systems. Langmuir, 25(13):7486–
7495, 2009.
[137] X. Zhang, X. Zhang, and S. Wen. Finite element modeling of the nano-scale adhesive contact and the
geometry-based pull-off force. Tribol. Lett., 41:65–72, 2011.
[138] B. Zhao, N. Pesika, K. Rosenberg, Y. Tian, H. Zeng, P. McGuiggan, K. Autumn, and J. Israelachvili.
Adhesion and friction force coupling of gecko setal arrays: Implications for structured adhesive surfaces.
Langmuir, 24:1517–1524, 2008.
25
| 5cs.CE
|
1
Variational Bayesian Inference of Line Spectral
Estimation with Multiple Measurement Vectors
Qi Zhang, Jiang Zhu, Peter Gerstoft, Mihai Alin Badiu and Zhiwei Xu
arXiv:1803.06497v1 [cs.IT] 17 Mar 2018
Abstract
In this paper, the line spectral estimation (LSE) problem with multiple measurement vectors (MMVs) is studied utilizing the
Bayesian methods. Motivated by the recently proposed variational line spectral estimation (VALSE) method, we extend it to deal
with the MMVs setting, which is especially important in array signal processing. The VALSE method can automatically estimate
the model order and nuisance parameters such as noise variance and weight variance. In addition, by approximating the probability
density function (PDF) of the frequencies with the mixture of von Mises PDFs, closed-form update equation and the uncertainty
degree of the estimates can be obtained. Interestingly, we find that the VALSE with MMVs can be viewed as applying the VALSE
with single measurement vector (SMV) to each snapshot, and combining the intermediate data appropriately. Furthermore, the
proposed prior distribution provides a good interpretation of tradeoff between grid and off-grid based methods. Finally, numerical
results demonstrate the effectiveness of the VALSE method, compared to the state-of-the-art methods in the MMVs setting.
Keywords: Variational Bayesian inference, von Mises PDF, multiple measurement vectors, line spectral estimation
I. I NTRODUCTION
Line spectral estimation (LSE) is one of the important problems in signal processing fields [1], which has many applications
such as channel estimation in wireless communications [2] and direction of arrival estimation in radar systems [3]. Traditional
methods for solving the LSE problem include periodogram, MUSIC and ESPRIT [1, 4, 5]. For periodogram method, it is
difficult to recover the closely separated frequencies [1]. While for subspace methods such as MUSIC and ESPRIT which
utilize the covariance matrix to estimate the frequencies, they perform well when the model order is known and the signal to
noise ratio (SNR) is high. Given that the model order is unknown in applications, some criterions such as Akaike information
criterion are adopted to estimate the model order [6].
In the past decades, sparse methods for LSE have been popular due to the development of sparse signal representation and
compressed sensing theory. By discretizing the continuous frequency into a finite set of grid points, the nonlinear problem
can be formulated as a linear problem. `1 optimization and sparse iterative covariance-based estimation (SPICE) are two
main sparse methods [7–10]. Compared to classical methods, the grid based methods perform better by utilizing the sparsity
in the spatial domain. Due to the grid mismatch, dictionary-based approaches suffer from spectral leakage. To mitigate the
drawbacks of static dictionary, some methods have been proposed to gradually refine the dynamic dictionary, such as iterative
grid refinement, joint sparse signal and parameter estimation [7, 11].
To avoid the model mismatch issues, off-grid compressed sensing methods have been proposed, such as atom norm
minimization and gridless SPICE (GLS) [12–15]. For the atomic norm-based methods, the variance of the noise is required
in advance. In contrast, GLS can be used to jointly estimate the frequencies and the noise variance. These off-grid based
methods involve solving a semidefinite programming (SDP) problem [16], whose computation complexity is prohibitively high
for large-scale problems.
A. Related Work
In [17], an exact discretization-free method called sparse and parametric approach (SPA) is proposed for uniform and sparse
linear arrays. This method is based on the well-established covariance fitting criterion, and the LSE is cast as a SDP problem.
In [18], a Newtonalized orthogonal matching pursuit (NOMP) method is proposed, where a Newton step and feedback are
utilized to refine the frequency estimation. In addition, the NOMP algorithm is also extended to deal with the MMVs setting
in [19]. Compared to the incremental step in updating the frequencies in NOMP approach, the iterative reweighted approach
(IRA) [20] estimates the frequencies in parallel. In [21], superfast LSE methods are proposed based on fast Toeplitz matrix
inversion algorithm.
The most closely related work is [22], which proposes off-grid based variational line spectral estimation (VALSE) algorithm.
While in [23], the mixture of the von Mises distributions is encoded as the prior, and maximum a posterior (MAP) methods are
proposed to estimate the frequencies with the number of frequencies being known. Compared to [22], we extend the VALSE
to deal with the MMVs setting, which is really important in array signal processing. In addition, we propose a flexible prior
and show that utilizing the prior appropriately improves the performance of the VALSE algorithm.
2
B. Main Contributions
The main contributions of this paper are summarized as follows: We rigorously develop the variational Bayesian inference
method for LSE in the MMVs setting, which is especially important in array signal processing. Meanwhile, the derived VALSE
reveals close relationship to the SMV VALSE algorithm, which is suitable for parallel processing. In addition, we propose
a flexible prior distribution which can be encoded naturally with the VALSE method. The prior information may be given
from past experience, and is particularly useful when the SNR is low or few samples are available. It is shown that the prior
distribution can be viewed as a tradeoff between grid and off-grid based methods. Furthermore, substantial experiments are
conducted to illustrate the competitive performance of the VALSE method, compared to other sparse based approaches.
Outline: We provide the problem model in Section II. Section III introduces some preliminary results. In Section IV, we
establish the probabilistic formulation of LSE and introduce the variational approach. Numerical results are presented in Section
V and conclusions are given in Section VI.
Notation: Let S ⊂ {1, · · · , N } be a subset of indices. For the matrix A ∈ CM ×N , let AS denote the submatrix by deleting
T
N ×L
T
the columns of A indexed by {1, · · · , N }\S. For the matrix H , [hT
and W , [w1T ; · · · ; wN
] ∈ CN ×L ,
1 ; · · · ; hN ] ∈ C
T
T
let hi and wi denote the ith row of H and W, respectively. Let HS and WS denote the submatrix by choosing the rows of
H and W indexed by S. For the matrix J ∈ CN ×N , let JS denote the submatrix by choosing both the rows and columns of
∗
T
H
J indexed by S. Let (·)S , (·)S and (·)S be the conjugate, transpose and Hermitian transpose operator of (·)S , respectively. Let
IL denote the identity matrix of dimension L. Let ||.||F denote the Frobenius norm. “ ∼ i” denotes the indices S excluding i
and <(·) returns the real part.
II. P ROBLEM S ETUP
For line spectral estimation problem with L snapshots, the measurements Y ∈ CM ×L consist of a superposition of K
complex sinusoids corrupted by the additive white Gaussian noise (AWGN), which is described by
Y=
K
X
e
e kT + U,
a(θek )w
(1)
k=1
e k and θek ∈ [−π, π) are the complex weight and frequency
where M is the number of measurements for each observation. w
h
iT
e represents the AWGN and a(θek ) = 1, ejθek , · · · , ej(M −1)θek . Since the number of
of the kth component, respectively. U
complex sinusoids K is generally unknown, we adopt the following model
Y=
N
X
a(θi )wiT + U = AW + U,
(2)
i=1
where A = [a(θ1 ), · · · , a(θN )] ∈ CM ×N , a(θi ) denotes the ith column of A, wiT denote the ith row of W ∈ CN ×L , and
N is the assumed number of complex sinusoids and satisfies N > K. We introduce binary hidden variables s = [s1 , ..., sN ]T ,
whose probability mass function is
p(si ) = λsi (1 − λ)(1−si ) ,
si ∈ {0, 1}.
(3)
We also assume (si , wi ) follows a Bernoulli-Gaussian distribution, that is
p(wi |si ; τ ) = (1 − si )δ(wi ) + si fCN (wi ; 0, τ IL ),
(4)
where
kwi k22
1
.
exp −
fCN (wi ; 0, τ IL ) =
(πτ )L
2τ
From (3) and (4), it can be seen that the parameter λ controls the active probability of the ith component and τ is a variance
Q
parameter. The variable θ = [θ1 , ..., θN ]T has the prior probability density function (PDF) p(θ) = N
. One can use
i=1 p(θi )
1
the uninformative prior distribution p(θi ) = 2π as shown in [22]. Another way is to use the following more flexible prior
distribution
p(θi ) =
N
X
αi,n fVM (θi ; µn , κp ),
i = 1, · · · , N,
(5)
n=1
PN
where fVM (θi ; µn , κp ) is the von Mises distribution defined later in (7), αi,n is unknown and αi,n ∈ {0, 1}, n=1 αi,n = 1,
PN
i=1 αi,n = 1, µn and κp are known mean direction and concentration parameters. For example, αi,n may be taken from the
3
(i, n)th element of IN . We emphasize that when inferring the frequencies, αi,n is unknown and needs to be estimated. As for
generating the frequencies in model (1), we use the prior distribution
p(θek ) =
N
X
k = 1, · · · , K,
α
ek,n fVM (θek ; µ
en , κ
e),
(6)
n=1
and α
ek,n can be taken from the (k, n)th element of a given permutation matrix.
We assume that the components of the noise U are independent and identically distributed (i.i.d.) complex Gaussian with
mean zero and variance ν, which gives the likelihood
1
||Y − AW||2F
p(Y|θ, W; ν) =
exp −
.
(πν)M L
ν
We use β = {ν, λ, τ } to denote the model parameters.
III. P RELIMINARY
In this section, the standard probability distribution presented in [22] are summarized in order to make this paper selfcontained. Firstly, the von Mises distribution is introduced. Then we present N -fold wrapped von Mises distribution. By using
characteristic function matching principle, the N -fold wrapped von Mises distribution can be approximated as a mixture of
von Mises distributions.
A. The von Mises distribution
Von Mises distribution is a useful distribution on the circle, and its form is similar to the Gaussian distribution on the line.
The PDF of the von Mises distribution is [24, p. 36]
fVM (θ; µ, κ) =
1
eκcos(θ−µ) ,
2πI0 (κ)
(7)
where µ is the mean direction, i.e., E(θ) = µ for PDF fVM (θ; µ, κ), κ is the measure of concentration, Ip (κ) is the modified
Bessel function of the first kind and the order p given by [24, p. 348]
Z 2π
1
Ip (κ) =
cos (pθ)eκ cos θ dθ.
(8)
2π 0
In complex notation, the PDF (7) is
fVM (θ; η) =
∗ jθ
1
e<{η e } ,
2πI0 (|η|)
(9)
where η = κejµ , and
fVM (θ; η1 )fVM (θ; η2 ) ∝ fVM (θ; η),
η = η1 + η2 .
(10)
Thus the von Mises PDFs are closed under multiplication. In Fig. 1, the PDF of von Mises distribution with means being
zero are presented. The concentrate parameters of the von Mises distributions are κ = 10, 102 , 103 and 104 , respectively. As
κ increases, the von Mises PDF becomes more concentrated around the means. For κ = 10, 102 , 103 and 104 , the standard
deviations of the von Mises distributions are 0.227, 0.071, 0.022 and 0.005, respectively.
Another way to describe the PDF is through the characteristic function. For distributions on the line, the characteristic
function of a random variable X is φX (t) = E[ejtX ]. Similarly, the characteristic function of distributions on the circle is
φp = E[ejpθ ], where p takes integer value as the distribution is invariant under θ 7→ θ + 2π. For von Mises distribution, the
characteristic function is [24, p. 40]
φp,VM = ejpµ
Ip (κ)
,
I0 (κ)
p ∈ Z.
(11)
4
40
35
von Mises PDF
30
25
20
15
10
5
0
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
Fig. 1. The PDF of von Mises distribution of mean 0 with different κ. We only plot the interval where θ ∈ [−0.6, 0.6].
B. N -fold wrapped von Mises distribution
In many scenarios, such as image analysis, biology and physics, various distributions are approximated as N -fold rotational
invariant whose PDFs are periodic functions with period 2π/N . The N -fold wrapped von Mises distribution is one of them,
and it has the PDF [24, pp. 52-54]
fWVM (N θ; µw , κw ) = fWVM (N θ; ηw )
∗ jN θ
1
e<{ηw e } ,
=
2πI0 (|ηw |)
(12)
where ηw = κw ejN µw . Since N -fold wrapped von Mises distributions are not closed under multiplication, we approximate
N -fold wrapped von Mises distribution as a mixture of von Mises distributions. The approximation is [22]
fWVM (N θ; ηw ) ≈ fMVM (θ; ηm ) ,
N
−1
X
n=0
1
fVM (θ; ηm,n ),
N
(13)
where ηm = [ηm,0 , · · · , ηm,N −1 ]T , ηm,k = κm ejµm,k and an additional restriction µm,n = µm + 2πn
N . µm and κm are calculated
as
µm = µw ,
I1 (κw )
IN (κm )
=
.
I0 (κm )
I0 (κw )
(14a)
(14b)
IV. A LGORITHM
In the Bayesian inference framework, one goal is to compute the MAP estimates or the minimum mean square error (MMSE)
estimates of θ, W and S. Given the observations Y and nuisance parameters β, the posterior distribution p(θ, W, s|Y; β) is
calculated as
!
N
Y
p(Y,θ,W,s;β)
p(θ, W, s|Y; β) =
∝
p(θi )p(si )p(wi |si ) p(Y|θ, W),
p(Y;β)
i=1
where we use p(wi |si ) and p(Y|θ, W) instead of p(wi |si ; τ ) and p(Y|θ, W; ν) for simplicity. Directly maximizing the above
posterior PDF with respect to θ, W and s to find the MAP estimates or calculating the MMSE estimates is intractable.
As a result, we use variational inference method to approximate p(θ, W, s|Y) with a given structured PDF q(θ, W, s|Y).
Variational Bayesian often uses the Kullback-Leibler (KL) divergence of p(θ, W, s|Y) from q(θ, W, s|Y) to describe their
dissimilarity, which is defined as [25, p. 732]
Z
q(θ, W, s|Y)
DKL (q(θ, W, s|Y)||p(θ, W, s|Y)) = q(θ, W, s|Y) log
dθdWds.
(15)
p(θ, W, s|Y)
5
In general, we choose the posterior PDF q(θ, W, s|Y) which minimizes the KL divergence. The log model evidence ln p(Y; β)
for any assumed PDF q(θ, W, s|Y) is [25, pp. 732-733]
ln p(Y; β) =DKL (q(θ, W, s|Y)||p(θ, W, s|Y)) + L(q(θ, W, s|Y)),
(16)
h
i
L(q(θ, W, s|Y)) = Eq(θ,W,s|Y) ln p(Y,θ,W,s;β)
.
q(θ,W,s|Y)
(17)
where
For a given data Y, ln p(Y; β) is a constant, thus minimizing the KL divergence is equivalent to maximizing L(q(θ, W, s|Y))
in (16). Therefore we maximize L(q(θ, W, s|Y)) in the sequel.
For the factored PDF q(θ, W, s|Y), we make the following assumptions: 1) The posterior of the frequencies q(θi |Y) are
independent of each other. 2) The posterior of the binary hidden variables q(s|Y) has all its mass at b
s, i.e., q(s|Y) = δ(s − b
s).
3) Given Y and s, the frequencies and weights are independent. As a result,
q(θ, W, s|Y) =
N
Y
q(θi |Y)q(W|Y, s)q(s|Y).
(18)
i=1
Due to the factorization property of (18), the frequencies θ can be estimated from the marginal distribution q(θ, W, s|Y) as
[24, pp. 26]
θbi = arg(Eq(θi |Y) [ejθi ]),
bi = Eq(θi |Y) (a(θi )), i ∈ {1, ..., N },
a
(19a)
(19b)
where arg(·) returns the angle. In IV-A, we will
q(θi |Y) as a von Mises distribution. For von Mises distribution
approximate
jθ
jµ I1 (κ)
fVM (θ; µ, κ) (7), arg(EfVM (θ;µ,κ) [e ]) = arg e I0 (κ) = µ = EfVM (θ;µ,κ) [θ]. Therefore, θbi is also the mean direction of θ
for von Mises distribution.
(κ)
bi (19b) can be calculated easily from (11). From (11), the magnitudes of Eq(θi |Y) [ejmθi ] are IIm0 (κ)
Besides, a
< 1 for
b
m ∈ {1, · · · , M − 1} because of Im (κ) < I0 (κ) for m > 0. An alternative approach is to estimate a(θi ) as a(θi ). For large κ,
the bias is small and the performance of utilising âi (19b) is similar to that of utilizing a(θbi ). Thus we use (19b) to estimate
a(θi ). For the given posterior PDF q(W|Y, b
s), the mean and covariance of the weights are estimated as
b i = Eq(W|Y) [wi ],
w
b
b iw
b jH , i, j ∈ {1, ..., N }.
Ci,j = Eq(W|Y) [wi wjH ] − w
(20a)
(20b)
Given that q(s|Y) = δ(s − b
s), the posterior PDF of W is
q(W|Y) = q(W|Y, s)δ(s − b
s) = q(W|Y, b
s).
(21)
Let S be the set of indices of the non-zero components of s, i.e.,
S = {i|1 ≤ i ≤ N, si = 1}.
b i.e.,
Analogously, we define Sb based on b
s. The model order is the cardinality of S,
b = |S|.
b
K
According to (2), the noise-free signal is reconstructed as
b =
Y
X
bi w
b iT
a
i∈Sb
Maximizing L(q(θ, W, s|Y)) with respect to all the factors is also intractable. Similar to the Gauss-Seidel method [26], we
optimize L over each factor q(θi |Y), i = 1, · · · , N and q(W, s|Y) separately with the others being fixed. In the following,
we detail the procedures.
A. Inferring the frequencies
For each i = 1, ..., N , we maximize L with respect to the factor q(θi |Y). For i ∈
/ S, we have q(θi |Y) = p(θi ). For i ∈ S,
the optimal factor q(θi |Y) can be calculated as [25, pp. 736-737]
ln q(θi |Y) =E∼θi [ln p(Y, θi , θ∼i , W, s; β)] + const,
(22)
6
where E∼θi [·] denotes the expectation with respect to q(W|Y, b
s)
of the PDF. Substituting (19) and (20) in (22), we obtain
Q
j6=i
q(θj |Y) except θi . The constant ensures normalization
ln q(θi |Y) =E∼θi [ln p(Y, θi , θ∼i , W, s; β)] + const = E∼θi [ln(p(θ)p(s)p(W|s)p(Y|θ, W))] + const
N
N
X
X
=E∼θi [
ln p(θj ) +
ln p(sj ) + ln p(W|s) + ln p(Y|θ, W)] + const
j=1
j=1
= ln p(θi ) + E∼θi [ν
−1
2
||Y − ASbWSb||F ] + const
H
= ln p(θi ) + E∼θi [2ν −1 <(tr(YH ASbWSb)) − ν −1 tr((ASbWSb) (ASbWSb))] + const
H
b iT YH a(θi )) − E∼θi [2ν −1 <(wiT WS\{i}
= ln p(θi ) + 2ν −1 <(w
AH
)]a(θi ) + const,
b
b
S\{i}
Therefore
∗
b iT YH p − E∼θi [wiT WS\{i}
q(θi |Y) ∝p(θi ) exp{<((2ν −1 (w
AH
]))a(θi ))}.
b
b
S\{i}
After further manipulation utilizing (19), (20) and (12), q(θi |y) is simplified as
q(θi |Y) ∝ p(θi ) exp{<(ηiH a(θi ))} = p(θi )
M
−1
Y
∗
exp{<(ηi,m
ejmθi } ∝ p(θi )
m=1
M
−1
Y
fWVM (mθi ; ηi,m ),
(23)
m=1
where p(θi ) is defined in (5), the complex vector ηi is given by
b b E∼θ [WT
b i∗ − A
ηi = 2ν −1 (Yw
wi∗ ]))
i
b
S\{i}
S\{i}
X
b i∗ −
bl E∼θi [wlT wi∗ ])
= 2ν −1 (Yw
a
b
l∈S\{i}
= 2ν
−1
b i∗
(Yw
−
X
bl E∼θi [wiH wl ])
a
b
l∈S\{i}
b i∗ −
= 2ν −1 (Yw
X
b l,i ) + w
bl (tr(C
b iH w
b l )),
a
(24)
b
l∈S\{i}
ηi,m denotes the mth element of ηi with the polar form ηi,m = κi,m ejµi,m .
The posterior distribution q(θi |Y) can be approximated as a von Mises distribution according to [22], which consists of two
steps. Firstly, we pick the most dominant von Mises distribution whose mean direction is θi and estimate αi,n (5) as α̂i,n .
Then we add a Newton step to refine the estimation, i.e.,
f 0 (θi )
,
θbi = µ = θi − 00
f (θi )
1
κ
bi = A−1 exp( 00
),
2f (θi )
(25a)
(25b)
(·)
the detail of the inverse of A(·) , II01 (·)
is given in [24]. f (θ) denotes the exponent of (23) with αi,n replaced with α̂i,n . As
a result, q(θi |Y) can be approximated as
q(θi |Y) ≈ fVM (θi ; ηbi ),
(26)
bi are the characteristic functions (11) of fVM (θi ; ηbi ) for p ∈ {0, 1, · · · , M − 1}, which
where ηbi = κ
bi ejθi . The elements of a
are
IM −1 (b
κi )
I1 (b
κi )
bi = diag 1,
,··· ,
a(θbi ).
(27)
a
I0 (b
κi )
I0 (b
κi )
b
1
Remark 1 Different form [22], we assume that the frequencies follow the distribution (5). When κp = 0, p(θi ) = 2π
, which
PN
is the uninformative prior. Interestingly, when κp tends to infinity, p(θi ) = n=1 αi,n δ(θi − µn ), where δ(·) denotes the delta
function. Consequently, the signal model (2) is a sum over deterministic frequencies µi , i.e.,
Y=
N
X
a(µi )wiT + U,
(28)
i=1
and the VALSE algorithm can be viewed as a grid based method. Thus, by varying κp , the prior of the VALSE algorithm
provides a trade-off between grid and off-grid based method.
7
B. Inferring the weights and support
Next we keep q(θi |Y), i = 1, ..., N fixed and maximize L w.r.t. q(W, s|Y). Plugging the postulated PDF (18) in (17) we
obtain
L(q(W|Y, s; b
s)) = const − Eq(W|Y,s) {ln q(W|Y, b
s) − Eq(θ|Y) [ln p(Y, θ, W, b
s; β)]}.
(29)
Let us introduce a PDF
t(W; b
s) =
1
Z(b
s)
exp{Eq(θ|Y) [ln p(Y, θ, W, b
s; β)]},
where Z(b
s) is the normalizing constant. (29) can be simplified as
L(q(W|Y, S; b
s)) = −DKL (q(W|Y, s)||t) + ln Z(b
s) + const.
(30)
Given that the KL divergence DKL (q(W|Y, s)||t) is nonnegative and equals to 0 if and only if q(W|Y, s) = t, the maximum
of L (30) over q(W, s|Y) is attained when
b
s = argmax ln Z(s)
and
q(W|Y, b
s) = t (W; b
s) .
s
Before further deriving q(W|Y, ŝ), we define two matrices J and H as
(
M,
i=j
, i, j ∈ {1, 2, · · · , N },
Jij =
bH
b
a
a
,
i
6= j
i j
H
b1 Y; ...; a
bH
H= a
NY ,
(31)
(32)
where Jij denotes the element of the ith row and the jth column of J. From (21) and (31) we obtain
q(W|Y) = t(W; b
s) =
=
+
1
b
Z(b
s) exp{Eq(θ|Y) [ln p(Y, θ, W, s; β)]}
1
s||0 ln λ − ||b
s||0 ln(1 − λ) − ||b
s||0 Lln(πτ )
Z(b
s) exp{const + ||b
−1
H
−1
H
−1
H
ν tr(WSb HSb) + ν tr(HSb WSb) − ν tr(WSb JSbWSb)},
− τ −1 tr(WSbH WSb)
(33)
c b, C
b b) Q b δ(wi ),
which is quadratic in W. Given that wi is Gaussian distributed for si = 1, we obtain q(W|Y) = fCN (WSb; W
S
S
i6∈S
c b, C
b b) ,
fCN (WSb; W
S
S
1
b
(π)||bs||0 L det(C
Sb)
c T ))H C
b −1 (vec(WT ) − vec(W
c T ))}.
exp{−((vec(WSTb ) − vec(W
Sb
Sb
Sb
Sb
c b in (33) should be equal to that of (IV-B) and the formula [27, p. 28]
According to the fact that the coefficients of W
S
tr(ABCD) = vecT (DT )(CT ⊗ A)vec(B),
we obtain
ν
I b )WSb)
τ |S|
ν
ν
H
−1
T
H
T b −1
T
=ν −1 vecT (WSTb )((JT
+ I|S|
vecH (WSTb )((JSb + I|S|
b ) ⊗ IL )vec(WSb ) = ν
b ) ⊗ IL )vec(WSb ) = vec (WSb )CSb vec(WSb )
Sb
τ
τ
and
ν −1 tr(WSHb (JSb +
ν −1 tr(HH
WSb) = ν −1 vecT (WSTb )((H∗Sb ⊗ IL )vec(IL )
Sb
= ν −1 vecT (IL )(HH
⊗ IL )vec(WSTb )
Sb
c T )C
b −1 vec(WT ),
= vecH (W
Sb
Sb
Sb
where ⊗ denotes the Kronecker product. The mean and covariance matrix of WSb are computed as
c T ) = ν −1 C
b b(Hb ⊗ IL )vec(IL ),
vec(W
S
S
Sb
b b = ν(J b + ν I b )−1 ⊗ IL .
C
S
S
τ |S|
(34a)
(34b)
b b = ν(J b + ν I b )−1 and utilizing the formulas [27, pp. 25-26]
Letting C
S,0
S
τ |S|
(A ⊗ B)(C ⊗ D) = (AC) ⊗ (BD)
and
vec(ABC) = (CT ⊗ A)vec(B),
(34) can be simplified as
b b Hb
c b = ν −1 C
W
S
S,0 S
bb=C
b b ⊗ IL .
and C
S
S,0
(35)
8
Then, we need to find the sequence b
s which maximizes
ln Z(s) = Eq(θ|Y) [ln p(Y,θ, W, s; β)] − ln t(W; s)
b H Y) + ν −1 tr(YH A
b S WS )
= const + ||s||0 ln λ − ||s||0 ln(1 − λ) − ||s||0 Lln(πτ ) − τ −1 tr(WSH WS ) + ν −1 tr(WSH A
S
−1
−1
H
−1
T
T
H
T
b S ) + (vec(W ) − vec(W
c )) C
b (vec(W ) − vec(W
c T ))
− ν tr(W JS WS )} + ||s||0 L ln(π) + ln det(C
S
S
S
S
S
S
ν
ν
ν
λ
−1
= ln det((JS + I|S| )−1 ) + ν −1 tr(HH
+ ||s||0 L ln + const.
(36)
HS ) + ||s||0 ln
S (JS + I|S| )
τ
τ
1−λ
τ
Finding the maximum of (36) is NP-hard. As a consequence, a greedy iterative search strategy similar to [22] is proposed to
find a local maximum of ln Z(s) in Appendix A.
C. Estimating the model parameters
After updating the frequencies and weights, we now estimate the model parameters β = {ν, λ, τ } via maximizing the
lower bound L(q(θ, W, s|Y); β) for fixed q(θ, W, s|Y), i.e.,
L(q(θ, W, s|Y); β) = Eq(θ,W,S|Y) [ln p(Y,θ,W,s;β)
q(θ,W,s|Y) ]
= Eq(θ,W,S|Y) [
N
X
ln p(si ) + ln p(W|s) + ln p(Y|θ, W)] + const
i=1
= ||b
s||0 ln λ − ||b
s||0 ln(1 − λ) + ||b
s||0 L ln
+
1
1
1
1
− Eq(W|Y) [ tr(WSbWSHb )] + M L ln
− tr(YH Y)
πτ
τ
πν
ν
2
c H H b)) − 1 Eq(W|Y) [tr(WH J bW b)] + const.
<(tr(W
S
Sb S
Sb S
ν
ν
(37)
Because
X
Eq(W|Y) [tr(WSbWSHb )] = tr(Eq(W|Y) [
wiH wi ])
i∈Sb
X
X
X
b ii )
= tr(
Eq(W|Y) [wiH wi ]) = tr(
wiH wi ) + tr(
C
i∈Sb
i∈Sb
i∈Sb
X
X
c b) + Ltr(C
b b ),
b ii,0 = tr(W
c HW
= tr(
wiH wi ) + L
C
S
S,0
Sb
i∈Sb
i∈Sb
and
Eq(W|Y) [tr(WSHb JSbWSb)] = tr(JSbEq(W|Y) [WSbWSHb ])
b ij ) = w
b ij,0 ,
b iT w
b j∗ + tr(C
b iT w
b j∗ + LC
Eq(W|Y) [WWH ]ij = Eq(W|Y) [wiT wj∗ ] = w
H
H
c bW
c + LC
b b )),
Eq(W|Y) [tr(W J bW b)] = tr(J b(W
Sb S
S
S
S
Sb
S,0
(38)
we obtain
1
c Hb H b)) − tr(YH Y)
[2<(tr(W
S
S
ν
c Hb J bW
c b) − Ltr(J bC
b b )] − M Llnν − 1 [tr(W
c bW
c Hb )
− tr(W
S
S S,0
S
S S
S
τ
λ
b
+ Ltr(CS,0
s||0 (ln
− Llnτ ) + N ln(1 − λ) + const.
b )] + ||b
1−λ
L(q(θ, W, s|Y); β) =
Setting
∂L
∂ν
= 0,
∂L
∂λ
= 0,
∂L
∂τ
= 0, we have
b bW
c b||2 /(M L) + tr(J bC
b b )/M +
νb = ||Y − A
S
S F
S S,0
L
XX
|w
bil |2 (1/L − ||b
ai ||22 /(LM )),
b l=1
i∈S
s||0
b = ||b
λ
,
N
c bW
c H ) + Ltr(C
bb )
tr(W
S
S,0
Sb
τb =
.
L||b
s||0
(39)
9
D. The MMV VALSE algorithm
Now the details of updating the assumed posterior q(θ, W, s|Y) have been given and summarized in Algorithm 1. For the
b τb
proposed algorithm, the initialization is important for the performance of the algorithm.
The schemes that we initialize νb, λ,
||Y H a(θ1 )||22
0
and q(θi |Y), i ∈ {1, · · · , N } are below. Firstly, we assign q(θ1 |Y) ∝ exp
. We define M = {m − n | m, n ∈
νM
0
{0, 1, · · · , M − 1}, m >Pn} with cardinality M 0 and a0 : [−π, π) → CM , θ → a0 (θ) , (ejθm | m ∈ M0 )T . For each
1
H
t = 1, · · · , M 0 , γt = M
(k,l)∈Tt Yk,: Yl,: with Tt = {(k, l) | 1 ≤ k, l ≤ M, mk − ml = t}, where Yl,: denote the lth row of
Y, we obtain
2 H 0
q(θ1 |Y) ∝ exp <
γ a (θ1 )
.
(40)
ν
For the latter in (40), we use γ to build a Toeplitz matrix as an estimation of E[YYH ]. We estimate Lb
ν as the average of the
b = 0.5 and let τb = (YH Y/M − Lb
b ). Then, to initial q(θi |Y),
lower quarter of the eigenvalues of that matrix. We set
λ
ν
)/(
λN
Pi−1
2
b k }i−1
bk w
b kT . Initialize q(θi |Y) ∝ exp{||ZH
we compute {w
i−1 a(θ)||2 /(νM )}, which can
k=1 a
k=1 and the residual Zi−1 = Y −
bi can be calculated straightforwardly (Sec.IV-A).
be approximately represented as a product of MVM similar to (40), and a
Algorithm 1 Outline of VALSE algorithm with MMVs setting.
Input: Signal matrix Y
c b and reconstructed signal
b frequencies estimate θb b, complex weights estimate W
Output: The model order estimate K,
S
S
b
X
b τb and qθ |Y , i ∈ {1, · · · , N }; compute a
bi
1: Initialize ν
b, λ,
i
2: repeat
c b and C
b b (Sec.IV-B)
3:
Update b
s, W
S
S
b
4:
Update νb, λ, τb (39)
bi for all i ∈ Sb (Sec.IV-A)
5:
Update ηi and a
6: until stopping criterion is satisfied
c b and X
b
7: return ||b
s||0 , θbSb, W
S
Now we compare MMV Algorithm 1 with the SMV VALSE algorithm in [22]. The MMVs can be decoupled as L SMVs.
For each SMV, we perform the VALSE algorithm and obtain ηi,l , for the lth snapshot. Then ηi (24) is obtained via summing
L
P
ηi,l for all the snapshots, i.e., ηi =
ηi,l , and now each ηi,l is updated as ηi . We use ηi to obtain estimates θbi (25a) and
l=1
T
bi (27). In addition, we update the weight and their covariance (35) by applying the SMV VALSE. Let w
b S,l
a
b be the estimated
T
T
c
b ;··· ;w
b
weight of the lth snapshot, the whole weight matrix W b (35) can be constructed as [w
]. It is worth noting that
S
b
S,1
b
S,L
the second equation of (35) reveals that for different snapshots, the weight vectors are uncorrelated. Besides, the covariance
of the weight for each snapshot is the same, which means that the common covariance of the weight can be fed to the SMV
VALSE. For updating S, (47) can be viewed as a sum of the results (equation (40) in [22]) from the VALSE in SMVs, minus
λ
an additional constant term (L − 1) ln 1−λ
, while for the deactive case, (52) can be viewed as a sum of the results (equation
λ
(44) in [22]) from the VALSE in SMVs, plus an additional constant term (L − 1) ln 1−λ
. The additional constant terms can
not be neglected because we need to determine the sign of (47) and (52). For the lth snapshot, model parameters estimates νb
L
L
P
P
and τb are updated as the average of their respective estimates, i.e., νb =
νbl /L and τb =
τbl /L, where νbl and τbl denote
l=1
l=1
b can be naturally estimated.
the estimate of the lth SMV VALSE, and λ
V. N UMERICAL S IMULATION
In this section, substantial numerical simulations are performed to substantiate the VALSE algorithm. We will evaluate
the VALSE algorithm in terms of the probability of correct model order estimation and the error of estimated frequency. To
generate the real signal, we will replace N = 20 in (2) with true number of complex sinusoids K = 3. The frequencies are
(N −1)π T
)π (3−N )π
e = [e
drawn from the prior distribution (6) unless stated otherwise, where µ
µ1 , · · · , µ
eN ]T = [ (1−N
N +1 , N +1 , · · · , N +1 ] ,
e without replacement. For each element, i.e., µ
κ
e = 100. We first uniformly pick K elements from µ
ei , we generate θi from the
von Mises distribution fVM (θi ; µ
ei , κ
e). We also evaluate the performance of the VALSE algorithm utlizing noninformative prior,
i.e., p(θi ) = 1/(2π), i = 1, · · · , N . Each element of W is a complex Gaussian distribution with mean 0 and variance 1. We
define the SNR as SNR = 20log(||A(θ)WT ||F /||U||F ), and the normalized MSE (NMSE) of θb as 20log(||θb − θ||2 /||θ||2 ).
b > K, we sort κ
b = [b
Given that the model order is over estimated such that K
κ1 , κ
b2 , · · · , κ
bN ] in a descending order and
b to calculate the NMSE, where κ
choose the frequencies corresponding to the top K elements of κ
bi is defined in (25b). We
b (t−1) − Y
b (t) ||2 /||Y
b (t−1) ||2 < 10−5 or the
b < K. The Algorithm 2 stops when ||Y
fill frequencies with zeros as θb when K
10
45
40
35
PDF
30
25
20
15
10
5
0
-3
-2
-1
0
1
2
3
Fig. 2. The posterior PDF (26) of the frequencies. To make comparison, we plot the true θe with different values corresponding to the posterior PDF with
noninfo. prior and prior.
number of iterations exceeds 200. All the results are averaged over 300 Monte Carlo (MC) trials unless stated otherwise. For
performance comparison, we choose the classical MUSIC method which is provided by MATLAB rootmusic function, the
SPA method [17] and the NOMP method [18, 19]. For the MUSIC method, we empirically find the optimal sliding window
W to ensure that the MUSIC algorithm performs best. In our setting we set W = d(M + K)/2e, where d.e denotes the ceil
operator. For the SPA approach, we obtain the denoised covariance matrix and use the MUSIC method to avoid frequency
splitting phenomenon. We also evaluate the Cramér-Rao bound (CRB) as derived in [19].
Before moving into investigating the NMSE performance of the frequency, we present a simple example and plot the posterior
PDF of θ as shown in Fig. 2. The parameters are set as follows: M = 20, N = 5, K = 3, L = 10, κ
e = 103 , SNR = −2
5
dB. It can be seen that the posterior PDF (26) is very peaked (κ̂i ≈ 10 ). Besides, the VALSE utilizing the prior distribution
(NMSE: −44 dB) performs better than that with noninformative prior (NMSE: −37 dB).
A. Estimation by varying SNR
The performance in terms of model order estimation accuracy and frequency estimation error by varying SNR is presented
in Fig. 3. We set the number of measurements M = 20 and snapshots L = 10. The wrap-around distance between any two
frequencies is larger than 4ω = 2π
N . It can be seen that utilizing the prior information improves the performance of the VALSE
algorithm when the SNR is low. In Fig. 3(a), the VALSE algorithm achieves the highest probability of correct model order
estimation, compared with NOMP and SPA algorithms.
For the frequency estimation error, it is seen that the SPA (assuming K is known ) approaches the CRB firstly. Then the
VALSE and NOMP algorithms begin to approach the CRB. The SPA with unknown K is the last one that approaches CRB.
The running time of the algorithms is presented in Table I. It can be seen that the NOMP is the fastest method, and the SPA
is the slowest. In addition, the VALSE algorithm which utilizes prior information runs faster than that with noninformative
prior.
TABLE I
T HE RUNNING TIME OF EACH ALGORITHM AVERAGED OVER 300 INDEPENDENT MC TRIALS (SNR = 10 D B).
Algorithm
Time [s]
SPA
0.915
NOMP
0.008
VALSE, noninfo. prior
0.460
VALSE, prior
0.251
B. Estimation by varying L
In this subsection, we examine the estimation performance by varying the number of snapshots L. The number of measurements M is 30 and the SNR = 0 dB. The results are presented in Fig. 4. The wrap-around distance between any two
frequencies is larger than 4ω = 2π
N . From Fig. 4(a) and 4(b), we can see that when L ≤ 4, the NOMP algorithm achieves
the highest probability of correct model order estimation, while its NMSE is higher than that of VALSE methods. As for SPA
with unknown K, the correct model order estimation probability is very low and the resulting NMSE is large, so we do not
plot it in Fig. 4(b). Compared with the VALSE algorithm with noninformative prior, the VALSE algorithm utilizing the true
prior distribution achieves lower probability of correct model order estimation and NMSE. The reason is that utilizing the true
11
1
10
0.9
SPA, K unknown
NOMP
VALSE, noninfo. prior
VALSE, prior
0.7
NMSE of freq. est. (dB)
0.8
SPA, K known
SPA, K unkonwn
NOMP
VALSE, noninfo. prior
VALSE, prior
CRB
0
0.6
0.5
0.4
0.3
-10
-20
-30
-40
-50
0.2
-60
0.1
0
-10
-5
0
5
10
15
-70
-10
20
-5
0
SNR(dB)
5
10
15
20
SNR(dB)
(a)
(b)
Fig. 3. Performance of algorithms by varying SNR. The number of measurements is M = 30 and the number of snapshots is L = 10. Simulation results
are averaged over 300 MC simulations.
1
-5
0.9
-10
0.8
SPA, K known
NOMP
VALSE, noninfo. prior
VALSE, prior
CRB
NMSE of freq. est. (dB)
-15
0.7
SPA, K unknown
NOMP
VALSE, noninfo. prior
VALSE, prior
0.6
0.5
0.4
0.3
-20
-25
-30
-35
-40
0.2
-45
0.1
0
-50
2
4
6
8
10
12
14
16
18
20
2
4
6
8
10
L
12
14
16
18
20
L
(a)
(b)
Fig. 4. Performance of algorithms by varying snapshots L. We set SNR = 0dB and the number of measurements M = 30. Simulation results are averaged
over 300 MC trials.
prior distribution makes the VALSE algorithm overestimate the model order, as shown in Table II. Meanwhile, the VALSE
algorithm utilizing the true prior distribution estimates the true frequencies more accurately than that of VALSE algorithm with
noninformative prior.
TABLE II
b > K OF VALSE ALGORITHM AVERAGED OVER 300 INDEPENDENT MC TRIALS .
T HE EMPIRICAL PROBABILITY OF K
snapshots L
VALSE, noninfo. prior
VALSE, prior
2
31%
95%
4
33%
44%
6
3%
5%
8
0
0
10
0
0
C. Estimation by varying M
The performance is examined by varying the number of measurements per snapshots, and the results are presented in Fig.
5. The wrap-around distance between any two frequencies is larger than 4ω = 2π
N . For the first subfigure, the observations
in Fig. 3(a) is also applicable in this scenario. For the last subfigure, the SPA algorithm is the first to approach the CRB, and
then VALSE follows. The VALSE with prior information has better performance than VALSE with noninformative prior. The
NOMP algorithm approaches CRB lastly.
12
1
-5
0.9
SPA, K unknown
NOMP
VALSE, noninfo. prior
VALSE, prior
-15
NMSE of freq. est. (dB)
0.8
SPA, K known
NOMP
VALSE, noninfo. prior
VALSE, prior
CRB
-10
0.7
0.6
0.5
0.4
0.3
-20
-25
-30
-35
-40
0.2
-45
0.1
0
-50
10
15
20
25
30
35
10
15
20
25
M
30
35
M
(a)
(b)
1
1
0.9
0.9
0.8
0.8
0.8
0.7
0.7
0.7
0.6
0.6
0.6
0.5
CCDF
1
0.9
CCDF
CCDF
Fig. 5. The performance of algorithms by varying measurements M . We set SNR = 0dB and the number of snapshots L = 10. Simulation results are
averaged over 300 MC trials.
0.5
0.5
0.4
0.4
0.4
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0
-60
-50
-40
-30
-20
-10
0
-60
0
-50
-40
-30
NMSE(dB)
-20
-10
0
NMSE(dB)
(a)
0
-60
-50
-40
-30
-20
-10
0
NMSE(dB)
(b)
(c)
Fig. 6. Simulation results are averaged over 1000 MC trials. The true κ
e of the three subfigures are 102 , 103 and 104 , respectively.
D. Estimation by encoding with different prior distribution
In this subsection, we investigate the VALSE algorithm with the assumed prior distribution. We set M = N = 20, L = 10,
K = 4, SNR = −2 dB. From Fig. 6, it can be seen that the VALSE with noninformative prior works best for κ
e = 102 in
5
terms of the distribution of the NMSE, while the performance of the VALSE with κa = 10 is the worst. As presented in Table
III, the VALSE with noninformative prior also achieves the highest correct model order estimation probability. For κ
e = 103 ,
3
4
VALSE with κa = 10 works best. In contrast, VALSE with noninformative prior is the worst. For κ
e = 10 , the VALSE
with κa = 105 works best, and VALSE with noninformative prior is the worst. In addition, for κ
e = 103 and κ
e = 104 , the
probability of correct model order estimation increases with κa , as shown in Table III. In summary, this numerical experiment
demonstrates that for small κ
e ≤ 102 , utilizing the assumed prior distribution with large κa makes the VALSE algorithm overfit
and deviate from the true frequencies. For medium κ
e ≈ 103 , utilizing the assumed prior distribution with marginally larger
κa will make the VALSE algorithm work better. In addition, the probability of correct model order estimation increases as
κa increases, while the performance degrades in terms of the distribution of the NMSE. For large κ
e ≈ 104 , it is almost a
grid-based method, and increasing κa improves the performance of VALSE algorithm.
TABLE III
MODEL ORDER ESTIMATION PROBABILITY AVERAGED OVER 1000 INDEPENDENT MC TRIALS OF VALSE FOR DIFFERENT MEASURE OF
CONCENTRATION . F OR EACH ROW, WE DRAW FREQUENCIES FROM MIXTURES OF THE VON M ISES DISTRIBUTIONS WITH κ = κ̃, AND WE USE THE
VALSE ALGORITHM WITH THE ASSUMED PRIOR DISTRIBUTION SUCH THAT κ = κa .
κ̃\κa
102
103
104
0
0.597
0.660
0.681
102
0.562
0.671
0.693
103
0.588
0.731
0.746
104
0.521
0.760
0.790
105
0.481
0.757
0.795
13
VI. C ONCLUSION
In this paper, the VALSE algorithm is extended to jointly estimate the frequencies and weight coefficients in the MMVs
setting. Different from the sparse optimization-based approach, the VALSE algorithm can easily incorporate the prior information of the frequencies, and provide the posterior mean and variance of the estimated frequencies. Meanwhile, the proposed
prior distribution provides good interpretation of the VALSE algorithm in terms of the tradeoff between grid and off-grid based
methods. Furthermore, substantial numerical experiments demonstrate the excellent performance of the VALSE algorithm.
VII. A PPENDIX
A. Finding a local maximum of ln Z(s)
Finding the globally optimal binary sequence s of (36) is hard in general. As a result, a greedy iterative search strategy is
adopted [22]. We proceed as follows: In the pth iteration, we obtain the kth test sequence tk by flipping the kth element of
(p)
(p)
s(p) . Then we calculate ∆k = ln Z(tk ) − ln Z(s(p) ) for each k = 1, · · · , N . If ∆k < 0 holds for all k we terminate the
(p)
algorithm and set b
s = s(p) , else we choose the tk corresponding to the maximum ∆k as s(p+1) in the next iteration.
0
When k 6∈ S, that is, sk = 0, we activate the kth component of s by setting sk = 1. Now, S 0 = S ∪ {k}.
∆k = ln Z(s0 ) − ln Z(s)
ν
λ
ν
ν
ν
−1
+ L ln + ν −1 tr((HH
HS 0
= ln det((JS 0 + I|S 0 | )−1 ) − ln det((JS + I|S| )−1 ) + ln
S 0 (JS 0 + I|S 0 | )
τ
τ
1−λ
τ
τ
ν
−1
HS )).
(41)
− HH
S (JS + I|S| )
τ
To compute ∆k , we introduce two formulas
!
!
1
1
1
D2
D2
0
D− 2 u
D u
1
1
,
(42)
=
1
uH 1
uH D− 2 (1 − uH D−1 u) 2
0
(1 − uH D−1 u) 2
D
uH
−1
u
=
1
−1
H
−1
D
D−1 + D1−uuu
T D−1 u
uH D−1
− 1−uH D−1 u
−1
− 1−uDH Du−1 u
!
.
1
1−uH D−1 u
(43)
Let jk = JS,k denote the kth column of JS and hk = Hk,: denote the kth row of H. By using formulas (42) and (43), we
obtain
ν
JS + τν I|S|
L ln(det(JS 0 + I|S 0 | )−1 ) = L ln det
jH
τ
k
J +ν I
−1
S
jk
τ |S|
ν
ν
ν 0
M+ τ
M+ τ
= L ln (M + )|S | det
jH
k
τ
1
ν
jk
M+
−1
ν
τ
M+ τ
ν
ν
ν
= L ln det(JS + I|S| )−1 (M + − jH
I|S| )−1 jk )−1
k (JS +
τ
τ
τ
ν
ν
ν
I|S| )−1 jk )),
= −L ln det(JS + I|S| ) − Lln((M + − jH
K (JS +
τ
τ
τ
JS + τν I|S|
ν
τ
M+
ν
jH
k
τ
M + τν
ν −1 HH
S 0 (JS 0
ν
1
+ I|S 0 | )−1 HS 0 = ν −1 HH
S0
τ
M+
=ν −1 HH
S (JS +
ν
I|S| )−1 HS +
τ
ν(M +
ν
τ
QH Q
,
−
+ τν I|S| )−1 jk )
jH
k (JS
ν
−1
where Q = (hk − jH
HS ). Plugging (44) and (45) in (41), and let
k (JS + τ I|S| )
ν
ν
−1
vk = ν(M + − jH
jk )−1 and
k (JS + I|S| )
τ
τ
ν
−1
uk = ν −1 vk (hk − jH
HS ),
k (JS + I|S| )
τ
∆k can be simplified as
∆k = L ln
jk
M + τν
vk
uk uH
λ
k
+
+ ln
.
τ
vk
1−λ
1
(44)
−1
HS 0
(45)
(46)
(47)
14
We update the mean and covariance of the weights after changing s
!
b0 0
b S,0
b
C
c0k,0
ν
C
−1
S \k,0
0
0
=
ν(J
+
I
)
=
S
|S |
0H
0
b
0
τ
b
ck,0
Ckk,0
to s0 ,
b S,0 jk
b S,0 jk H
0
ν −1 C
ν −1 C
+ vk
,
0
−1
−1
(48)
b 0 0 is obtained if b
b0
b0
C
c0k,0 , b
c0H
S ,0
k,0 and Ckk,0 are inserted appropriately in CS 0 \k,0 , and
c 0 0 = ν −1 C
b 0 0 HS0
W
S
S ,0
1
=
ν
b S,0 HS +
C
(
b i0 =
w
1
1
Hb
b
b
ν 2 vk CS,0 jk jk CS,0 HS − ν vk CS,0 jk hk
1
Hb
− ν vk jk CS,0 HS + vk hk
uT
k,
T
b i − ν −1b
w
cH
i,0 jk uk ,
!
i=k
,
i∈S
b S,0 .
where b
ci,0 denotes the ith column of C
For the deactive case with sk = 1, s0k = 0 and S 0 = S\{k}, (48) and (49) can be rewritten as
!
−1 0
−1 0
b S\k,0 b
b0 0
b 0 jk
b 0 jk H
C
ck,0
C
0
ν C
ν C
S
,0
S
,0
S
,0
+ vk
,
bkk,0 =
b
0
0
−1
−1
cH
C
,
(49)
(50)
k,0
(
uT
k,
bi =
w
T
b i0 − ν −1b
c0H
w
i,0 jk uk ,
i=k
,
i ∈ S0
(51)
b 0 0 , vk and uk are given by (46). From (50) and (51), we obtain vk = C
bkk,0 and
where b
c0i,0 denotes the ith column of C
S ,0
T
b k , respectively. ∆k (47) is given by
uk = w
∆k = −L ln
bkk,0
uk uH
C
w H wk
λ
vk
λ
k
−
= −L ln
− k
− ln
.
− ln
bkk,0
τ
vk
1−λ
τ
1−λ
C
(52)
T
b 0H0 , (32)
b0 0 = C
b 0 0 jk . According to C
b i0 = w
b i + ν −1b
c0H
ck,0 = ν −1 vk C
From (51), we obtain w
i,0 jk uk . From (50), we obtain b
S ,0
S ,0
S ,0
and (34), we get
b
ν −1 vk b
c0H
i,0 jk = Cik,0 .
(53)
bkk,0 and w
b kT and using (53), the mean of the weights can be updated as
By replacing vk and uk as C
b i0 = w
bi −
w
bik,0
C
b ,
w
bkk,0 k
C
i ∈ S 0.
(54)
b 0H0 , (32) and (34), C
b0 0 = C
b 0 0 jk )H . According to C
b0 0 = C
b S\k,0 − vk (ν −1 C
b 0 0 jk )(ν −1 C
b0
From (50), we obtain C
ij,0
S ,0
S ,0
S ,0
S ,0
S ,0
can be written as
!
!
b
vk b
c0H
jk
vk jH
c0j,0
1
i,0
k
0
bij,0 = C
bij,0 −
C
, i, j ∈ S 0 .
vk
ν
ν
−1
b
bH = C
bkj,0 . We also have vk = C
bkk,0 . Therefore,
According to (53), we obtain ν −1 vk b
c0H
vk jH
c0j,0 = C
i,0 jk = Cik,0 and ν
kb
jk,0
the covariance of the weights can be updated as
b
b
b0 = C
bij,0 − Cik,0 Ckj,0 ,
C
ij,0
bkk,0
C
i, j ∈ S 0 .
(55)
R EFERENCES
[1] P. Stoica and R. L. Moses, Spectral Analysis of Signals. Upper Saddle River, NJ, USA: Prentice-Hall, 2005.
[2] W. Bajwa, A. Sayeed, and R. Nowak, “Compressed channel sensing: A new approach to estimating sparse multipath
channels,” Proc. IEEE, vol. 98, pp. 1058-1076, Jun. 2010.
[3] B. Ottersten, M. Viberg and T. Kailath, “Analysis of subspace fitting and ML techniques for parameter estimation from
sensor array data,” IEEE Trans. Signal Process., vol. 40, pp. 590-600, Mar. 1992.
[4] R. Schmidt, “Multiple emitter location and signal parameter estimation,” IEEE Trans. on Antennas and Propagation, vol.
34, no. 3, pp. 276-280, 1986.
15
[5] R. Roy and T. Kailath, “ESPRIT - estimation of signal parameters via rotational invariance techniques,” IEEE Trans. on
Acoustics, Speech and Signal Processing, vol. 37, no. 7, pp. 984-995, 1989.
[6] P. Stoica and Y. Selen, “Model-order selection: a review of information criterion rules,” IEEE Signal Processing Magazine,
vol. 21, no. 4, pp. 36-47, July 2004.
[7] D. Malioutov, M. Cetin and A. Willsky, “A sparse signal reconstruction perspective for source localization with sensor
arrays,” IEEE Trans. Signal Process., vol. 53, no. 8, pp. 3010-2022, 2005.
[8] P. Stoica, P. Babu and J. Li, “New method of sparse parameter estimation in separable models and its use for spectral
analysis of irregularly sampled data,” IEEE Trans. Signal Process., vol. 59, no. 1, pp. 35-47, Jan. 2011.
[9] P. Stoica, P. Babu and J. Li, “SPICE: A sparse covariance-based estimation method for array processing,” IEEE Trans.
Signal Process., vol. 59, no. 2, pp. 629-638, Feb. 2011.
[10] P. Stoica and P. Babu, “SPICE and LIKES: Two hyperparameter-free methods for sparse-parameter estimation,” Signal
Process., vol. 92, no. 7, pp. 1580-1590, 2012.
[11] L. Hu, Z. Shi, J. Zhou and Q. Fu, “Compressed sensing of complex sinusoids: An approach based on dictionary refinement”
IEEE Trans. Signal Process., vol. 60, no. 7, pp. 3809-3822, 2012.
[12] G. Tang, B. Bhaskar, P. Shah and B. Recht, “Compressed sensing off the grid,” IEEE Trans. Inf. Theory, vol. 59, no. 11,
pp. 7465-7490, 2013.
[13] Z. Yang and L. Xie, “On gridless sparse methods for line spectral estimation from complete and incomplete data,” IEEE
Trans. Signal Process., vol. 63, no. 12, pp. 3139-3153, 2015.
[14] Z. Yang and L. Xie, “Continuous compressed sensing with a single or multiple measurement vectors,” IEEE Workshop
on Statistical Signal Processing, pp. 288-291, 2014.
[15] Y. Li and Y. Chi, “Off-the-grid line spectrum denoising and estimation with multiple measurement vectors,” IEEE Trans.
Signal Process., vol. 64, no. 5, pp. 1257-1269, 2016.
[16] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004.
[17] Z. Yang, L. Xie and C. Zhang, “A discretization-free sparse and parametric approach for linear array signal processing,”
IEEE Trans. Signal Process., vol. 62, no. 19, pp. 4959-4973, 2014.
[18] B. Mamandipoor, D. Ramasamy and U. Madhow, “Newtonized orthogonal matching pursuit: Frequency estimation over
the continuum,” IEEE Trans. Signal Process., vol. 64, no. 19, pp. 5066-5081, 2016.
[19] L. Han, J. Zhu, R. S. Blum and Z. Xu, “Newtonized orthogonal matching pursuit for line spectrum estimation with
multiple measurement vectors,” avaliable at https://arxiv.org/pdf/1802.01266.pdf.
[20] J. Fang, F. Wang, Y. Shen, H. Li and R. S. Blum, “Superresolution compressed sensing for line spectral estimation:an
iterative reweighted approach,” IEEE Trans. Signal Process., vol. 64, no. 18, pp. 4649-4662, 2016.
[21] T. L. Hansen, B. H. Fleury and B. D. Rao, “Superfast line spectral estimation,” avaliable at
https://arxiv.org/pdf/1705.06073.pdf.
[22] M. A. Badiu, T. L. Hansen and B. H. Fleury, “Variational Bayesian inference of line spectrum estimation,” IEEE Trans.
Signal Process., vol. 65, no. 9, pp. 2247-2261, 2017.
[23] D. Zachariah, P. Wirfält, M. Jansson and S. Chatterjee, “Line spectrum estimation with probabilistic priors,” Signal
Processing, vol. 93, no. 11, pp. 2969-2974, 2013.
[24] K. V. Mardia and P. E. Jupp, Directional Statistics. New York, NY, USA: Wiley, 2000.
[25] K. P. Murphy, Machine Learning A Probabilistic Perspective. MIT Press, 2012.
[26] D. P. Bertsekas and J. N. Tsitsiklis : Parallel and Distributed Computation: Numerical Methods, Athenan Scientific:
Massachusetts, 1997.
[27] A. Hjφrungnes, Complex-Valued Matrix Derivatives: With Applications in Signal Processing and Communications,
Cambridge University Press, 2011.
| 7cs.IT
|
Improved MapReduce and Streaming Algorithms for
k-Center Clustering (with Outliers)
arXiv:1802.09205v1 [cs.DC] 26 Feb 2018
Matteo Ceccarello, Andrea Pietracaprina, and Geppino Pucci
Department of Information Engineering
University of Padova, Italy
{ceccarel, capri, geppo}@dei.unipd.it
Abstract
We present efficient MapReduce and Streaming algorithms for the k-center problem
with and without outliers. Our algorithms exhibit an approximation factor which is arbitrarily close to the best possible, given enough resources.
1 Introduction
The k-center clustering problem requires to find, given a set S in a metric space, a subset
C ⊆ S of k points, called centers, minimizing the maximum distance between any point in S
and a center in C. Along with k-median and k-means (which require to minimize the sum
of all distances and the sum of all squared distances to the closest centers, respectively) kcenter is a popular clustering variant. All these clustering problems are NP-hard, therefore
we have to settle for approximate solutions. Also, the objective function of these clustering
problems is at risk of being severely influenced by a few very distant points, called outliers.
This is especially true for the k-center problem, since the objective function is a maximum.
The presence of outliers is inherent in many datasets, since these points are often artifacts of
data collection, or represent noisy measurements, or simply erroneous information. To cope
with these problems, the k-center clustering problem admits a formulation that takes into
account outliers: when computing the objective function, up to z points are allowed to be
discarded, where z is a user-defined input parameter.
In the last decade, the explosive growth of data that needs to be processed has ruled
out sequential approaches that are efficient on small-sized datasets, but impractical on large
ones. Therefore, recent research has focused on developing efficient parallel and distributed
algorithms that are able to cope with large volumes of data. In this setting, a very natural
approach entails partitioning the input and then using a sequential k-center approximation
algorithm in a bootstrapping fashion: we find a set of k centers on each subset of the partitioned input, and then we compute the final solution on the union of these sets of centers.
This is indeed the approach taken by Malkomes et al. [MKC+ 15], who show that it provides
a 4-approximation. A natural question arises: what if we select more centers from each partition in the first phase? Intuitively, we should get a better solution than if we just selected
k points. In fact, as we select more and more points from each partition we should obtain a
solution progressively closer to the one returned by the best sequential algorithm at the expense of larger space requirements. We will leverage the doubling dimension of metric spaces
to characterize the tradeoffs between approximation and space blowup.
1
1.1 Related work
The study of k-center clustering dates back to the 80’s, when Gonzalez proved that it is
impossible to achieve an approximation factor 2 − ε in general metric spaces, unless P =
NP [Gon85]. In the same paper, Gonzalez provided a simple greedy algorithm with running
time O (kn) and approximation factor 2, which is often referred to as GMM in the recent literature. Feder and Greene used a modified version of GMM to develop an algorithm with the
same approximation guarantee and with running time O (n log k) [FG88].
To deal with the problem of outliers, Charikar et al. [CKMN01] introduced the k-center
with outliers problem, where the radius of the clustering is determined by ignoring the z points
that are farthest from their centers, where z is a parameter provided
in input. For this problem, they give a 3-approximation algorithm which runs in O n3 time. Furthermore, they
prove that it is impossible to achieve an approximation factor 3 − ε in general metric spaces,
unless P = NP.
Nowadays, with the dramatic increase in the amount data that has to be processed, traditional sequential algorithms may no longer be a viable option. Recently, a lot of attention
has been devoted to the MapReduce model of computation, where a set of processors with
limited memory can process data in a sequence of parallel rounds [DG08]. In particular, in
the MapReduce model [PPR+ 12] there are two parameters: the amount of local memory ML
available to each processor, and the amount of aggregated memory MA , which is the sum
of local memories of all processors. The goal is to minimize the number of parallel rounds
while using O (ML ) memory locally and O (MA ) memory globally. k-center clustering under
this model has been first studied by Ene et al. [EIM11], who provided a randomized algorithm with approximation factor 10. This result was subsequently improved by Malkomes
et al. [MKC+ 15] with a deterministic 4-approximation algorithm. A similar algorithm has
been also developed by McClintock and Wirth [MW16]. As for the k-center with outliers
problem, Malkomes et al [MKC+ 15] proposed a deterministic 13-approximation MapReduce
algorithm.
In the MapReduce setting, a very useful tool is represented by composable coresets, introduced by Indyk et al. [IMMM14] as an extension of the work by Agarwal et al. [AHPV04].
For a given objective function, a coreset is small subset of the input such that the cost of the
optimal solution on the subset is close to the cost of the optimal solution on the original set.
The additional property of composability requires that, for a partition of the input set, if we
separately build a coreset for each subset, the cost of the optimal solution on the union of
the coresets is close to the optimum on the whole input. Clearly, this approach enables the
development of parallel algorithms where each processor computes the coreset relative to
a single subset of the partitioned input, delegating the computation of the final solution to
another processor that receives all the coresets. Composable coresets have been used for a
number of problems, including diversity maximization [IMMM14, AFZ15, CPPU17, CPP18],
submodular maximization [MZ15], and graph matching and vertex cover [AK17].
Another option when dealing with large amounts of data is to resort to streaming. The
streaming model requires algorithms to use a single processor with limited memory, allowing
only a few sequential passes over the input (ideally just one) [HRR98]. Originally developed
for the external memory setting, this model also captures the scenario in which data is generated on the fly and must be analyzed in real-time, for instance on a social media platform. Under this model, Charikar et al. [CCFM04] developed a single pass algorithm which requires
Θ (k) space and computes an 8-approximation, deterministically, or a 5.43-approximation,
probabilistically.
Later, this algorithm has been improved by McCutchen and Khuller [MK08]
to achieve a
2 + ε approximation factor, deterministically, using space Θ kε−1 log(ε−1 ) . In the same paper, the authors give a deterministic 4 + ε-approximation
streaming algorithm for the prob
lem with outliers, using space O k · z · ε−1 .
2
2 Preliminaries
Consider a metric space with distance function d(·, ·), and recall that d is symmetric, nonnegative, equal to 0 only on pairs of identical points, and satisfies the triangle inequality.
Also, for a point u of the metric space, the ball of radius r centered at u is the set of points at
distance at most r from u. The doubling dimension of a metric space is the smallest D such that
for any radius r and point u, all points in the ball of radius r centered at u are included in
the union of at most 2D balls of radius r/2 centered at suitable points. Consequently, for any
0 < ε 6 1, a ball of radius r can be covered with at most (1/ε)D balls of radius εr.
Consider now a set S of points from a metric space. For notational convenience, we define
the distance between a point s ∈ S and a set X as
d(s, X) = min d(s, x)
x∈X
Let C ⊆ S. We define the radius of S with respect to C as
rC (S) = max d(s, C)
(1)
s∈S
The k-center problem requires to find a subset C ⊆ S of size k such that rC (S) is minimized.
We define r∗k (S) the radius achieved by the optimal solution to the problem. Note that the
set of centers C induces a partition of S into k clusters by assigning each point to the cluster
associated to the closest point of C.
In the related work section, we mentioned the GMM algorithm [Gon85], which provides
a 2-approximation to the k-center problem. Here we briefly review how it works. Given a
set S, GMM builds the set of centers T incrementally in k iterations. An arbitrary point of S is
selected as the first center and is added to T . Then, the algorithm iteratively selects a point
maximizing the distance from T as a new center and adds it to T , until T contains k centers.
Note that, rather than setting k a priori, GMM can be used by growing the set T until a
target radius is achieved. In fact, the radius of the clustering induced by the set of centers
built by GMM is a non-increasing function of the iteration number. In this paper, we will
make use of the following property of GMM.
Lemma 1. Consider a set S, and let X ⊆ S. For a given k, let TX be the output of
have
rTX (X) 6 2 · r∗k (S)
GMM
on X. We
Proof. We prove this lemma by rephrasing the proof by [Gon85] in terms of subsets. We want
to prove that, ∀x ∈ X, d(x, TX ) 6 2 · r∗k (S). Assume by contradiction that this is not the case.
Then for some y ∈ X it holds that d(y, TX ) > 2 · r∗k (S). By the greedy choice of GMM, we have
that for any pair t1 , t2 ∈ TX , d(t1 , t2 ) > d(y, TX ), otherwise y would have been included in
TX . So we have that d(t1 , t2 ) > 2 · r∗k (S). Therefore, the set {y} ∪ TX consists of k + 1 points at
distance > 2 · r∗k (S) from each other. Consider now the optimal solution to k-center on the set
S. Since ({y} ∪ TX ) ⊆ S, two of the k + 1 points of {y} ∪ TX , say x1 and x2 , must be closest to the
same optimal center o∗ . By the triangle inequality we have d(x1 , x2 ) 6 d(x1 , o∗ ) + d(o∗ , x2 ) 6
2 · r∗k (S), which yields a contradiction, since d(x1 , x2 ) > 2 · r∗k (S).
The k-center with outliers problem requires to identify, for fixed k and z, a set of centers
C of size k and a set Z of z outliers such that
rC,Z (S) = max d(s, C)
s∈S\Z
(2)
is minimized. In other words, the problem allows to discard up to z points of the input when
computing the radius of the set of centers. We denote the radius of the optimal solution
of this problem, for given S, k, and z, with r∗k,z (S). The following lemma ensures that the
optimal solution of the problem without outliers with k + z centers has a smaller radius than
the optimal solution of the problem with k centers and z outliers.
3
Lemma 2. Given a set S and parameters k, and z
r∗k+z (S) 6 r∗k,z (S)
Proof. Consider Ok,z , an optimal solution to the problem with k centers and z outliers.
Clearly, it is an admissible solution to the problem with k + z centers without outliers. By
the optimality of r∗k+z (S) we have r∗k+z (S) 6 rOk,z (S) = r∗k,z (S).
We adopt the formulation of the MapReduce computational model introduced
in [PPR+ 12]. A MapReduce algorithm executes as in a sequence of parallel rounds, where,
in a round, a multiset X of key-value pairs is transformed into a new multiset Y of pairs by
applying a given reducer function (simply called reducer) independently to each subset of
pairs of X having the same key. The model features two parameters: ML is the local memory
available to each reducer, and MA is the aggregate memory across all reducers.
3 MapReduce algorithms for k-center and k-center with z outliers
In the following subsections we will develop MapReduce algorithms for the k-center problem
(Subsection 3.1) and the k-center problem with z outliers (Subsection 3.2). Both algorithms
adopt the composable coreset framework of [IMMM14], reviewed in the introduction, can be
viewed as improved variants of the algorithms in [MKC+ 15]. Specifically, we are able to
achieve an approximation quality arbitrarily close to the one guaranteed by the best known
sequential algorithms, at the expense of a local memory blowup which is modest for inputs
drawn from metric spaces of bounded doubling dimension.
3.1 MapReduce algorithms for k-center
Consider an instance S of the k-center problem and fix some ε ∈ (0, 1]. The MapReduce
algorithm works in two rounds. In the first round, S is partitioned into ℓ subsets Si , with
1 6 i 6 ℓ of equal size. In parallel, for each Si we run GMM incrementally and call Tij the set
of j centers selected in the first j iterations of the algorithm. Let rT k (Si ) denote the radius of
i
the clustering induced by the first k centers (Tik ) on Si . We continue to run GMM until the first
S
τ
iteration τi > k such that rT τi (Si ) 6 ε/2 · rT k (Si ), and let Ti = Ti i . Define T = ℓi=1 Ti .
i
i
In the second round, the set T is gathered into a single reducer and GMM is run on T to
compute the final set of k centers. In the following we will show that these centers are a good
solution for the k-center problem on S.
We define a proxy function p : S → T that maps each s ∈ Si into the closest point in Ti , for
every 1 6 i 6 ℓ. Observe that since S1 , . . . , Sℓ is a partition of S, the function p is defined for
every point of S and that each point of T is mapped into itself.
Lemma 3. The proxy function p : S → T is such that:
d(x, p(x)) 6 ε · r∗k (S) ∀x ∈ S
Proof. Fix i ∈ [1, ℓ], and consider Si ⊆ S, and the set Tik computed by the first k iterations of
GMM . Since Si is a subset of S, by Lemma 1 we have that rT k (Si ) 6 2 · r∗
k (S). By construction,
i
∗
we have that rTi (Si ) 6 ε/2 · rT k (Si ). It follows that rTi (Si ) 6 εrk (S). Consider now the proxy
i
function p. For every 1 66 ℓ and s ∈ Si , it holds that d(s, p(s)) 6 rTi (Si ) 6 εr∗k (S).
We can conveniently bound the size of T as a function of the doubling dimension of the
underlying metric space.
4
Lemma 4. If the points of S belong to a metric space of doubling dimension D, then the set T constructed with the above procedure is such that
D
8
|T | 6 ℓ · k ·
.
ε
Proof. Fix an i ∈ [1, ℓ]. We prove an upper bound on the number τi of iterations of GMM
which are needed to obtain rT τi (Si ) 6 (ε/2)rT k (Si ), thus obtaining a bound on the size
i
i
of Ti . Consider an optimal k-center clustering of Si , with radius r∗k (Si ). By the definition
of doubling dimension, we have that each of the k clusters can be covered using at most
(1/η)D balls of radius 6 η · r∗k (Si ), for any constant η ∈ (0, 1]. By selecting from each such
ball an arbitrary point as a center and using the triangle inequality, we can show that there
exists a clustering of k(1/η)D clusters with radius 6 2η · r∗k (Si ). Fix now η = ε/8 and let
h = k(1/η)D = k(8/ε)D . We have:
r∗h (Si ) 6 2η · r∗k (Si ) = ε/4 · r∗k (Si )
(3)
by the optimality of r∗h (Si ). Hence, after h iterations we are guaranteed that GMM finds a set
Tih such that
rT h (Si ) 6 2 · r∗h (Si )
(by the approximation factor of GMM)
i
6 ε/2 · r∗k (Si )
(by Inequality 3)
(by the optimality of r∗k (Si ))
6 ε/2 · rT k (Si )
i
which meets the stopping condition. Therefore, τi 6 h, hence the coreset Ti for Si has size at
most k(8/ε)D . The lemma follows, by summing over all i ∈ [1, ℓ].
We have:
Theorem 1. Let 0 < ε 6 1. If the points of S belong to a metric space of doubling dimension D, then
the above 2-round MapReduce algorithm computes a (2 + ε)-approximation for the k-center clustering
problem with memory requirements
• MA = O (|S|)
|S|
• ML = O ℓ + ℓ · k ·
8 D
ε
Proof. Let X be the solution found by GMM on T . Since T ⊆ S, from Lemma 1 it follows that
rX (T ) 6 2 · r∗k (S). Consider an arbitrary point s ∈ S, along with its proxy p(s) ∈ T , as defined
before. By Lemma 3 we know that d(s, p(s)) 6 ε · r∗k (S). Let x ∈ X be the center closest to
p(s). It holds that d(x, p(s)) 6 2 · r∗k (S). By applying the triangle inequality, we have that
d(x, s) 6 d(x, p(s)) + d(p(s), s) 6 2 · r∗k (S) + ε · r∗k (S) = (2 + ε)r∗k (S).
While computing the coresets in the first round, each processor
needs to store |S|/ℓ points
of the input, and computes a coreset of size O k · (8/ε)D , as per Lemma 4. In the second
round, a single processor needs enough memory to
store ℓ such coresets, therefore the memory needed in the second round is O ℓ · k · (8/ε)D , and the bound on ML follows. Finally, it
is immediate to see that aggregate memory proportional to the input size suffices.
3.2 MapReduce algorithms for k-center with outliers
Consider an instance S of the k-center problem with z outliers and fix some ε̂ ∈ (0, 1]. As in
the case with no outliers, our MapReduce algorithm works in two rounds. In the first round,
S is partitioned into ℓ subsets Si , with 1 6 i 6 ℓ of equal size. In parallel, for each Si we run
j
GMM incrementally and call Ti the set of the first j selected centers. Let rT k+z (Si ) denote the
i
5
radius of the clustering induced by the first k + z centers. We continue to run GMM until the
first iteration τi > k + z such that rT τi (Si ) 6 ε̂/2 · rT k+z (Si ). Let Ti = Tiτi and T = ∪ℓi=1 Ti . As
i
i
before, for each point s ∈ Si we define its proxy p(s) ∈ T to be the point of Ti closest to s, and
for each t ∈ Ti we define its weight wt as the number of points of Si for which t is the proxy.
In the second round, the union T of the weighted coresets is gathered into a single
reducer, which runs the following weighted variant (dubbed O UTLIERS C LUSTER) of the
sequential algorithm by [CKMN01] for the k-center problem with outliers.
Algorithm 1: O UTLIERS C LUSTER(T , k, r, ε̂)
T′ ← T
X←∅
while ((|X| < k) and (T ′ 6= ∅)) do
for t ∈ T do
⊲ Consider also already covered nodes as centers for balls
′
Bt ← {v : v ∈ T ∧ d(v, t) 6 (1 + 2ε̂) · r}
P
x ← arg maxt∈T v∈Bt wv
X ← X ∪ {x}
Ex ← {v : v ∈ T ′ ∧ d(v, x) 6 (3 + 4ε̂) · r}
T ′ ← T ′ \ Ex
return X, T ′
O UTLIERS C LUSTER(T , k, r, ε̂) returns two subsets X, T ′ ⊆ T such that X is a set of (at most) k
centers, and T ′ is the set of all points at distance greater than (3 + 4ε̂) · r from X. The set of
centers X is determined in |X| 6 k iterations as follows. The algorithm maintains a set T ′ of
uncovered points (initially T ′ = T ). In each iteration the next element x of X is chosen as the
point maximizing the number of uncovered points in its ball of radius (1 + 2ε̂) · r. Note that
x is not required to be an uncovered point. Then, all uncovered points at distance at most
(3 + 4ε̂) · r from x are removed from T ′ . The algorithm terminates when either k centers have
been added to X or no uncovered points are left. It is immediate to see that, at the end, T ′ is
exactly the set of points at distance greater than (3 + 4ε̂) · r from X.
We run O UTLIERS C LUSTER(T , k, r, ε̂) for O (log
|T |) different values of r, established
through a binary search protocol on the O |T |2 distances between points of T , until the
smallest r is found such that the aggregate weight of the points in the set T ′ returned by
O UTLIERS C LUSTER(T , k, r, ε̂) is at most
z. We remark that the next value of r to test can
be identified in sequential time O |T |2 and space O (|T |), through an immediate adaptation
the approximate median streaming algorithm by [MRL98] applied to the distances
between
points of T . Hence, the time required for this identification is within the O k|T |2 time taken
by O UTLIERS C LUSTER.
The set of centers X computed in the last call of O UTLIERS C LUSTER is finally returned as
solution of the k-center problem with z outliers for S.
The analysis of the algorithm proceeds as a sequence of lemmas. The first lemma establishes a bound on the distance between any point and its proxy.
Lemma 5. The proxy function p : S → T is such that:
d(x, p(x)) 6 ε̂ · r∗k,z (S)
∀x ∈ S
Proof. Consider any subset Si of the partition S1 , . . . , Sℓ of S. By construction, we have that
d(x, p(x)) 6 (ε̂/2) · rT k+z (Si ). Since Si is a subset of S, Lemma 1 ensures that rT k+z (Si ) 6
i
i
2r∗k+z (S). Hence, d(x, p(x)) 6 ε̂r∗k+z (S). Since r∗k+z (S) 6 r∗k,z (S), as established by Lemma 2,
we have d(x, p(x)) 6 ε̂ · r∗k,z (S).
6
We will now bound the size of the set T ∪ℓi=1 Ti computed in the first round.
Lemma 6. If the points of S belong to a metric space of doubling dimension D, then the set T built in
the first round of the algorithm is such that
D
1
|T | 6 ℓ · (k + z) ·
ε̂
Proof. The proof proceeds similarly to the second part of the proof of Lemma 4, with the
understanding that the definition of doubling dimension is applied to each of the (k + z)
clusters induced by the points of Tik+z on Si .
The next lemma characterizes the quality of the solution provided by O UTLIERS C LUSTER
when run on T and the optimal radius r = r∗k,z (S). For now we assume that the value r∗k,z (S)
is known and we will argue later how the binary search performed in the second round of
our algorithm computes an approximation to this value, which is suitable for our purposes.
Lemma 7. Given a set S and parameters k and z, consider the set T ⊆ S constructed in the first round
of the algorithm, and let X, T ′ ⊆ T be the sets returned by O UTLIERS C LUSTER (T , k, r∗k,z (S), ε̂). Let
also ST ′ = {s ∈ S : p(s) ∈ T ′ }. Then, we have that
d(s, X) 6 (3 + 5ε̂) · r∗k,z (S)
∀s ∈ S \ ST ′
and |ST ′ | 6 z.
Proof. Consider a point s ∈ S \ ST ′ . We have that p(s) ∈ T \ T ′ , hence d(p(s), X) 6 (3 +
4ε̂)r∗k,z (S). Since d(s, p(s)) 6 ε̂ · r∗k,z (S), by the triangle inequality we obtain that
d(s, X) 6 d(s, p(s)) + d(p(s), X) 6 (3 + 5ε̂) · r∗k,z (S),
which proves the first inequality. We are left to show that |ST ′ | 6 z. Suppose first that |X| < k.
In this case, it must be T ′ = ∅, hence |ST ′ | = 0, and the proof follows. We now concentrate on the case |X| = k. Consider the i-th iteration of the while loop of O UTLIERS C LUSTER
(T , k, r∗k,z (S), ε̂) and define xi as the center of X selected in the iteration, and Ti′ as the residual set of uncovered points at the beginning of the iteration. Recall that xi is the point of Ti′
which maximizes the cumulative weight of the set Bxi of uncovered points in Ti′ at distance
at most (1 + 2ε̂) · r∗k,z (S) from xi , and that the set Exi of all uncovered points at distance at
most (3 + 4ε̂) · r∗k,z (S) is removed from Ti′ at the end of the iteration.
We now show that
k X
X
wt > |S| − z,
i=1 t∈Exi
which will immediately imply that |ST ′ | 6 z. For this purpose, let O be an optimal set of
k centers for the problem instance under consideration, and let Z be the set of at most z
outliers at distance greater than r∗k,z (S) from O. For each o ∈ O, define Co ⊆ S \ Z as the
set of nonoutlier points which are closer to o than to any other center of O, with ties broken
arbitrarily. We will show that there exists an ordering o1 , o2 , . . . , ok of the centers in O so that,
for each 1 6 i 6 k,
i X
X
wt > |Co1 ∪ · · · ∪ Coi |.
j=1 t∈Exj
The proof proceeds by induction on i and uses a charging argument to assign each point in
S
Co1 ∪ · · · ∪ Coi to a point in ij=1 Exj , where each t in the latter set will be in charge of at
most wt points. We define two charging rules. A point can be either charged to its own proxy
(charging rule 1) or to another point of T (charging rule 2).
7
Coi
Exj
Bxj
6
v
(1 + 2ε̂)r∗k,z (S)
xj
6 r∗k,z (S)
6 r∗k,z (S)
oi
u
6 ε̂r∗k,z (S)
p(v)
6 ε̂r∗k,z (S)
p(u)
Figure 1: Path between a center xj and a the proxy of a node u ∈ Oi .
The basis of the induction (i = 0) is trivial. Assume, as inductive hypothesis, that the
S
points in Co1 ∪ · · · ∪ Coi−1 have been charged to points in i−1
j=1 Ej , for some i > 1 and some
choice of distinct optimal centers o1 , o2 , . . . , oi−1 . Consider now the i-th iteration of O UTLIERS C LUSTER. We have one of the following two cases.
Case 1 For this case we assume that among the still unchosen optimal centers, there is one,
call it oi , such that there exists a point v ∈ Coi whose proxy p(v) belongs to Bxj , for some
1 6 j 6 i. Hence d(xj , p(v)) 6 (1 + 2ε̂) · r∗k,z (S). By repeatedly applying the triangle inequality
(see Figure 1) we have that for each point u ∈ Coi
d(xj , p(u))
6 d(xj , p(v)) + d(p(v), v) + d(v, oi ) + d(oi , u) + d(u, p(u))
6 (3 + 4ε̂) · r∗k,z (S),
since each point is at distance at most ε̂r∗k,z (S) from its proxy, and each point of Coi is at
distance at most r∗k,z (S) from oi . The above inequality implies that p(u) ∈ Exj . Therefore we
can charge each point u ∈ Oi to its proxy, by charing rule 1.
Case 2 If the previous assumption does not hold, it means that for each unchosen optimal
center o and each v ∈ Co , p(v) does not belong to Bx1 ∪ Bx2 ∪ · · · ∪ Bxi . In this case, we set oi
to be the unchosen optimal center which maximizes the cardinality of {p(u) : u ∈ Coi } ∩ Ti′ ,
that is the number of points whose proxy is still uncovered. Figure 2 depicts the situation.
Clearly, there may be points u ∈ Coi such that p(u) ∈
/ T ′ , that is p(u) ∈ Ex1 ∪ · · · ∪ Exi−1 .
We charge these points to their own proxy by rule 1.
As for the other points, we will show that we can charge them to the points of the greedy
disk Bxi centered at xi . By repeatedly applying the triangle inequality (see Figure 3) we have
that for each point u ∈ Coi
d(p(oi ), p(u))
6 d(p(oi ), oi ) + d(oi , u) + d(u, p(u))
6 (1 + 2ε̂) · r∗k,z (S).
Consider Bp(oi ) , that is the set of points in Ti′ at distance at most (1 + 2ε̂)r∗k,z (S) from
poi . Consider all the points {u ∈ Coi : p(u) ∈ Ti′ }: by the above discussion they
are contained in Bp(oi ) , therefore the weight of the uncovered points of Bp(oi ) is at least
u ∈ Coi : p(u) ∈ Ti′ . The i-th iteration of the algorithm considers p(oi ) as a candidate for
8
points with their proxy not
covered are charged to Bxi
by rule 2
points with their proxy
covered by Exj , for some
j < i, are charged to their
proxy by rule 1
oi
Coi
xi
Bxi
Exj
Figure 2: Application of charging rules in case 2 of the proof. Round points are points of S,
whereas star-shaped points are proxy points in T . Arrows represent charging.
6 ε̂r∗k,z (S)
6 r∗k,z (S)
oi
p(oi )
6 ε̂r∗k,z (S)
u
p(u)
Figure 3: Distance between the proxies of oi and u ∈ Coi .
being selected as next center xi . Then
X
X
wz
wt >
z∈Bp(o
t∈Bxi
>
(by the greedy selection protocol)
i)
u ∈ Coi : p(u) ∈ Ti′
(by the discussion above)
Thus the points in Bxi have enough weight to be charged with each uncharged point of Coi .
Note that no further charging to points of Bxi by rule 1 will happen in subsequent iterations,
since we are applying rule 2 and thus Bxi does not intersect the set Co of any optimal center
o yet to be chosen. Also, in subsequent iterations charging through rule 2 will only target sets
Bxh with h > i.
Thus, through the above induction argument we showed that the charging rules ensure
k X
X
wt > |Co1 ∪ · · · ∪ Cok | > |S| − z,
j=1 t∈Exj
which completes the proof.
Finally, we have:
Theorem 2. Let 0 < ε 6 1 and set ε̂ = ε/5. If the points of S belong to a metric space of doubling
dimension D, then the above 2-round MapReduce algorithm computes a (3 + ε)-approximation for the
k-center clustering problem with z outliers with memory requirements
• MA = O (|S|)
9
• ML = O
|S|
ℓ
+ ℓ · (k + z) ·
D
1
ε
Proof. By Lemma 7 we know that if we invoke O UTLIERS C LUSTER on T with r = r∗k,z (S) and
ε̂ = ε/5, we will get a set of centers such that all but at most z points are at distance at most
(3 + ε)r∗k,z (S) from the centers. This implies that the binary search in the second round of the
algorithm will stop at a value r 6 r∗k,z (S), thus proving the approximation bound. As for the
memory requirements, in the first round each reducer needs to store |S|/ℓ
points of the input
and, based on Lemma 6, it computes a coreset of size O (k + z)(1/ε)D . In the second round,
the reducer which computes the final solution requires the memory to store the union
of the ℓ
coresets, therefore the memory needed in the second round is O ℓ · (k + z) · (1/ε)D , and the
bound on ML follows. Globally, the reducers need only sufficient memory to store the input,
hence MA = O (|S|).
Observation It is important to remark that, thanks to the incremental nature of GMM, our
coreset-based MapReduce algorithms for the k-center problem, both without and with outliers, need not know the doubling dimension D of the underlying metric space in order to
attain the claimed performance bounds. This is a very desirable property, since, in general, D
may not be known in advance.
4 Streaming algorithm for k-center with outliers
Consider once again an instance S of the k-center problem with z outliers and fix some ε̂ ∈
(0, 1]. Suppose that the points of S belong to a metric space of (known) doubling dimension
D. Our streaming algorithm also adopts a coreset-based approach. Specifically, in a pass over
the stream of points of S a suitable weighted coreset T is selected. Then, at the end of the pass,
the final set of centers is computed using O UTLIERS C LUSTER exactly as in the second round
of the MapReduce algorithm described in Subsection 3.2.
As before, the coreset T consists of τ > k + z centers, where τ depends on ε̂ and D, which
represent a good approximate solution to the τ-center problem on S (without outliers). The
main difference with the MapReduce algorithm is the fact that we cannot exploit the incremental approach provided by GMM, since no efficient implementation of GMM in the streaming setting is known. Hence, for the computation of T we resort to a weighted variant of the
doubling algorithm by Charikar et al. [CCFM04], which is described below.
For a given stream of points S and a target number of centers τ, the algorithm maintains
a set T of centers selected among the points of S processed so far, and a lower bound φ on
r∗τ (S). T is initialized with the first τ points of S while φ is initialized as half of the minimum
distance between the first τ + 1 points of S. For the sake of the analysis, we will define a proxy
function p : S → T which, however, will not be explicitly stored by the algorithm.
The points of S are processed one at a time maintaining the following invariants.
(a) The set T contains at most τ centers
(b) The distance between any two centers in T is > 4φ
(c) The maximum distance between any point s ∈ S processed so far and p(s) ∈ T is 6 8φ
(d) Each center t ∈ T has a weight wt equals to the number of points s ∈ S processed so far
such that p(s) = t.
(e) φ 6 r∗τ (S)
To maintain these invariants, the following two rules are applied to process each new point
s ∈ S. The update rule checks if d(s, T ) 6 8φ. If this is the case, we identify the center t ∈ T
10
closest to s, set p(s) ← t, increment wt by one, and discard s. If instead d(s, T ) > 8φ, then
s is added as a new center to T , setting ws ← 1 and p(s) ← s. Note that in this latter case,
the size of T may exceed τ, thus violating invariant (a). When this happens, the following
merge rule is invoked repeatedly until invariant (a) is re-established. Each invocation of this
rule first sets φ ← 2φ, which, in turn, may lead to a violation of invariant (b). For each pair
of points u and v violating invariant (b), we discard u and set wv ← wv + wu . Conceptually,
this corresponds to the update of the proxy function which sets p(x) ← v, for each point x for
which p(x) was equal to u. Doing so enforces invariant (d).
It is straightforward to see that the above two rules maintain invariants (a)-(d). The following lemma proves that invariant (e) is also maintained.
Lemma 8. At the end of the processing of any point s ∈ S, invariant (e) holds.
Proof. Consider the processing of a point s requiring the invocation of the merge rule and assume that all invariants hold at the beginning. After the update rule is applied, only invariant
(a) is violated and |T | = τ + 1. In particular, each pair of centers in T are at distance at least
4φ from one another (invariant (b)). Let φ ′ be the new value of φ resulting after the required
applications of the merging rule. It is easy to see that until the penultimate application of
the merge rule, T still contains τ + 1 points. Therefore each pair of these points must be at
distance at least 4(φ ′ /2) = 2φ ′ from one another. This implies, that φ ′ is still a lower bound
to r∗τ (S).
We now return to the k-center problem with z outliers. Fix τ = 16(k + z)(1/ε̂)D and let T
be the weighted coreset T of size τ returned by the above variant of the doubling algorithm
for the τ-center problem. Consider also the proxy function that maps S into T conceptually
computed along the way. We have:
Lemma 9. For every s ∈ S
d(s, p(s)) 6 ε̂ · r∗k,z (S).
Proof. Observe that S can be covered using k + z balls of radius r∗k+z (S). Since S comes from a
space of doubling dimension D, we know that S can also be covered using τ = 16(k + z)ε̂−D
balls (not necessarily centered at points in S) of radius 6 ε̂/16 · r∗k+z (S). Picking an arbitrary
center of S from each such ball induces a τ-clustering of S with radius at most ε̂/8 · r∗k+z (S).
Hence,
r∗τ (S) 6 ε̂/8 · r∗k+z (S).
By Lemma 2 we know that r∗k+z (S) 6 r∗k,z (S), therefore r∗τ (S) 6 ε̂/8 · r∗k,z (S). By invariants
(c) and (e) we have that for every s ∈ S
d(s, p(s)) 6 8φ 6 8 · r∗τ (S) 6 ε̂ · r∗k,z (S).
The above lemma states a property about the proxy function p defined on T that is equivalent to the one proved in Lemma 5. We have:
Theorem 3. Let 0 < ε 6 1 and set ε̂ = ε/5. If the points of S belong to a metric space of doubling
dimension D, then the above 1-pass streaming algorithm computes a (3 + ε)-approximation
for the
D
k-center clustering problem with z outliers using memory O (k + z)(1/ε) .
Proof. Let T be the weighted coreset of τ = 16(k + z)(1/ε̂)D points computed by our variant
of the doubling algorithm. By reasoning as in the proof of Lemma 7, we can show that O UTLIERS C LUSTER (T , k, r∗
k,z (S), ε̂) returns a set of centers X such that all but at most z points of
S are within distance (3 + 5ε̂)r∗k,z (S) from X. This implies that the runs of O UTLIERS C LUSTER
dictated by the binary search will stop at a value r 6 r∗k,z (S), thus proving the approximation bound. The memory usage comes directly from the fact that T has size τ and that the
streaming algorithm needs only memory proportional to |T |.
11
Note that by applying the same technique to the k-center problem (without outliers)
we
would obtain a streaming algorithm yielding a 2 + ε approximation using O k(1/ε)D space.
Given that the k-center algorithm by [MK08] provides a 2 + ε approximation using space
O kε−1 log ε−1 , for this problem, in this case our techniques would not improve upon the
state of the art.
References
[AFZ15]
Sepideh Aghamolaei, Majid Farhadi, and Hamid Zarrabi-Zadeh. Diversity maximization via composable coresets. In Proc. CCCG, 2015.
[AHPV04] Pankaj K. Agarwal, Sariel Har-Peled, and Kasturi R Varadarajan. Approximating
extent measures of points. Journal of the ACM, 51(4):606–635, 2004.
[AK17]
Sepehr Assadi and Sanjeev Khanna. Randomized Composable Coresets for
Matching and Vertex Cover. pages 3–12, 2017.
[CCFM04] Moses Charikar, Chandra Chekuri, Tomas Feder, and Rajeev Motwani. Incremental clustering and dynamic information retrieval. SIAM Journal on Computing,
33(6):1417–1440, 2004.
[CKMN01] Moses Charikar, Samir Khuller, David M Mount, and Giri Narasimhan. Algorithms for Facility Location Problems with Outliers. Proceedings of the Twelfth
Annual Symposium on Discrete Algorithms, pages 642–651, 2001.
[CPP18]
Matteo Ceccarello, Andrea Pietracaprina, and Geppino Pucci. Fast Coreset-based
Diversity Maximization under Matroid Constraints. In Proc. WSDM, to appear,
2018.
[CPPU17] Matteo Ceccarello, Andrea Pietracaprina, Geppino Pucci, and Eli Upfal. MapReduce and Streaming Algorithms for Diversity Maximization in Metric Spaces of
Bounded Doubling Dimension. PVLDB, 10(5):469–480, 2017.
[DG08]
Jeffrey Dean and Sanjay Ghemawat. Mapreduce: simplified data processing on
large clusters. Communications of the ACM, 51(1):107–113, 2008.
[EIM11]
Alina Ene, Sungjin Im, and Benjamin Moseley. Fast Clustering Using MapReduce. In Proc. KDD, pages 681–689. ACM, 2011.
[FG88]
Tomás Feder and Daniel Greene. Optimal algorithms for approximate clustering.
In Proc. STOC, pages 434–444. ACM, 1988.
[Gon85]
Teofilo F. Gonzalez. Clustering to minimize the maximum intercluster distance.
Theoretical Computer Science, 38:293 – 306, 1985.
[HRR98]
Monika Rauch Henzinger, Prabhakar Raghavan, and Sridhar Rajagopalan. Computing on data streams. In Proc. DIMACS, pages 107–118, 1998.
[IMMM14] Piotr Indyk, Sepideh Mahabadi, Mohammad Mahdian, and Vahab S. Mirrokni.
Composable core-sets for diversity and coverage maximization. In Proc. PODS,
pages 100–108, 2014.
[MK08]
Richard M. McCutchen and Samir Khuller. Streaming algorithms for k-center
clustering with outliers and with anonymity. Lecture Notes in Computer Science,
5171 LNCS:165–178, 2008.
12
[MKC+ 15] Gustavo Malkomes, Matt J Kusner, Wenlin Chen, Kilian Q Weinberger, and Benjamin Moseley. Fast Distributed k-Center Clustering with Outliers on Massive
Data. In NIPS, pages 1063–1071. Curran Associates, Inc., 2015.
[MRL98]
Gutmeet Singh Manku, Sridhar Rajagopalan, and Bruce G. Lindsay. Approximate medians and other quantiles in one pass and with limited memory. In Proc.
SIGMOD Conference, pages 426–435, 1998.
[MW16]
Jessica McClintock and Anthony Wirth. Efficient parallel algorithms for k-center
clustering. In Proc. ICPP, pages 133–138. IEEE, 2016.
[MZ15]
Vahab Mirrokni and Morteza Zadimoghaddam. Randomized Composable Coresets for Distributed Submodular Maximization. In Proc. STOC, pages 153–162.
ACM, 2015.
[PPR+ 12]
Andrea Pietracaprina, Geppino Pucci, Matteo Riondato, Francesco Silvestri, and
Eli Upfal. Space-round tradeoffs for MapReduce computations. In Proc. ICS,
pages 235–244. ACM, 2012.
13
| 8cs.DS
|
MEMORY-ADJUSTABLE NAVIGATION PILES
WITH APPLICATIONS TO SORTING AND CONVEX HULLS∗
arXiv:1510.07185v1 [cs.DS] 24 Oct 2015
OMAR DARWISH† , AMR ELMASRY‡ , AND JYRKI KATAJAINEN§
Abstract. We consider space-bounded computations on a random-access machine (RAM) where
the input is given on a read-only random-access medium, the output is to be produced to a writeonly sequential-access medium, and the available workspace allows random reads and writes but is
of limited capacity. The length of the input is N elements, the length of the output is limited by the
computation, and the capacity of the workspace is O(S) bits for some predetermined parameter S.
We present a state-of-the-art priority queue—called an adjustable navigation pile—for this restricted
RAM model. Under some reasonable assumptions, our priority queue supports minimum and insert
in O(1) worst-case time and extract in O(N/S + lg S) worst-case time for any S ≥ lg N . (We use lg x
as a shorthand for log2 (max {2, x}).) We show how to use this data structure to sort N elements and
to compute the convex hull of N points in the two-dimensional Euclidean space in O(N 2 /S + N lg S)
worst-case time for any S ≥ lg N . Following a known lower bound for the space-time product of
any branching program for finding unique elements, both our sorting and convex-hull algorithms are
optimal. The adjustable navigation pile has turned out to be useful when designing other spaceefficient algorithms, and we expect that it will find its way to yet other applications.
1. Introduction.
1.1. Problem Area. Consider a sequential-access machine (Turing machine)
that has three tapes: input tape, output tape, and work tape. In space-bounded
computations, the input tape is read-only, the output tape is write-only, and the aim
is to limit the amount of space used in the work tape. In this set-up, the theory
of language recognition and function computation requiring O(lg N ) bits of working
space for an input of size N is well established; people talk about log-space programs
[40, Section 3.9.3] and classes of problems that can be solved in log-space [40, Section
8.5.3]. Also, in this set-up, trade-offs between space and time have been extensively
studied [40, Chapter 10]. Although one would seldom be forced to rely on a log-space
program, it is still theoretically interesting to know what can be accomplished when
only a logarithmic number of extra bits are available.
In this paper we reconsider the space-time trade-offs on a random-access machine (RAM). Analogous with the sequential-access machine, we have separate storage media for read-only input, write-only output, and read-write workspace which
is of limited capacity. Now, however, both the input and workspace allow random
access, but the output is still to be produced sequentially. Over the years, starting
by a seminal paper of Munro and Paterson [33]—where a related model was used,
the space-time trade-offs in this restricted RAM model have been studied for many
problems including: sorting [21, 37], selection [19, 21], and various geometric problems
[2, 4, 12, 13]. The practical motivation for some of the previous work has been the
appearance of special devices, where the size of working space is limited (e.g. mobile
devices) and where writing is expensive (e.g. flash memories).
An algorithm or a data structure is said to be memory adjustable if it uses O(S)
bits of working space for a given parameter S. Naturally, we expect to use at least
a constant number of words, so Ω(w) is a lower bound for the space usage, w being
the size of the machine word in bits. Sorting is one of the few problems for which
∗ Parts
of this paper have appeared in preliminary form in [3] and [11]
Institute for Informatics, Saarbrücken, Germany
‡ Department of Computer Engineering and Systems, Alexandria University, Egypt
§ Department of Computer Science, University of Copenhagen, Denmark
† Max-Planck
1
the optimal space-time product has been settled: Beame showed [5] (see also [40,
Theorem 10.13.8]) that Ω(N 2 ) is a lower bound, and Pagter and Rauhe showed [37]
that an O(N 2 /S + N lg S) worst-case running time is achievable for any S ≥ lg N .
As for the convex-hull problem, Chan and Chen [9] gave an algorithm for computing
the convex hull of a planar set of N points, stored in a read-only array, that runs in
O((N 2 /S) · lg N + N lg S) worst-case time for any S ≥ lg N .
1.2. Model of Computation. We assume that the elements being manipulated
are on a read-only array, and use N to denote the number of elements stored there.
Observe that N does not need to be known beforehand. The output is sent to a
separate write-only stream, where the output printed cannot be read or rewritten.
In addition to the input and output media, a limited random-access workspace
is available. The data on this workspace is manipulated wordwise as on the word
RAM [23]. We assume that the word size w is at least dlg N e bits and that the
processor is able to execute the same arithmetic, logical, and bitwise operations as
those supported by contemporary imperative programming languages—like C [29].
It is a routine matter [31, Section 7.1.3] to store a bit vector of size N such that it
occupies dN/we words and any string of at most w bits can be accessed in O(1) worstcase time. That is, the time complexity of an algorithm is proportional to the number
of primitive operations plus the number of element accesses and element comparisons
performed. We do not assume the availability of any powerful memory-allocation
routines. The workspace is an infinite array (of words), and the space used by an
algorithm is the prefix of this array. Even though this prefix can have some unused
zones, the length of the whole prefix specifies the space complexity of the algorithm.
In our setting, the elements lie in a read-only array and the data structure only
constitutes references to these elements. We assume that each of the elements appears
in the data structure at most once, and it is the user’s responsibility to make sure that
this is the case. Also, all operations are position-based; the position of an element can
be specified by its index. Since the positions can be used to distinguish the elements,
we implicitly assume that the elements are distinct.
1.3. Our Results. Let A be a read-only array and let N denote its size. Consider a priority queue Q storing a subset of the elements in A. We use |Q| to denote
the number of elements in Q. In the adjustable set-up, a (min-)priority queue is a
data structure that supports the following operations:
Q.minimum(): Return the index of the minimum element in Q, as long as |Q| > 0.
Q.insert(i): Insert A[i] into Q, for some i ∈ {0, 1, . . . , N − 1}.
Q.extract(i): Extract A[i] from Q, for some i ∈ {0, 1, . . . , N − 1}.
A max-priority queue, which is defined to support the operation maximum instead of
minimum, is obtained from a min-priority queue by reversing the comparison function
used in element comparisons. In the non-adjustable set-up, any priority queue—like
a binary heap [41] or a queue of pennants [7] (that both operate in-place)—could be
used to store positions of the elements instead of the elements themselves.
In the first part of the paper, we improve and simplify the memory-adjustable
priority queue presented by Pagter and Rauhe [37] by introducing a kindred data
structure that we call an adjustable navigation pile. Compared to the navigation piles
of [28], that require Θ(N ) bits, our adjustable variant can achieve the same asymptotic
run-time performance with only Θ(N/ lg N ) bits. (Another priority queue that uses
Θ(N ) bits in addition to the input was given in [16].) In Table 1, we compare the
performance of the new data structure to some of its competitors. Note that the
stated bounds are valid under some reasonable assumptions declared in Section 2.
2
Table 1
The performance of adjustable navigation piles and their competitors in the restricted RAM
model; N is the size of the read-only input and S is an asymptotic target for the size of workspace
in bits where S ≥ lg N .
Reference
Space
minimum
insert
extract
[7]
Θ(N lg N )
O(1)
O(1)
O(lg N )
[28]
Θ(N )
O(1)
O(lg N )
O(lg N )
[21]
Θ(S)
O(1)
O((N lg N )/S + lg S)
O((N lg N )/S + lg S)
[37]
Θ(S)
O(N/S 2 + lg S) O(N/S + lg S) amortized
O(N/S + lg2 S)
[this paper]
Θ(S)
O(1)
O(1)
O(N/S + lg S)
In the second part of the paper, we use the adjustable navigation pile for sorting.
Our algorithm is simpler and more intuitive than that of Pagter and Rauhe [37], and
we also achieve the optimal O(N 2 /S + N lg S) running time for any S ≥ lg N . The
algorithm is priority-queue sort like heapsort [41]: Insert the N elements one by one
into a priority queue and extract the minimum from that priority queue N times.
In the third part of the paper, we improve the Chan-Chen bound for the convexhull problem by introducing an algorithm that runs in O(N 2 /S + N lg S) time for any
S ≥ lg N . To prove this result, we augment the adjustable navigation pile with extra
information while still using O(S) bits of workspace.
1.4. Related Models. The basic feature that distinguishes the model we use
from other related models is the capability of having random access to the input
data. In the context of sequential-access machines, the input is on a tape that only
allows single-pass algorithms. The so-called streaming model still enforces sequential
access, but allows multi-pass algorithms, and the goal is to minimize the number
of passes over the input when the size of the random-access workspace is limited.
Munro and Paterson [33] considered this model. For some problems, the restricted
RAM model is more powerful than the multi-pass streaming model. For example, for
the selection problem the lower bound known for the multi-pass streaming model [8]
can be bypassed in the restricted RAM model [19].
In the in-place model, the elements are to be stored at the beginning of an infinite
array, a constant number of additional variables are allowed, and the elements may be
swapped and overwritten, but not modified, still keeping this compact representation.
All the problems considered in this paper—maintaining priority queues (see, e.g. [7,
15, 41]), sorting (see, e.g. [15, 27, 41]), and computing convex hulls [6]—have been
studied in this classical setting. In the restore model [10], the input elements may
be temporarily rearranged, and even modified, during a computation, but at the end
the input must be restored to its original state. Again, sorting problems have been
central in the exploration of the power of this model (see [10, 26]). In general, any
reversible algorithm would work well in the restore model, so reversible computing is
distantly related to this study.
1.5. Bit Vectors with rank and select Support. Given a bit vector B of N
bits, of which n are 1 bits, consider the following operations:
B.access(i): Return B[i], i.e. the bit at index i for some i ∈ {0, 1, . . . , N − 1}.
B.rank (i): Return the number of 1 bits among the bits B[0], B[1], . . . , B[i], for some
i ∈ {0, 1, . . . , N − 1}.
B.select(j): Return the index of the jth 1 bit, for some j ∈ {1, 2, . . . , n}, i.e. if
B.select(j) = i, this means that B[i] = 1 and B.rank (i) = j.
3
The operations B.rank 0(i) and B.select0(j) are similarly defined considering the 0
bits instead of the 1 bits.
For a bit vector of size N , our requirements for an acceptable solution are that
all the operations should run in O(1) worst-case time, the space used should be O(N )
bits, and the construction of the data structure should take O(N ) worst-case time.
The problem of extending a bit vector with rank -select operations has been addressed
in several papers (see, for example, [24, 32, 39]). Most of the known solutions rely
on the idea of dividing the bit vector into blocks, precomputing rank and select
values for some specific positions, and calculating the other values on the fly using
the stored values, some precomputed tables, and bits in the bit vector itself. For
example, the solution presented in [39] would be suitable for our purposes; it requires
N + O(N lg lg N/ lg N ) bits, and O(1) worst-case time per operation.
2. Memory-Adjustable Priority Queues.
2.1. Assumptions. In this section two memory-adjustable priority queues are
described. The first structure is a straightforward adaptation of a tournament tree
(also called a selection tree [30, Section 5.4.1]) for read-only data. For a parameter
S, it uses O(S) words of workspace. The second structure is an improvement of a
navigation pile [28] for which the workspace is O(S) bits, for S ≥ lg N , where N is the
size of the read-only input. Both data structures can perform minimum and insert
in O(1) worst-case time and extract in O(N/S + lg S) worst-case time.
When describing the data structures, we make the following assumptions:
1. N is known beforehand.
2. Insertions and extractions are monotonic such that, at any given point of
time, there is a single element (if any) indicating that the elements smaller
than or equal to it are outside the data structure. We call such an element
the latest output, and say that an element is alive if it is larger than the latest
output. In particular, an extraction must remove the smallest element and
an insertion must add an element that is larger than the latest output.
3. Insertions are sequential —but insertions and extractions can be intermixed—
such that the elements are inserted one by one from consecutive input entries
starting from the first element stored in the read-only input.
These assumptions are valid when a priority queue is used for sorting. Actually, in
sorting all insertions are executed before extractions; a restriction that is not mandated by the data structure. At the end of this section, we show how to get rid of
these assumptions. The first assumption is not critical. But, when relaxing the second
assumption, the required size of workspace has to increase by N bits. When relaxing
the third assumption, the asymptotic worst-case running time of insert will become
the same as that required by extract.
2.2. Adjustable Tournament Trees. For an integer S, we use S̄ as a shorthand for 2dlg Se . It suffices that S ≤ N/ lg N ; even if S was larger, the operations
would not be asymptotically faster. The input array is divided into S̄ buckets each
containing dN/S̄e elements, except for the last bucket that may contain less. A complete binary tree is built above these buckets. Each leaf of this tree covers a single
bucket and each branch covers the buckets of the leaves in the subtree rooted at that
branch. We call the elements within the buckets covered by a node the covered range
of this node. Note that the covered range of a node is a sequence of elements stored
in consecutive locations of the input array. The data stored at each node is an index
specifying the position of the smallest alive element in the covered range of that node.
4
0
latest output:
1 at position 7
0
1
0
0
1
•
5
•
•
•
•
•
•
bucket index
1
0
2
•
•
•
•
•
•
9
•
1
2
3
4
8
•
•
•
•
•
•
•
•
•
2
•
•
•
•
•
2
3
3
5
•
•
•
•
•
•
•
6
•
•
7
•
•
•
4
7
•
•
•
•
•
•
•
6
•
•
•
•
•
•
•
•
•
3
4
5
6
7
8 elements per bucket
Fig. 1. An adjustable tournament tree for a set of N = 64 elements when S̄ = 8. Indices stored
at the nodes are visualized as pointers. Only the smallest alive element in each bucket is shown.
An adjustable tournament tree is an array of 2S̄ − 1 positions (indices). To make
the connection to our adjustable navigation piles clear, we store the indices in this
array in breadth-first order as in a binary heap [41]. At each level, we start the
indexing of the nodes from 0. For the sake of simplicity, we maintain a header that
stores the offsets to the beginning of each level (even though this information is easy
to calculate). For a node number i, its left child has number 2i at the level below,
the right child has number 2i + 1 at the level below, and the parent has number bi/2c
at the level above. When we know the current level and the index of a node at that
level, the information available at the header and the above formulas are enough to
get to a neighbouring node in constant time. In Figure 1, we give an illustration of
an adjustable tournament tree for a set of N = 64 elements when S̄ = 8.
To support insert efficiently, we partition the data structure into three components: tournament tree, submersion buffer, and insertion buffer. The submersion buffer
is the last full bucket that is being submerged into the tournament tree. The insertion
buffer is the bucket that embraces the new elements. One or both of these buffers
can be empty. The idea is to insert the elements into a buffer and, when the buffer
gets full, submerge it into the tournament tree. To make minimum straightforward,
we recall the position of the overall minimum of these three components.
In Figure 2, we describe in pseudo-code how a bucket is submerged into the
tournament tree at one go. In insert, the submersion is done incrementally. The
process starts from a leaf corresponding to the given bucket and proceeds in a bottomup manner to the root by following the implicit parent pointers. We call the path
of nodes visited the updating path. First, the index at the leaf is set to point to the
minimum of the bucket. Then, every branch node on the updating path inherits the
position of the smaller of the two elements pointed to via its two children. At the end
5
procedure: submersion
input: bucket-start: index of the beginning of the bucket to be submerged
bucket-min: index of the minimum alive element within this bucket
data: N : number of elements; S̄: workspace target rounded to a power of 2
A[0 . . N − 1]: read-only array of elements
T [0 . . 2S̄ − 1]: array of indices from {none, 0, 1, . . . , N − 1}
h
header [0 . . lg S̄]: array
of
offsets, header [h] = 2 − 1 for h ∈ 0, 1, . . . , lg S̄
current ← bucket-start/ N/S̄
leaf ← header [lg S̄] + current
T [leaf ] ←
bucket-min
for ` ∈ lg S̄, lg S̄ − 1, . . . , 1 :
parent ← bcurrent/2c
this ← header [`] + current
sibling ← if current mod 2 = 0 : this + 1 else this − 1
above ← header [` − 1] + parent
if T [sibling] = none:
T [above] ← T [this]
else if T [this] = none:
T [above] ← T [sibling]
else if A[T [this]] < A[T [sibling]]:
T [above] ← T [this]
else:
T [above] ← T [sibling]
current ← parent
Fig. 2. Submersion when done at one go.
the root stores the index of the overall minimum among all alive elements. Since, at
some point, some buckets may have no alive elements, we use an unspecified constant
none to indicate that the covered range of the node in question has no alive elements.
The pseudo-code of insert is given in Figure 3. At first, the next element from
the input array becomes part of the insertion buffer. If the new element is smaller
than the buffer minimum and/or the overall minimum, the positions of these minima
are updated. Once the insertion buffer becomes full, the submersion buffer must
have been already submerged into the tournament tree. At this point, we treat the
insertion buffer as the new submersion buffer and start a new incremental submersion
process that recomputes the indices at the nodes on the updating path bottom-up,
one by one. As long as the submersion is not finished, each insert carries out a
constant amount of the submersion work. Since the work needed to update this path
is O(lg S), which is O(N/S) when S ≤ N/ lg N , the process terminates before the
insertion buffer becomes again full. Clearly, insert takes O(1) worst-case time.
In extract, there are three cases depending on whether the bucket that contains
the element to be extracted is one of the buffers or is covered by a node of the tournament tree. However, these cases are quite similar. For a pseudo-code description,
see Figure 4. To begin with, the latest output is set up to date. The bucket index
of the given element can be determined by simple calculations. Because of the monotonicity assumption the smallest alive element is to be extracted, so the bucket must
be scanned to find its new minimum. If the current bucket is the insertion buffer, it
is just enough to update the position of its minimum. If the current bucket is the
6
procedure: insert
input: i: index of an element to be inserted into T
data: N : number of elements; S̄: workspace target rounded to a power of 2
A[0 . . N − 1]: read-only array of elements
T [0 . . 2S̄ − 1]: array of indices from {none, 0, 1, . . . , N − 1}
insertion-buffer -start: index of the beginning of the insertion buffer
insertion-buffer -size: number of elements in the insertion buffer
insertion-buffer -min: index of the minimum of the insertion buffer
assert: insertion-buffer -start + insertion-buffer -size = i
insertion-buffer -size ← insertion-buffer -size + 1
if insertion-buffer -min = none or A[i] < A[insertion-buffer -min]:
insertion-buffer -min ← i
if overall -min = none or A[i] < A[overall -min]:
overall -min ← i
if insertion-buffer -size = N/S̄ :
submersion-buffer -start ← insertion-buffer -start
submersion-buffer -size ← insertion-buffer -size
submersion-buffer -min ← insertion-buffer -min
insertion-buffer -start ← i + 1
insertion-buffer -size ← 0
insertion-buffer -min ← none
execute O(1) steps of the submersion process, if one is in progress
Fig. 3. Inserting an element into an adjustable tournament tree.
submersion buffer, the submersion process is completed by recomputing the indices
at the nodes on the updating path covering the submersion buffer. Hereafter the
submersion buffer ceases to exist. If the current bucket is covered by the tournament
tree, it is necessary to recompute the indices at nodes on the updating path covering
the current bucket. At the end, the position of the overall minimum is to be updated.
It is the scanning of a bucket that makes this operation expensive: The worst-case
running time is O(N/S + lg S), which is O(N/S) when S ≤ N/ lg N .
2.3. Adjustable Navigation Piles. In brief, an adjustable navigation pile is a
compact representation of an adjustable tournament tree. The main differences are
as follows (compare Figure 1 and Figure 5):
1. A bit vector of size 2S̄ − 1 is used to indicate whether or not the buckets
covered by each node contain any alive elements. As for the data in the
tournament tree, in this bit vector the bits of the nodes are stored in breadthfirst order. This bit vector can be used to emulate the constant none.
2. Only branch nodes, i.e. nodes whose heights are larger than 0, store some
additional information about the position of the smallest alive element in the
covered range of each of these nodes. In the complete binary tree built above
the buckets, the number of branch nodes is S̄ − 1.
3. The navigation information is stored in a bit vector of size 4S̄ (an explanation
will follow shortly). To save space, at the bottom of the tree, the position of
the smallest alive element is only specified approximately. Here the details
of our construction differ from those used in the navigation piles of [28] and
their precursors [37], although the techniques used are similar.
7
procedure: extract
input: j: index of an element to be extracted from T
data: N : number of elements; S̄: workspace target rounded to a power of 2
A[0 . . N − 1]: read-only array of elements
T [0 . . 2S̄ − 1]: array of indices from {none, 0, 1, . . . , N − 1}
insertion-buffer -start: index of the beginning of the insertion buffer
insertion-buffer -min: index of the minimum of the insertion buffer
submersion-buffer -start: index of the beginning of the submersion buffer
submersion-buffer -min: index of the minimum of the submersion buffer
latest-output ← A[j]
bucket-start ← N/S̄ · j/ N/S̄
bucket-min
← none
for i ∈ bucket-start, bucket-start + 1, . . . , min bucket-start + N/S̄ − 1, N − 1
if latest-output < A[i] and (bucket-min = none or A[i] < A[bucket-min]):
bucket-min ← i
if bucket-start = insertion-buffer -start:
insertion-buffer -min ← bucket-min
else if bucket-start = submersion-buffer -start:
submersion-buffer -min ← bucket-min
submersion(submersion-buffer -start, submersion-buffer -min)
submersion-buffer -min ← none
else:
submersion(bucket-start, bucket-min)
overall -min ← none
for k ∈ {T [0], submersion-buffer -min, insertion-buffer -min}:
if k 6= none and (overall -min = none or A[k] < A[overall -min]):
overall -min ← k
Fig. 4. Extracting an element from an adjustable tournament tree.
A branch node of height h ∈ 1, 2, . . . , lg S̄ covers 2h buckets. As in a navigation
pile [28], due to scarcity of bits, the bucket index is relative within the covered range.
We use h bits to specify in which bucket the smallest alive element is. A significant
new ingredient, borrowed from [37], is the concept of a quantile. For a branch node
of height h, every covered bucket is divided into 2h quantiles, and we store additional
h bits to specify in which quantile the smallest alive element
is. We
call this quantile
the active quantile of the node. A quantile contains N/(S̄ · 2h ) elements, except
that the last quantile can possibly be smaller. Thus, we use 2h bits per node; but if
2h ≥ dlg N e, we only use dlg N e bits (since this is enough to specify the exact position
of the smallest alive element). To sum up, since there are S̄/2h nodes of height h and
since at each node we store min{2h, dlg N e} bits, the total number of navigation bits
is bounded by
lg S̄
X
S̄ · min{2h, dlg N e}
< 4S̄ .
2h
h=1
The navigation bits are stored in a bit vector in breadth-first order. As before,
we maintain a header giving the position of the first bit at each level. The space
needed by the header is O(lg2 S̄) bits. Inside each level, the navigation information
8
:
0
latest output:
1 at position 7
011 | 010
0
1
11 | 01
0
1
0|0
•
5
•
•
•
•
•
•
bucket index
0
11 | 11
2
1|0
•
•
•
•
•
•
9
•
1
3
1|0
8
•
•
•
•
•
•
•
•
•
2
•
•
•
•
•
2
3
•
•
•
•
•
•
•
1|1
•
•
•
•
•
4
7
•
•
•
•
•
•
•
6
•
•
•
•
•
•
•
•
•
3
4
5
6
7
8 elements per bucket
Fig. 5. An adjustable navigation pile corresponding to the adjustable tournament tree in
Figure 1. In this snapshot all buckets contain one or more alive elements so the bit vector—that is
not shown—indicating their existence contains just 1 bits.
is stored compactly side by side, and the nodes are numbered at each level starting
from 0. Since the length of the navigation bits is fixed for all nodes at the same level,
using the height and the index of a node, it is easy to calculate the positions where
the navigation bits of that node are stored.
Let us now consider how to access the desired quantile for a branch node. The
procedure is summarized in pseudo-code form in Figure 6. Let the branch node be
at index v within its own level, and assume that its heightis h. The first element of
the covered range of this branch node is in position v · 2h · N/S̄ . The first h bits of
the navigation information give the desired bucket
inside
the covered range; let the
relative bucket index be b, so we have to go b · N/S̄ positions forward. The second
h bits of the navigation information give the desired quantile
inside
that bucket; let
this quantile be q, so we have to proceed another q · N/(S̄ · 2h ) positions forward
before we reach the beginning of the desired quantile. Obviously, these calculations
can be carried out in constant time.
The priority-queue operations are implemented in a similar way as for an adjustable tournament tree. To facilitate constant-time minimum, we keep in memory
the index of the overall minimum (since the root of an adjustable navigation pile does
not necessarily specify a single element). In insert and extract, the subtle difference is
dealing with quantiles. When updating the navigation information for a branch node,
at the bottom of an adjustable navigation pile, we do not have direct access to the
minimum among the alive elements covered. Instead, we have to scan the elements
in the quantiles specified for the sibling nodes of the nodes along the updating path.
Later on, a quantile is said to be active if it contains the minimum among the alive
elements covered by a node. After updating the navigation bits of a node v, we locate
its parent node u and its sibling node w. The navigation bits of w are used to locate
its active quantile. This quantile is scanned, and the minimum of the alive elements
is found and compared with the minimum of the alive elements covered by v. From
9
procedure: calculate-quantile
input: v: index of a branch node at its own level; h: height of that node
data: N : number of elements; S̄: workspace target rounded to a power of 2
B[0 . . 4S̄ − 1]: array of navigation bits
header [1 . . lg S̄]: array of offsets
λ ← min{2h, dlg N e}
info-start ← header [h] + v · λ
if λ = dlg N e:
quantile-start ← B[info-start . . info-start + dlg N e − 1]
quantile-size ← 1
return (quantile-start, quantile-size)
b ← B[info-start . . info-start + h − 1]
q ← B[info-start + h . . info-start
+ 2 · h − 1]
covered -range-start ← v · 2h · N/S̄
bucket-start ← covered -range-start + b · N/S̄
quantile-start ← bucket-start + q · N/(S̄ · 2h )
quantile-past ← min quantile-start + N/(S̄ · 2h ) , bucket-start + N/S̄
quantile-size ← quantile-past − quantile-start
return (quantile-start, quantile-size)
Fig. 6. Calculating the beginning and size of the active quantile of a node.
the bucket index and the position of the smaller of the two elements, the navigation
bits of u are then calculated and accordingly updated. If the active quantile of u has
only one element, the position of this single element can be stored as such.
The
key point
is that for a node of height h the size of the active quantile is at
most N/(S̄ · 2h ) , so the total work done in the scans of the quantiles of the siblings
Plg S̄
along the updating path is proportional to h=1 N/(S̄ · 2h ) , which is O(N/S+lg S).
It follows that the asymptotic efficiency of the priority-queue operations is the same
as for an adjustable tournament tree.
2.4. Getting Rid of the Assumptions. So far we have consciously ignored
the fact that the sizes of the buckets depend on N , and that we might not know this
value beforehand. The standard way of handling this is to rely on global rebuilding
[36, Chapter V]. We use an estimate N0 and initially set N0 = 8. We build two data
structures, one for N0 and another for 2N0 . The first structure is used to perform the
priority-queue operations, but insertions and extractions are mirrored in the second
structure (if the extracted element exists there). When the structure for N0 becomes
too small, we dismiss the smaller structure in use, double N0 , and in accordance start
building a new structure of size 2N0 . We should speed up the construction of the new
structure by inserting up to two alive elements into it at a time, instead of only one.
This guarantees that the new structure will be ready for use before the first one is
dismissed. Even though global rebuilding makes the construction more complicated,
the time and space bounds remain asymptotically the same.
Since we have random-access capability to the read-only input, it is not necessary
that elements are inserted by visiting the input sequentially, but insertions should still
be monotonic. If this is the case, in connection with each insert, we have to fix the
information related to the current bucket as in extract. That is, we have to find the
smallest alive element of the bucket and, if the inserted element is smaller than the
10
procedure: priority-queue-sort
input: A[0 . . N − 1]: read-only array of N elements; S: workspace target
P ← navigation-pile(A, S)
for i ∈ {0, 1, . . . , N − 1}:
P.insert(i)
while |P | > 0:
j ← P.minimum()
P.extract(j)
print(A[j])
Fig. 7. Priority-queue sort; the position of an element is specified by its index.
current minimum at that bucket, update the navigation information on the updating
path covering that bucket. The worst-case cost of insert then becomes the same as
that of extract, i.e. O(N/S + lg S).
In some applications, insertions and extractions may not be monotonic. To handle
this situation, we have to allocate one bit per array entry (N bits in total), indicating
whether the corresponding element is alive or not.
3. Sorting.
3.1. The Pagter-Rauhe Algorithm. Let us turn our attention to sorting.
Given N elements in a read-only array, the task is to print the elements in nondecreasing order. Assume that the asymptotic workspace target is S. In the basic
setting, Pagter and Rauhe [37] proved that the running time of their sorting algorithm
is O(N 2 /S + N lg2 S) using O(S) bits of workspace. Lagging behind the optimal
bound for the space-time product by a logarithmic factor when S = ω(N/ lg2 N ), they
suggested using their memory-adjustable data structure in Frederickson’s adjustable
binary heap [21] to handle subproblems of size (N lg N )/S using O(lg N ) bits for
each. In accordance, by combining the two data structures, they achieve an optimal
O(N 2 /S) running time for sorting when lg N ≤ S ≤ N/ lg N . In our treatment we
avoid the complication of plugging two data structures together.
3.2. Priority-Queue Sort. To sort the elements, we create an empty adjustable
navigation pile, insert the elements into this pile by scanning the read-only array from
beginning to end, and then repeatedly extract the minimum of the remaining elements
from the pile. The pseudo-code in Figure 7 implements this algorithm.
3.3. Analysis. From the bounds derived for the priority queue, the asymptotic
performance can be directly deduced: The worst-case running time is O(N 2 /S +
N lg S) and the size of workspace is O(S) bits, where S ≥ lg N . It is also easy
to count the number of element comparisons performed during the execution of the
algorithm. When inserting the N elements into the data structure, O(N ) element
comparisons are performed. We can assume that after these insertions, the buffers
are submerged into the main structure. In each extract we have to find the minimum
of a single bucket which requires at most N/S̄ element comparisons. In addition, we
have to update a single path in the complete binary tree. At each level, the minimum
below the current node is already known and we have to scan the quantiles of the
sibling nodes. During the path update, we have to perform at most N/S̄ +lg S̄ element
comparisons. Hence, the total number of element comparisons performed is bounded
by 2N 2 /S̄ + N lg S̄ + O(N ), which is 2N 2 /S + N lg S + O(N ) since S ≤ S̄ ≤ 2S.
11
4. Augmenting the Adjustable Navigation Pile.
4.1. Motivation for Augmentation. In our algorithm for computing the convex hull of a set of planar points (see Section 5), insertions are neither monotonic nor
sequential and extractions are not monotonic. Nevertheless, we want to keep the algorithm memory adjustable and use O(S) bits of extra space for S < N . To limit the
size of working space, our solution is to work with a subset of the input constituting
at most S elements at a time. In more details, the input is processed in a number of
rounds, where in each round we shall have two filter elements and only elements whose
values are between these filters are to be inserted or extracted from the adjustable
navigation pile. We refer to these S elements as the candidates. Note that the candidates need not be contiguous in the read-only input array. In every round, we use a
vector of S bits, one bit per candidate, to indicate whether each of these candidates is
still alive or has been deleted. Subsequently, we need to map the indices of the input
array to indices in the range [0 . . S − 1]. To be able to do that, the adjustable navigation pile needs to be augmented with additional information. We shall refer to the
adjustable navigation pile explained in Section 2 as the unaugmented navigation pile
to distinguish it from the augmented version to be described in this section. Next, we
introduce the additional structures used in the augmentation. Then, we explain how
to update and utilize these structures in the priority-queue operations. A schematic
view illustrating the components of an augmented navigation pile is given in Figure 8.
4.2. Additional Structures. For the sake of simplicity, assume that the required workspace target S is a power of 2. We augment the adjustable navigation pile
with the following data structures:
• A bit vector alive of size S is used to denote whether each candidate is
currently alive or not. The order of the candidates in this vector is identical
to their order in the read-only array. The alive vector is dynamically updated
by insert and extract operations.
• A bit vector start of size S—corresponding to the same elements, in the same
order, as alive—is used to denote whether a candidate is the first, among other
candidates, of an active quantile or not. We refer to the candidate corresponding to the first entry for a candidate from quantile q as first-candidate(q).
(Note that a candidate may simultaneously be the first candidate of more than
one active quantile.) The start vector is also dynamic, as every insert and
extract may change it. Every bucket will possibly map to a part of alive and
start, which obviously has at most dN/Se entries. We refer to the part of
start corresponding to bucket u as start.part(u).
• A static bit vector count is used to store the number of candidates contained
in each bucket. We encode these counts in unary, using a 0 bit to mark the
border between every two consecutive buckets. Since we are dealing with at
most S candidates, the vector contains at most S ones; and since we have
exactly S buckets, it contains S − 1 zeros. The count vector should efficiently
support rank and select queries. The bit vector and the accompanying rankselect structures thus consume Θ(S) bits. The objective is to efficiently locate
the first entry of any bucket in alive and start using the index of the bucket.
Assume the first bucket has index 0. Let u > 0 be the index of the bucket
whose first entry in alive and start is to be located. Then, t = count.select0(u)
is the index of the last 0 bit preceding bucket u in count. It follows that t−u+1
is the number of candidates lying in the buckets 0, 1, . . . , u−1, which precede
the first entry of bucket u in alive and start. The size (number of bits) of
12
count:
011101001110010
candidates:
16 13
9
10 15 14 11 12
alive:
1
1
1
0
1
0
1
1
start.part(4)
start:
1
0
1
1
0
1
0
001 | 010 | 010 | 00
1
start.part(4,q)
filter:
01 | 01 | 10 | 0
00 | 10 | 01 | 0
[ 9 ... 16 ]
v
1|0|0|0
0|0|0|0
16
•
•
•
•
•
•
•
•
bucket index
0
0|1|0|0
•
•
•
10
13
•
•
•
•
•
•
•
9
•
•
•
•
•
•
•
•
•
•
•
•
•
15
14
quantile
q
11
•
•
•
•
•
•
•
•
2
3
4
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
12
•
•
1
0|1|0|0
5
6
7
Fig. 8. A snapshot of an augmented navigation pile that has N = 64 and S = 8. Only elements
within the range of the filters are shown. The snapshot is taken considering the given alive vector.
For each node, the navigation bits are displayed in left-to-right order as follows: relative bucket
index, quantile index, start.part index, and start.before. As an example, we can see that node v
refers to the candidate 11 as its minimum alive element. Hence, its relative bucket index is 0,
which refers to bucket 4. Bucket 4 is partitioned into two quantiles in the view of node v, where
the candidate 11 lies in the second quantile q. So, the quantile index for node v is 1. start.part(4)
in start is dedicated to bucket 4. At node v, start.part(4) is partitioned into two parts. Since the
candidate 14 is first-candidate(q), and as this candidate is in the first part of start.part(4), the
start.part index of node v is set to 0. The part that contains first-candidate(q) is referred to as
start.part(4, q).
start.part(u) is calculated as count.select0(u + 1) − count.select0(u) − 1.
• For every node of height h, associated with bucket u and quantile q, start.part(u)
is further divided into 2h subparts (or less, if its size is less than 2h ). Another
h bits will be stored in every node of height h to indicate in which subpart
first-candidate(q) lies. We refer to this subpart as start.part(u, q). Since the
size of start.part(u)
is at most
dN/Se, the size of start.part(u, q) is bounded
from above by N/(S · 2h ) , which is the size of a quantile.
• For every branch node of height h, associated with bucket u and quantile q, an
additional dlg he bits will be used. These bits encode the number of first candidates of other active quantiles that exist in the same subpart start.part(u, q)
before first-candidate(q), i.e. the number of ones in start.part(u, q) preceding
the one representing first-candidate(q). Let us refer to the value of this number as start.before(u, q). The reason we need only dlg he bits to store this information for each node of height h is that only quantiles tied to nodes (whose
heights are less than h) on one and only one path from a leaf node to the
node associate with q can have their first candidates before first-candidate(q)
in start.part(u, q).
It directly follows, using simple calculations, that the space complexity of the
augmented navigation pile is still Θ(S) bits.
13
To keep the time complexity for insert and extract in O(N/S + lg S), we need to
know, in an efficient way, if a given candidate that belongs to an active quantile is alive
or not. Starting with the index of a candidate in the read-only array, we want to find,
without altering the time bounds, the index of the corresponding bit in alive. Given
a node v, associated with bucket u and quantile q, the index of first-candidate(q) in
alive is to be located. Using the count vector and the navigation bits of node v, we
can easily locate start.part(u, q), where the entry we are searching for lies. We also
get the value r = start.before(u, q) from the navigation bits of node v. As previously
stated, the size of start.part(u, q) is at most the size of quantile q. While visiting
node v, we shall be scanning quantile q anyhow. A scan of start.part(u, q) would
then not alter the worst-case asymptotic time complexity. We scan start.part(u, q) to
find the rth one bit, the index of the entry at which we find this bit is the index of
first-candidate(q) in alive.
We always access the elements of a quantile sequentially. To locate the corresponding elements in alive, we start at the first entry of the quantile in alive as
explained above. While scanning the quantile, we repeatedly check the elements one
after another. If the next element is a candidate (lying in the range of the filters), we
increase the current index to the next entry in alive to correspond to this candidate.
The same procedure can be applied on buckets. We get the first entry of the bucket in
alive using the count vector, and then move sequentially on the bucket and on alive,
increasing the alive index whenever we encounter the next candidate in the bucket.
4.3. Operations. We next explain how insert and extract can be performed in
our augmented navigation pile in O(N/S + lg S) worst-case time per operation.
We first find the bucket in which the element to be inserted or extracted lies; this
can be done with simple calculations in constant time once we have the array index
of the element. Let the index of this bucket be u. Using the count vector, we get
the first entry of this bucket in alive as explained earlier, and move sequentially on
bucket u and alive. We can then get the indices of the candidates lying in this bucket
within alive, and consequently know whether each of these candidates is currently
alive or not. After knowing the index of the element to be inserted or extracted in
alive, we should set the corresponding bit to 1 or 0 respectively. Also, while scanning
the bucket, we would know if this element is the minimum in the bucket or not. If
the minimum alive element in bucket u has changed due to the current operation, the
following updates need to be done.
As in an unaugmented navigation pile, the information in the nodes along the
updating path is to be fixed bottom up. The update will work as in the unaugmented
navigation pile, where we scan the quantiles associated with the nodes lying on or
hanging from the updating path. However, here we also want to know whether each of
the candidates in these quantiles is alive or not. Before accessing a quantile, we get its
first entry in alive; this can be done as previously described. Then, we simultaneously
scan both the quantile and the corresponding subpart in alive.
Next, we show how to update the start vector. Suppose that we are to handle a
node v, associated with quantile q, on the updating path, knowing that its child has
just been handled. If the index of first-candidate(q) is different from the index of the
first candidate of each of the two quantiles associated with the two children of node v,
we reset the bit for first-candidate(q) in start to 0. Note that we do not reset that bit
to 0 if it is the first entry of the quantile of a child of node v in start. Alternatively,
this bit may be temporarily reset to 0 in the previous step and then again set to 1
within the upcoming step. Assume that the child of node v that is associated with
14
the quantile that has the smaller element of the two children of v associated with
quantile q 0 . Now the quantile associated with node v should be either the first half or
the second half of quantile q 0 . If it is the first half, then the corresponding entry in
start must have been already set to 1 before. Else, we move sequentially on quantile
q 0 and start till we reach the first candidate in the second half of quantile q 0 , and set
its corresponding entry in start to 1. The above procedure is repeated for every node
on the updating path.
For the nodes along the updating path, we show next how the additional bits in
our augmented navigation pile will be updated. Consider a node v that is associated
with bucket u and quantile q after the update. While handling node v, as explained
earlier, we are able to know the entry in start for first-candidate(q) as well as the
size of start.part(u, q). Knowing these values, it is easy to update the bits indicating
start.part(u, q) in node v. Also, after getting these bits, we loop on start.part(u, q) to
count the number of 1 bits in this subpart preceding the entry for first-candidate(q)
in start, and store this value for node v in start.before(u, q).
It is obvious that the update will be performed on at most lg S nodes on the
updating path. Also, looping on the quantiles and the corresponding
parts of start
Plg S
for the nodes on the updating path would sum up to O( i=1 dN/(S · 2i )e). So the
time complexity for the update is in O(N/S + lg S), as claimed.
The following lemma summarizes the functionality of our data structure.
Lemma 1. Assume that we are given a read-only array of N elements, and two
filters that enclose at most S of these elements between their values. Using Θ(S)
bits of workspace, after spending O(N ) worst-case time on building the augmented
structure, insert and extract can be applied to any element of the array whose value
is between the filters in O(N/S + lg S) worst-case time per operation.
5. Convex Hulls.
5.1. The Chan-Chen Algorithm. Consider now the problem of computing the
convex hull of a set of N points in the plane given in a read-only array P . Without
loss of generality, we can assume that the points are unique such that no two points
have the same x-coordinate or y-coordinate. The task is to print the points on the
convex hull in the clockwise order of their appearance on the hull, starting from some
arbitrary point. As is standard, it is enough to show how to compute the upper hull
of the point set; the lower hull can be computed in a symmetric manner. We assume
the availability of the standard geometric primitive that tells whether or not there
is a right turn on point P [j] when going from point P [i] to point P [k] via P [j]; we
denote this predicate as right-turn(P [i], P [j], P [k]).
A high-level description of the algorithm by Chan and Chen [9] is given in Figure 9.
Our algorithm is similar but the details are different. When the working space of Θ(S)
bits is available, they set the space parameter s to S/ lg N and use Θ(s) indices to
recall the points being processed. The algorithm performs dN/se rounds and handles
the s points with the next smallest x-coordinates in each round; these points form a
vertical slab σ. In each round, the algorithm starts with a known hull vertex P [i0 ]
and computes the part of the upper hull for the points of σ starting from point P [i0 ]
and ending at the left endpoint of the hull edge crossing the right wall of σ.
When finding the s points with the next smallest x-coordinates among the remaining points, the algorithm uses space for 2s indices and maintains in the first half
the s indices of the points with the smallest x-coordinates among the points examined so far. Each time, the second half is refilled with another s indices from the
unexamined portion, the median of the x-coordinates of the 2s recorded points is
15
procedure: compute-convex -hull
input: P [0 . . N − 1]: read-only array of N points; s: workspace target
i0 ← none
for i ∈ {0, 1, . . . , N − 1}:
if i0 = none or x-coordinate(P [i]) < x-coordinate(P [i0 ]):
i0 ← i
while i0 6= none:
let σ be the vertical slab containing P [i0 ] and the next s − 1 points (if possible)
on the right of and closest to the wall determined by P [i0 ]
let {i0 , i1 , . . . , is−1 } be the indices of the s points in σ, sorted by x-coordinate
use Graham’s scan to compute the upper hull of these points
let {h0 , h1 , h2 , . . . , hs0 } be the indices of the points on this hull // h0 = i0
j 0 ← none // right endpoint of the hull edge crossing the right wall of σ
for each point P [j] to the right of the wall determined by P [is−1 ]:
if j 0 6= none and right-turn(P [hs0 ], P [j 0 ], P [j]):
continue
while s0 > 0 and not right-turn(P [hs0 −1 ], P [hs0 ], P [j]):
s0 ← s0 − 1
0
j ←j
for k ∈ {h0 , h1 , . . . , hs0 }:
print(P [k])
i0 ← j 0
Fig. 9. High-level description of the Chan-Chen algorithm.
found, and these points are partitioned around this median. This is repeated until
all the points are examined, leaving the s points within the slab σ. The upper hull
of these points can then be constructed using any of the known O(s lg s) convex-hull
algorithms [38, Chapter 3]; a natural choice is to use Graham’s scan since the rest
of the algorithm follows the same elimination strategy. The Chan-Chen algorithm
eliminates the points that are not on the convex hull by traversing the tentative hull
chain in reverse order, the points with the larger x-coordinates first. This procedure
is done through finding the hull edge crossing the right wall of σ by performing a pass
over the remaining points (those to the right of σ). Suppose a point P [j] from the
remaining points is currently being inspected, by imitating Graham’s scan, the point
P [j] is tentatively added to the hull if it is above the tentative hull edge found so far
crossing the right wall of σ. Also, adding P [j] to the hull might require removing some
points from the chain in Graham-scan fashion. In the description given in Figure 9,
we do not add any new points to the tentative hull chain. We just keep P [j 0 ] as the
point representing the right endpoint of the hull edge crossing the right wall of σ and
we update j 0 accordingly. Hence, points are only removed from the hull chain during
the elimination process. At the end of the pass, the leftover points on the chain are
indeed on the convex hull and are accordingly reported.
Finding the next s points with the smallest x-coordinates requires O(N ) time,
computing the upper hull for s points requires O(s lg s) time, and pruning the tentative
hull chain from points not on the upper hull requires O(N ) time. Since there are
O(N/s) rounds, the algorithm runs in O(N/s · (N + s lg s)) worst-case time, which is
O((N 2 /S) · lg N + N lg S).
16
5.2. Our Algorithm. Our convex-hull algorithm uses three navigation piles: a
(max-)augmented navigation pile that we call max -pile, and two (min-)unaugmented
navigation piles that we call min-pile 1 and min-pile 2 . Without loss of generality, we
assume that the orientation (either max or min) of the navigation piles is with respect
to the x-coordinates of the points. The algorithm works as follows:
1. Insert all the N points in both min-pile 1 and min-pile 2 .
2. Let i0 be the index of the point with the minimum x-coordinate, and none
if the input set is empty.
3. while i0 6= none:
(a) Extract the minimum S points, one by one, from min-pile 1 (or until
|min-pile 1 | = 0), and keep track of the first and last points. These
two values will be used as filters for max -pile; let us call them f1 and
f2 . These S points will be the candidates considered in this round,
determining the slab σ.
(b) Reinitialize an empty max -pile using the current candidates and filters,
without actually inserting the points. Reinitialize the count vector by
scanning the N points, bucket by bucket. Then, build the rank and
select structures for the count vector.
(c) Construct the upper hull for the current S points. This can be done by
using a space-efficient implementation of Graham’s scan. For example,
the in-place variant described in [6] could be modified to use a bit vector of alive elements, instead of swapping the input elements. In this
computation, max -pile and min-pile 2 can be deployed as follows.
i. Extract the two points with the minimum x-coordinates from min-pile 2
and insert both of them into max -pile.
ii. repeat S − 2 times (or until |min-pile 2 | = 0):
A. Extract the minimum point from min-pile 2 ; let its index be i0 .
B. while |max -pile| ≥ 2 and not right-turn(next to maximum of
max -pile, maximum of max -pile, P [i0 ]):
repeatedly extract the maximum point from max -pile.
C. Insert the point P [i0 ] into max -pile.
At this point, the alive points in max -pile form a tentative upper hull.
(d) The goal here is to eliminate the points that are not really on the hull
among the points in max -pile forming the tentative hull chain. We do
so by computing the hull edge crossing the right wall of σ.
i. Set j 0 to none, which represents the right endpoint of the hull edge
crossing the right wall of σ.
ii. for every point with index j whose x-coordinate is greater than f2 :
A. If j 0 6= none and right-turn(maximum of max -pile, P [j 0 ], P [j]),
continue.
B. while |max -pile| ≥ 2 and not right-turn(next to maximum of
max -pile, maximum of max -pile, P [j]):
repeatedly extract the maximum point from max -pile.
C. Set j 0 to j.
(e) Set i0 to j 0 , then extract and neglect points from min-pile 1 and min-pile 2
until we reach P [i0 ] as the current minimum for both.
(f) Extract all alive points from max -pile and report them as points on the
convex hull, but in reverse order. This reversal can be done by inserting
the points in an empty min-navigation pile, then extracting them.
17
By inspecting our algorithm and the algorithm of Chan and Chen, at high level,
the algorithms are quite similar. The main difference is that we use adjustable navigation piles in order to recall the points and deal with them. It should also be noted
that the full power of our techniques (augmenting the structure) is only needed in
step (3c). All other parts could have been handled efficiently without augmenting the
adjustable navigation pile.
5.3. Analysis. We need to prove that our convex-hull algorithm achieves a time
complexity of O(N 2 /S + N lg S) and a space complexity of Θ(S) bits. As the space
complexity, we deploy three navigation piles that are proved to require Θ(S) bits.
As for the time complexity, step (1) requires O(N ) time to build two unaugmented
navigation piles. Obviously, step (2) can be done in O(1) time. Now, we are going to
analyse step (3). There are dN/Se rounds in total. The work done in each round can
be summed up as follows:
• In step (3a), extracting S points requires O(N + S lg S) time.
• Step (3b) needs O(N ) time to construct the count vector and the data structures for answering rank and select queries.
• Constructing the upper hull of S points in step (3c) is done in O(N + S lg S)
time. Note that S points are extracted from min-pile 2 . Each of these points
will be inserted and may be extracted later from max -pile, but a point can
be inserted once and extracted once from max -pile, for a total of at most S
insertions and S extractions.
• The time complexity of step (3d) is O(N + S lg S) too. We need O(N ) to
loop on the whole input sequence. Also, we may be extracting points from
max -pile. Given that the number of alive points in max -pile is at most S,
then the extractions need at most O(N + S lg S) time.
• Step (3e) does not affect the time complexity of the algorithm. Here, we
extract points from min-pile 1 and min-pile 2 . Since these piles initially contain
the N points and no more insertions are done into them, throughout the
algorithm all extractions from these piles require O(N 2 /S + N lg S) time.
• Given that the number of points in max -pile is at most S, step (3f) (including
the reversal of the order of the points) will be done in O(N + S lg S) time.
As a conclusion, the total cost of step (3e) is O(N 2 /S + N lg S). Except for
step (3e), the worst-case time complexity of step (3) is O(N + S lg S) per round.
Multiplying this by the number of rounds dN/Se, the time complexity of our convexhull algorithm is O(N 2 /S + N lg S), as claimed.
6. Concluding Remarks.
6.1. Summary. When constructing adjustable navigation piles, four techniques
are important: 1) implicit links when indexing different types of objects, 2) bit packing
and unpacking, 3) buffering and incremental submersion, and 4) quantile thinning. In
addition to their connection to binary heaps [41], we pointed out the strong connection
between the navigation piles and tournament trees. Conclusively, our data structure is
shown to be a useful ingredient in space-efficient algorithms for problems that employ
incremental sorting within their engine. In general, we illustrated some conditions for
when a succinct implementation of a priority queue that uses a workspace constituting
a sublinear number of bits is possible, so that algorithm designers would be careful
when using the structure.
On the negative side, in practice, navigation piles are slow [25] for two reasons:
1) The bit-manipulation machinery is heavy and index calculations devour clock
18
cycles. 2) The cache behaviour is poor because the memory accesses lack locality.
It is expected that the situation is not better for adjustable navigation piles.
Our sorting is a heapsort algorithm [41] that uses an adjustable navigation pile
instead of a binary heap. For our convex-hull algorithm we had to augment the data
structure with extra information while still maintaining the same asymptotic memory
usage. In spite of the optimality of the asymptotic running time of our algorithms, one
could criticize the practicality of the model itself, since the memory-access patterns
may not always be friendly to contemporary computers while it is not allowed to move
the elements around.
6.2. Related Developments. The credit for the use of quantile thinning should
go to Pagter and Rauhe [37]. However, the way the technique is used in the adjustable
navigation pile leads to a simpler and more elegant data structure. Recently, quantile
thinning has also been used in a data structure to answer heavy-hitter queries for a
set of points on a line [18].
Buffering and incremental submersion is a general data-structural transformation
that can be used to speed up insert in priority queues [1]. Recently, this technique
has also been used in a space-efficient manner in weak heaps [14] and in strengthened
lazy heaps [15].
After introducing the adjustable navigation piles in the conference version of this
paper, our data structure has also found application in space-efficient graph algorithms
[17] and space-efficient plane-sweep geometric algorithms [20].
6.3. Other Data Structures for Read-Only Data. In our experience, very
few data structures can be made memory adjustable as elegantly as priority queues.
A stack is a gratifying companion [4]. As counterweight, a dictionary must maintain
a permutation of a set of size N ; this means that it is difficult to manage with much
less than N dlg N e bits. However, when the goal is to cope with about N bits, a bit
vector extended with rank and select facilities (for a survey, see [35]) is a relevant
data structure. Two related constructions are the wavelet stack used in [19] and the
wavelet tree introduced in [22] (for a survey, see [34]).
6.4. Open Problems. The optimality of our algorithm for computing convex
hulls follows from the fact that the sorting problem reduces to the problem of computing the convex hull of a planar point set [38, Section 3.2]. However, in the restricted
RAM model, for the following variants of the problem, the exact space-time trade-offs
are still unknown:
1. Compute the convex hull of a planar set of points given in lexicographic sorted
order according to their coordinates.
2. Compute the convex hull of a simple polygon.
3. Compute the extreme points of a planar set of N distinct points, i.e. the
points on the convex hull in any order, and express the complexity as the
function of N and h, where h denotes the size of the output.
For these problems, the space-time lower bound obtained via the unique-elements
problem [40, Section 10.13.7] does not hold any more. For the best known results, we
refer to [4, 9] (just be aware that in these papers the space bounds are expressed in
words, not in bits).
Acknowledgements. We thank Tetsuo Asano for introducing the restricted
RAM model to us and taking part in the initial discussions on the topic that led
to the post at a conference [3].
19
REFERENCES
[1] S. Alstrup, T. Husfeldt, T. Rauhe, and M. Thorup, Black box for constant-time insertion in
priority queues, ACM Trans. Algorithms 1, 1 (2005), 102–106.
[2] T. Asano, K. Buchin, M. Buchin, M. Korman, W. Mulzer, G. Rote, and A. Schulz, Memoryconstrained algorithms for simple polygons, E-print arXiv:1112.5904, arXiv.org, Ithaca
(2011).
[3] T. Asano, A. Elmasry, and J. Katajainen, Priority queues and sorting for read-only data,
TAMC 2013, LNCS 7876, Springer, Heidelberg (2013), 32–41.
[4] L. Barba, M. Korman, S. Langerman, K. Sadakane, and R. I. Silveira, Space–time trade-offs
for stack-based algorithms, Algorithmica 72, 4 (2015), 1097–1129.
[5] P. Beame, A general sequential time-space tradeoff for finding unique elements, SIAM J. Comput. 20, 2 (1991), 270–277.
[6] H. Brönnimann, J. Iacono, J. Katajainen, P. Morin, J. Morrison, and G. Toussaint, Spaceefficient planar convex hull algorithms, Theoret. Comput. Sci. 321, 1 (2004), 25–40.
[7] S. Carlsson, J. I. Munro, and P. V. Poblete, An implicit binomial queue with constant insertion
time, SWAT 1988, LNCS 318, Springer, Heidelberg (1988), 1–13.
[8] T. M. Chan, Comparison-based time-space lower bounds for selection, ACM Trans. Algorithms
6, 2 (2010), 26:1–26:16.
[9] T. M. Chan and E. Y. Chen, Multi-pass geometric algorithms, Discrete Comput. Geom. 37, 1
(2007), 79–102.
[10] T. M. Chan, J. I. Munro, and V. Raman, Selection and sorting in the “restore” model, SODA
2014, SIAM, Philadelphia (2014), 995–1004.
[11] O. Darwish and A. Elmasry, Optimal time-space tradeoff for the 2D convex-hull problem, ESA
2014, LNCS 8737, Springer, Heidelberg (2014), 284–295.
[12] M. De, S. C. Nandy, and S. Roy, Convex hull and linear programming in read-only setup with
limited work-space, E-print arXiv:1212.5353, arXiv.org, Ithaca (2012).
[13] M. De, S. C. Nandy, and S. Roy, Prune-and-search with limited work-space, J. Comput. System
Sci. 81, 2 (2015), 398–414.
[14] S. Edelkamp, A. Elmasry, and J. Katajainen, Weak heaps engineered, J. Discrete Algorithms
23 (2013), 83–97.
[15] S. Edelkamp, A. Elmasry, and J. Katajainen, An in-place priority queue with O(1) time for
push and lg n + O(1) comparisons for pop, CSR 2015, LNCS 9139, Springer, Heidelberg
(2015), 1–15.
[16] A. Elmasry, Three sorting algorithms using priority queues, ISAAC 2003, LNCS 2906,
Springer, Heidelberg (2003), 209–220.
[17] A. Elmasry, T. Hagerup, and F. Kammer, Space-efficient basic graph algorithms, STACS 2015,
LIPIcs 30, Schloss Dagstuhl—Leibniz Center for Informatics, Dagstuhl (2015), 288–301.
[18] A. Elmasry, M. He, J. I. Munro, and P. K. Nicholson, Dynamic range majority data structures,
ISAAC 2011, LNCS 7074, Springer, Heidelberg (2011), 150–159.
[19] A. Elmasry, D. D. Juhl, J. Katajainen, and S. R. Satti, Selection from read-only memory with
limited workspace, Theoret. Comput. Sci. 554 (2014), 64–73.
[20] A. Elmasry and F. Kammer, Space-efficient plane-sweep algorithms, E-print arXiv:1507.
01767, arXiv.org, Ithaca (2015).
[21] G. N. Frederickson, Upper bounds for time-space trade-offs in sorting and selection, J. Comput.
System Sci. 34, 1 (1987), 19–26.
[22] R. Grossi, A. Gupta, and J. S. Vitter, High-order entropy-compressed text indexes, SODA
2003, SIAM, Philadelphia (2003), 841–850.
[23] T. Hagerup, Sorting and searching on the word RAM, STACS 1998, LNCS 1373, Springer,
Heidelberg (1998), 366–398.
[24] G. Jacobson, Space-efficient static trees and graphs, FOCS 1989, IEEE Computer Society, Los
Alamitos (1989), 549–554.
[25] C. Jensen and J. Katajainen, An experimental evaluation of navigation piles, CPH STL Report
2006-3, Dept. Comput. Sci., Univ. Copenhagen, Copenhagen (2006).
[26] J. Katajainen and T. Pasanen, Sorting multisets stably in minimum space, Acta Inform. 31, 4
(1994), 301–313.
[27] J. Katajainen, T. Pasanen, and J. Teuhola, Practical in-place mergesort, Nordic J. Comput. 3,
1 (1996), 27–40.
[28] J. Katajainen and F. Vitale, Navigation piles with applications to sorting, priority queues, and
priority deques, Nordic J. Comput. 10, 3 (2003), 238–262.
[29] B. W. Kernighan and D. M. Ritchie, The C Programming Language, 2nd Edition, Prentice
Hall, Englewood Cliffs (1988).
20
[30] D. E. Knuth, Sorting and Searching, The Art of Computer Programming 3, 2nd Edition,
Addison-Wesley, Reading (1998).
[31] D. E. Knuth, Combinatorial Algorithms: Part 1, The Art of Computer Programming 4A,
Addison-Wesley, Boston (2011).
[32] J. I. Munro, Tables, FSTTCS 1996, LNCS 1180, Springer, Heidelberg (1996), 37–42.
[33] J. I. Munro and M. S. Paterson, Selection and sorting with limited storage, Theoret. Comput.
Sci. 12, 3 (1980), 315–323.
[34] G. Navarro, Wavelet trees for all, CPM 2012, LNCS 7354, Springer, Heidelberg (2012), 2–26.
[35] G. Navarro and E. Providel, Fast, small, simple rank/select on bitmaps, SEA 2012, LNCS
7276, Springer, Heidelberg (2012), 295–306.
[36] M. H. Overmars, The Design of Dynamic Data Structures, LNCS 156, Springer, Heidelberg
(1983).
[37] J. Pagter and T. Rauhe, Optimal time-space trade-offs for sorting, FOCS 1998, IEEE Computer
Society, Los Alamitos (1998), 264–268.
[38] F. P. Preparata and M. Shamos, Computational Geometry: An introduction, Springer, Heidelberg (1985).
[39] R. Raman, V. Raman, and S. R. Satti, Succinct indexable dictionaries with applications to
encoding k-ary trees, prefix sums and multisets, ACM Trans. Algorithms 3, 4 (2007),
43:1–43:25.
[40] J. E. Savage, Models of Computation: Exploring the Power of Computing (2008). The book
is released in electronic form under a CC-BY-NC-ND-3.0-US license and is available at
http://cs.brown.edu/~jes/book/home.html
[41] J. W. J. Williams, Algorithm 232: Heapsort, Commun. ACM 7, 6 (1964), 347–348.
21
| 8cs.DS
|
Incorporating Depth into both CNN and CRF for Indoor Semantic
Segmentation
Jindong Jiang, Zhijun Zhang∗, Member, IEEE, Yongqian Huang, Lunan Zheng
arXiv:1705.07383v3 [cs.CV] 4 Oct 2017
The School of Automation Science and Engineering, South China University of Technology
Guangzhou 510640, China
E-mail: [email protected], [email protected], [email protected], [email protected]
Abstract:
To improve segmentation performance, a novel neural network architecture (termed DFCN-DCRF) is proposed, which combines
an RGB-D fully convolutional neural network (DFCN) with a depth-sensitive fully-connected conditional random field (DCRF).
First, a DFCN architecture which fuses depth information into the early layers and applies dilated convolution for later contextual
reasoning is designed. Then, a depth-sensitive fully-connected conditional random field (DCRF) is proposed and combined with the
previous DFCN to refine the preliminary result. Comparative experiments show that the proposed DFCN-DCRF achieves competitive
performance compared with state-of-the-art methods.
Key Words: Convolutional neural networks, conditional random fields, RGB-D, semantic segmentation, transfer learning.
1
Introduction
In order to realize scene understanding, semantic segmentation plays a very important role and has attracted more and
more researchers interests [1–5]. Among existing methods,
convolutional neural networks (CNNs) have shown great advantages on semantic segmentation with RGB images. One
typical CNN, called fully convolutional neural network (FCN),
achieves remarkable performance over the past few years. As
reported in Ref. [5–7], encoder-decoder type FCNs dramatically improved the dense prediction accuracy by fusing different layer representations. In order to expand the receptive field
without losing resolution and generate a better performance
on multiple segmentation tasks, a dilated convolution operator
was applied to replace the encoder-decoder architecture [8, 9].
Despite many efforts had been taken on the improvement, the
result was still unsatisfactory, especially, on the boundary of
the objects. To remedy this problem, researchers started to
combine RGB model based fully-connected conditional random fields (CRFs) with CNN and gained improvements on
several semantic segmentation benchmarks [8–13]. However,
it is difficult to apply these methods in the indoor scene where
objects share similar colors.
Recently, some RGB-D image datasets [14–17] have been
released in public. Since the depth information includes 3D
positions and structures of the objects, utilizing depth channel
as complementary information to RGB channel may increase
the potential to implement accurate semantic segmentation.
This hypothesis is verified by Couprie et al. who interacted
the depth information into a multiscale convolutional network
[18]. Inspired by this work, a novel neural network (DFCN)
architecture with a depth-sensitive fully-connected conditional
random field (DCRF) is proposed in this paper. Different from
the existing FCNs, we incorporate the depth information into a
FCN with dilated operator and a CRF to improve the accuracy
greatly.
Before ending this section, the main contributions of this
work are listed as follow:
1. A novel neural network architecture (termed DFCNDCRF) is proposed, which combines an RGB-D fully
convolutional neural network (DFCN) with a depthsensitive fully-connected conditional random field
(DCRF). To the best of the authors’ knowledge, it is the
first time to incorporate the depth information into both a
FCN with dilated operator and a CRF.
2. The design process and theoretical analysis of the proposed DFCN-DCRF is presented in detail.
3. Two comparison experiments on SUN RGB-D benchmark verify the effectiveness of the proposed DFCNDCRF on semantic segmentation.
2
Related Work
In this section, the literature of deep CNN for semantic segmentation, fully-connected conditional random fields, and incorporation of depth information have been previewed in detail.
2.1 Deep Convolutional Neural Network for Semantic
Segmentation
In 2015, Long et al. proposed a fully convolutional neural
network model [5], which had a structure of encoder-decoder
architecture. In this work, a skip architecture was designed,
which combined semantic information from the deep coarse
layer with appearance information from a shallow fine layer.
The skip architecture is able to take advantage of all feature
spectra and showed an accurate segmentation result. As a further discussion, Noh et al. [6] proposed a novel FCN structure
which eliminates the limitation of fixed-size receptive field.
On the decoding step, it applied unpooling and convolution
transpose to allow the network to learn the upsample weights.
With the similar network structure of these two models, Badrinarayanan et al. [7] presented another architecture called SegNet, which comprised unpooling as well as the skip architecture. Besides, dropout [19] and batch normalization [20]
(a) normal convolution
(b) dilated convolution
Fig. 1: Receptive field of normal convolution and dilated convolution. Left to Right: the grids (marked in blue) contributes
to the calculation of the center grids (marked in red) through
three convolution layers with a kernel size of 3 × 3. (a) Receptive field of normal convolution through three layers. (b)
Receptive field of convolution layers with 1, 2, and 4 dilation
rate through three layers.
can also further improve the segmentation accuracy during test
time [7, 21].
Meanwhile, another approach for contextual reasoning
called dilated convolution was proposed. It aims to expands
the receptive field on the input image exponentially without
losing resolution of the result. As shown in Fig. 1, dilation
on convolution expands the receptive field on the input image
exponentially without losing resolution of the result. Based on
this architecture, Chen et al. proposed DeepLab system which
encoded objects as well as image context at multiple scales [9].
Yu et al. further applied dilated convolution on a novel FCN
architecture with no pooling layer [8]. In our work, we design
the core network with both pooling and dilated convolution
(discussed in Section 3.1).
2.2 Fully-Connected Conditional Random Fields
Recently, some semantic segmentation algorithms based
on CNN combine the FCN with conditional random fields
(CRFs). CRFs is able to model the contextual relationships
between different pixel so as to maximize the label agreement
of them. Koltun et al presented an efficient inference algorithm
for Gaussian Edge Potentials [4]. The inference method allows
a fully-connected CRF with pairwise connection over all pairs
of pixels to inference in a reasonable time. It has been proved
that the poor accuracy of boundary in the output of FCN can
be addressed by combining the responses in the last layer of
CNN with a fully-connected CRF model [8–10]. In particular,
Zheng et al. proposed a novel architecture, in which the mean
field approximation was modeled as a recurrent neural network
and integrated as a part of deep neural network [12].
Fully-connected CRFs with RGB information works well on
refining CNN output, based on the fact that different objects
have different colors or brightness. However, indoor scene
objects (e.g., bed, couch, pillow) often share similar color or
brightness. Therefore, it is reasonable to incorporate the depth
information into fully-connected CRF as a post-processing
method to provide additional information such as distance or
clear distinctive boundaries. The idea of incorporating depth
information into conditional random fields was first proposed
by Muller et al. [22]. They applied a super-pixel-based model
for semantic segmentation. Inspired by Ref. [22], we incorporate the depth information into a fully-connected CRF after
CNN architecture to generate a more accurate segmentation.
2.3 Incorporation of Depth Information
Based on some labeled RGB-D image datasets [14–17],
many studies have tried to incorporate the depth information
to generate a better performance. Long et al. [5] shows that
simply stack the depth with RGB as a 4-channels input cannot improve the performance in a significant way. Gupta et
al. [23] presented a different representation of depth information referred as HHA. It comprised horizontal disparity, height
from the ground and the angle between the local surface normal and gravity direction, and got a good results [5, 23]. However, Hazirbas et al. [24]. argued that the HHA representation did not contain more information than the raw depth itself, and it required a high computation cost. In their work,
a fusion-based CNN architecture was presented. The network
consisted of two branches of encoding networks, i.e., a depth
branch and an RGB branch. The feature representation in these
two branches was then fused into the master branch. There are
two ways of fusing approaches, i.e., sparse fusion and dense
fusion. It was proved that the sparse one is better. Therefore,
in our work, we fuse RGB and depth channel feature representation in a sparse way from Conv1 to Conv4.
3
Approach
In this section, the design process and theoretical analysis of
the proposed DFCN-DCRF is stated in detail.
3.1 RGB-D FCN for Unary Potential Generation
We propose a RGB-D FCN architecture (DFCN) to generate the response for the unary potential for each class on each
pixel. As shown in Fig. 2, the DFCN part has two major
blocks: 1) Convolution layers with three downsample pooling
for features extraction and depth fusion; 2) Dilated convolution
layer for contextual reasoning and dense prediction.
In the first block, we employ the 16-layer VGG net from first
layer Conv1 1 to Conv4 3 as a fundamental framework. This
fundamental framework is applied on both the RGB channel
and depth channel (i.e., master branch and depth branch) for
features extraction. In this stage, we take layers before every pooling in two branches of the network and fuse them together with element-wise summation. The fusion layers are
then added before every pooling layer in the master branch.
To prevent further resolution decrease of the feature maps, we
replace Pool4 in VGG-16 with a one-strike max-pooling and
Pool5 with a one-strike max-pooling and a one-strike averagepooling. To make the values in two branches compatible and
easier to train, we normalize the depth channel, which originally ranges from 0 to 65535, into the same range as color
RGB
Image
Output
(DFCN)
Conv1
Conv2
Conv3
Conv4
+
+
+
Conv5
Conv6
+
Unary
Master
Depth
Image
+
+
Pairwise
Depth
Dense-sensitive CRF (DCRF)
Dense-sensitive FCN (DFCN)
Conv+ReLU
Fusion
Pooling
Atrous+BN
Dropout
Score
Output
DFCN-DCRF)
Socre+CRF
Fig. 2: The proposed DFCN-DCRF architecture.
images, i.e., from 0 to 255.
In the second block, the dilated convolution is applied after
Pool4 with dilation rate of 2 for three layers in Conv5 and dilation rate of 12 for Conv6 1. All dilated convnets in the proposed DFCN-DCRF architecture are followed by batch normalization layers to avoid covariate shift [20]. Conv6 1 and
Conv6 2 are also followed by dropout layers when training to
avoid overfitting [19]. The final score map is upsampled with
a factor of 8 by using bilinear interpolation to recover the original resolution. This score map is also converted into a preliminary pixel-wise label prediction, that is, the result of DCRF.
3.2 Depth-Sensitive Fully-Connected Conditional Random Fields
We present a depth-sensitive fully-connected CRF (DCRF)
to refine the upsampled output from DFCN. Every pixel is
treated as a CRF node, and the energy function of the DCRF
is composed of a unary and a pairwise factors (also called first
and second order factors). Considering an image I has a size
of N , the energy function E(y) with y denoting label vector
is defined as
X
X
φij (yi , yj )
(1)
φi (yi ) +
E(y) =
i
ij
where y = [y1 , y2 , · · · , yi , · · · , yN ]T with i ∈ [1, N ], superscript T denoting the transpose of a matrix or a vector. Element yi is the label assigned of the ith pixel. The unary potential φi (yi ) = − log P (yi ) is computed from the last layer of
DFCN, where P (yi ) is the result of applying softmax on the
score map at pixel i. φij (yi , yj ) is a pairwise potential function
with Gaussian kernel over all pair of pixels in image I, which
can be represented as
φij (yi , yj ) = µ(yi , yj ) ω1 θa (fi , fj ) + ω2 θs (fi , fj )
(2)
where µ is the label compatibility function. In our model,
µ(yi , yj ) = [yi 6= yj ]. In Potts model [25], it means that we
have penalty for the assignment of different labels. fi and fj are
feature vectors of pixels in the ith and jth positions. θs (fi , fj )
is smoothness kernel, i.e.,
k pi − pj k 2
θs (fi , fj ) = exp −
(3)
2σγ2
where pi and pj denote the position vectors of the ith and jth
pixels. Parameter σγ controls the degrees of nearness of two
pixels. The smoothness kernel is used to eliminate small isolated regions. θa (fi , fj ) is appearance kernel. In this paper, we
present two kinds of appearance kernels, i.e.,
k Ii − Ij k2
k pi − pj k 2
−
θa (fi , fj ) = exp −
2σα2
2σβ2
(4)
k di − dj k2
−
2σν2
k pi − pj k 2
k Ii − Ij k2
−
−
2σα2
2σβ2
k pi − pj k 2
k di − dj k2
−
−
2
2σα
2σν2
θa (fi , fj ) = exp
+λ exp
(5)
where pi is defined the same as before, Ii is the color vector of the ith pixel and di is the depth vector of the ith pixel.
σα , σβ , and σν control the degrees of nearness and similarity
between two pixels. With this definition, pixels with close position, similar color and similar depth are forced as the same
label. The position, color and depth features are combined into
one Gaussian kernel in Equation (4) but two Gaussian kernels
in Equation (5), where λ controls the balance between two kernels. Equation (4) indicates that big differences in either RGB
channel or depth channel can cause the different assignments
of labels, and thus the penalty will be small. On the contrary,
Equation (5) only gives a small penalty for pixels whose RGB
information and depth information are alike. In practice, we
find that Equation (4) provides better performance than Equation (5) in the context of indoor semantic segmentation.
In order to balance the importance of the depth and RGB
information, the depth input to the fully-connected CRF must
be scaled into a compatible range referred to RGB channel.
The most accessible way to do so is to directly scale the depth
channel into the range of RGB channel. However, the depth
channel contains invalid values, which is always presented as
an extreme value, i.e., 0 or 65535. These invalid values might
prevent the scaled depth values from falling into an appropriate
range. Therefore, rather than rigidly scale the depth into 0 to
255, we scale and shift every depth image to have the same
mean value and standard deviation with its RGB counterpart.
This allows the depth image and RGB image have compatible
value range in CRF model.
4
Experimental Evaluation
In this section, the proposed DFCN-DCRF is tested on a
SUN RGB-D scene understanding benchmark suite [17]. This
dataset was captured by four different kinds of sensors with
different resolutions and fields of view. It also contains the data
from NYU Depth v2 [14], Berkeley B3DO [15] and SUN3D
[16] with totally 10,335 RGB-D images and their pixel-wise
semantic annotations. Moreover, it has a default trainval-test
split which comprises 5285 images for training/validation and
5050 images for testing. To improve the quality of depth channel, multiple frames are collected to obtain a refined depth
map. However, we find that if the invalid area in the raw depth
map is too large, the corresponding refined depth image still
contains invalid measurement or losing information on corresponding pixels. Thus 387 training images are excluded, as
they have more than 45% of invalid values in the raw depth
map. According to Ref. [26], since different classes of objects
have different instance-wise and pixel-wise present frequency,
we also use weighted losses for different classes.
Training For the CNN stage, bilinear interpolation on RGB
channel and nearest-neighbor interpolation on depth channel
are applied to get 480×480-size image inputs of two branches.
The loss function is the sum of softmax loss on each pixel in
the output map. The parameters before Pool4 are initialized
with the values from the 16-layer VGG model [27] pre-trained
on ImageNet dataset [28]. Since Conv1 1 layer in depth branch
only has one channel, we average the parameter values of VGG
Conv1 1 along 3 channels to get a single channel for its initialization. During training, the data is augmented by applying
random hue, brightness, contrast and saturation adjustment on
the original image, and we randomly scale and crop the image
as well as the label to generate more data.
The proposed DFCN-DCRF architecture is implemented on
the TensorFlow framework [29], and stochastic gradient descent (SGD) is applied for end-to-end training. We set the initial learning rate of layers before Pool5 in the master branch
as 0.0002, the final score layer as 0.005, and all other layers as
0.001. All those learning rates are decayed by a factor of 0.9 in
every 50,000 iterations. A momentum of 0.9 and weight decay
of 0.0005 are also applied. The network is continually trained
on a Nvidia Titan X Pascal GPU with a batch size of 5 until
the loss does not further decrease.
For the fully-connected CRF stage, we first obtain the
DFCN response on the score layer after it is fine-tuned on training. ω2 in Equation (2) is set as 3 and σγ in Equation (4) is set
as 3. Then a random search algorithm is employed to determine the best values for ω1 , σα , σβ and σν . More concretely,
we randomly search the best values of ω1 in a range from 5 to
11, σα in a range from 90 to 170, σβ and σν in a range from 7
to 12 , which iteratively refines the search step around the last
round’s best values.
Testing The network is performed on 5050 images testing
set with three criteria, i.e., the pixel accuracy, the mean accuracy and the intersection-over-union (IoU) score. Cij denotes
the number of pixels those are predicted as category j but actually belongs to category i. Cii denotes the number of pixels
with correct prediction of category i. Ti denotes the total number of pixels that belongs to category i in the ground truth. K
denotes the total number of categories in the dataset.
i) Pixel accuracy measures the percentage of correctly classified pixels:
P
Cii
Pixel = Pi
i Ti
ii) Mean accuracy measures the classwise pixel accuracy:
Mean =
1 X Cii
K i Ti
iii) Intersection-over-union calculates the average value of the
intersection between the ground truth and the prediction
regions:
IoU =
C
1 X
P ii
K i Ti + j Cij − Cii
Among these metrics, Pixel accuracy measurement is more
sensitive to the large objects such as bed, wall, and floor in
the dataset, so Pixel accuracy measurement will be misleading
when the network performs better on the large objects. Therefore, Mean accuracy measurement and IoU score measurement
are more informative.
4.1 Quantitative Results
Two comparison experiments are conducted in this section.
In the first experiment, the proposed DFCN and DFCN-DCRF
are compared with the start-of-the-art methods. The results
are shown in Table 1. For comparisons, we also illustrate
the results of pure DFCN. The segmentation results show that
both the DFCN and DFCN-DCRF outperform other existing
methods, except for Context-CRF [10]. However, it is worth
pointing out that Context-CRF requires CNNs with multi-scale
image input and pyramid pooling to generate CRF potentials
for primary results and another dense CRF to do refinement,
which is relatively difficult to train and inference compared
with our method. It is also worth noting that our CNN method,
which has no CRF post-processing stage, already outperforms
Table 1: Comparison of segmentation results among the proposed DFCN-DCRF, DFCN and the state-of-the-art on SUN
RGB-D benchmark [16].
FCN-32s [5]
FCN-16s [5]
SegNet [7]
Context-CRF [10]
FuseNet-SF5 [24]
DFCN
DFCN-DCRF
Pixel
68.35
67.51
71.2
78.4
76.27
76.1
76.6
Mean
41.13
38.65
45.9
53.4
48.3
50.8
50.6
Image
IoU
29.00
27.15
30.7
42.3
37.29
38.0
39.3
Depth
Ground Truth
DFCN noDepth
Table 2: Segmentation results of different depth incorporation
strategies.
DFCNnoDepth
DFCNnoDepth -CRFnoDepth
DFCNnoDepth -DCRF
DFCN
DFCN-CRFnoDepth
DFCN-DCRF
Pixel
72.4
73.8
73.5
76.1
76.2
76.6
Mean
44.7
44.1
44.2
50.8
48.7
50.6
DFCN
IoU
33.4
34.4
34.4
38.0
38.5
39.3
DFCN
CRF noDepth
DFCN noDepth
DCRF
DFCN-DCRF
FuseNet, which shares the same fusing strategy on a encoderdecoder based FCN architecture.
In the second experiment, we test the proposed DFCNDCRF with different depth incorporation cases, and the results
are shown in Table 2. The CRF with and without depth are
denoted as DCRF and CRFnoDepth . It can be seen from Row
2 and Row 3 of Table 2 that if the RGB based CRF is integrated into RGB based DFCN, the performance will become
a bit better but not much. That is to say, the RGB based CRF
after RGB based DFCN can only cause a little improvement in
Pixel accuracy measurement and IoU score measurement, and
even obtain a setback in Mean accuracy measurement. The
possible reason is that objects in the indoor environment often
have similar colors. Therefore an RGB based CRF cannot distinguish the differences between the objects with similar colors. In addition, we add the depth information into only DCRF
part of the model as shown in Row 4 of Table 2, the performance is further improved a bit. The result of adding depth
information into DFCN is shown in Row 5 of Table 2, which
shows that the performance increase a lot compared with the
previous three cases. It implies that the depth information has
a great effect if it is added into the DFCN. Furthermore, we
combine the DFCN with the RGB based CRF, we find that the
performance improves a bit but still not much as shown in Row
6 of Table 2. Compared with Row 2 and 3, as well as 5 and 6,
whether integrating the RGB-based CRF into DFCN has little
influence on the performance. What is more, we compared the
proposed DFCN-DCRF with all the other networks, it has the
best performance. The same conclusion can be obtained from
Fig. 3. Moreover, we also compared our method with one
state-of-the-art method, i.e., FuseNet-SF5 in classwise mIoU
score. The results are shown in Table 3. It shows that both the
proposed DFCN-DCRF and DFCN are better than FuseNetSF5 Ref. [24].
window
counter
dresser
picture
shelves
clothes
cabinet
person
night_stand
ceiling
mirror
curtain
paper
fridge
blinds
pillow
books
whiteboard
floor_mat chair
bathtub
lamp
towel
sink
toilet
desk
bath_curtain
wall
table
door
sofa
bookshelf
bed
bag
box
tv
floor
Fig. 3: Visualization result on SUN RGB-D testing data. The
first three rows contain the RGB images, depth channel and the
segmentation ground-truth respectively. The following rows
contain the segment results of different architectures.
5
Discussion
In this paper, a novel neural network architecture (termed
DFCN-DCRF) has been designed and proposed, which combines an RGB-D fully convolutional neural network (DFCN)
with a depth-sensitive fully-connected conditional random
fields (DCRF). Different from most of methods only adding
depth-information into FCN, we have added the depth information into both the DFCN and DCRF. In addition, the design
process and theoretical analysis of the proposed DFCN-DCRF
have been presented in detail. Two comparison experiments on
SUN RGB-D benchmark have verified the effectiveness and
advantages of the proposed DFCN-DCRF on semantic segmentation.
References
[1] X. He, R. S. Zemel, and M. Á. Carreira-Perpiñán, “Multiscale
conditional random fields for image labeling,” in Proceedings of
the 2004 IEEE computer society conference on Computer vision
and pattern recognition, vol. 2, pp. 695–703, IEEE, 2004.
[2] J. Shotton, J. Winn, C. Rother, and A. Criminisi, “Textonboost
for image understanding: Multi-class object recognition and
segmentation by jointly modeling texture, layout, and context,”
International Journal of Computer Vision, vol. 81, no. 1, pp. 2–
23, 2009.
Table 3: mIoU score classwise comparison of FuseNet-SF5 in [24], our DFCN, and DFCN-DCRF
SF5 [24]
DFCN
DFCN-DCRF
SF5 [24]
DFCN
DFCN-DCRF
SF5 [24]
DFCN
DFCN-DCRF
wall
74.94
74.72
74.29
desk
18.31
20.47
21.28
towel
16.55
28.43
29.77
floor
87.41
87.41
86.78
shelf
9.20
7.16
7.23
shwr
6.06
0.21
0
cabin
41.70
41.52
43.44
ctn
52.68
53.58
55.5
box
15.77
23.62
25.69
bed
66.53
62.49
64.25
drssr
34.61
35.65
39.49
board
49.23
45.03
45.21
chair
64.45
64.58
64.80
pillow
37.77
35.50
34.41
person
14.59
29.64
35.14
sofa
50.36
48.78
51.6
mirror
38.87
28.57
28.55
stand
19.55
16.27
18.52
[3] P. Kohli, L. Ladicky, and P. H. S. Torr, “Robust higher order potentials for enforcing label consistency,” International Journal
of Computer Vision, vol. 82, no. 3, pp. 302–324, 2009.
[4] P. Krähenbühl and V. Koltun, “Efficient inference in fully connected crfs with gaussian edge potentials,” in Advances in Neural Information Processing Systems 24, pp. 109–117, 2011.
[5] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional
networks for semantic segmentation,” in Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition,
pp. 3431–3440, 2015.
[6] H. Noh, S. Hong, and B. Han, “Learning deconvolution network for semantic segmentation,” in Proceedings of the IEEE
International Conference on Computer Vision, pp. 1520–1528,
2015.
[7] V. Badrinarayanan, A. Kendall, and R. Cipolla, “Segnet: A deep
convolutional encoder-decoder architecture for scene segmentation,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. PP, no. 99, pp. 1–1, 2015.
[8] F. Yu and V. Koltun, “Multi-scale context aggregation by dilated
convolutions,” arXiv preprint arXiv:1511.07122, 2015.
[9] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L.
Yuille, “Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs,”
arXiv preprint arXiv:1606.00915, 2016.
[10] G. Lin, C. Shen, A. v. d. Hengel, and I. Reid, “Exploring context
with deep structured models for semantic segmentation,” arXiv
preprint arXiv:1603.03183, 2016.
[11] A. G. Schwing and R. Urtasun, “Fully connected deep structured networks,” arXiv preprint arXiv:1503.02351, 2015.
[12] S. Zheng, S. Jayasumana, B. Romera-Paredes, V. Vineet, Z. Su,
D. Du, C. Huang, and P. H. Torr, “Conditional random fields as
recurrent neural networks,” in Proceedings of the IEEE International Conference on Computer Vision, pp. 1529–1537, 2015.
[13] J. Dai, K. He, and J. Sun, “Boxsup: Exploiting bounding boxes
to supervise convolutional networks for semantic segmentation,” in Proceedings of the IEEE International Conference on
Computer Vision, pp. 1635–1643, 2015.
[14] N. Silberman, D. Hoiem, P. Kohli, and R. Fergus, “Indoor segmentation and support inference from rgbd images,” Proceedings of the European Conference on Computer Vision, pp. 746–
760, 2012.
[15] A. Janoch, S. Karayev, Y. Jia, J. T. Barron, M. Fritz, K. Saenko,
and T. Darrell, “A category-level 3d object dataset: Putting the
kinect to work,” in Proceedings of the IEEE International Conference on Computer Vision Workshops on Consumer Depth
Cameras for Computer Vision, pp. 1168–1174, 2011.
[16] J. Xiao, A. Owens, and A. Torralba, “Sun3d: A database of big
spaces reconstructed using sfm and object labels,” in Proceedings of the IEEE International Conference on Computer Vision,
table
49.01
44.94
45.73
mat
0
0
0
toilet
67.06
65.94
67.72
door
33.35
31.54
31.67
clthes
16.67
26.18
28.64
sink
54.99
48.84
49.91
wdw
44.77
46.18
47.64
ceiling
67.34
64.46
63.11
lamp
35.07
33.74
33.24
bslf
28.12
31.08
32.55
books
27.29
33.32
33.12
btub
63.06
56.08
60.64
pic
46.84
47.71
46.43
fridge
31.31
37.82
42.33
bag
9.52
15.41
16.52
cnter
27.73
31.09
32.00
tv
31.64
36.34
42.96
mean
37.29
38.0
39.3
blinds
31.47
31.15
32.07
paper
16.01
22.21
23.03
pp. 1625–1632, 2013.
[17] S. Song, S. P. Lichtenberg, and J. Xiao, “Sun rgb-d: A rgb-d
scene understanding benchmark suite,” in Proceedings of the
IEEE conference on computer vision and pattern recognition,
pp. 567–576, 2015.
[18] C. Couprie, C. Farabet, L. Najman, and Y. LeCun, “Indoor semantic segmentation using depth information,” arXiv preprint
arXiv:1301.3572, 2013.
[19] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and
R. Salakhutdinov, “Dropout: a simple way to prevent neural
networks from overfitting,” Journal of Machine Learning Research, vol. 15, no. 1, pp. 1929–1958, 2014.
[20] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating
deep network training by reducing internal covariate shift,”
arXiv preprint arXiv:1502.03167, 2015.
[21] A. Paszke, A. Chaurasia, S. Kim, and E. Culurciello, “Enet:
A deep neural network architecture for real-time semantic segmentation,” arXiv preprint arXiv:1606.02147, 2016.
[22] A. C. Müller and S. Behnke, “Learning depth-sensitive conditional random fields for semantic segmentation of rgb-d images,” in Proceedings of the IEEE International Conference on
Robotics and Automation, pp. 6232–6237, IEEE, 2014.
[23] S. Gupta, R. Girshick, P. Arbeláez, and J. Malik, “Learning rich
features from rgb-d images for object detection and segmentation,” in Proceedings of the European Conference on Computer
Vision, pp. 345–360, Springer, 2014.
[24] C. Hazirbas, L. Ma, C. Domokos, and D. Cremers, “Fusenet:
Incorporating depth into semantic segmentation via fusionbased cnn architecture,” in Proceedings of the Asian Conference
on Computer Vision, vol. 2, 2016.
[25] P. Kohli, M. P. Kumar, and P. H. Torr, “P3 and beyond: Solving energies with higher order cliques,” in Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition,
pp. 1–8, IEEE, 2007.
[26] D. Eigen and R. Fergus, “Predicting depth, surface normals and
semantic labels with a common multi-scale convolutional architecture,” in Proceedings of the IEEE International Conference
on Computer Vision, pp. 2650–2658, 2015.
[27] K. Simonyan and A. Zisserman, “Very deep convolutional
networks for large-scale image recognition,” arXiv preprint
arXiv:1409.1556, 2014.
[28] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al., “Imagenet large scale visual recognition challenge,” International
Journal of Computer Vision, vol. 115, no. 3, pp. 211–252, 2015.
[29] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro,
G. S. Corrado, A. Davis, J. Dean, M. Devin, et al., “Tensorflow:
Large-scale machine learning on heterogeneous distributed systems,” arXiv preprint arXiv:1603.04467, 2016.
| 1cs.CV
|
Finite group actions on 3-manifolds and cyclic
branched covers of knots
arXiv:1506.01895v3 [math.GT] 22 Jan 2018
M. Boileau∗, C. Franchi, M. Mecchia†, L. Paoluzzi∗, and B. Zimmermann†
January 23, 2018
Abstract
As a consequence of a general result about finite group actions on
3-manifolds, we show that a hyperbolic 3-manifold can be the cyclic
branched cover of at most fifteen inequivalent knots in S3 (in fact, a main
motivation of the present paper is to establish the existence of such a
universal bound). A similar, though weaker, result holds for arbitrary irreducible 3-manifolds: an irreducible 3-manifold can be a cyclic branched
cover of odd prime order of at most six knots in S3 . We note that in most
other cases such a universal bound does not exist.
AMS classification:
57M25; 57M50.
Primary 57S17; Secondary 57M40; 57M60; 57M12;
Keywords: Finite group actions on 3-manifolds; cyclic branched covers of
knots; geometric structures on 3-manifolds; hyperbolic 3-manifolds.
1
Introduction
A classical way to construct closed, connected, orientable 3-manifolds is to consider cyclic covers of the 3-sphere branched along knots. A natural question in
this setting is to understand in how many different ways a closed, connected,
orientable 3-manifold can be presented as the (total space of a) cyclic branched
cover of a knot. There is an extensive literature on this problem, mainly focussing on the case where the branching order is fixed. For instance, it is known
that a closed hyperbolic 3-manifold is the n-fold cyclic branched cover of at
most two knots in S3 , provided n > 2 [Z2], and at most nine if n = 2 [Re]. For
arbitrary closed, connected, orientable, irreducible 3-manifolds some results are
known but only for prime orders of ramification. More precisely, such manifolds
can be 2-fold cyclic branched covers of arbitrary many knots [Mon1, Mon2]
but can cover at most two knots if the order is an odd prime [BoPa]. For
branching order equal to 2, several analyses of how the different quotient knots
are related can also be found (see, for instance, [MR] for hyperbolic manifolds,
[Mon1, Mon2, V, MonW] for toroidal ones, [P] for Conway reducible hyperbolic
knots, and [Gr] for alternating ones).
∗ Partially
supported by ANR project 12-BS01-0003-01
supported by the FRA 2015 grant “Geometria e topologia delle varietà ed applicazioni”, Università di Trieste
† Partially
1
Possibly due to the fact that the knots that are covered by the same manifold
but with different orders of branching are harder to relate, not much was known
so far on the general problem, even for hyperbolic manifolds (see [RZ]). The
prime motivation for this work was to understand whether it is possible to
establish bounds on the number of ways a 3-manifold can be presented as a
cyclic branched cover of a knot, without fixing the order of the cover. By the
above discussion, no universal bound can be given in general, however the first
main result of the present paper assures the existence of a universal bound for
the class of closed hyperbolic 3-manifolds.
Theorem 1. A closed hyperbolic 3-manifold is a cyclic branched cover of at
most fifteen inequivalent knots in S3 .
We call two knots equivalent if one is mapped to the other by an orientationpreserving diffeomorphism of S3 . In the present paper, all manifolds are closed,
connected and orientable, and all maps are smooth and orientation-preserving.
The orientation-preserving isometry group of a closed hyperbolic 3-manifold
M is finite, and every finite group occurs for some hyperbolic M . Suppose
that M is a cyclic branched cover of a knot in S3 ; then the group of covering
transformations acting on M is generated by a hyperelliptic rotation: we call
a periodic diffeomorphism of a closed 3-manifold a hyperelliptic rotation if all
of its non-trivial powers have connected, non-empty fixed-point set (a simple
closed curve), and its quotient (orbit) space is S3 . By the geometrization of
3-orbifolds, or of finite group actions on 3-manifolds ([BLP], [BoP], [DL]), the
group of covering transformations is conjugate to a subgroup of the isometry
group of M . Hence establishing a universal upper bound for hyperbolic 3manifolds as in Theorem 1 is equivalent to bounding the number of conjugacy
classes of cyclic subgroups generated by hyperelliptic rotations of the isometry
group of a hyperbolic 3-manifold. Now Theorem 1 is a consequence of the
following more general result on finite group actions on closed 3-manifolds.
Theorem 2. Let M be a closed 3-manifold not homeomorphic to S3 . Let G be
a finite group of orientation-preserving diffeomorphisms of M . Then G contains
at most fifteen conjugacy classes of cyclic subgroups generated by a hyperelliptic
rotation (at most six for cyclic subgroups whose order is not a power of two).
Note that the 3-sphere is the n-fold cyclic branched cover of the trivial knot
for any integer n ≥ 2 (and, by the solution to the Smith conjecture, only of the
trivial knot). It is well-known that, for any branching order n, a 3-manifold can
be the n-fold cyclic branched cover of an arbitrary number of non-prime knots
(such a manifold is not irreducible), and that an irreducible 3-manifold can be
the 2-fold branched cover of arbitrarily many prime knots (see [Mon1, Mon2, V]).
Moreover, Proposition 7 in Section 6.1 shows that Seifert fibred manifolds can
cover an arbitrary number of knots, all with non-prime orders. It is thus natural
to restrict our attention to covers of odd prime order.
For irreducible 3-manifolds, the following holds:
Theorem 3. Let M be a closed, irreducible 3-manifold. Then there are at most
six inequivalent knots in S3 having M as a cyclic branched cover of odd prime
order.
2
The proof of Theorem 3 uses Theorem 2 in connection with the equivariant
torus-decomposition of irreducible 3-manifolds into geometric pieces, see [BoP],
[BLP], [CHK], and [KL]. For arbitrary 3-manifolds, as a direct consequence
of Theorem 3 as well as the equivariant prime decomposition for 3-manifolds
[MSY], the following remains true.
Corollary 1. Let M be a closed 3-manifold not homeomorphic to S3 . Then
M is a p-fold cyclic branched cover of a knot in S3 for at most six distinct odd
prime numbers p.
Another consequence of Theorem 2 is the following characterisation of the
3-sphere which generalises the main result of [BPZ] from the case of Z-homology
3-spheres to arbitrary closed 3-manifolds.
Corollary 2. A closed 3-manifold M is homeomorphic to S3 if and only if
there is a finite group G of orientation-preserving diffeomorphisms of M such
that G contains sixteen conjugacy classes of subgroups generated by hyperelliptic
rotations.
A noteworthy aspect of the proof of Theorem 2 is the substantial use of finite
group theory, in particular of the classification of finite simple groups. For a
prime p, the p-fold cyclic branched cover of a knot in S3 is a rational homology
3-sphere, and we will prove in Section 7 that every finite group acts non-freely
on some rational homology 3-sphere, so the use of the classification seems to
be intrinsic to the proofs of our results. We note that the class of finite groups
acting on a Z/2-homology 3-sphere instead is quite restricted (see [MZ]), and
in this case the much shorter Gorenstein-Harada classification of finite simple
groups of sectional 2-rank at most four is sufficient for our proofs. The bounds
that can be derived for Z/2-homology 3-spheres are, however, precisely the same
as those we get for arbitrary manifolds.
We have tried to separate the algebraic, purely group theoretical parts of
the proof (Section 4) from the topological parts (Sections 3 and 5), so they can
be read independently.
We note that, although the upper bounds in our results are quite small, at
this point we do not know if they are really optimal. In the hyperbolic case,
one can easily construct manifolds that are covers of orders > 2 of three distinct
knots [RZ]. For hyperbolic double covers, the bound (nine) is sharp according
to a result of Kawauchi [Ka], but no explicit examples are known so far. For
general irreducible manifolds, Brieskorn spheres of type Σ(p, q, r), where p, q,
and r are three pairwise different odd primes, provide examples of manifolds
that cover four knots: Σ(p, q, r) is the p-fold (resp. q-fold and r-fold) cyclic
cover of S3 branched along the torus knot T (q, r) (resp. T (p, r) and T (p, q)) as
well as the double branched cover of a Montesinos knot.
The paper is organised as follows. In Section 2 we present a brief sketch of the
proof of Theorem 2. Hyperelliptic rotations and their properties are considered
in Section 3. Section 4 contains the main group-theoretical part of the paper,
Section 5 the proof of Theorem 2, and Section 6 the proof of Theorem 3 for the
irreducible case. Finally, in an Appendix we prove that every finite group acts
non-freely on some rational homology sphere (adapting the result of [CL] that
deals with free actions).
3
2
Sketch of the proof of Theorem 2
The proof of Theorem 2 is based on a series of preliminary results which are
presented in Sections 3 and 4. Our choice to present the group-theoretical part
of the proof in a separate section (Section 4) allows for the group-theoretical
results to be read independently of the other parts of the paper. In the following,
in order to make the paper more accessible, we explain the main steps of the
proof.
We begin with a more detailed definition of hyperelliptic rotation. Note that
throughout the paper, unless otherwise stated, 3-manifold will mean orientable,
connected, closed 3-manifold. Also, all finite group actions by diffeomorphisms
will be faithful and orientation-preserving.
Definition 1. Let ψ : M −→ M be a finite order diffeomorphism of a 3manifold M . We shall say that ψ is a rotation if it preserves the orientation
of M , F ix(ψ) is non-empty and connected, and F ix(ψ) = F ix(ψ k ) for all nontrivial powers ψ k of ψ. F ix(ψ) will be referred to as the axis of the rotation.
Note that if ψ is a periodic diffeomorphism of prime order, then ψ is a rotation
if and only if F ix(ψ) = S1 . We shall say that a rotation ψ is hyperelliptic if
the space of orbits M/ψ of its action is S3 , and a hyperelliptic group is a cyclic
group generated by a hyperelliptic rotation.
We start by observing that the case of hyperelliptic rotations whose order is
a power of two is already well-understood by work of Reni and Mecchia (see [Re]
and [Mec1]). In particular, there are at most nine conjugacy classes of cyclic
groups generated by such hyperelliptic rotations. As a consequence, from now
on, we exclude this case and consider only hyperelliptic groups whose order is
not a power of two.
Section 3 collects various simple facts on the geometry of hyperelliptic rotations. In particular, we prove that there are at most three hyperelliptic groups
commuting pairwise. This result implies the existence of a universal bound in
the solvable case. In fact, in the solvable case, the presence of Hall subgroups
assures that all hyperelliptic subgroups commute, up to conjugacy, implying
that there at most three conjugacy classes of such groups.
The case of non-solvable groups, where local approaches on the basis of pgroups fail, is more involved. We need a global description of the groups that
may arise, which is provided in Section 4. In that section, we introduce the
notion of an algebraically hyperelliptic collection of cyclic subgroups which have
the same algebraic properties as the hyperelliptic subgroups; this allows a purely
algebraic approach in Section 4. The main result there (Theorem 4) is that a
non-solvable finite group generated by an algebraically hyperelliptic collection
is of a very special type, in particular G has a quotient by a normal solvable
subgroup which is isomorphic to the direct product of at most two simple groups.
The next step is to cover the hyperelliptic subgroups by a bounded number
of conjugacy classes of solvable subgroups (i.e., to find a collection of solvable
subgroups such that each element of odd prime order in a hyperelliptic subgroup has a conjugate in one of these solvable subgroups); this concept of a
solvable cover of a finite group is central for the proofs in the present paper
since a bound on the number of elements of such a cover implies that there is
a bound on the number of conjugacy classes of hyperelliptic subgroups. Using
the classification of finite simple groups, we prove, in Proposition 4, that in any
4
finite simple group the hyperelliptic subgroups can be covered, up to conjugacy,
by at most four solvable subgroups. This result, together with the characterisation of groups generated by algebraically hyperelliptic collections, gives directly
the existence of a universal bound of fifty-seven, much larger than the bound
of fifteen obtained in Theorem 2 by exploiting extra topological considerations.
Note that the existence of a universal (although non explicit) bound is ensured
by the existence of only a finite number of simple sporadic groups.
In Section 5 we conclude the proof of Theorem 2. We can suppose that G is
generated by an algebraically hyperelliptic collection and that G is not solvable.
Under these hypotheses, the proof is divided into two cases.
In the first case, we suppose that G contains no rotation of order two. By
geometric motivations, an involution acting dihedrally by conjugation on a hyperelliptic subgroup is a rotation. The absence of this type of involutions induces
further restrictions on the structure of G; in particular, up to a quotient by a
solvable subgroup, G is a single simple group. By using these properties of G
and the solvable covers we prove that four is the upper bound in this case.
In the second case, G contains a rotation of order two. The groups that
may act on 3-manifolds containing such an involution are listed in [Mec2]. We
combine this with our result about groups generated by an algebraically hyperelliptic collection and we obtain that a quotient of G by a solvable subgroup
must be isomorphic to one of A8 , P SL2 (q) or P SL2 (q) × P SL2 (q ′ ). For these
groups we explicitly find a solvable cover with a bounded number of elements
and hence a universal bound as in Theorem 2.
3
Rotations and their properties
In this section we shall establish some properties of rotations in general and
hyperelliptic ones in particular.
Remark 1. Assume that ψ is a hyperelliptic rotation acting on a 3-manifold
M . The natural projection from M to the space of orbits M/ψ of ψ is a cyclic
cover of S3 branched along a knot K = F ix(ψ)/ψ. The converse is also true,
that is any deck transformation generating the automorphism group of a cyclic
covering of S3 branched along a knot is a hyperelliptic rotation.
We observe that cyclic branched covers of prime order are closely related to
Q-homology 3-spheres.
Remark 2.
1. If the order of ψ is a prime p, then M is a Z/p-homology sphere [Go].
2. By Smith theory, if f is a periodic diffeomorphism of order p, a prime
number, acting on a Z/p-homology sphere, then f either acts freely or is
a rotation.
We start with a somehow elementary remark which is however central to
determine constraints on finite groups acting on 3-manifolds.
Remark 3. Let G ⊂ Dif f + (M ) be a finite group of diffeomorphisms acting on
a 3-manifold M . One can choose a Riemannian metric on M which is invariant
by G and with respect to which G acts by isometries. Let now ψ ∈ G be
5
a rotation. Since the normaliser NG (hψi) of ψ in G consists precisely of those
diffeomorphisms that leave the circle F ix(ψ) invariant, we deduce that NG (hψi)
is a finite subgroup of Z/2 ⋉ (Q/Z ⊕ Q/Z), where the nontrivial element in Z/2
acts by conjugation sending each element of Q/Z⊕Q/Z to its inverse. Note that
the elements of NG (hψi) are precisely those that rotate about F ix(ψ), translate
along F ix(ψ), or invert the orientation of F ix(ψ); in the last case the elements
have order 2 and non-empty fixed-point set meeting F ix(ψ) in two points.
Remark 4. Let us consider the 3-sphere S3 . According to Smith’s theory, an
orientation-preserving finite-order diffeomorphism of S3 is a rotation if and only
if its fixed-point set is non-empty. Because of the positive solution to the Smith
conjecture the fixed-point set of a rotation of S3 is the trivial knot. Morover,
any group of symmetries of a non-trivial knot K (that is, any finite group of
orientation-preserving diffeomorphisms of S3 acting on the pair (S3 , K)) is either
cyclic or dihedral.
Lemma 1. Let ϕ and ψ be two rotations contained in a finite group of orientationpreserving diffeomorphisms of a 3-manifold M .
1. A non-trivial power of ψ of order different from 2 commutes with a nontrivial power of ϕ, if and only if ϕ and ψ commute.
2. Assume M 6= S3 . If ϕ and ψ are hyperelliptic and F ix(ϕ) = F ix(ψ), then
hϕi = hψi (in particular they have the same order).
3. Assume M 6= S3 . If ϕ and ψ are hyperelliptic, then hϕi and hψi are
conjugate if and only if some non-trivial power of ϕ is conjugate to some
non-trivial power of ψ.
Proof.
Part 1 The sufficiency of the condition being obvious, we only need to prove
the necessity. Remark that we can assume that both rotations act as isometries
for some fixed Riemannian metric on the manifold. Denote by f and g the
non trivial powers of ϕ and ψ, respectively. Note that, by definition, F ix(ψ) =
F ix(g) and F ix(ϕ) = F ix(f ). Since g and f commute, g leaves invariant
F ix(ϕ) = F ix(f ) and thus normalises every rotation about F ix(ϕ). Moreover
g and ϕ commute, for the order of g is not 2 (see Remark 3). In particular, ϕ
leaves F ix(ψ) = F ix(g) invariant and normalises every rotation about F ix(ψ).
The conclusion follows.
Part 2 Reasoning as in Part 1, one sees that the two rotations commute. Assume, by contradiction, that the subgroups they generate are different. Under
this assumption, at least one of the two subgroups is not contained in the other.
Without loss of generality we can assume that hϕi 6⊂ hψi. Take the quotient of
M by the action of ψ. The second rotation ϕ induces a non-trivial rotation of S3
which leaves the quotient knot K = F ix(ψ)/ψ ⊂ S3 invariant. Moreover, this
induced rotation fixes pointwise the knot K. The positive solution to the Smith
conjecture implies now that K is the trivial knot and thus M = S3 , against the
hypothesis.
Part 3 follows from 2 since the conjugate of a hyperelliptic rotation is again a
hyperelliptic rotation.
Corollary 3. Let G be a finite group of orientation-preserving diffeomorphisms
acting on a 3-manifold M 6= S3 . Let ψ be a rotation and let f ∈ G be an element
6
of odd prime order which is a power of ψ. Then we have CG (hf i) = CG (hψi)
and NG (hf i) = NG (hψi).
There is a natural bound on the number of hyperelliptic subgroups of order
not a power of two which commute pairwise; we begin analysing the situation
of the symmetry group of a knot.
Definition 2. A rotation of a knot K in S3 is a rotation ψ of S3 such that
ψ(K) = K and K ∩ F ix(ψ) = ∅. We shall say that ψ is a full rotation if K/ψ
in S3 = S3 /ψ is the trivial knot.
Remark 5. Let ψ and ϕ be two commuting rotations acting on some manifold
M and with orders not both equal to 2. Assume that ψ is hyperelliptic and
ϕ is not a power of ψ. According to Remark 3 we have two situations: Either
F ix(ψ)∩F ix(ϕ) = ∅ and ϕ induces a rotation φ of K = F ix(ψ)/ψ, or F ix(ψ) =
F ix(ϕ). In the former situation we have that ϕ is hyperelliptic if and only if
φ is a full rotation. This can be shown by considering the quotient of M by
the action of the group generated by ψ and ϕ. This quotient is S3 and the
projection onto it factors through M/ϕ, which can be seen as a cyclic cover of
S3 branched along K/φ. By the positive solution to the Smith conjecture, M/ϕ
is S3 if and only if K/φ is the trivial knot. In the latter situation, again because
of the positive solution to Smith’s conjecture, we have M = S3 (compare also
Part 2 of Lemma 1).
The following finiteness result about commuting rotations of a non-trivial
knot in S3 is one of the main ingredients in the proof of Theorem 2 (see [BoPa,
Proposition 2], and [BoPa, Theorem 2] for a stronger result where commutativity
is not required).
Proposition 1. Let K be a non-trivial knot in S3 . Let us consider a set of
pairwise commuting full rotations in Dif f + (S3 , K). The elements of the set
generate at most two pairwise distinct cyclic subgroups.
Proof. Note, first of all, that, according to Remark 4 the finite subgroup of
Dif f + (S3 , K) generated by a finite set of pairwise commuting full rotations is
cyclic.
Assume now, by contradiction, that there are three pairwise distinct cyclic
subgroups generated by commuting full rotations of K, ϕ, ψ and ρ respectively.
Note that such cyclic subgroups have distinct orders. Assume that two of them
-say ϕ, ψ- have the same axis. Fix the one with smaller order -say ψ-: since ψ is
a full rotation, the quotient K/ψ is the trivial knot, and ϕ induces a rotation of
K/ψ which is non-trivial since ϕ commutes with ψ and its order is larger than
that of ψ. The axis A of this induced symmetry is the image of F ix(ψ) in the
quotient S3 /ψ by the action of ψ. In particular K/ψ and A form a Hopf link
and K is the trivial knot: this follows from the equivariant Dehn lemma, see
[Hil].
We can thus assume that the axes are pairwise disjoint. Indeed, this follows
from Remark 3 taking into account that the rotations commute pairwise and
at most one of them has order 2. Therefore we would have that the axis of ρ,
which is a trivial knot, admits two commuting rotations, ϕ and ψ, with distinct
axes, which is impossible: this follows, for instance, from the fact (see [EL, Thm
5.2]) that one can find a fibration of the complement of the trivial knot which
is equivariant with respect to the two symmetries.
7
Observe that the proof of the proposition shows that two commuting full
rotations of a non-trivial knot either generate the same cyclic subgroup or have
disjoint axes.
We stress that if a knot K ⊂ S3 admits a full rotation, then it is a prime
knot, see [BoPa, Lemma 2].
Lemma 2. Let G be a finite group of diffeomorphisms acting on a 3-manifold
M 6= S3 and {H1 , . . . , Hm } be a set of hyperelliptic subgroups of G of order not
a power of two. Suppose that there exists an abelian subgroup of G, containing
at least an element of odd prime order of each Hi , then m ≤ 3. Moreover either
the orders of the Hi are pairwise coprime or m ≤ 2.
Proof.
By Lemma 1 we obtain that the subgroups Hi commute and have trivial
intersection pairwise. Consider the cyclic branched covering M → M/H1 ∼
= S3
over the knot F ix(H1 )/H1 . By projecting Hi with i ≥ 2 to M/H1 we obtain full
rotations of F ix(H1 )/H1 . Note that if two subgroups Hi have the same order
that is moreover a divisor of the order of H1 , a priori they might map to the
same subgroup in hH1 , . . . , Hm i/H1 . We claim, however, that the induced full
rotations are distinct so there are m − 1 of them. This follows from the fact
that, for different indices i and j, F ix(Hi ) and F ix(Hj ) are disjoint according
to Part 2 of Lemma 1. Now, since the subgroups commute, for all i = 1, . . . , m
we have H1 (F ix(Hi )) = F ix(Hi ), so that the fixed-point sets of the induced
full rotations are disjoint, too, and the full rotations are pairwise distinct. By
Proposition 1 we obtain m − 1 ≤ 2. Note that, by Remark 4 a non-trivial knot
cannot admit two distinct and commuting cyclic groups of symmetries of the
same order. This proves the latter part of the lemma.
The above lemma implies directly the following corollary.
Corollary 4. Let p be an odd prime and assume that H ∼
= Z/p ⊕ Z/p acts on a
3-manifold M 6= S3 , the group H contains at most two distinct cyclic subgroups
generated by non-trivial powers of hyperelliptic rotations.
We now collect several facts relative to hyperelliptic rotations that can be
deduced from the discussion in this section.
Remark 6. Let G be a finite group of orientation-preserving diffeomorphisms
of a 3-manifold M 6= S3 . Let ψ ∈ G be a hyperelliptic rotation of order not a
power of 2. Recall that the structures of the centraliser and of the normaliser
of hψi are described in Remark 3. Let N denote the normaliser NG (hψi).
1. The centraliser CG (hψi) of ψ in G satisfies 1 −→ hψi −→ CG (hψi) −→
H −→ 1, where H is cyclic, possibly trivial. This follows from Remark 4.
2. Since the symmetries of a knot not acting freely have connected fixedpoint set (see again Remark 4), the possible elements of N \ CG (hψi) are
rotations of order two.
3. According to last part of Lemma 2, the centraliser CG (hψi) contains at
most one more cyclic subgroup of the same order as hψi and generated by
a hyperelliptic rotation.
4. Observe that we have NG (N ) = NG (CG (hψi)).
8
5. Of course, the conjugate of a hyperelliptic rotation is again a hyperelliptic
rotation. Consider an element g ∈ NG (N ). Then either g conjugates hψi
to itself and so g belongs to N , or g does not normalise hψi. In this case,
CG (hψi) contains precisely two cyclic subgroups generated by hyperelliptic
rotations of the same order as ψ (hψi and another one) and g exchanges
them. It follows that CG (hψi) has index a divisor of 4 in NG (N ).
6. Let f ∈ G be an element of odd prime order which is a power of ψ. Then
we have CG (hf i) = CG (hψi) and NG (hf i) = N . This is just Corollary 3.
The following general observation will be useful in the sequel.
Remark 7. With the notation used in the previous remark, assume that there
are elements in NG (N ) which do not act in the same way on CG (hψi), then
necessarily the index of N in NG (N ) is two and CG (hψi) contains two distinct
subgroups conjugate to hψi.
We end this section by providing a dictionary translating between algebraic
properties of the structure of NG (N ) and the symmetries of the knot K =
F ix(ψ)/ψ. This will not be needed in the proofs of our results but will provide
a geometric interpretation of the different situations that occur in the proof of
Theorem 2 (see Remark 9). We start with the following definition.
Definition 3. A 2-component link is called exchangeable if there exists an
orientation-preserving diffeomorphism of S3 which exchanges the two components of the link.
Let K be a knot and ρ a rotation of K of order n and with axis A. Consider
the 2-component link K ∪ A consisting of the images of the knot K and of the
axis A in the quotient S3 /ρ of the 3-sphere by the action of ρ. Note that at
least one component of this link (i.e. A) is trivial. We call K n-self-symmetric
if K ∪ A is exchangeable. In this case ρ is a full rotation of K.
=
Figure 1: A 5-self-symmetric knot on the left, and its exchangeable quotient
link on the right.
Since the structure of the normaliser of hψi and of its centraliser, only depend
on the symmetries of K that lift to G, we introduce the following definitions:
Definition 4. Let G be a finite group of orientation preserving diffeomorphisms
of a closed connected 3-manifold M . Let ψ be a hyperelliptic rotation contained
in G with quotient knot K. We say that K is strongly invertible with respect
9
to G if K admits a strong inversion that lifts to G. Similarly we say that K
is self-symmetric with respect to G if G contains an element ψ ′ conjugate to ψ
such that the subgroup hψ, ψ ′ i is abelian of rank 2, i.e. not cyclic. Remark that
in the latter case K is n-self-symmetric, where n is the order of ψ.
The proof of the following facts is elementary and left to the reader.
Proposition 2. Let K, ψ, N , and G be as above.
• The centraliser of hψi in G is contained with index 2 in N if and only if
K is strongly invertible with respect to G.
• N is contained with index 2 in NG (N ) if and only if K is self-symmetric
with respect to G.
Moreover, if M is hyperbolic and G = Iso+ (M ), K is strongly invertible if and
only if it is strongly invertible with respect to G, and it is self-symmetric with
respect to G if and only if it is n-self-symmetric, where n is the order of the
hyperelliptic rotation ψ.
4
Group theoretical results
This section is devoted to the proofs of the group-theoretical results that are
used to obtain the bound provided by Theorem 2.
The first result, Theorem 4, is probably the more interesting of the two from
a group-theoretical point of view. It describes the finite groups G that may be
generated by hyperelliptic rotations. In this settings, a hyperelliptic rotation
is simply an element of G of order not a power of two, satisfying a purely
algebraic condition (see Definition 5, below) on the structure of the normaliser
of (a power of odd prime order of) the element. More precisely, Theorem 4
states that either G is solvable, or it admits a quotient by a normal solvable
subgroup so that the quotient is either the product of a simple group times a
cyclic group of odd order, or the product of two simple groups. We see that the
fact of being generated by hyperelliptic rotations puts very strong constraints
on the structure of G.
The second result, Proposition 4, although possibly not as striking as the
previous one, is however key to be able to carry out the strategy of bounding
the number of hyperelliptic subgroups by covering them with solvable subgroups
(see Definition 6). Indeed, in Proposition 4 we establish the fact that four
conjugacy classes of solvable subgroups are sufficient to contain, up to conjugacy,
all hyperelliptic subgroups of orders not a power of two that may sit inside a
finite simple group.
Note that these two main results imply that 4 × 4 solvable subgroups suffice
to cover all hyperelliptic groups in G, and together with Lemma 3, and Lemma 2
show that a finite group G can contain at most forty-eight conjugacy classes of
hyperelliptic subgroups of orders that are not powers of 2.
In the following G will denote a finite group. We start with some preliminary
observations and definitions.
10
Definition 5. A collection {Ci } of subgroups of G, each of odd prime order pi ,
is said to be algebraically hyperelliptic if, for each i, the following conditions are
satisfied:
1. the centraliser of Ci in G is abelian of rank at most two and has index at
most two in the normaliser of Ci in G;
2. each element belonging to the normaliser of Ci but not to the centraliser
inverts by conjugation each element in the centraliser;
3. if Ci is contained in a Sylow pi -subgroup Si of G, then Si contains at most
two distinct conjugates of Ci .
We remark that in this definition the primes pi are not necessarily pairwise
distinct.
Proposition 3. Let S be a Sylow pi -subgroup where pi is the order of a group
Ci belonging to an algebraically hyperelliptic collection.
1. S is either cyclic or the product of two cyclic groups, and
2. NG (S) contains with index at most 2 the normaliser of a conjugate of Ci ,
and contains an abelian subgroup of rank at most 2 with index a divisor
of 4. In particular NG (S) is solvable.
Proof.
Up to conjugation we can suppose that S contains Ci . By Properties 1 and
2 in Definition 5 and the fact that pi is odd, the normaliser NS (Ci ) is abelian
of rank at most two. Property 3 in Definition 5 implies that NS (NS (Ci )) =
NS (Ci ). Since S is a pi -subgroup we obtain that S = NS (Ci ) and we get the
first part of the thesis.
Since the subgroup NG (S) normalises the maximal elementary abelian subgroup of S, we obtain also the second part of the thesis.
Lemma 3. Let G be a solvable group containing an algebraically hyperelliptic
collection {C1 , · · · , Cm } of subgroups of odd prime order. Then there exists an
abelian subgroup of G containing a conjugate of Ci , for each i = 1, . . . , m. In
particular the subgroups Ci commute pairwise up to conjugacy.
Proof. Let π be the set of the orders of the Ci and let B be a Hall π-subgroup
of G. Each Ci is conjugate to a subgroup of B. Since π contains only odd primes,
Definition 5 yields that centraliser and normaliser of every Sylow p-subgroup of
B coincide. By Burnside’s p-complement theorem (see [Su, Theorem 2.10 page
144]), every Sylow p-subgroup of B has a normal complement, and hence B is
abelian.
Remark 8. Suppose that N is a normal subgroup of G and H is a p-subgroup
of G. If the order of N is coprime with p, then the normaliser of the projection
of H to G/N is the projection of the normaliser of H in G, that is
NG/N (HN/N ) = NG (H)N/N.
The inclusion ⊇ holds trivially. We prove briefly the other inclusion. Let f N
be an element of NG/N (HN/N ), then H f ⊆ HN . Both H f and H are Sylow
11
p-subgroups of HN and by the second Sylow theorem they are conjugate by an
element hn ∈ HN. We obtain that H f hn = H, and hence f ∈ NG (H)N.
Analogously if f is an element of prime order coprime with the order of N ,
then CG/N (f N ) = CG (f )N/N.
Recall that a finite group Q is quasisimple if it is perfect (the abelianised
group is trivial) and the factor group Q/Z of Q by its centre Z is a nonabelian
simple group (see [Su, chapter 6.6]). A group E is semisimple if it is perfect
and the factor group E/Z(E) is a direct product of nonabelian simple groups.
A semisimple group E is a central product of quasisimple groups which are
uniquely determined. Any finite group G has a unique maximal semisimple
normal subgroup E(G) (maybe trivial), which is characteristic in G. The subgroup E(G) is called the layer of G and the quasisimple factors of E(G) are
called the components of G.
The maximal normal nilpotent subgroup of a finite group G is called the Fitting subgroup and is usually denoted by F (G). The Fitting subgroup commutes
elementwise with the layer of G. The normal subgroup generated by E(G) and
by F (G) is called the generalised Fitting subgroup and is usually denoted by
F ∗ (G). The generalised Fitting subgroup has the important property to contain its centraliser in G, which thus coincides with the centre of F ∗ (G). For
further properties of the generalised Fitting subgroup see [Su, Section 6.6.].
Theorem 4. Suppose that G is generated by the algebraically hyperelliptic collection H := {C1 , · · · , Cm }. Denote by pi the order of Ci and by A the maximal
normal solvable subgroup of order coprime with every pi .
If G is non-solvable, then the following properties hold:
1. every pi divides the order of any component of G/A;
2. either G/A is the direct product of a cyclic group of odd order and a simple
group or it is the direct product of two simple groups;
3. if in addition G does not contain any involution acting dihedrally on any
Ci , then E(G/A) is simple, every pi divides the order of F (G/A) and a
Sylow pi -subgroup of G contains exactly two conjugates of Ci .
Proof. Let π be the set of the primes pi .
By Remark 8 we can suppose that A is trivial and F (G) is a π-group.
Claim 1. F (G) is cyclic and E(G) is not trivial.
Suppose first that F (G) contains an abelian pi -subgroup S of rank two.
Then S, being the maximal elementary abelian pi -subgroup contained in F (G),
is normal in G and contains Ci . This implies that G is solvable and we get a
contradiction. Hence F (G) is cyclic.
If E(G) is trivial, then F (G) = F ∗ (G) and G/F (G) is isomorphic to a
subgroup of AutF (G). Since F (G) is cyclic AutF (G) is solvable, and we get
again a contradiction.
Claim 2. Each pi divides the order of any component of G. Moreover the
components of G are simple groups and are at most two.
12
Since the Sylow pi -subgroups are abelian and A is trivial, by [Su, Exercise
1, page 161] the components of G have trivial centre.
Now we prove that each component of G is normalised by any Ci . Let fi be
a generator of Ci and Q a component of G. Suppose by contradiction that Q is
not normalised by fi . We define the following subgroup:
Qc = {xfi xfi−1 . . . fipi −1 xfipi −1 | x ∈ Q}.
Since the components of G commute elementwise, Qc is a subgroup of G isomorphic to Q. Moreover, each element of Qc commutes with fi and this gives
a contradiction.
We have that Ci normalises Q but cannot centralise it, so the action by
conjugation of fi on Q is not trivial.
Assume that Q is either sporadic or alternating. Since the order of the outer
automorphism group of any such simple group is a (possibly trivial) power of
2 (see [GLS, Section 5.2 and 5.3]), we conclude that f must induce an inner
automorphism of Q. In particular pi divides the order of Q.
We can thus assume that Q is a simple group of Lie type.
Recall that, by [GLS, Theorem 2.5.12], Aut(Q) is the semidirect product
of a normal subgroup Inndiag(Q), containing the subgroup Inn(Q) of inner
automorphisms, and a group ΦΓ, where, roughly speaking, Φ is the group of
automorphisms of Q induced by the automorphisms of the defining field and Γ
is the group of automorphisms of Q induced by the symmetries of the Dynkin
diagram associated to Q (see [GLS] for the exact definition). By [GLS, Theorem 2.5.12.(c)], every prime divisor of |Inndiag(Q)| divides |Q|. Thus we
can assume that the automorphism induced by fi on Q is not contained in
Inndiag(Q) and its projection θ on Aut(Q)/Inndiag(Q) ∼
= ΦΓ has order pi .
We will find a contradiction showing that in this case the centraliser of fi in Q
is not abelian.
Write θ = φγ, with φ ∈ Φ and γ ∈ Γ. If φ = 1, then γ is nontrivial and fi
induces a graph automorphism according to [GLS, Definition 2.5.13]. Since pi
is odd, the only possibility is that Q is D4 (q) and pi = 3 (see [GLS, Theorem
2.5.12 (e)]). The centraliser of fi in Q is nonabelian by [GLS, Table 4.7.3 and
Proposition 4.9.2.]. If φ 6= 1 and Q is not isomorphic to the group 3 D4 (q), then
the structure of the centraliser of fi in Q is described by [GLS, Theorem 4.9.1],
and it is nonabelian. Finally, if φ 6= 1 and Q ∼
= 3 D4 (q), the structure of the non
abelian centraliser of fi in Q follows from [GLS, Proposition 4.2.4]. We proved
that the automorphism induced by fi is contained in Inndiag(Q) and pi divides
|Q|.
Since pi divides the order of any component, G has at most two components.
Claim 3. G = E(G)F (G).
We prove first that G = E(G)CG (E(G)). Let us assume by contradiction
that there exists Ci with trivial intersection with E(G)CG (E(G)) and denote by
f a generator of Ci . Since pi divides the order of every component of G and the
Sylow pi -subgroup has rank at most 2, we get that E(G) has only one component
which we denote by Q. The Sylow pi -subgroups of Q are cyclic. Moreover, by the
first part of the proof, the automorphism induced by f on Q is inner-diagonal.
If it is inner, we obtain f as a product of an element that centralises Q and
an element in Q, a contradiction to our assumption; otherwise, we get again
13
contradiction, since, by [GLS, Theorem 2.5.12] and [A, (33.14)], a group of Lie
type with cyclic Sylow pi -subgroup cannot have a diagonal automorphism of
order pi .
Hence, all the subgroups Ci are contained in E(G)CG (E(G)) and, since they
generate G, we obtain that G = E(G)CG (E(G)).
Now if F (G) = 1, then F ∗ (G) = E(G) and hence CG (E(G)) = CG (F ∗ (G)) =
Z(E(G)) = 1 and the claim is proved. So suppose that F (G) 6= 1. Then,
since F (G) is a π-group, there is at least one subgroup Ci that is contained in
E(G)F (G). Hence there is a subgroup T1 of E(G) with order pi and a subgroup
T2 of F (G) with order pi such that Ci ≤ T1 T2 . Since F (G) is cyclic, T2 is
normal in G and so CG (E(G)) ≤ NG (T1 T2 ). But NG (T1 T2 ) has an abelian
normal subgroup of index a divisor of 4, so CG (E(G)) is solvable. This shows
that every Sylow p-subgroup of CG (E(G)) for p odd is contained in F (G), whence
it follows that G = E(G)F (G).
Claim 4. If no involution of G acts dihedrally on any Ci , then E(G) is simple,
and, for every i, pi divides the order of F (G) and a Sylow pi -subgroup of G
contains exactly two conjugates of Ci .
Suppose no involution of G acts dihedrally on any Ci . By Definition 5, it
follows that NG (Ci ) = CG (Ci ) for every i = 1, . . . , m. Let Q be a component of
E(G) and suppose by contradiction that Q contains a Sylow pi -subgroup S of
rank two, for some i. Up to conjugation we can suppose that S contains Ci . By
Definition 5, NQ (S) contains with index at most two the abelian group CQ (Ci ).
Since Q is simple, Burnside’s p-complement theorem (see [Su, Theorem 2.10
page 144]) yields that NQ (S) is not abelian. Therefore, NQ (S) contains with
index two NQ (Ci ) and the elements of NQ (S) \ NQ (Ci ) conjugate Ci to a cyclic
subgroup distinct from Ci . By using [Su, Exercise 1, page 161] and the fact that
Q is perfect we get again a contradiction. Hence, for every i ∈ {1, . . . , m}, the
Sylow pi -subgroups of Q are cyclic. Now, as above for every Sylow subgroup S
of Q we have NQ (S) 6= CQ (S). Since NG (Ci ) = CG (Ci ) for every i = 1, . . . , m,
that implies that the Sylow pi -subgroups of G are not cyclic and hence pi divides
the order of F (G) for every i.
To bound above the number of conjugacy classes of hyperelliptic rotations,
our strategy will consist of conjugating hyperelliptic rotations in solvable subgroups of G, where they are forced to commute, hence, in analogy with the
standard definition of normal covers, we introduce the notion of solvable normal π-cover and we prove the following lemma.
Definition 6. Let G be a finite group. Let π be a set of primes dividing
|G|. We will call a collection C of subgroups of G a solvable normal π-cover
of G if every element of G of prime order p belonging to π is contained in an
element of C and for every g ∈ G, H ∈ C we have that H g ∈ C. We denote
by γπs (G) the smallest number of conjugacy classes of subgroups in a solvable
normal π-cover of G. Here, the letter “s” stands for “solvable”, and is used to
distinguish this number from γ(G), that is the standard notation in the case
of covers by subgroups that are not requested to be solvable. Note that, since
Sylow subgroups are clearly solvable, γπs (G) ≤ |π|.
Proposition 4. Let G be a finite nonabelian simple group. If π is the set of odd
primes p such that G has cyclic Sylow p-subgroup, the centraliser of CG (g) is
14
abelian for every element g ∈ G of order p and the normaliser of any subgroup
of order p contains with index at most two its centraliser, then γπs (G) ≤ 4.
Proof.
If G is a sporadic group, the primes dividing the order of the group do not
satisfy the condition on the normaliser.
If G is isomorphic to the alternating group An and p ∈ π, then the condition
on the centraliser of the elements of order p implies that p > n − 4 and γπs (G) ≤
|π| ≤ 2.
The only remaining case is that of groups of Lie type.
Let G ∼
= Σn (q) or G ∼
= d Σn (q), where q is a power of a prime t. Here we
use the same notation as in [GLS]: the symbol Σ(q) (resp. d Σ(q)) may refer
to finite groups in different isomorphism classes, each of them is an untwisted
(resp. twisted) finite group of Lie type with root system Σ (see [GLS, Remark
2.2.5]). Any finite group of Lie type is quasisimple with the exception of the following groups: A1 (2), A1 (3), 2 A2 (2), 2 B2 (2), B2 (2), G2 (2), 2 F4 (2) and 2 G2 (3)
(see [GLS, Theorem 2.2.7]).
If t ∈ π, then by [GLS, Theorem 3.3.3], either t = 3 and G ∼
= (2 G2 (3))′ or
2
′
∼
G = A1 (s). In the former case the order of G2 (3) is divided only by two odd
primes, thus γπs (G) ≤ 2; in the latter case we have γπs (G) ≤ 2 (see for example
[H]).
Assume now that t 6∈ π. By [GLS, Paragraph 4.10], since the Sylow subgroups are cyclic, every element of order p ∈ π is contained in a maximal torus
of G, and clearly a maximal torus is abelian.
Therefore, we need only to bound the number of conjugacy classes of cyclic
maximal tori in G with abelian centraliser. Note that the number of conjugacy
classes of maximal tori in G is bounded by the number of different cyclotomic
polynomials evaluated in q appearing as factors of |G|. Moreover the power of
a cyclotomic polynomial in the order of G gives the rank of the corresponding
maximal torus (except possibly when the prime divides the order of the centre
but in this case the Sylow subgroup is not cyclic, see [A, (33.14)])
Recall Σ is the root system associated to G as in [GLS, 2.3.1]; let Π =
{α1 , . . . , αn } be a fundamental system for Σ as in Table 1.8 in [GLS], α∗ be
the lowest root relative to Π as defined in [GLS, Paragraph 1.8] and set Π∗ =
Π ∪ {α∗ }. We recall that |G| can be deduced from [GLS, Table 2.2] and the
Dynkin diagrams can be found in [GLS, Table 1.8]. Observe that, by [GLS,
Proposition 2.6.2], if Σ0 is a root subsystem of Σ, then G contains a subsystem
subgroup H, which is a central product of groups of Lie type corresponding
to the irreducible constituents of Σ0 . In order to prove the lemma, we shall
show that for every group G and for every element g of order a prime r lying in
a maximal torus belonging to any but four conjugacy classes of maximal tori,
either the Sylow r-subgroup is not cyclic or we find a subsystem subgroup H
that is a central product of two groups H1 and H2 such that H1 contains g and
H2 is not abelian. Note that for every prime power q, A1 (q) is a non-abelian
group (see [GLS, Theorem 2.2.7]).
We treat the case G ∼
= An (q) in detail as an example. All other cases can be
dealt similarly. Assume G ∼
= An (q). Let m be the minimum index i such that r
divides q i+1 −1 and let Σ0 be generated by Π∗ \{α1 , αn }. Then the corresponding
subsystem subgroup is H = H1 · H2 , where H1 ∼
= A1 (q).
= An−2 (q) and H2 ∼
Thus if m ≤ n − 1, then H1 contains an element g of order r and CG (g) contains
15
H2 which is not abelian. Therefore, since g has an abelian centraliser, r may
divide only (q n − 1)(q n+1 − 1), that is r divides Φn (q)Φn+1 (q). Hence we have
at most two conjugacy classes of maximal tori with abelian centraliser.
5
Proof of Theorem 2
Let G be a finite group of orientation preserving diffeomorphisms of M , a closed
orientable connected 3-manifold which is not homeomorphicto S3 .
As noted in Section 2, there at most nine conjugacy classes of cyclic subgroups generated by a hyperelliptic rotation of order 2n , so we concentrate on
hyperelliptic subgroups whose order is not a power of two and we will prove
that there are at most six conjugacy classes of such subgroups.
Let S = {C1 , . . . , Cm } be the set of cyclic subgroups of odd prime orders that
are generated by powers of the hyperelliptic rotations of G. We recall that the
conjugate of a power of a hyperelliptic rotation is the power of a hyperelliptic
rotation too. By Remark 3 and Corollary 4, S is an algebraically hyperelliptic
collection, and actually the definition of an algebraically hyperelliptic collection
was chosen precisely to capture the behaviour of cyclic subgroups of odd prime
orders generated by powers of the hyperelliptic rotations. For i ∈ {1, . . . , m},
let pi be the prime order of Ci and π be the set including every pi . We denote by
G0 the subgroup generated by the subgroups Ci . Let A be the maximal normal
solvable subgroup of G0 of order coprime with all pi . We denote by G0 the
quotient group G0 /A and by C i the projection of Ci to G0 .
Case 1. If G0 is solvable, we are in the hypotheses of Lemma 3 which
affirms that up to conjugacy the hyperelliptic rotations of G commute. Since
they commute we can apply Lemma 2 and conclude that there are at most three
conjugacy classes of subgroups generated by hyperelliptic rotations of order not
a power of two.
Even if its proof is easier than that of the remaining situations, the solvable
case is interesting in its own right, in particular for it plays an important role
also in the proof of Theorem 3. We then summarise the conclusions in the
following proposition.
Proposition 5. Let M be a closed 3-manifold not homeomorphic to S3 . Let
G be a finite group of orientation preserving diffeomorphisms of M . If G is
solvable, it contains at most three conjugacy classes of hyperelliptic subgroups of
order not a power of two, and any two such subgroups commute up to conjugacy.
Moreover, if there are three conjugacy classes, then their orders must be pairwise
coprime.
Case 2. Suppose G0 is not solvable and it has no rotation of order 2 outside
A. Then, because of the structure of the normaliser of a hyperelliptic rotation
as described in Remark 3, G0 has no involution acting dihedrally on any Ci and
by Part 3 of Theorem 4 describing the structure of G0 , E(G0 ) is simple and,
for each pi , any Sylow pi -subgroup contains exactly two distinct conjugates of
C̄i . By Remark 3 every hyperelliptic rotation commuting with one of these
two subgroups of order pi commutes also with the other one. From this fact,
Lemma 3, and Lemma 2, it follows that γπs (G0 ) bounds from above the number
of conjugacy classes of hyperelliptic subgroups of order not a power of two (see
16
also Proposition 5). It is easy to see that γπs (G0 ) ≤ γπs (E(G0 )). By Proposition 4
we have γπs (E(G0 )) ≤ 4 and so we get the thesis in this case.
Case 3. Suppose G0 is not solvable and it has a rotation of order 2 not
contained in A. The groups containing a rotation of order two are studied in
[Mec2] where the following result was proved.
Theorem 5. [Mec2] Let D be a finite group of orientation-preserving diffeomorphisms of a closed orientable 3-manifold. Let O be the maximal normal
subgroup of odd order and E(D̃) be the layer of D̃ = D/O. Suppose that D
contains an involution which is a rotation.
1. If E(D̃) is trivial, there exists a normal subgroup H of D such that H is
solvable and D/H is isomorphic to a subgroup of A8 , the alternating group
on 8 letters.
2. If the semisimple group E(D̃) is not trivial, it has at most two components
and the factor group of D̃/E(D̃) is solvable.
Moreover if D contains a rotation of order 2 such that its projection is contained
in E(D̃), then E(D̃) is isomorphic to one of the following groups:
P SL2 (q), SL2 (q) ×Z/2 SL2 (q ′ )
where q and q ′ are odd prime powers greater than 4.
Applying Theorem 5 and Theorem 4 to G0 , we get that E(G0 ) is isomorphic
either to a subgroup of A8 , or to P SL2 (q), or to P SL2 (q) × P SL2 (q ′ ). In the
first case, there are at most three odd primes dividing the order of E(G0 ) and
the thesis follows again from Lemma 2 and Lemma 3.
In the remaining cases, we will use a solvable normal π-cover to bound
the number of conjugacy classes. We have that γπs (P SL2 (q)) ≤ 2. In fact the
upper triangular matrices form a solvable subgroup of SL2 (q) of order (q −
1)q, moreover SL2 (q) contains a cyclic subgroup of order q + 1 (see [H]). The
conjugates of the projections of these two subgroups to P SL2 (q) give a solvable
normal π-cover of P SL2 (q). It is again easy to see that, if E(G0 ) is isomorphic
to P SL2 (q), then γπs (G) ≤ 2. As above, by Lemma 3 and Lemma 2, we get a
bound of six in this case.
Finally, if E(G) is isomorphic to P SL2 (q) × P SL2 (q ′ ), then it follows from
the discussion above that γπs (E(G0 )) ≤ 4, and hence γπs (G0 ) ≤ 4. Now, by Part
1 of Theorem 4, for each pi ∈ π, pi divides the order of both P SL2 (q) and
P SL2 (q ′ ), so that the Sylow pi -subgroup of G0 has rank 2. For each Sylow
pi -subgroup Si , one can find in E(G0 ) elements of the group which normalise
Si but do not act in the same way on all of its elements. Indeed, let H be a
cyclic subgroup of order pi in P SL2 (q). If pi does not divide q, then there is
an element of order 2 in P SL2 (q) which acts dihedrally on H but commutes
with all elements of order p in P SL2 (q ′ ). If pi divides q, then q = pi since
the Sylow pi -subgroup of the component must be cyclic. Since q > 3, we have
again elements in P SL2 (q) that normalise the Sylow p-subgroup but do not
centralise it (note that in this case the structure of the normaliser is compatible
with the description given in Remark 3 only if p = q = 5). We now deduce from
Remark 7 and Corollary 3 that Si contains two subgroups conjugate to Ci .
17
Reasoning as in Case 2 we get that γπs (G0 ) bounds from above the number
of conjugacy classes of hyperelliptic subgroups of order not a power of two. This
concludes the proof.
Remark 9. We wish to stress that the three cases that appear in the proof of
Theorem 2 correspond to different types of symmetries possessed by the knots
that are branched covered by the manifold. Let M be a closed, connected,
oriented 3-manifold and let G a finite group of diffeomorphisms of M generated
by a set {ψ1 , . . . , ψk } of hyperelliptic rotations. Denote by Ki the quotient
knot F ix(ψi )/ψi , i = 1, . . . , k. If there exists a Ki which is neither strongly
invertible nor self-symmetric with respect to G, then we are in Case 1, that is,
the hyperelliptic rotations commute up to conugacy (see [Su, Thm 2.10, page
144]). If there is a Ki which is strongly invertible with respect to G, then G
contains a rotation of order 2 and we are in Case 3. Otherwise, every Ki is
self-symmetric with respect to G but not strongly invertible with respect to G,
and we are in Case 2.
The proof of Theorem 2 shows that topology and geometry can impose extra
constraints on the conditions that can be derived by the algebra alone. In this
spirit, note that if M is a (closed, connected, oriented) reducible 3-manifold and
G a finite group of diffeomorphisms of M generated by hyperelliptic rotations,
then, by the equivariant sphere theorem, G is isomorphic to a finite subgroup of
SO(3), that is cyclic, dihedral or a spherical triangular group. It follows readily,
that, up to conjugacy, at most three cyclic hyperelliptic subgroups can be contained in any such G. We know, however, that the group of diffeomorphisms of
a reducible 3-manifolds can admit arbitrarily many conjugacy classes of cyclic
hyperelliptic subgroups.
The main point here is that, generically, one expects that two hyperelliptic
rotations in the group Dif f + (M ) generate an infinite subgroup. As a consequence, Theorem 2 cannot be directly exploited to obtain bounds for nonhyperbolic manifolds and new strategies must be developped, as we will see in
the next section.
6
Proof of Theorem 3
The statement of Theorem 3 is equivalent to the following:
Theorem 6. Let M be a closed, orientable, connected, irreducible 3-manifold
which is not homeomorphic to S3 , then the group Dif f + (M ) of orientation
preserving diffeomorphisms of M contains at most six conjugacy classes of hyperelliptic subgroups of odd prime order.
6.1
Proof of Theorem 6 for Seifert manifolds
In this section we prove Proposition 6 which implies Theorem 6 for closed Seifert
fibred 3-manifolds. We also show that the assumption that the hyperelliptic
rotations have odd prime orders cannot be avoided in general by exhibiting
examples of closed Seifert fibred 3-manifolds M such that Dif f + (M ) contains
an arbitrarily large number of conjugacy classes of hyperelliptic subgroups of
odd, but not prime, orders.
18
Proposition 6. Let M be a closed Seifert fibred 3-manifold which is not homeomorphic to S3 . Then the group Dif f + (M ) of orientation preserving diffeomorphisms of M contains at most one conjugacy class of hyperelliptic subgroups
of odd prime order except if M is a Brieskorn integral homology sphere with 3
exceptional fibres. In this latter case Dif f + (M ) contains at most three non
conjugate hyperelliptic subgroups of odd prime orders.
Proof.
By hypothesis M is a cyclic cover of S3 branched over a knot, so it is orientable and a rational homology sphere by Remark 2. Notably, M cannot be
S1 × S2 nor a Euclidean manifold, except for the Hantzsche-Wendt manifold,
see [Or, Chap. 8.2]. In particular, since M is prime it is also irreducible.
Consider a hyperelliptic rotation ψ on M of odd prime order p and let K be
the image of F ix(ψ) in the quotient S3 = M/ψ by the action of ψ. The knot
K must be hyperbolic or a torus knot, otherwise its exterior would be toroidal
and have a non-trivial JSJ-collection of essential tori which would lift to a nontrivial JSJ-collection of tori for M , since the order of ψ is p > 2 (see [JS, J] and
[BS]). By the orbifold theorem (see [BoP], [CHK]), the cyclic branched cover
with order p ≥ 3 of a hyperbolic knot is hyperbolic, with a single exception
for p = 3 when K is the the figure-eight knot and M is the Hantzsche-Wendt
Euclidean manifold. But then, by the orbifold theorem and the classification of
3-dimensional christallographic groups, ψ generates the unique, up to conjugacy,
Euclidean hyperelliptic cyclic subgroup of Dif f +(M ), see for example [Dun],
[Z1].
So we can assume that M is the p-fold cyclic cover of S3 branched along a
non-trivial torus knot K of type (a, b), where a > 1 and b > 1 are coprime integers. Then M is a Brieskorn-Pham manifold M = V (p, a, b) = {z p + xa + y b =
0 with (z, x, y) ∈ C3 and |z|2 + |x|2 + |y|2 = 1}. A simple computation shows
that M admits a Seifert fibration with 3, p or p + 1 exceptional fibres and base
space S2 , see [Ko, Lem. 2], or [BoPa, Lemma 6 and proof of Lemma 7]. In
particular M has a unique Seifert fibration, up to isotopy: by [Wa], [Sco] and
[BOt] the only possible exception with base S2 and at least 3 exceptional fibres
is the double of a twisted I-bundle, which is not a rational homology sphere,
since it fibers over the circle. We distinguish now two cases:
Case 1: The integers a and b are coprime with p, and there are three singular
fibres of pairwise relatively prime orders a, b and p. By the orbifold theorem
any hyperelliptic rotation of M of order > 2 is conjugate into the circle action
S1 ⊂ Dif f + (M ) inducing the Seifert fibration, hence the uniqueness of the
Seifert fibration, up to isotopy, implies that M admits at most 3 non conjugate
hyperelliptic groups of odd prime orders belonging to {a, b, p}. Indeed M is a
Brieskorn integral homology sphere, see [BPZ].
Case 2: Either a = p and M has p singular fibres of order b, or a = a′ p with
a′ > 1, and M has p singular fibres of order b and one extra singular fibre of
order a′ . In both situations, there are p ≥ 3 exceptional fibres of order b which
are cyclically permuted by the hyperelliptic rotation ψ. As before, M has a
unique Seifert fibration, up to isotopy. Therefore, up to conjugacy, ψ is the only
hyperelliptic rotation of order p on M , and by the discussion above M cannot
admit a hyperelliptic rotation of odd prime order q 6= p.
19
Remark 10. The requirement that the rotations are hyperelliptic is essential in
the proof of Proposition 6. The Brieskorn homology sphere Σ(p1 , . . . , pn ), n ≥
4, with n ≥ 4 exceptional fibres admits n rotations of pairwise distinct prime
orders but which are not hyperelliptic.
The hypothesis that the orders of the hyperelliptic rotations are 6= 2 cannot
be avoided either.
Indeed, Montesinos’s construction of fibre preserving hyperelliptic involutions on Seifert fibered rational homology spheres [Mon1], [Mon2], (see also
[BS, Appendix A], [BZH, Chapter 12]), shows that for any given integer n there
are infinitely many closed orientable Seifert fibred 3-manifolds with at least n
conjugacy classes of hyperelliptic rotations of order 2.
On the other hand, the hypothesis that the orders are odd primes is sufficient
but not necessary: A careful analysis of the Seifert invariants shows that if
M 6= S3 is a Seifert rational homology sphere, then M can be the cyclic branched
cover of a knot in S3 of order > 2 in at most three ways.
The hypotheses of Proposition 6 cannot be relaxed further, though: Proposition 7 below shows that there exist closed 3-dimensional circle bundles with
arbitrarily many conjugacy classes of hyperelliptic rotations of odd, but not
prime, orders.
Proposition 7. Let N be an odd prime integer. For any integer 1 ≤ q < N2
the Brieskorn-Pham manifold M = V ((2q + 1)(2(N −q) + 1), 2q + 1, 2(N −q) + 1)
is a circle bundle over a closed surface of genus g = 2N −1 with Euler class ±1.
Hence, up to homeomorphism (possibly reversing the orientation), M depends
only on the integer N and admits at least N 2−1 conjugacy classes of hyperelliptic
groups of odd orders.
Proof. We remark first that the integers q and N − q are relatively prime,
because N is prime. If k is a common prime divisor of 2q + 1 and 2(N −q) + 1, by
the Bezout identity we have 21 = 2aq+b(N −q) ≡ (−1)a+b mod k which implies
that k = 3. But then (−1)q ≡ (−1)(N −q) ≡ −1 mod 3 and thus (−1)N ≡ 1
mod 3 which is impossible since N is odd. Hence 2q + 1 and 2(N −q) + 1 are
relatively prime.
So the Brieskorn-Pham manifold M is the (2q + 1)(2(N −q) + 1)-fold cyclic
cover of S3 branched over the torus knot Kq = T (2q + 1, 2(N −q) + 1). It is
obtained by Dehn filling the (2q + 1)(2(N −q) + 1)-fold cyclic cover of the exterior
of the torus knot Kq along the lift of its meridian. The (2q + 1)(2(N −q) + 1)-fold
cyclic cover of the exterior of Kq is a trivial circle bundle over a once punctured
surface of genus g = 2N − 1. On the boundary of the torus-knot exterior the
algebraic intersection between a meridian and a fibre of the Seifert fibration of
the exterior is ±1 (the sign depends on a choice of orientation, see for example
[BZH, Chapter 3]). So on the torus boundary of the (2q + 1)(2(N −q) + 1)-fold
cyclic cover the algebraic intersection between the lift of a meridian of the torus
knot and an S1 -fiber is again ±1. Hence the circle bundle structure of the
(2q + 1)(2(N −q) + 1)-fold cyclic cover of the exterior of the torus knot Kq can be
extended with Euler class ±1 to the Dehn filling along the lift of the meridian.
So M is a circle bundle over a closed surface of genus g = 2N −1 with Euler class
±1.
Since the torus knots Kq = T (2q +1, 2(N −q) +1) are pairwise inequivalent for
1 ≤ q ≤ N2−1 , the hyperelliptic subgroups corresponding to the (2q +1)(2(N −q) +
20
1)-fold cyclic branched covers of the knots Kq are pairwise not conjugate in
Dif f + (M ).
Remark 11. Note that the Seifert manifolds M and their hyperelliptic rotations
constructed in Proposition 7 enjoy the following properties: If N > 8, then no
hyperelliptic rotation can commute up to conjugacy with all the remaining ones
(see Proposition 5 and [BoPa, Theorem 2]). If N > 14 no finite subgroup
of Dif f + (M ) can contain up to conjugacy all hyperelliptic rotations of M ,
according to Theorem 2.
6.2
Reduction to the finite group action case
The fact that Theorem 6 implies Corollary 1 follows from the existence of a decomposition of a closed, orientable 3-manifold M as a connected sum of prime
manifolds and the observation that a hyperelliptic rotation on M induces a hyperelliptic rotation on each of its prime summands. A 3-manifold admitting
a hypereliptic rotation must be a rational homology sphere, and so M cannot
have S2 × S1 summands. Hence all prime summands are irreducible and at least
one is not homeomorphic to S3 , since M itself is not homeomorphic to S3 . This
is enough to conclude.
The remaining of this section is devoted to the proof that Theorem 2 implies
Theorem 6.
We prove the following proposition:
Proposition 8. If M is a closed, orientable, irreducible 3-manifold such that
there are k ≥ 7 conjugacy classes of hyperelliptic subgroups of Dif f + (M ) whose
order is an odd prime, then M is homeomorphic to S3 .
Proof.
Let M be a closed, orientable, irreducible 3-manifold such that Dif f + (M )
contains k ≥ 7 conjugacy classes of hyperelliptic subgroups of odd prime orders.
According to the orbifold theorem (see [BoP], [BMP], [CHK]), a closed orientable irreducible manifold M admitting a rotation has geometric decomposition. This means that M can be split along a (possibly empty) finite collection
of π1 -injective embedded tori into submanifolds carrying either a hyperbolic or
a Seifert fibered structure. This splitting along tori is unique up to isotopy and
is called the JSJ-decomposition of M , see for example [NS], [BMP, chapter 3].
In particular, if its JSJ-decomposition is trivial, M admits either a hyperbolic
or a Seifert fibred structure.
First we see that M cannot be hyperbolic. Indeed, if the manifold M is
hyperbolic then, by the orbifold theorem, any hyperelliptic rotation is conjugate into the finite group Isom+ (M ) of orientation preserving isometries of M .
Hence, applying Theorem 2 to G = Isom+ (M ), we see that k ≤ 6 against the
hypothesis.
If the manifold M is Seifert fibred, it follows readily from Proposition 6 of
the previous section that M = S3 . So we are left to exclude the case where the
JSJ-decomposition of M is not empty.
Consider the JSJ-decomposition of M : each geometric piece admits either
a complete hyperbolic structure with finite volume or a Seifert fibred product
21
structure with orientable base. Moreover, the geometry of each piece is unique,
up to isotopy.
Let Ψ = {ψ1 , . . . , ψk , k ≥ 7} be the set of hyperelliptic rotations which
generate non conjugate cyclic subgroups in Dif f + (M ). By the orbifold theorem
[BoP], [BMP], [CHK], after conjugacy, one can assume that each hyperelliptic
rotation preserves the JSJ-decomposition, acts isometrically on the hyperbolic
pieces, and respects the product structure on the Seifert pieces. We say that
they are geometric.
Let Γ be the dual graph of the JSJ-decomposition: it is a tree, for M is a
rational homology sphere (in fact, the dual graph of the JSJ-decomposition for a
manifold which is the cyclic branched cover of a knot is always a tree, regardless
of the order of the covering). Let H ⊂ Dif f + (M ) be the group of diffeomorphisms of M generated by the set Ψ of geometric hyperelliptic rotations. By
[BoPa, Thm 1], there is a subset Ψ0 ⊂ Ψ of k0 ≥ 4 hyperelliptic rotations with
pairwise distinct odd prime orders, say Ψ0 = {ψi , i = 1, . . . , k0 }.
Let HΓ denote the image of the induced representation of H in Aut(Γ). Since
rotations of finite odd order cannot induce an inversion on any edge of Γ, the
finite group HΓ must fix pointwise a non-empty subtree Γf of Γ.
The idea of the proof is now analogous to the ones in [BoPa] and [BPZ]: we
start by showing that, up to conjugacy, the k0 ≥ 4 hyperelliptic rotations with
pairwise distinct odd prime orders can be chosen to commute on the submanifold Mf of M corresponding to the subtree Γf . We consider then the maximal
subtree corresponding to a submanifold of M on which these hyperelliptic rotations commute up to conjugacy and prove that such subtree is in fact Γ. Then
the conclusion follows by applying Lemma 2.
The first step of the proof is achieved by the following proposition:
Proposition 9. The hyperelliptic rotations in Ψ0 commute, up to conjugacy in
Dif f + (M ), on the submanifold Mf of M corresponding to the subtree Γf .
Proof.
Since the hyperelliptic rotations in Φ have odd orders, either Γf contains an
edge, or it consists of a single vertex. We shall analyse these two cases.
Case 1: Mf contains an edge.
Claim 5. Assume that Γf contains an edge and let T denote the corresponding
torus. The hyperelliptic rotations in Ψ0 commute, up to conjugacy in Dif f + (M ),
on the geometric pieces of M adjacent to T .
Proof.
The geometric pieces adjacent to T are left invariant by the hyperelliptic
rotations in Ψ0 , since their orders are odd. Let V denote one of the two adjacent
geometric pieces: each hyperelliptic rotation acts non-trivially on V with odd
prime order. We distinguish two cases according to the geometry of V .
V is hyperbolic. In this case all rotations act as isometries and leave a cusp
invariant. Since their order is odd, the rotations must act as translations along
horospheres, and thus commute.
Note that, even in the case where a rotation has order 3, its axis cannot
meet a torus of the JSJ-decomposition of M for each such torus is separating
and cannot meet the axis in an odd number of points.
22
V is Seifert fibred. In this case we can assume that the hyperelliptic rotations
in Ψ preserve the Seifert fibration with orientable base. Since their orders are
odd and prime, each one preserves the orientation of the fibres and of the base.
The conjugacy class of a fiber-preserving rotation of V with odd prime order
depends only on its combinatorial behaviour, i.e. its translation action along the
fibre and the induced permutation on cone points and boundary components of
the base. In particular, two geometric rotations with odd prime order having
the same combinatorial data are conjugate via a diffeomorphism isotopic to the
identity.
Since the hyperelliptic rotations in Ψ0 have pairwise distinct odd prime
orders, an analysis of the different cases described in Lemma 4 below shows that
at most one among these hyperelliptic rotations can induce a non-trivial action
on the base of the fibration, and thus the remaining ones act by translation
along the fibres and induce the identity on the base. Since the translation
along the fibres commutes with every fiber-preserving diffeomorphism of V , the
hyperelliptic rotations in Ψ0 commute on V .
Lemma 4 describes the Seifert fibred pieces of a manifold admitting a hyperelliptic rotation of odd prime order, as well as the action of the rotation on
the pieces. Its proof can be found in [BoPa, Lemma 6 and proof of Lemma 7],
see also [Ko, Lem. 2].
Lemma 4. Let M be an irreducible 3-manifold admitting a non-trivial JSJdecomposition. Assume that M admits a hyperelliptic rotation of prime odd
order p. Let V be a Seifert piece of the JSJ-decomposition for M . Then the
base B of V can be:
1. A disc with 2 cone points. In this case either the rotation freely permutes p
copies of V or leaves V invariant and acts by translating along the fibres.
2. A disc with p cone points. In this case the rotation leaves V invariant
and cyclically permutes the singular fibres, while leaving a regular one
invariant.
3. A disc with p + 1 cone points. In this case the rotation leaves V and a singular fibre invariant, while cyclically permuting the remaining p singular
fibres.
4. An annulus with 1 cone point. In this case either the rotation freely permutes p copies of V or leaves V invariant and acts by translating along
the fibres.
5. An annulus with p cone points. In this case the rotation leaves V invariant
and cyclically permutes the p singular fibres.
6. A disc with p − 1 holes and 1 cone point. In this case the rotation leaves V
invariant and cyclically permutes all p boundary components, while leaving
invariant the only singular fibre and a regular one.
7. A disc with k holes, k ≥ 2. In this case either the rotation freely permutes
p copies of V or leaves V invariant. In this latter case either the rotation
acts by translating along the fibres, or k = p − 1 and the rotation permutes
all the boundary components (while leaving invariant two fibres), or k = p
23
and the rotation permutes p boundary components, while leaving invariant
the remaining one and a regular fibre.
We conclude that the rotations in Ψ0 can be chosen to commute on the
submanifold Mf of M corresponding to Γf by using inductively at each edge of
Γf the gluing lemma below (see [Lemma 6][BPZ]). We give the proof for the
sake of completeness.
Lemma 5. If the rotations preserve a JSJ-torus T then they commute on the
union of the two geometric pieces adjacent to T .
Proof.
Let V and W be the two geometric pieces adjacent to T . By Claim 5, after
conjugacy in Dif f + (M ), the rotations in Ψ0 commute on V and W . Since
they have pairwise distinct odd prime orders, their restrictions on V and W
generate two cyclic groups of the same finite odd order. Let gV and gW be
generators of these two cyclic groups. Since they have odd order, they both act
by translation on T . We need the following result about the slope of translation
for such periodic transformation of the torus:
Claim 6. Let ψ be a periodic diffeomorphism of the product T 2 × [0, 1] which
is isotopic to the identity and whose restriction to each boundary torus T × {i},
i = 0, 1, is a translation with rational slopes α0 and α1 in H1 (T 2 ; Z). Then
α0 = α1 .
Proof.
By Meeks and Scott [MS, Thm 8.1], see also [BS, Prop. 12], there is a
Euclidean product structure on T 2 × [0, 1] preserved by ψ such that ψ acts by
translation on each fiber T × {t} with rational slope αt . By continuity, the
rational slopes αt are constant.
Now the the following claim shows that the actions of gW and gV can be
glued on T .
Claim 7. The translations gV |T and gW |T have the same slope in H1 (T 2 ; Z).
Proof.
Let Ψ0 = {ψi , i = 1, . . . , k0 }. Let pi the order of ψi and qi = Πj6=i pj . Then
the slopes αV and αW of gV |T and gW |T verify: qi αV = qi αW for i = 1, ..., k0 ,
by applying Claim 6 to each ψi . Since the GCD of the qi is 1, it follows that
αV = αW .
This finishes the proof of Lemma 5 and of Proposition 9 when Mf contains
an edge.
To complete the proof of Proposition 9 it remains to consider the case where
Γf is a single vertex.
Case 2: Mf is a vertex.
Claim 8. Assume that Γf consists of a single vertex and let V denote the
corresponding geometric piece. Then the hyperelliptic rotations in Ψ0 commute
on V , up to conjugacy in Dif f + (M ).
24
Proof.
We consider again two cases according to the geometry of V .
The case where V is Seifert fibred follows once more from Lemma 4.
We consider now the case where V is hyperbolic.
In this case, the hyperelliptic rotations in Ψ act non-trivially on V by isometries of odd prime orders. The restriction H|V ⊂ Isom+ (V ) of the action of the
subgroup H that they generate in Dif f + (M ) is finite.
If the action on V of the cyclic subgroups generated by two of the hyperelliptic rotations in Ψ are conjugate in H|V , one can conjugate the actions in
Dif f + (M ) to coincide on V , since any diffeomorphism in H|V extends to M .
Then by [BoPa, Lemma 10] these actions must coincide on M , contradicting the
hypothesis that the conjugacy classes of cyclic subgroups generated by the hyperelliptic rotations in Ψ are pairwise distinct in Dif f + (M ). Hence, the cyclic
subgroup generated by the k ≥ 7 hyperelliptic rotations in Ψ are pairwise not
conjugate in the finite group H|V ⊂ Isom+ (V ).
Since the dual graph of the JSJ-decomposition of M is a tree, a boundary
torus T ⊂ ∂V is separating and bounds a component UT of M \ int(V ). Since,
by hypothesis, Γf consists of a single vertex, no boundary component T is
setwise fixed by the finite group H|V . This means that there is a hyperelliptic
rotation ψi ∈ Ψ of odd prime order pi such that the orbit of UT under ψi is the
disjoint union of pi copies of UT . In particular UT projects homeomorphically
onto a knot exterior in the quotient S3 = M/ψi . Therefore on each boundary
torus T = ∂UT ⊂ ∂V , there is a simple closed curve λT , unique up to isotopy,
that bounds a properly embedded incompressible and ∂-incompressible Seifert
surface ST in the knot exterior UT .
By pinching the surface ST onto a disc D2 , in each component UT of M \
int(V ), one defines a degree-one map π : M −→ M ′ , where M ′ is the rational
homology sphere obtained by Dehn filling each boundary torus T ⊂ V along
the curve λT .
For each hyperelliptic rotation ψi in Ψ, of odd prime order pi , the ψi -orbit
of each component UT of M \ int(V ) consists of either one or pi elements. As a
consequence, by [Sa], ψi acts equivariantly on the set of isotopy classes of curves
λT ⊂ ∂V . Hence, each ψi extends to periodic diffeomorphism ψi′ of order pi on
M ′ . Moreover, M ′ is a Z/pi -homology sphere, since so is M and π : M −→ M ′
is a degree-one map. According to Smith theory, if F ix(ψ ′ ) is non-empty on
M ′ , then ψi′ is a rotation on M ′ . To see that F ix(ψ ′ ) 6= ∅ on M ′ it suffices to
observe that either F ix(ψ) ⊂ V or ψi is a rotation of some UT ; in this latter
case, ψi′ must have a fixed point on the disc D2 onto which the surface ST is
pinched. To show that ψi′ is hyperelliptic it remains to show that the quotient
M ′ /ψi′ is homeomorphic to S3 .
Since ψi acts equivariantly on the components UT of M \ int(V ) and on the
set of isotopy classes of curves λT ⊂ ∂V , the quotient S3 = M/ψi is obtained
from the compact 3-manifold V /ψi by gluing knot exteriors (maybe solid tori)
to its boundary components, in such a way that the boundaries of the Seifert
surfaces of the knot exteriors are glued to the curves λT /ψi ⊂ ∂V /ψi .
In the same way, the rotation ψi′ acts equivariantly on the components M ′ \
int(V ) and on the set of isotopy classes of curves λT ⊂ ∂V . By construction,
these components are solid tori, and either the axis of the rotation is contained
in V or there exists a unique torus T ∈ ∂V such that the solid torus glued to
T to obtain M ′ contains the axis. In the latter case, by [EL, Cor. 2.2], the
25
rotation ψi′ preserves a meridian disc of this solid torus and its axis is a core
of WT . It follows that the images in the quotient M ′ /ψi′ of the the solid tori
glued to ∂V are again solid tori. Hence M ′ /ψi′ is obtained from S3 by replacing
each components of S3 \ V /ψi by a solid torus, in such way that boundaries
of meridian discs of the solid tori are glued to the curves λT /ψi′ ⊂ ∂V /ψi′ . It
follows that M ′ /ψi′ is again S3 .
So far we have constructed a closed orientable 3-manifold M ′ with a finite
subgroup of orientation preserving diffeomorphisms HV that contains at least
seven conjugacy classes of hyperelliptic subgroups of odd prime orders. Theorem 2 implies that M ′ must be S3 , and thus by the orbifold theorem [BLP] HV
is conjugate to a finite subgroup of SO(4). In particular the subgroup H0 ⊂ HV
generated by the subset Ψ0 of at least 4 hyperelliptic rotations with pairwise
distinct odd prime orders must be solvable. Therefore, by Proposition 5 the induced rotations commute on M ′ and, by restriction, the hyperelliptic rotations
in Ψ0 commute on V .
In the final step of the proof we extend the commutativity on Mf to the
whole manifold M . The proof of this step is analogous to the one given in
[BPZ], since the proof there was not using the homology assumption. We give
the argument for the sake of completeness.
Proposition 10. The k0 ≥ 4 hyperelliptic rotations in Ψ0 commute, up to
conjugacy in Dif f + (M ), on M .
Proof.
Let Γc be the largest subtree of Γ containing Γf , such that, up to conjugacy
in Dif f + (M ), the rotations in Ψ0 commute on the corresponding invariant
submanifold Mc of M . We shall show that Γc = Γ. If this is not the case, we
can choose an edge contained in Γ corresponding to a boundary torus T of Mc .
Denote by UT the submanifold of M adjacent to T but not containing Mc and
by VT ⊂ UT the geometric piece adjacent to T .
Let H0 ⊂ Dif f + (M ) be the group of diffeomorphisms of M generated by
the set of geometric hyperelliptic rotations Ψ0 = {ψi , i = 1, . . . , k0 }. Since the
rotations in Ψ0 commute on Mc and have pairwise distinct odd prime orders,
the restriction of H0 on on Mc is a cyclic group with order the product of the
orders of the rotations. Since Γf ⊂ Γc , the H0 -orbit of T cannot be reduced to
only one element. Moreover each rotation ψ ∈ Ψ0 either fixes T or acts freely
on the orbit of T since its order is prime.
If no rotation in Ψ0 leaves T invariant, the H0 -orbit of T contains as many
elements as the product of the orders of the rotations, for they commute on Mc .
In particular, only the identity (which extends to U ) stabilises a torus in the
H0 -orbit of T . Note that all components of ∂Mc that are in the H0 -orbit of T
bound a manifold homeomorphic to UT .
Since the rotation ψi acts freely on the H0 -orbit of UT , UT is a knot exterior
in the quotient M/ψi = S3 . Hence there is a well defined meridian-longitude
system on T = ∂UT and also on each torus of the H0 -orbit of T . This set of
meridian-longitude systems is cyclically permuted by each ψi and thus equivariant under the action of H0 .
Let Mc /H0 be the quotient of Mc by the induced cyclic action of H0 on Mc .
Then there is a unique boundary component T ′ ⊂ ∂(Mc /H0 ) which is the image
of the H0 -orbit of T . We can glue a copy of the knot exterior UT to Mc /H0 along
26
T ′ by identifying the image of the meridian-longitude system on ∂UT with the
projection on T ′ of the equivariant meridian-longitude systems on the H0 -orbit
of T . Denote by N the resulting manifold. For all i Q
= 1, . . . , k0 , consider the
cyclic (possibly branched) cover of N of order qi = j6=i pj which is induced
by the cover πi : Mc /ψi −→ Mc /H0 . This makes sense because π1 (T ′ ) ⊂
πi ∗ (π1 (Mc /ψi )). Call Ñi the total space of such covering. By construction it
follows that Ñi is the quotient (Mc ∪ H0 · UT )/ψi . This implies that the ψi ’s
commute on Mc ∪ H0 · UT contradicting the maximality of Γc .
We can thus assume that some rotations fix T and some do not. Since all
rotations commute on Mc and have pairwise distinct odd prime orders, we see
that the orbit of T consists of as many elements as the product of the orders
of the rotations which do not fix T and each element of the orbit is fixed by
the rotations which leave T invariant. The rotations which fix T commute on
the orbit of VT according to Claim 5 and Lemma 5, and form a cyclic group
generated by, say, γ. The argument for the previous case shows that the rotations acting freely on the orbit of T commute on the orbit of UT and thus on
the orbit of VT , and form again a cyclic group generated by, say, η. To reach
a contradiction to the maximality of Mc , we shall show that γ, after perhaps
some conjugacy, commutes with η on the H0 -orbit of VT , in other words that
γ and ηγη −1 coincide on H0 · VT . Since η acts freely and transitively on the
H0 -orbit of VT there is a natural and well-defined way to identify each element
of the orbit H0 · VT to VT itself. Note that this is easily seen to be the case if
VT is hyperbolic: this follows from Claim 5 and Claim 6. We now consider the
case when VT is Seifert fibred.
Claim 9. Assume that VT is Seifert fibred and that the restriction of γ induces
a non-trivial action on the base of VT . Then γ induces a non-trivial action on
the base of each component of the H0 -orbit of VT . Moreover, up to conjugacy
on H0 · VT \ VT by diffeomorphisms which are the identity on H0 · T and extend
to M , we can assume that the restrictions of γ to these components induce the
same permutation of their boundary components and the same action on their
bases.
Proof.
By hypothesis γ and ηγη −1 coincide on ∂Mc . The action of γ on the base of
VT is non-trivial if and only if its restriction to the boundary circle of the base
corresponding to the fibres of the torus T is non-trivial. Therefore the action of
γ is non-trivial on the base of each component of H0 · VT .
By Lemma 4 and taking into account the fact that VT is a geometric piece
in the JSJ-decomposition of the knot exterior UT , the only situation in which
the action of γ on the base of VT is non-trivial is when the base of VT consists of
a disc with p holes, where p is the order of one of the rotations that generate γ.
Moreover, the restriction of γ to the elements of H0 ·VT cyclically permutes their
boundary components which are not adjacent to Mc . Up to performing Dehn
twists, along vertical tori inside the components of H0 · VT \ VT , which permute
the boundary components, we can assume that the restriction of γ induces the
same cyclic permutations on the boundary components of each element of the
orbit H0 · VT . We only need to check that Dehn twists permuting two boundary
components extend to the whole manifold M . This follows from the fact that
the manifolds adjacent to these components are all homeomorphic and that
Dehn twist act trivially on the homology of the boundary.
27
Since the actions of the restrictions of γ on the bases of the elements of
H0 · VT are combinatorially equivalent, after perhaps a further conjugacy by an
isotopy, the different restrictions can be chosen to coincide on the bases.
By Claim 5 and Claim 9 we can now deduce that the restrictions of γ and
ηγη −1 to the H0 -orbit of VT commute, up to conjugacy of γ which is the identity
on the H0 -orbit of T . Since γ and ηγη −1 coincide on this H0 -orbit of T , we can
conclude that they coincide on the H0 -orbit of VT . This finishes the proof of
Proposition 10.
Since there are at least four hyperelliptic rotations with paiwise distinct
odd prime orders in Ψ0 , Proposition 8 is consequence of Proposition 10 and
Proposition 1, like in the solvable case.
Remark 12. As we have seen, the strategy to prove that an irreducible manifold
M with non-trivial JSJ-decomposition cannot admit more than six conjugacy
classes of hyperelliptic subgroups of odd prime order inside Dif f + (M ) consists
in modifying by conjugacy any given set of hyperelliptic rotations so that the
new hyperelliptic rotations commute pairwise. Note that this strategy cannot
be carried out in general if the orders are not pairwise coprime (see, for instance,
[BoPa, Section 4.1] where the case of two hyperelliptic rotations of the same odd
prime order, generating non conjugate subgroups, is considered). Similarly, for
hyperelliptic rotations of arbitrary orders > 2 lack of commutativity might arise
on the Seifert fibred pieces of the decomposition, as it does for the Seifert fibred
manifolds constructed in Proposition 7.
7
Appendix: non-free finite group actions on rational homolgy spheres
In this section we will show that every finite group G admits a faithful action
by orientation preserving diffeomorphisms on some rational homology sphere so
that some elements of G have non-empty fixed-point sets.
Cooper and Long’s construction of G-actions on rational homolgy spheres in
[CL] consists in starting with a G-action on some 3-manifold and then modifying
the original manifold, notably by Dehn surgery, so that the new manifold inherits
a G-action but has “smaller” rational homology. Since their construction does
not require that the G-action is free, it can be used to prove the existence of
non-free G-actions. We will thus start by exhibiting non-free G-actions on some
3-manifold before pointing out what need to be taken into account in this setting
in order for Cooper and Long’s construction to work.
Since every (non-trivial) cyclic group acts as a group of rotations of S3 , for
simplicity we will assume that G is a finite non-cyclic group.
Claim 10. Let G be a finite non-cyclic group. There is a closed, connected,
orientable 3-manifold M on which G acts faitfully, by orientation preserving
diffeomorphisms so that there are g ∈ G \ {1} with the property that F ix(g) 6= ∅.
Proof.
Let k ≥ 2 and let {gi }1≤i≤k+1 be a system of generators for G satisfying the
following requirements:
• for all 1 ≤ i ≤ k + 1, the order of gi is ni ≥ 2;
28
• gk+1 = g1 g2 . . . gk .
Since G is not cyclic these conditions are not difficult to fulfill, and actually
they can be fulfilled even in the case when G is cyclic for an appropriate choice
of the set {gi }1≤i≤k+1 .
Consider the free group of rank k that we wish to see as the fundamental
group of a (k + 1)-punctured 2-sphere: each generator xi corresponds to a loop
around a puncture of the sphere so that a loop around the k + 1st puncture
corresponds to the element x1 x2 . . . xk . Build an orbifold O by compactifying
the punctured-sphere with cone points so that the ith puncture becomes a cone
point of order ni . The resulting orbifold has (orbifold) fundamental group with
the following presentation:
−1
i.
hx1 , x2 , . . . , xk , xk+1 | {xni i }1≤i≤k+1 , x1 . . . xk xk+1
Clearly this group surjects onto G. Such surjection gives rise to an orbifold
covering Σ −→ O, where Σ is an orientable surface on which G acts in such a
way that each element gi has non-empty fixed-point set. One can consider the
3-manifold Σ × S1 : the action of G on Σ extends to a product action of G on
Σ × S1 which is trivial on the S1 factor.
To be able to repeat Cooper and Long’s construction it is now easy to observe
that it is always possible to choose G-equivariant families of simple closed curves
in M so that they miss the fixed-point sets of elements of G and either their
homology classes generate H1 (M ; Q) (so that the hypothesis of [CL, Lemma
2.3] are fulfilled when we choose X to be the exterior of such families) or the
family is the G-orbit of a representative of some prescribed homology class (as
in the proof of [CL, Proposition 2.5]).
Acknowledgement The authors are indebted to M. Broué for valuable discussions on the topics of the paper.
References
[A]
M. Aschbacher, Finite groups theory, Cambridge studies in advanced mathematics
10, Cambridge University Press, 1986.
[BMP]
M. Boileau, S. Maillot, and J. Porti, Three-dimensional orbifolds and their geometric structures, Panoramas et Synthèses 15, Société Mathématique de France,
Paris, 2003.
[BLP]
M. Boileau, B. Leeb and J. Porti, Geometrization of 3-dimensional orbifolds,
Annals of Math. 162, (2005), 195-290.
[BOt]
M. Boileau and J.P. Otal. Scindements de Heegaard et groupe des homéotopies
des petites variétés de Seifert, Invent. Math. 106, (1991), 85-107.
[BoPa]
M. Boileau and L. Paoluzzi, On cyclic branched coverings of prime knots, J. Topol.
1, (2008), 557-583.
[BPZ]
M. Boileau, L. Paoluzzi and B. Zimmermann, A characterisation of S3 among
homology spheres, The Zieschang Gedenkschrift. Geom. Topol. Monogr. 14, Geom.
Topol. Publ., Coventry, (2008), 83-103.
[BoP]
M. Boileau and J. Porti, Geometrization of 3-orbifolds of cyclic type, Astérisque
Monograph, 272, 2001.
[BS]
F. Bonahon and L. Siebenmann, The characteristic splitting of irreducible compact
3-orbifolds, Math. Math. 278, (1987), 441-479.
29
[BZH]
G. Burde and H. Zieschang, M. Heusener, Knots, Third edition, De Gruyter Studies
in Mathematics, 5. De Gruyter, Berlin, 2014.
[CL]
D. Cooper and D. D. Long, Free actions of finite groups on rational homology
3-spheres, Topology Appl. 101, (2000), 143148.
[CHK]
D. Cooper, C. Hodgson, and S. Kerckhoff, Three-dimensional orbifolds and conemanifolds, MSJ Memoirs 5, 2000.
[DL]
J. Dinkelbach and B. Leeb, Equivariant Ricci flow with symmetry and applications
to finite group actions on 3-manifolds, Geom. Topol. 13, (2009), 11291173.
[Dun]
W. D. Dunbar, Geometric Orbifolds, Rev. Mat. Univ. Comp. Madrid, 1, 1998,
67-99.
[EL]
A. L. Edmonds and C. Livingston, Group actions on fibered three-manifolds,
Comm. Math. Helv. 58, (1983), 529-542.
[Gr]
J. E. Greene, Lattices, graphs, and Conway mutation. Invent. Math. 192 (2013),
no. 3, 717750.
[Go]
C. McA. Gordon, Some aspects of classical knot theory, Knot Theory, Proceedings,
Plans-sur-Bex, Switzerland (J.C. Hausmann ed.), Lect. Notes Math. 685, (1977),
Springer-Verlag, 1-60.
[GLS]
D. Gorenstein, R. Lyons, and R. Solomon, The classification of the finite simple
groups, Number 3, Math. Surveys Monogr. 40.3, Amer. Math. Soc., Providence,
RI, 1998.
[Hil]
J. Hillman, Links with infinitely many semifree periods are trivial, Arch. Math.
42, (1984), 568-572.
[H]
B. Huppert, Endlichen Gruppen I, Springer-Verlag, New York, 1968.
[KL]
B. Kleiner and J. Lott. Local Collapsing, Orbifolds, and Geometrization,
Astérisque Monograph 365, 2014.
[JS]
W. H. Jaco and P. B. Shalen, Seifert fibred spaces in 3-manifolds, Mem. Amer.
Math. Soc. 220, 1979.
[J]
K. Johannson, Homotopy equivalence of 3-manifolds with boundary, Lecture Notes
in Math. 761, Springer-Verlag, Berlin, 1979.
[Ka]
A. Kawauchi, Topological imitations and Reni-Mecchia-Zimmermann’s conjecture,
Kyungpook Math. J. 46 (2006), 19.
[Ko]
S. Kojima, Determining knots by branched covers, in Low Dimensional Topology
and Kleinian groups, London Math. Soc. Lecture Note Ser. 112, Cambridge Univ.
Press (1986), 193-207.
[Mec1]
M. Mecchia, How hyperbolic knots with homeomorphic cyclic branched coverings
are related, Topology Appl. 121, (2002), 521-533.
[Mec2]
M. Mecchia, Finite groups acting on 3-manifolds and cyclic branched coverings
of knots, The Zieschang Gedenkschrift. Geom. Topol. Monogr. 14, Geom. Topol.
Publ., Coventry, (2008), 393-416.
[MR]
M. Mecchia and M. Reni, Hyperbolic 2-fold branched coverings of links and their
quotients, Pacific J. Math. 202, (2002), 429447.
[MZ]
M. Mecchia and B. Zimmermann, On finite groups acting on Z2 -homology 3spheres, Math. Z. 248, (2004), 675-693.
[MSY]
W. H. Meeks, L. Simon, and S.-T. Yau, Embedded minimal surfaces, exotic spheres,
and manifolds with positive Ricci curvature, Ann. of Math. 116 (1983), 621659.
[MS]
W. H. Meeks and P. Scott, Finite group actions on 3-manifolds, Invent. Math. 86,
(1986), 287-346.
[Mon1]
J. M. Montesinos, Variedades de Seifert que son recubridores ciclicos ramificados
de dos hojas, Bol. Soc. Mat. Mexicana 18, (1973), 1-32.
[Mon2]
J. M. Montesinos, Revêtements ramifiés de noeuds, espaces fibrés de Seifert et
scindements de Heegaard, Publicaciones del Seminario Mathematico Garcia de
Galdeano, Serie II, Seccion 3, (1984).
[MonW]
J. M. Montesinos and W. Whitten Constructions of two-fold branched covering
spaces. Pac. J. Math. 125, 415446 (1986)
30
[NS]
W. Neumann and G. Swarup, Canonical decompositions of 3-manifolds, Geom.
Topol. 1, (1998), 21-40.
[Or]
P. Orlik, Seifert manifolds, Lecture Notes in Mathematics 291, Springer, 1972.
[P]
L. Paoluzzi, On hyperbolic type involutions, Rend. Istit. Mat. Univ. Trieste 32
suppl. 1, (2001), 257-288.
[Re]
M. Reni, On π-hyperbolic knots with the same 2-fold branched coverings, Math.
Ann. 316, (2000), no. 4, 681697
[RZ]
M. Reni and B. Zimmermann, Hyperbolic 3-manifolds as cyclic branched coverings,
Comment. Math. Helv. 76, (2001), 300-313.
[Sa]
M. Sakuma, Uniqueness of symmetries of knots, Math. Z. 192, (1986), 225-242.
[Sco]
P. Scott, Homotopy implies isotopy for some Seifert fibre spaces, Topology 24,
(1985), 341-351.
[Su]
M. Suzuki, Group theory II, Grundlehren Math. Wiss. 248, Springer-Verlag, New
York, 1982.
[V]
O. Ja. Viro Nonprojecting isotopies and knots with homeomorphic coverings, Zap.
Nau. Semin. POMI 66, 133147, 207208, Studies in topology, II (1976)
[Wa]
F. Waldhausen, Eine Klasse von 3-dimensionalen Mannigfaltigkeiten. I, II, Invent.
Math. 3, (1967), 308333; ibid. 4 (1967), 87117.
[Z1]
B. Zimmermann, On the Hantzsche-Wendt Manifold, Mh. Math. 110, (1990) ,
321-327.
[Z2]
B. Zimmermann, On hyperbolic knots with homeomorphic cyclic branched coverings, Math. Ann. 311, (1998), 665-673.
Aix-Marseille Université, CNRS, Centrale Marseille, I2M, UMR 7373, 13453
Marseille, France
[email protected]
Dipartimento di Matematica e Fisica “Niccolò Tartaglia”, Università Cattolica
del Sacro Cuore, Via Musei 41, 25121 Brescia, Italy
[email protected]
Dipartimento di Matematica e Geoscienze, Universitá degli studi di Trieste, Via
Valerio 12/1, 34127 Trieste
[email protected]
Aix-Marseille Université, CNRS, Centrale Marseille, I2M, UMR 7373, 13453
Marseille, France
[email protected]
Dipartimento di Matematica e Geoscienze, Universitá degli studi di Trieste, Via
Valerio 12/1, 34127 Trieste
[email protected]
31
| 4math.GR
|
arXiv:1801.05465v1 [math.ST] 16 Jan 2018
On a bimodal Birnbaum-Saunders distribution with applications to
lifetime data
Roberto Vila1 , Jeremias Leão2∗, Helton Saulo1 , Mirza Nabeed3 , and Manoel Santos-Neto4,5
1
Department of Statistics, University of Brası́lia, Brası́lia, BRA
2
Department of Statistics, Federal University of Amazonas, Manaus, BRA
3
Department of Statistics, University of Gujrat, Gujrat, PAK
4
Department of Statistics, Federal University of São Carlos, São Carlos, BRA
5
Department of Statistics, Federal University of Campina Grande, Campina Grande, BRA
January 18, 2018
Abstract
The Birnbaum-Saunders distribution is a flexible and useful model which has been used in several fields. In this
paper, a new bimodal version of this distribution based on the alpha-skew-normal distribution is established. We
discuss some of its mathematical and inferential properties. We consider likelihood-based methods to estimate the
model parameters. We carry out a Monte Carlo simulation study to evaluate the performance of the maximum
likelihood estimators. For illustrative purposes, three real data sets are analyzed. The results indicated that the
proposed model outperformed some existing models in the literature, in special, a recent bimodal extension of the
Birnbaum-Saunders distribution.
Keywords: Birnbaum-Saunders distribution; Alpha-skew-normal distribution; Bimodality; Maximum likelihood
estimation; Monte Carlo simulation.
1
Introduction
Despite its broad applicability in many fields, see, for example, Balakrishnan et al. (2007), Bhatti (2010), Vilca et al.
(2010), Paula et al. (2012), Saulo et al. (2013),Leiva et al. (2014a,b), Leiva (2016) and Leao et al. (2017), the BirnbaumSaunders (BS) distribution Birnbaum and Saunders (1969) is not suitable to model bimodal data. This distribution is
positively skewed with positive support and is related to the normal distribution through the stochastic representation
T =
i2
p
βh
αZ + (αZ)2 + 4 ,
4
(1)
where T ∼ BS(α, β), Z ∼ N(0, 1) and α > 0, β > 0 are shape and scale parameters, respectively. The BS(α, β)
probability density function (PDF) and cumulative distribution function (CDF) are respectively given by
f (t; α, β) = φ(a(t))
∗ Corresponding
t−3/2 (t + β)
2α β 1/2
and F (t; α, β) = Φ(a(t)),
author: Jeremias Leão. Email: [email protected]
1
t > 0,
where φ(·) and Φ(·) are standard normal PDF and CDF, respectively, and
"r
r #
1
t
β
.
−
a(t) =
α
β
t
Note that the k-th derivative of a(t), denoted by a(k) (t), satisfies a(k) (t) > 0
k > 1. Some special cases of these derivatives are
"
"
r #
1
1 β
1
1
1
′
′′
√ +
√ +
a (t) =
, a (t) = −
2α
t t
4αt
βt
βt
a′′′ (t) =
3
8αt2
"
r #
5 β
1
√ +
.
t t
βt
(2)
(or < 0) for k odd (or k even), where
3
t
r #
β
t
and
(3)
Note also that the function a(·) has inverse specified by (1). In order not to cause confusion, hereafter we will write
a−⊥ (·) to denote the inverse of function a(·).
The stochastic representation in (1) allows us to obtain several generalizations of the BS model. For example, Dı́azGarcı́a and Leiva (2005) assumed that Z follows a standard symmetric distribution in the real line and obtained the class
of generalized BS distributions. On the same line, Balakrishnan et al. (2009) proposed scale-mixture BS distributions
by assuming that Z belongs to the family of scale mixture of normal distributions. Many other generalizations can be
obtained in order to obtain a new distribution with domain on the positive numbers; see Leiva (2016).
In general, one uses mixtures of distributions for describing bimodal data. However, it may be troublesome as
identifiability problems may arise in the parameter estimation of the model; see Lin et al. (2007a,b) and Gómez et al.
(2011). In this sense, new mixture-free models which have the capacity to accommodate unimodal and bimodal data
are very important. Some asymmetric bimodal models in the real line have been discussed by Azzalini and Capitanio
(2003), Kim (2005), and Ma and Genton (2004), among others. In the context of bimodal BS models, Balakrishnan
et al. (2011) introduced different mixture models and studied their characteristics. Olmos et al. (2017) introduced a
bimodal extension of the BS distribution, denoted by BBSO, based on the approach described in Gómez et al. (2011).
Olmos et al. (2017) studied, amongst other things, the probabilistic properties and moments of the BBSO distribution,
and showed that this model can fit well both unimodal and bimodal data in comparison with the BS, log-normal and
skew-normal BS models. A thorough inference study on the parameters that index the BBSO distribution was addressed
by Fonseca and Cribari (2016).
In this paper, we introduce a new bimodal version of the BS distribution, denoted by BBS, by assuming that Z in (1)
follows a alpha-skew-normal (ANS) distribution discussed by Elal-Olivero (2010). We present a statistical methodology
based on the proposed BBS model including model formulation, mathematical properties, estimation and inference
based on the maximum likelihood (ML) method. We evaluate the performance of the ML estimators by Monte Carlo
(MC) simulations. Three real data illustrations indicate the good performance of the proposed model. Specially, the
proposed BBS model provides better adjustment compared to the BBSO model proposed by Olmos et al. (2017).
The rest of the paper proceeds as follows. In Section 2, we introduce the BBS distribution and discuss some related
results. In Section 3, we consider likelihood-based methods to estimate the model parameters and to perform inference.
In Section 4, we carry out a MC simulation study to evaluate the performance of the ML estimators. In Section 5, we
illustrate the proposed methodology with three real data sets. Finally, in Section 6, we make some concluding remarks
and discuss future research.
2
2
The BBS distribution
If a random variable (RV) X has an ASN distribution with parameter δ, denoted by X ∼ ASN(δ), then its PDF and
CDF are given by
(1 − δx)2 + 1
2 − δx
g(x) =
φ(x),
(4)
φ(x) and G(x) = Φ(x) + δ
2 + δ2
2 + δ2
where x, δ ∈ R and δ is an asymmetric parameter that controls the uni-bimodality effect; see Elal-Olivero (2010). The
PDF of the BS distribution, based on the alpha-skew-normal model, is given by
f (t; α, β, δ) =
(1 − δa(t))2 + 1
t−3/2 (t + β)
,
φ(a(t))
2 + δ2
2α β 1/2
t > 0,
(5)
where a(·) is as in (2) and the notation T ∼ BBS(α, β, δ) is used. If δ = 0, then the classical BS(α, β) distribution is
obtained. The corresponding BBS(α, β, δ) CDF is given by
2 − δa(t)
F (t; α, β, δ) = Φ(a(t)) + δ
φ(a(t)), t > 0.
(6)
2 + δ2
Note that f (t; α, β, δ) = g(a(t))a′ (t) = (G ◦ a)′ (t), F (t; α, β, δ) = (G ◦ a)(t) and limδ→±∞ {F (t; α, β, δ) + φ(a(t))a(t)} =
F (t; α, β).
Differentiating the PDF of the BBS distribution (5) we obtain
f ′ (t; α, β, δ) = g ′ (a(t))[a′ (t)]2 + g(a(t))a′′ (t) and
′′
′′
′
3
′
′
(7)
′′
′′′
f (t; α, β, δ) = g (a(t))[a (t)] + 3g (a(t))a (t)a (t) + g(a(t))a (t),
where g ′′ (x) = −xg ′ (x) + φ(x) −3δ 2 x2 + 4δx − 2(1 − δ 2 ) /(2 + δ 2 ) and
g ′ (x) =
(8)
φ(x)
−δ 2 x3 + 2δx2 − 2(1 − δ 2 )x − 2δ .
2
2+δ
The survival and hazard functions, denoted by SF and HR, respectively, of the BBS distribution are given by
S(t; α, β, δ) = 1 − (G ◦ a)(t) and
h(t; α, β, δ) =
f (t; α, β, δ)
(G ◦ a)′ (t)
=
,
1 − F (t; α, β, δ)
S(t; α, β, δ)
t > 0,
respectively. From Figure 1 we note some different shapes of the BBS PDF for different combinations of parameters.
These figures reveal clearly the bimodality effect caused by the parameter δ. Also, Figure 2 shows unimodal and bimodal
shapes for the BBS HR.
2.1
Some properties of the BBS distribution
Lemma 2.1. Let t0 = β[(α/δ) +
a decreasing function when
p
(α/δ)2 + 4]2 /4. The PDF of the BBS distribution t 7→ f (t; α, β, δ) defined in (5) is
1. δ = 0 and t > β, or
2. t < t0 ( t > t0 ), for each δ > 0 ( δ < 0 ).
3
4
0.00
1.2
3
PDF
2
1
0
1.5
0
0.0
1
0.5
PDF
3
PDF
2
1.0
PDF
0.0 0.5 1.0 1.5 2.0 2.5 3.0
0.05
15
0.00
1.2
0.10
2.0
2.5
3.0
0.0
0.5
1.0
1.5
t
2.0
2.5
3.0
2. cT ∼ BBS(α, cβ, δ), with c > 0;
3. T −1 ∼ BBS(α, β −1 , −δ).
Proof. Since P(a(T ) 6 t) = F (a−⊥ (t); α, β, δ), we have that the PDF of a(T ) is equal to f (a−⊥ (t); α, β, δ)/a′ (a−⊥ (t)) =
g(t). Then a(T ) ∼ ASN(δ). The proof of the Items 2 and 3 are immediate, after making convenient variables transformations.
Proposition 2. Let T ∼ BBS(α, β, δ) and X ∼ ASN(δ) and suppose E[T n ] exists, n > 1. Then, we have
2δ
;
1. E[T ] = β2 −α 2+δ
2 + ω0,1
2+3δ 2
2+δ 2
20
0.6
0.05
0.15
1.5
t
1. a(T ) ∼ ASN(δ);
4 + 2α2
15
0.10
0.20
1.0
Proposition 1. Let T ∼ BBS(α, β, δ) as defined in (4). Then,
β
2
20
0.6
0.15
0.25
= 10
= 10
= 10
= 10
= 10
Proof. 1. If δ = 0, we have f (t; α, β, δ) = f (t; α, β) = φ(a(t))a′ (t), t > 0. Since a′ (t) > 0 and a′′ (t) < 0 (see (3)) we
have that t 7→ φ(a(t)) and t 7→ a′ (t) are decreasing functions whenever t > β. Therefore, since the PDF of the BBS
distribution is a product of nonnegative decreasing functions, it is decreasing for each t > β.
2. Let r(x) = ((1 − δx)2 + 1)/(2 + δ 2 ), x ∈ R. Note that (r ◦ a)′ (t) = −2δ(1 − δa(t))a′ (t)/(2 + δ 2 ), t > 0. If δ > 0 (
δ < 0 ) then (r ◦ a)′ (t) < 0 whenever t < t0 ( t > t0 ). Since, by Item 1, the function t 7→ f (t; α, β) is decreasing and
f (t; α, β, δ) = r(a(t))f (t; α, β), t > 0, the proof follows.
2
30
0.20
0.30
3.0
Figure 1: BBS PDFs for some parameter values (β = 1.0).
2. E[T 2 ] =
25
0.8
0.25
2.5
α = 0.25, δ =
5
0.8
α = 0.50, δ = 5
= 3, δ
= 3, δ
= 3, δ
= 3, δ
= 3, δ
2.0
α = 1.00, δ =
5
1.2
α = 0.10, β
α = 0.25, β
α = 0.50, β
α = 1.00, β
α = 1.50, β
α = 1.50, δ = 5
1.5
t
α = 0.10, δ = −2
1.0
α = 0.25, δ = −2
α = 0.50, δ = −2
0.5
α = 1.00, δ = −2
α = 1.50, δ = −2
0.0
α = 0.10, δ = −10
0.5
25
0.8
0.30
α = 0.25, δ = −10
0.2
α = 0.10, δ = −10
α = 0.10, δ = 5
α = 0.25, δ = 5
α = 0.50, δ = 5
α = 1.00, δ = 5
α = 1.50, δ = 5
α = 0.25, δ = −10
α = 0.50, δ = −10
α = 1.00, δ = −10
α = 0.10, δ = 5
3.0
α = 1.50, δ = −10
2.5
α = 0.25, δ = 5
2.0
α = 0.50, δ = 5
1.5
t
α = 1.00, δ = 5
1.0
α = 0.10, δ = −2
α = 0.25, δ = −2
α = 0.50, δ = −2
α = 1.00, δ = −2
α = 1.50, δ = −2
0.0
30
α = 0.25, δ = −10
0.2
α = 0.10, δ = −10
α = 1.00, δ = −10
0.4
α = 0.50, δ = −10
0.3
α = 0.10, δ =
5
0.6
α = 1.50, δ = −10
α = 1.00, δ = −10
0.4
α = 1.50, δ = −10
α = 0.50, δ = −10
0.3
0.6
0.8
−10
−10
−10
−10
−10
1.2
α = 0.10, δ = −2
α = 0.25, δ = −2
α = 0.50, δ = −2
α = 1.00, δ = −2
α = 1.50, δ = −2
α = 0.10, β = 3, δ = 10
α = 0.25, β = 3, δ = 10
α = 1.00, β = 3, δ = 10
α = 0.50, β = 3, δ = 10
α = 1.50, β = 3, δ = 10
=
=
=
=
=
α = 0.10, δ = 5
0.10, δ
0.25, δ
0.50, δ
1.00, δ
1.50, δ
α = 0.25, δ = 5
α = 0.50, δ = 5
0.5
α = 1.50, δ = 5
α = 0.10, β = 3, δ = 10
α = 0.50, β = 3, δ = 10
α = 0.25, β = 3, δ = 10
α = 1.50, β = 3, δ = 10
α = 1.00, β = 3, δ = 10
0.0
=
=
=
=
=
α = 1.00, δ = 5
α = 1.50, δ = 5
α = 0.25, δ = −2
α = 0.10, δ = −2
α = 1.00, δ = −2
α = 0.50, δ = −2
α = 1.50, δ = −2
α = 0.10, β = 3, δ = 10
α = 0.50, β = 3, δ = 10
α = 0.25, β = 3, δ = 10
α = 1.50, β = 3, δ = 10
α = 1.00, β = 3, δ = 10
α
α
α
α
α
+ 2αω1,1 ;
4
10
4
HR
6
8
10
8
6
HR
0.0 0.5 1.0 1.5 2.0 2.5
2
0
20
4
2
0
0
5
HR
10
15
HR
15
1.0
1.5
2.0
2.5
0.5
3.0
1.0
1.5
2.0
α = 0.1, β = 0.8, δ = 2
δ
2
;
−24α(α2 + 1) 2+δ
2 + 3α ω2,1 + αω1,1 + ω0,3
16 + 8α2
8+12δ 2 +6α2 +15δ 2 α2
2+δ 2
+ 4α(α2 ω3,1 + ω1,3 ) and
2
2
β
+3δ 4
2
2α2 4+6δ
2
(2+δ 2 )2 + ω0,1 + 2α(ω1,1 + ω0,1
√
= E[X r ( α2 X 2 + 4)k ].
5. Var[T ] =
where ωr,k
15
0.00
1.2
β
2
20
0.6
0.00
1.2
0.05
4
20
0.6
0.05
0.10
4. E[T 4 ] =
25
0.8
0.10
0.15
2.5
= 2
= 2
= 2
= 2
= 2
Figure 2: BBS HRs for some parameter values (β = 1.0).
β
2
30
0.15
0.20
3.0
0.5
α = 0.1, β = 0.8, δ = 2
3
25
0.8
0.20
0.25
α = 0.10, δ = 10
α = 0.25, δ =0.2
10
= 0.5, δ
= 0.8, δ
= 1.0, δ
= 1.2, δ
= 1.5, δ
α = 0.50, δ =0.3
10
α = 0.1, β
α = 0.1, β
α = 0.1, β
α = 0.1, β
α = 0.1, β
α = 1.00, δ =0.4
10
3.0
α = 0.1, β = 0.8, δ = 2
α = 1.50, δ = 10
α = 0.10, δ = 0.6
0.1
2.5
α = 0.25, δ = 0.8
0.1
2.0
α = 0.50, δ = 0.1
α = 1.00, δ = 1.2
0.1
α = 0.5, δ = −2
1.5
α = 1.50, δ = 0.1
α = 0.5, δ = 0
1.0
α = 0.5, δ = 1
α = 0.5, δ = 2
t
0.5
α = 0.5, δ = 1
0.0
α = 0.10, δ = 10
α = 0.25, δ = 10
α = 0.5, δ = −2
α = 0.5, δ = 1
α = 0.5, δ = 0
α = 0.5, δ = 1
α = 0.5, δ = 2
0.0
3. E[T 3 ] =
30
0.25
0.30t
α = 0.10, δ = 0.1
α = 0.25, δ = 0.1
α = 0.50, δ = 0.1
α = 1.00, δ = 0.1
α = 1.50, δ = 0.1
α = 0.50, δ = 10
α = 1.00, δ = 10
3.0
α = 0.1, β = 0.8, δ = 2
α = 1.50, δ = 10
2.5
α = 0.10, δ = 0.1
2.0
α = 0.50, δ = 0.1
1.5
α = 0.25, δ = 0.1
1.0
α = 1.50, δ = 0.1
0.5
α = 1.00, δ = 0.1
0.0
α = 0.1, β = 0.5, δ = 2
α = 0.1, β = 1.0, δ = 2
α = 0.1, β = 1.5, δ = 2
α = 0.1, β = 1.2, δ = 2
0.30t
α = 0.10, δ = 10
α = 0.25, δ =0.2
10
α = 0.50, δ =0.3
10
α = 1.00, δ =0.4
10
α = 1.50, δ = 10
0.6
0.8
10
10
10
10
10
1.2
α = 0.5, δ = −2
α = 0.5, δ = 0
α = 0.5, δ = 1
α = 0.5, δ = 1
α = 0.5, δ = 2
t
α = 0.1, β = 0.5, δ = 2
α = 0.1, β = 1.0, δ = 2
α = 0.1, β = 1.2, δ = 2
α = 0.10, δ = 0.1
=
=
=
=
=
α = 0.1, β = 1.5, δ = 2
α = 0.50, δ = 0.1
0.10, δ
0.25, δ
0.50, δ
1.00, δ
1.50, δ
α = 0.25, δ = 0.1
α = 1.50, δ = 0.1
=
=
=
=
=
α = 1.00, δ = 0.1
α = 0.5, δ = 1
α = 0.5, δ = −2
α = 0.5, δ = 0
α = 0.5, δ = 1
α = 0.5, δ = 2
α = 0.1, β = 0.5, δ = 2
α = 0.1, β = 1.2, δ = 2
α = 0.1, β = 1.0, δ = 2
α = 0.1, β = 1.5, δ = 2
α
α
α
α
α
2δ
2+δ 2 )
,
Proof. By Proposition 1 Item 1 we have a(T ) ∼ ASN(δ) which implies that E[T n ] = E {a−⊥ (X)}n . Then, the proof
is immediate since
n h
n i
p
β
, n>1
(9)
E αX + (αX)2 + 4
E {a−⊥ (X)}n =
2
and E[X] = −2δ/(2 + δ 2 ), E[X 2 ] = 1 − δE[X], E[X 3 ] = 3E[X], and E[X 4 ] = 3(1 − 2δE[X]).
Remark. By using the Binomial Theorem and (9) note that E[T n ] exists iff ωr,n−r (defined in Proposition 2) exists,
with r = 0, . . . , n. By Jensen’s inequality (see, e.g., Chung (2001)) we obtain
s
p
2 + 3δ 2
|ω0,1 | 6 α2 E[X 2 ] + 4 = α2
+ 4 < +∞,
2 + δ2
and by Minkowski inequality (see, e.g., Natanson (1955)) we have
s
r
2
p
p
2
+
3δ
2 + 3δ 2
2
+ 4 < +∞.
|ω1,1 | 6 E[X 2 ] α2 E[X 2 ] + 4 =
α
2 + δ2
2 + δ2
5
Then, the expected value E[T ] and variance Var[T ] always exist. Note also that higher order moments can also be easily
obtained from the expression of E[T n ].
Proposition 3. Let X ∼ ASN(δ) and T = a−⊥ (X). Then, T ∼ BBS(α, β, δ).
Proof. Since P(a−⊥ (X) 6 t) = G(a(t)), we have that the PDF of the RV a−⊥ (X) is equal to g(a(t))a′ (t) = f (t; α, β, δ).
Proposition 4. Let T ∼ χ23 , where χ23 denotes the chi-squared distribution with 3 degrees of freedom. We have the
following relation
fa−⊥ (√T ) (t) (1/a(t) − δ)2 + (1/a(t))2 = 2(2 + δ 2 ) f (t; α, β, δ), t > 0,
√
where fa−⊥ (√T ) (·) denotes the PDF of the RV a−⊥ ( T ).
√
Proof. Since P(a−⊥ ( T ) 6 t) = P(T 6 a2 (t)), we have that fa−⊥ (√T ) (t) = 2a2 (t)φ(a(t))a′ (t), from where the proof
follows.
2.2
Some properties of the HR of the BBS distribution
Let s(t) = −f ′ (t; α, β, δ)/f (t; α, β, δ), where f (·; α, β, δ) denotes the PDF of the BBS distribution (5). It is straightforward to show that
a′ (t)m(t)
,
s(t) =
(1 − δa(t))2 + 1
where
a′′ (t)
.
[a′ (t)]2
p
p
Remark. By (3) we have that a′ (t) = (1/α) t/β − (1/2)a(t) /t and a′′ (t) = − (1/α) t/β − (3/4)a(t) /t2 . Using
these identities, note that m(t) = 0 iff
m(t) = δ 2 a3 (t) − 2δa2 (t) − 2a(t)(δ 2 − 1) + 2δ − (1 − δa(t))2 + 1
1 2 3
t δ a (t) − 2δa2 (t) − 2(δ 2 − 1)a(t) + 2δ
α2 β
1
+ a(t) δ 2 a4 (t) − 2δa3 (t) − (5δ 2 − 2)a2 (t) + 8δa(t) − 6
4
1 √ 2 4
t −δ a (t) + 2δa3 (t) + (3δ 2 − 2)a2 (t) − 4δa(t) + 2 = 0.
+ √
α β
Consider also the function spaces
(
B=
U=
(
differentiable :
ℓ′ (t) < 0 for t ∈ (0, t0 ), ℓ′ (t0 ) = 0,
ℓ′ (t) > 0 for t > t0
)
,
ℓ : R → R differentiable :
ℓ′ (t) > 0 for t ∈ (0, t0 ), ℓ′ (t0 ) = 0,
ℓ′ (t) < 0 for t > t0
)
.
+
ℓ:R →R
+
+
+
Each function ℓ ∈ B or ℓ ∈ U is said bathtub shaped or upside down bathtub shaped, respectively.
The following results due to Glaser (1980) helps us to characterize the shape of the failure rates, through the function
s(·).
6
1. If t 7→ s(t) is increasing, then, the HR is increasing in t.
2. If t 7→ s(t) is decreasing, then, the HR is decreasing in t.
3. If t 7→ s(t) ∈ B and if there exist a t∗ such that h′ (t∗ ; α, β, δ) = 0, then, the HR belongs to B, otherwise the HR
is increasing in t.
4. If t 7→ s(t) ∈ U and if there exist a t∗ such that h′ (t∗ ; α, β, δ) = 0, then, the HR belongs to U , otherwise the HR
is decreasing in t.
Using the expressions of the derivatives of a(·) in (3) we can get the monotonicity of the HR of the BBS distribution
from the following equation
s′ (t) =
m′ (t)a′ (t) + m(t)a′′ (t) + 2δ(1 − δa(t))a′ (t)s(t)
,
(1 − δa(t))2 + 1
where
m′ (t)
=3δ 2 a2 (t) − 4δa(t) − 2(δ 2 − 1)
a′ (t)
a′′′ (t)a′ (t) − 2[a′′ (t)]2
a′′ (t)
.
+ 2δ(1 − δa(t)) ′ 2 − (1 − δa(t))2 + 1
[a (t)]
[a′ (t)]4
For example, if δ = 0 and
α > 2, we have that m(t) > 0 iff t > β.
Defining the set Lβα = t : t4 +(4 − α2 )βt3 + 6(1 − α2 )β 2 t2 + (4 + 3α2 )β 3 t + β 4 < 0 note that Lβα is a non-empty
set (because t = β ∈ Lβα ) and m′ (t) = a′ (t) 2 − 2 (a′′′ (t)a′ (t) − 2[a′′ (t)]2 )/[a′ (t)]4 < 0 on Lβα . That is, s′ (t) < 0 on
{t : t > β} ∩ Lβα , where s′ (t) = (m′ (t)a′ (t) + m(t)a′′ (t)) /2 . Therefore, by Item 2 above, the HR t 7→ h(t; α, β, δ = 0)
is decreasing on {t : t > β} ∩ Lβα . On the other hand, if δ = 0 and α < 1, m(t) < 0 iff t < β. In this case, note that
m′ (t) > 0 on [Lβα ]c = R+ , hence s′ (t) > 0 for each t < β. Then, using the Item 1 above, the HR is increasing for each
t < β.
Another easy case to study is when δ = 1. In this case, m(t) > 0 iff t > β. Note also that m′ (t) < 0 on the set
2
Lα,β = {t : 3a2 (t)− 4a(t)+ 2(1 − a(t))a′′ (t)/[a′ (t)]√
< 0} and that Lα,β is a non-empty set because t = β ∈ Lα,β . Then,
′
s (t) < 0 on {t : t > t1 } ∩ Lα,β where t1 = β[α + α2 + 4]2 /4. Therefore, by Item 2 above, the HR t 7→ h(t; α, β, δ = 1)
is decreasing on {t : t > t1 } ∩ Lα,β . Similar analyzes can be done for the other possible cases.
We emphasize that h′ (t; α, β, δ) = 0 iff the PDF of the BBS distribution is a decreasing function.pBut, by Lemma
2.1 this happens when δ = 0 and t > β or t < t0 ( t > t0 ), for each δ > 0 ( δ < 0 ) with t0 = β[(α/δ) + (α/δ)2 + 4]2 /4.
So, to see if the HR belongs (or not) to B or to U it would be sufficient to verify that t 7→ s(t) belongs (or not) to B
or to U .
2.3
Bimodality Properties
In this subsection, some results on the bimodality properties of BBS distribution are obtained.
Proposition 5. A mode of the BBS(α, β, δ) is any point t0 = t0 (α, β, δ) that satisfies
α2 β 1
1 √
t0 = −
t pe4 (t) ,
a(t)p4 (t) + √
p3 (t) 4
α β
where p3 (t) = δ 2 a3 (t) − 2δa2 (t) − 2(δ 2 − 1)a(t) + 2δ, p4 (t) = δ 2 a4 (t) − 2δa3 (t) − (5δ 2 − 2)a2 (t) + 8δa(t) − 6 and
pe4 (t) = −δ 2 a4 (t) + 2δa3 (t) + (3δ 2 − 2)a2 (t) − 4δa(t) + 2.
7
Proof. A mode of the BBS(α, β, δ) is any point t that satisfies f ′ (t; α, β, δ) = 0. But this happens iff s(t) = 0 which is
equivalent to m(t) = 0, where s(t) and m(t) were defined in Subsection 2.2. Then, using Remark 2.2 and solving for t
gives the result.
Proposition 6. The function t 7→ (g ◦ a)(t) and the PDF of the BBS distribution (5) have different modes.
Proof. We will do the proof by contradiction. Let’s suppose that t0 is a mode for both (g ◦ a)(·) (which always exists,
since g is bimodal) and f (·; α, β, δ). Then g ′ (a(t0 ))a′ (t0 ) = 0 and g ′′ (a(t0 ))[a′ (t0 )]2 < 0.
Since f ′ (t0 ; α, β, δ) = 0 and f ′′ (t0 ; α, β, δ) < 0, using (7) and (8) we obtain that g(a(t0 )) = 0, which is impossible.
Then, the proof follows.
Remark. As a consequence of the proof of the Proposition 6 we have that, if t0 is a maximum point of t 7→ (g ◦ a)(t)
then, the maximum points of the BBS distribution must be to the left side of t0 . On the other hand, if t1 is a minimum
point of t 7→ (g ◦ a)(t) then, the minimum points of the BBS distribution must be to the right side of t1 .
Proposition 7. The PDF of the BBS distribution (5) has at most one mode when δ = 0.
Proof. If δ = 0, then the classical BS(α, β) distribution is obtained, i.e., f (t; α, β, δ) = f (t; α, β) = φ(a(t))a′ (t), t > 0.
Differentiating f (t; α, β), we obtain
f ′ (t; α, β) = φ(a(t)) a′′ (t) − a(t)[a′ (t)]2 .
Using (3), it is straightforward to show that f ′ (t; α, β) = 0 iff
t3 + β(1 + α2 )t2 − β 2 t − β 3 = 0.
(10)
The discriminant of a cubic polynomial ax3 + bx2 + cx + d is given by ∆3 = b2 c2 − 4ac3 − 4b3 d − 27a2 d2 + 18abcd. In
our case, we have
∆3 = β 6 4(1 + α2 )3 + (1 + α2 )2 + 18(1 + α2 ) − 23 .
Note that ∆3 > 0 for each α > 0, then the equation (10) has three distinct real roots.
Let t1 , t2 and t3 be the three distinct real roots of (10), by Vieta’s formula (see, e.g., Vinberg (2003)), it is valid
that
t1 + t2 + t3
=
t1 t2 + t1 t3 + t2 t3
=
t1 t2 t3
=
−β(1 + α2 )
−β 2
β3.
From the first and third equations above we conclude that there must be two negative and one positive roots, hence
f (t; α, β) has at most one mode.
Proposition 8. If δ = −α, then one of the modes of the BBS distribution (5) occurs at t = β.
√
Proof. Since a(β) = 0, a′ (β) = 1/αβ, a′′ (β) = −1/αβ 2 , g(0) = 2/ 2π(2 + δ 2 ) and g ′ (0) = −g(0)δ, by (7) we have that
δ
2
√
+ 1 = 0 since δ = −α.
(11)
f ′ (β; α, β, δ) = −
αβ 2 (2 + δ 2 ) 2π α
I.e., t = β is one of the critical points of f when δ = −α.
8
Since a′′′ (β) = 9/4αβ 3 and g ′′ (0) = −g ′ (0)(1 + δ 2 ), using (8), note that
δ
9
2
2 1
′′
√
δ(1 + δ ) 2 + 3 −
.
f (β; α, β, δ) =
α
α 4
αβ 3 (2 + δ 2 2π)
As δ = −α we obtain
2
√
f (β; α, β, δ) = −
3
αβ (2 + α2 ) 2π
′′
1 + α2
21
+
α
4
< 0.
Therefore, the PDF of the BBS distribution is concave downward when δ = −α.
Example 2.1 (Bimodality). Consider α = β = 1 and δ = −α. By Proposition 8 the point t = 1 is one of the modes
of f (·; α, β, δ). Using (7) note that f ′ (t; α, β, δ) = 0 iff
p(y) = y 10 + 2y 9 + y 6 − 4y 5 + 3y 4 − 8y 3 + 4y 2 + 2y − 1 = 0,
where y = t1/2 .
We have that p(0) = −1 < 0,
p(1/2) =
85
> 0,
1024
p(3/4) = −
252223
< 0 and p(5/4) ≈ 15.27127 > 0.
1048576
Therefore, p(y) has roots in the intervals (0, 1/2), (1/2, 3/4) and (3/4, 5/4). It is not hard to show that p(y) > 0 for
y > 1. Thus, all real roots of the polynomial p(y) lie in the interval (0, 5/4). Computationally it can be verified that
y0 ≈ 0.419703, y1 ≈ 0.646914 and y2 = 1 are the only roots of p(y) on {y : y > 0}. Hence, t0 = y02 ≈ 0.1761,
t1 = y12 ≈ 0.4184 and t2 = y22 = 1 are the only roots of f ′ (t; α, β, δ) = 0. It can be verified that
a(k) (t)
t = t0
t = t1
k=0
−1.9633
−0.8991
g (k) (t)
t = t0
t = t1
k=0
1.9219 φ(t0)/3
1.0101 φ(t1)/3
k=1
7.963
2.6204
k=2
−61.0996
−7.5471
k=1
1.8585 φ(t0)/3
1.1100 φ(t1)/3
k=3
848.6550
42.8876
and
k=2
−0.0616 φ(t0)/3
2.1692 φ(t1)/3
where a(0) ≡ a and g (0) ≡ g. Using (8) and the quantities above, we obtain
f ′′ (t0 ; α, β, δ) = g ′′ (a(t0 ))[a′ (t0 )]3 + 3g ′ (a(t0 ))a′ (t0 )a′′ (t0 ) + g(a(t0 ))a′′′ (t0 )
≈ −1107.6637
φ(t0 )
< 0,
3
and similarly f ′′ (t1 ; α, β, δ) ≈ 60.3992 φ(t1)/3 > 0.
Therefore, the PDF of the BBS distribution, with parameters α = β = 1 and δ = −α, has exactly two modes at
t = t0 and t = t2 .
Remark. Let α = β = 1 and δ = −α. It can be verified that the point tmax ≈ a−⊥ (0.83929) = 2.26240 is the only
maximum point of the function (g ◦ a)(·). The Remark 2.3 assures us that the maximum points of the PDF f (·; α, β, δ)
must be to the left side of tmax . This statement was verified in the previous example.
9
2.4
Shannon Entropy
For a continuous PDF f (t) on an interval I, its entropy is defined as
Z
H(f ) = − f (t) log f (t)dt.
I
This definition of entropy, introduced by Shannon and Weaver (1949), resembles a formula for a thermodynamic
notion of entropy. In our probabilistic context, if X is an absolutely continuous RV with PDF fX (t), the quantity
H(X) = H(fX ) = −E[log fX (X)] is viewed as a measure of uncertainty associated with a RV. Note that H(X) is not
necessarily well-defined, since the integral does not always exist.
Consider T ∼ BBS(α, β, δ). The Shannon entropy of T satisfies the following identity
Proposition 9. If T ∼ BBS(α, β, δ), there exists a constant C(α, β, δ) such that
T 3/2 /(T + β)
.
H(T ) = C(α, β, δ) + E log
(1 − δa(T ))2 + 1
(12)
Proof. It is straightforward to verify that
√
1
H(T ) = log(2 + δ 2 ) + log(2αβ 1/2 ) + log( 2π) + E[a2 (T )]
2
3/2
Z ∞
t /(t + β)
f (t; α, β, δ)dt.
log
+
(1 − δa(t))2 + 1
0
Since a(T ) ∼ ASN(δ), by Proposition 1 we have E[a2 (T√
)] = 1 + 2δ 2 /(2 + δ 2 ). Therefore, the identity (12) is verified
2
1/2
considering C(α, β, δ) = log(2 + δ ) + log(2αβ ) + log( 2π) + (1 + 2δ 2 /(2 + δ 2 ))/2.
Remark. If T ∼ BBS(α, β, δ) and T > 1, then, the Shannon entropy always exists. In fact, by Jensen’s inequality (see,
e.g., Chung (2001)), Minkosky inequality (see, e.g., Natanson (1955)) and Remark 2.1 we obtain
E[log(T 3/2 )] 6 log E[T 3/2 ] 6 log(E[T 2 ])1/2 + log(E[T ])1/2 < +∞,
|E[log(T + β)]| 6 log(E[T ] + β) < +∞ and
E[log (1 − δa(T )2 + 1 ] 6 log 2 + δ 2 E[a2 (T )] − 2δE[a(T )] < +∞,
since a(T ) ∼ ASN(δ). Then, using (12) and the above inequalities, the proof follows.
3
3.1
Estimation and inference
Maximum likelihood estimation
Let (t1 , . . . , tn ) be a random sample of size n from the BBS distribution with PDF in (5). Considering δ known, it
follows that the log-likelihood function, without the constant, is given by
ℓ(θ) = − n log(α) −
−
n
X
n
log 1 + (1 − δa(ti ))2
log(β) +
2
i=1
n
n
X
1X 2
log (ti + β) ,
a (ti ) +
2 i=1
i=1
10
where θ = (α, β). Since
∂
t−1/2
a(t) = − 2 1/2 (t − β)
∂α
α β
and
t−1/2 −3/2
∂
a(t) = −
(β
t + 2),
∂β
2α
taking the first derivatives with respect to α and β and equating them to zero, we have
n
2δ(1 − δa(ti ))
n X
∂
∂
+ a(ti )
ℓ(θ) = − −
a(ti ) = 0 and
2
∂α
α i=1 1 + (1 − δa(ti ))
∂α
n
X
∂
n
ℓ(θ) = −
−
∂β
2β i=1
(13)
n
X
1
2δ(1 − δa(ti ))
∂
+
a(t
)
a(t
)
+
= 0.
i
i
1 + (1 − δa(ti ))2
∂β
t
+
β
i
i=1
The ML estimates α
b and βb of α and β, respectively, are obtained by solving an iterative procedure for non-linear
optimization of the system of equations in (13), such as the Broyden-Fletcher-Goldfarb-Shanno (BFGS) quasi-Newton
method. We can use the profile log-likelihood for finding the value of δ. Generally, two steps are required:
i) Let δi = i and for each i = −20, . . . , 0, . . . , 20 compute the ML estimates of α and β by solving the system of
equations in (13);
ii) Select the final estimate of δ as the one which maximizes the log-likelihood function and also select the associated
estimates of α and β as final ones.
Case of random censoring. Suppose that the time to the event of interest is not completely observed and it may
be subject to right censoring. Let ci denote the censoring time and ti the time to the event of interest. We observe
yi = min{ti , ci }, whereas τi = I(ti ≤ ci ) is such that τi = 1 if yi is the time to the event of interest and τi = 0 if it
is right censored, for i = 1, . . . , n. Let θ = (α, β) denote the parameter vector of the BBS model given in (5) with
δ known. From n pairs of times and censoring indicators (t1 , τ1 ), . . . , (tn , τn ), the corresponding likelihood function
obtained under uninformative censoring can be expressed as
L(θ) =
n
Y
i=1
f (ti ; α, β, δ)τi (1 − F (ti ; α, β, δ))1−τi
)τ i
−3/2
1 + (1 − δa(ti ))2
(ti + β)
ti
=
φ(a(ti ))
2 + δ2
2α β 1/2
i=1
1−τi
2 − δti
.
φ(a(ti ))
× 1 − Φ(a(ti )) + δ
2 + α2
n
Y
(
(14)
Therefore, the log-likelihood function for the BBS model obtained from (14) is given by
ℓ(θ) = − ωi η(θ) +
n
X
i=1
τi log(1 + (1 − δa(ti ))2 ) +
n
X
τi log(φ(a(ti )))
i=1
n
n
X
3X
τi log(ti + β)
τi log(ti ) +
2 i=1
i=1
n
X
2 − δti
(1 − τi ) log 1 − Φ(a(ti )) + δ
+
φ(a(ti )) ,
2 + α2
i=1
−
11
(15)
Pn
where ωi = i=1 τi and η(θ) = log(2αβ 1/2 (2 + δ 2 )). The parameter vector θ may be estimated using an iterative
procedure for non-linear optimization (BFGS method) of the log-likelihood function (15). The estimation of δ can be
performed using the profile log-likelihood as mentioned earlier in Section 3.1.
Remark. The moments estimators based on BBS distribution can be obtained using the moments equations below
n
1 P
t
E[T
]
=
i
n
i=1
n
P
1
2
2
ti
E[T ] = n
g(α, β, δ) =
= g(t1 , t2 , . . . , tn ),
(16)
i=1
1
n
n
E[ T ] = P
1
i=1
ti
where E[g(T1 , T2 , . . . , Tn )] − g(α, β, δ) ≡ 0 and (T1 , T2 , . . . , Tn ) is a random sample from a BBS population. However,
equations in (16) do not have analytical solutions, and it is necessary to use some numerical method. When these
estimators present analytical solutions, they are usually used as initial values in the iterative processes used to obtain
ML estimators. In this work, as initial values, we fixed δ = 0 and we used the modified moment estimators proposed
by Ng et. al. (2003) for the BS distribution.
3.2
Confidence intervals
In this subsection we present confidence intervals (CIs) for S(t; α, β, δ), E[T ] and Var[T ], where T ∼ BBS(α, β, δ) and
δ is known.
Let {Tn , n ≥ 1} be a sequence of RVs. We will say that {Tn } is asymptotically normal (AN) with mean µn and
variance σn2 , and write Tn ∼ AN (µn , σn2 ), if σn > 0 and as n → ∞,
T n − µn
−→ N (0, 1).
σn
Here µn is not necessarily the mean of Tn and σn2 , not necessarily its variance. This is, for sufficiently large n, for each
t ∈ R we can approximate the probability P(Tn 6 t) by P(Z 6 ((t − µn )/σn )) where Z is N (0, 1).
Let θ = (α, β)⊤ in Θ and ρ ∈ (0, 1). The random interval (θ(T1 , . . . , Tn ), θ(T1 , . . . , Tn )) will be called a CI at
confidence level 1 − ρ for the parameter θ, provided that
P θ(T1 , . . . , Tn ) < θ < θ(T1 , . . . , Tn ) > 1 − ρ.
In what follows, we assume ℓ(θ) holds the following standard regularity conditions:
1. The parameter space, defined by Θ, is open and ℓ(θ) has a global maximum at Θ;
2. For almost all t, the fourth-order log-likelihood derivatives with respect to the model parameters exist and are
continuous in an open subset of Θ that contains the true parameter θ;
3. The support set of t 7→ f (t; θ, δ), for θ in Θ, does not depend on θ;
4. The expected information matrix I(θ) is positive definite and finite. We remember that the information matrix
I(θ) is a 2 × 2 matrix with elements Ij,k (θ) j, k = 1, 2, defined by
∂
∂
log f (T ; θ, δ),
log f (T ; θ, δ) , θj , θk ∈ {α, β}.
Ij,k (θ) = Cov
∂θj
∂θk
These regularity conditions are not restrictive and hold for the models cited in this work.
12
3.2.1
Confidence Interval for S(t; θ, δ).
b ⊤ has normal
Let α
b and βb be ML estimates of α and β, respectively. It is known that the ML estimate of θb = (b
α, β)
asymptotic distribution, with null mean vector and asymptotic covariance matrix given by the inverse of the information
matrix I(θ). I.e.,
√ b
n(θ − θ) ∼ AN 0, [I(θ)]−1 .
Since the function θ 7→ S(t; θ, δ), ∀t > 0, is continuously differentiable, by the Delta method we have
√
b δ) − S(t; θ, δ) ∼ AN 0, JS (θ) [I(θ)]−1 JS (θ)⊤ ,
n S(t; θ,
where JS (θ) = JS (θ; t) =
h
∂
∂α S(t; θ, δ)
i
∂
∂β S(t; θ, δ)
1×2
is the Jacobian of the function θ 7→ S(t; θ, δ).
b δ) can be estimated by
As θb is a ML estimate of θ, the asymptotic variance of S(t; θ,
h
i−1
b ⊤.
b δ) ≈ JS (θ)
b I(θ)
b
JS (θ)
Var S(t; θ,
As θb is consistent (because it is a ML estimate), by Slutsky’s theorem we have
√
b δ) − S(t; θ, δ) ∼ AN 0, Var S(t; θ,
b δ) .
n S(t; θ,
(17)
b δ)]. Then, the random interval
where zρ/2 is the ρ/2-quantile of the normal distribution and σ
b2 (t) = Var[S(t; θ,
b δ) − z√ρ/2 σ
b δ) + z√ρ/2 σ
b(t)
S(t; θ,
b(t) , S(t; θ,
(18)
If 0 < ρ < 1, using (17), the CI at confidence level 1 − ρ for S(t; θ, δ) is obtained from the following identity:
!
b δ) − S(t; θ, δ)|
zρ/2
|S(t; θ,
lim P
= P(|Z| < zρ/2 ) > 1 − ρ,
< √
n→∞
σ
b (t)
n
n
n
is a CI at confidence level 1 − ρ for S(t; θ, δ), ∀t > 0.
3.2.2
Confidence Interval for E[T |θ] = E[T ].
R∞
Since T ∼ BBS(θ, δ) is a positive RV, we have the identity E[T |θ] = 0 S(t; θ, δ)dt. Using this identity and denoting
b δ)] note that (18) implies that the set
σ
b2 (t) = Var[S(t; θ,
n
o
R
R
b − z√ρ/2 ∞ σ
b + z√ρ/2 ∞ σ
E[T |θ]
b
(t)dt
<
E[T
|θ]
<
E[T
|
θ]
b
(t)dt
n 0
n 0
contains the set
n
b δ) −
S(t; θ,
Therefore, the random interval
b −
E[T |θ]
zρ/2
√ σ
b(t)
n
zρ/2
√
n
R∞
0
b δ) +
< S(t; θ, δ) < S(t; θ,
b +
σ
b(t)dt , E[T |θ]
zρ/2
√
n
R∞
0
o
zρ/2
√ σ
b(t)
n
σ
b(t)dt
.
provides us a CI at confidence level 1 − ρ for E[T |θ]. If the lower limit of the CI is negative, we will replace it with zero.
13
Confidence Interval for Var[T |θ] = Var[T ].
√
b δ)], t > 0. Assume that L
b δ) ± zρ/2 σ
b ± (t) = S(t; θ,
b − (t) > 0, otherwise we replace
Let L
b (t)/ n where σ
b2 (t) = Var[S(t; θ,
this lower limit
n with zero.
o
b − (t) < S(t; θ, δ) < L
b + (t) and
Let A = L
3.2.3
B=
Using the identity E[T 2 |θ] = 2
R∞
0
nR
∞
0
b − (t)dt < E[T |θ] <
L
R∞
0
o
b + (t)dt .
L
n R
o
R∞
∞ b
2
b
tS(t; θ, δ)dt, let’s denote also C = 2 0 tL
− (t)dt < E[T |θ] < 2 0 tL+ (t)dt and
o
n R
R∞
∞ b
2
2
2
b
.
D = −( 0 L
+ (t)dt) < −(E[T |θ]) < −( 0 L− (t)dt)
b − (t), L
b + (t) is a random CI for S(t; θ, δ) with confidence coefficient
Note that A ⊆ B, C, D and B ∩ D = B. Hence, if L
R∞
R
R
R
b − (t)dt, ∞ L
b + (t)dt and 2 ∞ tL
b − (t)dt, 2 ∞ tL
b + (t)dt are also
1 − ρ (by Section 3.2.1), for each t > 0, then 0 L
0
0
0
(random) CIs for E[T |θ] and E[T 2 |θ] respectively, with confidence coefficient 1 − ρ each.
Since
n
o
b−, L
b+)
b+, L
b−)
IVar
=
2J(L
<
Var(T |θ)
<
2J(L
⊇
B ∩ C ∩ D
=
B ∩ C,
where J denotes the operator J(f, g) =
R∞
0
R∞
tf (t)dt − ( 0 g(t)dt)2 , we have
P(IVar ) > P(B ∩ C) > P(B) + P(C) − 1 > 1 − 2ρ.
b−, L
b + ), 2J(L
b+, L
b − ) is a (random) CI for Var(T |θ) with confidence coefficient 1 − 2ρ. Again, if the
Therefore, 2J(L
lower limit of the CI is negative, we will replace it with zero.
Remark. Analogously to that done in Subsection 3.2.1, we can construct a CI for the function log(− log(S(t; α, β, δ))).
4
Monte Carlo simulation study
A simulation study for evaluating the behavior of the estimators of the model parameters is carried out. The simulation
scenario assumes the sample sizes n ∈ {10, 50}, the values of the shape parameter as α ∈ {0.10, 0.50, 1.00, 1.50}, the
values of the asymmetric parameter as δ ∈ {−10, −5, −1, 1, 5, 10}, 10,000 MC replications, and without loss of generality,
we assume β = 1.00 in all cases. The censoring proportion is p ∈ {0.0, 0.1, 0.3}; see Section 3.1. Note that the values of
the shape parameter α have been chosen in order to study the performance under low, moderate and high skewness.
For each value of the parameter δ, sample size and censoring proportion, the empirical values for the bias (Bias)
and mean squared error (MSE) of the ML estimators are reported in Tables 1–2. From these tables, note that, as the
sample size increases, the ML estimators become more efficient, as expected. We can also note that, as the censoring
proportion increases, the performances of the estimators of α and β, deteriorate. It is interesting to note two points
b (i) when the skewness increases, the bias of βb increases, which is expected as the
on the increasing of the bias of β:
original distribution occurs in the BS, see for example Lemonte et al. (2008); and (ii) note that there seems to be an
increase in the bias of βb when we decrease the values of the parameter δ, see the cases δ = {−1, 1}. In general, all of
these results show the good performance of the proposed model.
14
Table 1: Simulated values of biases (MSEs within parentheses) of the estimators of the BBS model.
BBS(α = 0.1, β = 1.0, δ)
BBS(α = 0.5, β = 1.0, δ)
b
b
censoring % n
δ
Bias(b
α)
Bias(β)
Bias(b
α)
Bias(β)
0%
10%
30%
5
10
−10
−5
−1
1
5
10
−0.0011 (0.0002)
−0.0021 (0.0003)
−0.0240 (0.0012)
−0.0235 (0.0012)
−0.0022 (0.0003)
−0.0012 (0.0002)
−0.0010 (0.0009)
−0.0034 (0.0013)
−0.0418 (0.0054)
0.0480 (0.0066)
0.0050 (0.0015)
0.0020 (0.0010)
−0.0066 (0.0062)
−0.0117 (0.0075)
−0.1224 (0.0308)
−0.1192 (0.0303)
−0.0111 (0.0078)
−0.0056 (0.0063)
0.0028 (0.0184)
−0.0042 (0.0224)
−0.1581 (0.0824)
0.2824 (0.2153)
0.0293 (0.0376)
0.0169 (0.0244)
50
−10
−5
−1
1
5
10
−0.0001 (<0.0001)
−0.0003 (<0.0001)
−0.0024 (0.0002)
−0.0024 (0.0002)
−0.0003 (<0.0001)
−0.0001 (<0.0001)
−0.0001 (0.0001)
0.0002 (0.0001)
−0.0182 (0.0018)
0.0205 (0.0022)
0.0001 (0.0001)
0.0003 (0.0001)
−0.0012 (0.0004)
−0.0025 (0.0010)
−0.0139 (0.0040)
−0.0133 (0.0040)
−0.0019 (0.0010)
−0.0010 (0.0009)
0.0004 (0.0026)
0.0027 (0.0028)
−0.0597 (0.0254)
0.0927 (0.0475)
0.0019 (0.0028)
0.0034 (0.0026)
10
−10
−5
−1
1
5
10
0.0039 (0.0009)
0.0070 (0.0013)
−0.0006 (0.0011)
0.0012 (0.0011)
−0.0029 (0.0004)
−0.0007 (0.0004)
−0.0241 (0.0081)
−0.0466 (0.0133)
−0.1308 (0.0183)
0.1533 (0.0258)
0.0002 (0.0043)
−0.0047 (0.0033)
−0.0084 (0.0083)
−0.0184 (0.0112)
0.0023 (0.0301)
0.0431 (0.0381)
−0.0233 (0.0095)
−0.0059 (0.0079)
0.0393 (0.0392)
0.0276 (0.0575)
−0.4827 (0.2416)
1.0269 (1.1883)
0.0376 (0.1120)
0.0281 (0.1520)
50
−10
−5
−1
1
5
10
0.0019 (0.0002)
0.0070 (0.0007)
0.0091 (0.0003)
0.0116 (0.0004)
0.0015 (0.0003)
0.0006 (0.0001)
0.0109 (0.0022)
−0.0027 (0.0062)
−0.1387 (0.0195)
0.1632 (0.0271)
−0.0118 (0.0031)
−0.0094 (0.0008)
0.0053 (0.0011)
0.0095 (0.0015)
0.0391 (0.0083)
0.0904 (0.0171)
0.0001 (0.0011)
0.0003 (0.0011)
0.0659 (0.0074)
0.1094 (0.0170)
−0.5053 (0.2571)
1.0871 (1.2133)
−0.0680 (0.0078)
−0.0423 (0.0040)
10
−10
−5
−1
1
5
10
0.0875 (0.0123)
0.0811 (0.0106)
0.0031 (0.0013)
0.0004 (0.0012)
0.0410 (0.0062)
0.0679 (0.0102)
−0.2085 (0.0677)
−0.2206 (0.0638)
−0.1311 (0.0187)
0.1497 (0.0250)
−0.0509 (0.0390)
−0.1303 (0.0538)
0.0195 (0.0232)
0.0035 (0.0255)
−0.0680 (0.0286)
0.0521 (0.0550)
0.0283 (0.0343)
0.0307 (0.0290)
0.0854 (0.1273)
0.0741 (0.1459)
−0.4564 (0.2194)
1.0430 (1.4937)
0.1593 (0.4103)
0.0776 (0.2059)
50
−10
−5
−1
1
5
10
0.1319 (0.0185)
0.1144 (0.0139)
0.0126 (0.0005)
0.0124 (0.0005)
0.0657 (0.0097)
0.0913 (0.0145)
−0.2877 (0.0945)
−0.2838 (0.0828)
−0.1400 (0.0199)
0.1635 (0.0273)
−0.0825 (0.0538)
−0.1506 (0.0695)
0.0324 (0.0037)
0.0483 (0.0082)
−0.0200 (0.0056)
0.1065 (0.0219)
0.0101 (0.0039)
0.0189 (0.0066)
0.1056 (0.0234)
0.1750 (0.0597)
−0.4831 (0.2352)
1.0849 (1.2121)
−0.0336 (0.0613)
−0.0005 (0.0656)
Real data analysis
The proposed BBS model is now used to analyse three lifetime data sets. For comparison, the results of the bimodal
BBSO model (bimodal BS distribution proposed by Olmos et al. (2017)) and the mixture BS (MXBS) distribution
introduced by Balakrishnan et al. (2011), in addition to classical BS (BS) and log-normal (LN) models, are given as
well.
Example 5.1. The first data set corresponds to the duration of the eruption for the Old Faithful geyser in Yellowstone
National Park, Wyoming, USA; see Azzalini and Bowman (1990). Descriptive statistics for the Old Faithful data set are
the following: 272(sample size), 43(minimum), 96(maximum), 76(median), 70.897(mean), 13.595(standard deviation),
19.176(coefficient of variation), −0.414(coefficient of skewness) and −1.156(coefficient of kurtosis). Table 3 reports the
ML estimates, computed by the BFGS method, standard errors (SEs) and log-likelihood (log-lik) values for the BBS,
BBSO, MXBS, BS and LN models. Furthermore, we report the Akaike (AIC) and Bayesian information (BIC) criteria.
15
Table 2: Simulated values of biases (MSEs within parentheses) of the estimators of the BBS model.
BBS(α = 1.0, β = 1.0, δ)
BBS(α = 1.5, β = 1.0, δ)
b
b
censoring % n
δ
Bias(b
α)
Bias(β)
Bias(b
α)
Bias(β)
0%
10%
30%
10
−10
−5
−1
1
5
10
−0.0162 (0.0263)
−0.0258 (0.0326)
−0.2455 (0.1298)
−0.2441 (0.1295)
−0.0269 (0.0314)
−0.0160 (0.0253)
0.0202 (0.0507)
0.0070 (0.0575)
−0.2136 (0.1767)
0.5896 (0.9632)
0.0728 (0.1911)
0.0367 (0.0831)
−0.0244 (0.0608)
−0.0477 (0.0779)
−0.3764 (0.3075)
−0.3801 (0.3087)
−0.0411 (0.0714)
−0.0258 (0.0588)
0.0248 (0.0728)
0.0166 (0.0859)
−0.2438 (0.2449)
0.8931 (2.3962)
0.0980 (0.3867)
0.0537 (0.1826)
50
−10
−5
−1
1
5
10
−0.0023 (0.0037)
−0.0042 (0.0042)
−0.0391 (0.0179)
−0.0385 (0.0175)
−0.0035 (0.0042)
−0.0033 (0.0037)
0.0014 (0.0073)
0.0046 (0.0081)
−0.0630 (0.0486)
0.1410 (0.1272)
0.0029 (0.0078)
0.0057 (0.0072)
−0.0033 (0.0083)
−0.0049 (0.0093)
−0.0703 (0.0456)
−0.0728 (0.0448)
−0.0050 (0.1832)
−0.0030 (0.0115)
0.0012 (0.0105)
0.0058 (0.0114)
−0.0616 (0.0618)
0.1612 (0.1832)
0.0021 (0.0115)
0.0068 (0.0107)
10
−10
−5
−1
1
5
10
0.0088 (0.0575)
−0.0074 (0.0713)
0.0005 (0.1111)
0.1053 (0.1851)
−0.0250 (0.0393)
−0.0142 (0.0276)
0.1620 (0.4960)
0.2010 (0.2840)
−0.6838 (0.4792)
2.4846 (7.1727)
0.0338 (0.5171)
0.0270 (0.1993)
0.0456 (0.1813)
0.0573 (0.2973)
−0.0498 (0.1986)
0.1370 (0.3345)
−0.0236 (0.1127)
−0.0380 (0.0587)
0.2316 (0.6773)
0.4123 (1.6875)
−0.7573 (0.5941)
3.7838 (18.1242)
0.0605 (1.5485)
0.0121 (0.2071)
50
−10
−5
−1
1
5
10
0.0187 (0.0053)
0.0219 (0.0060)
0.0746 (0.0268)
0.2470 (0.0974)
0.0028 (0.0043)
0.0057 (0.0043)
0.1260 (0.0256)
0.2016 (0.0569)
−0.7159 (0.5141)
2.5919 (6.9146)
−0.1110 (0.0205)
−0.0619 (0.0103)
0.0272 (0.0115)
0.0314 (0.0134)
0.0963 (0.0492)
0.3733 (0.2098)
0.0004 (0.0092)
0.0077 (0.0093)
0.1458 (0.0361)
0.2368 (0.0820)
−0.7910 (0.6272)
3.4797 (12.8450)
−0.1286 (0.0275)
−0.0763 (0.0146)
10
−10
−5
−1
1
5
10
0.2393 (0.3985)
0.1985 (0.2985)
−0.1458 (0.1069)
0.1667 (0.3212)
0.0959 (0.3339)
0.0774 (0.1579)
0.6891 (4.0713)
0.6169 (1.9729)
−0.6624 (0.4552)
2.4961 (7.9861)
0.4870 (5.5723)
0.1944 (0.9839)
0.4426 (1.2441)
0.7262 (2.1753)
−0.3442 (0.2392)
0.2472 (0.5322)
0.2082 (1.0045)
0.1840 (0.6911)
1.0896 (7.0831)
1.9417 (12.9445)
−0.7627 (0.5970)
3.5858 (15.6473)
0.7895 (11.7672)
0.4070 (4.2938)
50
−10
−5
−1
1
5
10
0.1884 (0.3646)
0.2240 (0.1207)
−0.1144 (0.0284)
0.3004 (0.1388)
0.0436 (0.0771)
0.0347 (0.0223)
0.6260 (8.5780)
0.6080 (0.7797)
−0.6953 (0.4855)
2.5498 (6.7053)
0.0425 (1.5990)
−0.0014 (0.4995)
0.1899 (0.1168)
0.4197 (0.4973)
−0.2762 (0.1004)
0.4939 (0.3467)
0.0406 (0.0558)
0.0410 (0.0156)
0.3775 (0.4992)
0.9224 (2.5520)
−0.7957 (0.6349)
3.5056 (12.8719)
−0.0484 (0.9883)
−0.0309 (0.0156)
From this table, we note that the BBS and MXBS models provide better adjustments compared to the other models based
on the values of AIC and BIC. The null hypothesis of a BS distribution (δ = 0) against an alternative BBS distribution
b − ℓBBS (b
b δ)).
b In this case, we obtain
(δ 6= 0) can be tested by using the likelihood ratio (LR) test LR = −2(ℓBS (b
α, β)
α, β,
LR = −2(−1107.849 + 1050.592) = 114.514 and comparing it to the 5% critical value from the chi-square distribution
with one degree of freedom (χ21 = 3.84), it supports rejection of the null hypothesis, thus the BBS model outperforms,
in terms of fitting, the BS one for the data under study.
Figure 3 shows the histogram of the data set superimposed with the fitted curves of the BBS, BBSO, MXBS, BS and
LN distributions. From this figure, we clearly note that the BBS captures quite well the inherent bimodality of the data.
Example 5.2. The data used here, which are given by Andrews and Herzberg (1985) who attribute them to a study by
Barlow et al. (1984), present the stress rapture life in hours of Kevlar-49/epoxy strands when subjected to a constant
sustained pressure until failure. A descriptive summary for the Kevlar-49/epoxy data set provides the following values:
16
Table 3: ML estimates and model selection measures for fit to the Old Faithful data.
Model
Parameter
ML estimate
log-lik
AIC
BIC
0.0034
0.4739
−1050.592
2107.184
2118.001
BBSO
α
β
γ
0.0893
65.7730
−2.1803
0.0047
0.4128
0.1432
−1054.396
2114.792
2125.609
MXBS
α1
α2
β1
β2
p
0.1150
0.0697
54.8174
80.1850
0.3762
0.0046
0.0108
0.4824
0.7607
0.0317
−1032.681
2075.362
2093.391
BS
α
β
0.2055
69.4289
0.0088
0.8608
−1107.849
2221.698
2226.91
LN
µ
σ
4.2411
0.2048
0.0124
0.0087
−1108.300
2222.6
2227.812
0.000.010.020.030.04 0.05 0.060.07
α
β
δ
PSfrag replacements
BBS
PDF
0.1255
66.8612
−4
SE
BBS
BBSO
MXBS
BS
LN
40
50
60
70
80
waiting times
90
100
Figure 3: Histogram of waiting times until the next eruption (from the Old Faithful data) overlaid with the fitted
densities.
49(sample size), 1051(minimum), 17568(maximum), 8831(median), 8805.694(mean), 4553.915 (standard deviation),
51.176(coefficient of variation), 0.094(coefficient of skewness) and −0.915(coefficient of kurtosis).
Table 4 reports the ML estimates, SEs and log-lik values associated with the BBS, BBSO, MXBS, BS and LN models.
Furthermore, we report the values of AIC and BIC. From this table, observe that the proposed BBS model has the lowest
values for the AIC and BIC, suggesting that this model provides the best fit to Kevlar-49/epoxy data. To test the null
17
hypothesis of a BS distribution (δ = 0) against an alternative BBS distribution (δ 6= 0), we use the LR test. The result
LR = −2(−488.4345 + 480.049) = 16.771 supports the BBS model assumption and rejects the BS model for this data
set. This result suggest that the BBS distribution is indeed a good model for the Kevlar-49/epoxy data. A graphical
comparison of the fitted BBS, BBSO, MXBS, BS and LN distributions is given in Figure 4.
Table 4: ML estimates and model selection measures for fit to the Kevlar-49/epoxy data.
Model
Parameter
ML estimate
SE
log-lik
AIC
BIC
0.5933
4507.365
−2
0.0504
376.7557
−480.049
966.098
971.773
0.4679
5036.113
−1.5345
0.08655
437.0908
0.4708
−490.208
986.417
992.092
BBS
α
β
δ
BBSO
α
β
γ
MXBS
α1
α2
β1
β2
p
0.3412
0.0697
10278.38
4577.872
0.6708
0.0166
0.0123
979.1201
50.2859
0.0991
−593.987
1197.974
1207.433
BS
α
β
0.7520
6800.546
0.0759
679.8509
−488.434
982.869
984.652
LN
µ
σ
8.8925
0.7012
0.1001
0.0708
−487.873
981.746
983.530
Example 5.3. The third data set corresponds to the lifetimes of adult flies in days after exposure to a pest control
technique, which consists of using small portions of food laced with an insecticide that kills the flies. The experiment
was carried out at the Department of Entomology of the Luiz de Queiroz School of Agriculture, University of São Paulo,
Brazil. In this technique, the period was set at 51 days such that larvae that survived beyond this period are considered as
censored cases; see Silva et al. (2013) for more details about this experiment. Descriptive statistics for the Entomology
data are the following: sample size = 172 (four cases are lost), minimum = 1.000, maximum = 51.000, median = 21.000,
mean = 21.878, standard deviation = 11.674, coefficient of variation = 53.30, coefficient of skewness = 0.818 and
coefficient of kurtosis = 0.569.
The ML estimates and log-lik values for the BBS, BBSO, MXBS, BS and LN models are reported in Table 5.
Furthermore, the AIC and BIC values are also reported in this table. From Table 5, we note that the proposed BBS
model has the lowest AIC and BIC values, and therefore it could be chosen as the best model. Using the LR statistic
to compare the fits of the BS and BBS models, that is, the null hypothesis of a BS distribution (δ = 0) against an
alternative BBS distribution (δ 6= 0), we obtain LR = −2(−676.913 + 610.523) = 132.780 and then we could accept the
BBS model. Figure 5 shows the fitted PDFs and SFs (by Kaplan-Meier (KM) estimator) of the BBS, BBSO, BS and
LN distributions.
18
5e-5 10e-5 15e-5 20e-5 25e-5
PDF
0
log-lik
AIC
BIC
BBS
α
β
δ
0.6922
8.7636
−2
0.0385
0.5311
−610.523
1227.052
1236.494
BBSO
α
β
γ
0.4975
7.5959
−2.2434
0.0437
0.5250
0.2859
−663.144
1332.287
1341.730
MXBS
α1
α2
β1
β2
p
1.2349
0.2104
14.6760
19.9960
0.60
0.0115
0.0009
2.7919
0.4717
−631.137
1266.273
1272.568
BS
α
β
0.8912
16.1512
0.0500
1.0078
−676.913
1357.862
1364.157
LN
µ
σ
2.9139
0.7613
0.0583
0.0708
−660.230
1324.461
1330.756
19
0
SE
1
ML estimate
2
Parameter
3
Model
4
Table 5: ML estimates and model selection measures for fit to the Entomology data.
5
Figure 4: Histogram of Kevlar 49/epoxy strands failure times (70% pressure) overlaid with the fitted densities.
6
15000
8
10000
failure times
10
5000
12
15
20
25
30
0.00
0.05
0.10
0.15
0.20
0.25
0.0
0.30
0.2
0.3
0.4
0.5
0.6
0.8
1.0
1.2
1.5
2.0
2.5
3.0
BBS
BBSO
MXBS
BS
LN
1.00
0.80
0.25
0.20
SF
0.40 0.60
0.0
0.00
0.05
0.20
PDF
0.10 0.15
KM
BBS
BBSO
MXBS
BS
LN
0
10
20
30
40
survival times
50
0
10
20
30
40
survival times
50
Figure 5: Histogram and SF fitted by KM with the Entomology data.
6
Concluding remarks
In this work, we have introduced a bimodal generalization of the Birnbaum-Saunders distribution, based on the alphaskew-normal distribution. We have discussed some of its properties. We have considered estimation and inference based
on likelihood methods. We have carried out a Monte Carlo simulation study to evaluate the behavior of the maximum
likelihood estimators of the corresponding parameters. Three real data sets were considered to illustrate the potentiality
of the proposed model. In general, the results have shown that the proposed bimodal Birnbaum-Saunders distribution
outperforms some existing models in the literature. As part of future research, it is of interest to study univariate and
multivariate bimodal Birnbaum-Saunders regression models; see Rieck and Nedelman (1991), Balakrishnan and Zhu
(2015) and Marchant et al. (2016). Moreover, time series models based on the bimodal Birnbaum-Saunders distribution
with corresponding influence diagnostic tools can also be considered; see Saulo et al. (2017). Work on these problems
is currently under progress and we hope to report these findings in a future paper.
Acknowledgments
The authors thank the Editors and reviewers for their constructive comments on an earlier version of this manuscript.
The research was partially supported by CNPq and CAPES grants from the Brazilian government.
References
Andrews, D. F. and Herzberg, A. M. (1985) Data: A Collection of Problems from Many Fields for the Student and
Research Worker (Springer Series in Statistics). Springer: New York. 10.1007/978-1-4612-5098-2
Azzalini, A. and Bowman, A. W. (1990) A look at some data on the Old Faithful geyser. Applied Statistics, 39, 357–365.
doi:10.2307/2347385
Azzalini A. and Capitanio A. (2003) Distributions generate by perturbation of symmetry with emphasis on a multivariate
skew-t distribution. JJ. R. Stat. Soc. Ser. B Stat. Methodol, 65, 367–389.
Balakrishnan, N., Leiva V. and López, J. (2007) Acceptance sampling plans from truncated life tests based on the
generalized Birnbaum-Saunders distribution. Communications in Statistics - Simulation and Computation, 36, 643–
656. http://dx.doi.org/10.1080/03610910701207819
20
PSfrag replacements
0.00
0.05
0.10
0.15
0.20
0.25
PSfrag replacements
PDF
survival times
0.20
0.40
0.60
0.80
1.00
SF
KM
survival times
BBS
BBSO
MXBS
BS
LN
Balakrishnan, N., Leiva, V., Sanhueza, A. and Vilca, F. (2009) Estimation in the Birnbaum-Saunders distribution based
on scale-mixture of normals. Statistics and Operations Research Transactions, 33, 171–192.
Balakrishnan, N., Gupta, R. C., Kundu, D., Leiva, V. and Sanhueza, A. (2011) On some mixture models based on
the Birnbaum-Saunders distribution and associated inference. Journal of Statistical Planning and Inference, 141,
2175–2190.
Balakrishnan, N. and Zhu, X. (2015) Inference for the Birnbaum-Saunders lifetime regression model with applications. Communications in Statistics - Simulation and Computation, 44, 2073–2100. http://dx.doi.org/10.1080/
03610918.2013.844838
Barlow, R. E., Toland, R. H. and Freeman, T. (1984) A Bayesian analysis of stress-rupture life of Kevlar 49/epoxy
spherical pressure vessels. In T. D. Dwivedi (Ed.), Proceedings of the Canadian Conference in Applied Statistics.
Marcel Dekker: New York.
Bhatti, C. R. (2010) The Birnbaum-Saunders autoregressive conditional duration model. Mathematics and Computers
in Simulation, 80, 2062–2078.
Birnbaum, Z. W. and Saunders, S. C. (1969) A new family of life distributions. Journal of Applied Probability, 6,
319–327.
Chung, K. L. (2001) A Course in Probability Theory. Third edition. Academic Press, Inc.: San Diego, CA.
Dı́az-Garcı́a, J. A. and Leiva, V. (2005) A new family of life distributions based on elliptically contoured distributions.
Journal of Statistical Planning and Inference, 128, 445–457.
Elal-Olivero, D. (2010) Alpha-skew-normal distribution. Proyecciones Journal of Mathematics, 29, 224–240.
Fonseca, R. and Cribari-Neto, F. (2016) Inference in a bimodal Birnbaum-Saunders model. ArXiv e-prints. Available
at https://arxiv.org/pdf/1609.00339.pdf.
Glaser, R. E. (1980) Bathtub and related failure rate characterizations. Journal of the American Statistical Association,
75, 667–672.
Gómez, H. W., Elal-Olivero D., Salinas H. S. and Bolfarine H. (2011) Bimodal extension based on the skew-normal
distribution with application to pollen data. Environmetrics, 22, 50–62.
Kim, H. J. (2005) On a class of two-piece skew-normal distributions. Statistics, 39, 537–553.
Leao, J., Leiva, V., Saulo, H., Tomazella, V. (2017) Birnbaum-Saunders frailty regression models: Diagnostics and
application to medical data. Biometrical Journal, 59, 291–314.
Leiva, V. (2016) The Birnbaum-Saunders Distribution. Elsevier/Academic Press: Amsterdam.
Leiva, V., Marchant, C., Saulo, H., Aslam, M. and Rojas, F. (2014a) Capability indices for Birnbaum-Saunders processes
applied to electronic and food industries. Journal of Applied Statistics, 41, 1881–1902.
Leiva, V., Saulo, H., Leão, J. and Marchant, C. (2014c) A family of autoregressive conditional duration models applied
to financial data. Computational Statistics and Data Analysis, 79, 175–191.
Lemonte, A. J., Simas, A. B. and Cribari-Neto, F. (2008) Bootstrap-based improved estimators for the two-parameter
Birnbaum-Saunders distribution. Journal of Statistical Computation and Simulation, 78, 37–49.
21
Lin, T. I., Lee, J. C., Hsieh, W. J. (2007a) Robust mixture models using the skew-t distribution. Statistics and Computing, 17, 81–92.
Lin, T. I., Lee, J. C. and Yen, S. Y. (2007b) Finite mixture modeling using the skew-normal distribution. Statistica
Sinica, 17, 81–92.
Ling, J. and Pan, J. (1998) A new method for selection of population distribution and parameter estimation. Reliability
Engineering and System Safety, 60, 247–255. https://doi.org/10.1016/S0951-8320(97)00171-3
Ma, Y. and Genton, M. G. (2004) Flexible class of skew-symmetric distributions. Scandinavian Journal of Statistics,
31, 459–468.
Marchant, C., Leiva, V. and Cysneiros, F.J.A. (2016) A multivariate log-linear model for Birnbaum-Saunders distributions. IEEE Transactions on Reliability, 65, 816–827. 10.1109/TR.2015.2499964
Natanson, I. P. (1955) Theory of functions of a real variable (Translated by Leo F. Boron with the collaboration of Edwin
Hewitt). Frederick Ungar Publishing Co.: New York.
Ng. H.K.T., Kundu, D. and Balakrishnan (2003) Modified moments estimation for the two-parameter BirnbaumSaunders distribution Computational Statistics & Data Analysis, 43, 283–298.
Olmos, N. O., Martı́nez-Flórez and M., Bolfarine, H. (2017) Bimodal Birnbaum-Saunders distribution with applications
to non-negative measurements. Communications in Statistics - Theory and Methods, 46, 6240–6257.
Paula, G. A., Leiva, V., Barros M. and Liu, S. (2012) Robust statistical modeling using the Birnbaum-Saunders-t
distribution applied to insurance. Applied Stochastic Models in Business and Industry, 28, 16–34.
Rieck, J. R. and Nedelman, J. R. (1991) A log-linear model for the Birnbaum-Saunders distribution. Technometrics,
33, 51–60. 10.2307/1269007
Saulo, H., Leiva, V., Ziegelmann, F.A. and Marchant, C. (2010) A nonparametric method for estimating asymmetric
densities based on skewed Birnbaum-Saunders distributions applied to environmental data. Stochastic Environmental
Research and Risk Assessment, 27, 1479–1491. https://doi.org/10.1007/s00477-012-0684-8
Saulo, H., Leao, J., Leiva, V. and Aykroyd, R. G. (2017) Birnbaum-Saunders autoregressive conditional duration models applied to high-frequency financial data. Statistical Papers. Available at https://doi.org/10.1007/
s00362-017-0888-6.
Shannon C. E. and Weaver, W. (1949) The Mathematical Theory of Communication. Univ. of Illinois Press, Urbana,
IL.
Silvia M. A., Bezerra-Silva G. C. D, Vendramim J. D. and Mastrangelo T. (2013) Sublethal effect of neem extract on
Mediterranean fruit fly adults. Revista Brasileira de Fruticultura, 35, 93–101.
Vilca, F., Sanhueza, A., Leiva, V. and Christakos, G. (2010) An extended Birnbaum-Saunders model and its application
in environmental quality in Santiago. Stochastic Environmental Research and Risk Assessment, 24, 771–782. https:
//doi.org/10.1007/s00477-009-0363-6
Vinberg, E. B. (2003) A course in algebra. American Mathematical Society, Providence, R.I., ISBN 0-8218-3413-4 .
22
| 10math.ST
|
Covert Communications on Poisson Packet
Channels
Ramin Soltani∗ , Dennis Goeckel∗ , Don Towsley† , and Amir Houmansadr†
∗
Electrical and Computer Engineering Department, University of Massachusetts, Amherst,
arXiv:1610.00381v2 [cs.IT] 28 Nov 2017
{soltani, goeckel}@ecs.umass.edu
†
College of Information and Computer Sciences, University of Massachusetts, Amherst,
{towsley, amir}@cs.umass.edu
Abstract
Consider a channel where authorized transmitter Jack sends packets to authorized receiver Steve according
to a Poisson process with rate λ packets per second for a time period T . Suppose that covert transmitter Alice
wishes to communicate information to covert receiver Bob on the same channel without being detected by a
watchful adversary Willie. We consider two scenarios. In the first scenario, we assume that warden Willie cannot
look at packet contents but rather can only observe packet timings, and Alice must send information by inserting
her own packets into the channel. We show that the number of packets that Alice can covertly transmit to Bob
is on the order of the square root of the number of packets that Jack transmits to Steve; conversely, if Alice
transmits more than that, she will be detected by Willie with high probability. In the second scenario, we assume
that Willie can look at packet contents but that Alice can communicate across an M/M/1 queue to Bob by
altering the timings of the packets going from Jack to Steve. First, Alice builds a codebook, with each codeword
consisting of a sequence of packet timings to be employed for conveying the information associated with that
codeword. However, to successfully employ this codebook, Alice must always have a packet to send at the
appropriate time. Hence, leveraging our result from the first scenario, we propose a construction where Alice
covertly slows down the packet stream so as to buffer packets to use during a succeeding codeword transmission
This work has been supported by the National Science Foundation under grants ECCS-1309573 and CNS-1525642.
This work has been presented at the 53rd Annual Allerton Conference on Communication, Control, and Computing, October 2015.
The extension of this work to Renewal packet channels is [1], which has been presented at the 54th Annual Allerton Conference on
Communication, Control, and Computing, October 2016.
Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media,
including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. DOI: 10.1109/ALLERTON.2015.7447124
1
phase. Using this approach, Alice can covertly and reliably transmit O(λT ) covert bits to Bob in time period T
over an M/M/1 queue with service rate µ > λ.
Keywords: Covert Bits Through Queues, Covert Communication, Covert Communication over Queues,
Covert Computer Network, Covert Wired Communication, Low Probability of Detection, LPD, Covert
Channel, Covert Timing Channel, Covert Packet Insertion, Poisson Point Process, Information Theory.
I. I NTRODUCTION
The provisioning of security and privacy is critical in modern communication systems. The vast
majority of research in this area has focused on protecting the message content through encryption [2] or
emerging methods in information-theoretic security [3]. However, there are applications where the very
existence of the message must be hidden from a watchful adversary. For example, in the organization of
social unrest against an authoritarian regime, the mere existence of an encrypted message would likely
lead to the shut down of that communication and possible punishment of the users. Other applications
include hiding the presence or scale of military operations, or removing the ability of users to be tracked
in their everyday activities.
Whereas the provisioning of undetectable communications in objects with symbols drawn from
a finite field and transmitted over noiseless channels (e.g. images) has been extensively studied in
the steganogaphic community over a number of years [4], the consideration of hiding messages on
continuous-valued noisy channels has only recently drawn considerable attention. In particular, despite
the extensive historical use of spread spectrum systems for low probability of detection (LPD) communications, the fundamental limits of covert communications over the additive white Gaussian noise
(AWGN) channel were only recently obtained [5], [6]. The work of [5], [6] motivated significant further
work (e.g. [7]–[19]), with these works taken together providing an almost complete characterization
of the foundational limits of covert communications over discrete memoryless channels (DMCs) and
AWGN channels.
In this paper, we turn our attention to the consideration of a covert timing channel appropriate
for packet-based communication channels, e.g., computer networks. More specifically, we study how
packet insertion or inter-packet delays can be used for covert communications. The use of inter-packet
delays for covert communication was first explored by Girling [20] and was later studied by a number of
other authors [21]–[24]. In particular, Anantharam et al. [25] derived the Shannon capacity of the timing
channel with a single-server queue, and Dunn et al. [26] analyzed the secrecy capacity of such a system.
Considerable work has focused on quantifying and optimizing the capacity of timing channels [27]–
[32] by leveraging information theoretic analysis and the use of various coding techniques [33]–[35].
2
Moreover, methods for detecting timing channels [36] as well as eluding the detection by leveraging
the statistical properties of the legitimate channel have been proposed [37].
In this paper, we consider a channel where an authorized (overt) transmitter, Jack, sends packets to
an authorized (overt) receiver, Steve, within a time period T , where the timings of packet transmission
are modeled by a Poisson point process with rate λ packets per second. Covert transmitter Alice may
wish to transmit data to a covert receiver, Bob, on this channel in the presence of a warden, Willie, who
is monitoring the channel between Alice and Bob precisely to detect such transmissions. We consider
two scenarios in detail. In the first scenario, we assume: (1) the warden Willie is not able to see packet
contents, and therefore cannot verify the authenticity of the packets (e.g., whether they are actually sent
by Jack); and (2) Alice is restricted to packet insertion. Willie is aware that the timing of the packets
of the overt communication link follows a Poisson point process with rate λ, so he seeks to apply
hypothesis testing to verify whether the packet process has the proper characteristics. We show that (see
√
λT packets to Bob, she can keep Willie’s sum of error
Theorem 1) if Alice decides to transmit O
probabilities PF A + PM D > 1 − for any 0 < < 1, where PF A and PM D are Willie’s probability of
√
λT
false alarm and missed detection, respectively. Conversely, we prove that if Alice transmits ω
packets, she will be detected by Willie with high probability.
In the second scenario, we assume that Willie can look at packet contents and therefore can verify
packets’ authenticity. Thus, Alice is not able to insert packets, but we allow Alice in this scenario
to alter the packet timings to convey information to Bob, whom is receiving the packets through a
M/M/1 queue with service rate µ > λ. To do such, Alice designs an efficient code [25], where a
codeword consists of a sequence of packet timings drawn from the same process as the overt traffic;
hence, a codeword transmitted with those packet timings is undetectable. However, there is a causality
constraint, as Alice clearly cannot send the next packet (i.e. codeword symbol) unless she has a packet
from the Jack to Steve link available to transmit. This suggests the following two-stage process. In
the first stage, Alice covertly slows down the transmission of the packets from Jack to Steve so as to
buffer some number of packets. In the second stage, Alice continues to add packets transmitted by Jack
to her buffer while releasing packets with the inter-packet delay appropriate for the chosen codeword.
Alice’s scheme breaks down when her buffer is empty at any point before completing the codeword
transmission. Hence, the question becomes: how long must Alice collect packets during the first stage
so as to guarantee (with high probability) that she will not run out of packets before the completion of
codeword transmission during the second stage?
First in Lemma 1, we show that Alice can collect and store O
√
λT
packets in an interval of length
T while keeping Willie’s sum of error probabilities PF A +PM D > 1− for any 0 < < 1; conversely, if
3
she collects more, she will be detected by Willie with high probability. Building on Lemma 1, we prove
(Theorem 2) that, using our two-stage covert communications approach, Alice can reliably transmit
O(λT ) covert bits to Bob within the time period T while keeping Willie’s sum of error probabilities
bounded as PF A + PM D > 1 − for any 0 < < 1.
The remainder of the paper is organized as follows. In Section II, we present the system model and
the metrics employed for the two scenarios of interest. Then, we provide constructions and their analysis
for the two covert communication scenarios in Sections III and IV.
II. S YSTEM M ODEL AND M ETRICS
A. System Model
Suppose that Jack transmit packets to Steve on the interval [0, T ]. We model the packet transmission
times by a Poisson point process with parameter λ. Hence, the average number of packets that is
sent from Jack to Steve in a duration of time T is λT . Assume Alice wishes to use the channel to
communicate with Bob without being detected by warden Willie, whom is observing the channel from
Alice to Bob and attempting to detect any covert transmissions.
In Scenario 1, which is analyzed in Section III, we assume that Willie does not have access to the
contents of the packets and therefore is not able to perform authentication and determine whether a
packet is from Jack or not. We assume that Alice, with knowledge of λ, is allowed to use the channel
only by inserting and transmitting her own packets to Bob, and Bob is able to authenticate, receive and
remove the covert packets. Therefore Steve does not observe the covert packets. We consider the rate ∆ at
which Alice can insert packets while remaining covert. Willie knows that the legitimate communication
is modeled by a Poisson point process with parameter λ, and he knows all of the characteristics of
Alice’s insertion scheme (rate, method of insertion, etc.)
In Scenario 2, which is analyzed in Section IV, we assume that Willie is able to access packet contents
and hence authenticate whether a packet is coming from Jack. Hence, Alice is not able to insert packets
into the channel. However, we allow Alice to buffer packets and release them when she desires into the
channel, thereby enabling covert communication through packet timing control. The problem is made
non-trivial by assuming that Bob has access to the resulting packet stream only through an M/M/1
queue with service rate µ > λ; in other words, the service times of the packets are independently
and identically distributed (i.i.d) random variables according to an exponential random variable with
average 1/µ. Again, Willie knows that the legitimate communication is modeled by a Poisson point
process with parameter λ, and he knows all of the characteristics of Alice’s packet buffering and release
scheme except a secret key that is pre-shared between Alice and Bob. Fig. 2 depicts the system model
4
Fig. 1. System configuration for Scenario 1: Alice adds her packets to the channel between Jack and Steve to communicate covertly with
Bob. The average number of packets from Jack in a duration of time T is λT , and the average number of total packets observed by either
Bob and Willie in T is (λ + ∆) T . The blue color shows the legitimate communication and the green shows the covert communication.
Fig. 2. System configuration for Scenario 2: Alice is able to buffer packets in order to alter packet timings.
for this scenario. In this scenario, we consider the rate at which Alice can reliably transmit information
to Bob without detection by Willie.
B. Hypothesis Testing
Willie is faced with a binary hypothesis test: the null hypothesis (H0 ) corresponds to the case that Alice
is not transmitting, and the alternative hypothesis H1 corresponds to the case that Alice is transmitting.
We denote the distribution Willie observes under H1 by P1 , and the distribution Willie observes under
H0 by P0 . Also, we denote by PF A the probability of rejecting H0 when it is true (type I error or false
alarm), and PM D the probability of rejecting H1 when it is true (type II error or missed detection). We
assume that Willie uses classical hypothesis testing with equal prior probabilities and seeks to minimize
PF A + PM D ; the generalization to arbitrary prior probabilities is straightforward [6].
C. Covertness
Alice’s transmission is covert if and only if she can lower bound Willie’s sum of probabilities of
error (PF A + PM D ) by 1 − for any > 0 [6].
5
D. Reliability
A transmission scheme is reliable if and only if the probability that a codeword transmission from
Alice to Bob is unsuccessful is upper bounded by ζ for any ζ > 0. Note that this metric applies only
in Scenario 2.
III. C OVERT C OMMUNICATION VIA PACKET I NSERTION (S CENARIO 1)
In this section, we consider Scenario 1: Willie cannot authenticate packets to see whether they are
from Jack or Alice, and Alice is only allowed to send information to Bob by adding packets to the
channel.
√
Theorem 1. Consider Scenario 1. Alice can covertly insert O( λT ) packets in the period of time T .
√
λT packets in the period of time T , there exists a detector
Conversely, if Alice attempts to insert ω
that Willie can use to detect her with arbitrarily low sum of error probabilities PF A + PM D .
Proof. (Achievability)
Construction: Alice generates a Poisson point process with parameter ∆ and based on this, adds her
covert packets to the channel between Jack and Steve. Bob collects and removes the packets generated
by Alice.
Analysis: First we show that the number of packets observed during the time period of length T is
a sufficient statistic by which Willie can perform the optimal hypothesis test to decide whether Alice
is transmitting or not. The observation of the Poisson point process at Willie is characterized by the
collection of inter-packet delays. Let x̄ = (x1 , x2 , · · · ) be this observation vector, where xi is the delay
between the arrival time of packet i and packet i+1. The optimal hypothesis test is based on a likelihood
ratio (LRT) between the null hypothesis H0 (Alice is not transmitting) and the alternative hypothesis
H1 (Alice is transmitting):
P
f (x̄|H1 , n1 )P1 (n1 )
f (x̄|H1 )
Λ(x̄) =
= Pn1
f (x̄|H0 )
n0 f (x̄|H0 , n0 )P0 (n0 )
where P0 is the probability distribution function for the number of packets that Willie observes under
the null hypothesis H0 (Alice is not transmitting) corresponding to a Poisson point process with rate
λ, and P1 is the probability distribution function for the number of packets that Willie observes under
hypothesis H1 (Alice is transmitting) corresponding to a Poisson point process with rate λ + ∆. Since
the processes under H0 and H1 are Poisson, f (x̄ |H0 , n) = f (x̄|H1 , n) for all n, and
P
f (x̄|n1 )P1 (n1 )
Λ(x̄) = Pn1
n0 f (x̄|n0 )P0 (n0 )
6
(1)
Suppose that x̄ has n non-zero entries for a given observation. Since f (x̄|n1 ) = 0 when n1 6= n,
Λ(x̄) =
f (x̄|n)P1 (n)
P1 (n)
=
f (x̄|n)P0 (n)
P0 (n)
(2)
Therefore, the number of packets that Willie observes during time interval T is a sufficient statistic for
Willie’s decision.
Now, Willie applies the optimal hypothesis test on the number of packets during time T . Then [6]
r
1
PF A + PM D ≥ 1 −
D(P0 ||P1 )
(3)
2
r
1
PF A + PM D ≥ 1 −
D(P1 ||P0 )
(4)
2
where E[·] denotes the expected value, and D(P0 ||P1 ) is the relative entropy between P0 and P1 .
We next show how Alice can lower bound the sum of average error probabilities by upper bounding
q
1
D(P0 ||P1 ). Consider the following fact (proved in the Appendix):
2
Fact 1. For two Poisson distributions Pλ1 (n), Pλ2 (n):
D(Pλ1 (n)||Pλ2 (n)) = λ2 − λ1 + λ1 log λ1 /λ2
By Fact. 1, for the given P1 and P0 the relative entropy is:
∆
D(P0 ||P1 ) = ∆ · T − λT ln 1 +
λ
(5)
(6)
We prove in the Appendix that:
ln(1 + u) ≥ u −
u2
for u ≥ 0,
2
(7)
Applying (7) on (6) yields:
D(P0 ||P1 ) ≤
∆2
T.
λ
(8)
Suppose Alice sets
r
∆=
2λ
T
(9)
where 0 < < 1 is a constant. Then,
r
1
D(P0 ||P1 ) ≤
2
√
and E[PF A + PM D ] ≥ 1 − as long as ∆ · T = O
λT .
(Converse) To establish the converse, we provide an explicit detector for Willie that is sufficient to
limit Alice’s throughput across all potential transmission schemes (i.e. not necessarily insertion according
to a Poisson point process). Suppose that Willie observes a time interval of length T and wishes to
7
detect whether Alice is transmitting or not. Since he knows that the packet arrival process for the link
between Jack and Steve is a Poisson point process with parameter λ, he knows the expected number of
packets that he will observe. Therefore, he counts the number of packets S in this duration of time and
performs a hypothesis test by setting a threshold U and comparing S with λT + U . If S < λT + U ,
Willie accepts H0 ; otherwise, he accepts H1 . Consider PF A
PF A = P (S > λT + U |H0 ) = P (S − λT > U |H0 ) ≤ P (|S − λT | > U |H0 )
(10)
When H0 is true, Willie observes a Poisson point process with parameter λ; hence,
E [S |H0 ] = λT
(11)
Var [S |H0 ] = λT
(12)
. Thus, Willie can achieve PF A ≤ α
Therefore, applying Chebyshev’s inequality on (10) yields PF A ≤ λT
U2
q
√
for any 0 < α < 1 if he sets U = λT
.
Next,
we
will
show
that
if
Alice
transmits
∆
·
T
=
ω
λT
α
covert packets, she will be detected by Willie with high probability. Consider PM D
PM D = P (S ≤ λT + U |H1 ) = P (S − (λ + ∆) T ≤ U − ∆ · T |H1 )
= P (− (S − (λ + ∆) T ) ≥ − (U − ∆ · T ) |H1 )
≤ P (|S − (λ + ∆) T | ≥ |U − ∆ · T | |H1 )
(13)
When H1 is true, since Willies observes a Poisson point process with parameter λ + ∆,
E [S |H1 ] = (λ + ∆) T
(14)
Var [S |H1 ] = (λ + ∆) T
(15)
q
λT
, if Alice
Therefore, applying Chebyshev’s inequality on (13) yields PM D ≤ (U(λ+∆)T
.
Since
U
=
α
−∆·T )2
√
sets ∆ · T = ω
λT , Willie can achieve PM D < β for any 0 < β < 1. Combined with the results
q
√
for probability of false alarm above, if Alice sets ∆.T = ω
to
λT , Willie can choose a U = λT
α
achieve any (small) α > 0 and β > 0 desired.
IV. C OVERT T IMING C HANNEL U SING I NTER -PACKET D ELAYS (S CENARIO 2)
In this section, we consider Scenario 2: Willie can authenticate packets to determine whether or not
they were generated by the legitimate transmitter Jack. Therefore, Alice cannot insert packets into the
channel; rather, we assume that Alice is able to buffer packets and release them when she desires; hence,
she can encode information in the inter-packet delays by using a secret codebook shared with Bob.
8
In the construction below, each of Alice’s codewords will consist of a sequence of inter-packet delays
to be employed to convey the corresponding message. But this immediately presents a problem for
Alice, as she must have a packet in her buffer to transmit at the appropriate time. To clarify this issue,
suppose Alice were to attempt to start codeword transmission immediately. Suppose that the first letter
of Alice’s codeword for the current message is 10µs, and thus she has to apply this delay between the
first and second packet. To send this letter, Alice receives the first packet from Jack and immediately
transmits it to Bob. Then, she waits for the second packet. If the second packet arrives earlier than 10µs,
she stores the packet and transmits it 10µs after the first transmission. But, if the second packet arrives
from Jack after the 10µs interval, she cannot accurately transmit the codeword. Hence, the codeword
failure rate would be quite high with such a scheme.
This suggests that Alice must build up some number of packets in her buffer before starting codeword
transmission, and, in fact, this is the scheme that we employ. In particular, Alice will employ a twophase system. In the first phase, she will (slightly) slow down the transmission of packets from Jack
to Steve so as to build up a backlog of packets in her buffer. Then, during the codeword transmission
phase, she will release packets from her buffer with the inter-packet delays prescribed by the codeword
corresponding to the message, while continuing to buffer arriving packets from Jack. Two questions
dictate the resulting throughput of the scheme: (1) Phase I: how much can Alice slow down the packet
stream from Jack to Steve without it being detected by warden Willie?; (2) Phase II: how many packets
must Alice accumulate in her buffer by the start of Phase II so as to, with high probability, have a packet
in her buffer at all of the times required by the codeword? To answer the first question, we calculate
the number of packets that Alice can covertly collect by slowing down the packet stream from Jack
to Steve in a duration of time T and present the results in Lemma 1. Then, in Theorem 2, we answer
the second question to present an achievability result for the number of bits that Alice can reliably and
covertly transmit to Bob through packet timings.
√
Lemma 1. If Alice can buffer packets on the link from Jack to Steve, she can covertly buffer O( λT )
√
packets in a period of time of length T . Conversely, if Alice buffers ω
λT packets in the period of
time T , there exists a detector that Willie can use to detect such buffering with arbitrarily low sum of
error probabilities PF A + PM D .
Proof. For the achievability result, Alice buffers packets by slowing down the incoming process of rate
λ to one of rate λ−∆ (i.e. slightly scaling up the packet timings) and buffering the packets of the (faster:
of rate λ) incoming process that have yet to exit the buffer by the (slower: of rate λ − ∆) outgoing
process. P0 is the probability distribution function for Willie’s observation under the null hypothesis H0
9
and corresponds to a Poisson point process with rate λ, and P1 is the probability distribution function
for observations under the hypothesis H1 and corresponds to a Poisson point process with rate λ − ∆.
Similar to (6), by Fact 1:
∆
D(P1 ||P0 ) = ∆ · T − (λ − ∆)T ln 1 +
λ−∆
2 !
(a)
1
∆
∆
−
≤ ∆ · T − (λ − ∆)T
λ−∆ 2 λ−∆
∆2
T ∆2
=∆·T −T ∆−
=
(16)
2(λ − ∆)
2(λ − ∆)
p
where (a) follows from inequality (7). If ∆ = 2λ/ , then D(P1 ||P0 ) ≤ 22 as T → ∞. Therefore,
√
by (4) and (16), Alice can achieve PF A + PM D ≥ 1 − as for any 0 < < 1 long as ∆ · T = 2λT =
√
λT . The proof of the converse follows that of Theorem 1 with minor modifications.
O
Theorem 2. Consider Scenario 2. By embedding information in the inter-packet delays, Alice can can
covertly and reliably transmit O (λT ) bits to Bob in a time interval of length T .
Proof. Construction: To establish covert communication over the timing channel, Alice and Bob share
a secret key (codebook) to which Willie does not have access. To build a codebook, a set of M
l=M
independently generated codewords {C(Wl )}l=M
l=1 are generated for messages {Wl }l=1 according to
realizations of a Poisson process with parameter λ that mimics the overt traffic on the channel between
Jack and Steve, where M is defined later. In particular, to generate a codeword C(Wl ), first a random
variable N is generated according to a Poisson distribution with mean λT . Then, N inter-packet delays
are generated by placing N points uniformly and independently on an interval of length T [25] (see
Fig. 3). For each message transmission, Alice uses a new codebook to encode the message into a
codeword. According to the codebook, each message corresponds to a codeword that is a series of
inter-packet delays. To send a codeword, Alice applies the inter-packet delays to the packets that are
being transmitted from Jack to Steve.
Per above, Alice’s communication includes two phases: a buffering phase and a transmission phase.
During the buffering phase of length ψT , where 0 < ψ < 1 is a parameter to be defined later, Alice
slows down the packet transmission process to λ − ∆ in order to build up packets in her buffer. In
particular, Alice’s purpose in the first phase is to buffer enough packets to ensure she will not run out
of packets during the transmission phase, of length T 0 = (1 − ψ)T , with high probability (see Fig. 4).
Analysis: (Covertness) Suppose that Willie knows when each of the two phases start and end.
According to Lemma 1, during the first phase, Alice can lower bound the sum of Willie’s error probability
10
Fig. 3. Codebook generation: Alice and Bob share a codebook (secret), which specifies the sequence of inter-packet delays corresponding
to each message. To generate each codeword, a number N is generated according to the Poisson distribution with parameter λT , and
then N points are placed uniformly and randomly on the time interval [0, T ].
Fig. 4. Two-phased construction: Alice divides the duration of time T into two phases with lengths ψT and (1 − ψ)T . In the first phase,
Alice slows down the transmission to λ − ∆ and buffers the excess packets. In the next phase, she transmits packets to Bob at rate λ
according to the inter-packet delays in the codeword corresponding to the message to be transmitted.
by 1 − and collect m = ∆ · ψT packets. Thus, in this phase, Alice’s buffering is covert. By (9),
p
m = 2λψT
(17)
During the second phase, the packet timings corresponding to the selected codeword are an instantiation
of a Poisson point process with rate λ and hence the traffic pattern is indistinguishable from the pattern
that Willie expects on the link from Jack to Steve. Hence, the scheme is covert.
(Reliability) Next we show that Alice will have a reliable communication to Bob. The notion of
reliability is tied to two events. First, Bob should be able to decode the message with arbitrarily low
probability of error, which follows by adapting the coding scheme of [25]. Second, Alice needs to
avoid a “failure” event, in which Alice is unable to create the packet timings for the selected codeword
because she has run out of packets in her buffer at some point in the codeword transmission process.
Therefore, in the transmission phase, we need to demonstrate that Alice can achieve Pf < ζ for any
ζ > 0, where Pf is probability of the event “failure”.
Suppose that Alice has accumulated m packets and wishes to communicate with Bob during an
interval of length T 0 . There are two processes which impact the probability of running out of packets
in Alice’s buffer: the transmission of packets to Bob and the reception of packets from Jack (that are
11
intended for Steve). Each of these processes is a Poisson point process with parameter λ. Consider
the union of these two processes as a single process, termed “T-R Process”. The T-R Process is a
Poisson point process with parameter 2λ. Note that the probability of an event in this aggregate process
corresponding to Alice buffering a packet is the same as the probability that it corresponds to her
sending a packet to convey the next code symbol. Therefore, we toss a fair coin for each packet; if it
is heads, Alice buffers a received packet; otherwise, she releases a packet from her buffer.
Let K be the total number of received and transmitted packets during interval T 0 .
Pf = P (F ∩ {K > 4λT 0 }) + P (F ∩ {K ≤ 4λT 0 })
≤ P (K ≥ 4λT 0 ) + P (F ∩ {K ≤ 4λT 0 })
(18)
(19)
where F is the “failure” event, and (18) is due to the law of total probability. Consider P (K > 4λT 0 ).
Using the Chernoff bound for the T-R Poisson point process yields:
0
0
e 2λT 0
e−2λT (2λT 0 e)4λT
0
P (K ≥ 4λT ) ≤
=
0
4
(4λT 0 )4λT
(20)
Thus, for any 0 < ζ < 1, if λT 0 is large enough,
P (K ≥ 4λT 0 ) <
ζ
2
(21)
Now, consider P (F ∩ {K ≤ 4λT 0 })
P (F ∩ {K ≤ 4λT 0 }) = P (F |K ≤ 4λT 0 ) P (K ≤ 4λT 0 )
(22)
≤ P (F |K ≤ 4λT 0 )
(23)
≤ P (F|K = k 0 ) |k0 =4λT 0
(24)
where P (F|K = k 0 ) is the probability of “failure” given the total number of received and transmitted
packets in T 0 is k 0 , and (24) follows from (23) because P (F|K = k 0 ) is a monotonically increasing
function of k 0 . This is true since with m initial packets, if the total number of received and transmitted
packets increases, the probability of “failure” increases.
To calculate P (F|K = k 0 ), we model the T-R Process by a one-dimensional random walk. Suppose
that there is a random walker on the z axis that is initially located at the origin z = 0, and walks
at each step from z to either z + 1 or z − 1 with equal probability. If Alice transmits a packet, the
walker’s location is increased by 1, and if Alice receives a packet, the walker’s location is decrease by
1. Given Alice has m packets before starting this process, the event of “failure” in the T-R Process is
the same as the event that, at some point, the random walker (that has started from origin z = 0) hits
12
state z = m + 1. Clearly, this can only happen if the walker steps from state z = m to state z = m + 1
at some point. Since the number of times that Alice either receives or transmits a packets in a duration
of time T 0 is the same as the number of steps of the walker K:
P (F|K = k 0 ) = 1 − Pm (k 0 )
(25)
where Pm (k 0 ) is the probability that the walker’s location remains x ≤ m during the first k 0 steps. From
(Eq. (12.13) in [38]):
0
lim Pm (k ) = erf
k0 →∞
m
√
2k 0
(26)
where erf (·) is the error function. By (17),
lim Pm (k 0 )|k0 =4λT 0 = erf
T →∞
m
√
8λT 0
√
2λψT
= erf
= erf
!
(27)
p
8λ(1 − ψ)T
s
!
ψ
2 1−ψ
(28)
By (24), (25), (28)
0
0
0
lim P (F ∩ {K ≤ 4λT }) ≤ lim P (F|K = 4λT ) = 1 − lim Pm (k )
T →∞
T →∞
T →∞
1 − erf
k0 =4λT 0
2
s
ψ
1−ψ
!
.
Therefore, if
2
2 −1
ζ
erf
1−
2
(29)
ζ
lim P (F |{K ≤ 4λT 0 }) ≤ .
T →∞
2
(30)
ψ
=
1−ψ
where erf −1 (·) is the inverse error function,
Therefore, by (21), (30) Alice can achieve Pf < ζ for any 0 < ζ < 1 and thus the transmission is
reliable.
(Number of Covert Bits) Alice’s rate of packet transmission in the second phase is λ and the capacity
of the queue for conveying information through inter-packet delays is [25, Theorems 4 and 6].
C(λ) = λ log
µ
nats/sec
λ
(31)
Since the length of the second phase is T 0 = T (1 − ψ), the amount of information that Alice sends
through inter-packet delays is C(λ)T (1 − ψ) = O (λT ).
(Size of the Codebook) Finally, to complete the construction, we find the number M of codewords
in the codebook. According to [25, Definition 1], the rate of the codebook is
13
log M
T0
=
log M
.
T (1−ψ)
Since
the capacity of the queue at output rate λ is the maximum achievable rate at output rate λ (see [25,
Definition 2]), by (31), the size of the codebook is
M = e(1−ψ)λT log (µ/λ)
where 1 − ψ =
2
(32)
−1
2
erf −1 (1 − ζ) + 1
.
V. C ONCLUSION
In this paper, we have considered two scenarios of covert communication on a timing channel. In
the first one, Alice is able to insert her own packets onto the channel but not modify the timing of
other packets. We established that, if the packet transmission between legitimate nodes is modeled as
√
λT packets to Bob in a period of
a Poisson point process with parameter λ, Alice can transmit O
duration T without being detected by warden Willie. In the converse, we showed that if Alice inserts
√
more than O
λT packets, she will be detected by Willie. Next, we analyzed the scenario where
Alice cannot insert packets but instead is able to buffer packets and release them onto the channel at
a later time. We showed that if Alice waits ψT before codeword transmission, where ψ = Θ (1), and
buffers packets, she is able to transmit O (λT ) bits to Bob reliably and covertly in time period T , given
Alice and Bob share a secret of sufficient length (codebook).
R EFERENCES
[1] R. Soltani, D. Goeckel, D. Towsley, and A. Houmansadr, “Covert communications on renewal packet channels.” https://arxiv.org/
abs/1610.00368, 2016.
[2] J. Talbot and D. Welsh, Complexity and Cryptography: An Introduction. Cambridge University Press, 2006.
[3] M. Bloch and J. Barros, Physical-Layer Security. Cambridge, UK: Cambridge University Press, 2011.
[4] A. D. Ker, “Batch steganography and pooled steganalysis,” vol. 4437 of Lecture Notes in Computer Science, pp. 265–281, Springer
Berlin Heidelberg, 2007.
[5] B. A. Bash, D. Goeckel, and D. Towsley, “Square root law for communication with low probability of detection on AWGN channels,”
in Proc. IEEE Int. Symp. Inform. Theory (ISIT), (Cambridge, MA, USA), pp. 448–452, July 2012.
[6] B. Bash, D. Goeckel, and D. Towsley, “Limits of reliable communication with low probability of detection on AWGN channels,”
Selected Areas in Communications, IEEE Journal on, vol. 31, pp. 1921–1930, September 2013.
[7] B. A. Bash, D. Goeckel, D. Towsley, and S. Guha, “Hiding information in noise: Fundamental limits of covert wireless
communication,” IEEE Communications Magazine: Special Issue on Wireless Physical Layer Security, Dec. 2015. to appear.
[8] D. Goeckel, B. Bash, S. Guha, and D. Towsley, “Covert communications when the warden does not know the background noise
power,” IEEE Communication Letters, Aug 2015. submitted.
[9] B. A. Bash, S. Guha, D. Goeckel, and D. Towsley, “Quantum Noise Limited Communication with Low Probability of Detection,”
in Proc. IEEE Int. Symp. Inform. Theory (ISIT), (Istanbul, Turkey), pp. 1715–1719, July 2013.
[10] B. Bash, D. Goeckel, and D. Towsley, “LPD communication when the warden does not know when,” in Information Theory (ISIT),
2014 IEEE International Symposium on, pp. 606–610, 2014.
14
[11] R. Soltani, B. A. Bash, D. Goeckel, S. Guha, and D. Towsley, “Covert single-hop communication in a wireless network with distributed
artificial noise generation,” in Proc. Conf. on Commun., Control, Comp. (Allerton), (Monticello, IL, USA), pp. 1078–1085, 2014.
[12] T. Sobers, D. Goeckel, B. Bash, S. Guha, and D. Towsley, “Covert communication with the help of an uninformed jammer achieves
positive rate,” in Proc. Asilomar Conf. on Signals, Systems, and Computers, (Monterey, CA, USA), 2015.
[13] P. H. Che, M. Bakshi, and S. Jaggi, “Reliable deniable communication: Hiding messages in noise,” in Proc. IEEE Int. Symp. Inform.
Theory (ISIT), (Istanbul, Turkey), pp. 2945–2949, July 2013.
[14] S. Kadhe, S. Jaggi, M. Bakshi, and A. Sprintson, “Reliable, deniable, and hidable communication over multipath networks,” in Proc.
IEEE Int. Symp. Inform. Theory (ISIT), (Honolulu, HI, USA), pp. 611–615, July 2014.
[15] J. Hou and G. Kramer, “Effective secrecy: Reliability, confusion and stealth,” in Proc. IEEE Int. Symp. Inform. Theory (ISIT),
(Honolulu, HI, USA), pp. 601–605, July 2014.
[16] M. Bloch, “Covert communication over noisy memoryless channels: A resolvability perspective,” in Proc. IEEE Int. Symp. Inform.
Theory (ISIT), (Hong Kong, China), 2015.
[17] L. Wang, G. W. Wornell, and L. Zhang, “Limits of low-probability-of-detection communication over a discrete memoryless channel,”
in Proc. IEEE Int. Symp. Inform. Theory (ISIT), (Hong Kong, China), 2015.
[18] S. Lee and R. Baxley, “Achieving positive rate with undetectable communication over AWGN and Rayleigh channels,” in Proc.
IEEE Int. Conf. Commun. (ICC), pp. 780–785, June 2014.
[19] P. H. Che, M. Bakshi, C. Chan, and S. Jaggi, “Reliable deniable communication with channel uncertainty,” in Proc. Information
Theory Workshop (ITW), 2014.
[20] C. Girling, “Covert channels in lan’s,” Software Engineering, IEEE Transactions on, vol. SE-13, pp. 292–296, Feb 1987.
[21] V. Berk, A. Giani, and G. Cybenko, “Detection of covert channel encoding in network packet delays (tech. rep. tr2005-536),”
Department of Computer Science, Dartmouth College (November 2005), 2005.
[22] G. Shah, A. Molina, M. Blaze, et al., “Keyboards and covert channels.,” in USENIX Security, 2006.
[23] Y. Liu, D. Ghosal, F. Armknecht, A.-R. Sadeghi, S. Schulz, and S. Katzenbeisser, “Hide and seek in time–robust covert timing
channels,” in Computer Security–ESORICS 2009, pp. 120–135, Springer, 2009.
[24] A. Houmansadr and N. Borisov, “CoCo: coding-based covert timing channels for network flows,” in Information Hiding, pp. 314–328,
Springer, 2011.
[25] V. Anantharam and S. Verdu, “Bits through queues,” Information Theory, IEEE Transactions on, vol. 42, no. 1, pp. 4–18, 1996.
[26] B. P. Dunn, M. Bloch, and J. N. Laneman, “Secure bits through queues,” in Networking and Information Theory, 2009. ITW 2009.
IEEE Information Theory Workshop on, pp. 37–41, IEEE, 2009.
[27] J. K. Millen, “Finite-state noiseless covert channels,” in Computer Security Foundations Workshop II, 1989., Proceedings of the,
pp. 81–86, IEEE, 1989.
[28] A. S. Bedekar and M. Azizoglu, “The information-theoretic capacity of discrete-time queues,” Information Theory, IEEE Transactions
on, vol. 44, no. 2, pp. 446–461, 1998.
[29] S. H. Sellke, C.-C. Wang, N. Shroff, and S. Bagchi, “Capacity bounds on timing channels with bounded service times,” in Information
Theory, 2007. ISIT 2007. IEEE International Symposium on, pp. 981–985, IEEE, 2007.
[30] S. Sellke, C.-C. Wang, S. Bagchi, and N. Shroff, “Tcp/ip timing channels: Theory to implementation,” in INFOCOM 2009, IEEE,
pp. 2204–2212, April 2009.
[31] I. Moskowitz and A. Miller, “The channel capacity of a certain noisy timing channel,” Information Theory, IEEE Transactions on,
vol. 38, pp. 1339–1344, Jul 1992.
[32] I. Moskowitz and A. Miller, “Simple timing channels,” in Research in Security and Privacy, 1994. Proceedings., 1994 IEEE Computer
Society Symposium on, pp. 56–64, May 1994.
[33] N. Kiyavash, T. Coleman, and M. Rodrigues, “Novel shaping and complexity-reduction techniques for approaching capacity over
queuing timing channels,” in Communications, 2009. ICC ’09. IEEE International Conference on, pp. 1–5, June 2009.
15
[34] N. Kiyavash and T. Coleman, “Covert timing channels codes for communication over interactive traffic,” in Acoustics, Speech and
Signal Processing, 2009. ICASSP 2009. IEEE International Conference on, pp. 1485–1488, April 2009.
[35] R. Archibald and D. Ghosal, “A covert timing channel based on fountain codes,” in Trust, Security and Privacy in Computing and
Communications (TrustCom), 2012 IEEE 11th International Conference on, pp. 970–977, June 2012.
[36] S. Gianvecchio and H. Wang, “Detecting covert timing channels: an entropy-based approach,” in Proceedings of the 14th ACM
conference on Computer and communications security, pp. 307–316, ACM, 2007.
[37] S. Gianvecchio, H. Wang, D. Wijesekera, and S. Jajodia, “Model-based covert timing channels: Automated modeling and evasion,”
in Recent Advances in Intrusion Detection, pp. 211–230, Springer, 2008.
[38] A. Gut, Stopped random walks. Springer, 1998.
A PPENDIX
a) Proof of fact 1:
D(Pλ1 (n)||Pλ2 (n)) =
∞
X
Pλ1 (n) log(Pλ1 (n))/Pλ2 (n)) =
n=0
X
Pλ1 (n) log
n
λn1 e−λ1 /n!
λn2 e−λ2 /n!
λn1
=
Pλ1 (n) log n + λ2 − λ1
λ2
n
X
λ1
Pλ1 (n)n log
= (λ2 − λ1 ) +
λ2
n
X
λ1
E1 [N ]
λ2
λ1
= (λ2 − λ1 ) + λ1 log
λ2
= (λ2 − λ1 ) + log
where E1 [K] is the expected value of the Poisson distribution with parameter λ1 .
b) Proof of (7): Consider x ≥ 0, f (x) = ln(1 + x), and g(x) = x −
f 0 (x) − g 0 (x) =
1
x2
− (1 − x) =
≥0
1+x
1+x
x2
.
2
Therefore
(33)
On the other hand f (0) = g(0) = 0, therefore
Zx
f (x) − g(x) =
(f 0 (x) − g 0 (x)) dx ≥ 0
0
Thus, ln(1 + x) ≥ x −
x2
2
for x ≥ 0.
16
(34)
| 7cs.IT
|
Hierarchical Predictive Control Algorithms for
Optimal Design and Operation of Microgrids
arXiv:1803.06705v1 [math.OC] 18 Mar 2018
Sai Krishna Kanth Hari† , Kaarthik Sundar∗ , Harsha Nagarajan‡ , Russell Bent∗ , Scott Backhaus∗
† Department of Mechanical Engineering, Texas A&M University, College Station, USA.
‡ Theoretical Division (T-5), Los Alamos National Laboratory, NM, USA. Contact: [email protected]
∗ Center for Nonlinear Studies, Los Alamos National Laboratory, NM, USA.
Abstract—In recent years, microgrids, i.e., disconnected
distribution systems, have received increasing interest from
power system utilities to support the economic and resiliency
posture of their systems. The economics of long distance
transmission lines prevent many remote communities from
connecting to bulk transmission systems and these communities rely on off-grid microgrid technology. Furthermore,
communities that are connected to the bulk transmission system are investigating microgrid technologies that will support
their ability to disconnect and operate independently during
extreme events. In each of these cases, it is important to
develop methodologies that support the capability to design
and operate microgrids in the absence of transmission over
long periods of time. Unfortunately, such planning problems
tend to be computationally difficult to solve and those that are
straightforward to solve often lack the modeling fidelity that
inspires confidence in the results. To address these issues, we
first develop a high fidelity model for design and operations
of a microgrid that include component efficiencies, component
operating limits, battery modeling, unit commitment, capacity
expansion, and power flow physics; the resulting model is
a mixed-integer quadratically-constrained quadratic program
(MIQCQP). We then develop an iterative algorithm, referred
to as the Model Predictive Control (MPC) algorithm, that
allows us to solve the resulting MIQCQP. We show, through
extensive computational experiments, that the MPC-based
method can scale to problems that have a very long planning
horizon and provide high quality solutions that lie within 5%
of optimal.
Index Terms—microgrid, model predictive control, long-time
horizon, capacity planning, battery efficiency modeling
I. I NTRODUCTION
In recent years, there has been increasing concerns about
the susceptibility of modern electric power systems to
extreme events that cause large-scale black outs. Even in
2017, hurricanes such as Harvey, Irma, and Maria have
demonstrated the ability of such events to leave populations
in the United States and the Caribbean without power for
days or even weeks. As a result, the power engineering
community has a vested interest in exploring solutions that
can mitigate the impacts of these events. One proposed
solution is microgrids, i.e. a decentralized subsystem that
is capable of operating in either a grid-connected mode or
an islanded mode to satisfy the power demand of a local
community during large-scale black outs.
One of the biggest obstacles for rapid deployment of
microgrids is their high initial outlay. This makes optimizing
the design and operation of microgrids an economically
very important problem. However, this problem is generally
very difficult to solve because of the underlying nonlinear,
nonconvex physics and discrete decision variables. The
challenges in solving this problem are further exacerbated
by desires to ensure that islanded microgrids can operate
for long periods of time without a connection to bulk
transmission systems. In spite of these difficulties, recent
papers [1]–[5] have considered and modeled various important details of this problem, i.e. communication-constrained
expansion planning, component efficiency, linearized AC
physics, and N-1 security. However, the problem of planning and operating a microgrid over a finite time horizon,
i.e. months, or even years, [6], remains intractable. Even
planning for peak or average load conditions (i.e. typicalday) is computationally very difficult.
In this paper, we focus on the computational tractability
of long-time horizon planning and design of microgrids and
push the boundaries of scalability beyond what is possible
for existing approaches. In this context, we remark that for
the purposes of this article, we have ignored uncontrollable
sources like wind and solar in order to isolate and examine
the difficulty in developing efficient methods. Traditionally,
long-time horizon problems in the control literature are
handled by decomposing the time-horizon into smaller timestages, i.e. receding-horizon (RH) methods [7], [8]. In RH,
each sub-problem is sequentially solved to optimality and
the solutions from previous stages serve as initial conditions
for subsequent stages. Generally speaking, this approach
is computationally attractive and such methods have been
applied to microgrid design [9]–[11]. However, optimality
of individual stages does not guarantee global optimality,
even when the problem is convex [12]. Furthermore, the
RH approach is short-sighted in the sense that it does not
take into account for any fluctuations in the load (or the
slow dynamics) in the subsequent stages. To address this
issue, recent work has developed a first-order method that
uses dual information from a coarse long-horizon problem
[13] to communicate information between sub problems.
Under certain assumptions, such as convexity, this method
is guaranteed to converge to global optimality and has been
shown to be effective on long-time horizon, continuous
problems.
In contrast, high fidelity models of microgrid design
that include discrete installation of technologies, AC power
flow physics, engineering limits, and battery storage device
efficiencies is a long-time horizon mixed-integer, nonlinear
problem (MINLP) [1]. To address these key modeling details, this paper generalizes the existing first-order methods
[13] to a mixed-integer setting. In short, the key contributions of this paper are (i) formulation of the Microgrid
Design and Operation Problem (MDOP) as a mixed-integer
quadratically constrained quadratic program (MIQCQP), (ii)
a convex relaxation of battery-efficiency modeling without
the use of disjunctive binary variables, (iii) an iterative firstorder-based MPC algorithm to solve to the MDOP for longtime horizon; the algorithm is very general and can be applied to any MIQCQP that permits a decomposition in space
or time, (iv) an extensive computational experimentation of
the algorithm on an IEEE 13-node test feeder to corroborate
its effectiveness in finding high quality solutions (typically,
within 5% of the optimal) for the MDOP.
The rest of the paper is organized as follows: the next
section details the nomenclature and presents the MIQCQP
model for the MDOP. Sec. III presents the MPC algorithm
by starting with a QCQP and then extending it to an
MIQCQP. Finally, Sec. IV presents extensive computational
results followed by conclusions and future work in Sec. V.
II. M ATHEMATICAL FORMULATION
A. Nomenclature
Throughout the rest of the article, boldface symbols are
used to denote constants.
Sets:
N - set of nodes (buses), indexed by i
E - set of from edges (lines), indexed by (i, j)
Er - set of to edges (lines)
T - set of time periods, indexed by t
B - set of batteries, indexed by b
D - set of diesel generators, indexed by d
D(i) - set of diesel generators at bus i
B(i) - set of batteries at bus i
R B ∪ D - set or resources, indexed by r
R(i) - set of resources at bus i
Constants:
fb , fd - fixed cost for b ∈ B, d ∈ D, $
gb - variable capacity cost for b ∈ B, $/MVA
c0d , c1d , c2d - operational cost coefficients for d ∈ D
U Td , DTd - minimum up-time and down-time for d ∈ D
RUd , RDd - ramp-up and ramp-down limit for d ∈ D
si j - thermal limit for line (i, j), MVA
lpti + jlqit - AC power demand at bus i ∈ N, MVA
ri j + jxi j - impedance of line (i, j) ∈ E
gl
gu
pr , pr - real power generation limits for r ∈ R, MW
gl
gu
qr , qr - reactive power generation limits for r ∈ R,
MVAr
vil , viu - bounds for voltage magnitude square at bus i, kV
mb - maximum installable capacity for b ∈ B, MVA
sc b - maximum energy storage capacity of b ∈ B, MW-hr
τ b - energy storage capacity of b ∈ B, MVA
b
b
- charging, discharging efficiency for battery
, ηdis
ηch
b∈B
η d - efficiency of diesel generator d ∈ D
µ - load shedding penalty, $/MW
hb (i) - number of batteries that can be built at bus i
hd (i) - number of generators that can be built at bus i
∆t - duration of time step, hrs
Build decision variables:
z b - binary build variable for each battery b ∈ B
z d - binary build variable for each generator d ∈ D
Operation variables for diesel generators:
x dt - active/inactive status of d ∈ D for t ∈ T
y dt - start-up status of d ∈ D for t ∈ T
w dt - shut-down status of d ∈ D for t ∈ T
gt
gt
p d + jq d - apparent power generated by d ∈ D for t ∈ T
gt
b
p d - active power generated by d ∈ D for t ∈ T before
losses
Operation variables for batteries:
gt
gt
p b + jq b - apparent power entering/leaving b ∈ B for
t∈T
gt
b
p b - active power stored in b ∈ B for t ∈ T
sc bt - state-of-charge (energy stored) in b ∈ B at t ∈ T,
MW-hr
s b - maximum apparent power generation for b ∈ B, MVA
Other variables:
v it - squared voltage on node i ∈ N at t ∈ T
p et , q et - active/reactive power flow on line e ∈ E at t ∈ T
pt
qt
` i , ` i - active/reactive load shed on node i ∈ N at t ∈ T
B. Optimization problem
With the notations presented in Sec. II-A, we next formulate a mixed-integer quadratically constrained quadratic
programming formulation for the MDOP. The objective
function of the MDOP, Eq. (1), minimizes the build cost,
the total load shed, and the operating cost of the resources
(batteries and diesel generators). The operating cost of
diesel generators includes a no-load cost and a quadratic
generation cost.
min
Õ
(fb z b + gb s b ) +
b∈B
+
Õ
fd z d +
c0d · x dt + c1d ·
µ
pt
`i
+
qt
`i
t∈T i∈N
d∈D
ÕÕ
ÕÕ
gt
b
pd
+ c2d ·
gt
b
pd
2
t∈T d∈D
(1)
The MDOP is subject to the following constraints:
Power flow physics
The physics of power flow, thermal limits, and voltage
bounds are described by the following set of constraints:
Õ
gt
pt
p r + ` i − lpti
Õ
p itj
r∈R(i)
(i, j)∈E∪Er
Õ
Õ
∀i ∈ N, t ∈ T,
gt
qt
q itj
∀i ∈ N, t ∈ T,
(2b)
v it − 2 · (ri j p itj +
2 2
p itj + q itj 6 s2i j
v tj
vil 6 v it 6 viu
xi j q itj )
∀(i, j) ∈ E, t ∈ T,
∀(i, j) ∈ E, t ∈ T,
∀i ∈ N, t ∈ T.
(2c)
(2d)
(2e)
Eqs. (2a)–(2b) enforce Kirchoff’s current law and Eq. (2c)
models the “LinDistFlow” approximation of the nonconvex
DistFlow AC power flow equations [14]. Eqs. (2d) and (2e)
enforce the thermal limit for each transmission line and
bounds on the voltage magnitude at each bus, respectively.
Resource limits
The following constraints enforce the capacity limits on the
build and generation decisions for all the resources in R.
Õ
z b 6 hb (i) ∀i ∈ N,
z d 6 hd (i)
∀i ∈ N,
(3b)
d∈D(i)
s b 6 z b mb
∀b ∈ B.
The constraints on diesel generator operation are then:
x dt 6 z d
∀d ∈ D, t ∈ T,
(4a)
y dt − w dt x dt − x dt−1 ∀d ∈ D, t ∈ T,
y dt + w dt 6 1 ∀d ∈ D, t ∈ T,
gu
gl
gt
∀d ∈ D, t ∈ T,
p d 6 x dt pd
x dt pd 6 b
gt
t gu
t gl
∀d ∈ D, t ∈ T,
x d qd 6 q d 6 x d qd
gt
d gt
p d ∀d ∈ D, t ∈ T,
pd η b
g(t−1)
gt
RDd > p d
− p d ∀d ∈ D, t ∈ T,
gt
g(t−1)
RUd > p d − p d
∀d ∈ D, t ∈ T,
Õ
k
t
y d 6 x d ∀d ∈ D, t ∈ T,
k∈A(d,t)
w dk 6 1 − x dt
∀d ∈ D, t ∈ T.
(4b)
(4c)
(4d)
(4e)
(4f)
(4g)
(4h)
(4i)
(4j)
Eq. (4a) forces a diesel generator to be committed only if
it is built. Eq. (4b) determines if the generator is started
up or shut down at time period t based of its on-off status
between time periods t and t − 1. Eq. (4c) ensures that a
generator d is not started up and shut down in the same
time period, t. Eqs. (4d) and (4e) enforce the generation
limits of diesel generators. Real power limits are applied
gt
to b
p d , i.e. the generation produced before losses. This
loss is quantified by Eq. (4f). Eqs. (4g) and (4h) enforce
the ramping limits on consecutive time periods on every
generator d ∈ D. Finally, Eqs. (4i) and (4i) enforce the
minimum up time and minimum down time constraints for
every diesel generator. Usage of distinct binary variables,
(x dt , y dt , z dt ), to represent a tighter unit commitment polytope
is akin to the formulation described in [15].
Battery operation
The constraints on battery operations are given by:
gt
(3a)
b∈B(i)
Õ
B(d, t) { t̃ ∈ T : t − DTd + 1 6 t̃} ∀d ∈ D, t ∈ T.
k∈B(d,t)
(i, j)∈E∪Er
r∈R(i)
A(d, t) { t̃ ∈ T : t − U Td + 1 6 t̃} ∀d ∈ D, t ∈ T,
Õ
(2a)
q r + ` i − lqit
Operation of diesel generators
We first introduce sets A(d, t) and B(d, t) that support the
formulation of minimum up-time and down-time constraints
for each diesel generator d ∈ D at each time period t ∈ T.
(3c)
Eqs. (3a) and (3b) restrict the number of batteries and diesel
generators that can be installed at each bus, respectively. Eq.
(3c) limits the capacity of each battery.
gt
(p b )2 + (q b )2 6 (s b )2
∀b ∈ B, t ∈ T
(5a)
gt
sc bt sc bt−1 − b
pb
0 6 sc bt 6 z b sc b
∀b ∈ B, t ∈ T,
(5b)
· ∆t,
∀b ∈ B, t ∈ T.
(5c)
Eq. (5a) enforces the apparent power limits on charging and
discharging for each battery. Eq. (5b) determines the current
state-of-charge of the battery, sc bt , based on whether the
gt
gt
battery is being charged (b
p b < 0) or discharged (b
pb >
0). Finally, Eq. (5c) imposes bounds on the energy storage
gt
gt
gt
gt
discharging: p b > 0, b
p b > 0, | b
p b | > |p b |,
gt
gt
gt
(6a)
gt
charging: p b < 0, b
p b < 0, | b
p b | < |p b |.
(6b)
gt
For any battery b ∈ B and time period t ∈ T, b
p b is the
gt
power stored in the battery and p b is the amount of power
entering (charging) or leaving (discharging) b. Then, the
battery efficiency is modeled using this equation:
gt
gt
1
gt
gt
b
pb
p b ª ©p b ηb b
©p bgt ηgtdis b
ª
ch
p , b
® gt p gt < 0,®
b p b > 0, ® ∨ p b , b
®
b
gt
gt
gt
gt
b
|
p
|
>
|p
|
p b | < |p b | ¬
« b
« |b
b ¬
∀b ∈ B, t ∈ T.
(7)
Traditionally, Eq. (7) is reformulated into disjunctive, linear
constraints (i.e., [17]). In this article, we introduce the
following convex relaxation:
gt
gt
b
b
p b 6 ηdis
pb
1 gt
gt
p
pb 6 b b
ηch b
∀b ∈ B, t ∈ T,
∀b ∈ B, t ∈ T.
di
s
⌘b
=
10
0%
pgt
b
10
0%
=
1
b
⌘ch
ch
<latexit sha1_base64="CeZBMXh/zbiS9q+Qw/EI0+NH9UA=">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</latexit>
the variations in load over a full year can be reduced to
a few weeks, based on the typical-day approach [1], [2],
with a time discretization of 15 minutes. Using state-ofthe-art solvers such as Gurobi/CPLEX, the MDOP, even
for a time horizon of 1 day at 15 minute time intervals, is
challenging to solve to optimality because of the integer
variables. To address the computational difficulty of the
MDOP, this section describes an efficient heuristic that
adapts hierarchical model-predictive control architectures
[13] to discrete problems. We first present the MPC algorithm for a continuous relaxation of the MDOP, i.e. a QCQP,
and then extend it to MIQCQPs. For ease of exposition, we
recast the formulation of Sec. II-B as:
(F1 ) :
min
Õ
ϕ(x t , y t ) subject to:
(9a)
where τ ⊆ {0, . . . , |T|},
(9b)
t∈T
g(x τ , y τ ) 0
x0 x, y0 y,
(8b)
x t ∈ R , y t ∈ {0, 1}
III. A LGORITHMS
The formulation presented for the MDOP is an MIQCQP.
A typical planning horizon for the MDOP ranges from days
to a few years. For the purposes of this paper, we assume
convex
relaxation
pbgt
b
Figure 1. An illustrative example of convex relaxation (shaded) of
piecewise linear efficiency curve (blue) for batteries.
(8a)
The convex relaxation of Eq. (7) is given by Eqs. (8a),
(8b), (5a), and (5c). This convex relaxation of the efficiency
curves is also illustrated in Fig. 1. Though we assume the
piecewise efficiency curve with only two pieces, generalization of the proposed convex relaxation approach is straightforward for the case when there are multiple charging and
discharging efficiencies (see [1]). We remark that although
Eqs. (8a) and (8b) yield a relaxation, the relaxed solutions
are often empirically tight (see the computational results).
b
⌘dis
(0, 0)
⌘b
of the battery. The model of charging and discharging for
batteries is adapted from [16].
Battery efficiency modeling
We next discuss the constraints that model battery’s charging and discharging efficiency without the use of disjunctive
binary variables, typically used in the literature []. A preliminary version of the battery model, not utilizing disjunctive
binary variables, without a formal explanation has been
developed in [1]. In this article, we formalize the model as a
convex relaxation. For ease of exposition, we present battery
efficiency modeling with one charging and one dischargb
b
, respectively. Without
and ηdis
ing efficiency values, ηch
loss of generality, this model is extendable to piece-wise
linear charging and discharging curves with monotonically
decreasing charging and discharging efficiency values (see
[1]). We first define the charging and discharging regimes
for battery b ∈ B and time period t ∈ T as:
n
(9c)
m
∀t ∈ T.
(9d)
The variables x t and y t model the vectors of continuous and
binary variables in MDOP’s formulation. Eq. (9b) defines
the feasible region of the optimization problem, as specified
by the linear and quadratic constraints in Sec. II-B. Eq.
(9c) specifies the initial condition and Eq. (9d) specifies the
variable domains.
One of the main challenges with solving MDOP to
optimality is the large cardinality of T. To address this
challenge, we partition the time horizon {0, 1, .., |T|} into S
stages. Each stage consists of a subset of consecutive time
periods. For ease of exposition, we assume, without loss of
generality, that every stage has K time periods numbered,
K {1, .., K}. By convention, the zeroth time period for
every stage defines the initial conditions for that stage. We
let S {1, .., S}. Using this notation, F1 is reformulated as
F2 :
(F2 ) :
min
ÕÕ
ϕ(x s,k , y s,k ) subject to:
(10a)
s∈S k∈K
g(x s,τ , y s,τ ) 0
∀s ∈ S, τ K,
(10b)
x0,K x, y0,K y,
(10c)
x s,0 x s−1,K , y s,0 y s−1,K
x s,k ∈ R , y s,k ∈ {0, 1}
n
∀s ∈ S, k ∈ K,
∀s ∈ S, k ∈ K.
m
To simplify the presentation of the algorithm, without
loss of generality, we ignore the stage constraints and
focus on the coupling constraints. We first group the variables by stages by defining vectors of variables x®s
(x s,0 , x s,1 , . . . , x s,K ) and y®s (y s,0 , y s,1 , . . . , y s,K ). The
block form of F2 , without the stage constraints, is then
defined as:
(F3 ) :
min
s∈S
2 y®s
(λ s )
Πs
Qs
(11a)
∀s ∈ S,
(11b)
x®s−1
x®s
Πs
,
y®s−1
y®s
x®s ∈ Rn×K , y®s ∈ {0, 1} m×K
|
x®s
|
− Πs λ s + Πs+1 λ s+1 cs
y®s
∀s ∈ S \ {S},
(12a)
|
x®s
− Πs λ s cs
y®s
Πs
x®s
x®s−1
Πs
y®s
y®s−1
|
QS
ΠS
−ΠS
0
0
cS
−
0
ΠS
0
0
x®`+1
s−1
y® `+1
s−1
λ `+1
S
.
(13c)
One key observation of Eq. (13) is that each of the decentralized iteration equations are first-order KKT conditions
for the problems associated with each stage. The partitioned
problems are defined as:
|
1 x®1
x®1`+1
argmin
®1
y®1`+1
x®1 , y®1 2 y
x®1
x®1
−
y®1
y®1
x®`
x®1
Π1 0`
y®0
y®1
Π1
subject to:
|
Q1
c1 − Π2 λ 2`
|
(λ 1`+1 ).
(14a)
|
1 x®s
x®s`+1
argmin
`+1
®s
y®s
x®s , y®s 2 y
Qs
Πs
subject to:
|
x®s
x®s
−
y®s
y®s
x®`+1
x®s
Πs s−1
`+1
y®s−1
y®s
cs − Πs+1 λ `s+1
|
(λ `+1
s ).
(11c)
Qs
(13b)
x®`+1
S`+1
y®
S
λ `+1
S
(14b)
∀s ∈ S.
Qs
Πs
`
x®`+1
|
1`+1
x®0
y® c1 − 0 Π2 y® `
(13a)
1
0
Π1 0 0`
λ `+1
λ
1
2
`+1
x®`+1
®
| x
|
0 Πs+1 s−1
cs
−Πs s`+1
`+1
y®s
y®s−1
−
0
Πs
0 `+1
0 `+1
λ
s
λ s+1
∀s ∈ S \ {1, S},
|
−Π1
0
Eq. (11b) is equivalent to Eqs. (10c) and (10d) in matrix
form. The λ s are the dual variables of constraints (11b).
The matrices Πs and Πs are the coefficient matrices; they
correspond to the constraints (10d), rewritten for each stage
s ∈ S in a matrix form. The variables for s 0 are fixed
to the initial conditions of the MDOP. When the binary
constraints of F3 are relaxed, the first-order optimality
conditions (KKT) are given by
Qs
®s
x®s
| x
− cs
y®s
y®s
Q1
Π1
(10e)
A. MPC algorithm for the continuous relaxation of the
MDOP
(10d)
In formulation F2 , Eq. (10b) defines all the constraints of
stage s ∈ S. These constraints are referred to as “stage constraints”. Eqs. (10c) and (10d) specify the initial conditions
for each stage. Specifically, Eq. (10c) specifies the initial
conditions for the first stage and Eq. (10d) states that the
initial condition of stage s ∈ S is equal to the K th (final)
time period of the previous stage. The constraints in Eq.
(10d) are referred to as “coupling constraints”. Finally, Eq.
(10e) specifies the binary restrictions on the variables.
Õ 1 x® |
s
Any technique to solve the KKT system given in Eq. (12)
is iterative (for instance, Gauss Seidal). To avoid centrally
solving this KKT system using an iterative method, we use
a decentralized technique with an update index ` ∈ Z+ ; the
update equations used for each iteration is given below:
∀s ∈ {S},
(12b)
∀s ∈ S.
(12c)
1 x®S
x®S`+1
argmin
y®S`+1
y®S
2
x®S , y®S
|
QS
subject to:
ΠS
x®S
y®S
x®`+1
x®S
ΠS s−1
`+1
y®s−1
y®S
(λ S`+1 ).
(14c)
`
`
Given s ∈ S, we use P( x®s−1
, y®s−1
, λ `s+1 ) to define the
partitioned problem in Eq. (14) corresponding to stage s.
Since the decentralized KKT system in (13) are first-order
`
`
KKT conditions of P( x®s−1
, y®s−1
, λ `s+1 ), the KKT system
is solved by solving the smaller (partitioned) optimization
problems and communicating primal (x®s , y®s ) and dual (λ s )
variables between partitions. Convexity of the relaxation
of MDOP (convex QCQP) guarantees convergence of this
decentralized technique to the optimal solution (see [13]).
~x 0 , ~y0
" #|
min
~x 1 ,~y1
1 ~x 1
2 ~y1
~x1
~x0
Π1
~y1
~y0
" #
s.t.:
Π1
" #|
~x1
~x1
−
~y1
~y1
" #
Q1
c1 − Π2 λ 2`
|
" #
λ1`+1 ( ~y1`+1 )
`+1
`+1
~x1 , ~y1
" #|
min
1 ~x2
2 ~y2
s.t.:
Π2
~x 2 ,~y2
~x2
~x `+1
Π2 1`+1
~y1
~y2
" #
λ 2`+1
" #|
~x2
~x2
−
~y2
~y2
" #
Q2
"
c2 − Π3 λ 3`
|
`+1
`+1
~x2 , ~y2
#
λ 2`+1 ( ~y2`+1 )
λ 3`+1
Figure 2. Sketch of the primal and dual variables being communicated to neighboring stages at iteration (` + 1).
B. MPC algorithm for the MDOP
The heuristic for solving the MDOP, a MIQCQP, is
derived by making the following two changes to the technique presented in Sec. III-A: (i) the partitioned problem
corresponding to stage s ∈ S in Eq. (14) is solved as a
MIQCQP and the primal solution (x®s , y®s ) is communicated
forward to the next stage (see step 5 of Algorithm 1 and
Fig. 2) and (ii) the dual values of the coupling constraints
(λ s ) are computed by fixing the values of the binary
variables, y®s to the solution of the MIQCQP for stage s
(see step 6 of Algorithm 1 and Fig. 2). In other words,
the dual value for iteration ` and stage s, λ `+1
s , of the
linking constraint is computed as a function of the binary
variable solution values by solving the optimization problem
`
`
P( x®s−1
, y®s−1
, λ `s+1 )∪{ y®s y®s`+1 }. These values are denoted
`+1
`+1
by λ s ( y®s ). Unlike solving the continuous relaxation of
the MDOP, this algorithm has no convergence guarantees
to a global optimum. Hence, we stop the algorithm after a
finite number of iterations. A pseudo-code of the MPC is
shown in algorithm 1. The update scheme at iteration ` is
illustrated in the Fig. 2.
Algorithm 1 Psuedocode for the MPC algorithm
the slow-changing dynamics at low frequencies. In MDOP,
the real-time load fluctuations and slow raise to peak loads
over 24 hours horizon can be viewed as the high and low
frequency disturbances in the system, respectively. Thus, we
initialize the duals of the coupling constraints by solving
the continuous relaxation of MDOP to optimality. This
hierarchical MPC procedure (indirectly) captures and passes
the low frequency fluctuations as cost-to-go values via the
duals in the objective of each stage. Thus, MPC can be
viewed as an iterative generalization of the RH scheme.
IV. C OMPUTATIONAL R ESULTS
We first present the test-system specifications and parameter values used in the formulation and in the MPC
algorithm.
A. Data
For all the computational experiments, we use a standard
IEEE 13-node radial distribution test feeder network [18]
that was modified to support a positive-sequence model. A
schematic of the 13-node feeder network is shown in Fig.
3. The algorithm is allowed to build diesel generators or
batteries at the buses shown in the schematic (based on
[1]). To test the effectiveness of the MPC algorithm on
1: ` ← 1
2: Solve continuous relaxation of MDOP and initialize λ `s , for
all s ∈ S
3: while ` 6 N do
4:
for s 1, . . . , S do
` ,y
` , λ` )
®s−1
5:
x®s`+1 , y®s`+1 ← P(x®s−1
s+1
` ,y
` , λ` )
®s−1
6:
λ `+1
← getdual P(x®s−1
s
s+1
7:
end for
8:
` ← `+1
9: end while
7
6
Ð
{ y®s y®s`+1 }
diesel generator - build option
battery - build option
5
2
3
12
9
13
4
Initialization of dual variable values: Though there are
many ways of initializing the dual values, setting them to
zero values would be a trivial initialization. It is interesting
to note that when the initial dual values are set to zero,
the first iteration of MPC is equivalent to the standard
Receding Horizon (RH) algorithm [13]. However, one of
the known drawbacks of RH is it’s inability to capture the
future time horizon’s information in the current stage. To
address this, MPC methods are applied in a more general
setting of the hierarchical multigrid control, which captures
10
1
8
11
Figure 3. IEEE 13-node radial distribution network schematic.
different loading conditions, two load profiles based on data
from a New Mexico distribution utility (Kit Carson Electric
Cooperative) were generated. Both high and low frequency
fluctuations were utilized. The typical daily and weekly
load profiles are shown in the Fig. 4. A time discretization
of ∆t 15 minutes was chosen for all the runs of the
algorithm. This value concurs with the time discretization
of [1]. The load profiles are also assumed to be available at
every time discretization point.
run times of the MPC algorithm are detailed in Table I.
The average run time per iteration increases as the time
TABLE I
RUN TIMES
0.04
Case-1
Case-2
0.00
0
20
40
60
Time slots
80
0.15
Demand (MW)
Case-1
Case-2
0.10
0.05
0.00
200
400
Time slots
Avg. time per iteration
MPC (sec.)
3 days
7 days
14 days
67.50
249.17
700.30
44.98
84.15
136.97
3 days
7 days
14 days
233.79
313.03
1285.07
31.20
82.01
429.00
100
(a) Typical daily profile
0
Initialization
step (sec.)
horizon increases because the algorithm always decomposes
the problem into 6 stages irrespective of the time horizon.
The initialization step in Table I is the time taken to compute
the initial set of dual values for the first iteration of the MPC.
The dual values are based on the solution to the continuous
relaxation of the MDOP.
Table II presents the solution quality measured as a
relative gap from the objective value of a continuous,
perspective-based relaxation of the MDOP. The perspectivebased relaxation is not presented in this article due to
page limitations. For the sake of comparison with existing
approaches in the literature, we also present the results
obtained by using the RH algorithm. Whenever the RH
600
(b) Typical weekly profile
Figure 4. Typical load profiles on a particular bus.
For every battery, the charging and discharging efficiency
b
b
values, ηch
and ηdis
respectively, were set to 80% and 70%
(from previous work [17]). The battery’s fixed cost (fb ) and
variable cost (gb ) were $100 and 300$/MVA, respectively.
The efficiency of diesel generators was set to 50% (η d ).
The fixed cost and variable cost coefficients (fd , c0d , c1d , c2d )
for three different options of diesel generators were (200,
6, 35, 50), (300, 3, 10, 20), (350, 2, 5, 10).
Each run of the MPC algorithm was limited to 3 iterations, i.e., N 3. The value of the load-shedding penalty in
Eq. (1) was set to 107 $/MW. The number of stages for every
run of the algorithm was set to S 6. All algorithms were
implemented using Julia/JuMP [19] and Gurobi v7.0 was
used to solve the MIQCQP. All the runs were performed
on a Dell Precision T5500 workstation (Intel Xeon E5360
processor @2.53 GHz, 12 GB RAM).
B. Performance of the MPC algorithm
We first examine the run times of the MPC algorithm
for solving the MDOP with different time horizons. The
TABLE II
S OLUTION QUALITY. L OAD SHED (LS) VALUES ARE IN MW.
Load
profile
Time
horizon
# discretization
points
Case 1
0.02
Time
horizon
Case 1
0.06
Load
profile
3 days
7 days
14 days
288
672
1344
4.34
3.17
3.67
LS (0.35)
LS (2.77)
LS (0.53)
Case 2
Demand (MW)
0.08
Case 2
0.10
Relative gap (%)
MPC
RH
3 days
7 days
14 days
288
672
1344
4.64
1.30
0.97
5.04
1.35
0.98
algorithm sheds load, “LS” is reported in the relative gap
column of Table II. The actual real load shed is specified
in parentheses. It is clear from the Table II that the MPC
produces solutions that are within 5% of the lower bound
for all the cases. The RH is only able to produce comparable
solutions for case-2 load profiles. These have a faster change
in frequencies than case-1.
C. Battery usage
This section compares the battery usage of the MPC
and RH approaches. Fig. 5(b) shows the total state-ofcharge of all the batteries that were built in the system
D. Exactness of convex relaxation for battery efficiencies
We remark that despite using a relaxation for modeling
battery efficiencies in Sec. II to reduce the computational
burden, the solutions obtained using the MPC algorithm for
all the runs were always on the charging or discharging
efficiency curves in Fig. 1. This observation is primarily due
to the cost-minimization objective which ensures no further
losses during charging/discharging states of the battery.
The theoretical analysis and justification is an interesting
direction of future work.
7,000
5,000
0.16 MW
6,000
0.19 MW
Aggregated generation cost ($)
MPC (Total=$33265.3)
RH (Total=$33941.5)
8,000
4,000
1
2
3
4
Stages
5
6
(a) Total generation cost; the value inside the bars are the real
power shed.
50
Battery state-of-charge (MW-hr)
by the MPC and RH solutions for the case-1 load profile
(slow frequencies) with a time horizon of 3 days. It is
clearly evident from Fig. 5(b) that the MPC algorithm
builds and utilizes the batteries in a much more efficient
way than the RH algorithm. This is attributed to the fact
that the cost-to-go-like term in the objective of the stage
problems in Eq. (14) better approximates the slow dynamics
in the system for non-zero dual values. The non-zero values
provide the MPC algorithm with a look-ahead feature. The
RH algorithm lacks this look-ahead feature because it is a
one-pass algorithm that does not use dual values to update
the solution. The short-sightedness of the RH algorithm is
also evident from Fig. 5(a) which shows the total generation
cost for all the stages. Despite the RH having very low
generation cost in the initial few stages, the lack of a feature
in the RH algorithm to foresee the load profiles in the future
stages leads to a higher overall generation cost in stages 5
and 6 of Fig. 5(a). Despite real load shedding of 0.35MW,
the aggregated generation cost of RH algorithm is 2.04%
higher than the MPC’s solution.
MPC
RH
40
30
20
10
0
0
50
100 150 200
Time slots
250
300
(b) Aggregated state-of-charge of the batteries
Figure 5. State-of-charge of batteries based on solutions with MPC vs.
RH algorithms
R EFERENCES
V. C ONCLUSIONS
In this paper, we presented an MIQCQP formulation
for long time-horizon planning and operation of a microgrid. An elegant convex relaxation to model battery
efficiencies that was empirically exact was developed. We
also develop a fast, scalable, hierarchical predictive control
algorithm to compute feasible solutions for an MIQCQP.
Extensive computational experiments illustrate that the algorithm scales well with increases in the time horizon and
is able to compute feasible solutions that are within 5%
of a lower bound. Future research should consider N-1
security constraints, battery degradation models, inclusion
of fluctuating renewable energy sources like wind and solar,
and algorithmic enhancements such as hierarchical coarsegrid-based dual updates and sliding horizon schemes with
stage overlaps.
[1] S. C. Madathil, E. Yamangil, H. Nagarajan, A. Barnes, R. Bent,
S. Backhaus, S. J. Mason, S. Mashayekh, and M. Stadler, “Resilient
off-grid microgrids: Capacity planning and N-1 security,” IEEE
Trans. on Smart Grid, vol. PP, no. 99, pp. 1–1, 2017.
[2] S. Mashayekh, M. Stadler, G. Cardoso, M. Heleno, S. C. Madathil,
H. Nagarajan, R. Bent, M. Mueller-Stoffels, X. Lu, and J. Wang,
“Security-constrained design of isolated multi-energy microgrids,”
IEEE Trans. on Power Systems, vol. PP, no. 99, pp. 1–1, 2017.
[3] A. Barnes, H. Nagarajan, E. Yamangil, R. Bent, and S. Backhaus,
“Tools for improving resilience of electric distribution systems with
networked microgrids,” arXiv preprint arXiv:1705.08229, 2017.
[4] G. Cardoso, N. DeForest, L. Le Gall, C. Gehbauer, M. Hartner,
S. Mashayekh, C. Milan, T. Schittekatte, M. Stadler, D. Steen, and
J. Tjaeder, “DER-CAM user manual full DER web optimization service,” Lawrence Berkeley National Laboratory (LBNL), 1 Cyclotron
Rd, Berkeley, CA 94720, Tech. Rep., 2015.
[5] G. Byeon, P. Van Hentenryck, R. Bent, and H. Nagarajan,
“Communication-constrained expansion planning for resilient distribution systems,” arXiv preprint arXiv:1801.03520, 2018.
[6] D. T. Ton and M. A. Smith, “The US department of energy’s
microgrid initiative,” The Electricity Journal, vol. 25, no. 8, pp. 84–
94, 2012.
[7] E. F. Camacho and C. B. Alba, Model predictive control. Springer
Science & Business Media, 2013.
[8] R. Scattolini and P. Colaneri, “Hierarchical model predictive control,”
in Decision and Control, 2007 46th IEEE Conference on. IEEE,
2007, pp. 4803–4808.
[9] A. Bidram and A. Davoudi, “Hierarchical structure of microgrids
control system,” IEEE Trans. on Smart Grid, vol. 3, no. 4, pp. 1963–
1976, 2012.
[10] A. Parisio, E. Rikos, and L. Glielmo, “A model predictive control
approach to microgrid operation optimization,” IEEE Trans. on
Control Systems Technology, vol. 22, no. 5, pp. 1813–1827, 2014.
[11] P. Fortenbacher, A. Ulbig, and G. Andersson, “Optimal placement
and sizing of distributed battery storage in low voltage grids using
receding horizon control strategies,” IEEE Trans. on Power Systems,
2017.
[12] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
optimization and statistical learning via the alternating direction
method of multipliers,” Foundations and Trends® in Machine Learning, vol. 3, no. 1, pp. 1–122, 2011.
[13] V. M. Zavala, “New architectures for hierarchical predictive control,”
IFAC-PapersOnLine, vol. 49, no. 7, pp. 43–48, 2016.
[14] L. Gan, N. Li, U. Topcu, and S. H. Low, “Exact convex relaxation of
optimal power flow in radial networks,” IEEE Trans. on Automatic
Control, vol. 60, no. 1, pp. 72–87, 2015.
[15] K. Sundar, H. Nagarajan, M. Lubin, L. Roald, S. Misra, R. Bent,
and D. Bienstock, “Unit commitment with N-1 security and wind
uncertainty,” in Power Systems Computation Conference (PSCC),
2016. IEEE, 2016, pp. 1–7.
[16] I. Koutsopoulos, V. Hatzi, and L. Tassiulas, “Optimal energy storage
control policies for the smart power grid,” in IEEE International
Conference on Smart Grid Communications, 2011, pp. 475–480.
[17] J. K. Felder and I. A. Hiskens, “Optimal power flow with storage,”
in Power Systems Computation Conference (PSCC), 2014. IEEE,
2014, pp. 1–7.
[18] W. H. Kersting, “Radial distribution test feeders,” in Power Engineering Society Winter Meeting, 2001. IEEE, vol. 2. IEEE, 2001, pp.
908–912.
[19] I. Dunning, J. Huchette, and M. Lubin, “JuMP: A modeling language
for mathematical optimization,” arXiv preprint arXiv:1508.01982,
2015.
| 3cs.SY
|
DYNAMIC E VALUATION OF N EURAL S EQUENCE M ODELS
Ben Krause, Emmanuel Kahembwe, Iain Murray, & Steve Renals
School of Informatics, University of Edinburgh
Edinburgh, Scotland, UK
ben.krause,e.kahembwe,i.murray,[email protected]
arXiv:1709.07432v2 [cs.NE] 25 Oct 2017
A BSTRACT
We present methodology for using dynamic evaluation to improve neural sequence
models. Models are adapted to recent history via a gradient descent based mechanism, causing them to assign higher probabilities to re-occurring sequential patterns.
Dynamic evaluation outperforms existing adaptation approaches in our comparisons. Dynamic evaluation improves the state-of-the-art word-level perplexities
on the Penn Treebank and WikiText-2 datasets to 51.1 and 44.3 respectively, and
the state-of-the-art character-level cross-entropies on the text8 and Hutter Prize
datasets to 1.19 bits/char and 1.08 bits/char respectively.
1
I NTRODUCTION
Sequence generation and prediction tasks span many modes of data, ranging from audio and language
modelling, to more general timeseries prediction tasks. Applications of such models include speech
recognition, machine translation, dialogue generation, speech synthesis, forecasting, and music
generation, among others. Neural networks can be applied to these tasks by predicting sequence
elements one-by-one, conditioning on the history of sequence elements, forming an autoregressive
model. Convolutional neural networks (CNNs) and recurrent neural networks (RNNs), including
long-short term memory (LSTM) networks (Hochreiter & Schmidhuber, 1997) in particular, have
achieved many successes at these tasks. However, in their basic form, these models have a limited
ability to adapt to recently observed parts of a sequence.
Many sequences contain repetition; a pattern that occurs once is more likely to occur again. For
instance, a word that occurs once in a document is much more likely to occur again. A sequence of
handwriting will generally stay in the same handwriting style. A sequence of speech will generally
stay in the same voice. Although RNNs have a hidden state that can summarize the recent past, they
are often unable to exploit new patterns that occur repeatedly in a test sequence.
This paper concerns dynamic evaluation, which we investigate as a candidate solution to this problem.
Our approach adapts models to recent sequences using gradient descent based mechanisms. We show
several ways to improve on past dynamic evaluation approaches in Section 5, and use our improved
methodology to achieve state-of-the-art results in Section 7.1 and Section 7.2. In Section 6 we design
a method to dramatically to reduce the number of adaptation parameters in dynamic evaluation,
making it practical in a wider range of situations. In Section 7.3 we analyse dynamic evaluation’s
performance over varying time-scales and distribution shifts, and demonstrate that dynamically
evaluated models can generate conditional samples that repeat many patterns from the conditioning
data.
2
M OTIVATION
Generative models can assign probabilities to sequences by modelling each term in the factorization
given by the product rule. The probability of a sequence x1:T = {x1 , . . . , xT } factorizes as
P (x1:T ) = P (x1 )P (x2 |x1 )P (x3 |x2 , x1 ) · · · P (xT |x1 . . . xT −1 ).
(1)
Methods that apply this factorization either use a fixed context when predicting P (xt |x1:t−1 ), for
instance as in N-grams or CNNs, or use a recurrent hidden state to summarize the context, as in an
1
RNN. However, for longer sequences, the history x1:t−1 often contains re-occurring patterns that are
difficult to capture using models with fixed parameters (static models).
In many domains, in a dataset of sequences {x11:T , x21:T , ..., xn1:T }, each sequence xi1:T is generated
from a slightly different distribution P (xi1:T ). At any point in time t, the history of a sequence xi1:t−1
contains useful information about the generating distribution for that specific sequence P (xi1:T ).
Therefore adapting the model parameters learned during training θg is justified. We aim to infer a set
of model parameters θl from xi1:t−1 that will better approximate P (xit |xi1:t−1 ) within sequence i.
Many sequence modelling tasks are characterised by sequences generated from slightly different
distributions as in the scenario described above. The generating distribution may also change
continuously across a single sequence; for instance, a text excerpt may change topic. Furthermore,
many machine learning benchmarks do not distinguish between sequence boundaries, and concatenate
all sequences into one continuous sequence. Thus, many sequence modellingR tasks could be seen as
having a local distribution Pl (x) as well as a global distribution Pg (x) := P (l)Pl (x) dl. During
training time, the goal is to find the best fixed model possible for Pg (x). However, during evaluation
time, a model that can infer the current Pl (x) from the recent history has an advantage.
3
DYNAMIC EVALUATION
Dynamic evaluation methods continuously adapt the model parameters θg , learned at training time,
to parts of a sequence during evaluation. The goal is to learn adapted parameters θl that provide a
better model of the local sequence distribution, Pl (x). When dynamic evaluation is applied in the
present work, a long test sequence x1:T is divided up into shorter sequences of length n. We define
s1:M to be a sequence of shorter sequence segments si
s1:M = {s1 = x1:n , s2 = xn+1:2n , s3 = x2n+1:3n , ..., sM }.
(2)
The initial adapted parameters θl0 are set to θg , and used to compute the probability of the first
segment, P (s1 |θl0 ). This probability gives a cross entropy loss L(s1 ), with gradient ∇L(s1 ), which
is computed using truncated back-propagation through time (Werbos, 1990). The gradient ∇L(s1 ) is
used to update the model, resulting in adapted parameters θl1 , before evaluating P (s2 |θl1 ). The same
procedure is then repeated for s2 , and for each si in the sequence as shown in Figure 1. Gradients
for each loss L(si ) are only backpropagated to the beginning of si , so computation is linear in the
sequence length. Each update applies one maximum likelihood training step to approximate the
current local distribution Pl (x). The computational cost of dynamic evaluation is one forward pass
and one gradient computation through the data, with some slight overhead to apply the update rule
for every sequence segment.
As in all autoregressive models, dynamic evaluation only conditions on sequence elements that it has
already predicted, and so evaluates a valid log-probability for each sequence. Dynamic evaluation can
also be used while generating sequences. In this case, the model generates each sequence segment
si using fixed weights, and performs a gradient descent based update step on L(si ). Applying
dynamic evaluation for sequence generation could result in generated sequences with more consistent
regularities, meaning that patterns that occur in the generated sequence are more likely to occur again.
4
4.1
BACKGROUND
R ELATED APPROACHES
Adaptive language modelling was first considered for n-grams, adapting to recent history via caching
(Jelinek et al., 1991; Kuhn, 1988), and other methods Bellegarda (2004). More recently, the neural
cache approach (Grave et al., 2017) and the closely related pointer sentinel-LSTM (Merity et al.,
2017b) have been used to for adaptive neural language modelling. Neural caching has recently been
used to improve the state-of-the-art at word-level language modelling (Merity et al., 2017a).
The neural cache model learns a type of non-parametric output layer on the fly at test time, which
allows the network to adapt to recent observations. Each past hidden state hi is paired with the next
input xi+1 , and is stored as a tuple (hi , xi+1 ). When a new hidden state ht is observed, the output
probabilities are adjusted to give a higher weight to output words that coincided with past hidden
2
P(s1|θl0)
P(s2|s1,θl1)
model(s1,θl )
hidden
state
s1=x1:n
)
s2
L(
)
s1
L(
0
P(s3|s2,s1,θl2)
1
model(s2,θl )
hidden
state
s2=xn+1:2n
model(s3,θl2)
s3=x2n+1:3n
Figure 1: Illustration of dynamic evaluation. The model evaluates the probability of sequence
segments si . The gradient ∇L(si ) with respect to the log probability of si is used to update the model
parameters θli−1 to θli before the model progresses to the next sequence segment. Dashed edges are
what distinguish dynamic evaluation from static (normal) evaluation.
states with a large inner product (hTt hi ).
Pcache (xt+1 |x1:t , h1:t ) ∝
t−1
X
e(xi+1 ) exp(ωhTt hi ),
(3)
i=1
where e(xi+1 ) is a one hot encoding of xi+1 , and ω is a scaling parameter. The cache probabilities are
interpolated with the base network probabilities to adapt the base network at test time.
The neural cache closely relates to dynamic evaluation, as both methods can be added on top of a base
model for adaptation at test time. The main difference is the mechanism used to fit to recent history:
the neural cache approach uses a non-parametric, nearest neighbours-like method, whereas dynamic
evaluation uses a gradient descent based method to change model parameters dynamically. Both
methods rely on an autoregressive factorisation, as they depend on observing sequence elements after
they are predicted in order to perform adaptation. Dynamic evaluation and neural caching methods
are therefore both applicable to sequence prediction and generation tasks, but not directly to more
general supervised learning tasks.
One drawback of the neural cache method is that it cannot adjust the recurrent hidden state dynamics.
As a result, the neural cache’s ability to capture information that occurs jointly between successive
sequence elements is limited. This capability is critical for adapting to sequences where each element
has very little independent meaning, e.g. character level language modelling.
4.2
DYNAMIC EVALUATION IN NEURAL NETWORKS
Dynamic evaluation of neural language models was proposed by Mikolov et al. (2010). Their
approach simply used stochastic gradient descent (SGD) updates at every time step, computing the
gradient with fully truncated backpropagation through time, which is equivalent to setting n = 1
3
in equation (2). Dynamic evaluation has since been applied to character and word-level language
models (Graves, 2013; Krause et al., 2017; Ororbia II et al., 2017; Fortunato et al., 2017). Previous
work using dynamic evaluation considered it as an aside, and did not explore it in depth.
5
U PDATE RULE METHODOLOGY FOR DYNAMIC EVALUATION
We propose several changes to Mikolov et al. (2010)’s dynamic evaluation method with SGD and
fully truncated backpropagation, which we refer to as traditional dynamic evaluation. The first
modification reduces the update frequency, so that gradients are backpropagated over more timesteps.
This change provides more accurate gradient information, and also improves the computational
efficiency of dynamic evaluation, since the update rule is applied much less often. We use sequence
segments of length 5 for word-level tasks and 20 for character-level tasks.
Next, we add a global decay prior to bias the model towards the parameters θg learned during
training. Our motivation for dynamic evaluation assumes that the local generating distribution Pl (x)
is constantly changing, so it is potentially desirable to weight recent sequence history higher in
adaptation. Adding a global decay prior accomplishes this by causing previous adaptation updates to
decay exponentially over time. For SGD with a global prior, learning rate η and decay rate λ; we
form the update rule
θi ← θi−1 − η∇L(si ) + λ(θg − θli−1 ).
(4)
We then consider using an RMSprop (Tieleman & Hinton, 2012) derived update rule for the learning
rule in place of SGD. RMSprop uses a moving average of recent squared gradients to scale learning
rates for each weight. In dynamic evaluation, near the start of a test sequence, RMSprop has had very
few gradients to average, and therefore may not be able to leverage its updates as effectively. For this
reason, we collect mean squared gradients, MSg , on the training data rather than on recent test data
(which is what RMSprop would do). MSg is given by
MSg =
Nb
1 X
(∇Lk )2 ,
Nb
(5)
k=1
where Nb is the number of training batches and ∇Lk is the gradient on the kth training batch. The
mini-batch size for this computation becomes a hyper-parameter, as larger mini-batches will result
in smaller mean squared gradients. The update rule, which we call RMS with a global prior in our
experiments, is then
∇L(si )
+ λ(θg − θli−1 ),
(6)
θli ← θli−1 − η p
MSg +
where is a stabilization parameter. For the decay step
p of our update rule, we also consider scaling
the decay rate for each parameter proportionally to MSg . Parameters with a high RMS gradient
p
affect the dynamics of the network more, so it makes sense to decay them faster. RMS norm is MSg
p
divided by its mean, resulting in a normalized version of MSg with a mean of 1:
p
MSg
p
RMS norm =
.
(7)
avg( MSg )
We clip the values of RMS norm to be no greater than 1/λ to be sure that the decay rate does not
exceed 1 for any parameter. Combining the learning component and the regularization component
results in the final update equation, which we refer to as RMS with an RMS global prior
∇L(si )
θli ← θli−1 − η p
+ λ(θg − θli−1 )
MSg +
6
RMS norm .
(8)
S PARSE DYNAMIC EVALUATION
Mini-batching over sequences is desirable for some test-time sequence modelling applications because
it allows faster processing of multiple sequences in parallel. Dynamic evaluation has a high memory
4
cost for mini-batching because it is necessary to store a different set of parameters for each sequence
in the mini-batch. Therefore, we consider a sparse dynamic evaluation variant that updates a smaller
number of parameters. We introduce a new adaptation matrix M which is initialized to zeros. M
multiplies hidden state vector ht of an RNN at every time-step to get a new hidden state h0t , via
h0t = ht + Mht .
(9)
h0t
then replaces ht and is propagated throughout the network via both recurrent and feed-forward
connections. Applying dynamic evaluation to M avoids the need to apply dynamic evaluation to
the original parameters of the network, reduces the number of adaptation parameters, and makes
mini-batching less memory intensive. We reduce the number of adaptation parameters further by
only using M to transform an arbitrary subset of H hidden units. This results in M being an H ×H
matrix with d = H 2 adaptation parameters. If H is chosen to be much less than the number of hidden
units, this reduces the number of adaptation parameters dramatically. In Section 7.2 we experiment
with sparse dynamic evaluation for character-level language models.
7
E XPERIMENTS
We applied dynamic evaluation to word-level and character-level language modelling. In all tasks,
we evaluate dynamic evaluation on top of a base model. After training the base model, we tune
hyper-parameters for dynamic evaluation on the validation set, and evaluate both the static and
dynamic versions of the model on the test set. We also consider follow up experiments that analyse
the sequence lengths for which dynamic evaluation is useful. Code for our dynamic evaluation
methodology is available1 .
7.1
W ORD - LEVEL LANGUAGE MODELLING
We train base models on the Penn Treebank (PTB, Marcus et al., 1993) and WikiText-2 (Merity et al.,
2017b) datasets, and compare the performance of static and dynamic evaluation. These experiments
compare dynamic evaluation against past approaches such as the neural cache and measure dynamic
evaluation’s general performance across different models and datasets.
PTB is derived from articles of the Wall Street Journal. It contains 929k training tokens and a vocab
size limited to 10k words. It is one of the most commonly used benchmarks in language modelling.
We consider two baseline models on PTB, a standard LSTM implementation with recurrent dropout
(Zaremba et al., 2014), and the recent state-of-the-art AWD-LSTM (Merity et al., 2017a).
Our standard LSTM was taken from the Chainer tutorial on language modelling2 , and used two
LSTM layers with 650 units each, trained with SGD and regularized with recurrent dropout. On
our standard LSTM, we experiment with traditional dynamic evaluation as applied by Mikolov et al.
(2010), as well as each modification we make building up to our final update rule as described in
Section 5. As our final update rule (RMS + RMS global prior) worked best, we use this for all other
experiments and use “dynamic eval” by default to refer to this update rule in tables.
We applied dynamic evaluation on a more powerful model, the ASGD weight-dropped LSTM
(AWD-LSTM, Merity et al., 2017a). The AWD-LSTM is a vanilla LSTM that combines the use of
drop-connect (Wan et al., 2013) on recurrent weights for regularization, and a variant of averaged
stochastic gradient descent (Polyak & Juditsky, 1992) for optimisation. Our model, which used 3
layers and tied input and output embeddings (Press & Wolf, 2017; Inan et al., 2017), was intended to
be a direct replication of AWD-LSTM, using code from their implementation3 . Results are given in
Table 1.
Dynamic evaluation gives significant overall improvements to both models on this dataset. Dynamic
evaluation also achieves better final results than the neural cache on both a standard LSTM and the
AWD-LSTM reimplementation, and improves the state-of-the-art on PTB.
WikiText-2 is roughly twice the size of PTB, with 2 million training tokens and a vocab size of 33k.
It features articles in a non-shuffled order, with dependencies across articles that adaptive methods
1
https://github.com/benkrause/dynamic-evaluation
https://github.com/chainer/chainer/tree/master/examples/ptb
3
https://github.com/salesforce/awd-lstm-lm
2
5
model
parameters
RNN+LDA+kN-5+cache (Mikolov & Zweig, 2012)
CharCNN (Kim et al., 2016)
LSTM (Zaremba et al., 2014)
Variational LSTM (Gal & Ghahramani, 2016)
Pointer sentinel-LSTM (Merity et al., 2017b)
Variational LSTM + augmented loss (Inan et al., 2017)
Variational RHN (Zilly et al., 2017)
NAS cell (Zoph & Le, 2017)
Variational LSTM + gradual learning (Aharoni et al., 2017)
LSTM + BB tuning (Melis et al., 2017)
19M
66M
66M
21M
51M
23M
54M
105M
24M
valid
test
60.9
92.0
78.9
78.4
73.4
70.9
68.5
65.4
62.4
61.7
58.3
82.3
72.1
85.6
76.2
75.6
74.8
72.2
71.7
57.3
52.8
57.7
51.1
82.2
72.4
71.1
67.9
LSTM (Grave et al., 2017)
LSTM + neural cache (Grave et al., 2017)
LSTM (ours)
LSTM + traditional dynamic eval (sgd, bptt=1)
LSTM + dynamic eval (sgd, bptt=5)
LSTM + dynamic eval (sgd, bptt=5, global prior)
LSTM + dynamic eval (RMS, bptt=5, global prior)
LSTM + dynamic eval (RMS, bptt=5, RMS global prior)
20M
20M
20M
20M
20M
20M
86.9
74.6
88.0
78.6
78.0
77.4
74.3
73.5
AWD-LSTM (Merity et al., 2017a)
AWD-LSTM +neural cache (Merity et al., 2017a)
AWD-LSTM (ours)
AWD-LSTM + dynamic eval
24M
24M
24M
24M
60.0
53.9
59.8
51.6
Table 1: Penn Treebank perplexities. bptt refers to sequence segment lengths.
model
parameters
valid
test
Byte mLSTM (Krause et al., 2016)
Variational LSTM (Inan et al., 2017)
Pointer sentinel-LSTM (Merity et al., 2017b)
LSTM + BB tuning (Melis et al., 2017)
46M
28M
24M
92.8
91.5
84.8
69.1
88.8
87.0
80.8
65.9
LSTM (Grave et al., 2017)
LSTM + neural cache (Grave et al., 2017)
LSTM (ours)
LSTM + dynamic eval
50M
50M
104.2
72.1
109.1
63.7
99.3
68.9
103.4
59.8
AWD-LSTM (Merity et al., 2017a)
AWD-LSTM + neural cache (Merity et al., 2017a)
AWD-LSTM (ours)
AWD-LSTM + dynamic eval
33M
33M
33M
33M
68.6
53.8
68.9
46.4
65.8
52.0
66.1
44.3
Table 2: WikiText-2 perplexities.
should be able to exploit. For this dataset, we use the same baseline LSTM implementation and
AWD-LSTM re-implementation as on PTB. Results are given in Table 2.
Dynamic evaluation improves the state-of-the-art perplexity on WikiText-2, and provides a significantly greater improvement than neural caching to both base models. This suggests that dynamic
evaluation is effective at exploiting regularities that co-occur across non-shuffled documents.
6
model
parameters
test
Stacked LSTM (Graves, 2013)
Stacked LSTM + traditional dynamic eval (Graves, 2013)
Multiplicative integration LSTM (Wu et al., 2016)
HyperLSTM (Ha et al., 2017)
Hierarchical multiscale LSTM (Chung et al., 2017)
Bytenet decoder (Kalchbrenner et al., 2016)
LSTM + BB tuning (Melis et al., 2017)
Recurrent highway networks (Zilly et al., 2017)
Fast-slow LSTM (Mujika et al., 2017)
21M
21M
17M
27M
46M
46M
47M
1.67
1.33
1.44
1.34
1.32
1.31
1.30
1.27
1.25
mLSTM (Krause et al., 2016)
mLSTM + sparse dynamic eval (d = 250k)
mLSTM + dynamic eval
46M
46M
46M
1.24
1.13
1.08
Table 3: Hutter Prize test set error in bits/char.
model
parameters
test
Multiplicative RNN (Mikolov et al., 2012)
Multiplicative integration LSTM (Wu et al., 2016)
LSTM (Cooijmans et al., 2017)
Batch normalised LSTM (Cooijmans et al., 2017)
Hierarchical multiscale LSTM (Chung et al., 2017)
Recurrent highway networks (Zilly et al., 2017)
5M
4M
45M
1.54
1.44
1.43
1.36
1.29
1.27
mLSTM (Krause et al., 2016)
mLSTM + dynamic eval
45M
45M
1.27
1.19
Table 4: text8 test set error in bits/char.
7.2
C HARACTER - LEVEL LANGUAGE MODELLING
We consider dynamic evaluation on the text84 , and Hutter Prize (Hutter, 2006) datasets. The Hutter
Prize dataset is comprised of Wikipedia text, and includes XML and characters from non-Latin
languages. It is 100 million UTF-8 bytes long and contains 205 unique bytes. Similarly to other
reported results, we use a 90-5-5 split for training, validation, and testing. The text8 dataset is also
derived from Wikipedia text, but has all XML removed, and is lower cased to only have 26 characters
of English text plus spaces. As with Hutter Prize, we use the standard 90-5-5 split for training,
validation, and testing for text8. We used a multiplicative LSTM (mLSTM) (Krause et al., 2016)5 as
our base model for both datasets. The mLSTMs for both tasks used 2800 hidden units, an embedding
layer of 400 units, weight normalization (Salimans & Kingma, 2016), variational dropout (Gal &
Ghahramani, 2016), and ADAM (Kingma & Ba, 2014) for training.
We also consider sparse dynamic evaluation, as described in Section 6, on the Hutter Prize dataset.
For sparse dynamic evaluation, we adapted a subset of 500 hidden units, resulting in a 500×500
adaptation matrix and 250k adaptation parameters. All of our dynamic evaluation results in this
section use the final update rule given in Section 5. Results for Hutter Prize are given in Table 3, and
results for text8 are given in Table 4.
Dynamic evaluation achieves large improvements to our base models and state-of-the-art results on
both datasets. Sparse dynamic evaluation also achieves significant improvements on Hutter Prize
using only 0.5% of the adaptation parameters of regular dynamic evaluation.
4
5
http://mattmahoney.net/dc/textdata
https://github.com/benkrause/mLSTM
7
1.6
3.2
static
dynamic
static
dynamic
3
bits per character
bits per character
1.5
1.4
1.3
1.2
1.1
1
2.8
2.6
2.4
2.2
2
0
2000
4000
6000
8000
10000
0
2000
4000
6000
8000
number of characters
number of characters
(a) Hutter data
(b) Spanish data
10000
Figure 2: Average losses in bits/char of dynamic evaluation and static evaluation plotted against
number of characters processed; on sequences from the Hutter Prize test set (left) and European
Parliament dataset in Spanish (right), averaged over 500 trials for each. Losses at each data point are
averaged over sequence segments of length 100, and are not cumulative. Note the different y-axis
scales in the two plots.
7.3
T IME - SCALES OF DYNAMIC EVALUATION
We measure time-scales at which dynamic evaluation gains an advantage over static evaluation.
Starting from the model trained on Hutter Prize, we plot the performance of static and dynamic
evaluation against the number of characters processed on sequences from the Hutter Prize test set,
and sequences in Spanish from the European Parliament dataset (Koehn, 2005).
The Hutter Prize data experiments show the timescales at which dynamic evaluation gained the
advantage observed in Table 3. We divided the Hutter Prize test set into 500 sequences of length
10000, and applied static and dynamic evaluation to these sequences using the same model and
methodology used to obtain results in Table 3. Losses were averaged across these 500 sequences to
obtain average losses at each time step. Plots of the average cross-entropy errors against the number
of Hutter characters sequenced are given in Figure 2a.
The Spanish experiments measure how dynamic evaluation handles large distribution shifts between
training and test time, as Hutter Prize contains very little Spanish. We used the first 5 million
characters of the Spanish European Parliament data in place of the Hutter Prize test set. The Spanish
experiments used the same base model and dynamic evaluation settings as Hutter Prize. Plots of
the average cross-entropy errors against the number of Spanish characters sequenced are given in
Figure 2b.
On both datasets, dynamic evaluation gave a very noticeable advantage after a few hundred characters.
For Spanish this advantage continued to grow as more of the sequence was processed, whereas
for Hutter, this advantage was maximized after viewing around 2-3k characters. The advantage of
dynamic evaluation was also much greater on Spanish sequences than Hutter sequences.
We also drew 300 character conditional samples from the static and dynamic versions of our model
after viewing 10k characters of Spanish. For the dynamic model, we continued to apply dynamic
evaluation during sampling as well, by the process described in Section 3. The conditional samples
are given in the appendix. The static samples quickly switched to English that resembled Hutter Prize
data. The dynamic model generated data with some Spanish words and a number of made up words
with characteristics of Spanish words for the entirety of the sample. This is an example of the kinds
of features that dynamic evaluation was able to learn to model on the fly.
8
C ONCLUSION
This work explores and develops methodology for applying dynamic evaluation to sequence modelling
tasks. Experiments show that the proposed dynamic evaluation methodology gives large test time
improvements across character and word level language modelling. Our improvements to language
8
modelling have applications to speech recognition and machine translation over longer contexts,
including broadcast speech recognition and paragraph level machine translation. Overall, dynamic
evaluation is shown to be an effective method for exploiting pattern re-occurrence in sequences.
R EFERENCES
Z. Aharoni, G. Rattner, and H. Permuter. Gradual learning of deep recurrent neural networks. arXiv preprint
arXiv:1708.08863, 2017.
J. R. Bellegarda. Statistical language model adaptation: review and perspectives. Speech Communication, 42(1):
93–108, 2004.
J. Chung, S. Ahn, and Y. Bengio. Hierarchical multiscale recurrent neural networks. ICLR, 2017.
T. Cooijmans, N. Ballas, C. Laurent, and A. Courville. Recurrent batch normalization. ICLR, 2017.
M. Fortunato, C. Blundell, and O. Vinyals. Bayesian recurrent neural networks. arXiv preprint arXiv:1704.02798,
2017.
Y. Gal and Z. Ghahramani. A theoretically grounded application of dropout in recurrent neural networks. In
Advances in neural information processing systems, pp. 1019–1027, 2016.
E. Grave, A. Joulin, and N. Usunier. Improving neural language models with a continuous cache. ICLR, 2017.
A. Graves. Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850, 2013.
D. Ha, A. Dai, and Q. Lee. Hypernetworks. ICLR, 2017.
S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation, 9:1735–1780, 1997.
M. Hutter. The human knowledge compression prize. URL http://prize.hutter1.net, 2006.
H. Inan, K. Khosravi, and R. Socher. Tying word vectors and word classifiers: A loss framework for language
modeling. ICLR, 2017.
F. Jelinek, B. Merialdo, S. Roukos, and M. Strauss. A dynamic language model for speech recognition. In HLT,
volume 91, pp. 293–295, 1991.
N. Kalchbrenner, L. Espeholt, K. Simonyan, A. Oord, A. Graves, and K. Kavukcuoglu. Neural machine
translation in linear time. arXiv preprint arXiv:1610.10099, 2016.
Y. Kim, Y. Jernite, D. Sontag, and A. M. Rush. Character-aware neural language models. In Thirtieth AAAI
Conference on Artificial Intelligence, 2016.
D. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
P. Koehn. Europarl: A parallel corpus for statistical machine translation. In MT Summit, volume 5, pp. 79–86,
2005.
B. Krause, L. Lu, I. Murray, and S. Renals. Multiplicative LSTM for sequence modelling. arXiv preprint
arXiv:1609.07959, 2016.
B. Krause, I. Murray, S. Renals, and L. Lu. Multiplicative LSTM for sequence modelling. ICLR Workshop track,
2017. URL https://openreview.net/forum?id=SJCS5rXFl.
R. Kuhn. Speech recognition and the frequency of recently used words: A modified Markov model for natural
language. In Proceedings of the 12th conference on Computational linguistics-Volume 1, pp. 348–350.
Association for Computational Linguistics, 1988.
M. P. Marcus, M. A. Marcinkiewicz, and B. Santorini. Building a large annotated corpus of English: The Penn
Treebank. Computational linguistics, 19(2):313–330, 1993.
G. Melis, C. Dyer, and P. Blunsom. On the state of the art of evaluation in neural language models. arXiv
preprint arXiv:1707.05589, 2017.
S. Merity, N. S. Keskar, and R. Socher. Regularizing and optimizing LSTM language models. arXiv preprint
arXiv:1708.02182, 2017a.
S. Merity, C. Xiong, J. Bradbury, and R. Socher. Pointer sentinel mixture models. ICLR, 2017b.
T. Mikolov and G. Zweig. Context dependent recurrent neural network language model. SLT, 12:234–239, 2012.
T. Mikolov, M. Karafiát, L. Burget, J. Cernockỳ, and S. Khudanpur. Recurrent neural network based language
model. In Interspeech, volume 2, pp. 3, 2010.
T. Mikolov, I. Sutskever, A. Deoras, H. Le, S. Kombrink, and J. Cernocky. Subword language modeling with
neural networks. preprint (http://www. fit. vutbr. cz/imikolov/rnnlm/char. pdf), 2012.
A. Mujika, F. Meier, and A. Steger. Fast-slow recurrent neural networks. arXiv preprint arXiv:1705.08639,
2017.
A. G. Ororbia II, T. Mikolov, and D. Reitter. Learning simpler language models with the differential state
framework. Neural Computation, 2017.
B. T. Polyak and A. B. Juditsky. Acceleration of stochastic approximation by averaging. SIAM Journal on
Control and Optimization, 30(4):838–855, 1992.
O. Press and L. Wolf. Using the output embedding to improve language models. EACL 2017, pp. 157, 2017.
T. Salimans and D. P. Kingma. Weight normalization: A simple reparameterization to accelerate training of deep
neural networks. In Advances in Neural Information Processing Systems, pp. 901–909, 2016.
T. Tieleman and G. E. Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent
magnitude. COURSERA: Neural Networks for Machine Learning, 4(2), 2012.
L. Wan, M. Zeiler, S. Zhang, Yann L. Cun, and R. Fergus. Regularization of neural networks using dropconnect.
In Proceedings of the 30th international conference on machine learning (ICML-13), pp. 1058–1066, 2013.
9
P. J. Werbos. Backpropagation through time: what it does and how to do it. Proceedings of the IEEE, 78:
1550–1560, 1990.
Y. Wu, S. Zhang, Y. Zhang, Y. Bengio, and R. Salakhutdinov. On multiplicative integration with recurrent neural
networks. In NIPS, 2016.
W. Zaremba, I. Sutskever, and O. Vinyals. Recurrent neural network regularization. arXiv preprint
arXiv:1409.2329, 2014.
J. G. Zilly, R. K. Srivastava, J. Koutník, and J. Schmidhuber. Recurrent highway networks. ICLR, 2017.
B. Zoph and Quoc V Le. Neural architecture search with reinforcement learning. ICLR, 2017.
A
A.1
A PPENDIX
DYNAMIC SAMPLES CONDITIONED ON S PANISH
300 character samples generated from the dynamic version of the model trained on Hutter Prize,
conditioned on 10k of Spanish characters. The final sentence fragment of the 10k conditioning
characters is given to the reader, with the generated text given in bold:
Tiene importancia este compromiso en la medida en que la Comisión es un organismo que
tiene el montembre tas procedíns la conscriptione se ha Tesalo del Pómienda que et hanemos
que Pe la Siemina.
De la Pedrera Orden es Señora Presidente civil, Orden de siemin presente relevante frónmida
que esculdad pludiore e formidad President de la Presidenta Antidorne Adamirmidad i
ciemano de el 200’. Fo
A.2
S TATIC SAMPLES CONDITIONED ON S PANISH
300 character samples generated from the static version of the model trained on Hutter Prize,
conditioned on 10k of Spanish characters. The final sentence fragment of the 10k conditioning
characters is given to the reader, with the generated text given in bold:
Tiene importancia este compromiso en la medida en que la Comisión es un organismo que
tiene el monde,
<br>There is a secret act in the world except Cape Town, seen in now flat comalo
and ball market and has seen the closure of the eagle as imprints in a dallas within the
country." Is a topic for an increasingly small contract saying Allan Roth acquired the
government in [[1916]].
===
10
| 9cs.NE
|
WORD EQUATIONS I:
arXiv:1607.05431v1 [math.GR] 19 Jul 2016
PAIRS AND THEIR MAKANIN-RAZBOROV DIAGRAMS
Z. Sela1,2
This paper is the first in a sequence on the structure of sets of solutions to systems
of equations over a free semigroup. To describe the structure, we present a MakaninRazborov diagram that encodes the set of solutions to such system of equations. In
the sequel we show how this diagram, and the tools that are used in constructing it,
can be applied to analyze fragments of the first order theory of free semigroups.
A free semigroup can be viewed as the most basic formal language. This connection, and the analogy with Tarski’s problem on the first order theories of free groups,
led W. Quine to study the first order theory of a free semigroup. In 1946 Quine
proved that arithmetic can be defined in a free semigroup. Hence, by Godels incompleteness theorems, the theory is not axiomatizable, nor decidable [Qu]. Later,
smaller and smaller fragments of the theory have been shown to be undecidable, including sentences with only two quantifiers, by Durnev ([Du1],[Du2]), Marchenkov
[Mar], and others.
On the positive side, the Diophantine theory of a free semigroup has been shown
to be decidable by G. S. Makanin [Ma1]. Makanin presented an algorithm that
decides if a given system of equations over a free semigroup has a solution. Several
years later Makanin was able to modify his algorithm to decide if a given system
of equations over a free group has a solution [Ma2].
In 1987, A. A. Razborov managed to use Makanins techniques and gave a combinatorial description of the set of solutions to a system of equations (variety) over
a free group. This description was further developed by O. Kharlampovich and
A. Myasnikov [Kh-My], and a more geometric approach was given by the author
in [Se1]. The description of varieties over a free group is the starting point to a
structure theory that finally led to quantifier elimination ([Se3],[Se4]), and to the
solution of Tarski’s problem on the first order theories of free groups [Se5].
The search for a description of the set of solutions to a system of equations
over free semigroups has a long history (see [Di]). For equations with one variable,
and then for equations with 3 variables, the structure of the set of solutions over
free semigroup was achieved before the analogous structures over free groups (see
[Khm]). In these cases the structure that was found was purely combinatorial.
Since the mid 1980’s and in particular after Razborov’s thesis, there have been
quite a few attempts to study sets of solutions over a free semigroup, usually for
particular families of systems, that are often either with small number of variables,
1 Hebrew
University, Jerusalem 91904, Israel.
supported by an Israel academy of sciences fellowship.
2 Partially
or that are of rather particular type, mostly connected to quadratic equations over
a free group (e.g. [Ma3],[Ma4],[Ma5],[Ly],[Di] and their references).
In 2013 an algorithm to enumerate and encode the set of solutions to a general
system of equations over a free semigroup was found (by A. Jez). This algorithm
that enumerates the solutions efficiently is based on variants of the Ziv-Lempel
algorithm from information theory ([Je],[DJP]).
In this paper we present a geometric approach to study varieties over a free
semigroup. We use the combinatorial techniques that were introduced by Makanin
in proving the decidability of the Diophantine theory [Ma1], and we combine them
with geometric techniques that were used in the construction of the JSJ decomposition of finitely presented groups ([Se7],[Ri-Se]), and with techniques that appear
in the solution to Tarski’s problem.
Unfortunately, even though our approach is based on the construction of the
JSJ decomposition for groups, we were not able to find an analogue of the JSJ for
studying varieties over free semigroups (we were able to get such analogue in some
special cases). However, we are able to find an analogue of Razborov’s work over
free groups, and associate what we call a Makanin-Razborov diagram with each
variety over a free semigroup.
The MR diagram that we construct is not canonical, but it encodes all the
points in a given variety over a free semigroup. Furthermore, given a path in
the diagram, that we call a resolution, there exists a sequence of points in the
associated variety that factor through the resolution, and such that the sequence
converges in the Gromov-Hausdorff topology to an object from which the resolution
can be reconstructed. Such sequences are viewed as generic points in the variety
(a replacement for test sequences that exist over a free group), and are used in the
sequel to prove model theoretic results, that include generalizations of Merzlyakov’s
theorem (over varieties) for free semigroups (see [Me] and [Se2]).
The MR diagram over semigroups, its properties and the way it is constructed,
suggest that very basic questions about words (in formal languages) are connected
to concepts, objects and tools from low dimensional topology. These include the
JSJ decomposition, the geometry and topology of surfaces, the dynamics of automorphisms of surfaces and of free groups, and finally the dynamics of actions of
groups on real trees. We expect that this combination of techniques and structure
can be modified to describe sets of solutions to systems of equations over free objects in other (associative and non-commutative) algebraic categories, and we plan
to continue in these directions.
Given a system of equations Σ over a free semigroup:
u1 (x1 , . . . , xn ) = v1 (x1 , . . . , xn )
..
.
us (x1 , . . . , xn ) = vs (x1 , . . . , xn )
we naturally associate with it a f.p. semigroup:
S(Σ) = h x1 , . . . , xn | u1 = v1 , . . . , us = vs i
The set of solutions of Σ over a free semigroup, F Sk =< a1 , . . . , ak >, is in
bijection with the set of semigroup homomorphisms: {h : S(Σ) → F Sk }. Hence,
2
studying the variety of solutions to Σ is equivalent to studying the structure of the
set of homomorphisms from S(Σ) to the free semigroup F Sk .
With a f.g. semigroup S we associate a group G(S), that is obtained by forcing
all the elements in S to have inverses, or alternatively, by looking at a presentation
of S as a semigroup as if it is a presentation of a group. Naturally, S is mapped
into G(S), but in general it is not embedded in it. Let Ŝ be the (semigroup) image
of S in G(S).
The free semigroup F Sk embeds in the free group Fk in a standard way. A simple
observation shows that every semigroup homomorphism from S to F Sk extends
uniquely to a group homomorphism from G(S) to Fk (see section 1). Hence, one
can replace the set of semigroup homomorphisms from S to the free semigroup
F Sk , with the set of pair homomorphisms: {η : (Ŝ, G(S)) → (F Sk , Fk )}. i.e.,
those group homomorphisms from G(S) to Fk that map Ŝ into the standard free
semigroup F Sk .
Our whole approach is based on studying the structure of these pair homomorphisms. An immediate application of the techniques that were used over free
groups, shows that the set of pair homomorphisms from (Ŝ, G(S)) to (F Sk , Fk ) is
canonically a finite union of the sets of pair homomorphisms from pairs of the form
(Si , Li ) to (F Sk , Fk ), where for each i, Li is a limit group, Si is a semigroup that
generates Li as a group, and (Si , Li ) is a (limit) quotient of the pair (Ŝ, G(S)).
A free semigroup, and the semigroups that we need to consider, have usually very
few automorphisms. The ability to replace a semigroup with a pair in which the
ambient group is a limit group, enables one to work with a large group of automorphisms (of the limit group), that usually do not preserve the embedded semigroup.
In section 2 we describe an analogue of the shortening procedure for homomorphisms of pairs. Technically, the shortening procedure for homomorphisms of pairs
is much more involved than its analogue for homomorphisms of groups.
In section 3 we describe a construction of a JSJ decomposition for a pair (S, L),
where L is a limit group, and S generates it as a group. Unfortunately, the construction applies only in special cases. Note that free products exist in the categories of
groups, of semigroups, and of pairs. However, it may be that a pair (S, L), where L
is a limit group, is freely indecomposable as a pair, but the limit group L is freely
decomposable. This simple fact implies that any attempt to borrow concepts from
the JSJ theory and from the construction of the MR diagram over free groups must
be further refined. In section 6 we describe the construction of the MR diagram
for pairs in the freely indecomposable case (theorem 6.8). In section 7 we finally
describe the construction of the MR diagram for pairs in the general case (theorems
7.17 and 7.18).
As over free (and hyperbolic) groups, there are pairs, (Si , Li ), associated with
the nodes of the MR diagram, together with their associated decompositions and
modular groups. However, unlike the case of groups, the abelian decompositions
that are associated with the pairs that appear in the nodes of the diagram, need to
recall not only the algebraic structure of the group in question, but rather dynamical
properties and the associated modular groups. Hence, we need to extend the classes
of vertex groups that we borrow from the JSJ theory of groups. In particular, the
abelian decompositions that appear along the MR diagram for pairs contain a new
type of vertex groups that we call Levitt. These Levitt vertex groups are free
factors of the ambient limit groups that are connected to other vertex groups by
3
edges with trivial stabilizers, and each such vertex group contributes the group of
its automorphisms (that are automorphisms of free groups that are not necessarily
geometric) to the modular group of the pair with which it is associated.
In the MR diagram over free groups, when we follow a path (resolution) in
the diagram, the groups that are associated with the nodes along the path, form a
finite sequence of successive proper quotients (maximal shortening quotients as they
appear in [Se1]). In the MR diagram over free semigroups that we construct this is
not true. When we go over a resolution in the diagram, the pairs (and groups) that
appear along the nodes of the diagram are quotients, but not necessarily proper
quotients.
The resolutions in the MR diagram over free groups end with free groups of various ranks. In the MR diagram over free semigroups, resolutions end with marked
graphs with directed edges, that have free groups as their fundamental groups.
Group homomorphisms from a given free group into the coefficient free group are
easily understood using substitutions, and they are in bijection with a product set
of the coefficient group. In the case of semigroups, the set of homomorphisms that
are associated with a terminal node are also obtained by substitutions, and these
sets can be viewed as natural projections of affine spaces, i.e., natural projections
of product sets of the coefficient semigroup.
Finally I would like to thank Eliyahu Rips who encouraged me to proceed along
this long project, that somehow involved repeated mistakes and misconceptions.
§1. Maximal Pairs
In a similar way to the study of equations over groups [Ra1], with a finite system
of equations Φ over a free semigroup F Sk = < a1 , . . . , ak > it is natural to associate
a f.p. semigroup S(Φ). If the system Φ is defined by the coefficients a1 , . . . , ak , the
unknowns x1 , . . . , xn and the equations:
u1 (a1 , . . . , ak , x1 , . . . , xn ) = v1 (a1 , . . . , ak , x1 , . . . , xn )
..
.
us (a1 , . . . , ak , x1 , . . . , xn ) = vs (a1 , . . . , ak , x1 , . . . , xn )
we set the associated f.p. semigroup S(Φ) to be:
S(Φ) = < a1 , . . . , ak , x1 , . . . , xn | u1 = v1 , . . . , us = vs >
Clearly, every solution of the system Φ corresponds to a homomorphism (of semigroups) h : S(Φ) → F Sk for which h(ai ) = ai , and every such homomorphism
corresponds to a solution of the system Φ. Therefore, the study of sets of solutions
to systems of equations in a free semigroup is equivalent to the study of all homomorphisms from a fixed f.p. semigroup S into a free semigroup F Sk , for which a
given prescribed set of elements in S is mapped to a fixed basis of the free semigroup
F Sk .
Hence, as in studying sets of solutions to systems of equations over a free or a
hyperbolic group [Se1], to study sets of solutions to systems of equations over a free
semigroup, we fix a f.p. (or even a f.g.) semigroup S, and study the structure of its
set of homomorphisms into a free semigroup, F Sk , that we denote, Hom(S, F Sk ).
4
Given a f.g. semigroup, S, we can naturally associate a group with it. Given
a presentation of S as a semigroup, we set the f.g. group Gr(S) to be the group
with the presentation of S interpreted as a presentation of a group. Clearly, the
semigroup S is naturally mapped into the group, Gr(S), and the image of S in
Gr(S) generates Gr(S). We set ηS : S → Gr(S) to be this natural homomorphism
of semigroups.
The free semigroup, F Sk , naturally embeds into a free group, Fk . By the construction of the group, Gr(S), every homomorphism of semigroups, h : S → F Sk ,
extends to a homomorphism of groups, hG : Gr(S) → Fk , so that: h = hG ◦ ηS .
Our goal is to study the structure of the set of homomorphisms (of semigroups),
Hom(S, F Sk ). By construction, every homomorphism (of semigroups), h : S →
F Sk , extends to a homomorphism (of groups), hG : Gr(S) → Fk . Therefore, the
study of the structure of Hom(S, F Sk ), is equivalent to the study of the structure
of the collection of homomorphisms of groups, Hom(Gr(S), Fk ), that restrict to
homomorphisms of (the semigroup) S into the free semigroup (the positive cone),
F Sk .
By the techniques of section 5 in [Se1], with any given collection of homomorphisms of a f.g. group into a free group, we can associate its Zariski closure, and
with the Zariski closure one can associate canonically a finite collection of limit
groups, that are all (maximal) limit quotients of the given f.g. group, so that every homomorphism from the given collection factors through at least one of the
quotient maps from the given f.g. group into the (finitely many) limit quotients.
By (canonically) associating a finite collection of maximal limit quotients with the
set of homomorphisms, Hom(Gr(S), Fk ), that restrict to (semigroup) homomorphisms from S to F Sk , we get the following basic theorem, which is the basis for
our approach to study the structure of Hom(S, F Sk ).
Theorem 1.1. Let S be a f.g. semigroup, and let Gr(S) be the f.g. group that is
associated with S, by interpreting a semigroup presentation of S, as a presentation
of a group. Let ηS : S → Gr(S) be the natural semigroup homomorphism, and note
that ηS (S) generates Gr(S) as a group.
There exists a finite canonical collection of (limit) pairs, (S1 , L1 ), . . . , (Sm , Lm ),
where the Li ’s are limit quotients of Gr(S), and the semigroups, Si , are quotients
of the semigroup S that generate the limit groups Li as groups, with the following
properties:
(1) for each index i, 1 ≤ i ≤ m, there exists a (canonical) quotient map of
pairs, τi : (S, Gr(S)) → (Si , Li ).
(2) by construction, every homomorphism of semigroups, h : S → F Sk , extends
to a map of pairs, hP : (S, Gr(S)) → (F Sk , Fk ). For each such homomorphism of pairs, there exists an index i, 1 ≤ i ≤ m, and a homomorphism of
pairs, uh : (Si , Li ) → (F Sk , Fk ), for which: hP = uh ◦ τi .
Proof:
Identical to the proof of theorem 7.2 in [Se1].
In theorem 7.2 in [Se1], it is shown that with each f.g. group it is possible to
associate a canonical finite collection of limit groups, so that each homomorphism
from the f.g. group into a free group factors through at least one of the finitely
many limit quotients. Theorem 1.1 is the analogue of that theorem for semigroups.
5
It reduces (canonically) the study of the structure of the set of semigroup homomorphisms from a given semigroup to a free semigroup, Hom(S, F Sk ), to the study
of the structure of homomorphisms of finitely many pairs {(Si , Li )} into (F Sk , Fk ),
where the Li ’s are limit groups, and the Si ’s are subsemigroups of the Li ’s that
generate the limit groups Li ’s as groups.
§2. Positive Cones, their embeddings in Standard Cones, and Shortenings
To analyze homomorphisms of a f.g. semigroup into the free semigroup we analyzed sequences of homomorphisms of pairs (S, G) into the standard pair (F Sk , Fk ).
In studying such homomorphisms of pairs, the subsemigroup S is viewed as the positive cone in the ambient group G. To get a structure theory for the entire collection
of pair homomorphisms we are led to replacing the given cone with a standard cone.
The structure of a standard cone depends on the structure of the ambient group,
and more specifically on the structure of the tree that is obtained as a limit of a
sequence of homomorphisms, and the dynamics of the action of the ambient group
on that tree.
In general, a standard cone can not be embedded into the ambient group G,
but rather into an extension of G. Still, a standard cone is essential in applying
the shortening argument for semigroups, for any attempt to construct JSJ decompositions, for separating free factors, and in general for constructing a MakaninRazborov diagram that is associated with a pair (S, L).
Let (S, L) be a pair of a limit group, L, and its subsemigroup, S, that generates
it as a group. Let {hn : (S, L) → (F Sk , Fk )} be a sequence of pair homomorphisms
that converges into a faithful action of L on a real tree Y . In this section we
show how to extend the subsemigroup S to a bigger semigroup, by adding to it
standard generating sets of the various components of the real tree Y . We further
make sure that the (finitely many) generators of the original subsemigroup S can
be expressed as positive words in the standard generating sets that are associated
with the components of the limit tree Y .
The standard generating sets that we add depend on the dynamics of the action
of the limit group L on the limit tree Y . We start with the rather basic case, of an
axial action of a free abelian group on a line.
Theorem 2.1. Let (S, A) be a pair of a free abelian group group A and a subsemigroup S that generates A. Let {hn } be a sequence of pair homomorphisms
of (S, A) into (F Sk , Fk ) that converges into a faithful axial action of A on a line
Y . Let A0 be the direct summand of A that acts trivially on Y , and suppose that:
rk(A) − rk(A0 ) = ℓ.
Then there exists a positive collection of generators, a1 , . . . , aℓ , in A \ A0 so
that:
(1) A = A0 + < a1 > + . . . + < aℓ >.
(2) there exists some index n0 , such that for every n > n0 : hn (a1 ), . . . , hn (aℓ ) ∈
F Sk .
(1) (3) for each of the generators s1 , . . . , sr of the semigroup S, sj can be
expressed as a positive word in a1 , . . . , aℓ modulo an element in A0 . i.e.,
for every j, 1 ≤ j ≤ r:
mj
mj
sj = a1 1 . . . aℓ ℓ a0 (j)
6
where mji ≥ 0 and a0 (j) ∈ A0 .
(4) Wlog we can assume that for every j, 1 ≤ j ≤ r, and every n > n0 ,
hn (a0 (j)) ∈ F Sk .
Proof:
Let a1 , . . . , aℓ be elements in A, for which A = A0 + < a1 > + . . . + < aℓ >, and
so that the elements a1 , . . . , aℓ translate along the axis of A (the real tree Y ) in a
positive direction. Since the action of A on its axis is an axial action, aj translates
a positive distance αj , and the real numbers αj , 1 ≤ j ≤ ℓ, are independent over
the rationals.
Each of the fixed set of generators of the semigroup S, s1 , . . . , sr , can be written
as a word in the elements a1 , . . . , Aℓ times an element in A0 . We modify the
elements a1 , . . . , aℓ iteratively, so that the elements s1 , . . . , sr can be represented
as positive words in the modified set of generators. If an element sj is already
a positive word in the generators, then the modifications of the generators that
we perform, will change the positive word to another positive word in the new
generators.
Suppose that after rearranging the order of the given generators, s, one of the
elements, s1 , . . . , sf , is represented by the word:
ℓ
s = Σm
j=1 kj aj − Σj=m+1 tj aj
Wlog we may assume that tm+1 is one of the maximal elements from the set
tm+1 , . . . , tℓ . To prove the theorem we show that after finitely many modifications of the set of generators, s can be represented by another word in a new set
of positive generators, so that the absolute value of the negative coefficients is still
bounded by tm+1 , and the number of appearances of the coefficient −tm+1 is strictly
smaller than the number of appearances of it in the original word.
Suppose that one of the elements aj , j = 1, . . . , m, satisfies aj > am+1 . In that
case we replace the (positive) generator aj with a positive generator âj = aj −am+1 .
With respect to the new set of generators the coefficients kj are unchanged, −tm+1
is replaced by −tm+1 + kj and tj , j = m + 2, . . . , ℓ are unchanged. In particular, all
the new negative coefficients are bounded below by the previous −tm+1 , and the
number of negative coefficients that are equal to −tm+1 is reduced by 1.
Suppose that all the elements aj , j = 1, . . . , m, satisfy aj < am+1 . We start
by replacing the generator am+1 with âm+1 = am+1 − a1 . This replaces k1 by
k̂1 = k1 − tm+1 and keeps all the other coefficients unchanged. If k̂j = 0 we have
reduced the number of generators that participate in the word that represents the
element s. If k̂j > 0 we do the following:
(1) If there exists an index j, j = 1, . . . , m, for which aj > âm+1 , we replace
the generator aj with aj − am+1 , and hence replace the coefficient −tm+1
with −tm+1 + kj . This reduces the number of negative coefficients that are
equal to the previous −tm+1 by 1.
(2) Suppose that for all j, j = 1, . . . , m, aj < âm+1 . In that case we replace
âm+1 by âm+1 − a1 , and hence replace k̂1 by k̂1 − tm+1 .
In that case after finitely many steps either the coefficient of the first generator
becomes 0 or negative, or we reduced the number of negative coefficients that are
the biggest in their absolute value (i.e., equal to −tm+1 ).
7
If k̂1 < 0 then we reduced the number of generators with positive coefficients
in the representation of the element s. Hence, we repeat the same steps where
the elements that have positive coefficients are only a2 , . . . , am . Therefore, after
finitely many steps the number of generators with negative coefficient that are the
biggest in their absolute value is strictly smaller. Finally, after finitely many steps
we obtain a new set of generators for which the element s is represented by a
positive word. Continuing to the other elements s1 , . . . , sr , we get a generating set
for which all these elements are represented by positive words.
Theorem 2.1 constructs a positive basis in case the ambient limit group is abelian
and its limit action is axial. The next case that we consider is the case of a pair
(S, L), in which the ambient limit group L is a surface group, and L acts freely
on the limit (real) tree. In this case of an IET action of the surface group L, we
were not able to prove the existence of a positive standard cone. Instead we prove
a weaker statement that enables one to apply shortening arguments in the sequel
(shortening arguments are essential in the constructions of the JSJ decompositions,
and the Makanin-Razborov diagrams in the sequel).
Theorem 2.2. Let Q be a (closed) surface group, and let (S, Q) be a pair where S
is a f.g. subsemigroup that generates Q as a group. Let {hn } be a sequence of pair
homomorphisms of (S, Q) into (F Sk , Fk ) that converges into a free (IET) action of
Q on a real tree Y .
Then there exists a sequence of automorphisms {ϕℓ ∈ ACT (Q)}∞
ℓ=1 with the
following properties:
(1) for each index ℓ, there exists nℓ > 0, so that for every n > nℓ , and all the
generators of the subsemigroup S, s1 , . . . , sr , hn ◦ ϕℓ (sj ) ∈ F Sk . (note that
the automorphisms ϕℓ do not preserve the subsemigroup S in general, but
the image of S under the twisted homomorphisms hn ◦ ϕℓ remains in the
standard positive cone in Fk ).
(2) for every n > nℓ and every j, 1 ≤ j ≤ r:
1
dT (hn ◦ ϕℓ (sj ), id)
<
max1≤j≤r dT (hn (sj ), id.)
ℓ
where T is a Cayley graph of the coefficient free group Fk with respect to a
fixed set of generators.
Proof: The generators s1 , . . . , sr of the subsemigroup S generate the surface group
Q. The action of Q on the limit tree Y is an IET action, which is in particular a
geometric action. Hence, there exists a finite subtree R of Y , such that if one applies
the Rips machine to the action of the pseudogroup generated by the restrictions of
the actions of the generators s1 , . . . , sr to the finite tree R, the output is a standard
IET pseudogroup. i.e., the output is supported on a union of finitely many intervals
that are all part of one positively oriented interval, and these finitely many intervals
are divided into finitely many subintervals, a permutation of these subintervals, and
a standard set of generators of Q that map each subinterval to its image (dual)
under the permutation. Note that the generators of Q that are associated with
the IET presentation, are obtained by gluing the finitely many intervals on which
the IET is supported into a (finite) tree, and then the generators of Q are the
8
elements that map a subinterval into its dual. Hence, the generators of Q that are
associated with the IET transformation need not be positively oriented in case the
IET is supported on more than one connected interval. This is in contrast with
the positive orientation of the original generators s1 , . . . , sr , and of each of the
subintervals that define the IET transformation.
For presentation purposes we first assume that the endpoints of the intervals in
the interval exchange (that is obtained by the Rips machine from the action of Q
on Y ) belong to a single orbit under Q (i.e., that the generators of the interval
exchange act transitively on the endpoints of the intervals in the IET), and that
the base point in Y (i.e., the point in Y which is a limit of the identity element in
the Cayley graph of Fk ) belongs to that orbit.
The action of the surface group Q on the tree Y is reduced to an interval exchange
transformation that is based on positively oriented intervals I1 , . . . , Ig . By our
assumption, the points that are at the beginning and at the end of the subintervals
that define the interval exchange are all in the orbit of the basepoint in Y . Let
q0 , . . . , qf ∈ I be the points that are at the beginnings and at the ends of the
subintervals that define the IET transformation on the intervals I1 , . . . , Ig . We
look at those indices i, 1 ≤ i ≤ f , for which the segment [qi−1 , qi ] is supported on
one of the segments I1 , . . . , Ig . For each such index i, there is a unique element in
Q that maps the subinterval [qi−1 , qi ] to a (positively oriented) subinterval [y0 , yi ],
where y0 is the basepoint in Y . For each i, 1 ≤ i ≤ f , we set vi ∈ Q to be the
element that maps y0 to yi .
By construction, for large enough index n, hn (vi ) ∈ F Sk , for every i, 1 ≤
i ≤ f . The elements in Q, that restrict to the elements that are associated with
the interval exchange transformation, i.e., the elements that map subintervals to
their companions (after gluing the intervals I1 , . . . , Ig and obtaining a finite tree),
generate the surface group Q. Each of these elements is in the subgroup that is
generated by v1 , . . . , vf . Hence, v1 , . . . , vf generate Q.
For each j, 1 ≤ j ≤ r, the segment [y0 , sj (y0 )] is positively oriented, and sj
can be presented as a word in the elements v1 , . . . , vf , sj = wj (v1 , . . . , vf ). The
words wj need not be positive words in the elements v1 , . . . , vf . For each index j,
wj (v1 , . . . , vf ) represents a finite path in the tree Y , that starts at y0 and ends at
sj (y0 ). We denote this path pwj . Since the words wj need not be positive words,
pwj need not be positively oriented.
Each path pwj is supported on a finite subtree of the real tree Y that we denote
Twj . We view Twj as a combinatorial tree and not as a metric tree. To Twj we add
a finite collection of vertices:
(1) a vertex for the base point of Y (the initial point of pwj ), and for the point
sj (y0 ) (the terminal point of pwj ).
(2) a vertex for each root and each branching point in Twj .
(3) each segment in Twj , being a subsegment of Y , can be divided into finitely
many segments with either positive or negative orientation. If there exists
an edge in Twj which is not oriented (i.e., the edge can be cut into subsegments with both positive and negative orientations), then the word wj can
be strictly shortened as a word in v1 , . . . , vf , and still represents sj ∈ Q.
Therefore, we can assume that every edge in Twj is oriented.
(3) pwj starts at the base point in Y , and ends in sj (y0 ). While moving along
pwj , at certain segments of the path the distance to the base point y0 ∈ Y
9
increases, and at other segments it decreases. At each point which is the
boundary of such (increasing or decreasing) segments along pwj (i.e., points
in which the distance to y0 changes from increasing to decreasing or vice
versa), we add a vertex to Twj .
In the sequel we denote the paths pwj that are associated with the words wj on
Twj .
At this point we apply a sequence of Dehn twists on the generators of the interval
exchange (which is the output of the Rips machine from the action of the surface
group Q on the real tree Y ). We perform the sequence of Dehn twists from the
positive side of the interval on which the IET is based. Hence, the sequence of
IETs that we obtain are supported on a decreasing sequence of subintervals of the
original interval I that supports the original IET. All these supporting subintervals
share the endpoint y0 which is the base point in Y . Since the action of Q on Y was
assumed to be a minimal IET action, the lengths of the supporting intervals has to
approach 0.
With each of the obtained IETs we associate a positive generating set for a subsemigroup of Q, in a similar way to the association of v1 , . . . , vf with the original
IET. We denote the corresponding set of generators v1t , . . . , vft (note that the number of generators of the corresponding semigroups can only decrease, so by omitting
finitely many of them, we can assume that their number is fixed).
Lemma 2.3. For every index t ≥ 1, each of the elements v1 , . . . , vf is contained
in the subsemigroup that is generated by: v1t , . . . , vft .
Proof: By the definition of a Dehn twists that is performed from the positive side
of the supporting interval, for every index t ≥ 1, each of the elements v1t+1 , . . . , vft+1
can be written as a (concrete) positive word in the elements v1t , . . . , vft . Hence, the
claim of the lemma follows by induction.
By lemma 2.3 each of the elements v1 , . . . , vf can be written as a positive word
in the elements v1t , . . . , vft ). By substituting these positive words in the words wj
we obtain presentations of the elements sj , 1 ≤ j ≤ r, in terms of the generating
sets v1t , . . . , vft :
sj = wj (v1 , . . . , vf ) = wjt (v1t , . . . , vft ).
Since the elements v1 , . . . , vf and v1t , . . . , vft are positively oriented, and the elements vj are presented as positive words in the generators v1t , . . . , vft , the paths that
are associated with the words wjt in the tree Y , that we denote pwjt , are identical
to the paths pwj , that are associated with the words wj .
While shortening the lengths of the elements v1t , . . . , vft , the ratios between these
lengths may not be bounded. To prove theorem 2.2 we first need to replace the
elements v1t , . . . , vft by elements that are not longer than them, and for which the
ratios between their lengths are universally bounded. We start by proving the
existence of such generators assuming a global bound on the periodicity of the
images of the given set of generators of the semigroup S, s1 , . . . , sr , under the
sequence of homomorphisms {hn }. In the sequel we prove a general version of the
proposition omitting the bounded periodicity assumption.
Proposition 2.4. Suppose that there exists an integer cp , such that the periodicity
10
of the elements hn (s1 ), . . . , hn (sr ) is bounded by cp for all integers n. i.e., for
every n, and every j, 1 ≤ j ≤ r, hn (sj ) can not be written as a word of the form:
sj = w1 αcp +1 w2 , where w1 , w2 , α ∈ F Sk , and α is not the empty word.
For every t ≥ 1, the elements v1t , . . . , vft ∈ Q can be replaced by elements
ut1 , . . . , utgt ∈ Q, with the following properties:
(1) f ≤ gt ≤ e(f ).
(2) for every index t, the semigroup that is generated by ut1 , . . . , utgt in Q contains the semigroup that is generated by v1t , . . . , vft in Q.
(3) for every t ≥ 1, and for large enough index n, hn (uti ) ∈ F Sk for every index
i, 1 ≤ i ≤ gt ,
(4) the tuples ut1 , . . . , utgt belong to finitely many isomorphism classes (under
the action of Aut(Q)).
(5) because of (2) each of the elements v1t , . . . , vft can be represented as a positive
word in the elements ut1 , . . . , utgt . Substituting these words in the words wjt ,
for each t ≥ 1 and each j, 1 ≤ j ≤ r, sj = zjt (ut1 , . . . , utgt ). With each word
zjt (ut1 , . . . , utgt ) we can associate a path, pzjt in the tree Y , and this path is
identical to the path pwjt , that is identical to pwj .
(6) there exist positive constants d1 , d2 that depend only on f and cp , such
that for every t ≥ 1, there exist a sequence of indices (that depend on t)
1 ≤ i1 < i2 < . . . < ib(t) ≤ gt , such that for every 1 ≤ m1 < m2 ≤ b(t):
d1 · length(utim ) ≤ length(utim ) ≤ d2 · length(utim ).
1
2
1
Furthermore, for every index i, 1 ≤ i ≤ gt for which i 6= im , m = 1, . . . , b(t):
10gt · d2 · length(uti ) ≤ length(uti1 )
(7) For each index t, in the words zjt , 1 ≤ j ≤ r, in a distance bounded by g(t)
(that is bounded by the function e(f )), either before or after the occurrence
of an element uti , for i 6= i1 , . . . , ib(t) , appears one of the elements utim ,
1 ≤ m ≤ b(t).
Proof: To get a new set of elements that satisfy the conclusion of the proposition
we need to modify the standard sequence of Dehn twists that we used to get the
elements v1t , . . . , vft (which are Dehn twists that are performed on the pair of bases
that are adjacent to the vertex at the positive end of the interval on which the
interval exchange transformation is supported).
We start by dividing the generators v1t , . . . , vft into finitely many sets according
to their length. We order the elements from the longest to the shortest. We place a
separator between consecutive sets whenever there is a pair of consecutive elements
(ordered according to length) that satisfy:
t
)
length(vit ) ≥ c1 (f, cp ) · length(vi+
1
where c1 (f, cp ) = 4f cp .
Clearly the elements vit are divided into at most f sets. If there are no separators,
we set gt = f , and uti = vit , i = 1, . . . , f . Suppose that there is a separator. In that
case our goal is to construct a procedure that will iteratively reduce the number
11
of separators. We call the elements that are in the first (longest) set long and the
elements in all the other (shorter) sets short.
At this point we modify the standard Dehn twists that are performed on an
interval exchange basis, so that the performed Dehn twists do not change short
bases (short elements can be modified in a controlled way). Let I be the interval
that supports the IET transformation, and let pv be the vertex at the positive end
of I. On the given IET transformation we perform the following operations:
(1) Suppose that the two bases that are adjacent to pv are long, i.e., each of
them contains a long element. Let b1 be the longer base that is adjacent
to pv. Suppose that the other end of b1 is not covered by a long base. In
this case we cut b1 at the end of the last long base that completely overlaps
with b1 and perform Makanin’s entire transf ormation over the base b1
(see [Ca-Ka] for Makanin’s entire transformation). The length of what is
left from b1 after the entire transformation is at most f times the length of
the longest short element. For the rest of this part, before we change the
set of long elements, this part of b1 is declared to be a short base. In this
case we reduced the number of pairs of long bases by 1 and added a pair of
short bases.
By performing an entire transformation over the base b1 we transfered
several long bases (and elements), and several short bases (and elements)
that now replace the base b̂1 that was previously paired with the base b1 .
In the original interval exchange, that is based on the entire interval I,
every point is covered exactly twice. Hence, the subinterval that supports
the base b̂1 supports another sequence of (possibly) long and short bases
(and elements). The endpoints of these bases is now used together with the
endpoints of bases that were transfered using the entire transformation to
define the generators of a new semigroup, that contains the semigroup that
is associated with the original IET that is supported on I.
The generators of the semigroup that is associated with the new IET
contains some generators of the previous IET (that is based on the entire
interval I), and new generators that are located between endpoints of bases
that were transformed over the base b1 and bases that overlapped with b̂1
in the IET that we started with. We divide the new elements according to
the following rules.
Suppose that a previously long element (generator) is cut into finitely
many new elements by the new endpoints of bases. If the length of all
the fractions of the previously long element are at least c1 (f, cp ) times the
maximal length of a short element we consider all the new fractions to be
c (f,c )
long. Otherwise, the length of at least one of the fractions is at least 1 3 p
times the maximal length of a short element. If there are fractions that are
of length at least c1 (f, cp ) times the maximal length of a short element we
declare them to be long and the rest to be secondary short elements. If
c (f,c )
there are no such fractions, we declare the fractions of length at least 1 3 p
times the length of the maximal short element to be short and the rest to
be secondary short.
Suppose that a previously short element (generator) is cut into finitely
many new elements by the new endpoints of bases. The length of at least
one of the fractions is at least f1 times the previous length of that short
12
(2)
(3)
(4)
(5)
element. We declare the fraction with maximal length short and all the
other fractions secondary short.
Suppose that the other end (the beginning) of b1 is covered by another long
base. In this case we perform Makanin’s entire transformation over the
carrying base b1 , and continue as in part (1). If what left from b1 after
the entire transformation is a long element we continue to the next step,
and what left from b1 remains a long element. If what left is bounded by
c1 (f, cp ) times the length of the maximal short element it is set to be a
secondary short element, and the corresponding base to be short. In this
last case the number of pairs of long bases is reduced by 1 and the number
of pairs of short bases is increased by 1.
Suppose that the bases that are adjacent to pv are a short and a long base.
Suppose that the length of the long base that is adjacent to pv, that we
denote b1 , is bigger than the sum of the lengths of the short bases that are
adjacent to pv and the first long base that is adjacent to them. In that case
we perform Makanin’s entire transformation precisely as in cases (1) and
(2).
Suppose that a long and a short base are adjacent to pv, and that the length
of the long base that is adjacent to pv, that we denote b1 , is smaller than
the sum of the short bases that are adjacent to pv and the first long base
that is adjacent to these short bases. We set b2 to be the first long base
that is adjacent to the short bases that are adjacent to pv.
If b1 and b2 are not a pair of bases we do the following. We first transfer
the short bases that are adjacent to pv using the base b1 , and then perform
an entire transformation over the base b2 precisely as we did in cases (1)
and (2).
Suppose that b1 and b2 are a pair of bases. This implies that b1 (and
b2 ) contains a large periodic word, which is a power of the chain of short
bases that are adjacent to pv. We assumed that the periodicity of the words
hn (sj ) is bounded by cp for all n, and all j, 1 ≤ j ≤ r. Hence, the length
of b1 (and b2 ) is bounded by f · (cp + 1) times the maximal length of a
small element. This implies that there can not be a separator between the
length of an element that is supported by b1 or b2 and the collection of
small elements. Therefore, in this case b1 (and b2 ) can not be long bases, a
contradiction, so in case (4) b1 and b2 can not be a pair of bases (assuming
bounded periodicity).
Suppose that two short bases are adjacent to pv. Let b1 and b2 be the
two long bases that are adjacent to the two sequences of short bases that
are adjacent to pv. b1 and b2 can not be a pair of bases because of our
bounded periodicity assumption, as we argued in part (4). Suppose that
b1 is longer than b2 , and that the sequence of short bases that are adjacent
to pv and b1 is longer than the sequence that is adjacent to pv and b2 . In
that case, we first cut b2 (and the base that is paired with it) into two bases
at the endpoint of b1 . We declare the part of b2 that overlaps with short
bases before it overlaps with b1 to be short. Then we perform an entire
transformation along b1 precisely as we did in cases (1) and (2). In this
case we either reduce the number of long pairs of bases (if the beginning
of b1 overlaps with a short base), or we added a short base and left behind
at least two short bases, so the number of short bases that participate in
13
the next iterations in this part (before we change the set of long bases) is
reduced by at least 1.
(6) With the notation and the assumptions of part (5), suppose that the sequence of short bases that are adjacent to pv and b1 is shorter than the
sequence that is adjacent to pv and b2 . In that case, we first cut b1 (and
the base that is paired with it) into two bases at the endpoint of b2 . We
declare the part of b2 that overlaps with short bases before it overlaps with
b1 to be short. Then we perform an entire transformation along the longer
between the remaining of b1 and b2 , precisely as we did in cases (1) and
(2). As in part (5), in this case we either reduce the number of long pairs
of bases, or we added a short base and left behind at least two short bases,
so the number of short bases is reduced by at least 1.
In step (1) the number of pairs of long bases is reduced by at least one, and
the number of short bases is increased by at most the reduction in the number of
long bases. In step (2) there may be no change in the number of bases, and if
there is a change the number of short bases is increased by at most the reduction
in the number of long bases. The outcome of steps (3) and (4) in terms of the
number of bases is identical to that of steps (1) and (2). The outcome of steps
(5) and (6) on the number of bases is similar to that of steps (1) and (2) with an
additional increase of the number of short bases by 1, but an additional reduction
of the number of active short bases (i.e., the number of bases that take part in the
next steps of the iterative procedure) by at least 1.
Hence, when we run this procedure until there are no long elements. i.e., until
all the elements are either short or secondary short, the number of bases can grow
to at most 2f , and therefore the number of elements can grow to at most 4f .
Along the procedure whenever a long base is cut into a finite collection of (only)
short and secondary short elements, then the short elements have a length which
c (f,c )
is bounded below by 1 2f p times the maximal length of a previous short element.
When a short base is cut into a collection of short and secondary short elements,
then the length of the short (and not necessarily the secondary short) fractions is
1
bounded below by 4f
times the length of the original short base. Furthermore, the
obtained short fractions can be effected only by steps (5) and (6) before the set of
long elements is changed, and these steps do not effect these short fractions.
When the procedure reached the state in which there are no long elements, we
change the place of the separators between the elements. Once again we divide the
new collection of elements into finitely many sets according to their length. We
order the short and secondary short elements from the longest to the shortest. We
place a separator between consecutive sets whenever there is a pair of consecutive
elements (ordered according to length) that satisfy:
t
length(v̂it ) ≥ c2 (f, cp ) · length(v̂i+
)
1
where c2 (f, cp ) = cp (4f )2 .
The new elements v̂it may be divided (by the separators) into a larger number
of sets than the previous separated sets. However, the number of such sets that
contain at least one short element (and not only secondary short elements) is at
most the previous number of (separated) sets minus 1, and in particular is bounded
by f − 1. Also, by the construction of the procedure, the (separated) set that
14
contains the longest elements must contain a short element and not only secondary
short elements.
If there is only one separated set that contains a short element (which must be
the separated set of longest elements), we reached a terminal state of the iterative
procedure. Otherwise, if there is more than one such set, we declare a new collection
of long elements. We declare the set of long elements to be all the elements in the
separated set with the longest elements, and the elements in all the next (according
to length of elements) separated sets, until the next separating set that contains a
short element (where this last separated set is excluded).
We continue itartively. After each round of iterations the number of separated
sets that contain short elements reduces by at least one, so after at most f rounds
we end up with a unique set, the one that contains the longest elements, that
contains short elements. All the other separated sets contain only secondary short
elements.
Since the number of bases can multiply by at most 4 in each iteration, the number
of bases when the procedure terminates is bounded above by 4f · f . In particular,
the number of elements in the set that contains the longest elements (the only set
that contains short and not only secondary short elements) is bounded by 4f · f .
The ratios between consecutive elements in this set is bounded above by cp ·(4f ·f )f .
Hence, the ratio between the longest and the shortest element in this set is bounded
f
above by {cp · (4f · f )f }(4 ·f ) .
We set the elements that the iterative procedure ends up with, to be the generators ut1 , . . . , utg(t) . g(t), the number of elements is bounded above by e(f ) = 4f · f .
We set the elements in the separated set with the longest elements, to be the elements ut1 , . . . , utb(t) , and the ratios between their lengths is bounded above by
f
{cp · (4f · f )f }(4 ·f ) . The elements that are not in the separated set with the longest
elements, are all secondary short elements. By construction, in representing the
original elements vit , in a distance bounded by g(t), hence by e(f ), to the appearance of a secondary short element appears one of the short elements, so one of the
elements, uti1 , . . . , utib(t) . Therefore, part (6) of the proposition holds.
The next proposition is the key in our proof of theorem 2.2. Since we proved
proposition 2.4 under a bounded periodicity assumption, we state and prove it
under the same (bounded periodicity) assumption (and generalize it in the sequel).
It implies that the words zjt (in part (5) of the statement of proposition 2.4) can be
replaced by words ẑjt with uniformly bounded cancellation.
Proposition 2.5. Suppose that there exists an integer cp , such that the periodicity
of the elements hn (s1 ), . . . , hn (sr ) is bounded by cp for all integers n.
With the notation of proposition 2.4, there exists a constant C > 0, so that for
every index t ≥ 1 the words zjt , 1 ≤ j ≤ r, can be replaced by words: ẑjt with the
following properties:
(1) As elements in Q: ẑjt (ut1 , . . . , utgt ) = zjt (ut1 , . . . , utgt ).
(2) ẑjt is obtained from zjt by eliminating distinct pairs of subwords. Each pair
of eliminated subwords corresponds to two subpaths of the path pzjt that lie
over the same segment in the tree Twj , where the two subpaths have opposite
orientations.
(3) With the word ẑjt (ut1 , . . . , utgt ) we can naturally associate a path in the tree
15
Y , that we denote, pẑjt . The path pẑjt can be naturally divided into subsegments according to the appearances of the subwords uti in the word ẑjt .
Let DBẑjt be the number of such subsegments that are associated with
subwords uti in pẑjt , that at least part of them is covered more than once by
the path pẑjt . Then for every t ≥ 1 and every j, 1 ≤ j ≤ r, DBẑjt ≤ C.
Proof: Suppose that such a constant C does not exist. Then for every positive
integer m, there exists an index jm , 1 ≤ jm ≤ r, and an index tm > 1, so that for
m
that satisfy parts (1) and (2), part (3) is false
every possible choice of words ẑjtm
for the constant C = m.
By passing to a subsequence, we may assume that jm is fixed, and we denote it
j. We can further assume that for the subsequence the integer gtm , that counts the
number of generators of the semigroup that is constructed in step tm according to
the procedure that is described in proposition 2.4 is fixed and we denote it g. We
may also assume that the integer b(tm ) and the sequences of indices i1 , . . . , ib(tm ) ,
that are associated with the sets of generators of the semigroups that are constructed according to proposition 2.4 along the subsequence {tm } are fixed, and we
denote them b and i1 , . . . , ib .
The tree Twj has finitely many edges. By traveling along the path pwj we pair
subsegments of the path that cover the same edge in Twj with opposite orientations.
We start from the basepoint in Twj . Given an edge in Twj that is covered more
than once by the path pwj , we pair the first and the second subsegments of pwj
that pass through this edge. Note that since Twj is a tree, these two subsegments
have to be of opposite orientations. If pwj passes more than 3 times through an
edge in Twj , we further pair the 3rd and 4th subsegments of pwj that pass through
such an edge and so on.
The number of pairs of subsegments of pwj that we obtained depends only on
the original tree Twj , and the path pwj . The paths pzjt are the same as the path
pwj (as paths in the real tree Y ) for every index t. Our goal is to show that there
exists a subsequence of the given sequence, so that for every m in the subsequence,
the subwords of zjtm that are associated with subsegments of the path pz tm that
j
were paired together, can be replaced by eliminating distinct pair of subwords (part
(2) in the statement of the proposition), to subwords of uniformly bounded word
length, and so that by eliminating these distinct pairs the element zjtm is replaced
by an element ẑjtm that represents the same element in the surface group Q and
satisfies part (3) of the proposition for some constant C.
Let e be an edge in Twj , and let p1 and p2 be subpaths of pwj that were paired
together and are supported on e. Let pt1m and pt2m be the corresponding subpaths of
pz tm . By construction, the word lengths of the paths pt1m and pt2m is not bounded.
j
For each index m, we set scm to be the maximal length of the elements uti1m , . . . , utibm
(note that scm approaches 0 when m grows to infinity). We set Ym to be the real
tree Y equipped with the metric that is obtained from the metric on Y by dividing
it by scm . Q acts isometrically on Y , so it naturally acts isometrically on Ym . By
proposition 2.4 the length of each of the elements uti1m , . . . , utibm is bounded above
by 1, and below by d1 (the constant d1 is defined in part (6) of proposition 2.4).
From the actions of Q on the trees Ym it is possible to extract a subsequence
(still denoted {tm }) that converges into a (non-trivial) action of Q on some real
16
tree Y∞ . Q acts faithfully on Y∞ and the length of each of the elements ui1 , . . . , uib
on Y∞ is bounded below by d1 and above by 1. Since the action of Q on Y∞ is
faithful, and Q is a surface group, the action of Q on Y∞ contains only IET and
discrete components. Since we assumed a global bound on the periodicity of the
images: hn (sj ) for every n, and every j, 1 ≤ j ≤ r, and the elements s1 , . . . , sr
generate Q, the action of Q on the limit tree Y∞ contains no discrete components.
In particular the stabilizer of every non-degenerate segment in Y∞ has to be trivial.
Let e be an edge in Twj with subpaths p1 and p2 that were paired together,
and let pt1m and pt2m be the corresponding subpaths of pẑ tm . As the lengths of
j
the elements uti1m , . . . , utibm (when acting on the tree Y ) approaches 0 with m, the
combinatorial lengths of the paths pt1m and pt2m grows to ∞ with m.
Given the appearance of a generator utism , s = 1, . . . , b in pt1m , and the appearance
of a generator (utiŝm )−1 in pt2m that overlap in a non-degenerate segment, we define
their dual position to be the subsegment in which they overlap, and its image in the
segments that are associated with both utism and utiŝm . Although the combinatorial
lengths of the paths pt1m and pt2m grows to infinity with m, the number of possible
dual positions between the appearances of the various generators remain bounded.
Lemma 2.6. With the assumptions of proposition 2.5 (in particular, the bounded
periodicity of the images hn (sj )), there exists a global bound R, such that for every
index m (from the chosen convergent subsequence), the number of dual positions of
overlapping subwords utism in pt1m and (utiŝm )−1 in pt2m (s, ŝ = 1, . . . , b) is bounded
by R.
Proof: Suppose that the number of dual positions is not universally bounded.
Then there exists a subsequence (still denoted {tm }), pair of indices is and iŝ ,
s, ŝ = 1, . . . , b, such that the number of dual positions of utism in pt1m and (utiŝm )−1
in pt2m is bigger than m.
In that case, a simple piegon hole argument implies that in the action of utism on
Ym , there exists a subinterval Jm ⊂ [ym , utism (ym )] (where ym is the base point of
Ym ) so that:
(i) the length of Jm satisfies: 1 ≥ length(Jm ) > ǫ0 > 0 for every m.
(ii) there exists some non-trivial element qm ∈ Q, for which qm (Jm ) overlaps
with Jm in an interval of length that is at least (1 − δm ) · length(Jm ), and
the sequence δm approaches 0 with m.
Parts (i) and (ii) clearly imply that the periodicity of the elements hn (sj ) can
not be globally bounded, a contradiction to the assumptions of propositions 2.4
and 2.5.
tm
Lemma 2.6 proves that in the chosen subsequence of words, zj , the number of
possible dual positions between the appearances of the possible pairs of generators,
utism and (utiŝm )−1 , in overlapping paths, pt1m and pt2m , remain bounded. Given a
word zjtm , its associated path pz tm , and two of its overlapping subpaths pt1m and
j
pt2m , we use the bound on the number of dual positions to replace zjtm with a shorter
word z̃jtm , so that:
).
) = z̃jtm (ut1m , . . . , utgm
zjtm (ut1m , . . . , utgm
tm
tm
17
When two appearances of a pair utism in pt1m and (utiŝm )−1 in pt2m belong to the
same dual position, we can trim the paths pt1m and pt2m , by erasing the (identical)
subpaths between these two appearances from pt1m and pt2m . We set z̃jtm to be the
word that is obtained from zjtm after erasing the corresponding identical subpaths.
We further repeat this erasing for all the appearances of repeating pairs in the
same dual position in all the (paired) overlapping subpaths pt1m and pt2m along the
path pz tm . We denote the words that are obtained after this erasing, ẑjtm . Since
j
by lemma 2.6 there are at most R dual positions for every pair of generators, and
there are at most (2f )2 pairs of generators, overlapping subpaths in the path pẑ tm
j
have combinatorial length bounded by R · (2f )2 . As the number of overlapping
subpaths in pẑ tm is bounded by the number of overlapping subpaths in pz tm , which
j
j
is identical to the number of overlapping subpaths in pwj , the total combinatorial
length of the overlapping subpaths in pẑ tm is universally bounded. Therefore, the
j
words ẑjtm satisfy the conclusions of proposition 2.5.
belong
to
finitely
many
By part (3) of proposition 2.4 the tuples ut1m , . . . , utgm
tm
isomorphism classes. By proposition 2.5 the total combinatorial lengths of the
overlapping subpaths in the paths pẑ tm are universally bounded. Hence, we can
j
pass to a further subsequence (that we still denote {tm }) for which the isomorphism
is identical. In particular gtm = g is fixed, and the
class of the tuples ut1m , . . . , utgm
tm
overlapping subpaths in the paths pẑ tm represent the same words in the generators
j
(in correspondence).
ut1m , . . . , utgm
tm
For a fixed index tm0 and each index tm from this subsequence we set:
t
ŝtjm = ẑjm0 (ut1m , . . . , utgm ).
Since the tuples ut1m , . . . , utjm belong to the same isomorphism class, the tuples
ŝt1m , . . . , ŝtrm belong to the same isomorphism class as the tuple s1 , . . . , sr as sj =
t
ŝjm0 . Since the overlapping subpaths in the paths pẑ tm represent the same words in
j
the generators ut1m , . . . , utgm the paths [y0 , ŝtjm ] are all positively oriented. Since the
lengths of the elements utim approaches 0 when m grows to ∞, so are the lengths
of the elements ŝtjm . Therefore, if we denote the automorphism that maps the
tuple s1 , . . . , sr to ŝt1m , . . . , ŝtrm by ϕm , then a subsequence of the automorphisms
{varphim } that we denote {ϕℓ } satisfy the conclusions of theorem 2.2.
So far we proved theorem 2.2 in case in the reduction of the action of the surface
group Q on the real tree Y to an interval exchange transformation on a (positive)
interval, all the endpoints of the intervals that define the IET transformation belong
to the same orbit under the action of Q, and the periodicity of the images: hn (sj )
is globally bounded.
Suppose that the periodicity of the images hn (sj ), for every n and j, 1 ≤ j ≤ r,
is globally bounded, and not all the endpoints of the intervals belong to the same
orbit. First, by possibly cutting one of the intervals into two intervals, we can
always assume that one of the endpoints is in the orbit of the base point y0 of the
real tree Y , so we may assume that the interval on which the intervals exchange is
supported is a positively oriented interval, that starts in the base point y0 (as we
did in case all the endpoints were in the same orbit).
18
With the interval exchange (on a positively oriented) interval that starts at the
base point y0 , we associate the elements v1 , . . . , vf , that map an endpoint of a
subinterval to an endpoint of the next subinterval. Since the endpoints of the
subintervals belong to more than one orbit, the elements v1 , . . . , vf belong in a
nontrivial free product Q̂ = Q̃∗ < e1 , . . . , ec > where Q̃ is a surface group that
contains Q as a subgroup of finite index (i.e., Q̃ is associated with a surface that
finitely covers the surface that is associated with Q), and < e1 , . . . , ec > is a free
group. The statements and the proofs of lemma 2.3 and proposition 2.4, that do
not use any properties of the group that is generated by v1 , . . . , vf remain valid
in the general case of more than one orbit of endpoints of subintervals. To prove
proposition 2.5 we applied lemma 2.6. The proof of lemma 2.6 can be modified in
a straightforward way to include the more general case of more than one orbit of
endpoints of segments. This concludes the proof of theorem 2.2 under the bounded
periodicity assumption.
To prove theorem 2.2 omitting the bounded periodicity assumption, we use the
same strategy of proof, but modify the statements and the arguments to include
long periodic subwords, or in the limit, to include non-degenerate segments with
non-trivial stabilizers (that are contained in the discrete or simplicial part of the
limit action). We start with a generalization of proposition 2.4.
Recall that we started the proof of theorem 2.2 with an iterative process of Dehn
twists. v1t , . . . , vft are the (positive) elements between consecutive initial and final
points of the bases that generate the IET after t Dehn twists iterations. The aim
of proposition 2.4 was to replace these generators by a new (possibly larger set of)
generators so that the ratios between their lengths is globally bounded.
Proposition 2.7. For every t ≥ 1, the elements v1t , . . . , vft ∈ Q can be replaced by
elements ut1 , . . . , utgt ∈ Q that satisfy properties (1)-(5) in proposition 2.4. Properties (6) and (7) in proposition 2.4 are replaced by the following properties:
(6) there exist a real number d2 > 1 and a subset of indices (that depend on t)
1 ≤ i1 < i2 < . . . < ib(t) ≤ gt , such that for every index i, 1 ≤ i ≤ gt for
which i 6= im , m = 1, . . . , b(t):
10gt · d2 · length(uti ) ≤ length(uti1 )
(7) there exists an integer ℓ(t), 0 ≤ ℓ(t) ≤ b(t), and a positive real number d1
such that for every ℓ(t) + 1 ≤ m1 < m2 ≤ b(t):
d1 · length(utim ) ≤ length(utim ) ≤ d2 · length(utim )
1
2
1
For every m1 ≤ ℓ(t) and ℓ(t) + 1 ≤ m2 ≤ b(t):
d1 · length(utim ) ≤ length(utim )
2
1
(8) For each t, and every index m, 1 ≤ m ≤ ℓ(t), there exist distinct indices
1 ≤ j1 , . . . , jem (t) ≤ g(t) (that depend on t) that do not belong to the set
pm
i1 , . . . , ib(t) , such that: wm = utj1 . . . utjem (t) , and utm = αwm
where α is a
suffix of wm .
(9) for each index t, in the words zjt , 1 ≤ j ≤ r, in a bounded distance (where the
bound depends only on f ) either before or after the occurrence of an element
uti , for i 6= i1 , . . . , ib(t) , appears one of the elements utim , 1 ≤ m ≤ b(t).
19
Proof: We use the same iterative procedure that was used to prove proposition 2.4.
We divide the elements v1t , . . . , vft into finitely many sets according to their length.
We order the elements from the longest to the shortest. We place a separator
between consecutive sets whenever there is a pair of consecutive elements (ordered
t
according to length) that satisfy: length(vit ) ≥ c1 (f ) · length(vi+
) where c1 (f ) =
1
4f . Note that the constant c1 (f ) in the current proposition depends only on the
number of generators f , and not on the periodicity bound, that we didn’t assume
exists.
We use the same iterative procedure that we used in the proof of proposition 2.4,
and we refer to its various cases and notions as it appears in the proof of proposition
2.4. In cases (1)-(3) of this procedure we proceed precisely as in proposition 2.4
(the bounded periodicity case).
In case (4) of the procedure, we assumed that a long and a short base are adjacent
to pv, the terminal point of the interval that supports the IET, and that the length
of the long base that is adjacent to pv, that we denote b1 , is smaller than the sum
of the short bases that are adjacent to pv and the first long base that is adjacent to
these short bases. We set b2 to be the first long base that is adjacent to the short
bases that are adjacent to pv.
If b1 and b2 are not a pair of bases we do what we did in proposition 2.4 (essentially what we did in cases (1) and (2)). Suppose that b1 and b2 are a pair of
bases. In this case b1 and b2 are a pair of long bases, such that b1 is obtained by
a shift of b2 by a short word. Hence, b1 (and b2 ) can be presented as a short word
times a high power of another short word, precisely as described in case (8). In
this case we leave the bases b1 and b2 as they are, and continue analyzing the rest
of the IET transformation, i.e., the bases that are supported on the interval which
the complement of the union of b1 and b2 . The number of pairs of long bases is
reduced by 1. The number of active short bases, i.e., the number of short bases
that participates in the rest of the procedure is reduced by at least 1.
In case (5) let b1 and b2 be the two long bases that are adjacent to the two
sequences of short bases that are adjacent to pv. If b1 and b2 is not a pair of bases
we do what we did in the proof of proposition 2.4. If b1 and b2 are a pair of bases
we do what we did in case (4) when b1 and b2 are a pair of bases. Once again the
number of pairs of long bases is reduced by 1, and the number of active short bases
is reduced by at least 2.
In case (6) we act precsiely as we did in the proof of proposition 2.4. As in
proposition 2.4, when we run this procedure until there are no long elements. i.e.,
until all the elements are either short, secondary short, and a product of a short
and a high power of short elements (part (8) of the proposition), the number of
bases can grow to at most 2f , and therefore the number of elements can grow to
at most 4f .
When the procedure reached the state in which there are no long elements that
do not fall into the description in part (8) of the proposition, we change the place
of the separators between the elements. Once again we divide the new collection
of elements into finitely many sets according to their length. We order the short
and secondary short elements from the longest to the shortest (note that the elements that satisfy part (8) of the proposition, the semi-periodic elements, are not
contained in the set that we order). We place a separator between consecutive sets
whenever there is a pair of consecutive elements (ordered according to length) that
20
satisfy:
t
length(v̂it ) ≥ c2 (f ) · length(v̂i+
)
1
where c2 (f ) = (4f )2 .
If there is only one separated set that contains a short element, we reached a
terminal state of the iterative procedure. In that case if there is a semi-periodic
element (an element that satisfy part (8)), and its period contains a small element,
we further perform Dehn twist and shorten the semi-periodic element, so that it
becomes small as well. If its period does not contain small elements (only secondary
small), we do not change it. Note that such a semi-periodic element can be much
larger than the small elements.
Otherwise, if there is more than one such set, we declare a new collection of long
elements. We declare the set of long elements to be all the elements that satisfy part
(8) of the proposition (i.e., previously long semiperiodic elements), and the elements
in the separated set with the longest short elements, and the elements in all the
next (according to length of elements) separated sets, until the next separating set
that contains a short element (where this last separated set is excluded).
We continue itartively. After each round of iterations the number of separated
sets that contain short elements reduces by at least one, so after at most f rounds
we end up with a unique separated set that contains short elements. All the other
separated sets contain only secondary short elements.
The conclusions of the proposition follow from the termination of the iterative
procedure, precisely as in the proof of proposition 2.4.
Proposition 2.7 replaces proposition 2.4 in the general case (i.e., when there is
no periodicity assumption). To obtain the same conclusions as in proposition 2.5,
we further modify the tuples, ut1 , . . . , utgt .
Proposition 2.8. With the notation of proposition 2.4, it is possible to further
modify the tuple of elements ut1 , . . . , utgt , by performing Dehn twists on some of the
semi-periodic elements (the elements ut1 , . . . , utℓ(t) that satisfy part (8) in proposition
2.7), so that there exists a constant C > 0, for which for the modified tuples, that
we still denote: ut1 , . . . , utgt , for every index t ≥ 1 the words zjt , 1 ≤ j ≤ r, can be
replaced by words: ẑjt that satisfy properties (1)-(3) in proposition 2.5.
Proof: Suppose that such a constant C does not exist (for any possible application
of Dehn twists on the semiperiodic elements in the tuples: ut1 , . . . , utgt ). Then for
every positive integer m, there exists an index tm > 1, so that for every possible
choice of (applications of) Dehn twists to the semiperiodic elements in the tuple,
ut1m , . . . , utgm
, at least one of the words ẑjtm that satisfy parts (1) and (2) (in
tm
proposition 2.5), part (3) is false for the constant C = m.
By passing to a subsequence, we may assume that for the subsequence the integer gtm , that counts the number of generators of the semigroup that is constructed
in step tm according to the procedure that is described in proposition 2.7 is fixed
and we denote it g. We may also assume that the integers ℓ(tm ) and b(tm ) and
the sequences of indices i1 , . . . , ib(tm ) , that are associated with the sets of generators of the semigroups that are constructed according to proposition 2.7 along the
subsequence {tm } are fixed, and we denote them ℓ, b and i1 , . . . , ib .
21
For each index tm , we denote by lengthm , the minimal length of a long element
(i.e., the elements ut1m , . . . , utbm . For each semi-periodic element ut1m , . . . , utℓm we
i
denote the length of its period by lperm
.
lperi
For each index m, and every i, 1 ≤ i ≤ ℓ, we look at the ratios: lengthmm . We
can pass to a subsequence of the indices m, for which (up to a change of order
lperi
of indices): 0 < ǫ < lengthmm for some positive ǫ > 0, and i = 1, . . . , ℓ′ . And for
lperi
every i, ℓ′ < i ≤ ℓ, the ratios lengthmm approaches 0. We perform Dehn twists along
the semiperiodic elements ut1m , . . . , utℓm
(from the subsequence of indices m), so
′
that all these semiperiodic elements have lengths bounded by a constant times the
length of a long element. These elements will be treated as long elements and not
as semiperiodic elements in the sequel.
First, suppose that ℓ′ = ℓ, i.e., that there exists an ǫ > 0, such that for every
lperi
i, 1 ≤ i ≤ ℓ, 0 < ǫ < lengthmm . In that case, after applying Dehn twists to the
semiperiodic elements, all the elements ut1m , . . . , utgm are either long or secondary
short. By the argument that was used to prove proposition 2.5, either:
(i) the number of dual positions of the different elements is globally bounded
(for the entire subsequence {tm }).
(ii) there exists a subsequence (still denoted {tm }), and a fixed positive word
in a positive number of long elements and possibly some secondary short
elements, which is a periodic word, and ratio between the length of the
period and the length of the element that is represented by the positive
word approaches 0.
Part (ii) implies that either the surface group Q contains a free abelian group of
rank at least 2, or that Q is freely decomposable, and we get a contradiction.
Hence, in case ℓ = ℓ′ there is a global bound on the number of dual positions of
the different elements for the entire subsequence {tm }, and the conclusion of the
proposition follows by the same argument that was used to prove proposition 2.5.
The same argument remains valid if ℓ′ < ℓ but the lengths of the semiperiodic
elements ut1m , . . . , utℓm can be bounded by a constant times the length of a long
element.
Suppose that there exists a subsequence of indices (still denoted {tm }), for which
along a paired subpaths p1 and p2 , at least one of the appearances of a semiperiodic
element utim , ℓ < i ≤ ℓ (along p1 or p2 ), overlaps with an unbounded number of
elements (along p2 or p1 in correspondence). In that case, like part (ii) in the case in
which the lengths of the semiperiodic elements are bounded by a (global) constant
times the length of a long element, we get:
(ii’) there exists a subsequence (still denoted {tm }), and a fixed positive word
in a positive number of either long elements or semiperiodic elements and
possibly some secondary short elements, which is a periodic word, and ratio
between the length of the period and the length of the element that is
represented by the positive word approaches 0.
Hence, either the surface group Q contains a free abelian group of rank at least
2, or Q is freely decomposable, and we get a contradiction. Therefore, there exists a
global bound on the number of elements that overlaps with a semiperiodic element
that appears along a paired subpaths p1 and p2 . In this last case, once again either
part (i) or part (ii’) holds, and if part (ii’) holds we get a contradiction.
22
Finally, in all the cases we obtained a global bound on the number of dual
positions of the different elements, so the proposition follows by the same argument
that was used to prove proposition 2.5.
Given propositions 2.7 and 2.8, that generalize propositions 2.4 and 2.5, the rest
of the proof of theorem 2.2 follows precisely as in the bounded periodicity case.
So far we studied axial and IET components in a limit action of pair on a real
tree. The next theorem constructs a positive free basis in case the ambient group
is free (non-abelian), and its action on the limiting tree is free and is either discrete
or non-geometric. The remaining case, which is the Levitt action of a free group
will be analyzed in the sequel. The existence of such a positive free basis is crucial
in our general approach to the structure of varieties over a free semigroup.
Theorem 2.9. Let F be a free group, and let {hn : (S, F ) → (F Sk , Fk ) be a
sequence of homomorphisms of pairs that converges into a free action of the limit
pair (S, F ) on a real tree Y . Suppose that the action of F on the limit tree Y does
not contain any Levitt components.
Then there exists a directed finite graph Θ with the following properties:
(1) Θ contains a base point, and the free group F is identified with the fundamental group of the graph Θ. With each (positively oriented) edge in Θ we
associate a label, and each of the given set of generators of the subsemigroup
S corresponds to a positive path in Θ, that starts and ends at the base point,
and can be expressed by a positive word in the labels that are associated with
its edges.
(2) there exists an index n0 such that for every index n > n0 the homomorphisms {hn } are encoded by the graph Θ. i.e., each of these homomorphisms is obtained by substituting elements from the free semigroup F Sk to
the elements that are associated with the various positive edges in Θ (these
substitutions define a homomorphism from the fundamental group F of Θ
to the coefficient free group Fk , and this homomorphism is precisely the
homomorphism from the subsequence).
Proof: If the action of F on the limit tree Y is discrete then the conclusion of the
theorem is immediate. Hence, we may assume that Y contains non-discrete parts
and no Levitt components. In particular, the action is not geometric (Levitt (or
thin) and non-geometric actions on a real tree are defined in [Be-Fe]).
Suppose that the action of F on Y is free and contains only discrete and nongeometric components in the sense of Bestvina-Feighn (i.e., there are no Levitt
components). In that case every resolution (in the sense of [Be-Fe]) of the action
of F on the real tree Y is discrete.
Suppose that F × T → T is such a resolution. In particular, the action of F on
T is free and discrete, and there exists an F -equivariant map from T onto Y . By
the construction of a resolution (see [Be-Fe]), we can assume that the resolution
was constructed so that the (finitely many) segments that connect between the
base point in T and the images (in T ) of the base point under the action of the
given (finite) set of generators of the semigroup S are embedded isometrically into
the limit tree Y . Since the finite union of the orbits of these segments in both
23
trees T and Y cover these trees, and each of these segments is positively oriented,
the (equivariant) orientation of segments in Y lifts to an equivariant orientation of
segments in T .
The action of F on T is free and discrete. Hence, by Bass-Serre theory it is
possible to associate with this action a finite graph of groups with a basepoint.
We denote this graph Θ. Since the action is free, Θ has trivial vertex and edge
stabilizers, and its fundamental group is identified with F . Furthermore, the segments in T are oriented equivariantly, and every edge in Θ is contained in (an orbit
of) a segment that is associated with one of the given generators of the semigroup
S, hence, the orientation of segments in T gives an orientation of the edges in Θ.
Finally, since the segments that connect the base point to the images of the base
point under the action of the given set of generators of S in T are all positively
oriented, the loops in Θ that correspond to elements in the semigroup S are all
positively oriented, so Θ satisfies the properties that are listed in part (1).
With each positive segment in the graph Θ we associate a label. Every substitution of values from the free semigroup F Sk to these labels gives a homomorphism of
pairs from (S, L) to (F Sk , Fk ). Since the action of F on T resolves the action of F
on Y , and the action of F on Y is discrete and free, given a sequence of homomorphisms that converges into the action of F on the tree Y , there is an index n0 , for
which for every n > n0 positively oriented segments in Θ are mapped to elements in
F Sk . Therefore, these homomorphisms are obtained from Θ by substituting these
values to the labels that are associated with positively oriented edges, and we get
part (2).
To conclude this section we need to analyze Levitt components, that play an
essential role in our analysis of homomorphisms of pairs. The analysis of Levitt
components that we use, is similar to our analysis of IET components, as it appears
in the proof of theorem 2.2.
Theorem 2.10. Let F be a f.g. free group, and let (S, F ) be a pair where S is
a f.g. subsemigroup that generates F as a group. Let {hn } be a sequence of pair
homomorphisms of (S, F ) into (F Sk , Fk ) that converges into an indecomposable free
action of F on a real tree Y of Levitt (thin) type (for the notion of an indecomposable
action see definition 1.17 in [Gu]).
Then the conclusions of theorem 2.2 (for the IET case) are valid. There exists
a sequence of automorphisms {ϕℓ ∈ Aut(F )}∞
ℓ=1 with the following properties:
(1) for each index ℓ, there exists nℓ > 0, so that for every n > nℓ , and all the
generators of the subsemigroup S, s1 , . . . , sr , hn ◦ ϕℓ (sj ) ∈ F Sk .
(2) for every n > nℓ :
1
dT (hn ◦ ϕℓ (sj ), id)
<
max1≤j≤r dT (hn (sj ), id.)
ℓ
where T is a Cayley graph of the coefficient free group Fk with respect to a
fixed set of generators.
Proof: The argument that we use is an adaptation to the Levitt case of the argument in the IET case (theorem 2.2). Let s1 , . . . , sr be the generators of the
subsemigroup S that generates the free group F that acts freely on the limit tree
24
Y , and the limit action is of Levitt type. Hence, there exists a finite subtree R
of Y , such that if one applies the Rips machine to the action of the pseudogroup
generated by the restrictions of the actions of the generators s1 , . . . , sr to the finite
tree R, the output is a Levitt type pseudogroup (see [Be-Fe]).
A Levitt type pseudogroup that is based on an oriented interval I, is generated
by finitely many pairs of bases, where the union of the supports of these bases is
a finite union of subintervals of I. Since it is a Levitt type pseudogroup, there are
subintervals in I that are covered only once by the union of bases. Since every
segment in the tree Y can be divided into finitely many oriented segments, and
since a Levitt pseudogroup is mixing in the sense of [Mo], we can assume that the
Levitt pseudogroup is supported on an oriented subinterval I of Y .
Levitt pseudo groups are analyzed in [Be-Fe], [GLP] and in [Ra]. Given a Levitt
pseudogroup that is supported on some (subintervals of an) oriented interval I, there
are some subintervals that are covered exactly once by the bases of the pseudogroup.
Starting with the original f.g. pseudogroup one applies to it a sequence of moves (see
[Be-Fe]). There are finitely many subintervals that are covered only once. In each
move, one cuts such a subinterval that is contained in a base, hence, possibly cut
the base that is supported on this subinterval (if the subinterval does not contain an
endpoint of the base), and cuts a corresponding subinterval from the paired base.
Following section 7 in [Be-Fe], subintervals that are covered exactly once and
removed along the process are divided into 3 classes:
(1) an isolated base, that is a base that its interior cuts no other bases. In this
case the number of pairs of bases after removing the isolated base decreases
by 1.
(2) a semi-isolated base, that is a subinterval that is covered only once, contained in a base and contains an endpoint of that base. Removing such a
subinterval does not increase the number of bases.
(3) a subinterval that is covered only once and is contained in the interior of a
base. In this case the number of pairs of bases after removing the subinterval
increases by 1, and the number of connected components of subintervals that
are covered by bases increases by at least 1.
In addition to removing subintervals that are covered exactly once, one performs
the following operation:
(4) if there exists a subintervals that supports exactly two bases, and the support of both is the entire subinterval, one removes that subinterval and the
bases that it supports. If the bases are paired, they are erased and the
number of pairs of bases reduces by 1. If these bases are not paired, one
further pairs the the two bases that were previously paired with the bases
that were supported on that subinterval.
The Rips’ machine applies moves of types (1) and (4) as long as possible (i.e.,
removes subintervals of type (1) and (4)). Then applies moves of type (2) as long
as possible. If there are no more moves of types (1), (4) and (2), then one applies
a move of type (3). Since we assumed that the action of the given group F on the
tree Y is of Levitt type, and the given pseudogroup generates the action, moves
of type (1) can not occur (since if a subinterval of type (1) exists, the action of
the pseudogroup on the interior of that subinterval can not have a dense orbit).
Furthermore, in all moves of type (4), the two bases that are supported on that
subinterval can not be a pair of bases (by the same argument). By proposition 7.2
25
in [Be-Fe] moves of type (3) occur infinitely many times, i.e., a sequence of moves
of type (4) and (2) has always a finite length. The Rips’ machine does not give a
priority to the order of the subintervals of type (3) that are being treated, as long
as any such subinterval is treated after a finite time.
With the pseudogroup and the interval that supports it we can naturally associate a graph. The vertices in this graph are maximal subintervals that are covered
at least once by bases of the pseudogroup. Each such maximal subinterval starts
and ends with an endpoint of one of the bases that is not contained in the interior
of another base. The edges in the graph are associated with the pairs of bases of the
pseudogroup. For each pair of bases there is an edge that connects between a maximal subinterval that supports a base to a maximal subinterval (possibly the same
one) that supports its paired base. Clearly the Euler characteristic of the graph
that is associated with the initial Levitt pseudogroup is negative and is bounded
below by 1 − b, where b is the number of paired bases.
The graph that is associated with a pseudogroup that is obtained from the
original one after a sequence of moves of types (2)-(4) can have only a bigger
(smaller in absolute value) Euler characteristic. After performing a further finite
sequence of all the possible moves of types (2) and (4) (there are finitely many such
by proposition 7.2 in [Be-Fe]), the valency of each vertex in the associated graph
is at least 3, so at each step of the process in which all the moves of types (2)
and (4) were performed, the number of generators of the obtained pseudogroup is
universally bounded (in terms of the original Euler characteristic).
The sum of the lengths of the bases in the pseudogroups along the process that
starts with a Levitt pseudogroup and applies moves of type (2)-(4) approaches 0
(see proposition 8.12 in [Be-Fe]). Since the number of bases is totally bounded,
the (infinite) intersection of the unions of the subintervals that support these bases
consists of finitely many points.
Lemma 2.11. Let U be a f.g. Levitt pseudogroup that is based on some oriented
interval I that generates a Levitt type action of a free group F on some real tree
Y . Let U1 , . . . be the pseudogroups that are obtained from U by applying the moves
(2)-(4), where in each step we apply a single move of type (3) (if possible) and then
all the possible moves of types (2) and (4). Then for some index t0 and for all
t > t0 :
(1) the Euler characteristics of the graphs that are associated with the pseudogroups Ut are equal to the Euler characteristic of the graph that is associated with Ut0 .
(2) there is no nontrivial word w in the generators of the pseudogroup Ut , that
is defined on a non-degenerate subinterval of I, and acts trivially on that
subinterval.
(3) only moves of types (3) and (4) are applied in the process at each step t.
Proof: For every index t, each of the the operations (2)-(4) do not reduce the
absolute value of the Euler characteristic of the graph that is associated with the
pseudogroup Ut , i.e., these operations can not increase the absolute value of the Euler characteristic. Hence, after finitely many steps the Euler characteristic stabilizes
and part (1) follows.
Suppose that at some step t, there is a nontrivial word w in the generators
of Ut , that is defined on some non-degenerate subsegment of I, and w fixes that
26
subinterval pointwise. In that case there exists a point q ∈ I, that is contained
in the interior of a subinterval J, J ⊂ I, such that the interval J has a periodic
orbit under the action of the pseudogroup Ut . We can further assume that all the
translates of J in that periodic orbit are disjoint.
Suppose that the disjoint translates of J in that periodic orbit are J0 , . . . , Jℓ .
All these subintervals are covered by at least 2 bases from Ut , and for each index i
there exists a pair of bases that map Ji to Ji+1 , i = 0, . . . , ℓ − 1, and Jℓ to J0 . The
subinterval J (or rather the subinterval to which it is going to move using moves of
type (4)) will stay covered by at least two bases along the process, unless at some
step t1 > t, at least one of the subintervals in its orbit will be covered twice by a
base and its paired base, and these two paired bases have the same support (i.e.,
the map from that base to its paired base is the identity map). Since the lengths of
the bases approaches 0 when t → ∞, such a step t1 > t must exist. In this case we
can remove this pair of bases. Since the two paired bases that we remove can not
be supported on a subinterval that does not support another base, the removal of
this pair increases the Euler characteristic of the graph that is associated with the
pseudogroup Ut1 . By part (1) such a reduction can occur only finitely many times
and part (2) follows.
By parts (1) and (2) there is a step t0 such that for every t > t0 the Euler characteristic of the graph that is associated with the generating sets of the pseudogroups
Ut is constant, and there is no non-degenerate subsegment of the interval I that is
fixed pointwise by a non-trivial word w in the pseudogroup Ut . In particular, for
all t > t0 the pseudogroup Ut can not contain a pair of bases that are supported on
the same subinterval of I (as otherwise this pair can be removed and increase the
Euler characteristic, or leave a subinterval of I with trivial (simplicial) dynamics).
By the structure of the process, at step t0 + 1 we start with Ut0 and apply
to it all the possible moves of types (2) and (4) (there are finitely many such by
proposition 7.2 in [Be-Fe]). At this point the endpoints of all the bases in the
obtained pseudogroups are covered at least twice (otherwise a move of type (2) is
still possible).
Now we apply move (3) along a subinterval J1 ⊂ I that is covered only once
by a base b1 , and since it is a move of type (3), J is contained in the interior of
the subinterval that supports the base b1 . The base b1 is paired with a base b2 , so
by move (3) we cut b1 along J1 and b2 along a (paired) subinterval J2 ⊂ I, and
increase the number of paired bases by 1. We denoted the obtained pseudogroup
Ut0 +1 .
The subintervals J1 and J2 have to be disjoint. J1 was covered exactly once
before we erased the corresponding part of b1 , so every point in J2 was covered by
at least two bases before we erased the corresponding part of b2 , since otherwise
there is a subinterval of J2 on which the pseudogroup Ut0 act discretely, and a
Levitt pseudogroup is mixing.
The endpoints of all the bases that are not supported by J2 are covered at least
twice, since they were covered twice before the move of type (3) was applied, and
this move does not effect these endpoints. Suppose that there exists an endpoint of
a base that is supported by J2 , and this endpoint is supported by a single base in
the pseudogroup Ut0 +1 . Since every point in the subinterval J2 is covered at least
once, in case there is a point in J2 (including its endpoints) that is covered exactly
once, there must exist a point in J2 (including its endpoints) that is not contained
in the interior of a base. Hence, the graph that is associated with Ut0 +1 contains at
27
least 2 new vertices of valency at least 2 in addition to the vertices that existed in
the graph that was associated with the pseudogroup before the move of type (3).
Since the number of paired bases was increased only by 1, the Euler characteristic
of Ut0 +1 is bigger by at least 1 from that of Ut0 , and we got a contradiction to part
(1).
Therefore, all the endpoints of bases in Ut0 +1 are covered by at least 2 bases.
Moves of type (4) do not change this property, so no move of type (2) is required
before a move of type (3) is applied. Hence, part (3) follows by a straightforward
induction.
As was proved by Thurston for laminations on surfaces, and further generalized
by Morgan-Shalen to codimension 1 laminations of a 3-manifold, the foliation of a
band complex contains finitely many compact leaves up to isotopy.
Proposition 2.12. Let X be a band complex (see section 5 in [Be-Fe] for the
definition of a band complex). Then X contains finitely many isotopy classes of
periodic leaves.
Proof: See proposition 4.8 in [Be-Fe] and theorem 3.2 in [MS] for the analogous
claim for codimension-1 laminations.
For a Levitt pseudogroup U , that is associated with a free action of a free group
F on a real tree Y , proposition 2.12 gives presentations of the free group F in terms
of the groups that are generated by the generators of the pseudogroup U and the
pseudogroups Ut that are derived from it using the Rips machine.
Lemma 2.13. Let U be a f.g. Levitt pseudogroup that is associated with a free
action of a free group F on a real tree Y . Let U1 , . . . be the pseudogroups that are
derived from it using the Rips machine. Then:
(1) for each index t, the group F that is generated by the generators of Ut can
be presented as the quotient of a free group generated by generators that
correspond to pairs of bases from Ut , divided by a normal subgroup that is
(normally) generated by finitely many elements that are associated with the
periodic leaves in the band complex that is associated with the pseudogroup
Ut .
(2) the number of distinct presentations of the free group F that are associated
with the pseudogroups U1 , . . . is finite.
Proof: Part (1) is true for the initial pseudogroup, by construction. The moves of
the Rips machine do not change that, as can be seen in section 6 in [Be-Fe].
By part (1), one can read a presentation for the free group F from the pseudogroup Ut0 . By proposition 2.12 the pseudogroup Ut0 has finitely many isotopy
classes of periodic orbits. In particular, the length of a (simple) periodic orbit is
bounded. By part (3) of proposition 2.11, the pseudogroups, Ut0 +1 , . . . , are obtained from Ut0 by successive applications of moves of types (3) and (4). A move
of type (4) can only reduce the length of a periodic orbit. A move of type (3) does
not change the lengths of the periodic orbits. Therefore, the lengths of the periodic
orbits in all the band complexes, Ut0 +1 , . . . , are uniformly bounded.
By part (1) the periodic orbits determine the relations in the presentation of
F that is associated with each of the pseudogroups, Ut0 +1 , . . . . Since the num28
ber of generators in these pseudogroups are uniformly bounded, and the lengths
of the periodic orbits is uniformly bounded, hence, the lengths of relations that
appear in the presentations that are associated with these pseudogroups are uniformly bounded, the number of distinct presentations that can be read from the
pseudogroups, Ut0 +1 , . . . , is finite.
Propositions 2.11-2.13 enable one to modify the argument that was used in the
proof of theorem 2.2 (the IET case), to prove theorem 2.10 (the Levitt case).
We started with the pseudogroup U , and applied the Rips machine to obtain
the pseudogroups U1 , . . . . By proposition 2.11 starting from Ut0 , there is no nondegenerate subinterval in I that is stabilized pointwise by a non-trivial word in the
pseudogroup Ut , t > t0 . Furthermore, only moves (3) and (4) are applied along the
Rips machine at all the steps t > t0 .
Each of the pseudogroups Ut is supported on the oriented interval I, or rather on
a finite union of maximal connected subintervals of I. As in the IET case, with the
pseudogroup Ut0 we can associate a semigroup that its generators are elements in
a group that contains the free group F , namely the elements between consecutive
endpoints of bases that are supported on the same maximal connected subinterval
of I. We denote the generators of this semigroup v1t0 , . . . , vft0t .
0
In a similar way we can associate a set of generators of a semigroup with each of
the pseudogroups Ut , t > t0 . We denote these sets of generators: v1t , . . . , vft t . The
conclusion of lemma 2.3 is valid for the semigroups that are associated with Ut for
every t > t0 .
Lemma 2.14. For every pair of indices: t2 > t1 ≥ t0 , each of the elements that
generate the semigroup at step t1 : v1t1 , . . . , vft1t is contained in the semigroup that
1
is constructed at step t2 and is generated by: v1t2 , . . . , vft2t .
2
, are
Proof: By proposition 2.11, for every t > t0 the elements: v1t+1 , . . . , vft+1
t+1
t
t
obtained from the elements: v1 , . . . , vft by a finite (possibly empty) sequence of
moves of type (4) and a move of type (3). A move of type (4) does not change the
semigroup that is generated by the elements: v1t , . . . , vft t . In a move of type (3) a
subinterval J1 is erased from a base b1 , that is covered exactly once along J1 , and a
corresponding subinterval J2 is erased from the base that is paired with b1 , that we
denote b2 . Since each point in the subinterval J2 has to be covered at least twice
before it is erased from b2 , each of the elements: v1t , . . . , vft t can be written as a
positive word in the new elements: v1t+1 , . . . , vft+1
and the lemma follows.
t+1
Furthermore, for all t > t0 , the groups that are generated by the elements:
v1t , . . . , vft t , that strictly contain the free group F , satisfy similar properties to the
ones that are listed in lemma 2.13.
Lemma 2.15. For every t > t0 , let Vt be the group that is generated by the elements: v1t , . . . , vft t . Then:
(1) The groups Vt are all isomorphic and are all free.
(2) From the pseudogroup Ut and the structure of its periodic orbits one can
read a presentation of the free group Vt with generators: v1t , . . . , vft t . The
presentations of the groups Vt that are are obtained in this way belong to
29
finitely many classes, where the presentations in each class are similar. i.e.,
if t1 and t2 are in the same class, then the presentation that is associated
with Vt1 is obtained from that associated with Vt2 by replacing vit2 with vit1 ,
i = 1, . . . , ft1 = ft2 .
Proof:
By lemma 2.14 for every t > t0 and every t2 > t1 ≥ t0 , v1t1 , . . . , vft1t can be
1
written as positive words in v1t2 , . . . , vft2t . In particular, Vt1 < Vt2 . By part (3) of
2
proposition 2.11, for every t > t0 one applies only moves of type (3) and (4). Moves
of type (4) clearly do not change Vt . When one applies move (3) a subinterval J1 is
deleted from the interior of a base b1 , and a disjoint subinterval J2 is deleted from
a base b2 that is paired with b1 . The interval J1 is covered only once, where every
point in the interval J2 is covered by at least two bases before J2 was erased from
b2 .
Hence, from the two parts of the base b1 that are left after deleting J1 no new
generators are added to the new generating set. The two parts of b2 that are left
after deleting J2 can be presented as positive words in the generating set before
the deletion. Therefore, all the new generators can be presented as words (not
necessarily positive) in the previous generating set. So Vt+1 = Vt .
The group F acts on a real tree Y , and the action is indecomposable (in the sense
of [Gu]) and of Levitt type. With the action of F on Y once can associate finitely
many orbits of branching points, and with each orbit of branching point finitely
many orbits of germs. Each of the elements vti starts at a germ of a branching point
and ends at a germ of a branching point.
If all the starting and the ending branching points are in the same orbit under
the action of F , then Vt = F for every t, and Vt is free. Otherwise we divide the
branching points in Y into finitely many orbits under the action of F . If all the
elements, vit , from the generating set of Vt starts and ends at branching points
from the same orbit under the action of F , then all are elements in F and Vt = F .
Hence, we look at those elements vit that starts and ends at vertices from different
orbits under the action of F . If vit maps a germ of its starting point to a germ
of its ending point, then vit preserves the tree Y , i.e., it belongs to the stabilizer
of the same indecomposable component under the action of Vt . By a result of P.
Reynolds [Re], the subgroup that is generated by F and the elements that preserve
the tree Y , that we denote V̂t , generate a free group that contains F as a finite
index subgroup, Y is indecomposable under this action which is of Levitt type.
we look at the finite set of orbits of branching points of Y under the action of V̂t .
Each of the generators vti that is not contained in V̂t , map such an orbit to another
orbit, and no germ of the staring orbit is mapped to a germ of the ending orbit by
vit . We order the orbits of branching points of Y under V̂t . We look at all pairs
(j1 , j2 ), j1 < j2 , for which there is an element, vit , that maps a branching point in
orbit j1 to a branching point in orbit j2 or vice versa.
In [Gu] a graph of groups is associated with an action of a group on a real tree,
that is obtained from the analysis of its indecomposable components. The group
Vt admits an action on a real tree, in which the only indecomposable components
are the orbits of the subtree Y , that is stabilized by the subgroup V̂t . The graph
of groups that is associated with this action has one vertex stabilized by V̂t , and
edges that are associated with the different pairs (j1 , j2 ), j1 < j2 , with which one
can associate a generator vit . Therefore, Vt is free, and we get part (1).
30
As in the proof of theorem 2.2 in the IET case, while shortening the lengths of
the elements v1t , . . . , vft t using the Rips machine (moves (2)-(4)), the ratios between
the lengths of the elements v1t , . . . , vft t may not be bounded. Hence, we need to
replace the elements v1t , . . . , vft t by elements that are not longer than them, and for
which the ratios between their lengths are universally bounded. As in the IET case
(proposition 2.4), We start by proving the existence of such generators assuming
a global bound on the periodicity of the images of the given set of generators of
the semigroup S, s1 , . . . , sr , under the sequence of homomorphisms {hn }, and then
generalize the argument omitting the bounded periodicity assumption. Note that
in order to find a replacement with a global bound on the ratios between elements,
we may need to replace the original sequence v1t , . . . , vft t by another sequence that
satisfies the same properties (and in particular, all the claims 2.11-2.15).
Proposition 2.16 (cf. proposition 2.4). Suppose that there exists an integer
cp , such that the periodicity of the elements hn (s1 ), . . . , hn (sr ) is bounded by cp for
all integers n.
After possibly replacing the sequence of systems of generators, v1t , . . . , vft t , by a
sequence of systems of generators that satisfy claims 2.11-2.15, that we still denote
v1t , . . . , vft t , there exists a subsequence of indices, that for brevity we still denote t,
such that for every index t (from the subsequence), the elements v1t , . . . , vft t can be
replaced by elements ut1 , . . . , utg , with the following properties:
(1) for every index t (from the subsequence), the semigroup that is generated by
ut1 , . . . , utg contains the semigroup that is generated by v1t , . . . , vft t .
(2) for every t, and for large enough index n, hn (uti ) ∈ F Sk for every index i,
1 ≤ i ≤ g.
(3) for every t it is possible to associate naturally a presentation with the generators: ut1 , . . . , utg . The presentations are similar for all t. i.e., for t1 , t2
the presentations are identical if we replace uti1 with uti2 , for i = 1, . . . , g.
(4) by part (1) for every t each of the elements v1t0 , . . . , vft0t can be represented
0
as a positive word in the elements ut1 , . . . , utg . As in the IET case, each of the
generators sj , j = 1, . . . , r, can be written as a word, sj = wjt0 (v1t0 , . . . , vft0t .
0
By lemma 2.14, for every t each of the elements vit0 can be written as a
positive word in the elements, v1t , . . . , vft (t) . Hence, by substituting these
positive words instead of the elements, vit0 , each of the elements sj can be
written as a word: sj = wjt (v1t , . . . , vft t .
By part (1) each of the elements vit can be presented as a positive word
in the new generators: ut1 , . . . , utg . When we substitute these positive words
in the words wjt , we get the words: sj = zjt (ut1 , . . . , utgt ).
As in the IET case (part (5) in proposition 2.4), with each word, wjt
and zjt , we can associate a path in the tree Y , pwjt and pzjt . For a fixed j,
1 ≤ j ≤ r, pwjt and pzjt are identical, and they are identical paths for all t.
(5) there exist positive constants d1 , d2 , such that for every t, there exist a
sequence of indices: 1 ≤ i1 < i2 < . . . < ib ≤ g, such that for every
1 ≤ m1 < m2 ≤ b:
d1 · length(utim ) ≤ length(utim ) ≤ d2 · length(utim ).
1
2
1
31
Furthermore, for every index i, 1 ≤ i ≤ g for which i 6= im , m = 1, . . . , b:
10g · d2 · length(uti ) ≤ length(uti1 )
(6) For each index t, in the words zjt , 1 ≤ j ≤ r, in a distance bounded by
some constant c, either before or after the occurrence of an element uti , for
i 6= i1 , . . . , ib , appears one of the elements utim , 1 ≤ m ≤ b.
Proof: As in the IET case, to get the set of new generators ut1 , . . . , utg we need
to modify the Rips machine or the Makanin combinatorial algorithm. The Rips
machine runs two processes. The first is a process that erases subsegments that
are covered exactly once, and if that process terminates, the machine runs a second
process that applies a sequence of entire transformations until there is a subsegment
that is covered exactly once, or until every segment is covered exactly twice.
The pseudogroups Ut are all Levitt pseudogroups, hence, if one applies the Rips
machine to them, the first process doesn’t terminate. Since our aim is to construct
generators with comparable lengths, we need to separate between long bases (or
elements) and short ones. When we start with such separation, and modify accordingly the Rips machine, it may happen that even though we start with a Levitt
pseudogroup, the first process terminates after finitely many steps (or doesn’t apply
at all), and then one applies the second process finitely many times.
Recall that by lemma 2.11, for t > t0 no non-trivial word in the generators,
v1t , . . . , vft t , fix pointwise a non-degenerate subinterval of the interval I. Furthermore, in applying the first process in the Rips machine to the pseudogroup Ut , only
moves (3) and (4) of this process are applied.
As in the IET case, we start by dividing the generators v1t , . . . , vft t into finitely
many sets according to their length. We order the elements from the longest to the
shortest. We place a separator between consecutive sets whenever there is a pair of
consecutive elements (ordered according to length) that satisfy:
t
)
length(vit ) ≥ c1 (ft , cp ) · length(vi+1
where c1 (ft , cp ) = 4ft cp .
If there are no separators, we set gt = ft , and uti = vit , i = 1, . . . , ft . Suppose
that there is a separator. In that case we call the elements that are in the last
(shortest) set short and the elements in all the other (longer) sets long.
The construction of the new set of generators starts with (possibly none) applications of a modification of the first process in the Rips machine. By lemma 2.11,
for every t > t0 the process that is used to construct the pseudogroups Ut , applies
only moves of type (4) and (3). In each step of the modified process we first apply
moves of type (4) as long as possible. Since each such move erases a pair of bases,
one can apply only finitely many such moves. We continue a step by applying a
modified move of type (3) if such modified move is possible.
Let b1 be a base with a paired base b2 , and suppose that J1 is a maximal
subinterval of b1 (J1 is contained in the interior of b1 ) that is covered only once. J1
is one of the elements v1t , . . . , vft t . If it is not a long element, we do not perform a
modified move of type (3) along J1 . If J1 is long we perform move (3).
By the proof of lemma 2.11 every point in J2 is covered by b2 and at least one
additional base. As in the proof of lemma 2.11, if there is a point in J2 which is the
32
endpoint of a base, and this endpoint is covered only once after erasing J2 from b2 ,
the Euler characteristic of the graph that is associated with the new pseudogroup
has increased by at least 1. In this case, we replace the system of generators
v1t , . . . , vft t , by the new system of generators that is obtained after the modified
move (3), and start the whole process again. The graphs that are associated with
the new pseudogroups that we obtain have strictly bigger Euler characteristics
than the previous ones, and they will satisfy all the properties that are listed in
claims 2.11-2.15. Since their Euler characteristic is strictly bigger, and it has to
be negative, such a replacement can occur only finitely many times. Hence, in the
sequel we may assume that such an increase in the Euler characteristic does not
happen along the modified process that we present.
After performing (the modified) move (3), we get a new set of elements. The
endpoints of all the new bases are covered at least twice (so no move of type (2)
is required as we argued in the proof of lemma 2.11). There are at most 4 new
elements that have one endpoints at one of the two ends of J2 . A new element that
is contained in a previous long element, and its length is smaller than the maximal
length of (a previous) short element, is said to be secondary short. Every previous
long element contains at least a single new element of length that is bounded below
by ft times the length of a maximal (previous) short element. There may be (at
most two) short elements that were cut into two new elements. We just denote the
cut points on these short elements, but continue to the next steps with the lengths
of the previous short elements.
At this point we consider all the elements that are either short elements from
the original pseudogroup, or new elements that are contained in previous long
elements, and are not secondary short. We divide them into short and long elements
precisely as we divided them before the first step. Note that every element that
was short before the first step, and was perhaps cut into two new elements is still
regarded short. Also, note that all the new short elements have lengths that are
bounded below by the length of a maximal previous short element. Every previous
long element either stayed long, or can be written as a positive word in the new
generators, and this word contains at least a one new short or new long generator.
We repeat these steps iteratively. We first perform all the possible moves of
type (4), and then perform a modified step of type (3). i.e., we perform move
(3) only along a long element. As in lemma 2.11, since the endpoints of all the
new bases are covered at least twice, no move of type (2) is required along the
modified process. Afterwards we set some of the new elements (possibly none) to
be secondary short, and redefine the elements that are long and short as we did in
the first step. In considering lengths, we only denote new cutpoints on previously
defined short elements, so the lengths of short elements can only increase along the
process.
The modified process terminates after finitely many steps, since at each step we
erase a long element, i.e., an element of length bounded below by 4 · cp · ft times
the maximal length of the original short elements. The process terminates when
there are no long elements that are covered exactly once. Long elements may be
cut along the process, and at least one of the new elements that are obtained after
the new cuts are either long or short, and in the last case, their length is bounded
below by the maximal length of the original short elements, i.e., the short elements
from the set of generators: v1t , . . . , vft t .
Short elements may be cut along the process, but the number of cuts is bounded
33
by 3 times the number of the original long elements, which means that the number
of cuts is bounded by 3 · ft . Therefore, every short element is cut into at most 3 · ft
1
times
new elements, so the length of at least one of them is bounded below by 3·f
t
the minimal length of an original short element, i.e., a short element from the set
v1t , . . . , vft t . Therefore, every element from the set v1t , . . . , vft t can be written as a
positive word in the elements that are obtained after the first (modified) process,
and there exists a constant c2 (cp , ft ), so that every subword of combinatorial length
1
c2 in these words, contains either a long element, or an element of length 3·f
times
t
t
t
the minimal length of a short element from the original set v1 , . . . , vft .
Although after the first (modified) process there are no long elements that are
covered exactly once, there may still remain long elements after that process. Hence,
although we analyze pseudogroups of Levitt type, what is left to analyze are long
elements, so that in the intervals that support them every point is covered at least
twice. To analyze these long elements and shorten them we use a variation of the
process that is used in the proof of proposition 2.4.
If there exists a subsequence of the indices t, for which there exists a constant
c3 > 0, so that the maximal length of a long element (after the first process) is
bounded by c3 times the maximal length of a short element, the conclusion of the
theorem follows by the same argument that was used in the proof of proposition
2.4. Hence, in the sequel we will assume that there is no such subsequence.
We will denote the elements that are obtained after the first (modified) process,
t
v̂1 , . . . , v̂ftˆ , and the pseudogroup that they generate by Ût . By (the proof of) parts
t
(2) of lemmas 2.13 and 2.15, with this infinite sets of generators, there are associated
only finitely many presentations of the group F (lemma 2.13), and of the group that
they generate (lemma 2.15). Hence, by passing to a subsequence we may assume
that the presentations are all the same. Also note that since in the first process we
only applied finitely many moves of types (3) and (4), lemma 2.14 remains valid for
the semigroup that is associated with Ût . i.e., the semigroup that is generated by
the elements: v̂1t , . . . , v̂ftˆ contains the semigroup that is generated by the elements:
t
v1t , . . . , vftˆ .
t
At this point we pass to a further subsequence, and divide the elements v̂1t , . . . , v̂ft t
into finitely many equivalence classes. First, by passing to a subsequence, we may
assume that ft , the number of elements, is fixed and equals f . Moreover, we may
assume that the short and secondary short elements (after changing the order) are
the elements v̂1t , . . . , v̂ℓt1 .
In the lowest equivalence class we put all the short and the secondary short
elements. Note that by construction, the ratios between the lengths of any two
short elements is uniformly bounded (i.e., bounded by a constant independent of t),
and the ratios between the lengths of secondary short elements and short elements
are uniformly bounded from above.
We are left with long elements. The second equivalence class, that after change
of order we may assume to include the elements v̂ℓt1 +1 , . . . v̂ℓt2 . An element v̂it is
included in this class if the ratio between the lengths of v̂it and the maximal length
of a short element approaches infinity, the ratios between elements in this class are
uniformly bounded, and the ratio between the length of an element in this class
and long elements that are not in the class approaches 0.
The other classes are defined iteratively. Each class contains elements so that
34
the ratios between their lengths and the lengths of elements in lower classes approaches infinity, the ratios between the lengths of elements in the class are uniformly bounded, and the ratios between their lengths and lengths of elements in
higher classes approaches 0. By passing to a further subsequence we may assume
that the ratios between elements in the same equivalence class (above the lowest
one), and between short elements converge to positive constants.
Suppose first that all the bases that contain elements that belong to all the
equivalence classes above the lowest class (the one that contains short elements),
are covered exactly twice. i.e., that two such bases overlap only in elements that
are in the lowest class. In that case we modify the procedure that was used in the
proof of proposition 2.4 (the IET case).
We modify the Dehn twists that appear in the procedure that was used in the
IET case, to gradually reduce the number of long elements, and while doing that
add a bounded number of secondary short elements.
Let I be the interval that supports the pseudogroup Ût (Ût is in fact supported
on finitely many subintervals that are contained in I). We consider all the elements
that belong to classes that are above the lowest class to be long, and all the elements
in the lowest class to be short or secondary short.
We further divide the short and secondary short elements in Ût into several
classes. We say that two such elements v̂it1 , v̂it2 are in the same class if there
is a sequence of short and secondary short elements that leads from v̂it1 to v̂it2
and the length between the endpoints of consecutive elements in the sequence is
bounded by 2c4 (cp , ft ) times the maximal length of a short element. This divides
the short and secondary short elements into at most ft equivalence classes. Each
such equivalence class is supported on some subinterval of I. Our strategy along
the process that we present (that modifies the one that is used in the proof of
proposition 2.4), is to gradually reduce the number of these equivalence classes
(of short and secondary short elements) and the number of long bases (i.e. bases
that contain long elements), while creating a bounded number of new secondary
short and short elements, such that that the lengths of the new short elements are
bounded below by a fixed fraction of the the lengths of the previous short elements
where the fraction and the the numbers of new short and secondary short elements
are bounded by a constant that depends only on the number of elements in the
pseudogroup Ût , with which we start this part of the procedure.
Let pv be the vertex at the positive end of I. On the given pseudogroup Ût we
perform the following operations:
(1) Suppose that pv does not belong to a subinterval that supports one of the
equivalence classes of short and secondary short elements. In particular,
there are only two long bases that are adjacent to pv. Suppose that one
of them is b1 and the other is b2 . b1 and b2 can not be a pair of bases,
since by part (2) of lemma 2.11 no non-degenerate subinterval is fixed by a
non-trivial word.
Suppose that b1 is longer than b2 , and that the other endpoint of b1 is
contained in a subinterval that supports an equivalence class of short and
secondary short elements that contains the endpoint of just one additional
long base, and this long base does not overlap with b1 in a long element.
Let b3 be the base that supports the subinterval that supports the class that
contains the (other, not pv) endpoint of b1 .
35
In that case we first transfer all the long elements, starting with b2 until
b3 (not including b3 ), and the short and secondary short elements that are
supported on the interval that supports b1 , until the class that includes
the endpoint of b3 that we denote q (q is contained in the subinterval that
supports b1 ). We transfer these elements from the subinterval that supports
b1 to the subinterval that supports its paired base.
We cut b1 at the point q, the endpoint of b3 , and throw away the part
of b1 between pv and q from b1 and from its paired base. We further add a
marking point on what was left from the base b1 , at the point that was the
limit of the subinterval that supported the equivalence class of short and
secondary short elements, that contained q. Note that this point is marked
on the base that is paired with the new (what was left from) b1 . Note that
by adding this marking, and the additional element that is associated with
it, we guarantee that the semigroup that is generated by the elements that
are associated with the new pseudogroup contains the semigroup that is
associated with the previous semigroup. The group that is generated by
the two semigroups remains the same.
We end this step by checking if there are new elements of length that are
bounded by c4 times the length of a short element. In case there are such
elements we declare them to be secondary short. We also divide the short
and secondary short elements into equivalence classes as we did before this
step. The number of such equivalence classes did not increase after this step
(it may decreased). The number of bases did not change after this step.
(2) Suppose that pv does not belong to a subinterval that supports one of
the equivalence classes of short and secondary short elements (as in case
(1)). Suppose the b1 is longer than b2 , and that the other endpoint of b1 is
contained in a subinterval that supports an equivalence class of short and
secondary short elements that contains the endpoint of a long base that
overlaps with b1 in a long element. Let b3 be the long base that overlaps
with b1 along some long elements and such that its endpoint is supported by
the same subinterval that supports the class of short and secondary short
elements and supports the endpoint of b1 .
Like in part (1) we first transfer all the long bases, starting with b2 until
b3 (not including b3 ), and the short and secondary short elements that are
supported on the interval that supports b1 , until the class that includes the
endpoints of b3 and b1 , to the subinterval that supports the base that is
paired with b1 .
Let q1 be the endpoint of b1 and q3 be the endpoint of b3 , be the endpoints
that are supported by the same subinterval that supports an equivalence
class of short and secondary short elements. Suppose that q3 is supported
on the interval that supports b1 . In that case we transfer b3 using b1 . We
further cut b1 (and its paired base) into 3 new (paired) bases. One from pv
to the beginning of the subinterval that supports the equivalence class of
short and secondary short elements that contain q1 and q3 , the second from
this point to q3 and the third from q3 to q1 . We erase the first part of b1
and its paired base, i.e., the part from pv to the beginning of the subinterval
that supports the equivalence class that contains q1 and q3 . The other two
parts that are left from b1 are set to be secondary short.
Suppose that q1 is supported by the subinterval that supports b3 . In
36
that case we cut the base b3 at q1 and declare the part between q1 and q3
to be secondary short. We do that same to the base is paired with b3 . We
transfer the part of b3 that contains a long element using b1 . We further
cut b1 into two bases, One from pv to the beginning of the subinterval that
supports the equivalence class of short and secondary short elements that
contains q1 and q3 , and the other from that point to q1 . We erase the first
part of b1 and its paired base.
In both cases, we added two new pairs of secondary short bases, and get
rid of one long pair. Hence, step of the type (2) can occur only boundedly
many times (bounded by the number of long pairs in Ût ). We end this step
by checking if there are new short or secondary short elements, and dividing
the short and secondary short elements to equivalence classes as we did
before this step (and at the end of step (1)). The number of such equivalence
classes did not increase after this step (it may decreased). In both cases
the semigroup that is associated with the new pseudogroup contains the
semigroup that is associated with the previous semigroup. the groups that
are generated by both semigroups are identical.
(3) Suppose that pv belongs to a subinterval that supports one of the equivalence classes of short and secondary short elements. Let b1 and b2 be the
long bases that are adjacent to the subinterval that supports the equivalence
class that contains pv.
Suppose that b1 is longer than b2 , and as in part (1), the other endpoint
of b1 is contained in a subinterval that supports an equivalence class of short
and secondary short elements that does not contain the endpoint of a long
base that overlaps with b1 along a long element. Let b3 be the base that
supports the subinterval that supports the class that contains the (other,
not pv) endpoint of b1 .
Let p1 be the endpoint of b1 and p2 be the endpoint of b2 that are closer
to pv. Suppose that p2 is supported on the interval that supports b1 . In
that case we transfer b2 using b1 . We transfer all the long bases that overlap
with b1 , from b2 to b3 (not including b3 , and all the short and secondary
bases and elements that are in the equivalence classes that are supported
on b1 from pv until the endpoint of b3 , and do not include the equivalence
class of pv and the equivalence class of the endpoint of b3 that is supported
on b1 , using b1 . We cut b1 into 3 bases, one from p1 to p2 , and the second
from p2 to the endpoint of the subinterval that supports the equivalence
class of short and secondary short elements that contains p1 and p2 . The
third is from that point to the endpoint of b1 . We set the first two parts of
b1 to be secondary short.
Suppose that p1 is supported by the subinterval that supports b2 . In that
case we cut the base b2 at p1 and declare the part between p1 and p2 to be
secondary short. We do the same to the base that is paired with b2 . We
transfer the part of b2 that contains a long element using b1 , and transfer
the other long elements that are supported on b1 (not including b3 ), and
the short and secondary short elements that are in equivalence classes that
are supported on b1 from pv until the endpoint of b3 , and do not contain p1
or the endpoint of b3 , using b1 . Finally we cut b1 into two bases, one from
p1 to the endpoint of the subinterval that supports the equivalence class of
short and secondary short elements that contains p1 and p2 . The second is
37
from this point to the endpoint of b1 .
We continue as in part (1). Let q be the endpoint of b3 that is supported
by b1 . We cut b1 at the point q, the endpoint of b3 , and throw away the part
of b1 between the endpoint of the subinterval that supports the equivalence
class of short and secondary short elements that contains pv and q from b1
and from its paired base. We further add a marking point on what was left
from the base b1 , at the point that was the limit of the subinterval that
supported the equivalence class of short and secondary short elements, that
contained q.
As in the first two parts, the semigroup that is generated by the elements
that are associated with the new pseudogroup contains the semigroup that
is associated with the previous semigroup. The group that is generated by
the two semigroups remains the same.
We end this step by checking if there are new short or secondary short
elements, and dividing the short and secondary short elements to equivalence classes as we did before this step. The number of such equivalence
classes that participate in the next steps of the process decreased by at least
1 after this step, hence, it may occur only boundedly many times (bounded
by the number of such classes in the original pseudogroup Ût ).
(4) Suppose that pv belongs to a subinterval that supports one of the equivalence classes of short and secondary short elements. Let b1 and b2 be the
long bases that are adjacent to the subinterval that supports the equivalence
class that contains pv.
Suppose that b1 is longer than b2 , and as in part (2), the other endpoint of
b1 is contained in a subinterval that supports an equivalence class of short
and secondary short elements that contains the endpoint of a base that
overlaps with b1 along a long element. Let b3 be the long base that overlaps
with b1 along some long elements and such that its endpoint is supported by
the same subinterval that supports the class of short and secondary short
elements and supports the endpoint of b1 . In this case we combine what we
did in step (3) on one side of b1 , and in step (2) in the other side of b1 .
First we use the endpoint p1 of b1 , and p2 of b2 , to cut b1 and b2 as we
did in part (3). Then we transfer all the long bases from b2 to b3 (excluding
b3 ), and all the short and secondary short bases that belong to equivalence
classes, from the one that contains pv to the one that contains the endpoints
of b1 and b3 , excluding the initial class that includes pv and excluding the
terminal one that contains the endpoints of b1 and b3 .
We continue as in part (2). We use q1 , the endpoint of b1 , and q3 the
endpoint of b3 , to cut the bases b1 and b3 as we did in part (2). Then we
transfer the long part of (what left from) b3 using b1 , and erase the long
part of (what is left from) b1 , precisely as we did in part (2).
In this case, we added at most 4 new pairs of short and secondary short
bases, and got rid from at least one pair of long bases (b1 and its paired
base). Also, in the active part of the pseudogroup, the number of equivalence
classes of short and secondary short elements is reduced by at least 1. Hence,
part (4) can occur boundedly many times (bounded by the number of long
pairs in Ũt ). As in the previous cases, the semigroup that is associated with
the new pseudogroup contains the semigroup that is associated with the
38
previous pseudogroup. the groups that are generated by both semigroups
is identical.
As in the previous parts, we end this step by checking if there are new
secondary short elements and finally dividing the short and secondary short
elements to equivalence classes as we did before this step. The number of
such equivalence classes in the active part of the pseudogroup decreased by
at least 1 after this step.
Since in parts (2) and (4) the number of pairs of long bases is reduced, and in
step (3) the number of equivalence relations of short and secondary short elements
is reduced, these steps can occur only boundedly many times (where the bound
can be taken to be the number of elements in the initial pseudogroup Ût , which is
uniformly bounded for all t). Therefore, only step of type (1) can occur finitely but
unboundedly many times. In part (1), the endpoint of the interval I that supports
the bases of the corresponding pseudogroup, pv, is the endpoint of the two long
bases that are adjacent to it, and the equivalence classes of short and secondary
short bases are supported on subintervals that do not contain the endpoint pv.
By construction, a step of type (1) produces no new bases, but it may add marks
on a long base. In case a base that is adjacent to pv was marked, and the marked
part of the base is erased, then the marking is inherited by the other base that is
adjacent to pv. Note that the number of markings on each long base is bounded by
the number of endpoints of long bases, hence, it is bounded by twice the number
of long bases in the initial pseudogroup Ût , which is uniformly bounded.
We denote the pseudogroup that is obtained when the second process terminates
Utf n . It contains no long elements, and the total number of elements is bounded in
terms of the number of elements in Ût , so by passing to a further subsequence we
may assume that it is fixed. The terminal pseudogroup Ût and its set of generators
satisfy the conclusions of proposition 2.16, precisely in the same way it is was argued
in the IET case in proposition 2.4.
So far we assumed that in the initial pseudogroup, Ût , there is no element, that
is not in the lowest equivalence class that consists of short and secondary short
elements, and is contained in more than 2 bases. Suppose that there exists such
element.
Suppose first that there exists a subsequence of indices t, for which there exists
a long element in the highest equivalence class that is contained in more than 2
bases. With each of the pseudogroups we can naturally associate an action of the
free group F on a real tree, and after appropriate rescaling and further passing
to a subsequence, we can assume that these actions converge into an action of F
on a real tree Y∞ . By the Gromov-Hausdorff convergence, The pseudogroups Ut
converge into a pseudogroup U∞ , from which it is possible to reconstruct the action
of F on U∞ .
In the pseudogroup U∞ , every element that is not in the highest class in the
pseudogroups Ût degenerates to a point, and every element in the highest class
converges into a non-degenerate element in U∞ . By our assumptions on the pseudogroups Ût , there exists a non-degenerate element in U∞ that is covered by at
least 3 bases (of U∞ ).
At this point we apply the second process of the Rips (or the Makanin) machine
to the pseudogroup U∞ . Since each segment is covered at least twice, the second
process in the Rips machine is applied (see section 7.3 in [Be-Fe]). By part (2) of
39
lemma 2.11 no non-trivial word in the generators of the pseudogroups Ut (hence,
also Ût ) fixes a non-degenerate subsegment in the original tree. The periodicity
of the words that are associated with the various bases in Ût is assumed to be
bounded. Hence, when the second process of the Rips machine is applied to the
pseudogroup U∞ there can not be a step in which a base is supported on precisely
the same subinterval as its paired base.
Therefore, along the applications of the second process of the Rips machine to
U∞ only entire transformations are applied at each step. This means that every
point in the interval that supports the pseudogroup that is constructed at each step
is covered at least twice, and since there are subintervals that are covered more than
twice by bases in U∞ , there are non-degenerate subintervals that covered more than
twice in the pseudogroups that are constructed in each step of the process.
With the pseudogroup U∞ there is an associated faithful action of the free group
F on some real tree. Since F is free, the action contains no axial components.
Hence, when the second process of the Rips machine is applied to the pseudogroup
U∞ , no toral (axial) component can be discovered at any step of the process. Therefore, by proposition 7.6 in [Be-Fe] after finitely many steps each point which is not
an endpoint of one of the bases in the constructed pseudogroup must be covered
exactly twice. However, the in the process that is applied to U∞ there are always non-degenerate subintervals that are covered more than twice, and we get a
contradiction.
So far we can deduce that in the pseudogroups Ût (perhaps after passing to a
further subsequence), every element that belongs to the highest equivalence class
is covered exactly twice. Let vi0 be an element that is covered more than twice, for
which there are no elements in a higher class that are covered more than twice. In
that case we modify the procedure that we used in case every element that is not in
the lowest class is covered twice, to reduce the longer elements to the length of the
elements in the equivalence class of the element that is covered more than twice.
We divide the equivalence classes of the elements in Ût into 3 categories. The
elements in classes that are lower than the class of vi are set to be small. The
elements that are in classes that are higher then vi are set to be big, and the
elements in the class of vi are set to be intermediate.
We further divide the small and intermediate elements into several classes. We
say that two small elements v̂it1 , v̂it2 are in the same class if there is a sequence of
small elements that leads from v̂it1 to v̂it2 and the length between the endpoints of
consecutive elements in the sequence is bounded by 2c4 (cp , ft ) times the maximal
length of a small element. We will refer to these classes as small classes of elements.
Similarly we divide the collection of small and intermediate elements into classes.
We say that two such elements are in the same class if there exists a sequence of
small and intermediate elements that leads from one such element to another, and
the length between the endpoints of consecutive elements is bounded by 2c4 times
the length of the maximal intermediate element. We will refer to such classes as
intermediate classes. Clearly, the numbers of small and intermediate classes is
bounded by ft , and every small class is contained in an intermediate class.
Our strategy along the process is to gradually reduce the number of long elements
and the number of intermediate classes, while keeping the intermediate elements
that are covered more than twice. When the procedure terminates, the length of
the terminal long elements is bounded by a constant times the maximal length
40
of an intermediate element, hence, intermediate elements belong to the highest
equivalence class, at least one of them is covered more than twice, and we will get
a contradiction by the argument that was used in case an element in the highest
equivalence class is covered more than twice.
Let pv be the vertex at the positive end of I. On the given pseudogroup Ût we
perform the following operations:
(1) Suppose that pv does not belong to a subinterval that supports one of the
intermediate equivalence classes. Let b1 and b2 be the two big bases that
end at pv, suppose that b1 is longer, and that b1 ends in a point that belongs
to an equivalence class of intermediate elements that contains no endpoint
of a big base that overlaps with b1 along a big element.
In that case we act precisely as in part (1) of the procedure in which
every element which is not short is covered exactly twice. Note that if an
intermediate element was covered more than twice (by intermediate and big
bases), then that same element (perhaps after it was transfered) is covered
more than twice (by intermediate and big bases) after the move.
We end this step by checking if there are new elements of length that are
bounded by c4 times the maximal length of an intermediate element. If the
length of such new element is bounded by c4 times the maximal length of a
small element, we consider it to be small. Otherwise it will be considered as
an intermediate element. Finally we divide the small and the intermediate
elements into equivalence classes as we did before the initial step. Note that
if the number of equivalence classes is the same as before the step, no new
small or intermediate elements are created. Also, that we don’t add any
markings to long elements, as our goal is to get a contradiction and deduce
that the whole process that we use in this case is in fact not needed.
(2) Suppose that pv does not belong to a subinterval that supports one of the
equivalence classes of intermediate equivalence classes. Let b1 and b2 the
two big bases that have pv as an endpoint, let b1 be the longer, and suppose
that b1 ends in a point that belongs to an equivalence class of intermediate
elements, and that class contains the endpoint of a big base that overlaps
with b1 along a big element, that we denote b3 .
In that case we act in a similar way to what we did in part (2) of the
previous procedure. First, we transfer all the big bases, starting with b2
until b3 (not including b3 ), and all the small and intermediate bases that
are supported on the interval that supports b1 , until the intermediate class
that includes the endpoints of b3 and b1 , to the subinterval that supports
the base that is paired with b1 .
Let q1 be the endpoint of b1 and q3 be the endpoint of b3 , that are
contained in a subinterval that supports the same intermediate class. Let
p be the endpoint of that intermediate class that is closer to pv. We cut b1
and b3 at the point p, and cut their paired bases accordingly. We transfer
the part of b3 that starts at the endpoint that is not q3 and ends at p using
b1 . We further erase the part of b1 from pv to p, and its paired base.
In this case we added two pairs of intermediate (or small) bases, but
erased a big pair of bases. Hence, step of the type (2) can occur only
boundedly many times. Note that if there was an intermediate element
that was covered more then twice before this step, then there exists an
41
intermediate element that is covered more than twice after applying step
(2). As at the end of step (1) we check if there are new small or intermediate
elements, and update the division of small and intermediate elements into
equivalence classes.
(3) Suppose that pv belongs to a subinterval that supports one of the equivalence classes of small or intermediate equivalence classes of elements. Let b1
and b2 be the big bases that are adjacent to the subinterval that supports
the small or intermediate equivalence class that contains pv.
Suppose that b1 is longer than b2 , and as in part (1), the other endpoint
of b1 is contained in a subinterval that supports an intermediate class does
not contain the endpoint of a big base that overlaps with b1 along a big
element. Let b3 be the big base that supports the subinterval that supports
the intermediate class that contains the (other, not pv) endpoint of b1 .
Let p1 be the endpoint of b1 and p2 be the endpoint of b2 that are closer to
pv, and let p be the endpoint of the interval that supports the intermediate
class that supports pv (and is not pv). We first cut b1 and b2 at the point p,
and accordingly their paired bases. We transfer all the big bases from what
is left from b2 until b3 (not including b3 , using b1 . We further transfer all
the small and intermediate classes from the class the intermediate class that
includes pv (but not including this class of pv) until the class that includes
the endpoints of both b1 and b3 using b1 .
We continue as in part (1). Let q be the endpoint of b3 that is supported
by b1 . We cut what is left from b1 at q. We throw away the part of b1
between the points p and q (and the corresponding part from its paired
base). Note that since there was an intermediate element that is covered by
at least 3 bases, there is still such an intermediate element after applying
step (3).
As in steps (1) and (2), we check if there are new small or intermediate
elements after this step. We further update the equivalence classes of small
and intermediate elements. The number of intermediate classes that participate in the next steps of the process decreased by at least 1 after applying
step (3), so it may occur only boundedly many times.
(4) Suppose that pv belongs to a subinterval that supports one of the intermediate classes. Let b1 and b2 be the big bases that are adjacent to the
subinterval that supports the intermediate class that contains pv.
Suppose that b1 is longer than b2 , and as in part (2), the other endpoint
of b1 is contained in a subinterval that supports an intermediate class that
contains the endpoint of a big base, that we denote b3 , that overlaps with
b1 along a big element.
In this case we combine what we did in part (3), along the subinterval
that supports that class of pv, and in part (2), along the subinterval that
supports the class that contains the endpoints of b1 and b3 (cf. part (4) in
case there was no long element that is covered more than twice).
In this case, we added at most 4 new pairs of small or intermediate bases,
and got rid from at least one pair of big bases (b1 and its paired base). Also,
in the active part of the pseudogroup, the number of intermediate classes
is reduced by at least 1. Hence, part (4) can occur boundedly many times.
Note that as in the previous steps, since there was an intermediate element
that is covered more than twice, there such an intermediate element after
42
applying step (4).
As in the previous parts, we end this step by checking if there are new
small or intermediate elements. We also update the collections of small and
intermediate equivalence classes of elements.
Parts (2)-(4) can occur boundedly many times, so as the number of big elements
or the number of intermediate classes reduces in each of them. All the steps preserve
the existence of intermediate elements that are covered more than twice. When the
procedure terminates the lengths of all the elements are bounded by a constant
times the length of an intermediate element. Hence, when we rescale the length
so that the maximal length of an element is 1, and pass to a convergent sequence
of pseudogroups, the convergent sequence converges into a sequence in there exists a non-degenerate subinterval that is covered more than twice. Therefore, the
argument that leads to a contradiction, in case there is an element in the highest
equivalence class that is covered more than twice (and the periodicity is bounded)
leads to a contradiction in this case as well. This implies that in the bounded
periodicity case, all the elements that are not in the lowest class in Ût , i.e., the
class that contains only short and secondary short elements, are all covered exactly
twice, and the procedure that web used in this case constructs a subsequence of
pseudogroups that satisfy the conclusions of proposition 2.16.
As in the IET case, proposition 2.16 enables the proof of the key claim in the
proof of theorem 2.10 - the combinatorial bounded cancellation along the process
that constructs the new sets of generators: ut1 , . . . , utg . Since proposition 2.16
is proved under the bounded periodicity assumption, we first prove the bounded
cancellation assuming bounded periodicity.
Proposition 2.17. Suppose that there exists an integer cp , such that the periodicity
of the elements hn (s1 ), . . . , hn (sr ) is bounded by cp for all integers n.
With the notation of proposition 2.16, there exists a constant C > 0, so that
for a subsequence of the indices t, that for brevity we still denote t, the words zjt ,
1 ≤ j ≤ r, can be replaced by words: ẑjt with the following properties:
(1) As elements in ambient free group F : ẑjt (ut1 , . . . , utg ) = zjt (ut1 , . . . , utg ).
(2) ẑjt is obtained from zjt by eliminating distinct pairs of subwords. Each pair
of eliminated subwords corresponds to two subpaths of the path pzjt that lie
over the same segment in the tree Twj , where the two subpaths have opposite
orientations.
(3) With the word ẑjt (ut1 , . . . , utg ) we can naturally associate a path in the tree Y ,
that we denote, pẑjt . The path pẑjt can be naturally divided into subsegments
according to the appearances of the subwords uti in the word ẑjt .
Let DBẑjt be the number of such subsegments that are associated with
subwords uti in pẑjt , that at least part of them is covered more than once by
the path pẑjt . Then for every t > t0 and every j, 1 ≤ j ≤ r, DBẑjt ≤ C.
Proof: In case of bounded periodicity, proposition 2.17 follows from proposition
2.16 by exactly the same argument that proposition 2.5 follows from proposition
2.4 in the IET case.
43
Under the bounded periodicity assumption, proposition 2.17 and part (3) of
proposition 2.16, that proves that the tuples ut1 , . . . , utg generate groups with similar
presentations, imply the conclusions of theorem 2.10 by exactly the same argument
that was used in the IET case (theorem 2.2).
As in the IET case, to omit the bounded periodicity assumption, we modify the
statements and the arguments that were used in the proofs of propositions 2.16
and 2.17, in a similar but a slightly different way than in the IET case, to include
long periodic subwords, or in the limit, to include non-degenerate segments with
non-trivial stabilizers.
We start with a generalization of proposition 2.16, which is the analogue of
proposition 2.7 in the Levitt case. Recall that the aim of proposition 2.16 was to
replace the generators v1t , . . . , vft t by a new (possibly larger set of) generators so
that the ratios between their lengths is globally bounded.
Proposition 2.18 (cf. proposition 2.7). There exists a subsequence of indices t,
that for brevity we still denote t, for which the finite set of generators: v1t , . . . , vft t
can be replaced by elements ut1 , . . . , utg that satisfy properties (1)-(4) in proposition 2.16. Properties (5) and (6) in proposition 2.16 are replaced by the following
properties:
(6) there exist a real number d2 > 1 and a subset of indices 1 ≤ i1 < i2 <
. . . < ib ≤ g, such that for every index i, 1 ≤ i ≤ g for which i 6= im ,
m = 1, . . . , b:
10g · d2 · length(uti ) ≤ length(uti1 )
(7) there exists an integer ℓ, 0 ≤ ℓ ≤ b, and a positive real number d1 such that
for every ℓ + 1 ≤ m1 < m2 ≤ b:
d1 · length(utim1 ) ≤ length(utim2 ) ≤ d2 · length(utim1 )
For every m1 ≤ ℓ and ℓ + 1 ≤ m2 ≤ b:
d1 · length(utim ) ≤ length(utim )
2
1
(8) For every t, and every index m, 1 ≤ m ≤ ℓ, there exist distinct indices
1 ≤ j1 , . . . , jem ≤ g that do not belong to the set i1 , . . . , ib , such that:
pm
wm = utj1 . . . utjem , and utm = αwm
where α is a suffix of wm .
t
(9) for each index t, in the words zj , 1 ≤ j ≤ r, in a bounded distance either
before or after the occurrence of an element uti , for i 6= i1 , . . . , ib , appears
one of the elements utim , 1 ≤ m ≤ b.
Proof: To prove proposition 2.18 we start with the same (first) procedure that
was used in the proof of proposition 2.16 (the bounded periodicity case). Hence,
we divide the generators v1t , . . . , vft t into finitely many sets according to their length,
call only the elements in the shortest group short and all the other (longer) elements
long.
At this point we apply the modification of the first process in the Rips machine
(the Makanin algorithm) that was used in the bounded periodicity case. Note that
44
the bounded periodicity assumption is not used nor mentioned along this modified
first process. Once the modified first process terminates, every long element is
covered at least twice, though there may still be short or secondary short elements
that are covered only once.
To analyze the pseudogroup that is the output of the modified first process,
we use modifications that combine the procedures that were used in the proofs of
propositions 2.16 and 2.8. If there exists a subsequence of the indices t, for which
there exists a constant c3 > 0, so that the maximal length of a long element (after
the first process) is bounded by c3 times the maximal length of a short element, the
conclusion of the theorem follows. Hence, in the sequel we will assume that there
is no such subsequence.
As in the proof of proposition 2.16, we pass to a further subsequence, and divide
the elements v̂1t , . . . , v̂ftˆ , the generators of the semigroup that is associated with the
t
pseudogroup that was constructed after the first modified procedure, into finitely
many equivalence classes according to their lengths. We may also assume that fˆt ,
the number of elements, is fixed.
Suppose first that all the bases that contain elements that belong to all the
equivalence classes that are longer than the lowest equivalence class (the one that
contains short elements), are covered exactly twice. i.e., that two such bases overlap
only in elements that are in the lowest class. In that case the procedure that was
used in the proof of proposition 2.16, that modifies the procedures that were used
in the proofs of proposition 2.4 (in the bounded periodicity case) and in proposition
2.8 (in the general IET case) prove the conclusions of the proposition.
Suppose that there exists an element that belongs to an equivalence class that is
longer than the lowest class that contains the short and secondary short elements,
that is covered by more than 2 bases. Suppose first that there exists a long element in the highest equivalence class that is contained in more than 2 bases. As
in the proof of proposition 2.16, after an appropriate rescaling and passing to a
subsequence, the given pseudogroups, Ût , converge into a pseudogroup U∞ , that is
associated with a faithful action of the free group F on some real tree Y∞ .
In the pseudogroup U∞ , every element that is not in the highest class in the
pseudogroups Ût degenerates to a point, and every element in the highest class
converges into a non-degenerate element in U∞ . By our assumptions on the pseudogroups Ût , every non-degenerate element that supports some of the bases of U∞
supports at least two bases, and there exists a non-degenerate element in U∞ that
is covered by at least 3 bases (of U∞ ).
Starting with U∞ we apply the second process of the Rips (or the Makanin)
machine, since by our assumption every non-degenerate segment in the interval
that supports U∞ is covered at least twice. The second process apply a sequence
of entire transformations, until there is base that is identified with its dual.
Since the free group F is free, the action of F on the real tree Y∞ contains no
axial components. Hence, by proposition 7.6 in [Be-Fe], either after a finite time
we get a pseudogroup in which a base is identified with its dual, or every nondegenerate segment is covered exactly twice. Since we started with U∞ in which
there was a non-degenerate segment that is covered more than twice, after finitely
many applications of entire transformations we must get to a pseudogroup in which
a base is identified with its dual.
Recall that by part (2) of lemma 2.11 no non-trivial word in the generators of the
45
pseudogroups Ut (hence, also Ût ) fixes a non-degenerate subsegment in the original
tree. Hence, when a base is identified with its dual, it follows that the part of the
words that correspond to such a base has to be unboundedely periodic (the length
of the period is not bounded below by a positive constant times the length of the
base).
Let b be the base that is identified with its dual after finitely many entire transformations. If the subinterval that supports b, supports only b and its dual, we
continue with the rest of the pseudogroup, and by our assumption there must be
a non-degenerate subsegment that supports at least 3 bases that are not b nor its
dual.
Let Ib be the subinterval that supports the base b and its dual. Suppose that
there exists a non-degenerate subinterval of Ib that is transfered by a non-trivial
word in the that involve all the bases except for b and its dual to a non-degenerate
subinterval of Ib . Then there exists a non-trivial word, the commutator of this
(non-trivial) word and the transfer from b to its dual, that acts trivially on a nondegenerate segment in the original tree, a contradiction to part (2) in lemma 2.11.
Note that the subinterval Ib is contained in the simplicial (discrete) part of the
real tree Y∞ , and it has a non-trivial stabilizer. Hence, the part of every base that
is supported on Ib has a non-trivial stabilizer, and is contained in the simplicial
part of Y∞ . At this point we continue applying the Rips machine. We erase the
bases b and its dual. and if there are subintervals of Ib that are covered only once
(after the erasing of b and its dual), we apply the first process in the Rips machine.
Since Ib is contained in the simplicial part of Y∞ the first process of the Rips
machine terminates after finitely many steps (it erases subintervals that are all
contained in the simplicial part). We start the process by erasing subintervals from
basis that are (partly) supported on Ib , and we never get back to a subinterval that
is supported on Ib , since as we argued before, in such a case we get a non-trivial
word that stabilizes a non-degenerate segment, a contradiction to part (2) in lemma
2.11. By the same reason the subintervals that we erase from the various bases must
have disjoint supports except (possibly) for their endpoints.
Since the subintervals that we erase must have distinct supports, and the erasing
procedure terminates after finitely many steps, we are left with a new pseudogroup,
that has a bigger Euler characteristic (smaller in absolute value, smaller complexity
in Makanin’s terminology), in which every point that is not an endpoint of a base
is covered at least twice, and in which there exists a subsegment that supports a
subinterval of a base that was erased, and that subsegment can be mapped (using a
word in the generators of the old pseudogroup) to the subinterval Ib . In particular,
this subinterval has non-trivial stabilizer, and it belongs to the simplicial part of
Y∞ .
We continue by applying the second process of the Rips machine. After finitely
many steps there must exist a new basis that is identified with its dual. Let Ic be
the subinterval that supports that new basis that is identified with its dual. If a
subinterval of Ic is identified with another subinterval of Ic using a non-trivial word
that does not involve the new base and its dual, we get a non-trivial word that acts
trivially on a non-degenerate segment, a contradiction to part (2) of lemma 2.11.
If a non-degenerate subinterval of Ic can me mapped into a subinterval of Ic using
elements that do not include the base b and its dual, we also get a non-trivial word
that acts trivially on a non-degenerate segment.
We erase the base and its dual that Ic supports, and apply the first part the
46
Rips machine. The Euler characteristic of the remaining pseudogroup increases
(Makanin’s complexity decreases). As we argued after erasing Ib , the application
of the first part of the Rips machine terminates after finitely many steps. If Ic
supports parts of more bases, the supports of the subintervals that are erased are
disjoint, and they must have trivial intersection with the subintervals that can be
mapped into Ib .
We continue iteratively. After erasing Ic and the subintervals that are erased
after the application of the first part of the Rips machine, there are still subintervals
that covered at least twice, and be mapped into Ib , hence, belong to the simplicial
part of Y∞ . Hence, after finitely many entire transformations (the second part of
the Rips machine), there must exist a new base that is identified with its dual.
Each time such a base and its dual are identified, they are removed and the Euler
characteristic increases (Makanin’s complexity decreases). Hence, this process has
to stop after finitely many steps. By the arguments that we already used, when
it stops there must be exist a non-trivial word in the elements of the original
pseudogroup Ût , that acts trivially on a non-degenerate segment, a contradiction
to part (2) in lemma 2.11. Therefore, all the long elements in the pseudogroup Ût
must be covered exactly twice, except perhaps at their endpoints.
So far we can deduce that in the pseudogroups Ût (perhaps after passing to a
further subsequence), every element that belongs to the highest equivalence class
is covered exactly twice. Suppose that there is an element that belongs to an
intermediate class, i.e., not to the highest class and not to the class that contains
the short and secondary short elements, that is covered more than twice.
In that case we use the procedure that was applied in the proof of proposition
2.16 in that case. After applying this procedure we replace the generators of the
pseudogroup Ût (possibly after passing to a subsequence) by generators in which
the highest class contains the intermediate class that by assumption contains an
element that is covered more than twice. By the argument that we presented above,
the highest class can not contain such elements. Hence, in the pseudogroups Ût all
the elements in all the equivalence classes that are longer than the lowest one, that
contains short and secondary short elements, are covered exactly twice. Since we
already treated this case, the conclusion of the proposition follows.
In a similar way to the IET case, Proposition 2.18 replaces proposition 2.16 in
the general case (i.e., when there is no periodicity assumption). To obtain the same
conclusions as in proposition 2.17, we further modify the tuples, ut1 , . . . , utg in a
similar way to what we did in proposition 2.8.
Proposition 2.19. With the notation of proposition 2.16, for a subsequence of the
indices t, that for brevity we still denote t, it is possible to further modify the tuple
of elements ut1 , . . . , utg , by performing Dehn twists on some of the semi-periodic
elements (the elements ut1 , . . . , utℓ that satisfy part (8) in proposition 2.18), so that
there exists a constant C > 0, for which for the modified tuples, that we still denote:
ut1 , . . . , utg , for every index t the words zjt , 1 ≤ j ≤ r, can be replaced by words: ẑjt
that satisfy properties (1)-(3) in proposition 2.17.
Proof: The proof is similar to the proof of proposition 2.8, though it needs to be
modified since in the proof of proposition 2.8 we used the fact that a surface group
is freely indecomposable.
47
Suppose that such a constant C does not exist (for any possible application of
Dehn twists on the semiperiodic elements in the tuples: ut1 , . . . , utg ). Then for
every positive integer m, there exists an index tm , so that for every possible choice
of Dehn twists to the semiperiodic elements in the tuple, ut1m , . . . , utgm , at least one
of the words ẑjtm that satisfy parts (1) and (2) (in proposition 2.5), part (3) is false
for the constant C = m.
For each index tm , we denote by lengthm the minimal length of a long element.
For each semi-periodic element ut1m , . . . , utℓm we denote the length of its period by
i
lperm
.
lperi
For each index m, and every i, 1 ≤ i ≤ ℓ, we look at the ratios: lengthmm . We
can pass to a subsequence of the indices m, for which (up to a change of order of
lperi
indices): 0 < ǫ < lengthmm for some positive ǫ > 0, and i = 1, . . . , ℓ′ . And for every
lperi
i, ℓ′ < i ≤ ℓ, the ratios lengthmm approaches 0. We perform Dehn twists along the
semiperiodic elements ut1m , . . . , utℓm
′ , so that all these semiperiodic elements have
lengths bounded by a constant times the length of a long element. These elements
will be treated as long elements and not as semiperiodic elements in the sequel.
First, suppose that ℓ′ = ℓ, i.e., that there exists an ǫ > 0, such that for every
lperi
i, 1 ≤ i ≤ ℓ, 0 < ǫ < lengthmm . In that case, after applying Dehn twists to the
semiperiodic elements, all the elements ut1m , . . . , utgm are either long or secondary
short. By the argument that was used to prove proposition 2.5, either:
(i) the number of dual positions of the different elements is globally bounded
(for the entire subsequence {tm }).
(ii) there exists a subsequence (still denoted {tm }), and a fixed positive word
in the free group F that has roots of unbounded order.
Since there is a bound on the order of a root of a fixed element in a free group, part
(ii) does not happen. Therefore, in case ℓ = ℓ′ the conclusion of the proposition
follows as in the bounded periodicity case, and the same argument remains valid if
ℓ′ < ℓ but the lengths of the semiperiodic elements ut1m , . . . , utℓm can be bounded
by a constant times the length of a long element.
Suppose that there exists a subsequence of indices (still denoted {tm }), for which
along a paired subpaths p1 and p2 , at least one of the appearances of a semiperiodic
element utim , ℓ′ < i ≤ ℓ (along p1 or p2 ), overlaps with an unbounded number of
elements (along p2 or p1 in correspondence). In that case, there exists a subsequence
(still denoted {tm }), and a fixed positive word in a positive number of either long
elements or semiperiodic elements and possibly some secondary short elements,
which is a periodic word, and the ratio between the length of the period and the
length of the element that is represented by the positive word approaches 0.
By a theorem of P. Reynolds [Re] if a f.g. group G acts indecomposably on a
real tree, and H is a f.g. subgroup of G that acts indecomposably on its minimal
subtree, then H is finite index in G. Hence, if F acts on a real tree and the action
is of Levitt type, the action of F extends to an indecomposable action of a group
G, then F is of finite index in G, The action of G is of Levitt type as well, and in
particular, G is free.
A given f.g. free group can be a finite index subgroup in finitely many free groups
(that are all of strictly smaller rank). In particular, if the Levitt action of F extends
to an indecomposable action of G, then there is a bound on the index of F in G
48
(the bound depends only on the rank of F ).
Since F is of finite index in the group that include both F and the period, F
contains a subgroup of bounded index in the subgroup that is generated by the
period. Hence, if ℓ′ < ℓ and there is no bound on the number of elements that
overlap with a semi-periodic element, there exists an element in the free group F
with an unbounded root, a contradiction.
Therefore, there exists a global bound on the number of elements that overlap
with a semiperiodic element that appears along a paired subpaths p1 and p2 . In
this last case, once again either part (i) or part (ii) hold, and since a free group
contain no non-trivial elements with a root of unbounded order, part (i) holds.
Given part (i), i.e., a global bound on the number of dual positions between two
overlapping elements, the proposition follows by the same argument that was used
to prove proposition 2.17 (in the bounded periodicity case).
Given propositions 2.18 and 2.19, that generalize propositions 2.15 and 2.16 and
are the analogue of propositions 2.7 and 2.8 in the Levitt case, the rest of the proof
of theorem 2.10 (the Levitt case) follows precisely as in the bounded periodicity
case, and precisely as in the IET case.
§3. The (Canonical) JSJ decompositions of (some) Pairs
In theorem 1.1 we have shown that with any given f.g. semigroup, S, and its
set of homomorphisms into the free semigroup F Sk , Hom(S, F Sk ), there is an
associated (canonical) finite collection of pairs, (S1 , L1 ), . . . , (Sm , Lm ), where each
pair consists of a limit group, Li , and a semigroup, Si , that is embedded in the limit
group Li (as a subsemigroup), and generates Li as a group. By the construction of
the pairs, (Si , Li ), each of them is obtained as a (maximal) limit from a sequence
of homomorphisms from the set, Hom(Si , F Sk ).
Once we associated the canonical set of maximal pairs, (S1 , L1 ), . . . , (Sm , Lm ),
with a f.g. semigroup S, to analyze the structure of the set of homomorphisms,
Hom(S, F Sk ), it is sufficient (and equivalent) to analyze the set of homomorphisms
of each of the limit groups, Li , into the free group, Fk , that restrict to homomorphisms of the subsemigroups, Si , into the free semigroup, F Sk . We denote this set
of homomorphisms (of pairs), Hom((Si , Li ), (F Sk , Fk )).
In the case of free groups, to analyze the set of homomorphisms, Hom(L, Fk ),
where L is a free group, we used Grushko free decomposition to factor L into a
free product, and then associated the canonical JSJ decomposition with each of the
factors. With the JSJ decomposition we used its associated modular group to twist
(or shorten) homomorphisms, that allowed us to associate finitely many (maximal)
shortening quotients with the limit group. Repeating this procedure iteratively we
obtained the Makanin-Razborov diagram, in which every path (called a resolution)
terminates in a free group, and the set of homomorphisms from a free group into
the coefficient group Fk , can be naturally presented as (Fk )s , where s is the rank
of the free group.
In the case of a semigroup, the geometric tools that are needed in order to analyze
the set of homomorphisms, Hom((Si , Li ), (F Sk , Fk )), are based on the analogous
tools for groups, but they need to be further refined, as the modular automorphisms
that can be used to modify (shorten) automorphisms are required to ensure that
49
the image of the subsemigroup Si remains a subsemigroup of the standard free
semigroup F Sk . To ensure that we analyzed axial and IET actions of groups (or
rather pairs) on oriented trees in the previous section.
As in analyzing homomorphisms into a free group, the basic object that we need
to associate with a pair, (S, L), where L is a freely indecomposable limit group and
S is a subsemigroup of L that generates L, is an analogue of a JSJ decomposition.
Unfortunately, we manage to construct a direct analogue of the JSJ decomposition for groups only under further restrictions on a pair (S, L). For general pairs
(S, L) in which the limit group L is freely indecomposable, we replace the JSJ
decomposition with finitely many sequences of decompositions, i.e., with finitely
many resolutions or towers.
For pairs for which we construct a JSJ decomposition, the JSJ decomposition is
(at least partly) canonical, but unlike the group analogue it is not unique, i.e., we
associate a (canonical) finite collection of decompositions with a pair (S, L).
Let (S, L) be a pair consisting of a freely-indecomposable limit group, L, and a
subsemigroup S of L that generates L as a group. We look at all the sequences of
homomorphisms, {hn : L → Fk }, that restrict to semigroup homomorphisms of S
into F Sk , and converge into the pair (S, L). By a theorem of F. Paulin [Pa], each
such sequence subconverges into an action of the limit group L on a real tree Y ,
and this action is stable (and even super stable in the sense of Guirardel [Gu]) by
lemma 1.3 in [Se1]. By works of M. Bestvina and M. Feighn [Be-Fe1], [Se3], and
V. Guirardel [Gu], with a superstable action of the limit group L on the real tree
Y it is possible to associate (canonically) a graph of groups decomposition.
Therefore, with the pair (S, L) we can associate a collection of graphs of groups
decompositions of L, i.e., those graphs of groups that are associated with actions
of L on real trees, where these actions are obtained as a limit from convergent
sequences of homomorphisms. Note that since all of these graphs of groups are
abelian decompositions, they can be all obtained from the abelian JSJ decomposition of the limit group L, by cutting QH vertex groups along some s.c.c. and then
possibly collapse and fold some parts of the obtained graphs of groups.
Also, note that if a cyclic subgroup C of L stabilizes a non-degenerate segment in
a real tree that is obtained as a limit of actions of L that correspond to homomorphisms of L into Fk , then the unique maximal cyclic subgroup of the limit group
L that contains C stabilizes this segment as well. Similarly, if an abelian subgroup
A < L stabilizes a non-degenerate segment in such a real tree, then the unique direct summand that contains A as a subgroup of finite index, in the unique maximal
abelian subgroup of L that contains A, stabilizes this non-degenerate segment as
well.
On these graphs of groups we can naturally define a partial order. We say that
given two graphs of groups, Λ1 and Λ2 , Λ1 > Λ2 , if Λ1 is a proper refinement
of Λ2 , or alternatively, Λ2 is obtained from Λ1 by (possibly) cutting some QH
vertex groups along a finite collection of s.c.c. and then (possibly) performing some
collapses and foldings.
Proposition 3.1. Let (S, L) be a pair of a freely indecomposable limit group L, and
its subsemigroup S that generates L. Then there exist maximal abelian decompositions of the pair (S, L). Every strictly increasing sequence of abelian decompositions
that are associated with (S, L), Λ1 < Λ2 < Λ3 < . . . , terminates after finitely
50
many steps.
Proof: All the abelian decompositions, Λi , are obtained from the abelian JSJ
decomposition of the freely-indecomposable limit group L, by cutting QH vertex
groups along a finite (possibly empty) collection of s.c.c. and then (possibly) collapse
and fold the obtained abelian decomposition. Since there is a bound on the size
of a collection of disjoint non-homotopic non null homotopic s.c.c. on the surfaces
that are associated with the QH vertex groups in the JSJ decomposition of L, and
the abelian edge groups of the JSJ decomposition are all finitely generated, the
proposition follows.
Proposition 3.1 proves the existence of maximal elements in the set of abelian
decompositions that are associated with the pair, (S, L), with their natural partial
order. To construct an analogue of a JSJ decomposition for the pair, (S, L), we
further prove that under further assumptions on the pair (S, L), there are only
finitely many (equivalence classes) of such maximal elements.
Theorem 3.2. Suppose that L is freely indecomposable, and that all the maximal
abelian decompositions, {Λi }, that are associated with the pair (S, L), correspond
to simplicial actions of (S, L) on real trees. Then there exist only finitely many
(equivalence classes of ) maximal abelian decompositions of the pair (S, L).
Proof: Suppose that there are infinitely many (equivalence classes of) maximal
abelian decompositions of a pair (S, L). Let {Λi }∞
i=1 be the collection of these
(inequivalent) maximal decompositions. Under the assumptions of the theorem,
they are all simplicial.
With the maximal pair (S, L) we associate a sequence of homomorphisms, {hn :
(S, L) → (F Sk , Fk )}, that converges into (S, L). To define the sequence of homomorphisms, we fix a (symmetric) generating set of the limit group L, {g1 , . . . , gm },
that contains a generating set, s1 , . . . , sr , of the semigroup, S. With the given
generating set of the pair, (S, L), we naturally associate its Cayley graph, that we
denote X. For every positive integer n, we denote the ball of radius n in the Cayley
graph X, Bn .
Each maximal decomposition, Λi , that is associated with the pair (S, L), is obtained from a sequence of homomorphisms, {fi (j)}∞
j=1 , from (S, L) into (F Sk , Fk ).
For each index i, the sequence of actions of (S, L) on the Cayley graph of (F Sk , Fk ),
that are associated with the sequence of homomorphisms, {fi (j)}∞
j=1 , converges in
the Gromov-Hausdorff topology, after rescaling the metric so that the maximal
length of the image of a generator (under fi (j)) is 1, to an action of the pair (S, L)
on a real tree, that we denote Ti .
For each index n, we choose a homomorphism hn : (S, L) → (F Sk , Fk ) that
satisfies the following conditions:
(1) hn is one of the homomorphisms, {fn (j)}∞
j=1 .
(2) let g be a word of length at most n in the fixed set of generators, g1 , . . . , gm ,
of the limit group L. Then hn (g) = 1 if and only if g = 1 in L.
(3) Let Yn be the Cayley graph of (F Sk , Fk ) after rescaling the metric so that
the maximal length of the image of a generator (under hn ) is 1, and let tn
be the base point in Tn . Then for every ℓ1 , ℓ2 ∈ Bn :
1
1
≤ dYn (hn (ℓ1 ), hn (ℓ2 )) ≤ dTn (ℓ1 (tn ), ℓ2 (tn )) + .
dTn (ℓ1 (tn ), ℓ2 (tn )) −
n
n
51
With each homomorphism hn , there is naturally an associated action of the
pair (S, L) on the simplicial tree Yn that is obtained from the Cayley graph of the
coefficient free group Fk by rescaling the metric so that the maximal length of the
image of a generator (under hn ) is 1. From the sequence of homomorphisms, {hn },
it is possible to extract a subsequence, so that the sequence of actions of the pair
(S, L) on the trees {Yn }, converges into a faithful action of the pair (S, L) on a
limit real tree Y . Note that this action is precisely the limit of the corresponding
subsequence of actions of the limit group L on the limit real trees Tn .
The action of (S, L) on the limit tree Y is non-trivial, has abelian stabilizers of
non-trivial segments, and is super-stable in the sense of [Gu]. Hence, with this action it is possible to associate a (graph of groups) decomposition ∆1 with trivial and
abelian edge stabilizers. Since L was assumed to be freely indecomposable, all the
edge stabilizers in ∆1 are non-trivial abelian. Furthermore, under the assumptions
of the theorem, the abelian decompositions, {Λi } and ∆1 , must be simplicial, and
they are all dominated by the JSJ decomposition of the (freely indecomposable)
limit group L.
We divide the edges in the abelian decomposition ∆1 , and the vertices that are
adjacent to these edges into families. The stabilizer of every edge group in ∆1 is an
abelian subgroup of L, and each vertex group in ∆1 is finitely presented. We fix a
finite generating set for each of the vertex groups in ∆1 .
Definition 3.3. Let E be an edge in ∆1 , and let A be its edge group. We say that
an edge group A in ∆1 is elliptic, if every element a ∈ A, is elliptic in almost all the
abelian decompositions Λn (i.e., in all but at most finitely many decompositions).
Otherwise we say that A is hyperbolic.
Suppose that A is a hyperbolic edge group and let V be a vertex group that is
adjacent to the edge E. We fix a generating set v1 , . . . , vℓ of V . For an element
f ∈ Fn we denote the length of the conjugacy class of f by |f |. We say that the edge
group A is periodic in the vertex group V , if for every element a ∈ A, there exists
a positive constant ca > 0, such that for almost every index n (i.e., for all except
finitely many values of n), every point in the Cayley graph T of the coefficient group
Fk is moved by at least one of the elements, hn (v1 ), . . . , hn (vℓ ), a distance of at
least ca · n · |hn (a)|.
We say that a hyperbolic edge group A is non-periodic in V , if there exists a
non-trivial element a ∈ A, and a positive constant ca > 0, so that for all but finitely
many indices n, there exists a point in the Cayley graph T of the coefficient group
Fk that is moved by each of the elements, hn (v1 ), . . . , hn (vℓ ), a distance that is
bounded by ca · |hn (a)|.
By passing to a subsequence of the maximal abelian decompositions, {Λn }, we
may assume that every edge group in ∆1 is either elliptic or hyperbolic. By passing
to a further subsequence, we may assume that every hyperbolic edge group in ∆1
is either periodic or non-periodic in the one or two vertices that are adjacent to it.
We continue with such a subsequence of the abelian decompositions {Λn }.
At this point we gradually refine the abelian decomposition ∆1 . Let V be a vertex
group in ∆1 , that is connected only to either elliptic or periodic edge groups. If V
is not elliptic in a subsequence of the abelian decompositions,{Λn }, we pass to this
subsequence, and analyze the actions of the f.g. subgroup V on the Cayley graph
of the coefficient group Fk via the homomorphisms {hn }. By passing to a further
52
subsequence, these actions of V do converge into a non-trivial action of V on a real
tree. Since all the edge groups that are connected to V in ∆1 are either elliptic or
periodic, the abelian decomposition that is associated with the limit action of V
can be further extended to an abelian decomposition of the pair (S, L), that strictly
refines the abelian decomposition ∆1 .
We repeat this refinement procedure for every non-elliptic vertex group in the
obtained (refined) abelian decomposition, that is connected only to elliptic or periodic edge groups, and is a point stabilizer in the corresponding action on a real
tree, i.e., that is not a vertex group that is associated with an IET or an axial
component. By the accessibility for small splittings of f.p. groups [Be-Fe1] (or alternatively, by acylindrical accessibility ([Se],[De],[We])), this refinement procedure
terminates after finitely (in fact boundedly) many steps, and we obtain an abelian
decomposition that we denote ∆2 . In ∆2 every vertex group V that is connected
only to elliptic and periodic edge groups, is either associated with an IET or an
axial component, or it is elliptic itself, in which case all the edge groups that are
connected to it are elliptic as well.
If all the vertex groups in ∆2 are elliptic, then for almost all the indices n,
the abelian decomposition Λn is dominated by the abelian decomposition ∆2 (i.e.,
∆2 is a (possibly trivial) refinement of Λn for almost every index n). But in this
case ∆2 dominates only finitely many non-equivalent abelian decompositions, a
contradiction to the existence of the infinite sequence of non-equivalent maximal
abelian decompositions {Λn } (from which ∆2 was obtained).
If ∆2 contains a vertex that is associated with an axial or an IET component.
Note that such a vertex group is connected to only elliptic and periodic edge groups.
Then there exists a sequence of homomorphisms of pairs that converges into a nonsimplicial faithful action of L on a real tree (and contradicts the assumption of
theorem 3.2).
Proposition 3.4. Suppose that ∆2 contains a vertex group that is associated with
either an axial or an IET component. Then there exists a sequence of homomorphisms of pairs: {νn : (S, L) → (F Sk , Fk )} that converges into a faithful action of
L on a real tree, and this action contains either an axial or an IET component.
Proof: Suppose that ∆2 contains a vertex group A that is associated with an
axial component. In that case L = V ∗A0 A. By theorem 2.1 the vertex group
A can be written as A = A0 + < a1 , . . . , aℓ >, where A0 is the point stabilizer of
the axial component, ℓ ≥ 2, there exists some index n0 so that for every n > n0 ,
hn (ai ) ∈ F Sk , and for each j, 1 ≤ j ≤ r, sj can be written as a word:
sj = v1j w1j v2j w2j . . . vbjj wbjj
where each of the elements wij is a positive word in the basis elements a1 , . . . , aℓ ,
vij ∈ V , and for every index n > n0 , hn (vij ) ∈ F Sk .
In that case, for every n > n0 , we can modify the homomorphisms hn , by
preserving the images of elements in the vertex group V , and modifying the images
of the elements a1 , . . . , aℓ . By the properties of the homomorphisms hn , for every
n > n0 we can set νn (ai ) to be arbitrary elements in F Sk that commute with
hn (A0 ). Clearly, we can choose these images of hn (a1 ), . . . , hn (aℓ ), so that in the
limit they will be independent over the rationals, and so that in the limit the
53
whole vertex group V will stabilize a point. Therefore, the limit of the constructed
homomorphisms {νn } will contain a single axial component with an associated
group A. In particular the limit action is faithful and contains an axial component.
Suppose that ∆2 contains an IET component. Let Q be a (hyperbolic) surface
group and let (S, Q) be a pair. Let {hn : (S, Q) → (F Sk , Fk )} be a sequence of pair
homomorphisms that converges to a free action of Q on a real tree. Let s1 , . . . , sr
be a given set of generators of the subsemigroup S.
If Q, the subgroup that is associated with the IET component in ∆2 is a (closed)
surface group, then since L is freely indecomposable, L = Q, and L admits an IET
action on a real tree, so proposition 3.4 follows. Hence, we may assume that Q is
a punctured surface group. Let Q be a punctured surface group that is associated
with an IET component in ∆2 . Let VQ be the vertex that is stabilized by Q in ∆2 .
By proposition 2.4 it is possible to find a sequence of tuples of elements {ut1 , . . . , utg }
such that:
(1) the tuples belong to the same isomorphism class, and satisfy the properties
that are listed in proposition 2.4.
(2) for a fixed t, for large enough index n, hn (uti ) ∈ F Sk , 1 ≤ i ≤ g.
(3) the elements uti generate some natural extension of Q, that we denote Q̂.
The generators s1 , . . . , sr can be written as words in elements of Q̂ that we
denote v1 , . . . , vf , and elements that lie outside the vertex that is stabilized
by Q in ∆2 .
(4) for each t, and each index j, 1 ≤ j ≤ f , there exists a word ẑjt , such that
vj = ẑjt (ut1 , . . . , utj ).
(5) For every pair of indices t1 , t2 , and every j, 1 ≤ j ≤ f , the elements
hn (ẑjt1 (ut12 , . . . , utj2 )) ∈ F Sk for large enough index n.
Given an index t > 1, we define an automorphism of the natural extension
of Q that restricts to an automorphism of Q, by sending the tuple of generators
ut1 , . . . , utg to u11 , . . . , u1g . We denote this automorphism (of the natural extension
of Q) ψt . ψt maps each of the elements vj , 1 ≤ j ≤ f to: ψt (vj ) = ẑjt (u11 , . . . , u1j ).
By part (5), for large enough n, hn ◦ ψt (vj ) ∈ F Sk , 1 ≤ j ≤ f . Suppose that for
every index t we choose an index nt , such that nt grows to ∞, and hnt ◦ ψt (vj ) ∈
F Sk , 1 ≤ j ≤ f . A subsequence of the sequence hnt ◦ ψt converges into an action of
Q on some real tree. By construction and the properties of the elements ut1 , . . . , utg
(that are listed in proposition 2.4), this action is faithful, and the action has to be
bi-Lipschitz equivalent to the limit IET action of Q on a real tree that is obtained
from the restrictions of the homomorphisms {hn } to Q. In particular, the limit
action has to be an IET action of Q on a real tree.
The restrictions of the automorphisms ψt to Q extend naturally to automorphisms of the limit group L (viewed as the fundamental group of the abelian decomposition Λ). We denote these automorphisms of L, ϕt . By construction, the
sequence of homomorphisms hnt ◦ϕt has a subsequence that converges into a faithful
action of L on a real tree, and this action contains an IET component that is associated with the action of the subgroup Q. This concludes the proof of proposition
3.4.
After constructing the refined abelian decomposition ∆2 , we use (part of) its
modular group to shorten the sequence of homomorphisms, {hn }, that were used
54
to construct the abelian decompositions ∆1 and ∆2 , to obtain a new abelian decomposition of the pair (S, L). After repeating this shortening procedure iteratively finitely many times, we are able to replace the ”machine” that is used in
constructing the JSJ decomposition for groups [Ri-Se], and by the existence of such
a machine, eventually deduce the finiteness of the maximal decompositions of the
given pair (S, L).
In the abelian decomposition ∆2 no non-elliptic vertex group is connected to
only elliptic and periodic edge groups. We fix a finite set of generators for each
of the vertex groups in ∆2 . We order the vertex groups in ∆2 by the order of
magnitude of the displacements of their fixed sets of generators. Let v1 , . . . , vℓ be
the fixed set of generators of a vertex group V . For each index n, we associate with
V the minimal displacement of a point in the Cayley graph T of the coefficient
group Fk , under the action of the tuple of elements hn (v1 ), . . . , hn (vℓ ). We denote
this minimal displacement dispn (V ).
After passing to a subsequence, and up to certain equivalence relation on the
displacement functions, the displacement functions define an order on the vertex groups in ∆2 . We say that two displacement functions, dispn (V1 ), dispn (V2 ),
are comparable if there exists positive constants c1 , c2 , so that for every index
n, c1 dispn (V1 ) < dispn (V2 ) < c2 dispn (V1 ). We say that dispn (V1 ) dominates
dispn (V2 ) if dispn (V2 ) = o(dispn (V1 ).
Since there are only finitely many vertex groups in ∆2 , there is a finite collection
of vertex groups in ∆2 with comparable displacement functions, that dominate the
displacement functions of all the other vertex groups in ∆2 . Note that none of the
vertex groups in the dominating set can be elliptic. By our construction of ∆2 ,
for each vertex group in this dominating subset of vertex groups, there exists a
non-periodic edge group that is connected to it. Furthermore, a non-periodic edge
group that is connected to a vertex group in the dominating set , must be connected
only to vertex groups in the dominating subset.
We set M od(∆2 ) to be the modular group of the pair (S, L) that is associated
with the abelian decomposition ∆2 . We set M XM od(∆2 ) to be the subgroup of
M od(∆2 ), that is generated by Dehn twists only along non-periodic (hyperbolic)
edge groups for which their corresponding edges connect between dominating vertex
groups.
In the graph of groups ∆2 , the edges with non-periodic hyperbolic edge groups
that connect between dominating vertex groups, are grouped in several connected
components, that we denote: Γ1 , . . . , Γu . We fix finite sets of generators for the
fundamental groups of each of the connected subgraphs, Γ1 , . . . , Γu .
For each index n, we replace the homomorphism, hn , with a homomorphism,
1
hn , that has similar properties to those of hn , and choose an automorphism ϕn ∈
M XM od(∆2 ), so that h1n , ϕn and the twisted homomorphism, h2n = h1n ◦ ϕn , have
the following properties:
(1) Let Tn be the Bass-Serre tree that is associated with the abelian decomposition Λn . ϕn is chosen so that:
(i) for large enough j, fn (j) ◦ ϕn : L → Fk , is a homomorphism of the
pair (S, L) into the pair (F Sk , Fk ) (i.e., fn (j) ◦ ϕn sends the positive
cone in L into the standard positive cone F Sk in Fk ).
(ii) for each connected component, Γi , i = 1, . . . , u, there exists a point
ti (n) ∈ Tn , so that the displacement of ti (n) under the action of the
55
fixed finite collection of generators of the fundamental group of Γi
twisted by ϕn , is the shortest among all the points in Tn and all the
automorphisms ϕ ∈ M XM od(∆2 ) that satisfy property (i).
(2) Having fixed ϕn , we choose h1n , to be one of the homomorphisms fn (j), for
which both h1n and h1n ◦ ϕn satisfy the conditions that hn was required to
satisfy, with respect to the given action of L on Tn , λn : L × Tn → Tn , and
with respect to the twisted action λn ◦ ϕn : L ×Tn → Tn , in correspondence.
The sequence of homomorphisms, {h1n }, converges into a faithful action of the
limit group L on the same real tree as the sequence, {hn }, i.e., to the action of
L on the real tree Y . With the action of L on Y we have associated the abelian
decomposition, ∆2 . We set h2n = h1n ◦ ϕn . As the automorphisms ϕn were chosen to
be from the modular group M XM od(∆2 ), the sequence h2n converges into a faithful
action of L on a real tree Y1 .
With the action of L on the real tree Y1 we can naturally associate an abelian
decomposition ∆3 . Since we assumed that all the maximal abelian decompositions
that are associated with the pair (S, L) are simplicial, the abelian decomposition ∆3
has to be simplicial as well. Starting with ∆3 , we can possibly successively refine
it and obtain an abelian decomposition ∆4 , in the same way that we successively
refined the abelian decomposition ∆1 and obtained the abelian decomposition ∆2 .
By construction, all the edges in ∆2 that are in the complement of the connected
subgraphs, Γ1 , . . . , Γu , remain edges in ∆4 . Hence, the edge groups of these edges
are elliptic in ∆4 . Unlike the abelian decomposition ∆2 , it may be that all the
vertex and edge groups in ∆4 are elliptic. Since we assumed that all the maximal
abelian decompositions of the pair (S, L) are simplicial, proposition 3.4 implies that
it can not be that ∆4 contains an axial or a QH vertex group.
Since ∆3 is obtained from ∆2 by shortening edges that are stabilized by nonperiodic edge groups and connect between dominating vertex groups, the edge
groups in the connected subgraphs, Γ1 , . . . , Γu , are not contained in vertex groups
in ∆4 . Therefore, there exists at least one edge group in ∆2 that is not elliptic in
∆4 .
The abelian decomposition ∆2 of the limit group L was obtained from an action
of L on a real tree Y , that is itself obtained as a limit of a sequence of homomorphisms of L into the free group Fk . Hence, every edge group in ∆2 that is
associated with a stabilizer of a non-degenerate segment in the real tree Y , is either
the centralizer of itself, or it is a direct summand in its centralizer.
We start by assuming that all the edge groups in ∆2 that have cyclic centralizers
can be conjugated into vertex groups in ∆4 . Let E1 be an edge in ∆2 with a
non-periodic abelian edge group A1 , so that the edge E1 is contained in one of
the connected subgraphs of groups Γ1 , . . . , Γu (so E1 connects between dominating
vertex groups, and A1 is not contained in a vertex group in ∆4 ). By our assumption
the centralizer of A1 in L is non-cyclic. First, we further assume that A1 is a strict
direct summand in its centralizer. Since the action of L on the real tree Y is assumed
to be simplicial, and A1 stabilizes pointwise a line in Y , ∆2 contains a circle so that
all the edge groups in this circle are stabilized by A1 , and the Bass-Serre generator
that is associated with that circle, that we denote t1 , commutes with A1 , and the
centralizer of A1 is the direct sum of A1 and < t1 >.
Since the abelian decompositions, ∆3 and ∆4 , were obtained by shortening the
homomorphisms, {h1n }, using the subgroup of modular automorphisms, M XM od(∆2 ),
56
that shortens only edges with non-periodic edge groups that connect between dominating vertex groups, the maximal abelian subgroup A = A1 + < t1 >, and in
particular its subgroup A1 , do not fix a vertex in ∆4 . Hence, there is a circle in ∆4
so that all the edges in this circle are stabilized by a direct summand A2 < A, and
A1 ∩ A2 is a direct summand of A that is also a strict summand in both A1 and A2 .
Therefore, there exists an element t2 < A1 , such that A1 = (A1 ∩ A2 )+ < t2 >.
If there are no vertex groups along the circle that is stabilized by A1 in ∆2 , or
along the circle that is stabilized by A2 in ∆4 , then the limit group L has to be a
non-cyclic free abelian group. In that case L admits an axial action on a real tree
that is obtained as a limit of a sequence of homomorphisms of pairs from (S, L)
into (F Sk , Fk ), and such an action contradicts our assumption that every action of
(S, L) that is obtained as a limit of a sequence of homomorphisms from (S, L) into
(F Sk , Fk ) is simplicial. Hence, we may suppose that the circle that is stabilized by
A1 in ∆2 and by A2 in ∆4 contain non-trivial vertex groups.
If A1 ∩ A2 is the trivial subgroup of A, then the fundamental groups of the
connected components that are obtained from ∆2 by deleting the circle that is
stabilized by A1 , inherit non-trivial free decompositions from ∆4 , and these free
decompositions extend to a (non-trivial) free decomposition of the ambient limit
group L, contradicting our assumption that L is freely indecomposable. Hence, we
may assume that A1 ∩ A2 is a non-trivial direct summand of A.
Let V1 , . . . , Vℓ be the vertex groups in ∆2 , that are placed in the circle that is
stabilized by A1 in ∆2 . Each of these vertex groups inherits a graph of groups
decomposition from the abelian decomposition (of the ambient group L) ∆4 . Each
of these graphs of groups of the vertex groups, V1 , . . . , Vℓ , contains a circle, so that
each edge in that circle is stabilized by the direct summand A1 ∩ A2 .
The limit group L is assumed to be freely indecomposable, and the centralizer
of the edge group A1 ∩ A2 is the non-cyclic abelian subgroup A. Hence, by the
existence of an abelian JSJ decomposition for limit groups, an edge group C in
∆2 , which is not conjugate to A1 , and is a subgroup of a vertex group Vi (that is
contained in the circle that is stabilized by A1 in ∆2 ), is elliptic with respect to a
decomposition of L along edges that are stabilized by A1 ∩ A2 . Therefore, C, can
be conjugated into a connected subgraph of the graph of groups that Vi inherits
from ∆4 , where this subgraph does not contain any of the edges in the circle that
is stabilized by A1 ∩ A2 in the graph of groups that is inherited by Vi .
Each of the given set of generators s1 , . . . , sr of the subsemigroup S of the limit
group L, can be written in a normal form with respect to the graph of groups ∆2 .
These normal forms contain elements from the vertex groups V1 , . . . , Vℓ of ∆2 . We
can further write each of the elements of V1 , . . . , Vℓ , that appear in the normal forms
of s1 , . . . , sr , in a normal form with respect to the graph of groups that the vertex
groups V1 , . . . , Vℓ inherit from the abelian decomposition ∆4 . If we substitute each
of these last normal forms in the normal forms of s1 , . . . , sr with respect to ∆2 ,
we represented each of the elements s1 , . . . , sr as (fixed) words in vertex groups
and Bass-Serre generators of ∆2 and of the vertex groups in the graphs of groups
that are inherited by V1 , . . . , Vℓ from ∆4 . In particular, these fixed words contain
powers of the elements t1 and t2 that are contained in the abelian subgroup A, that
are Bass-Serre generators in ∆2 and ∆4 , in correspondence.
At this point we can finally modify the sequence of homomorphisms, {h1n }, in
order to get a new sequence of homomorphisms of the pair (S, L) into the pair
(F Sk , Fk ) that converges into a faithful action of the pair (S, L) on a real tree, and
57
this action contains an axial component, a contradiction to our assumption that
every such limit action is simplicial.
With each of the elements, s1 , . . . , sr , we have associated a (fixed) word, that
was constructed from a normal form of the element si with respect to the abelian
decomposition ∆2 , and normal forms of elements from the vertex groups V1 , . . . , Vℓ
with respect to the graphs of groups that these vertex groups inherit from ∆4 .
These words contain powers of the elements t1 and t2 . We set d to be the sum of
the absolute values of powers of the element t2 that appear in all the (fixed) words
that we have associated with the elements s1 , . . . , sr .
We choose ℓ−1 positive irrational numbers, α1 , . . . , αℓ−1 , so that 1, α1 , . . . , αℓ−1
are independent over the rationals, and additional positive real numbers αℓ and β
so that:
(1) α1 + α2 + . . . + αℓ = 1.
(2) for every i, 1 ≤ i ≤ ℓ:
1
1
.
|αi − | <,
ℓ
10ℓ
1
1
(3) 4dℓ
< β < 3dℓ
.
(4) β is not in the subspace that is spanned by 1, α1 , . . . , αℓ−1 over the rationals.
Given the positive irrational numbers, α1 , . . . , αℓ and β, we modify the sequence
of homomorphisms {h1n }. First, for every index n, we precompose the homomorphism h1n with a modular automorphism ψn ∈ M XM od(∆2 ). The automorphisms
ψn only conjugate the vertex groups V2 , . . . , Vℓ , and the vertex groups that are
connected to them in ∆2 and are located outside the circle that is stabilized by
A1 , by different powers of the element t2 . The conjugations that determine the
automorphisms {ψn }, are chosen to guarantee that in the limit tree that is obtained from the sequence of homomorphisms {h1n ◦ ψn }, by using the same rescaling constants as those that were used for the sequence {h1n }, the vertex groups
V1 , . . . , Vℓ fix points, and the distances in the limit tree (that we denote) Ŷ satisfy,
dŶ (F ix(Vi ), F ix(Vi+1 )) = αi tr(t1 ), for every i, i = 1, . . . , ℓ − 1, and where tr(t1 )
is the displacement of the element t1 along its axis (that is stabilized by A1 ) in Ŷ .
Since the homomorphisms {h1n } are homomorphisms of the pair (S, L), for large
enough index n, the homomorphisms h1n ◦ ψn are homomorphisms of the pair (S, L)
as well, and in the limit, we get a faithful action of the pair (S, L) on the real tree Ŷ
(note that the action of (S, L) on the real tree Ŷ is similar to the action of (S, L) on
the real tree Y , where the difference is only in the lengths of the segments between
the points that are stabilized by the vertex groups V1 , . . . , Vℓ ).
At this point we further modify the homomorphisms {h1n ◦ ψn }, by changing
only the image of the element t2 . Each of the vertex groups, V1 , . . . , Vℓ , inherits
an abelian decomposition from the graph of groups ∆4 . In each of these inherited
graphs of groups there is a circle that is stabilized by the subgroup A1 ∩ A2 , and
with a Bass-Serre generator t2 . Hence, we can modify each of these graphs of groups
of the vertex groups, V1 , . . . , Vℓ , and replace the circle that is stabilized by A1 ∩ A2 ,
with a new vertex group that is stabilized by A1 , that is connected to the vertex
Vℓ with an edge that is stabilized by A1 ∩ A2 .
Therefore, each of the vertex groups, V1 , . . . , Vℓ , can be written as an amalgamated product, Vi = Wi ∗A1 ∩A2 A1 . Furthermore, all the edge groups in ∆2 of
58
edges that are connected to the vertex that is stabilized by Vi in ∆2 can be conjugated into Wi . This implies that the graph of groups ∆2 can be modified and
collapsed to an amalgamated product: L = W ∗A1 ∩A2 A, where A is the centralizer
of A1 and A2 , and A = (A1 ∩ A2 )+ < t2 > + < t1 >.
Now, we finally modify the homomorphisms {h1n ◦ ψn } by changing the images of
the element t2 . We define a sequence of homomorphisms {h2n : (S, L) → (F Sk , Fk )}
as follows. For each index n and every element w ∈ W we set: h2n (w) = h1n ◦
ψn (w). We further set h2n (t1 ) = h1n ◦ ψn (t1 ), and h2n (t2 ) to be an element that
commutes with h1n (t2 ), and so that in the limit tree Ỹ , that is obtained from the
sequence of homomorphisms {h2n }, by using the same rescaling constants as those
that were used for the sequence {h1n }, tr(t2 ) = βtr(t1 ). By our choice of the
constants α1 , . . . , αℓ−1 and β, the homomorphisms h2n map the fixed generators
of the semigroup S, s1 , . . . , sr , to elements in the standard free semigroup F Sk ,
hence, h2n are homomorphisms of the pair (S, L) into the pair (F Sk , Fk ).
By construction, the subgroup W acts faithfully on the limit tree Ỹ , since it
acts faithfully on the real tree Ŷ , and the restrictions of the homomorphisms h2n
and h1n to the subgroup W are identical. Since β was chosen to be irrational, the
subgroup < t1 , t2 > acts indiscretely, with a dense orbit, on a line in Ỹ . Since
the real numbers, 1, α1 , . . . , αℓ−1 , β, are independent over the rationals, L modulo
the kernel of the action of it on Ỹ can be written as an amalgamated product:
W ∗A1 ∩A2 A, hence, L acts faithfully on the real tree Ỹ . Finally, the real tree Ỹ was
constructed from a sequence of homomorphisms of the pair (S, L) into (F Sk , Fk ),
that we denoted {h2n }, and Ỹ contains an axial component (the component that
contains the axis of t1 and t2 ), a contradiction to the assumption (of proposition
2.3) that every such limit action is discrete.
So far we assumed that all the edge groups in ∆2 that have cyclic centralizers can
be conjugated into vertex groups in ∆4 , and that there exists an edge in ∆2 with
a non-periodic abelian edge group A1 , that is contained in one of the connected
subgraphs, Γ1 , . . . , Γu (i.e., that this non-periodic edge group connects between
dominating vertex groups), and so that A1 is a (strict) direct summand in its
centralizer, i.e., that the centralizer of A1 is not contained in a vertex in ∆2 .
We have already pointed out that since we have started with an infinite sequence
of inequivalent graphs of groups of L, not all the edge groups in ∆2 are elliptic.
Suppose that all the edge groups in ∆2 that have cyclic centralizers, are not in
the connected subgraphs of groups, Γ1 , . . . , Γu , so they can all be conjugated into
vertex groups in ∆4 . Hence, the subgraphs Γ1 , . . . , Γu contain only edges with edge
groups that have non-cyclic centralizers. Suppose that there is no edge E1 in ∆2 ,
with an edge group A1 , for which:
(1) E1 is contained in one of the subgraphs Γ1 , . . . , Γu (hence, A1 has a noncyclic centralizer).
(2) A1 is a proper subgroup (a proper direct summand) in its centralizer.
In this case ∆4 does contain edge groups with non-cyclic stabilizers, so that their
centralizers can not be conjugated into vertex groups in ∆4 . At this stage we can
not apply the argument that we used previously and we continue as follows. If
all the edge groups in ∆4 are elliptic, then the abelian decompositions {Λn }, from
which we constructed the abelian decompositions ∆i , i = 1, . . . , 4, belong to only
finitely many equivalence classes, a contradiction to our assumptions. Hence, at
least one edge group in ∆4 is not elliptic. Therefore, we can start with ∆4 , and
59
apply the construction of ∆3 and ∆4 , to get new abelian decompositions of L, that
we denote ∆5 and ∆6 .
If there exists an edge E in ∆4 with an edge group A that has a cyclic centralizer,
so that A is non-periodic and E connects between two dominating vertex groups
in ∆4 , A can not be conjugated into a vertex group in ∆6 . In that case we show in
the sequel (proposition 3.5) that it’s possible to find a sequence of homomorphisms
of (S, L) into (F Sk , Fk ) that converges into a faithful action of L on a real tree,
and this action contains an axial or an IET component, a contradiction to the
assumptions of theorem 3.2. Hence, we may assume that all the edges in ∆4 that
have edge groups with cyclic centralizers, are either periodic or they don’t connect
between dominating vertex groups.
If there exists an edge E in ∆4 with an abelian edge group A that has a noncyclic centralizer, the centralizer can not be conjugated into a vertex group in ∆4 ,
A is non-periodic and connects between two dominating vertex groups in ∆4 , we
can apply the previous argument to the edge group A and its centralizer, and to
the two abelian decompositions, ∆4 and ∆6 , and obtain an action of L on a real
tree with an axial component, a contradiction to the assumptions of theorem 3.2.
Therefore, we may assume that all the non-periodic edge groups in ∆4 that
connect between dominating vertex groups, have non-cyclic centralizers, and their
centralizers can be conjugated into a vertex group in ∆4 . In that case, there exists
at least one edge in ∆4 and at least 2 edges in ∆6 , that have edge groups with noncyclic stabilizers, and these stabilizers can not be conjugated into vertex groups in
∆4 and ∆6 in correspondence.
By starting with ∆6 and iteratively constructing abelian decompositions, we
either obtain:
(1) an abelian decomposition in which all the edge and vertex groups are elliptic - a contradiction to the assumption that there are infinitely many
inequivalent maximal abelian decompositions of (S, L), {Λn }.
(2) and edge E with a stabilizer A that connects between two dominating vertex
group in ∆2i , such that A has non-cyclic centralizer, A is non-periodic,
and the stabilizer of A is not contained in a vertex group in ∆2i . In that
case we apply the construction that we used in case ∆2 has such an edge,
and associate a sequence of pair homomorphisms with the pair (S, L) that
converges into an action of L on a real tree and the action contains an axial
component. A contradiction to theorem 3.2.
(3) an edge group with cyclic centralizer in ∆2i , so that this centralizer can not
be conjugated into a vertex group in ∆2(i+1) - in that case we prove in the
sequel (proposition 3.5) that there exists a sequence of homomorphisms of
the pair (S, L) into (F Sk , Fk ) that converges to a faithful action of (S, L) on
a real tree, and this limit real tree contains an axial or an IET component,
a contradiction to the assumption of theorem 3.2.
(4) if cases (1)- (3) do not occur, then the abelian decomposition ∆2i contains
at least (i − 1) edges with edge groups that have non-cyclic centralizers, and
these centralizers can not be conjugated into vertex groups in ∆2i .
By the accessibility for small splittings of a f.p. group [Be-Fe], or by acylindrical
accessibility ([Se],[De],[We]), there is a global bound on the number of edges in
the abelian decompositions ∆2i , that depends only on L (in fact on the number of
generators of L [We]). Hence, if cases (1) - (3) do not occur, there is a global bound
60
on the number of steps that case (4) can occur.
Therefore, to complete the proof of theorem 3.2, we need the following basic
proposition, that motivates our approach to the construction of the JSJ decompositions for pairs.
Proposition 3.5. Suppose that for some index i, there exists an edge E in ∆2i
with an edge group A that has a cyclic centralizer, so that A is non-periodic and E
connects between two dominating vertex groups in ∆2i , and A can not be conjugated
into a vertex group in ∆2(i+1) . Then there exists a sequence of homomorphisms of
(S, L) into (F Sk , Fk ) that converges into a faithful action of L on a real tree, and
this action contains an IET component.
Proof: By the properties of the JSJ decomposition of the limit group L, the subgroup A corresponds to a s.c.c. on some maximal QH subgroup M SQ in the JSJ
of L. Hence, to prove proposition 3.5 it is possible to use geometric properties of
(oriented) curves and arcs on surfaces (i.e., on the surface that is associated with
M SQ), together with properties of the simplicial actions of L on the real trees that
are associated with the abelian decompositions ∆2i and ∆2(i+1) .
The abelian decomposition ∆2i was obtained from a limit of a sequence of homomorphisms of pairs that we denote: {hn : (S, L) → (F Sk , Fk )}. The abelian decomposition ∆2(i+1) is obtained from a sequence of shortened homomorphisms, {hsn :
(S, L) → (F Sk , Fk 0}, i.e., from the homomorphisms {hn } that are pre-composed
with automorphisms in the modular group that is associated with ∆2i (generated
by Dehn twists along the edges in ∆2i ), hsn = hn ◦ ϕn , where ϕn ∈ M XM od(∆2i.
Since the sequnce of shortened homomorphisms, {hsn }, converges into an action
of L on a real tree with an associated abelian decomposition, ∆2(i+1) , after possibly
passing to a subsequence, each of the shortened homomorphisms, hsn , can be written
as a composition: hsn = hbn ◦ νn , where:
(1) νn ∈ M od(∆2(i+1) ).
(2) hbn is a pair homomorphism: hbn (S, L) → (F Sk , Fk ).
(3) for each index n:
max length(hsn (sj )) ≥ n · max length(hbn (sj )).
1≤j≤r
1≤j≤r
Let A1 , . . . , At be the edge groups in ∆2i that are non-periodic, have cyclic stabilizers, and can not be conjugated into a vertex group in ∆2i+1 . Let a1 , . . . , at
be the positive generators of A1 , . . . , At , i.e., the generators of the cyclic groups
A1 , . . . , At that have conjugates that are mapped to F Sk , by each of the homomorphisms {hn }.
With the limit group L there is an associated JSJ decomposition. Since the
subgroups A1 , . . . , Aℓ have cyclic centralizers, and they can not be conjugated into
vertex groups in ∆2(i+1) , their positive generators a1 , . . . , at correspond to oriented
s.c.c. C1 , . . . , Ct on surfaces (that are associated with M QH subgroups) in the JSJ
decomposition of L.
Each of edge subgroups A1 , . . . , At can not be conjugated into a vertex group in
∆2(i+1) . Hence, with each of the edge group A1 , . . . , At we can associate a minimal,
non-degenerate subgraph of ∆2(i+1) , that contains a conjugate of it. Let B1 , . . . , Bℓ
be the edge groups in the union of these minimal non-degenerate subgraphs. By
the properties of the JSJ decomposition of the limit group L, each of the subgroups
61
B1 , . . . , Bℓ is cyclic, has a cyclic centralizer, and is associated with (the subgroup
generated by) a s.c.c. on a surface in the JSJ decomposition of L (as a limit group).
Let b1 , . . . , bℓ be the positive generators of the cyclic groups B1 , . . . , Bℓ . These
positive generators correspond to oriented s.c.c. c1 , . . . , cℓ on surfaces in the JSJ
decomposition of the limit group L. Furthermore, since the generators a1 , . . . , at
and b1 , . . . , bℓ are positive, if two oriented curves Ci and cj intersect, each of their
intersection points is positively oriented.
With each of the given generators s1 , . . . , sr we can naturally associate a (possibly empty) finite collection of oriented arcs and curves on the surfaces in the JSJ
decomposition of the limit group L. Since the graph of groups ∆2i and ∆2(i+1)
were constructed from pair homomorphisms, the elements s1 , . . . , sr act positively
on both ∆2i and ∆2(i+1) . Hence, if an oriented arc and or curve from the finite
collections that are associated with each of the elements s1 , . . . , sr intersect with
any of the oriented s.c.c. C1 , . . . , Ct or c1 , . . . , cℓ , then every such intersection point
is positively oriented.
With each homomorphism from a (punctured) surface group into a free group
it is possible to associate a map from the (punctured) surface into a bouquet of
circles. With a pair homomorphism the image of such a map is a bouquet of
oriented circles. By homotoping such a map to be transversal at the midpoints of
the (oriented) circles, we may assume that the preimages of the midpoints of the
oriented circles are finite collections of oriented arcs and s.c.c. on the (punctured)
surface.
The pair homomorphisms hsn are compositions hsn = hbn ◦νn , where νn ∈ M od(∆2(i+1) ,
and the homomorphisms hbn are much shorter than the homomorphisms hsn . With
each of the homomorphisms hbn there is an associated map from each of the (punctured) surfaces in the JSJ decomposition of L into a bouquet of oriented circles. As
we indicated, we may assume that the preimages of the midpoints of these oriented
circles are finite collections of oriented s.c.c. and arcs. Up to conjugacy, each of
the oriented s.c.c. cj (that are associated with the positive generators b1 , . . . , bℓ of
edge groups in ∆2(i+1) ) are mapped by hbn into F Sk . Hence, we may homotope the
map from the surface to the bouquet of oriented circles so that the preimages of
the midpoints of the oriented circles are oriented s.c.c. and arcs that intersect the
s.c.c. c1 , . . . , cℓ only in positive orientation.
We set the (modular) automorphism ψ ∈ M XM od(∆2i to be the composition
of positive Dehn twists along the edges that are associated with the edge groups
A1 , . . . , At in ∆2i . We set the (modular) automorphism ϕ ∈ M od(∆2(i+1) to be
the composition of positive Dehn twists along the edges that are associated with
the edge groups B1 , . . . , Bℓ .
We look at homomorphisms of the form hbn ◦ ϕm for large m. Since the homomorphisms hsn map the elements s1 , . . . , sr and C1 , . . . , Ct into F Sk , so do the
homomorphisms hbn ◦ ϕm . We look at the map from the surfaces that are associated with the JSJ decomposition of L into the bouquet of oriented circles that is
associated with homomorphisms hbn ◦ ϕm . The preimages of the midpoints of the
oriented circles under this map, are obtained from the preimages of the midpoints
of these circles under the maps that are associated with the homomorphisms hbn ,
by performing m powers of Dehn twists along the s.c.c. c1 , . . . , cℓ .
Since the homomorphisms hsn ) where obtained as shortenings, we may assume
that hsn ◦ ψ v (Ci ) ∈ F Sk and hsn ◦ ψ v (sj ) ∈ F Sk , for 1 ≤ i ≤ t, 1 ≤ j ≤ r, and
62
v = 0, 1, . . . , v0 for some fixed (previously chosen) positive integer m0 . Hence, for
every every large m, hbn ◦ ϕm ◦ ψ v (Ci ) ∈ F Sk and hsn ◦ ϕm ◦ ψ v (sj ) ∈ F Sk , for
1 ≤ i ≤ t, 1 ≤ j ≤ r, and every positive integer v.
Because the intersection points between the s.c.c. ci and Cj , 1 ≤ i ≤ t, 1 ≤ j ≤ ℓ,
are all positively oriented, and because the high power ϕm of positive Dehn twists
along the curves {cj }, surrounds each such s.c.c. with long positive (periodic) words,
it follows that for every large positive pair m1 , m2 , and for every large m that is
much bigger than both m1 and m2 : hbn ◦ ϕm ◦ ψ m2 ◦ ϕm1 (cj ) ∈ F Sk , 1 ≤ j ≤ ℓ, and
hbn ◦ ϕm ◦ ψ m2 ◦ ϕm1 (Ci ) ∈ F Sk , 1 ≤ i ≤ t.
Now, with each of the elements s1 , . . . , sr we can associate oriented arcs and
curves on the surfaces that are associated with the JSJ decomposition of L. Whenever these curves intersect the curves C1 , . . . , Cℓ and c1 , . . . , ct , they intersect them
positively. Since hbn (sj ) ∈ F Sk , and hsn (sj ) ∈ F Sk , for 1 ≤ j ≤ r, for every large
positive pair m1 , m2 , and for every large m that is much bigger than both m1 and
m2 : hbn ◦ ϕm ◦ ψ m2 ◦ ϕm1 (sj ) ∈ F Sk , 1 ≤ j ≤ r.
Furthermore, the positive intersection numbers, and the high power of positive
Dehn twists along the curves {cj }, imply that for every positive integer p, and every
large tuple of positive integers: e1 , f1 , . . . , ep , fp , and for every large m that is much
bigger than the sum of these positive integers:
hbn ◦ ϕm ◦ ψ fp ◦ ϕep ◦ ψ f1 ◦ ϕe1 (Ci ) ∈ F Sk , 1 ≤ i ≤ t
hbn ◦ ϕm ◦ ψ fp ◦ ϕep ◦ ψ f1 ◦ ϕe1 (cj ) ∈ F Sk , 1 ≤ j ≤ ℓ
hbn ◦ ϕm ◦ ψ fp ◦ ϕep ◦ ψ f1 ◦ ϕe1 (sj ) ∈ F Sk , 1 ≤ j ≤ r.
Therefore, the sequence of powers that is used in constructing the JSJ decomposition for groups (theorem 4.5 in [Ri-Se]), when taken to consist of only positive
powers (so that the twisted homomorphisms are indeed pair homomorphisms), can
be used to construct a sequence of homomorphisms that converges into a faithful
action of the limit group L on a real tree, and this limit action contains an IET
component, as proposition 3.5 claims.
Proposition 3.5 completes the proof of theorem 3.2.
Theorem 3.2 proves that if every faithful action of a freely indecomposable limit
group L on a real tree Y , that is obtained as a limit from a sequence of homomorphisms of the pair (S, L) into (F Sk , Fk ), is discrete (or simplicial), then the pair
(S, L) has only finitely many maximal abelian decompositions, that we view as its
(finite collection of) canonical JSJ decompositions. A similar statement is valid in
case every such action contains only simplicial and axial components.
Theorem 3.6. Suppose that L is freely indecomposable, and that all the maximal
abelian decompositions, {Λi }, that are associated with the pair (S, L), correspond
to faithful actions of L on a real tree, where these actions contain only simplicial
and axial components.
Then there exist only finitely many (equivalence classes of ) maximal abelian
decompositions of the pair (S, L).
63
Proof: Suppose that there are infinitely many (equivalence classes of) maximal
abelian decompositions of a pair (S, L). Let {Λi }∞
i=1 be the collection of these
(inequivalent) maximal decompositions.
All the abelian decompositions, Λi , are dominated by the JSJ decomposition of
the freely indecomposable limit group L. If there exists an infinite subsequence of
the decompositions, {Λi }, that is dominated by an abelian decomposition, Θ1 , that
is strictly dominated by the JSJ decomposition of L, we pass to that subsequence.
We continue iteratively. If the infinite subsequence of the decompositions, {Λi },
that is dominated by Θ1 , has a further infinite subsequence that is dominated by
an abelian decomposition, Θ2 , that is strictly dominated by Θ1 , we pass to that
subsequence. Since every strictly decreasing sequence of abelian decompositions of
the limit group L terminates after finitely many steps, the sequence of maximal
abelian decompositions of the pair, (S, L), contains an infinite subsequence, that
is dominated by an abelian decomposition, Θ, and no infinite subsequence of that
infinite subsequence is dominated by an abelian decomposition, Θ′ , that is strictly
dominated by Θ. We (still) denote this infinite subsequence of maximal abelian
decompositions, {Λi }, and their dominating abelian decomposition of L, Θ.
Our goal is to show that the abelian decomposition, Θ, that strictly dominates
the entire sequence of maximal abelian decompositions of the pair (S, L), {Λi }, is
itself a maximal abelian decomposition of (S, L), a contradiction to the maximality
of each of the abelian decompositions, Λi , hence, a contradiction to the existence
of an infinite sequence of maximal abelian decompositions that are associated with
the pair (S, L), and the theorem follows.
We start with the infinite sequence of inequivalent maximal abelian decompositions {Λi }, that are all dominated by the abelian decomposition Θ. With the
sequence of abelian decompositions, {Λi }, we have associated a sequence of homomorphisms {hn } that satisfy properties (1)-(3) that are listed in the beginning of
the proof of theorem 3.3. By possibly passing to a subsequence of the homomorphisms {hn } (still denoted {hn }), we obtain a convergent sequence, that converges
into a superstable action of the limit group L on a real tree Y , and with this action
there is an associated abelian decomposition of L that we denoted ∆1 .
Starting with ∆1 we (possibly) refine it to get an abelian decomposition ∆2 ,
precisely as we did in case all the actions on real trees are simplicial, i.e., precisely
as we did in the course of proving theorem 3.2. Since all the abelian decompositions
Λi are dominated by Θ, ∆2 is dominated by Θ as well.
Proposition 3.7. If ∆2 is equivalent to the abelian decomposition Θ, then there
exists a sequence of homomorphisms un : (S, L) → (F Sk , Fk ) that converges into a
superstable action of the limit group L on a real tree Ŷ , and the abelian decomposition that is dual to this action is Θ itself.
Proof: Θ can not be simplicial since it dominates an infinite sequence of inequivalent maximal abelian decompositions of L. Hence, Θ contains vertex groups that are
associated with either axial or IET components. ∆2 that is assumed to be equivalent to Θ is obtained from a sequence of pair homomorphisms: {hn : (S, L) →
(F Sk , Fk )}. If Θ does not contain vertex groups that are associated with IET
components, we vary the homomorphisms {hn } by precomposing them with Dehn
twists along edge groups (that connect between non-axial vertex groups), and vary
the values of a prefered basis of generators of vertex groups that are associated with
axial components using theorem 2.1, as we did in the proof of proposition 3.4 in
64
case ∆2 contains an axial component. Such modifications gives a new sequence of
homomorphisms {ĥn } that converges into a faithful action of the limit tree L on a
real tree, where the abelian decomposition that is associated with the limit action
is ∆2 that is equivalent to Θ.
If ∆2 contains vertex groups that are associated with IET components, then
we modify the homomorphisms {hn } by performing Dehn twists along edges that
connect between non-axial non-QH vertex groups, vary the values of prefered basis
of axial vertex groups using theorem 2.1, and precomposing the obtained homomorphisms with automorphisms that are extensions of automorphisms of the QH
vertex groups - precisely as we did in proving proposition 3.4 (in case ∆2 contains
a QH vertex group). Again, the obtained sequence of homomorphisms converges
into a faithful action of L on a real tree, where the abelian decomposition that is
associated with this action is ∆2 that is assumed to be equivalent to Θ.
Since Θ dominates all the maximal abelian decompositions, {Λn }, proposition
3.7 implies that if ∆2 is equivalent to Θ we obtained a contradiction, and theorem
3.2 follows. Hence, we may assume that Θ strictly dominates ∆2 .
We continue in a similar way to what we did in proving theorem 3.2. If all
the edge groups in ∆2 are elliptic, and the only non-elliptic vertex groups are
axial, then for almost all the indices i, the abelian decomposition Λi is dominated
by the abelian decomposition ∆2 (i.e., ∆2 is a (possibly trivial) refinement of Λi
for almost every index i). Hence, there is a subsequence of the maximal abelian
decompositions, {Λi }, that is dominated by the abelian decomposition ∆2 , that is
strictly dominated by the abelian decomposition Θ, a contradiction to our choice
of Θ. Therefore, ∆2 must contain a hyperbolic edge group.
Suppose that ∆2 contains an axial vertex group A = A1 +A2 , so that ∆2 collapses
to the abelian decomposition L = L1 ∗A1 A. If in addition A1 is elliptic, then it is
elliptic in almost all the abelian decomposition {Λi }, so it must be elliptic in the
dominating abelian decomposition Θ, and therefore, A is an axial vertex group in
Θ as well. If A1 is not elliptic, it can not be elliptic in Θ, and therefore, rk(A1 ) ≥ 2.
In both cases we continue by essentially analyzing the subgraph that is obtained
from ∆2 by taking out the vertex group A (and continue with the limit group L1 ).
Conversely, if A is an axial vertex group in ∆2 , A = A1 + A2 , so that ∆2
collapses to an abelian decomposition L = L1 ∗A1 A, and A is conjugate to an axial
vertex group in Θ, that is connected to other vertex groups in Θ by subgroups that
generate a conjugate to A1 , then A1 is elliptic in all the abelian decompositions
{Λi }.
From the graph of groups ∆2 we take out the edges with elliptic edge groups,
and the axial vertex groups. What left are several connected subgraphs of ∆2 , that
we denote: C1 , . . . , Cu . For each of these subgraphs we fix a finite generating set
of its fundamental group.
As in the proof of theorem 3.2, we shorten the homomorphisms {hn } (using
Dehn twists along dominating edge groups), and pass to a subsequence of them
for which the shortenings converge into a (new) faithful action of the limit group
L on some real tree. We further restrict each of the shortened homomorphisms to
the fundamental groups of the connected components (that are subgraphs in ∆2 ),
Ci , i = 1, . . . , u. After possibly passing to a subsequence, the restrictions of the
shortened homomorphisms converge to a faithful action of that group on a real tree,
that we denote Yi . We denote the abelian decomposition that is associated with
65
these actions, ∆i3 , i = 1, . . . , u. Given each of the abelian decompositions ∆i3 we
further refine it in the same way in which we refined the abelian decomposition ∆1
to obtain the abelian decomposition ∆2 . We denote the obtained refined abelian
decompositions, ∆i4 , 1 ≤ i ≤ u.
The edges that were taken out from ∆2 to obtain the connected components,
Ci , i = 1, . . . , u, are either elliptic or contained in abelian subgroups of rank at
least 2 in the fundamental groups of the connected components Ci . Hence, these
edge groups can be assumed to be either elliptic in the abelian decomposition, ∆i4 ,
i = 1, . . . , u, or the abelian decompositions ∆i4 can be modified to guarantee that
all the non-cyclic abelian subgroups are elliptic, so that the edge groups that were
taken out from ∆2 are elliptic. Therefore, from the abelian decomposition ∆i4 ,
1 ≤ i ≤ u, possibly after a modification that guarantees that all the non-cyclic
abelian subgroups are elliptic, it is possible to obtain an abelian decomposition of
the ambient limit group L, by adding the elliptic edge groups, and the QH and
(axial) abelian edge groups that were taken out from ∆2 . We denote the obtained
abelian decomposition ∆4 .
Since the abelian decomposition Θ strictly dominates ∆2 , the construction of
∆4 from (almost) shortest homomorphisms of the connected components Ci of ∆2
implies that at least one of the following possibilities must hold:
(1) there exists a QH vertex group in ∆4 .
(2) there exists an edge group with a cyclic centralizer in ∆4 that can not be
conjugated into an edge group nor into a vertex group in ∆2 . Equivalently,
there is an edge group with cyclic centralizer in ∆2 that can not be conjugated into a vertex group or into an edge group in ∆4 .
(3) there exists a non-trivial abelian subgroup A1 of an abelian subgroup A of
rank at least 2 in L, so that A1 was contained in a non-QH, non-axial vertex
group in ∆2 and A was not contained in such a vertex group in ∆2 , and A1
can not be conjugated into a non-QH, non-axial vertex group in any of the
abelian decompositions ∆i4 .
(4) there exists a non-trivial abelian subgroup A of rank at least 2 in L, so that
A was contained in a non-QH, non-axial vertex group in ∆2 , and A can not
be conjugated into a vertex group in any of the abelian decompositions ∆i4 .
If case (1) occurs we get a contradiction to the assumptions of theorem 3.6
according to proposition 3.4. If cases (3) or (4) occur we apply the argument that
was used in the proof of theorem 3.2, and obtain a proper refinement of ∆2 . If
case (2) occurs it is possible to construct a sequence of homomorphisms of the pair
(S, L) that converge into a faithful action of L on a real tree, where this real tree
contains an IET component, a contradiction to the assumptions of theorem 3.6.
As long as the obtained abelian decomposition is strictly dominated by Θ we can
continue refining the obtained abelian decomposition iteratively. Therefore, after
finitely many refinements we get the abelian decomposition Θ, and a sequence of
homomorphisms of the pair (S, L) that converge into a faithful action of L on
a real tree with an associated abelian decomposition Θ. Since Θ dominates the
sequence of maximal abelian decompositions {Λi }, we obtained a contradiction to
their maximality, and theorem 3.6 follows.
In case L is freely indecomposable, and all the faithful actions of L on a real tree
that are obtained as limit of sequences of homomorphisms of the pair (S, L) contain
66
only simplicial and axial components, we set each of the finitely many maximal
abelian decompositions that are associated with such a pair (S, L), to be a JSJ
decomposition of the pair (S, L). Hence, with such a pair (S, L) we have canonically
assigned finitely many (abelian) JSJ decompositions, that are all obtained from the
abelian JSJ decomposition of the limit group L, by possibly cutting some of the
QH vertex groups along finitely many s.c.c. and further collapsing and folding.
Unfortunately, we were not able to generalize theorem 3.6 to all pairs (S, L) in
which the limit group L is freely indecomposable. i.e., we were not able to prove
that given a pair (S, L) in which L is freely indecomposable, there are only finitely
many maximal abelian decompositions that are associated with it. Such a statement
will enable one to associate canonically finitely many abelian JSJ decompositions
with such a pair (that are all dominated by the abelian JSJ decomposition of
the limit group L). The existence of these JSJ decompositions will simplify (and
enrich) considerably the structure theory that we develop in the sequel, including
the structure of the Makanin-Razborov diagram that we associate with a pair (S, L).
Question. Let (S, L) be a pair in which L is freely indecomposable. Are there only
finitely maximal abelian decompositions that are associated with the pair (S, L)? (see
proposition 3.1 for the definition and existence of maximal abelian decompositions).
§4. Limit sets
One of the fundamental objects that is associated with a Kleinian group is its
limit set. For such a (non-elementary) group the limit set is a non-empty closed
subset of S 2 , that often has a fractal structure. In this section we define a natural
limit set that is associated with a pair (S, L), where L is a freely indecomposable
limit group, and one of its associated (canonical) JSJ decompositions.
Given a (Gromov) hyperbolic group or a limit group one can study all the small
stable faithful actions of such a group on a real tree. For a freely indecomposable
torsion-free hyperbolic group, the properties of its canonical JSJ decompositions,
together with a theorem of Skora [Sk], and Thurston’s compactification of the
Teichmuller space, imply that the set of small stable faithful actions of such a group
on a real tree up to dilatation, can be naturally viewed as a topological space, that
is homeomorphic to a finite union of (possibly trivial) products of finitely many
spheres and a (possibly trivial) Euclidean factor. The properties of the abelian
decomposition of a limit group, implies the same conclusion for the structure of the
set of small stable faithful actions of a freely indecomposable limit group on a real
tree up to dilatation.
Definition 4.1. Let (S, L) be a pair in which L is a freely indecomposable limit
group. Look at all the sequences of pair homomorphisms {hn : (S, L) → (F Sk , Fk )}
that converge into a faithful action of L on a real tree. Note that the action must
be small and (super) stable.
The set of all these limit actions up do dilatation, is a subset of all the faithful
small stable actions of L on real trees up to dilatation, and can naturally be equipped
with the induced topology. We call this topological space the limit set of the pair
(S, L).
It is natural to ask if such a limit set is homeomorphic to a finite simplicial
67
complex. If so it may be described more explicitly, perhaps even computed. In the
sequel, for the purposes of encoding the structure of all pair homomorphisms of
a given pair, we don’t need a detailed understanding of this limit set, apart from
some of its global properties.
If it is homeomorphic to a simplicial complex, when is freely indecomposable,
one can ask the same question for general limit groups L, omitting the indecomposability requirement.
§5. Enlarging the Positive Cone
The JSJ decomposition a freely indecomposable limit group plays an essential
role in the geometric construction of the Makanin-Razborov diagram that encodes
the set of solutions to a system of equations over a free group ([Se1]). In section 3
we managed to associate canonically finitely many JSJ decompositions with pairs
(S, L) only is some special cases. Therefore, to construct a Makanin-Razborov
diagram that encodes the solutions to a system of equations over a free semigroup,
we need to modify the construction that is used over groups.
The main object that we use in the case of semigroups is a resolution. A
resolution has finitely many steps, it starts with the given pair (S, L), and continues
with quotients of it that are not necessarily proper quotients. Later on we show that
there exist finitely many such resolutions that encode all the pair homomorphisms
of a given pair (S, L).
To construct resolutions we start with the extended cone that we associate with
an abelian decomposition, that was already used in the shortenings in section 2.
We start with a strengthening of theorem 2.1 which is not needed in the sequel,
but is of independent interest.
Theorem 5.1. Let A be a f.g. free abelian group, let rk(A) = ℓ, and let S be a f.g.
subsemigroup of A that generates A as a group. Let s1 , . . . , sr be a fixed generating
set of S.
Then there exist finitely many positive collections of (free) bases of A, ai1 , . . . , aiℓ ,
1 ≤ i ≤ c, such that for any sequence of homomorphisms hn : (S, A) → (F Sk , Fk ),
that converges into a free action of A on a real tree, there exists an index i, 1 ≤ i ≤ c,
and an index n0 , for which for every index n > n0 :
(1) (1) hn (aij ) ∈ F Sk for 1 ≤ j ≤ ℓ.
(2) (2) for each of the (fixed set of ) generators s1 , . . . , sr of the semigroup S,
there are fixed words, that depend only on the index i, 1 ≤ i ≤ c, so that for
every m, 1 ≤ m ≤ r:
sm = wi,m (ai1 , . . . , aiℓ )
where the words wi,m are (fixed) positive words in the elements ai1 , . . . , aiℓ .
Proof: The theorem is immediate if ℓ = 1, hence, we may assume that ℓ > 1. Given
a sequence of homomorphisms {hn : (S, A) → (F Sk , Fk )}, that converges into a free
action of A on a real tree, theorem 2.1 implies that there exists a free basis of A,
a1 , . . . , aℓ , and an index n0 , so that for every n > n0 , the homomorphisms, {hn },
that satisfy properties (1) and (2) in the statement of the theorem.
To prove that there exists a finite collection of positive words that suffices for all
the convergent sequences we use a compactness argument. We have already shown
68
that given a sequence of homomorphisms {hn : (S, A) → (F Sk , Fk )} that converges
into a free action of A on a real tree Y , there exists a collection of positive words
and a basis of A for which the conclusions of the theorem hold.
Suppose that finitely many such collections of free bases together with positive
words do not suffice. There exist only countably many possible collections of positive words, so we order the infinite set of the collections of free bases and positive
words that are associated with convergent sequences of homomorphisms that satisfy
the assumptions of the theorem.
For each positive integer t, let {htn : (S, A) → (F Sk , Fk )} be a sequence of
homomorphisms that satisfies the assumption of the theorem (i.e., it converges to
a free action of A on a real tree). Suppose further that the sequence {htn } does not
satisfy the conclusions of the theorem with respect to first t collections of free bases
and positive words.
For each positive integer t, the sequence {htn } converges into a free action of the
(free) abelian group A on an oriented line Yt , λt : A × Yt → Yt . From the sequence
of actions λtn it is a possible to extract a convergent subsequence, that converge into
a non-trivial, not necessarily faithful, action of the abelian group A on a directed
real line, α0 : A × L0 → L0 .
If the action α0 is not free, let A1 be the kernel of the action α0 . A1 is a direct
summand in A. If A1 is non-trivial, then we can pass to a further subsequence, for
which the sequence of actions λt restricted to the direct summand A1 converges into
a non-trivial, not necessarily faithful action of A0 on a directed line: α1 : A0 ×L1 →
L1 .
If the action α1 is not free, we continue iteratively by restricting the actions λt
to the kernel of the previous action (which is a direct summand in the previous
summand), and pass to a convergent subsequence. Since A is a f.g. free abelian
group this iterative procedure terminates after finitely many steps.
Let {λtd } be the final convergent subsequence. By the argument that was used
to prove theorem 2.1, the conclusions of theorem 2.1 hold for the subsequence λtd .
Hence, there exists a fixed free basis of A that act positively on Ytd for large enough
d, and a fixed set of words {wm }, 1 ≤ m ≤ r, that satisfy part (ii) in the statement
of the theorem.
This free basis of A, together with the words {wm }, appear in the ordered list of
bases and words. Hence, for large enough index d, the sequences of homomorphisms
{htnd } are assumed not to satisfy the conclusion of theorem 2.1 with respect to
the free basis and the collection of words that we associated with the convergent
sequence {λtd }. But since this free basis and the collection of words {wm } were
associated using the argument of theorem 2.1 with the convergent sequence {λtd },
for large enough d, and large enough n, the sequence of homomorphisms {htnd } does
satisfy the conclusion of theorem 2.1 with respect to this free basis and collection
of words, a contradiction.
Therefore, finitely many free bases together with a finite collection of words
suffice for all the convergent sequences that satisfy the assumption of the theorem,
and the conclusion of the theorem follows.
For a pair (S, Q) in which Q is a (closed) surface group, S is a f.g. subsemigroup
that generates Q as a group, and (S, Q) is obtained as a limit of a sequence of
pair homomorphisms into (F Sk , Fk ), we state a weaker statement that associates
69
a standard cone with such a pair.
Lemma 5.2. Let (S, Q) be a pair of a closed (hyperbolic) surface group Q and a
subsemigroup S that generates Q as a group. Let s1 , . . . , sr be a fixed generating
set of S.
Let {hn : (S, Q) → (F Sk , Fk )} be a sequence of pair homomorphisms that converge into a free action of Q on a real tree. Then there exists a set of standard
generators, q1 , . . . , qℓ of Q, such that:
(1) the presentation of Q with respect to q1 , . . . , qℓ is one of the finitely many
possible (positively oriented) interval exchange type generating sets of Q
(i.e., they generate Q and the presentation is the one obtained from a permutation of finitely many positively oriented subintervals (that intersect only
in their endpoints) of an ambient positively oriented interval).
(2) there exists an index n0 such that for every n > n0 : hn (qb ) ∈ F Sk for
b = 1, . . . , ℓ.
Proof: Since the sequence of homomorphisms {hn } converges into a minimal IET
action on a real tree, by the Rips machine, or the Makanin procedure, there exists
a standard set of generators that satisfies the conclusion of the lemma.
Theorem 5.1 and lemma 5.2 generalize to pairs (S, L) in which the ambient limit
group L is freely indecomposable. The generalization is crucial in constructing the
Makanin-Razborov diagram of a pair.
Theorem 5.3. Let (S, L) be a pair of a freely indecomposable limit group L and
a subsemigroup S that generates L as a group. Let s1 , . . . , sr be a fixed generating
set of S. Let {hn : (S, L) → (F Sk , Fk )} be a sequence of pair homomorphisms
that converge into a faithful action of L on a real tree with an associated abelian
decomposition Λ. Suppose that A is the stabilizer of an axial component in the real
tree (and a vertex group in Λ), and Q is a QH vertex group in Λ that is associated
with an IET component.
Let A0 < A be the direct summand of A that contains the subgroup that is
generated by the edge groups that are connected to the vertex that is stabilized by A
in Λ as a subgroup of finite index. Suppose that: rk(A) − rk(A0 ) = ℓ. Let ĝ be the
number of vertices in Λ, that are not stabilized by the abelian group A, for which
their vertex groups have conjugates that stabilize points in the axis of A (these are
all the vertices that are adjacent to the vertex that is stabilized by A in Λ). Let
g = max(ĝ − 1, 0). Let EA = A+ < e1 , . . . , eg > be a free abelian group of rank
rk(A) + g.
With the QH subgroup Q we associate a natural extension. The associated natural
extension of Q is generated by Q and finitely many elements that map a point with
non-trivial stabilizer in the subtree that is stabilized by Q, to another point with a
non-trivial stabilizer in that subtree that is not in the same orbit under the action of
Q. By adding these elements all the points with non-trivial stabilizers in the subtree
that is stabilized by Q are in the same orbit under the natural extension of Q.
Each of the given generators of the semigroup S, s1 , . . . , sr , can be written as a
word (in a normal form) in elements that lie in the natural extensions of the surface
group Q (that are associated with the subpaths of the paths that are associated with
s1 , . . . , sr that are contained in subtrees that are stabilized by conjugates of Q), and
70
elements that lie outside these extensions. Let t1 , . . . , td be the elements that lie
in the natural extensions of Q that appear as subwords in the normal form of the
elements s1 , . . . , sr .
Then there exist:
(i) ℓ + g elements that are part from a (free) basis of EA, a1 , . . . , aℓ+g , such
that EA =< a1 , . . . , aℓ+g > +A0 .
(ii) standard generators, q1 , . . . , qℓ , of a natural extension of Q, that satisfy
properties that are analogous to properties (1) and (2) in the statement of
lemma 5.2.
iii) each of the generators sj can be written as a word in terms of elements {tjd }
in the natural extension of Q (i.e., tjd itself is a word in q1 , . . . , qℓ , elements
{uje } that are positive words in a1 , . . . , aell+g , and elements in the other
vertex and edge groups in Λ.
Such that there exists an index n0 for which for every n > n0 the following
properties hold:
(1) hn (ab ) ∈ F Sk for 1 ≤ b ≤ ℓ + g.
(2) hn (qb ) ∈ F Sk , 1 ≤ b ≤ ℓ+g, hn (ai ) ∈ F Sk , 1 ≤ i ≤ ℓ+g, and hn (tjd ) ∈ F Sk
for every possible pair of indices (j, d).
Proof: Follows by the same arguments that were used to prove theorem 5.1 and
lemma 5.2.
§6. A Makanin-Razborov diagram - the freely indecomposable case
In the previous section we associated a standard set of generators with an abelian
decomposition. In this section we use shortenings that were constructed in section 2,
together with the machine for the construction of the JSJ decomposition for groups,
to associate finitely many resolutions with a given pair (S, L) in which the limit
group L is freely indecomposable. These resolutions enable one to encode the set
of all the pair homomorphisms of such a pair using the set of pair homomorphisms
of finitely many proper quotient pairs.
For presentation purposes we will start by proving the main theorem in case
the limit group L is freely indecomposable and contains no non-cyclic abelian subgroups, and then combine it with arguments that were used in the construction
of the JSJ decompositions of pairs (in special cases) in section 3 to omit the assumption on abelian subgroups. In the sequel we generalize the construction of
resolutions to all pairs, omitting the freely indecomposable assumption. These resolutions are the building block of the Makanin-Razborov diagram that encodes all
the pair homomorphisms of a given pair, or alternatively, all the solutions to a
system of equations over a free semigroup.
Theorem 6.1. Let (S, L) be a pair, where L is a freely indecomposable limit group,
and let s1 , . . . , sr be a fixed generating set of the semigroup S. Suppose that the limit
group L contains no non-cyclic abelian subgroup. Let {hn : (S, L) → (F Sk , Fk )} be
a sequence of pair homomorphisms that converges into a faithful action of L on a
real tree Y .
71
Then there exists a resolution:
(S1 , L1 ) → (S2 , L2 ) → . . . → (Sm , Lm ) → (Sf , Lf )
that satisfies the following properties:
(1) (S1 , L1 ) = (S, L), and ηi : (Si , Li ) → (Si+1 , Li+1 ) is an isomorphism for
i = 1, . . . , m − 1 and ηm : (Sm , Lm ) → (Sf , Lf ) is a proper quotient map.
(2) with each of the pairs (Si , Li ), 1 ≤ i ≤ m, there is an associated abelian
decomposition that we denote Λi .
(3) there exists a subsequence of the homomorphisms {hn } that factors through
the resolution. i.e., each homomorphism hnr from the subsequence, can be
written in the form:
1
hnr = ĥr ◦ ϕm
r ◦ . . . ◦ ϕr
where ĥr : (Sf , Lf ) → (F Sk , Fk ), each of the automorphisms ϕir ∈ M od(Λi ),
and each of the homomorphisms:
i
hinr = ĥr ◦ ϕm
r ◦ . . . ◦ ϕr
is a pair homomorphism hinr : (Si , Li ) → (F Sk , Fk ).
Proof: If the abelian decomposition Λ that is associated with the action of L on
the limit tree Y is equivalent to the abelian JSJ decomposition of the limit group
L (as a group), then the theorem follows from the possibility to shorten using
M od(Λ) that was proved in section 2, and claim 5.3 in [Se1]. i.e., in that case
every shortening quotient is a proper quotient, so a subsequence of shortenings of
the homomorphisms {hn } converge into a proper quotient of the pair (S, L), and
the conclusion of the theorem follows with a resolution of length 1: η1 : (S, L) →
(Sf , Lf ).
Hence, we may assume that the abelian decomposition Λ, that is associated
with the action of L on the limit tree Y , is strictly dominated by the abelian JSJ
decomposition of L. In particular, Λ does not correspond to a closed surface, so it
contains edges.
Assume that the sequence {hn } converges into a faithful action of the limit
group L on some real tree Y with an associated abelian decomposition Λ, L is
freely indecomposable, and contains no non-cyclic abelian subgroups.
We start by refining the abelian decomposition Λ precisely as we did in proving
theorem 3.2. Using definition 3.3 we divide the edges that are adjacent to a given
non-QH vertex group in Λ to periodic and non-periodic. If there is a non-QH vertex
group in Λ that is adjacent only to periodic edge groups, we pass to a subsequence of
the homomorphisms {hn } for which their restrictions to this vertex group converges
to a non-trivial faithful action on a real tree. The abelian decomposition that is
associated with this action (in which all the edge groups that are connected to
this vertex group in Λ are elliptic), enable us to further refine Λ. By BestvinaFeighn accessibility [Be-Fe1], or alternatively by acylindrical accessibility [Se3], this
refinement process terminates after finitely many steps, and we obtain an abelian
decomposition Λ1 . Note that in Λ1 every non-QH vertex group is adjacent to at
least one non-periodic edge group.
72
In section 3 we divided the edges and vertices in Λ1 into several equivalence
classes, according to the growth of the translation lengths of fixed sets of generators of the corresponding edge groups. We fix a finite set of generators for
each of the vertex groups and edge groups in Λ1 . Recall that we say that two
displacement functions of two vertex groups or edge groups, dispn (V1 ), dispn (V2 ),
are comparable if there exists positive constants c1 , c2 , so that for every index n,
c1 · dispn (V1 ) < dispn (V2 ) < c2 · dispn (V1 ). We say that dispn (V1 ) dominates
dispn (V2 ) if dispn (V2 ) = o(dispn (V1 )).
After passing to a subsequence of the homomorphisms {hn } this defines an order on the equivalence classes of the edge groups and vertex groups in Λ1 . In
particular, there exists a collection of edge groups and vertex groups with comparable displacement functions that dominate all the other (displacement functions
of) vertex groups and edge groups in Λ1 .
Suppose first that Λ1 does not contain QH vertex groups. In that case we set
M od(Λ1 ) to be the modular group of the pair (S, L) that is associated with the
abelian decomposition Λ1 , and M XM od(Λ1 ) to be the subgroup of M od(Λ1 ), that
is generated by Dehn twists only along dominating edge groups.
We start by using the full modular group M od(Λ1 ). For each index n, we set
the pair homomorphism: h1n : (S, L) → (F Sk , Fk ) and h1n = hn ◦ ϕn where ϕn ∈
M od(Λ1 ), to be the shortest pair homomorphism that is obtained from hn by
precomposing it with a modular automorphism from M od(Λ1 ). If there exists a
subsequence of the homomorphisms {h1n } that converge into a proper quotient of the
pair (S, L), we set the limit of this subsequence to be (Sf , Lf ), and the conclusion
of the theorem follows with a resolution of length 1.
Therefore, we may assume that every convergent subsequence of the homomorphisms {h1n } converges into a faithful action of the limit group L on some real tree.
In that case we use only Dehn twists along dominant edge groups. For each index
n, we set h1n = hn ◦ ϕn , where ϕn ∈ M XM od(Λ1 ), and h1n is one of the shortest homomorphisms that is obtained by precomposing hn with a modular automorphism
from M XM od(Λ1 ). We pass to a subsequence of the homomorphisms {h1n } that
converges into (necessarily faithful) action of L on some real tree with an associated
abelian decomposition ∆2 . We further refine ∆2 by analyzing actions of non-QH
vertex groups in ∆2 that are connected only to periodic edge groups and obtain a
(possibly) refinement of ∆2 that we denote Λ2 .
For presentation purposes we assume that Λ2 contains no QH vertex groups.
Note that by construction every dominant edge group in Λ1 is not elliptic in Λ2 .
With Λ2 we associate its modular group M od(Λ2 ), and the modular group that is
associated only with dominant edge groups that we denote M XM od(Λ2 ).
We continue iteratively. First we shorten using the full modular group, M od(Λi ),
and check if there a subsequence of shortened homomorphisms that converge into
a proper quotient of the pair (S, L). If there is such a subsequence we obtained a
finite resolution. If not we shorten only along edges with dominant edge groups,
i.e., using automorphisms from M XM od(Λi ), pass to a convergent subsequence
and further refine the obtained abelian decomposition. For presentation purposes
we assume that all the obtained abelian decompositions, {Λi } do not contain QH
vertex groups.
Definition 6.2. Let Λ1 , . . . , Λi , . . . be the infinite sequence of abelian decompositions of the limit groups L that are constructed along the iterative procedure. Note
73
that we assume that all these abelian decompositions do not contain QH vertex
groups, and that the limit group L in the pair (S, L) is freely indecomposable and
contains no non-cyclic abelian subgroups. By construction, a dominant edge group
in Λi is not elliptic in Λi+1 .
All the abelian decompositions Λi are dominated by the abelian JSJ decomposition
of L. For each index i, we set Θi to be the minimal abelian decomposition (w.r.t.
to the natural partial order that we defined on abelian decompositions of L) that
dominates all the abelian decompositions, Λi , Λi+1 , . . . .
The sequence of abelian decompositions {Θi } is non-increasing, and every strictly
decreasing sequence of abelian decompositions of L has to terminate, hence, there
exists some index i0 , such that for every i > i0 , Θi0 = Θi . We call Θi0 the stable
dominant abelain decomposition of the sequence Λ1 , Λ2 , . . . .
The abelian decompositions Λi were assumed to have no QH vertex groups.
Since every dominant edge group in Λi is hyperbolic in Λi+1 , each of the dominant
abelian decompositions, Θi , dominates a pair of hyperbolic-hyperbolic splittings of
the limit group L. Hence, it must contain a QH vertex group. In particular, the
stable dominant abelian decomposition, Θi0 , must contain a QH vertex group.
The next proposition is crucial in our approach to proving theorem 6.1. It enables the substitution of the entire suffix of the sequence of abelian decompositions
{Λi }, Λi0 , Λi0 +1 , . . . , with a single abelian decomposition - the stable dominant
decomposition, Θi0 .
Proposition 6.3. There exists a subsequence of shortened pair homomorphisms
{hini }, i ≥ i0 , and a sequence of automorphisms: ψ i , ν i ∈ M od(Θi0 ), i ≥ i0 , with
the following properties:
(1) hini is obtained from the pair homomorphism hni , by shortening hni using a
sequence of elements from the dominant modular groups: M XM od(Λ1 ), . . . , M XM od(Λi).
i.e.,
hni = hini ◦ ϕini ◦ . . . ◦ ϕ1ni
where ϕjni ∈ M XM od(Λj ), 1 ≤ j ≤ i and for each j, 1 ≤ j ≤ i − 1, the
homomorphism:
hjni = hini ◦ ϕini ◦ . . . ◦ ϕj+1
ni
is a pair homomorphism hjni : (S, L) → (F Sk , Fk ).
(2) For each index i we define a pair homomorphism:
fi = hini ◦ ν i+1 ◦ ψ i ◦ . . . ◦ ψ i0 .
The sequence of pair homomorphisms, {fi }, converges into a faithful action
of the limit group L on a real tree with an associated abelian decomposition
Θi0 . In particular, the limit action contains an IET component.
Proof: To prove the proposition we basically imitate the construction of the JSJ
decomposition as it appears in [Se1] and [Ri-Se]. We fix a generating set <
s1 , . . . , sr > of the semigroup S, that also generate the limit group L (as a group).
We start with the subsequence of pair homomorphisms that are obtained using a
(finite) iterative sequence of shortenings from a subsequence of the given sequence
74
of homomorphisms {hn }, using the dominant modular groups M XM od that are
associated with the abelian decompositions Λ1 , . . . , Λi0 −1 . By our assumptions
the sequence {hni0t−1 } converges into a faithful action of L on a real tree, with an
associated abelian decomposition that can be further refined (by restricting the
homomorphisms to non-QH vertex groups that are connected only to periodic edge
groups) to the abelian decomposition Λi0 .
We set the automorphisms νni0t ∈ M od(Λi0 ) to be automorphisms that guarantee
that the compositions hni0t−1 ◦ νni0t converge into a faithful action of L on a real tree
Yi0 , in which:
(1) the abelian decomposition that is dual to the action of L on the new real tree
Yi0 is Λi0 . In particular, all the edges in Λi0 correspond to non-degenerate
segments in Yi0 .
(2) for some pre-chosen (and arbitrary) positive constant ǫi0 , and after rescalings, the lengths of the segments [y0 , sj (y0 )], 1 ≤ j ≤ r, differ by at most
ǫi0 from the length of that segment in the original real tree, which is the
limit of the sequence of homomorphisms hin0t−1 . Furthermore, the length of
each segment in the quotient graph of groups that is associated with the
action of L on Yi0 differ by at most ǫi0 from the length of the corresponding
segment in the graph of groups that is associated with the action of L on
the real tree that is obtained as a limit of the sequence hni0t−1 .
(3) the segments [y0 , sj (y0 )] are degenerate or positively oriented, and for each
t, hni0t−1 ◦ νni0t ∈ F Sk .
By the properties of the abelian decomposition Λi0 , such automorphisms νni0t can
be constructed.
We continue iteratively. In analyzing the next level, the one in which the sequence
i0
{hnt } converges into a limit action on a real tree that is associated with Λi0 +1 , we set
the automorphisms ψ i0 ∈ M od(Λi0 ) and νni0t+1 ∈ M od(Λi0 +1 ), to be automorphisms
that guarantee that the compositions hin0t ◦ νni0t+1 converge into a faithful action of
L on a real tree Yi0 +1 , in which:
(1) the abelian decomposition that is dual to the action of L on the new real
tree Yi0 +1 is Λi0 +1 . In particular, all the edges in Λi0 +1 correspond to
non-degenerate segments in Yi0 +1 .
(2) we fix finite generating sets for all the vertex groups in Λi0 that are connected to dominant edge groups, and for all the dominant edge groups in
Λi0 .
For some pre-chosen (and arbitrary) positive constant ǫi0 +1 , after rescalings, the lengths and the translation lengths of all the elements in these
fixed finite generating sets when acting on Yi0 +1 , differ by at most ǫi0 +1
from the corresponding lengths and translation lengths in the limit tree
that is obtained from the sequence of homomorphisms {hin0t }.
(3) after rescaling, the lengths of the segments [y0 , ψ i0 (sj )(y0 )], 1 ≤ j ≤ r,
in Yi0 +1 , differ by at most ǫi0 +1 from the lengths of the corresponding
segments, [y0 , sj (y0 )] in the limit tree Yi0 .
(4) the segments [y0 , sj (y0 )] are degenerate or positively oriented in Yi0 +1 , and
for each t, hin0t ◦ νni0t+1 (sj ) ∈ F Sk and hin0t ◦ νni0t+1 ◦ ψ i0 (sj ) ∈ F Sk .
In analyzing the next level, the one in which the sequence {hin0t+1 } converges
into a limit action on a real tree that is associated with Λi0 +2 , we consider all the
75
elements in the ball of radius 2, B2 , in the Cayley graph of the limit group L w.r.t.
the generating set s1 , . . . , sr .
We set the automorphisms ψ i0 +1 ∈ M XM od(Λi0 +1 ) and νni0t+2 ∈ M od(Λi0 +2 )
to be automorphisms that guarantee that the compositions hin0t+1 ◦ νni0t+2 converge
into a faithful action of L on a real tree Yi0 +2 , in which:
(1) the abelian decomposition that is dual to the action of L on the new real
tree Yi0 +2 is Λi0 +2 . In particular, all the edges in Λi0 +2 correspond to
non-degenerate segments in Yi0 +2 .
(2) we fix finite generating sets for all the vertex groups in Λi0 +1 that are
connected to dominant edge groups, and for all the dominant edge groups
in Λi0 +1 .
For some pre-chosen (and arbitrary) positive constant ǫi0 +2 , after rescalings, the lengths and the translation lengths of all the elements in these
fixed finite generating sets when acting on Yi0 +2 , differ by at most ǫi0 +2
from the corresponding lengths and translation lengths in the limit tree
that is obtained from the sequence of homomorphisms {hin0t+1 }.
(3) after rescaling, for all the elements u ∈ B2 , the lengths of the segments
[y0 , ψ i0 +1 ◦ ψ i0 (u)(y0 )] in Yi0 +2 , differ by at most ǫi0 +2 from the lengths of
the corresponding segments, [y0 , ψ i0 (u)(y0 )] in the limit tree Yi0 +1 .
(4) the segments [y0 , sj (y0 )] are degenerate or positively oriented in Yi0 +2 , and
for each t, hin0t+1 ◦ νni0t+2 (sj ) ∈ F Sk and hin0t+1 ◦ νni0t+2 ◦ ψ i0 +1 ◦ ψ i0 (sj ) ∈
F Sk .
We continue iteratively, constructing at each level i the automorphisms {νni+1
},
t
i
i
and ψ . Note that in the iterative construction the automorphisms ψ are fixed at
step i, and the sequence {νni+1
} is constructed at step i. All these automorphisms
t
are from the modular group M od(Θi0 ).
We set the homomorphisms fi , the sequence of homomorphisms that appear in
part (2) of the statement of the proposition, to be:
fi = hini ◦ ν i+1 ◦ ψ i ◦ . . . ◦ ψ i0
for a suitable strictly increasing sequence of indices: {ni }, such that the sequence
{fi } converges into an action of the limit group L on a real tree Y . By construction,
every non-trivial element of L is mapped to a non-trivial element of Fk for large
enough i, since the homomorphisms: hint ◦ νni+1
converge into a faithful action of L
t
on a real tree. Hence, the action of L on the limit tree Y is faithfull. Furthermore,
by the construction of the automorphisms ψ i and νni+1
, every element g ∈ L that
t
is not elliptic in the abelian decomposition Θi0 acts hyperbolically on the real tree
Y . Clearly, every element that is elliptic in Θi0 fixes a point in Y , since it is
elliptic in all the abelian decompositions Λi , i ≥ i0 , and the autmorphisms ψ i
map it to a conjugate. Therefore, the abelian decomposition that is associated
with the (faithful) action of L on the limit tree Y is the stable dominant abelian
decomposition Θi0 .
Proposition 6.3 enables one to replace a suffix of the sequence of abelian decompositions: Λ1 , . . . that was constructed by iteratively shortening the given sequence
of homomorphisms {hn } (and further pass to subsequences), with a single abelian
decomposition that minimally dominates the suffix.
76
Proposition 6.3 assumes that all the constructed abelian decompositions {Λi }
do not contain QH vertex groups. To prove theorem 6.1 in case the limit group L
contains no non-cyclic abelian vertex groups, we need to generalize the construction
of the abelian decompositions {Λi }, and the conclusion of proposition 6.3, in case
the abelian decompositions Λi do contain QH vertex groups.
Suppose that L is freely indecomposable with no non-cyclic abelian subgroups.
Recall that Λ is the abelian decomposition that L inherits from its action on the
real tree Y , that it is obtained as a limit from the convergent sequence of homomorphisms {hn }. We further refine Λ. If a non-QH vertex group is connected
only to periodic edge groups (and in particular it is not connected to a QH vertex
group), we restrict the homomorphisms {hn } to such a vertex group and obtain a
non-trivial splitting of it in which all the previous (periodic) edge groups are elliptic. Hence, the obtained abelian decomposition of the non-QH vertex group can be
used to refine the abelian decomposition Λ. Repeating this refinement procedure
iteratively, we get an abelian decomposition that we denote Λ1 .
We fix finite generating sets of all the edge groups and all the non-QH vertex
groups in Λ1 . We divide the edge and non-QH vertex groups in Λ1 into finitely
many equivalence classes of their growth rates as we did in case Λ1 contained no
QH vertex groups. Two non-QH vertex groups or edge groups (or an edge and a
non-QH vertex group) are said to be in the same equivalence class if the maximal
length of the image of their finite set of generators have comparable lengths, i.e.,
the length of a maximal length image of one is bounded by a (global) constant
times the maximal length of an image of other, and vice versa. After passing to a
subsequence of the homomorphisms {hn } (that we still denote {hn }), the non-QH
and edge groups in Λ1 are divided into finitely many equivalence classes. We say
that one class dominates another, if the maximal length of an image of a generator
of a group from the first class dominates the maximal length of an image of the
second, but not vice versa. By definition, the classes are linearly ordered (possibly
after passing to a further subsequence). There exists a class that dominates all the
other classes, that we call the dominant class, that includes (possibly) dominant
edge groups and (possibly) dominant non-QH vertex groups.
As in the case in which Λ1 contains no QH vertex groups, we denote by M od(Λ1 )
the modular group that is associated with Λ1 . We set M XM od(Λ1 ) to be the
dominant subgroup that is generated by Dehn twists along dominant edge groups,
and modular groups of those QH vertex groups that the lengths of the images of
their (fixed) set of generators grows faster than a constant times the length of the
images of the generators of a dominant edge or vertex group. We call these QH
vertex groups, dominant QH vertex groups.
As we did in the simplicial case, we start by using the full modular group
M od(Λ1 ). For each index n, we set the pair homomorphism: h1n : (S, L) →
(F Sk , Fk ), h1n = hn ◦ϕn , where ϕn ∈ M od(Λ1 ), to be a shortest pair homomorphism
that is obtained from hn by precomposing it with a modular automorphism from
M od(Λ1 ). If there exists a subsequence of the homomorphisms {h1n } that converges
into a proper quotient of the pair (S, L), we set the limit of this subsequence to be
(Sf , Lf ), and the conclusion of the theorem follows with a resolution of length 1.
Therefore, we may assume that every convergent subsequence of the homomorphisms {h1n } converges into a faithful action of the limit group L on some real tree.
In that case we use only elements from the dominant modular group, M XM od(Λ1 ).
77
First, we shorten the action of each of the dominant QH vertex groups using the
procedure that is used in the proof of propositions 2.7 and 2.8. For each of the
dominant QH vertex groups, the procedures that are used in the proofs of these
m
propositions give us an infinite collections of positive generators, um
1 , . . . , ug , with
similar presentations of the corresponding QH vertex groups, which means that the
sequence of sets of generators belong to the same isomorphism class. Furthermore,
with each of these sets of generators there are associated words, wjm , j = 1, . . . , r, of
lengths that increase with m, such that a given (fixed) set of positive elements can
m
m
be presented as: yj = wjm (um
1 , . . . , ug ). The words wj are words in the generators
um
i and their inverses. However, they can be presented as positive words in the
generators um
j , and unique appearances of words tℓ , that are fixed words in the
elements um
j and their inverses (i.e., the words do not depend on m), and these
m m
elements tℓ (u1 , . . . , um
g ) are positive for every m.
Given the output of the procedures that were used in propositions 2.7 and 2.8,
with each of the dominant QH vertex groups we associate a system of generators
u11 , . . . , u1g (the integer g depends on the QH vertex group). Since an IET action
of a QH vertex group is indecomposable in the sense of [Gu], finitely many (fixed)
translates of each of the positive paths that are associated with the positive paths,
u1i1 , cover the positive path that is associated with u1i2 , and the positive paths that
are associated the words tℓ (u11 , . . . , u1g ). The covering of the elements u1 i1 by finitely
many translates of elements u1i2 guarantee that the ratios between their lengths
along the iterative (shortening) procedure that we use remain globally bounded.
With a QH vertex group Q, and the abelian decomposition Λ1 , we can naturally
associate an abelian decomposition ΓQ , that is obtained by collapsing all the edge
groups that are not connected to Q in Λ1 , ΓQ contains one QH vertex group, Q,
and all the other vertex groups are connected only to the vertex stabilized by Q. If
a generator sj from the fixed set of generators of the semigroup S, s1 , . . . , sr , is not
elliptic in ΓQ , or it is contained in a conjugate of Q, then finitely many translates
of the path that is associated with sj in the limit tree Y covers the paths that are
associated with the elements: tℓ and u11 , . . . , u1g that generate Q.
These coverings of the paths that are associated with the elements u11 , . . . , u1g by
translates of the path that is associated with sj , will guarantee that the lengths of
the paths that are associated with sj , along the entire shortening process that we
present, are bounded below by some (global) positive constant times the lengths of
the paths that are associated with the elements, u11 , . . . , u1g along the process.
For all the homomorphisms {hn }, except perhaps finitely many of them, the
images of the elements u11 , . . . , u1g that are associated with the various dominant
QH vertex groups, and the elements tℓ (u11 , . . . , u1g ), and the generators of the edge
groups in Λ1 , are all in free semigroup F Sk .
At this point we shorten the homomorphisms {hn } using the dominant modular
group M XM od(Λ1 ). For each homomorphism hn , we pick the shortest homomorphism after precomposing with an element from M XM od(Λ1 ) that keeps the
positivity of the given set of the images of the generators s1 , . . . , sr , and the positivity of the elements u11 , . . . , u1g and tℓ , and keeps their lengths to be at least the
maximal length of the image of a generator of a dominant edge or vertex group
(the generators are chosen from the fixed finite sets of generators of each of the
vertex and edge groups). We further require that after the shortening, the image
of each of the elements u11 , . . . , u1g and tℓ , will be covered by the finitely many
78
translates of them and of the paths that are associated with the relevant generators
s1 , . . . , sr , that cover them in the limit action that is associated with Λ1 . We further require that if a path that is associated with one of the generators, s1 , . . . , sr ,
passes through an edge with a dominant edge group in Λ1 , then after shortening
the path that is associated with such a generator contains at least a subpath that
is associated with the dominant edge group. This guarantees that the length of
such a generator remains bigger than the length of the dominant edge group along
the entire procedure. We (still) denote the obtained (shortened) homomorphisms
{h1n }.
By the shortening arguments that are proved in propositions 2.7 and 2.8, the
lengths of the images, under the shortened homomorphisms {h1n }, of the elements
u11 , . . . , u1g and tℓ , that are associated with the various dominant QH vertex groups,
and the lengths of the elements, s1 , . . . , sr , are bounded by some constant c1 (that
is independent of n) times the maximal length of the images of the fixed generators
of the dominant vertex and edge groups.
We pass to a subsequence of the homomorphisms {h1n } that converges into an
action of L on some real tree with an associated abelian decomposition ∆2 . If the
action of L is not faithful, the conclusions of theorem 6.1 follow, hence, we may
assume that the action of L is faithful. We further refine ∆2 , by restricting (a
convergent subsequence of) the homomorphisms to non-QH vertex groups that are
connected only to periodic edge groups, as we did with Λ, and construct an abelian
ˆ 2.
decomposition that we denote ∆
Suppose that there exists a QH vertex group Q in Λ1 , so that all its boundary
ˆ 2 . i.e., every boundary element is contained in either an
elements are elliptic in ∆
ˆ 2 . In that case, by the properties of
edge group or in a non-QH vertex group in ∆
the JSJ decomposition of the freely indecomposable (limit) group L, there exists
ˆ 2 and contains Q as a QH vertex
an abelian decomposition that possibly refines ∆
ˆ 2 , so that it contains all the QH vertex groups in Λ1 for
group. We further refine ∆
which all their boundary components are elliptic in D̂elta2 . We denote the obtained
decomposition (refinement), Λ2 .
By construction, a dominant edge group, a dominant boundary component of a
QH vertex group, a dominant non-QH vertex group, and a QH vertex group with
a dominant boundary component in Λ1 can not be elliptic in Λ2 (i.e., they can not
be contained in a non-QH vertex group or an edge group in Λ2 ). Furthermore, a
QH vertex group that has a dominant boundary element in Λ1 , is not elliptic nor
a QH vertex group in Λ2 .
With Λ2 we associate its modular group M od(Λ2 ). We further associate with
Λ2 its dominant edge groups and non-QH vertex groups, and the modular group
that is associated only with dominant edge groups, and with dominant QH vertex
groups, that we denote M XM od(Λ2 ).
We continue in a similar way to what we did with the sequence of homomorphisms {hn } and with Λ1 . We first shorten using automorphisms from the ambient
modular group M od(Λ2 ). If there exists a subsequence of shortened homomorphisms that converges into a proper quotient of L, the conclusion of theorem 6.1
follows. If there is no such subsequence, we restrict the shortenings to automorphisms from the dominant modular group M XM od(Λ2 ), and modify what we did
in shortening the homomorphisms {hn } (and the abelian decomposition Λ1 ).
First, we associate sets of positive generators with each of the new QH vertex
79
groups in Λ2 , i.e., those QH vertex groups in Λ2 that are not QH vertex groups in
Λ1 . We further choose finitely many translates of the paths that are associated with
each generator that cover the paths that are associated with the other generators.
These translates guarantee that the ratios between the lengths of the paths that
are associated with these generators along the iterative (shortening) procedure that
we use remain globally bounded.
With a QH vertex group Q, and the abelian decomposition Λ2 , we can naturally
associate an abelian decomposition Γ2Q , as we associated with QH vertex groups in
Λ1 . Γ2Q is obtained by collapsing Λ2 , and it contains one QH vertex group, Q, and
all the other vertex groups are connected only to the vertex stabilized by Q.
If any of the fixed set of generators u11 , . . . , u1g of a QH vertex groups in Λ1 , or a
generator of a dominant edge group in Λ1 , or an element in the ball of radius 2 in
the Cayley graph of L w.r.t. the generating set: s1 , . . . , sr , is not elliptic in ΓQ , or
it is contained in a conjugate of Q, then finitely many translates of the path that
is associated with sj in the limit tree Y covers the paths that are associated with
the elements: tℓ and u11 , . . . , u1g that generate Q.
Note that unlike the (fixed) positive generators of the QH vertex groups and
dominant edge groups in λ1 , and unlike the generating set, s1 , . . . , sr , path that
are associated with elements in the ball of radius 2 in the Cayley graph of L may
contain positively and negatively oriented subpaths. In case such an element is
not elliptic with respect to ΓQ , finitely many translates of a fixed positively or
negatively oriented subpath of the path that is associated with such element suffice
to cover the (positively oriented) paths that are associated with the fixed set of
generators of the QH vertex group Q.
These coverings of the paths that are associated with the fixed set of generators
of the QH vertex groups in Λ2 by translates of the paths that are associated with
generators of QH vertex groups in Λ1 , generators of dominant edge groups in Λ1 ,
and elements in the ball of radius 2 in the Cayley graph of L, will guarantee that
the lengths of the paths that are associated with these elements along the entire
shortening process that we present, are bounded below by some (global) positive
constant times the lengths of the paths that are associated with the fixed set of
generators of the QH vertex group.
At we did in the first step, we shorten the homomorphisms {h1n } using the
dominant modular group M XM od(Λ1 ). For each homomorphism h1n , we pick the
shortest homomorphism after precomposing with an element from M XM od(Λ1 )
that keeps the positivity of the given set of the images of the generators s1 , . . . , sr ,
and the positivity of the fixed sets of generators of the QH vertex groups in both
Λ1 and Λ2 , and the generators of the dominant edge groups in Λ1 . We further keep
the positivity and negativity of the positively and negatively oriented subpaths in
the paths that are associated with elements in the ball of radius 2 in the Cayley
graph of L w.r.t. s1 , . . . , sr . We require the shortened homomorphisms to keep the
lengths of the fixed set of generators of the QH vertex groups in Λ2 to be at least
the maximal length of the image of a generator of a dominant edge or vertex group
in Λ2 (the generators are chosen from the fixed finite sets of generators of each of
the vertex and edge groups).
We further require that after the shortening, if a path in the limit tree Y2 that
is associated with Λ2 and with one of the following:
(1) a generator of a QH vertex group in Λ1 or a generator of the dominant edge
80
groups in Λ1 .
(2) a positively or negatively oriented subpath in the paths that are associated with elements in the ball of radius 2 in the Cayley graph of L w.r.t.
s1 , . . . , sr .
passes through an edge with a dominant edge group in Λ2 , then after shortening
the path that is associated with such a generator contains at least a subpath that
is associated with the dominant edge group. This guarantees that the length of the
corresponding element remains bigger than the length of the dominant edge group
along the entire procedure. We denote the obtained (shortened) homomorphisms
{h2n }.
Note that the shortening procedures that we presented in section 2 preserve the
positivity of positively oriented paths. However, the way they are constructed can
be used to keep the positivity and the negativity of finitely many subpaths in a given
path (that may not be oriented, but can be divided into finitely many positively
and negatively oriented subpaths). Furthermore, the shortening procedure is constructed to keep the non-cancellability between the positive and negative subpaths
of a given (embedded) path.
By the shortening arguments that are proved in propositions 2.7 and 2.8, the
lengths of the images, under the shortened homomorphisms {h2n }, of the fixed set
of generators of the dominant QH vertex groups are bounded by some constant
c1 (that is independent of n) times the maximal length of the images of the fixed
generators of the dominant vertex and edge groups.
We pass to a subsequence of the homomorphisms {h2n } that converges into an
action of L on some real tree with an associated abelian decomposition ∆3 . If the
action of L is not faithful, the conclusions of theorem 6.1 follow. Hence, we may
assume that the action of L is faithful. We refine ∆3 to an abelian decomposition
Λ3 , precisely as we refined ∆2 to obtain Λ2 .
With Λ3 we associate its modular group M od(Λ3 ) and dominant modular group,
M XM od(Λ3 ). We first shorten using M od(Λ3 ), and if every convergent shortened
subsequence converges into a faithful action of L, we further use the dominant
modular group M XM od(Λ3 ).
In shortening using M XM od(Λ3 ), we keep the positivity of all the fixed (positive)
sets of generators of the QH vertex groups in Λ1 and Λ2 , and the generators of the
dominant edge groups in Λ1 and Λ2 . We also keep the positivity and negativity of all
the finitely many positive and negative subpaths of the paths that are associated
with the elements in the ball of radius 3 in the Cayley graph of L w.r.t. the
generating set s1 , . . . , sr .
As we did in shortening using Λ2 , for each of the above elements (generators
of QH vertex groups and dominant edge groups in Λ1 and Λ2 , and elements in
the ball of radius 3 in the Cayley graph of L) that are not elliptic in an abelian
decomposition ΓQ , that is associated with a QH vertex group Q in Λ3 , and is
obtained by collapsing Λ3 , we use elements from L to demonstrate that each of
the fixed set of generators of Q is covered by finitely many translates of the path
that are associated with these elements. These translates will demonstrate that
the lengths of the paths that are associated with these elements will be at least a
(fixed) positive constant times the length of the path that are associated with the
fixed generators of Q along the rest of the procedure.
We denote the obtained (shortened) homomorphisms h3n , and continue itera81
tively. If in all steps the obtained actions are faithful, we get an infinite sequence of
abelian decompositions, Λ1 , Λ2 , . . . . Given the infinite sequence of abelian decompositions, we define the stable dominant abelian decomposition, Θi0 , in the same
way as we did in the simplicial case (definition 6.2). From the convergent sequences
{hin }∞
n=1 , we choose a subsequence {fi } that has a subsequence that converges into
a faithful action of the limit groups L, with an associated abelian decomposition
Θi0 .
Proposition 6.4. Suppose that the limit group L is freely indecomposable, and
contains no non-cyclic abelian subgroups. It is possible to choose a sequence of pair
homomorphisms, fi : (S, L) → (F Sk , Fk ), such that:
(1) for each i, fi is a homomorphism from the sequence {hin }.
(2) the sequence fi has a subsequence that converges into a faithful action of L
on a real tree Y .
(3) With the action of L on the limit tree Y there is an associated abelian
decomposition ∆. By possibly refining ∆, using restrictions of the homomorphisms {fi } to some of its non-QH vertex groups, it is possible to obtain
the stable abelian decomposition, Θi0 .
Proof: For each index i, we choose fi to be a homomorphism hin (from the sequence
{hin }), that maps the fixed positive generators of all the QH vertex groups and the
generators of all the edge groups in Λ1 , . . . , Λi to F Sk . We further require that
the ratios between the lengths of the images of these generators, and the ratios
between the lengths of the images of all the elements in the ball of radius i in the
Cayley graph of L w.r.t. the generating set s1 , . . . , sr will be approximately the
ratios between the lengths of the paths that are associated with these elements in
the corresponding limit tree Yi+1 (the trees that are obtained as the limits of the
sequences {hin }. We further require that fi maps the elements in a ball of radius i
in the Cayley graph of L monomorphically into Fk .
The sequence {fi } has a subsequence that converges into a faithful action of L
on a real tree Y . Let ∆ be the abelian decomposition that is associated with the
action of L on Y . Since every elliptic element in the stable abelian decomposition
Θi0 must fix a point in Y , ∆ is dominated by Θi0 .
Lemma 6.5. Let Q be a QH vertex group in Θi0 that does not appear in any of
the abelian decompositions Λi , for i ≥ i0 .
If there is an non-peripheral element in Q that fixes a point in Y , then the entire
QH vertex group Q fixes a point in Y .
Proof: Suppose that a non-peripheral element q ∈ Q fixes a point in Y . q is contained in some ball Bm in the Cayley graph of L w.r.t. s1 , . . . , sr . Θi0 is the stable
dominant abelian decomposition of the sequence of abelian decompositions: Λ1 , . . . .
Hence, there must exist an abelian decomposition Λi , for some i > max(i0 , m), for
which either:
(i) q is a non-peripheral element in some QH vertex group in Λi .
(ii) q is hyperbolic in the abelian decomposition Λi .
According to the procedure that was used to construct the abelian decompositions,
Λ1 , . . . , if either (i) or (ii) hold for Λi , then the traces and the lengths of q in its
actions on the tress that are associated with the abelian decompositions Λi+1 , . . .
82
(i.e., the limit trees Yi+1 , . . . ), are bounded below by either a (global) positive
constant times the lengths of the fixed set of generators of the QH vertex group
that contains q in Λi (in case (i)), or by either a a (global) positive constant times
the lengths of a fixed set of generators in a QH vertex group in Λi , or the length of
a generator of an edge group in Λi that becomes a dominant edge group in some
Λi′ for some i′ ≥ i in case (ii).
By the properties of the stable abelian decomposition Θi0 , and the structure of
the procedure for the construction of the abelian decompositions Λ1 , . . . , for some
index j0 > i, the lengths of the fixed set of generators of a QH vertex group in Λi ,
or an edge group in Λi , that are contained in a QH vertex group Q in Θi0 , that does
not appear as a QH vertex group in any of the abelian decompositions Λj , j ≥ i0 ,
multiplied by some positive constants, have to be bigger than the lengths of the
fixed sets of generators of all the QH vertex groups and the lengths of generators of
all the edge groups in the abelian decompositions Λj , j ≥ j0 , for edge groups and
QH vertex groups that are contained in the QH vertex group Q in Θi0 .
Hence, if the element q is elliptic in the limit action on the real tree Y , all the
QH vertex groups and all the edge groups that appear in Λj for j ≥ j0 , and are
contained in Q, must fix points in Y , which means that the entire QH vertex group
Q in Θi0 that contains the non-peripheral element q must fix a point in Y .
Edge groups and the boundary elements of QH vertex groups in Θi0 , that appear
in some abelian decomposition Λj , for some j ≥ i0 , have to be fix points in the
limit tree Y . From the graph of groups Θi0 we take out all the edge groups and
all the QH vertex groups that appear in some abelian decomposition Λj for some
j ≥ i0 . We denote the obtained (possibly disconnected) graph of groups Θ̂. Θ̂
contains QH vertex groups and non-QH vertex groups that are connected to them.
We restrict the homomorphisms {fi } to the fundamental groups of each of the
connected components in Θ̂. By lemma 6.5 if a component of Θ̂ contains a QH
vertex group, then the restrictions of the sequence {fi } to the fundamental group
of that component subconverges into an action of a real tree, such that the abelian
decomposition that is associated with that action, contains one or more QH vertex
groups from Θ̂, that are also QH vertex groups in Θi0 .
We continue the refinement process by erasing these QH vertex groups from Θ̂,
and restrict the homomorphisms {fi } to the fundamental components of connected
components in the remaining graph of groups. By lemma 6.5, after finitely many
such revisions of the graph of groups Θ̂ (i.e. erasing the QH vertex groups that
were visible), we uncover all the QH vertex groups in Θi0 that do not appear as
QH vertex groups in an abelian decomposition Λj for j ≥ i0 .
Now, starting with the abelian decomposition ∆, that was read from the original
faithful action of the limit group L on the limit tree Y , possibly refining ∆ using
the QH vertex groups and the edge groups from Θi0 that appear in some abelian
decomposition Λj for some j ≥ i0 , and further refining the obtained decomposition
using the uncovered QH vertex groups, we finally obtain the stable dominant abelian
decomposition Θi0 . This concludes the proof of proposition 6.4.
Let j0 ≥ i0 be an index for which all the QH vertex groups in Θi0 that appear in
some Λj for some j ≥ i0 , already appear in Λj for some j0 ≤ j ≤ i0 . Proposition 6.4
enables us to replace the suffix of the sequence of abelian decompositions, Λi0 , . . . ,
83
with a finite resolution Λi0 , . . . , Λj0 , Θi0 , and hence the entire sequence Λ1 , . . .
with the finite resolution: Λ1 , . . . , Λj0 , Θi0 . Furthermore, we may continue with
the sequence of pair homomorphisms {fi }, that are obtained from a subsequence
of the pair homomorphisms that we started with {hn }, by precomposing them
with automorphisms from the modular groups that are associated with the abelian
decompositions that appear along the finite resolution.
We continue to the next step starting with the sequence of homomorphisms
{fi }, and the abelian decomposition Θi0 . In the abelian decomposition that is obtained from a subsequence of homomorphisms that were obtained from the homomorphisms {fi } by precomposing them with automorphisms from the (dominant)
modular group of Θi0 , either a dominant edge group or a dominant non-QH vertex
group is not elliptic.
At this point we repeat the whole construction of a sequence of abelian decompositions. If the sequence terminates after a finite number of steps the conclusion
of theorem 6.1 follows. Suppose it ends up with an infinite sequence of abelian
decompositions.
Note that for any abelian decomposition Λi along this sequence, after finitely
many steps there exists an abelian decomposition Λi′ for some i′ ≥ i, with a dominant edge group. This means that the dominant edge group in Λi′ is not elliptic
in Λi′ +1 . Furthermore, as long as the boundary elements of a QH vertex group Q
in Λi remain elliptic in the next abelian decompositions, the QH vertex group Q
remains a QH vertex group in the next abelian decompositions.
Using proposition 6.4, we can replace the suffix of the sequence with a finite
resolution that stably dominates the entire sequence (see definition 6.2).
We continue iteratively. At each step we start with a sequence of homomorphisms
that are obtained from a subsequence of the homomorphisms {hn } by precomposing
them with automorphisms from the modular groups of the abelian decompositions
that appear along the previously constructed finite resolution. We continue for a
single step, so that at least one dominant edge group or a dominant vertex group
in the previous abelian decomposition is not elliptic. Then we either associate with
the shortened sequence of homomorphisms a finite resolution, that completes the
proof of theorem 6.1, or we associate with it an infinite sequence of abelian decompositions. By proposition 6.4 a suffix of this last infinite sequence can be replaced
by a finite resolution that ends with a stable dominant abelian decomposition of it.
If this iterative procedure terminates after finitely many steps, the conclusion
of theorem 6.1 follows. Otherwise we obtained an infinite sequence of abelian decompositions. Note that the infinite sequence of abelian decompositions contains a
subsequence of abelian decompositions that contain QH vertex groups, and dominant edge groups in the abelian decompositions from the subsequence are not
elliptic in the abelian decomposition that appears afterwards in the sequence.
Now, we apply proposition 6.4 to the sequence of abelian decompositions that
we constructed. By proposition 6.4 a suffix of the sequence can be replaced with a
finite resolution that terminates with the abelian decomposition that stably dominates the original suffix of the sequence. Since the sequence of abelian decompositions contains a subsequence with QH vertex groups, the stable dominant abelian
decomposition must contain a QH vertex group as well. Since every abelian decomposition in this subsequence contains a (dominant) edge that is not elliptic in the
next abelian decomposition, the stable dominant abelian decomposition contains
either 2 QH vertex groups, or a single QH vertex group with an associated surface
84
group S that is either:
(i) an orientable surface with χ(S) ≤ −2.
(ii) a non-orientable surface with genus(S) + bnd(S) ≥ 3, where bnd(S) is the
number of boundary components of the surface S.
We repeat the whole construction starting with the abelian decomposition that
we obtained and the subsequence of homomorphisms that is associated with it according to the construction that is used in the proof of proposition 6.4. Either the
construction terminates in finitely many steps, or we get a sequence of abelian decompositions that has a subsequence that satisfies the properties that the previous
stably dominant abelian decomposition satisfied, and in each abelian decomposition
from the subsequence there exists a dominant edge group that is not elliptic in the
next abelian decomposition.
Once again we apply proposition 6.4 to the constructed sequence of abelian
decompositions. The stable dominant abelian decomposition of the sequence has
to contain QH vertex groups with topological complexity that is bounded below by
a larger lower bound. Hence, it either contains at least 3 QH vertex groups, or 2
QH vertex groups so that at least one of them satisfies properties (1) or (2), or a
single QH vertex group with an associated surface group S that is either orientable
with χ(S) ≤ −3 or non-orientable with genus(S) + bnd(S) ≥ 4.
We continue iteratively. Since after each iteration (of the entire construction) we
obtain an abelian decomposition that contain QH vertex groups with topological
complexity that is bounded below by larger and larger bounds, and since the obtained abelian decompositions are all dominated by the JSJ decomposition of the
freely-indecomposable limit group L, the procedure has to terminate after finitely
many iterations, and the obtained resolution satisfies the conclusions of theorem
6.1.
So far we assumed that L is freely indecomposable and contains no non-cyclic
abelian groups. To get a conclusion similar to the one that appears in theorem
6.1 in the presence of non-cyclic abelian subgroups, we need to slightly modify it.
Instead of using only the modular group or the dominant modular group, we need
to further allow generalized Dehn twists in the presence of a non-cyclic abelian
group, i.e., we further allow Dehn twists in roots of the the values of a generator
of an edge group (in case the edge group belongs to a non-cyclic maximal abelian
subgroup), and not just modular automorphisms.
Definition 6.6. Let (S, L) be a pair in which L is freely indecomposable limit
group, and let Λ be an abelian decomposition that is associated with this pair (L is
the fundamental group of Λ). Let h : (S, L) → (F Sk , Fk ) and f : (S, L) → (F Sk , Fk )
be two pair homomorphisms. We say that f is obtained from h using generalized
Dehn twists if:
(i) there exists a pair homomorphism ĥ : (S, L) → (F Sk , Fk ) that is obtained
from h by precomposing it with a modular automorphism of Λ: ĥ = h ◦ ϕ,
ϕ ∈ M od(Λ).
(ii) suppose that A < L is a non-cyclic maximal abelian subgroup, and A is
not elliptic in Λ. Let A0 be the maximal abelian subgroup in A that is
elliptic in Λ, and let A = A0 + < a1 , . . . , aℓ >. Then f is obtained from ĥ
by modifying the values of the (non-elliptic) generators a1 , . . . , aℓ , that is
85
replacing ĥ(ai ) by elements in Fk that are in the maximal cyclic subgroup
that contains ĥ(A), in (possibly) all the (finitely many conjugacy classes of )
non-cyclic maximal abelian subgroups A < L that are not elliptic in Λ.
We say that a pair homomorphism f is obtained from a homomorphism h using
dominant generalized Dehn twists if:
(iii) there exists a pair homomorphism ĥ : (S, L) → (F Sk , Fk ) that is obtained
from h by precomposing it with a dominant modular automorphism of Λ:
ĥ = h ◦ ϕ, ϕ ∈ M XM od(Λ).
(iv) f is obtained from ĥ by modifying the values of the (non-elliptic) generators
a1 , . . . , aℓ , that is replacing ĥ(ai ) by elements in Fk that are in the maximal cyclic subgroup that contains ĥ(A), in (possibly) all the (finitely many
conjugacy classes of ) non-cyclic maximal abelian subgroups A < L that are
not elliptic in Λ, and in which the maximal elliptic subgroup A0 < A is
dominant.
Note that in general f is not obtained from h by a precomposition with a modular
automorphism, and that apart from the degenerate case in which f (A) is trivial, the
relation of being obtained using (dominant) generalized Dehn twists is symmetric.
Also, note that generalized Dehn twists are used in studying systems of equations
with parameters over a free group (see sections 9-10 in [Se1]).
The addition of generalized Dehn twists enable us to use our treatment of noncyclic abelian subgroups in sections 2 (theorem 2.1) and 3 and generalize the statement of theorem 6.1 to include all pairs (S, L) with L a general freely indecomposable limit group.
Theorem 6.7. Let (S, L) be a pair, where L is a freely indecomposable limit group,
and let s1 , . . . , sr be a fixed generating set of the semigroup S. Let {hn : (S, L) →
(F Sk , Fk )} be a sequence of pair homomorphisms that converges into a faithful
action of L on a real tree Y .
Then there exists a resolution:
(S1 , L1 ) → (S2 , L2 ) → . . . → (Sm , Lm ) → (Sf , Lf )
that satisfies the following properties:
(1) (S1 , L1 ) = (S, L), and ηi : (Si , Li ) → (Si+1 , Li+1 ) is an isomorphism for
i = 1, . . . , m − 1 and ηm : (Sm , Lm ) → (Sf , Lf ) is a proper quotient map.
(2) with each of the pairs (Si , Li ), 1 ≤ i ≤ m, there is an associated abelian
decomposition that we denote Λi .
(3) there exists a subsequence of the homomorphisms {hn } that factors through
the resolution. i.e., each homomorphism hnr from the subsequence, is obtained from a homomorphism of the terminal pair (Sf , Lf ) using a composition of a modification that uses generalized Dehn twists that are associated with Λm , and modular automorphisms that are associated with
Λ1 , . . . , Λm−1 .
Proof: Suppose that L is freely indecomposable and does contain a non-cyclic
abelian group. If all the non-cyclic abelian groups in L remain elliptic in all the
abelian decompositions that are constructed along the iterative procedure that was
86
used in the proof of theorem 6.1 in case there is no non-cyclic abelian groups, i.e., in
the iterative applications of the construction that is used in the proof of proposition
6.4, the same construction proves the conclusion of theorem 6.1.
Suppose that at some step along an application of the iterative procedure that
is used in the proof of proposition 6.4, a non-cyclic (maximal) abelian subgroup
A < L is non-elliptic in an abelian decomposition Λi , that is associated with a
corresponding (faithful) action of L on a corresponding limit tree Yi .
In that case A is either the set stabilizer of an axial component or it is the set
stabilizer of a line in the simplicial part of Yi . Let A0 < A be the point stabilizer
of the axial component with set stabilizer A or of the axis of A.
A0 is the stabilizer of an edge in Λi . If A0 is not dominant, we don’t include
Dehn twists along elements of A in the dominant modular group of Λi , and proceed
with the procedure that is used in the proof of theorem 6.1 as long as A0 is not
dominant. The abelian decompositions that we consider in all the steps of the
procedure in which A0 is not dominant, are abelian decompositions relative to A0 .
If A0 remains not dominant along the entire procedure, the procedure that is
used in proving proposition 6.4 and its conclusions remain valid. Suppose that at
some step j, A0 is dominant. A = A0 + < a1 , . . . , aℓ >. By theorem 2.1 in the
axial case, and in case A acts simplicially, when A0 is dominant we can modify
a subsequence of the homomorphisms {hn } using generalized Dehn twists along
A (that do not change A0 ), so that the images of a1 , . . . , aℓ under the modified
homomorphisms are identified with fixed elements in A0 . Hence the modified sequence of homomorphisms converges into a limit group in which the image of A is
A0 . Therefore, the modified sequence of homomorphisms converges into a proper
quotient of the limit group L, and the conclusion of theorem 6.6 follows.
Theorem 6.7 proves that given a pair (S, L) in which L is freely indecomposable,
and a sequence of pair homomorphisms {hn : (S, L) → (F Sk , Fk )}, that converges
into a faithful action of L on some real tree, it is possible to extract a subsequence
that factors through a finite resolution of the pair (S, L) that terminates in a proper
quotient of the pair (S, L). Using a compactness argument it is not difficult to apply
the conclusion of theorem 6.7 and get a Makanin-Razborov diagram for a pair
(S, L), when L is a restricted limit group with no freely decomposable restricted
limit quotients.
Let F Sk =< a1 , . . . , ak > be a free semigroup that generates the free group Fk ,
and let (S, L) be a restricted pair, i.e., a pair that contains the subpair (F Sk , Fk ).
Suppose that the pair (S, L) does not admit a quotient restricted map η : (S, L) →
(Ŝ, L̂) with the following properties:
(1) η maps the subpair (F Sk , Fk ) < (S, L) monomorphically onto the corresponding subpair (F Sk , Fk ) < (Ŝ, L̂).
(2) L̂ admits a non-trivial free decomposition in which η(Fk ) is contained in a
factor.
The restricted pair (S, L) is in particular freely indecomposable with respect
to the subpair (F Sk , Fk ), and so is every restricted quotient of (S, L). Given a
sequence of homomorphisms {hn : (S, L) → (F Sk , Fk )}, it is possible to extract
a subsequence that converges into a faithful action of some restricted quotient
(S0 , L0 ) of (S, L). By theorem 6.7 it is possible to further extract a subsequence
that factors through a resolution that terminates in a proper quotient of (S0 , L0 ).
87
Applying theorem 6.7 iteratively, finitely many times, it is possible to extract a
further subsequence that factors through a resolution:
(S0 , L0 ) → (S2 , L2 ) → . . . → (Sm , Lm ) → (F Sk , Fk )
that terminates in the standard pair (F Sk , Fk ), and in which some of the epimorphisms: ηi : (Si , Li ) → (Si+1 , Li+1 ) are isomorphisms and some are proper quotient
maps. with each of the pairs (Si , Li ), 1 ≤ i ≤ m, there is an associated abelian
decomposition that we denote Λi , with which we naturally associate a modular
group.
Although f.g. subsemigroups of limit groups need not be f.p. in general, a pair
(S, L) in which L is a limit group and S is a f.g. subsemigroup of L is naturally a
finitely presented object. Hence, a resolution of the form:
(S0 , L0 ) → (S2 , L2 ) → . . . → (Sm , Lm ) → (F Sk , Fk )
together with the abelian decompositions that are associated with the various restricted pairs (Si , Li ) can be encoded using a finite amount of data. In particular,
there are only countably many such resolutions (that are associated with convergent restricted pair homomorphisms of a given pair (S, L)), and we can order these
resolutions using their finite encoding.
Theorem 6.8. Let (S, L) be a restricted pair, that has no restricted quotient (Ŝ, L̂)
in which L̂ admits a free decomposition relative to (the embedding of ) Fk . Then
there are finitely many resolutions of the form that is constructed in theorem 6.7:
(S0 , L0 ) → (S2 , L2 ) → . . . → (Sm , Lm ) → (F Sk , Fk )
where (S0 , L0 ) is restricted quotient pair of (S, L) , such that:
(1) every restricted pair homomorphism, h : (S, L) → (F Sk , Fk ), factors through
at least one of these finitely many resolutions.
(2) for each of the resolutions in the collection, there exists a sequence of homomorphisms: {hn : (S, L) → (F Sk , Fk )}, that converges into a faithful
action of the initial pair (S0 , L0 ) on a real tree with an associated abelian
decomposition (after the refinement that is used in the proof of proposition
6.4) Λ0 .
Furthermore, the sequence of homomorphisms {hn } can be modified using modular automorphisms and generalized Dehn twists that are associated
with the abelian decompositions: Λ0 , . . . , Λm , to get sequences of pair homomorphisms {h1n }, . . . , {hm
n }. Each of these modified sequences of homomorphisms {hin } converges into a faithful action of the pair (Si , Li ) on a
real tree.
Proof: Let {h | h : (S, L) → (F Sk , Fk )} be the collection of all the restricted pair
homomorphisms of the pair (S, L). We look at all the possible subsequences of such
homomorphisms {hn }, that converge into quotients of (S, L), with which we can
associate a resolution:
(S0 , L0 ) → (S2 , L2 ) → . . . → (Sm , Lm ) → (F Sk , Fk )
88
by iteratively applying theorem 6.7. Note that each of these constructed resolutions
satisfy part (2) of the theorem.
Each such resolution can be encoded using finite amount of data, in particular,
there are countably many such resolutions, and we can order them using the finite
encoding. We argue that the collection of all the pair homomorphisms of (S, L)
factor through a finite collection of these resolutions.
Suppose that finitely many do not suffice. Then there exists a sequence of restricted pair homomorphisms: {hn }, such that for every index n, hn does not factor
through the first n resolutions from the ordered countable set of resolutions that
were constructed. By iteratively applying theorem 6.7, from the sequence {hn }
we can extract a subsequence {hnr }, that factors through a resolution of form we
previously constructed:
(S0 , L0 ) → (S2 , L2 ) → . . . → (Sm , Lm ) → (F Sk , Fk )
But this resolution appears in the countable set of resolutions we associated with
the pair (S, L), hence, it appears in the ordered list. Therefore, for some index r0 ,
and for every r > r0 , the subsequence of homomorphisms: {hnr } factors through
a resolution that appears in the ordered list of resolutions, a contradiction to the
choice of the homomorphisms {hn }.
Theorem 6.8 constructs a diagram that encodes all the homomorphisms from
a pair (S, L) into the standard pair (F Sk , Fk ). However, the construction of the
resolutions in the diagram, that mainly uses the iterative procedure that was used
in proving theorem 6.1 and proposition 6.4, does not guarantee that there exist sequences of homomorphisms that factor through them for which the corresponding
shortened homomorphisms converge into the pairs that appear along the resolutions. i.e., the construction of the resolutions does not guarantee the existence of
test sequences or generic points for the resolutions in the diagram. To guarantee
the existence of such test sequences, we need to slightly modify the sequences of
homomorphisms that are used in the construction of the resolutions, i.e., those that
are used in the proof of proposition 6.4.
Theorem 6.9. Let (S, L) be a restricted pair, that has no restricted quotient (Ŝ, L̂)
in which L̂ admits a free decomposition relative to (the embedding of ) Fk . Then
there are finitely many resolutions of the form that is constructed in theorem 6.7:
(S0 , L0 ) → (S2 , L2 ) → . . . → (Sm , Lm ) → (F Sk , Fk )
where (S0 , L0 ) is restricted quotient pair of (S, L) , such that parts (1) and (2) in
theorem 6.8 hold for these resolutions, and in addition:
(3) for each of the resolutions in the collection, there exists a sequence of homomorphisms: {hn : (S, L) → (F Sk , Fk )}, that can be modified using modular automorphisms and generalized Dehn twists that are associated with
the abelian decompositions: Λ0 , . . . , Λm , to get sequences of pair homomorphisms {h1n }, . . . , {hm
n }. Each of these modified sequences of homomori
phisms {hn } converges into a faithful action of the pair (Si , Li ) on a real tree
with an associated abelian decomposition (after an appropriate refinement)
Λi .
89
Proof:
We call a finite collection of resolution that satisfies properties (1) and (2) in
theorem 6.8 and part (3) in theorem 6.9 a M akanin-Razborov diagram of the
restricted pair (S, L). Note that such a diagram is not canonical in general. We
view sequences of pair homomorphisms that satisfy part (3) with respect to one of
the resolutions in the diagram as generic points in the variety that is associated with
the pair (S, L). We later use such sequences of homomorphisms as a replacement
to test sequences that were used in [Se2] to construct formal solutions and obtain
generalized Merzlyakov theorems for AE sentences and formulas that are defined
over a given variety over a free semigroup.
§7. A Makanin-Razborov diagram
In section 6 we analyzed the collection of homomorphisms from a restricted pair
(S, L) into the standard pair (F Sk , Fk ), in case the pair (S, L) has no restricted
quotients (Ŝ, L̂), in which the restricted limit group L̂ is freely decomposable (with
respect to the coefficient subgroup Fk ), i.e., in which L is freely decomposable as a
restricted limit group.
In that case we managed to associate a Makanin-Razborov diagram with such a
restricted pair, that encodes all its pair homomorphisms, such that every resolution
in the diagram has a collection of generic points of homomorphisms that factor
through it (see theorem 6.9). Each of the resolutions in such a diagram terminates
with the standard pair (F Sk , Fk ).
In this section we generalize the construction of the Makanin-Razborov diagram
to include all possible pairs (S, L). To do that we need to generalize theorems 6.1
and 6.6 to include pairs with freely decomposable limit groups. We first present such
generalizations in case there are no Levitt components in the actions on the real
trees that we consider, and then omit this assumption, and allow Levitt components.
In both cases we use the machinery that was presented in section 6, to construct
a JSJ-like decompositions, that in the general case considers and encodes Levitt
components.
Definition 7.1. Let (S, L) be a pair. We say that (S, L) is Levitt-f ree if every
sequence of pair homomorphisms: {hn : (S, L) → (F Sk , Fk )} that converges into a
faithful action of L on some real tree, contains no Levitt (thin) components (for a
detailed description of Levitt components see [Be-Fe1] who call them thin).
Theorem 7.2. Let (S, L) be a Levitt-free pair, and suppose that the limit group
L contains no non-cyclic abelian subgroup. Let {hn : (S, L) → (F Sk , Fk )} be a
sequence of pair homomorphisms that converges into a faithful action of L on a real
tree Y .
Then there exists a resolution:
(S1 , L1 ) → . . . → (Sf , Lf )
that satisfies the following properties:
(1) (S1 , L1 ) = (S, L), and ηi : (Si , Li ) → (Si+1 , Li+1 ) is an isomorphism for
i = 1, . . . , f − 2 and ηf −1 : (Sf −1 , Lf −1 ) → (Sf , Lf ) is a quotient map.
90
(2) with each of the pairs (Si , Li ), 1 ≤ i ≤ f , there is an associated abelian
decomposition that we denote Λi . The abelian decompositions Λ1 , . . . , Λf −1
contain edges with trivial and cyclic edge stabilizers, QH vertex groups that
are associated with IET components, and rigid vertex groups.
(3) either ηf −1 is a proper quotient map, or the abelian decomposition Λf contains separating edges with trivial edge groups. Each separating edge is
oriented.
(4) there exists a subsequence of the homomorphisms {hn } that factors through
the resolution. i.e., each homomorphism hnr from the subsequence, can be
written in the form:
hnr = ĥr ◦ ϕrf −1 ◦ . . . ◦ ϕ1r
where ĥr : (Sf , Lf ) → (F Sk , Fk ), each of the automorphisms ϕir ∈ M od(Λi ),
where M od(Λi ) is generated by the modular groups that are associated with
the QH vertex groups, and by Dehn twists along edge groups with cyclic
stabilizers in Λi .
Each of the homomorphisms:
i
hinr = ĥr ◦ ϕm
r ◦ . . . ◦ ϕr
is a pair homomorphism hinr : (Si , Li ) → (F Sk , Fk ).
(5) if (Sf , Lf ) is not a proper quotient of (S, L), then the pair homomorphisms
ĥr are compatible with Λf . Let R1 , . . . , Rv be the connected components of
Λf after deleting its (oriented) separating edges. The homomorphisms ĥr
are composed from homomorphisms of the fundamental groups of the connected components R1 , . . . , Rv , together with assignments of values from
F Sk to the oriented separating edges. The homomorphisms of the fundamental groups of the connected components R1 , . . . , Rv converge into a
faithful action of these groups on real trees with associated abelian decompositions: R1 , . . . , Rv .
Proof: We modify the procedure that was used in proving theorem 6.1. Let Λ be
the abelian decomposition that is associated with the action of L on the limit tree
Y (that is obtained from the convergent sequence {hn }). If Λ contains a segment
in its simplicial part, and that segment has a trivial stabilizer, the conclusions of
the theorem follow.
By our assumptions, (S, L) is Levitt free, so as long as (S, L) is not replaced
by a proper quotient, none of the faithful actions of L on the limit trees that
are constructed along the procedure contain Levitt components. Suppose that
the action of L on the corresponding real tree is not geometric (in the sense of
[Be-Fe]). In that case we associate with the non-geometric action (with no Levitt
components), an approximating resolution (in the sense of [Be-Fe]) according to
the one that we constructed in theorem 2.9. The graph of groups that is associated
with that approximating resolution contains edges with trivial stabilizers, and a
subsequence of the given sequence of homomorphisms satisfies parts (4) and (5) of
the theorem according to theorem 2.9.
Therefore, we may assume that the action of L on the limit tree Y is geometric.
Hence, Y contains only IET and discrete components, and every segment in the
91
discrete part of Y can be divided into finitely many non-degenerate segments with
(non-trivial) cyclic stabilizers. Therefore, if L is a free product of non-cyclic, freely
indecomposable limit groups, and Λ is the JSJ decomposition of L, L is replaced by
a proper quotient after shortening along the modular group of L, and the conclusion
of the theorem follows.
We start by refining the abelian decomposition Λ in a similar way to what we
did proving theorems 3.2 and 6.1. If a non-QH vertex group is connected only to
periodic edge groups (and in particular it is not connected to a QH vertex group),
we restrict the homomorphisms {hn } to such a vertex group and obtain a nontrivial splitting of it in which all the previous (periodic) edge groups are elliptic.
Hence, the obtained abelian decomposition of the non-QH vertex group can be
used to refine the abelian decomposition Λ. Repeating this refinement procedure
iteratively, we get an abelian decomposition that we denote Λ1 .
We fix finite generating sets of all the edge groups and all the non-QH vertex
groups in Λ1 . We divide the edge and non-QH vertex groups in Λ1 into finitely many
equivalence classes of their growth rates as we did the proof of theorem 6.1. Two
non-QH vertex groups or edge groups (or an edge and a non-QH vertex group) are
said to be in the same equivalence class if the maximal length of the images of their
finite set of generators have comparable lengths, i.e., the length of a maximal length
image of one is bounded by a (global) positive constant times the maximal length
of an image of other, and vice versa. After possibly passing to a subsequence, there
exists a class that dominates all the other classes, that we call the dominant class,
that includes (possibly) dominant edge groups and (possibly) dominant non-QH
vertex groups.
As in the freely indecomposable case, we denote by M od(Λ1 ) the modular group
that is associated with Λ1 . We set M XM od(Λ1 ) to be the dominant modular
group that is generated by Dehn twists along dominant edge groups and modular
groups of dominant QH vertex groups, i.e., those QH vertex groups that the lengths
of the images of their fixed sets of generators grow faster than the lengths of the
images of the fixed generators of dominant edge and vertex groups.
We start by using the full modular group M od(Λ1 ). For each index n, we set the
pair homomorphism: h1n : (S, L) → (F Sk , Fk ), h1n = hn ◦ ϕn , where ϕn ∈ M od(Λ1 ),
to be a shortest pair homomorphism that is obtained from hn by precomposing
it with a modular automorphism from M od(Λ1 ). If there exists a subsequence
of the homomorphisms {h1n } that converges into a proper quotient of the pair
(S, L), or that converges into a non-geometric action of L on a real tree, or into an
action that contains a segment in its simplicial part, and this segment can not be
divided into finitely many subsegments with trivial stabilizers, we set the limit of
this subsequence to be (Sf , Lf ), and the conclusion of the theorem follows with a
resolution of length 1.
Therefore, we may assume that every convergent subsequence of the homomorphisms {h1n } converges into a faithful action of the limit group L on some real tree.
In that case we use only elements from the dominant modular group, M XM od(Λ1 ).
Before we continue to the next abelian decomposition, we need to check if the
original sequence of homomorphisms, {hn }, does not contain a subsequence of
separable homomorphisms. We are going to look for subsequences of separable
homomorphisms in every step of the iterative procedure for the construction of the
sequence of abelian decompositions Λ1 , . . . . The existence of such subsequence will
lead to a termination of the procedure, with an abelian decomposition that satisfies
92
the conclusions of theorem 7.2.
Definition 7.3. Let (S, L) be a pair, and let: {un : (S, L) → (F Sk , Fk )} be a
sequence of pair homomorphisms that converges into a faithful action of L on some
real tree Y . We say that the sequence {un } is separable if there exists a (reduced)
graph of groups ∆ with fundamental group L with the following properties and
additional data:
(1) the graph of groups ∆ is non-trivial and all its edges have trivial stabilizers.
We further assign orientation with each edge in ∆.
(2) with each non-trivial vertex group in ∆ we associate a base point. In addition there is a base point for the fundamental group L, that is placed in the
interior of one of the edges or it is one of the basepoints that are associated
with the vertex groups.
(3) with each edge in ∆ we associate a label. If the basepoint of L is in the interior of an edge, then with that edge there are two labels that are associated
with the two parts of the edge.
(4) the homomorphisms {un } are composed from homomorphisms of the vertex
groups into Fk , and assignments of values in Fk to the labels that are associated with the edges in ∆. Each element in L can be considered as a path
in ∆ that starts and ends in its basepoint. Hence, from the values that are
assigned to the labels, and the homomorphisms of the vertex groups (with
their basepoints), it is possible to read (uniquely) a homomorphism of L into
Fk .
(5) we extend the set of generators of the standard pair (F Sk , Fk ), by adding a
new free generator to the standard semigroup F Sk for each label in ∆. We
denote the extended standard pair (F Sm , Fm ).
Given each of the homomorphisms un , it is possible to replace the values that are assigned with each of the labels that are associated with the
edges in ∆, to values that contain a single (positively oriented) appearance
of the generator that is associated with each label, and no appearances of
generators that are associated with the other labels, without changing the
homomorphisms of the vertex groups in ∆ to obtain a homomorphism ûn .
The homomorphism ûn should coincide with the original homomorphism
un if we map the (new) generators that are associated with the labels to the
identity. Furthermore, and for each index n, the homomorphism ûn is a
pair homomorphism: ûn : (S, L) → (F Sm , Fm ).
If the sequence of homomorphisms, {hn }, contains a separable subsequence, the
conclusion of theorem 7.2 follows.
Proposition 7.4. Let (S, L) be a Levitt-free pair, and suppose that the limit group
L contains no non-cyclic abelian subgroup. Let {hn : (S, L) → (F Sk , Fk )} be a
sequence of pair homomorphisms that converges into a faithful action of L on a real
tree Y .
Let {ut : (S, L) → (F Sk , Fk )} be a separable subsequence of the sequence {hn }.
Then part (5) of theorem 7.2 holds for the subsequence {ut }.
Proof: With this separable subsequence there is an associated abelian decomposition ∆ that satisfies all the properties that are listed in definition 7.3. Given ∆,
93
and the separable sequence, {ut }, for every index t, we can associate with the homomorphism ut , an extended pair homomorphism: ût : (S, L) → (F Sm , Fm ) that
satisfies the properties that are listed in part (5) of definition 7.3.
The sequence {ût } has a convergent subsequence (still denote {ût }). This convergent subsequence converges into a faithful action of L on a real tree, Ŷ . The
abelian decomposition, Θ, that is associated with the action of L on Ŷ contains separating edges, that appear because of the extra free generators that appear (once)
in the values that are assigned to the labels that are associated with the edges in
∆. In particular, with each edge in ∆ there is a corresponding separating edge in
Θ. Hence, the conclusion of theorem 7.2 holds for the abelian decomposition, Θ.
For presentation purposes we first assume that Λ1 , and all the next abelian
decompositions that are obtained along the iterative procedure do not contain any
QH vertex groups. In that case we continue iteratively, precisely as we did in the
freely indecomposable case.
First we shorten the homomorphisms {hn } using the dominant modular group,
M XM od(Λ1 ). We denote the shortened sequence of homomorphisms, {h1n }, and
after passing to a subsequence, assume that the obtained sequence converges into a
faithful action on a real tree Y2 . If the action of the limit group L on the real tree
Y2 is not faithful, or if there is a segment in the simplicial part of Y2 that can not
be divided into subsegments with non-trivial stabilizers, or if the action of L on Y2
is not geometric, the conclusions of theorem 7.2 follow.
Hence, we may assume that the action of L on Y2 is geometric and faithful.
With the action of L on Y2 there is an associated abelian decomposition ∆2 . By
our standard procedure ∆2 can be (possibly) refined to an abelian decomposition
Λ2 . By our assumptions, Λ2 contains no QH vertex groups.
We first check if the sequence of abelian decomposition {h1n } contains a separable
subsequence. If it does theorem 7.2 follows. Otherwise, we associate with Λ2 its
modular group, M od(Λ2 ), and its dominant modular group, M XM od(Λ2 ). We
first shorten using the ambient modular group, and if the action is faithful, we
shorten using the dominant modular group.
We continue iteratively, and either terminate after finite time, or obtain the
sequence of abelian decompositions, Λ1 , . . . , and the corresponding sequences of
convergent shortened homomorphisms, that have no separable subsequences. As
in the freely indecomposable case, with the sequence of abelian decompositions,
Λ1 , . . . we associate its stable dominant (abelian) decomposition.
Definition 7.5. Let (S, L) be a pair, and let Λ1 , . . . be a sequence of abelian decompositions of L. With the sequence we associate its stable dominant abelian
decompositions, that generalizes the one presented in definition 6.2, in the case of
freely decomposable groups.
Note that given two abelian decompositions ∆1 and ∆2 of a limit group L, their
common refinement is the multi-graded abelian JSJ decomposition of L with respect
to the collection of (finitely many conjugacy classes of ) subgroups that are elliptic
in both splittings, ∆1 and ∆2 . The multi-graded abelian JSJ decomposition is an
abelian decomposition of a freely decomposable group, and hence includes free products, and starts with the multi-graded Grushko decomposition. For the definition of
the multi-graded JSJ decomposition see section 12 in [Se1].
94
Associating an abelian decomposition (a common refinement) with every pair of
abelian decompositions of L, enables one to associate such a common refinement
with every finite sequence of abelian decompositions of L. Given a sequence of
abelian decompositions, Λ1 , . . . , the (natural) complexities of the common refinements of the prefixes, Λ1 , . . . , Λℓ , does not decrease with ℓ, and all these decompositions are bounded by the abelian JSJ decomposition of the limit group L (the
abelian JSJ of a freely decomposable group that starts with the Grushko decomposition, and then associates the abelian JSJ decomposition with each of the non-cyclic
freely indecomposable factors). Hence, the common refinements of the prefixes of
the sequence Λ1 , . . . stabilize, and it is possible to associate a common refinement
with the entire sequence.
Therefore, it is possible to associate an abelian decomposition (common refinement) with every suffix of the sequence, Λℓ , . . . . The complexities of these abelian
decompositions do not increase with ℓ, and every sequence of strictly decreasing
abelian decompositions of L terminates after a finite (in fact, bounded) time. We
define the stable dominant abelian decomposition of the sequence Λ1 , . . . , to be the
minimal abelian decomposition that is associated with a suffix Λℓ , . . . , where the
minimum is with respect to all the suffixes of the sequence.
Let Θi0 be the stable dominant abelian decomposition of the sequence Λ1 , . . . .
Suppose that Θi0 does not collapse to any free factors, i.e., no edges with trivial
stabilizers (only QH and rigid vertex groups, and edges with non-trivial cyclic
stabilizers).
We assumed that the sequence of abelian decompositions Λ1 , . . . contains no QH
vertex groups. Hence, in case Θi0 contains no free products, the procedure that was
used to prove theorem 6.1 in case the abelian decompositions Λ1 , . . . contain no QH
vertex groups (cf. proposition 6.3), proves the existence of a sequence of homomorphisms that converges into an action of L on a real tree, with an associated abelain
decomposition, that after a possible refinement, is identical to Θi0 . Therefore, in
this case the infinite sequence Λ1 , . . . can be replaced by a finite sequence that ends
with Θi0 , precisely as we did in the freely indecomposable case.
Suppose that Θi0 collapses to free products. In that case, by the structure of
the iterative procedure for the construction of the abelian decompositions Λ1 , . . .
(proposition 6.3), there exists a sequence of shortened homomorphisms that converges into a geometric and faithful action of L on a real tree with an associated
abelian decomposition that can be further refined to be Θi0 . Since (S, L) is Levittfree the action on L on the limit tree contains only simplicial part and IET components. Using an argument that we explain in more detail in the presence of QH
vertex groups in the sequel, the limit action of L must be geometric and the limit
tree contains no segment in its simplicial part, that can not be divided into segments
with non-trivial stabilizers. Furthermore, there must exist an index i1 ≥ i0 , such
that all the edge groups in the abelian decompositions Λi , i > i1 , either appear as
edges in Θi0 , or they correspond to s.c.c. in a QH vertex groups in Θi0 . In particular, the modular groups that are associated with all the abelian decompositions,
Λi1 , . . . , are contained in the modular group that is associated with Θi0 . Therefore,
the sequence of abelian decompositions, Λ1 , . . . can be replaced by a finite sequence
that terminates with Θi0 .
At this point we still assume that the pair (S, L) is Levitt-free, but allow the
abelian decompositions in the sequel to be arbitrary, i.e., to contain QH vertex
95
groups. We have already assumed that starting with the sequence {hn } contains
no separable subsequence, and that shortening it using the ambient modular group
M od(Λ1 ), we get a sequence for which every subsequence is not separable (definition
7.3), and every convergent subsequence converges into a faithful action of L.
We use only the dominant modular group, M XM od(Λ1 ). We modify what we
did in the freely indecomposable case. First, we shorten the action of each of the
dominant QH vertex groups using the procedure that is used in the proof of propositions 2.7 and 2.8. For each of the dominant QH vertex groups, the procedures that
are used in the proofs of these propositions give us an infinite collections of posm
itive generators, um
1 , . . . , ug , with similar presentations of the corresponding QH
vertex groups, which means that the sequence of sets of generators belong to the
same isomorphism class. Furthermore, with each of these sets of generators there
associated words, wjm , j = 1, . . . , r, of lengths that increase with m, such that a
m
given (fixed) set of positive elements can be presented as: yj = wjm (um
1 , . . . , ug ).
m
m
The words wj are words in the generators ui and their inverses. However, they
can be presented as positive words in the generators um
j , and unique appearances
m
of words tℓ , that are fixed words in the elements uj and their inverses (i.e., the
m
m
words do not depend on m), and these elements tm
ℓ (u1 , . . . , ug ) are positive for
every m.
Given the output of the procedures that were used in propositions 2.7 and 2.8,
with each of the dominant QH vertex groups we associate a system of generators
u11 , . . . , u1g (the integer g depends on the QH vertex group). Since an IET action
of a QH vertex group is indecomposable in the sense of [Gu], finitely many (fixed)
translates of each of the positive paths that are associated with the positive paths,
u1i1 , cover the positive path that is associated with u1i2 , and the positive paths
that are associated the words tℓ (u11 , . . . , u1g ). As in the procedure in the freely
indecomposable case, the covering of the elements u1 i1 by finitely many translates
of elements u1i2 guarantee that the ratios between their lengths along the iterative
(shortening) procedure that we use remain globally bounded.
Let Q be a QH vertex group in Λ1 . With each of the generators sj , 1 ≤ j ≤ r,
we associate the (positive) path in the limit tree Y from the base point to the image
of the base point under sj (the path may be degenerate). Since the action of L on
the limit tree Y is geometric, such a path contains finitely many subpaths that are
contained in the orbit of an IET component that is associated with the QH vertex
group Q. Since the action of Q on its associated IET component is indecomposable,
given each non-degenerate subpath of the path that corresponds to a generator sj
and is contained in the IET component that is associated with Q, finitely many
translates of this non-degenerate subpath cover the paths that are associated with
the elements: tℓ and u11 , . . . , u1g that generate Q.
These coverings of the paths that are associated with the elements u11 , . . . , u1g by
translates of the subpaths that is associated with sj , will guarantee that the lengths
of the subpaths that are associated with sj , along the entire shortening process that
we present, are bounded below by some (global) positive constant times the lengths
of the paths that are associated with the elements, u11 , . . . , u1g along the process.
At this point we shorten the homomorphisms {hn } using the dominant modular
group M XM od(Λ1 ). For each homomorphism hn , we pick the shortest homomorphism after precomposing with an element from M XM od(Λ1 ) that keeps the
positivity of the given set of the images of the generators s1 , . . . , sr , and the posi96
tivity of the elements u11 , . . . , u1g and tℓ , and keeps their lengths to be at least the
maximal length of the image of a generator of a dominant edge or vertex group
(the generators are chosen from the fixed finite sets of generators of each of the
vertex and edge groups). We further require that after the shortening, the image
of each of the elements u11 , . . . , u1g and tℓ , will be covered by the finitely many
translates of them and of the paths that are associated with the relevant generators
s1 , . . . , sr , that cover them in the limit action that is associated with Λ1 . We further require that if a path that is associated with one of the generators, s1 , . . . , sr ,
passes through an edge with a dominant edge group in Λ1 , then after shortening
the path that is associated with such a generator contains at least a subpath that
is associated with the dominant edge group. This guarantees that the length of
such a generator remains bigger than the length of the dominant edge group along
the entire procedure. We (still) denote the obtained (shortened) homomorphisms
{h1n }.
By the shortening arguments that are proved in propositions 2.7 and 2.8, the
lengths of the images, under the shortened homomorphisms {h1n }, of the elements
u11 , . . . , u1g and tℓ , that are associated with the various dominant QH vertex groups,
and the lengths of the elements, s1 , . . . , sr , are bounded by some constant c1 (that
is independent of n) times the maximal length of the images of the fixed generators
of the dominant vertex and edge groups.
As in the freely indecomposable case, we pass to a subsequence of the homomorphisms {h1n } that converges into an action of L on some real tree with an associated
abelian decomposition ∆2 . If the action of L is not faithful, the conclusions of theorem 7.2 follow, hence, we may assume that the action of L is faithful. If the action
of L is not geometric or contains a non-degenerate subsegment in its simplicial part,
that can not be divided into finitely many subsegments with non-trivial stabilizers,
the conclusion of theorem 7.2 follow. If there exists a subsequence of the homomorphisms {h1n } that are separable (see definition 7.3), theorem 7.2 follow. Hence,
we may assume that there {h1n } does not contain a separable subsequence, that
the action of L is geometric, and that every segment in the simplicial part can be
divided into finitely many segments with non-trivial stabilizers.
We further refine ∆2 as we did in the freely indecomposable case., We restrict
(a convergent subsequence of) the homomorphisms {h1n } to non-QH vertex groups
that are connected only to periodic edge groups, as we did with Λ, and construct
ˆ 2.
an abelian decomposition that we denote ∆
Let Q be a dominant QH vertex group in Λ1 , so that all its boundary eleˆ 2 . i.e., every boundary element is contained in either an
ments are elliptic in ∆
ˆ 2 . With the original sequence of
edge group or in a non-QH vertex group in ∆
homomorphisms, {hn }, and the shortened sequence, {h1n }, we associate a new (intermediate) sequence of homomorphisms. Each homomorphism, h1n , is obtained
from hn , by precomposition with a (shortened) automorphism from the dominant
modular group, ϕn ∈ M XM od(Λ1 ). ϕn is a composition of elements from the
modular groups of dominant QH vertex group in Λ, and Dehn twists along edges
with dominant edge groups. We set τn ∈ M XM od(Λ1 ) to be a composition of the
same elements from the modular groups of dominant QH vertex groups in Λ1 , that
are not the dominant QH vertex group Q, and the same Dehn twists along edges
with dominant edge group as the shortened homomorphism ϕn . Let µn = hn ◦ τn .
The sequence of homomorphisms µn , converges into a faithful action of L on a
97
real tree, with an associated abelian decomposition ΓQ , that has a single QH vertex
group, a conjugate of Q, and possibly several edges with non-trivial edge groups,
ˆ 2.
that are edge in both Λ1 and ∆
ˆ 2 and ΓQ , have a common refineSuppose that the abelian decompositions, ∆
ment, in which a conjugate of Q appears as a QH vertex group, and all the edges and
ˆ 2 that do not correspond to s.c.c. or proper QH subgroups
QH vertex groups in ∆
of Q also appear in the common refinement (the elliptic elements in the common
ˆ 2 ). This,
refinement are precisely those elements that are elliptic in both ΓQ and ∆
in particular, implies that all the QH vertex groups that can not be conjugated into
ˆ 2 that can not be conjugated into non-peripheral
Q, and all the edge groups in ∆
elements in Q are elliptic in ΓQ . In case there exists such a common refinement,
ˆ 2 with this common refinement. We repeat this possible refinement
we replace ∆
for all the QH vertex groups in Λ1 that satisfy these conditions (the refinement
procedure does not depend on the order of the QH vertex groups in Λ1 that satisfy
the refinement conditions). We denote the obtained abelian decomposition, Λ2 .
At this point, we modify the shortened homomorphisms, {h1n }, that converge into
a faithful limit action of L on a limit tree from which the abelian decomposition Λ2 is
obtained, by precomposing them with a fixed automorphism, ψ1 , from the dominant
modular group M XM od(Λ1 ). This precomposition is needed to guarantee the
validity of certain inequalities between the lengths, or the ratios of lengths, of a
finite set of elements. This finite set include:
(1) the fixed set of generators of dominant QH vertex groups and generators of
dominant edge groups in Λ1 .
(2) the (fixed) generators, s1 , . . . , sr .
(3) the fixed set of generators of some of the QH vertex groups and some of the
edge groups in Λ2 .
Let Q21 , . . . , Q2f , and E12 , . . . , Ev2 , be those QH vertex groups and edge groups in
Λ2 , for which at least one of the fixed set of generators of the dominant QH vertex
groups, and dominant edge groups in Λ1 , is not elliptic with respect to them. i.e.,
if we collapse Λ2 to contain a single QH vertex group Q2i , ΓQ2i , or a single edge
group Ei2 , ΓEi2 , then there exists a generator of a dominant QH vertex group or
a dominant edge group in Λ1 that is mapped to a non-elliptic element in that
collapsed abelian decomposition.
Note that every non-dominant QH vertex group or edge group in Λ1 is a QH
vertex group or an edge group in Λ2 . Hence, in choosing a fixed automorphism
that we use in order to modify the shortened homomorphisms, {h1n }, we are not
concerned with these non-dominant vertex and edge groups.
Each generator sj , j = 1, . . . , r, can be written in a normal form with respect
to Λ1 . Let b1 , . . . , bt ∈ L be the collection of elements in the (fixed) normal forms
of s1 , . . . , sr , that are contained in non-QH vertex groups, or in an edge group
that is adjacent only to QH vertex groups in Λ1 . Each of the elements b1 , . . . , bt
can be represented in a normal form with respect to the abelian decomposition
Λ2 . In particular, with each of the elements b1 , . . . , bt , it is possible to associate a
(possibly empty) collection of paths in the IET components that are associated with
Q21 , . . . , Q2f , and segments in the simplicial part of Y2 (the tree that is associated
with Λ2 ), and are associated with E12 , . . . , Ev2 .
Before shortening the fixed set of generators of the dominant QH vertex groups
98
in Λ1 , and the generators, s1 , . . . , sr , we used the shortening procedure that was
applied in proving propositions 2.7 and 2.8. These propositions give us a sequence of
automorphisms of these QH vertex groups, that preserve positivity and the demonstration of the indecomposibility by a cover of finitely many (fixed) translates, that
we can now use to make the set of generators of the QH vertex groups longer, as
well the paths in the corresponding IET components in the associated limit tree
Y1 , that are associated with the set of generators, s1 , . . . , sr .
For each dominant QH vertex group Q in Λ1 we denote such an automorphism
ϕQ . For each dominant edge E in Λ1 we denote the corresponding (positive) Dehn
twist by ϕE . We set the (fixed) automorphism ψ1 ∈ M XM od(Λ1 ), that we precompose with the sequence of shortened homomorphisms {h1n }, to satisfy the following
properties:
(1) Let Q1 , . . . , Qℓ be the dominant QH vertex groups in Λ1 , and E1 , . . . , Es
be the dominant edge groups in Λ1 . Then for some positive integers α1 , αℓ
and β1 , . . . , βs :
βs
β1
αℓ
1
ψ1 = ϕα
Q1 ◦ . . . ◦ ϕQℓ ◦ ϕE1 ◦ . . . ◦ ϕEs .
(2) with each element g ∈ L it is possible to associate (possibly empty) finite collection of paths in the IET components that are associated with
(conjugates of) Q21 , . . . , Q2f (QH vertex groups in Λ2 ), and segments in the
simplicial part that are associated with E12 , . . . , Ev2 .
Let u1 , . . . , ug be the fixed set of generators of a dominant QH vertex
group Q in Λ1 . If there are some non-degenerate segments that are associated with uj in the IET component that is associated with a (conjugate
of a) QH vertex group Q2i , then the total lengths of the segments that are
α
associated with ϕQQ (uj ) in the IET components that are associated with
conjugates of Q2i , are at least twice the total length of the segments that
are associated with all the elements b1 , . . . , bt in the IET component that
are associated with conjugates of Q2i . Furthermore, this lower bound on the
ratios between the total lengths of paths can be demonstrated by finitely
many translations of subpaths in these paths. This demonstration guarantees that the lower bounds on the ratios remain valid along the entire
process.
If the path that is associated with uj in the limit tree that is associated
with Λ2 , contains a subsegment that is associated with (a conjugate of) an
edge group Ei2 , then the number of such subsegments (that are associated
α
with conjugates of Ei2 ) in the path that is associated with ϕQQ (uj ) is at least
twice the total appearances of such subsegments (associated with conjugates
of Ei2 ) in the paths that are associated with: b1 , . . . , bt .
(3) Let sj be one of the fixed set of generators s1 , . . . , sr . Suppose that there
are some non-degenerate subsegments that are contained in the path that
is associated with sj in the IET components that are associated with conjugates of a dominant QH vertex group Q in Λ1 .
Suppose that in the path that is associated a generator from the fixed
(finite) set of generators of Q in the real tree that is associated with Λ2 ,
there exist some subpath in an IET component that is associated with a
QH vertex group Q2i in Λ2 .
99
α
Then the total lengths of the subpaths that are associated with ϕQQ (sj )
in the IET components that are associated with conjugates of Q2i , are at
least twice the total length of the subpaths that are associated with all
the elements b1 , . . . , bt in the IET components that are associated with
conjugates of Q2i . Furthermore, this lower bound on the ratios between the
total lengths of paths can be demonstrated by finitely many translations
of subpaths in these subpaths. This demonstration guarantees that these
lower bounds on the ratios remain valid along the entire process.
Suppose that in the path that is associated with a generator from the
fixed (finite) set of generators of Q in the real tree that is associated with
Λ2 , there is a non-degenerate subsegment in the simplicial part of the real
tree that is associated with a conjugate of an edge group Ei2 in Λ2 .
Then the number of such subsegments (that are associated with conjuα
gates of Ei2 ) in the path that is associated with ϕQQ (sj ) is at least twice the
total appearances of such subsegments (associated with conjugates of Ei2 )
in the paths that are associated with: b1 , . . . , bt .
(4) Suppose that the path that is associated with sj in the tree that is associated with Λ1 , contains non-degenerate subsegments that are associated
with conjugates of a dominant edge group E in Λ1 . Suppose that in the
path that is associated a generator of E in the real tree that is associated
with Λ2 , there exist some subpath in an IET component that is associated
with a QH vertex group Q2i in Λ2 .
Then with the path that is associated with the element ϕβEE (sj ) in the
limit tree that is associated with Λ2 , the total lengths of the subpaths that
are in the IET components that are associated with conjugates of Q2i , are
at least twice the total length of the subpaths that are associated with
all the elements b1 , . . . , bt in the IET components that are associated with
conjugates of Q2i . Furthermore, this lower bound on the ratios between the
total lengths of paths can be demonstrated by finitely many translations of
subpaths in these subpaths.
Suppose that in the path that is associated a generator of E in the real
tree that is associated with Λ2 , there is a non-degenerate subsegment in the
simplicial part of the real tree that is associated with a conjugate of an edge
group Ei2 in Λ2 .
Then the number of such subsegments (that are associated with conjugates of Ei2 ) in the path that is associated with ϕβEE (sj ) is at least twice the
total appearances of such subsegments (associated with conjugates of Ei2 )
in the paths that are associated with: b1 , . . . , bt .
This concludes the construction of the homomorphisms that are associated with
the second level, that are set to be: {h1n ◦ ψ1 }. Note that each homomorphism h1n
can be presented as: h1n = hn ◦ νn1 , where νn1 and ψ1 are automorphisms from the
dominant modular group M XM od(Λ1 ).
We assign weight 1 with every QH vertex group and every edge group in Λ2 that
are also a QH vertex group or an edge group in Λ1 . We also assign weight 1 with
every QH vertex group and every edge group in Λ2 , for which there a generator of
a dominant QH vertex group or a dominant edge group in Λ1 , such that the path
that is associated with that generator in the tree that is associated with Λ2 , has a
subpath in an IET component that is associated with a conjugate of the QH vertex
100
group in Λ2 , or a subpath in the simplicial part of that tree that is associated with
the edge group in Λ2 . We assign weight 2 with every other QH vertex group or
edge group in Λ2 .
With the abelian decomposition Λ2 we associate its modular group, M od(Λ2 ),
and its dominant modular group, M XM od(Λ2 ). As we did in the freely indecomposable case, and in the first case, we first shorten the homomorphisms, {h1n ◦ ψ1 },
using the ambient modular group, M od(Λ2 ). We denote the obtained sequence
of homomorphisms, {h2n }. If there exists a subsequence of the homomorphisms,
{h2n }, that is either separable, or that converges into a non-faithful action of L on
some real tree, the conclusions of theorem 7.2 follow for this subsequence, and the
procedure terminates.
Hence, in the sequel we assume that {h2n } has no separable subsequence or a
subsequence that converges into a non-faithful action of L on a real tree. In this
case we use only the dominant modular group, M XM od(Λ2 ), and modify what we
did in the first step.
First, the shortening of the homomorphisms, {h1n ◦ ψ1 }, using the dominant
modular group, M XM od(Λ2 ), requires to preserve the positivity of the fixed set
of generators of the semigroup S, s1 , . . . , sr , and the fixed sets of generators of the
QH and edge groups in the abelian decompositions, Λ1 and Λ2 . In particular, it
is required to preserve the positivity of the paths that are associated with these
elements and are contained in IET components that are associated with conjugates
of dominant QH vertex groups in Λ2 . In addition the shortening is required to
preserve the finitely many equivariance restrictions that we imposed on these paths,
i.e., the overlaps of these paths with translates of themselves, that sample the
indecomposibility of the IET components, and guarantee that certain inequalities
between lengths of these paths will hold in the sequel, and in particular after the
shortenings.
As we did in the freely indecomposable case, in addition to these elements, in
shortening using M XM od(Λ2 ), we also consider the elements in the ball of radius 2
in the Cayley graph of L w.r.t. the generating set s1 , . . . , sr . Note that elements of
length 2 need not be positive nor negative, but they may rather contain a positive
and a negative subsegments. Still, the shortening procedure that is presented in
section 2 for abelian, QH, and Levitt components, works precisely in the same way,
when the path that is associated with an element contains finitely many oriented
(positively or negatively) subsegments, and not only in case the entire path is
oriented.
As in the freely indecomposable and in the first step, we first shorten the action
of each of the dominant QH vertex groups in Λ2 using the procedure that is used
in the proof of propositions 2.7 and 2.8. This, in particular, associates a fixed set
of generators with each dominant QH vertex group, and a further finite collection
of words, so that their positivity implies the positivity of a finite collection of given
paths under an infinite sequence of modular automorphisms (see the description in
the freely indecomposable case in the proof of proposition 6.4).
As in the first step, using the indecomposibility of an IET components, finitely
many (fixed) translates of each of the positive paths that are associated with each of
the fixed generators of a QH vertex group cover the paths that are associated with
the other generators. This coverings guarantee that the ratios between their lengths
along the iterative (shortening) procedure that we use remain globally bounded. We
101
do the same to paths that are associated with elements in the ball of radius 2 of
L and pass through an IET component, and to paths that are associated with
generators of QH vertex groups in Λ1 and pass through IET components.
At this point we shorten the homomorphisms {h1n ◦ ψ1 } using the dominant
modular group M XM od(Λ2 ), precisely as we did in the first step, but keeping the
positivity and the equivariance of a larger set of elements that is specified in the
beginning of this step. We (still) denote the obtained (shortened) homomorphisms
{h2n }. We pass to a subsequence of the homomorphisms {h2n } that converges into an
action of L on some real tree with an associated abelian decomposition ∆3 . If the
action of L is not faithful, or if the sequence {h2n } contains a separable subsequence,
the conclusions of theorem 7.2 follow and the procedure terminates. Hence, we may
assume that the action of L is faithful, that {h2n } contains no separable subsequence,
and that the action of L on the limit tree is geometric, and the tree contains no
segment in its simplicial part that can not be divided into finitely many subsegments
ˆ 3 and finally to an abelian
with non-trivial stabilizers. We further refine ∆3 to ∆
decomposition Λ3 , precisely as we did in the first step.
As in the first step, we modify the shortened homomorphisms, {h2n }, that converge into a faithful limit action of L on a limit tree from which the abelian decomposition Λ3 is obtained, by precomposing them with a fixed automorphism, ψ2 ,
from the dominant modular group M XM od(Λ1 ). This precomposition is needed
to guarantee the validity of certain inequalities between the lengths, or the ratios
of lengths, of a finite set of elements. This finite set include:
(1) the fixed set of generators of dominant QH vertex groups and generators of
dominant edge groups in Λ1 and Λ2 .
(2) the elements in a ball of radius 2 in the Cayley graph of L w.r.t. the generators: s1 , . . . , sr .
(3) the fixed set of generators of some of the QH vertex groups and some of the
edge groups in Λ3 .
Let Q31 , . . . , Q3f2 , and E13 , . . . , Ev32 , be those QH vertex groups and edge groups
in Λ3 , for which at least one of the fixed set of generators of the dominant QH vertex
groups of weight 2, and dominant edge groups of weight 2 in Λ2 , is not elliptic with
respect to them.
Each element in the ball of radius 2 in the Cayley graph of L (w.r.t. s1 , . . . , sr ),
and each of the fixed generators of a QH vertex group or an edge group in Λ1 , can be
written in a normal form with respect to Λ2 . Let b21 , . . . , b2t2 ∈ L be the collection
of elements in the (fixed) normal forms of these elements, that are contained in
non-QH vertex groups, or in an edge group that is adjacent only to QH vertex
groups in Λ2 . Each of the elements b21 , . . . , b2t2 can be represented in a normal
form with respect to the abelian decomposition Λ3 . In particular, with each of the
elements b21 , . . . , b2t2 , it is possible to associate a (possibly empty) collection of paths
in the IET components that are associated with Q31 , . . . , Q3f2 , and segments in the
simplicial part of Y3 (the tree that is associated with Λ3 ), and are associated with
E13 , . . . , Ev32 .
As in the first step, before shortening the fixed set of generators of the dominant
QH vertex groups in Λ2 , we used the shortening procedure that was applied in
proving propositions 2.7 and 2.8. These propositions give us a sequence of automorphisms of these QH vertex groups, that preserve positivity of positive elements
and the orientation (positive or negative) of subpaths of paths that are associated
102
with elements from the finite set that include elements from the ball of radius 2 in
the Cayley graph of L, and (fixed) generators of QH and abelian vertex groups in
Λ1 and Λ2 .
For each dominant QH vertex group Q2 in Λ2 we denote such an automorphism
ϕQ2 . For each dominant edge group E 2 in Λ2 we denote the corresponding (positive)
Dehn twist by ϕE 2 . Before constructing a fixed automorphism from the dominant
modular group M XM od(Λ2 ) that is going to be used to twist (precompose) the
sequence of homomorphisms, {h2n }, We construct a fixed automorphism that is
associated only with the dominant QH vertex groups and dominant edge groups in
Λ2 that are of weight 2. We denote this automorphism ψ22 , and in a similar way to
what we did in the first step we require it to satisfy the following properties:
(1) Let Q21 , . . . , Q2ℓ2 be the dominant QH vertex groups of weight 2 in Λ2 , and
E12 , . . . , Es22 be the dominant edge groups of weight 2 in Λ2 . Then for some
positive integers α21 , α2ℓ2 and β12 , . . . , βs22 :
α2
α2
β2
β2
ψ22 = ϕQ12 ◦ . . . ◦ ϕQℓ22 ◦ ϕE12 ◦ . . . ◦ ϕEs22 .
1
ℓ2
1
s2
(2) with each element in L it is possible to associate (possibly empty) finite collection of paths in the IET components that are associated with (conjugates
of) Q31 , . . . , Q3f2 (QH vertex groups in Λ3 ), and segments in the simplicial
part that are associated with E13 , . . . , Ev32 .
As in (part (2) in) the first step, if there are some non-degenerate segments, that are associated with a fixed generator, uj , of a dominant QH
vertex group, Q2 , of weight 2 in Λ2 , in the IET component that is associated with a (conjugate of a) QH vertex group Q3i , then the total lengths of
α2 2
the segments that are associated with ϕQQ2 (uj ) in the IET components that
are associated with conjugates of Q3i , are at least four times the total length
of the segments that are associated with all the elements b21 , . . . , b2t2 in the
IET components that are associated with conjugates of Q3i . Furthermore,
as in the first step, this lower bound on the ratios between the total lengths
of paths can be demonstrated by finitely many translations of subpaths in
these paths. This demonstration guarantees that these lower bounds on the
ratios remain valid along the entire process.
If the path that is associated with uj in the limit tree that is associated
with Λ3 , contains a subsegment that is associated with (a conjugate of) an
edge group Ei3 , then the number of such subsegments (that are associated
α2 2
with conjugates of Ei3 ) in the path that is associated with ϕQQ2 (uj ) is at
least four times the total appearances of such subsegments (associated with
conjugates of Ei3 ) in the paths that are associated with: b21 , . . . , b2t2 .
(3) Let g be one of the elements in the ball of radius 2 in the Cayley graph
of L (w.r.t. s1 , . . . , sr ), or one of the fixed set of generators of QH vertex
groups in Λ1 or a generator of an edge group in Λ1 . Note that g need not
be an oriented element, but the path that is associated with g may contain
a positive and a negative subpaths.
Suppose that there are some non-degenerate subsegments that are contained in the path that is associated with g in the IET components that are
103
associated with conjugates of a dominant QH vertex group Q2 of weight 2
in Λ2 .
Suppose that in the path that is associated with a generator from the
fixed (finite) set of generators of Q2 in the real tree that is associated with
Λ3 , there exist some subpath in an IET component that is associated with
a QH vertex group Q3i in Λ3 .
α2 2
With the path that is associated with the element ϕQQ2 (g) in the limit
tree that is associated with Λ2 , one can associate finitely many subpaths
that are contained in (finitely many) IET components that are associated
with conjugates of Q2 . Note that the number of such subpaths is the same
as the number of such subpaths in the path that is associated with g.
Then the total lengths of the subpaths that are associated with the image
of each of these subpaths in the IET components that are associated with
conjugates of Q3i , are at least four times the total length of the subpaths
that are associated with all the elements b21 , . . . , b2t2 in the IET components
that are associated with conjugates of Q3i . Furthermore, this lower bound
on the ratios between the total lengths of paths can be demonstrated by
finitely many translations of subpaths in these subpaths. This demonstration guarantees that these lower bounds on the ratios remain valid along
the entire process.
Suppose that in the path that is associated with a generator from the
fixed (finite) set of generators of Q2 in the real tree that is associated with
Λ3 , there is a non-degenerate subsegment in the simplicial part of the real
tree that is associated with a conjugate of an edge group Ei3 in Λ3 .
Then the number of such subsegments (that are associated with conjuα2 2
gates of Ei3 ) in the path that is associated with ϕQQ2 (g) is at least four times
the total appearances of such subsegments (associated with conjugates of
Ei3 ) in the paths that are associated with: b21 , . . . , b2t2 .
(4) Suppose that the path that is associated with g in the tree that is associated with Λ2 , contains non-degenerate subsegments that are associated
with conjugates of a dominant edge group E 2 of weight 2 in Λ2 . Suppose
that in the path that is associated with a generator of E 2 in the real tree
that is associated with Λ3 , there exist some subpath in an IET component
that is associated with a QH vertex group Q3i in Λ3 .
β2
Then with the path that is associated with the element ϕEE22 (g) in the
limit tree that is associated with Λ2 , the image of the subpaths that are
stabilized by conjugates of the dominant edge group E 2 , contain subpaths
that are in the IET components that are associated with conjugates of Q3i ,
and the total lengths of these subpaths are at four times the total length
of the subpaths that are associated with all the elements b21 , . . . , b2t2 in the
IET components that are associated with conjugates of Q3i . Furthermore,
this lower bound on the ratios between the total lengths of paths can be
demonstrated by finitely many translations of subpaths in these subpaths.
Suppose that in the path that is associated a generator of E 2 in the real
tree that is associated with Λ3 , there is a non-degenerate subsegment in the
simplicial part of the real tree that is associated with a conjugate of an edge
group Ei3 in Λ3 .
104
Then the number of such subsegments (that are associated with conjuβ2
gates of Ei3 ) in the path that is associated with ϕEE22 (g) is at least four times
the total appearances of such subsegments (associated with conjugates of
Ei3 ) in the paths that are associated with: b21 , . . . , b2t2 .
This concludes the treatment of dominant QH vertex groups and dominant edge
groups of weight 2 in Λ2 , and the construction of a fixed automorphism, ψ22 ∈
M XM od(λ2 ), that is associated with them. After ψ22 is constructed we can treat
in a similar way the dominant QH vertex groups and dominant edge groups of
weight 1 in Λ2 , and finally construct the fixed automorphism ψ2 , that is used in
precomposing the sequence of shortened homomorphisms, {h2n }.
Let Q31 , . . . , Q3f1 , and E13 , . . . , Ev31 , be those QH vertex groups and edge groups
in Λ3 , for which at least one of the fixed set of generators of the dominant QH vertex
groups of weight 1, and dominant edge groups of weight 1 in Λ2 , is not elliptic with
respect to them.
Let ColΛ2 be the abelian decomposition that is obtained by collapsing all the
QH vertex groups of weight 2 in Λ2 and the edges that are connected to them, and
all the edge groups of weight 2 in Λ2 . ColΛ2 contains only QH vertex groups and
edge groups of weight 1.
Each element in the ball of radius 2 in the Cayley graph of L, and each of the
fixed generators of a QH vertex group or an edge group in Λ1 , can be written
in a normal form with respect to ColΛ2 . Let b11 , . . . , b1t1 ∈ L be the collection
of elements in the (fixed) normal forms of these elements, that are contained in
non-QH vertex groups, or in an edge group that is adjacent only to QH vertex
groups in ColΛ2 . Each of the elements b11 , . . . , b1t1 can be represented in a normal
form with respect to the abelian decomposition Λ3 . In particular, with each of the
elements b11 , . . . , b1t1 , it is possible to associate a (possibly empty) collection of paths
in the IET components that are associated with Q31 , . . . , Q3f1 , and segments in the
simplicial part of Y3 (the tree that is associated with Λ3 ), and are associated with
E13 , . . . , Ev31 .
At this point we construct the automorphism ψ2 in a similar way to the construction of the automorphism ψ22 , and the automorphism ψ1 in the first step. Recall
that given a dominant QH vertex group, Q, in Λ2 , we used the procedure that was
applied in proving propositions 2.7 and 2.8 to associate with Q an automorphism,
ϕQ , that preserve the orientation of the oriented subpaths in the paths that are
associated with the elements from the finite set that include elements from the ball
of radius 2 in the Cayley graph of L, and (fixed) generators of QH and abelian
vertex groups in Λ1 and Λ2 .
The automorphism ψ2 is required to satisfy similar properties as ψ22 and ψ1 :
(1) Let Q11 , . . . , Q1ℓ1 be the dominant QH vertex groups of weight 1 in Λ2 , and
E11 , . . . , Es11 be the dominant edge groups of weight 1 in Λ2 . Then for some
positive integers α11 , α1ℓ1 and β11 , . . . , βs11 :
α1
α1
β1
β1
ψ21 = ϕQ11 ◦ . . . ◦ ϕQℓ11 ◦ ϕE11 ◦ . . . ◦ ϕEs11
1
ℓ1
1
s1
and ψ2 = ψ21 ◦ ψ22 .
(2) with each element in L it is possible to associate (possibly empty) finite collection of paths in the IET components that are associated with (conjugates
105
of) Q31 , . . . , Q3f1 (QH vertex groups in Λ3 ), and segments in the simplicial
part that are associated with E13 , . . . , Ev31 .
As in (part (2) in) the first step, if there are some non-degenerate segments, that are associated with a fixed generator, uj , of a dominant QH
vertex group, Q1 , of weight 1 in Λ2 , in the IET component that is associated
with a (conjugate of a) QH vertex group Q3i , then the total lengths of the
α1 1
segments that are associated with ϕQQ1 (uj ) in the IET components that are
associated with conjugates of Q3i , are at least four times the total length of
the segments that are associated with all the elements ψ22 (b11 ), . . . , ψ22 (b1t1 )
in the IET components that are associated with conjugates of Q3i . Furthermore, as in the first step, this lower bound on the ratios between the
total lengths of paths can be demonstrated by finitely many translations of
subpaths in these paths. This demonstration guarantees that these lower
bounds on the ratios remain valid along the entire process.
If the path that is associated with uj in the limit tree that is associated
with Λ3 , contains a subsegment that is associated with (a conjugate of) an
edge group Ei3 , then the number of such subsegments (that are associated
α1 1
with conjugates of Ei3 ) in the path that is associated with ϕQQ1 (uj ) is at
least four times the total appearances of such subsegments (associated with
conjugates of Ei3 ) in the paths that are associated with: ψ22 (b11 ), . . . , ψ22 (b1t1 ).
(3) Let g be one of the elements in the ball of radius 2 in the Cayley graph of L
(w.r.t. s1 , . . . , sr ), or one of the fixed set of generators of QH vertex groups
in Λ1 or a generator of an edge group in Λ1 .
Suppose that there are some non-degenerate subsegments that are contained in the path that is associated with g in the IET components that are
associated with conjugates of a dominant QH vertex group Q1 of weight 1
in Λ2 .
Suppose that in the path that is associated with a generator from the
fixed (finite) set of generators of Q1 in the real tree that is associated with
Λ3 , there exist some subpath in an IET component that is associated with
a QH vertex group Q3i in Λ3 .
α1Q1
With the path that is associated with the element ϕQ1 (g) in the limit
tree that is associated with Λ2 , one can associate finitely many subpaths
that are contained in (finitely many) IET components that are associated
with conjugates of Q1 .
Then the total lengths of the subpaths that are associated with the image of each of these subpaths in the IET components that are associated
with conjugates of Q3i , are at least four times the total length of the subpaths that are associated with all the elements ψ22 9b11 ), . . . , ψ22 (b1t1 ) in the
IET components that are associated with conjugates of Q3i . Furthermore,
this lower bound on the ratios between the total lengths of paths can be
demonstrated by finitely many translations of subpaths in these subpaths.
This demonstration guarantees that these lower bounds on the ratios remain
valid along the entire process.
Suppose that in the path that is associated with a generator from the
fixed (finite) set of generators of Q1 in the real tree that is associated with
Λ3 , there is a non-degenerate subsegment in the simplicial part of the real
106
tree that is associated with a conjugate of an edge group Ei3 in Λ3 .
Then the number of such subsegments (that are associated with conjuEi3 )
α1Q1
gates of
in the path that is associated with ϕQ1 (g) is at least four times
the total appearances of such subsegments (associated with conjugates of
Ei3 ) in the paths that are associated with: ψ22 (b21 ), . . . , ψ22 (b2t2 ).
(4) Suppose that the path that is associated with g in the tree that is associated with Λ2 , contains non-degenerate subsegments that are associated
with conjugates of a dominant edge group E 1 of weight 1 in Λ2 . Suppose
that in the path that is associated with a generator of E 1 in the real tree
that is associated with Λ3 , there exist some subpath in an IET component
that is associated with a QH vertex group Q3i in Λ3 .
β1
Then with the path that is associated with the element ϕEE11 (g) in the
limit tree that is associated with Λ2 , the image of the subpaths that are stabilized by conjugates of the dominant edge group E 1 , contain subpaths that
are in the IET components that are associated with conjugates of Q3i , and
the total lengths of these subpaths are at four times the total length of the
subpaths that are associated with all the elements ψ22 (b11 ), . . . , ψ22 (b1t1 ) in the
IET components that are associated with conjugates of Q3i . Furthermore,
this lower bound on the ratios between the total lengths of paths can be
demonstrated by finitely many translations of subpaths in these subpaths.
Suppose that in the path that is associated a generator of E 1 in the real
tree that is associated with Λ3 , there is a non-degenerate subsegment in the
simplicial part of the real tree that is associated with a conjugate of an edge
group Ei3 in Λ3 .
Then the number of such subsegments (that are associated with conjuβ1
gates of Ei3 ) in the path that is associated with ϕEE11 (g) is at least four times
the total appearances of such subsegments (associated with conjugates of
Ei3 ) in the paths that are associated with: ψ22 (b11 ), . . . , ψ22 (b1t1 ).
This concludes the construction of the homomorphisms that are associated with
the third level, that are set to be: {h2n ◦ ψ2 }. Note that each homomorphism h2n
can be presented as: h2n = h1n ◦ νn2 , where νn2 and ψ2 are automorphisms from the
dominant modular group M XM od(Λ2 ).
We assign weight 1 or 2 with every QH vertex group and every edge group in
Λ3 that are also a QH vertex group or an edge group of a similar weight in Λ2 . We
assign weight 1 with every QH vertex group and every edge group in Λ3 , for which
there is a generator of a dominant QH vertex group or a dominant edge group of
weight 1 in Λ2 , such that the path that is associated with that generator in the tree
that is associated with Λ3 , has a subpath in an IET component that is associated
with a conjugate of the QH vertex group in Λ3 , or a subpath in the simplicial part
of that tree that is associated with the edge group in Λ3 .
We assign weight 2 with every QH vertex group and every edge group in Λ3 ,
which is not already of weight 1, and for which there is a generator of a dominant
QH vertex group or a dominant edge group of weight 2 in Λ2 , such that the path
that is associated with that generator in the tree that is associated with Λ3 , has a
subpath in an IET component that is associated with a conjugate of the QH vertex
group in Λ3 , or a subpath in the simplicial part of that tree that is associated with
the edge group in Λ3 . We assign weight 3 with every other QH vertex group or
107
edge group in Λ3 , i.e., with the other QH vertex groups or edge groups that were
not assigned weight 1 or 2.
We continue iteratively. The abelian decompositions at step i, is constructed
from shortened homomorphisms {hi−1
n }, that are further twisted (or precomposed)
with a fixed automorphism ψi−1 that is contained in the dominant modular group
M XM od(Λi−1 ). The QH vertex groups and edge groups in Λi have weights in the
set 1, . . . , i.
With the abelian decomposition Λi we associate its modular group, M od(Λ2 ),
and its dominant modular group, M XM od(Λ2 ). As we did in first steps, we first
shorten the homomorphisms, {hi−1
◦ ψi−1 }, using the ambient modular group,
n
M od(Λi ). We denote the obtained sequence of homomorphisms, {hin }. If there
exists a subsequence of the homomorphisms, {hin }, that is either separable, or
that converges into a non-faithful action of L on some real tree, the conclusions of
theorem 7.2 follow for this subsequence, and the procedure terminates.
If {hin } has no separable subsequence or a subsequence that converges into a
non-faithful action of L on a real tree, we use only the dominant modular group,
M XM od(Λi ). The shortening of the sequence of homomorphisms, {hi−1
n ◦ ψi−1 }, is
required to preserve the positivity of the fixed set of generators, s1 , . . . , sr , and all
the fixed sets of generators of QH vertex groups and abelian edge groups in all the
abelian decompositions, Λ1 , . . . , Λi . It further requires to preserve the orientation
of the oriented (positive or negative) subpaths in the paths that are associated with
all the elements in a ball of radius i in the Cayley graph of L w.r.t. s1 , . . . , sr . The
shortening procedure is also suppose to preserve the finite equivariance that was
used to (finitely) demonstrate the indecomposability of paths that are associated
with the fixed generators of QH vertex groups in Λ1 , . . . , Λi , and the finite equivariance that was used to demonstrate certain inequalities between subpaths in the
trees that are associated with Λ1 , . . . , Λi .
Before shortening the (fixed) set of generators of each of the dominant QH vertex
groups in Λi , we use the procedure that was applied in the proofs of propositions
2.7 and 2.8, to associate an automorphism ϕQ with each such dominant QH vertex group in Λi . We further use the indecomposability of IET components, and
associate finitely many translates of the fixed generators of each of the dominant
QH vertex groups, that demonstrate the indecomposability, and provide lower and
upper bounds on the ratios between lengths of the fixed set of generators that will
be kept along the next steps of the procedure.
After shortening the homomorphisms, {hi−1
n ◦ψi−1 }, using the dominant modular
group M XM od(Λi), we pass to a convergent subsequence that we denote {hin } (note
that at each step the shortening is required to preserve the orientation of a larger
number of subpaths that are associated a larger number of elements). We pass to a
subsequence that we still denote, {hin } that converges into an action of L on some
real tree with an associated abelian decomposition ∆i+1 . If the action of L is not
faithful, or if the sequence {hin } contains a separable subsequence, the conclusions
of theorem 7.2 follow and the procedure terminates. Hence, we may assume that
the action of L is faithful, and we further refine ∆i+1 , precisely as we did in the
first two steps, and obtain the abelian decomposition Λi+1 .
As in the first two steps, we further construct a fixed automorphism, ψi , that
is used in modifying (by precomposition) the homomorphisms, {hin }. This precomposition is needed to guarantee the validity of certain inequalities between the
108
lengths, or the ratios of lengths, of a finite set of elements. This finite set include:
(1) the fixed set of generators of dominant QH vertex groups and generators of
dominant edge groups in Λ1 , . . . , Λi .
(2) the elements in a ball of radius i in the Cayley graph of L w.r.t. the generators: s1 , . . . , sr .
(3) the fixed set of generators of some of the QH vertex groups and some of the
edge groups in Λi+1 .
We start with dominant QH vertex groups and dominant edge groups in Λi that
are of highest weight (in Λi ). We look at all the QH vertex groups and all the edge
groups in Λi+1 for which at least one of the fixed set of generators of the dominant
QH vertex groups of highest weight in Λi , or of dominant edge groups of highest
weight in Λi , is not elliptic with respect to them.
Each element in the ball of radius i in the Cayley graph of L (w.r.t. s1 , . . . , sr ),
and each of the fixed generators of a QH vertex group or an edge group in Λ1 , . . . , Λi ,
can be written in a normal form with respect to Λi . Let w1 , 1 ≤ w1 ≤ i, be the
w1
1
highest weight in Λi , and let bw
1 , . . . , btw1 ∈ L be the collection of elements in
the (fixed) normal forms of these elements, that are contained in non-QH vertex
groups, or in an edge group that is adjacent only to QH vertex groups in Λi . We
w1
∈ M XM od(Λi ) will satisfy properties (1)require that a fixed automorphism, ψm
(4) that are listed in the first and second steps, w.r.t. the QH vertex groups and
the edge groups in Λi+1 for which at least one of the fixed set of generators of the
dominant QH vertex groups of highest weight in Λi , or of dominant edge groups of
highest weight in Λi , is not elliptic with respect to them, and w.r.t. the elements,
w1
1
bw
1 , . . . , btw ∈ L.
1
Once the automorphism ψiw1 is fixed, we treat dominant QH vertex groups and
dominant edge groups in Λi that are of the next highest weight, w2 , 1 ≤ w2 < w1 .
Let Colw1 Λi be the abelian decomposition that is obtained by collapsing all the
QH vertex groups of weight w1 in Λi and the edges that are connected to them, and
all the edge groups of weight w1 in Λi . Colw1 Λi contains only QH vertex groups
and edge groups of lower weight.
Each element in the ball of radius i in the Cayley graph of L, and each of the fixed
generators of a QH vertex group or an edge group in Λ1 , . . . , Λi , can be written
w2
2
in a normal form with respect to Colw1 Λi . Let bw
1 , . . . , btw2 ∈ L be the collection
of elements in the (fixed) normal forms of these elements, that are contained in
non-QH vertex groups, or in an edge group that is adjacent only to QH vertex
w2
groups in Colw1 Λi . We require that a fixed automorphism, ψm
∈ M XM od(Λi )
will satisfy properties (1)-(4) that are listed in the first and second steps, w.r.t.
the QH vertex groups and the edge groups in Λi+1 for which at least one of the
fixed set of generators of the dominant QH vertex groups of weight w2 in Λi , or of
dominant edge groups of weight w2 in Λi , is not elliptic with respect to them, and
w1 w1
1
w.r.t. the elements, ψiw1 (bw
1 ), . . . , ψi (btw1 ) ∈ L.
We continue iteratively to lower weight dominant QH vertex groups and dominant edge groups in Λi , until a a fixed automorphism ψi ∈ M XM od(Λi) is
constructed. This automorphism enable us to replace the sequence of homomorphisms, {hin } by precomposing them with ψi : {hin ◦ ψni }. Hence, the new sequence of homomorphisms is obtained from the sequence in the previous level as
{hi−1
◦ νni ◦ ψi , where νni and ψi are automorphisms from the dominant modular
n
group M XM od(Λi).
109
We assign the weight in Λi to every QH vertex group and every edge group in
Λi+1 that are also a QH vertex group or an edge group in Λi . We assign weight
w with every QH vertex group and every edge group in Λi+1 , for which there is a
generator of a dominant QH vertex group or a dominant edge group of weight w
in Λi , such that the path that is associated with that generator in the tree that is
associated with Λi+1 , has a subpath in an IET component that is associated with
a conjugate of the QH vertex group in Λi+1 , or a subpath in the simplicial part
of that tree that is associated with the edge group in Λi+1 , and there is no such
generator of a dominant QH vertex group or a dominant edge group in Λi that has
weight smaller than w. We assign weight i + 1 with every other QH vertex group
or edge group in Λi+1 , i.e., with those QH vertex groups and edge groups for which
no weight (bounded by i) was assigned to them.
We continue iteratively. If in all steps the obtained actions are faithful and
geometric, and the sequences of homomorphisms contain no separable subsequence,
we get an infinite sequence of abelian decompositions, Λ1 , Λ2 , . . . . Given the infinite
sequence of abelian decompositions, we associate with it its stable dominant abelian
decomposition (see definition 7.5), Θi0 .
Given the sequence of abelian decompositions, Λ1 , . . . , and its stable abelian
decomposition, Θi0 , our goal is to show that there exists an index i1 ≥ i0 , such
that all the abelian decompositions, Λi1 , . . . , are dominated by Θi0 . i.e., all the
modular groups, M od(Λi ), i ≥ i1 , are contained in the modular group M od(Θi0 ),
where M od(Θi0 is generated by the modular groups of the QH vertex groups in
Θi0 , and the Dehn twists along edges (with non-trivial stabilizers) in Θi0 . This will
enable us to replace the infinite sequence of abelian decompositions, Λ1 , . . . , with
the finite sequence, Λ1 , . . . , Λi1 −1 , Θi0 .
To show the existence of such an index i1 , we first need to modify the sequences
of homomorphisms, {hin }, that will enable us to appropriately collapse some of
the QH vertex groups and some of the edge groups in the abelian decompositions,
Λi0 , . . . .
By the accessibility of f.p. groups [Be-Fe], or by acylindrical accessibility ([Se],[De],[We]),
the number of edge (and vertex groups) in each of the abelian decompositions Λi
is universally bounded. With each QH vertex group and each edge group in the
abelian decompositions, Λi , we have associated a weight. A QH vertex group or
an edge group in Λi that are not dominant pass to the next level, as the same
QH vertex group or an edge group with the same weight. Hence, all the abelian
decompositions, Λi , must have either a QH vertex group or an edge group with
weight 1.
Definition 7.6. Let w be a positive integer. We say that w is a stable weight of the
sequence of abelian decompositions: Λ1 , . . . , if the sequence contains a subsequence
that contains either a QH vertex group or an edge group of weight w. This is
equivalent to the existence of a QH vertex group or an edge group of weight w in
each of the abelian decompositions: Λw , . . . . 1 is always a stable weight, and the
accessibility of f.p. groups implies that there are only finitely many stable weights.
Given Λi , we define the unstable modular group, U SM Od(Λi ), to be the modular group that is generated by modular groups of QH vertex groups of unstable
weight in Λi , and Dehn twists along edge groups with unstable weight in Λi .
Given the stable weights we modify the sequences of homomorphisms, {hin }.
110
For each index i, we precompose the homomorphisms {hin } with automorphisms
τni ∈ U SM od(Λi ), such that hin ◦ τni preserve all the positivity and the (finite)
equivariance properties that hin was supposed to preserve, and is the shortest among
all the homomorphisms of the form: hin ◦ τ that preserve the positivity and the
finite equivariance requirements, and in which: τ ∈ U SM od(Λi ). We denote the
homomorphisms hin ◦ τn , shin .
If for some index i, the sequence {shin } contains a separable subsequence, or if
it contains a subsequence that converges into a proper quotient of the limit group
L, the conclusions of theorem 7.2 follow. Hence, in the sequel we may assume that
{shin } contain no such subsequences.
Lemma 7.7. The stable dominant abelian decomposition Θi0 (possibly) contains
several QH vertex groups and (possibly) several edge groups with non-trivial stabilizers.
(i) for every QH vertex group Q in Θi0 there exists at least one stable weight
w, so that for all i ≥ max(w, i0 ), Λi contains a QH vertex group or an edge
group with weight w that are contained in Q.
(ii) for any edge E in Θi0 that has a non-trivial stabilizer, there exists a stable
weight w, so that (a conjugate of ) E appears as an edge group in Λi for all
i ≥ max(w, i0 ).
Proof: Let Q be a QH vertex group in Θi0 . Q inherits an abelian decomposition
from each of the abelian decompositions, Λi , i ≥ i0 , an abelian decomposition
in which the boundary elements in Q are elliptic (elliptic elements in the stable
dominant abelian decomposition Θi0 are elliptic in all the abelian decompositions,
Λi , i ≥ i0 ).
Non-peripheral elements in Q that do not have roots in Q do not have roots in
the ambient limit group L. Hence, every QH vertex group that appears in one of
the abelian decompositions, Λi , can either be conjugated into a subsurface of Q, or
every conjugate of that QH vertex group intersects Q trivially or only in conjugates
of peripheral elements of Q. Similarly every edge group in Λi , i ≥ i0 , is either
conjugate into a non-peripheral s.c.c. in Q, or it can not be conjugated into Q, or
it can be conjugated only to peripheral subgroups in Q.
Since all the non-peripheral elements in Q are not elliptic in the stable dominant
abelian decomposition, Θi0 , there must exists some index j ≥ i0 , so that Λj contains
a QH vertex group that can be conjugated into a subsurface of Q, or Λj contains
an edge group that can be conjugated into a non-peripheral s.c.c. in Q. Such a QH
vertex group or an edge group, has weight w that is bounded by j. Furthermore,
such a QH vertex group or an edge group are not elliptic in an abelian decomposition
Λi , i > j, only with respect to QH vertex groups or edge groups that can be
conjugated into subsurfaces or non-peripheral s.c.c. in Q.
Therefore, by the structure of the procedure that constructs the abelian decompositions, {Λi }, if such a QH vertex group or an edge group of weight w is not
elliptic in Λi , i > j, the next abelian decomposition, Λi+1 , must contain QH vertex
groups QH vertex groups or edge groups that can be conjugated into subsurfaces
or non-peripheral s.c.c. in Q, and have weight that is bounded above by w. This
implies part (i).
To prove (ii), note that an edge group E in Θi0 must be elliptic in all the abelian
decompositions Λi , i ≥ i0 . Since there is an edge e with stabilizer E in Θi0 , and E is
111
elliptic in Λi for all i ≥ i0 , there must exist an index i1 ≥ i0 , for which Λi1 contains
the (splitting that corresponds to the) edge e. The edge group E is elliptic in all
the abelian decompositions Λi , i ≥ i1 . Hence, it is not a dominant edge group in
any of these splittings, so the edge e is inherited by all the abelian decompositions,
λi , for i ≥ i1 .
Lemma 7.8. Let w be a stable weight. Suppose that for every large enough i, Λi
contains a QH vertex group or an edge group of weight w that are not conjugate to a
QH vertex group or an edge group in Θi0 , and w is the minimal such stable weight.
Then for every large i, there must exist a dominant QH vertex group of weight w
in Λi , or an edge group of weight w in Λi , such that the length of a generator of
a dominant edge group or the length of a generator of a non-QH dominant vertex
group in Λi is infinitesimal in comparison to the the length of the edge of weight w
in the real tree from which Λi was obtained.
Furthermore, if for every large i there exist a QH vertex group or an edge group
of weight w in Λi that are properly contained in a QH vertex group Q in Θi0 , and
w is the minimal such stable weight, then for every large i, there exists a dominant
QH vertex group of weight w in λi , or an edge of weight w in Λi , for which the
length of a generator of a dominant edge group or a dominant non-QH vertex group
is infinitesimal with respect to its length.
Proof: If Λi , i ≥ w, contains a QH vertex group or and edge group of stable weight
w or an edge group of stable weight w, and these are not conjugate to an edge group
or a QH vertex group in Θi0 , then for some i1 ≥ w, Λi1 contains s dominant such
edge group or QH vertex group.
Since w is assumed to be the minimal stable weight with these properties, there
exists some index i2 ≥ i1 , such that for every i ≥ i2 , if Λi contains a dominant QH
vertex group or a dominant edge group of weight w, that are not conjugate to a QH
vertex group or an edge group in Θi0 , then at least one of the QH vertex groups
in Λi+1 is dominant, or the length of all the fixed generators of the dominant edge
groups and the non-QH vertex groups in Λi+1 is infinitesimal in comparison with
the length of an edge with weight w that is not conjugate to an edge in Θi0 .
Given a QH vertex group in Θi0 , we apply the same argument to the QH vertex
groups and edge groups that can be conjugated into that QH vertex group in Θi0 ,
and same consequence holds for the minimal weight stable weight w w.r.t. a fixed
QH vertex group in Θi0 .
i
Lemmas 7.7 and 7.8 imply that the sequence of homomorphisms, {shn }, subconverges into an action on a real tree from which an abelian decomposition, sΛi ,
can be obtained (using the refinement procedure that was used in construction Λi ),
where sΛi is obtained from Λi by collapsing the following:
(1) QH vertex groups and edge groups with unstable weight.
(2) non-dominant QH vertex groups of stable weight that are not conjugate to
QH vertex groups in Θi0 .
(3) edge groups of stable weight that are not conjugate to edges in Θi0 , and
for which the length of the corresponding edge in the real tree into which
{hin } converges, is bounded by a constant times the length of a generator
of a dominant edge group or a dominant non-QH vertex group.
112
For each index i, it is possible to choose a homomorphism, fi , from the sequnce
{hin }, with the following properties:
(1) the set of abelian decompositions, ∆, with fundamental group L, and in
which all the edges have trivial stabilizers, is clearly countable. Hence, we
can define an (arbitrary but fixed) order on this set, and we denote the
corresponding sequence of abelian decompositions {∆m }. Note that every
abelian decomposition with fundamental group L and trivial edge groups
appears in this sequence.
For each index i ≥ i0 , we choose the homomorphism fi , to be a homomorphism from the sequence {shin } that is not separable with respect to
the abelian decompositions: ∆1 , . . . , ∆i (see definition 7.3).
(2) The sequence of homomorphisms {shin } subconverges into a faithful action
of L on the limit tree sYi+1 . We require fi to approximate the action on
the limit tree sYi+1 , of all the elements in a ball of radius i in the Cayley
graph of L (w.r.t. the given generating set s1 , . . . , sr ), of all the fixed sets
of generators of the QH vertex groups and edge groups in Λ1 , . . . , Λi , and of
all the (finitely many) elements that were chosen to demonstrate the mixing
property of the dominant QH vertex groups and dominant edge groups in
Λ1 , . . . , Λi .
By construction, the sequence of homomorphisms, {fi }, subconverges into a
faithful action of the limit group L on a real tree Y∞ . Since the sequence {fi }
contains no separable subsequence, the action of L on Y∞ must be geometric, and
contains no non-degenerate segments in its simplicial part that can not be divided
into finitely many segments with non-trivial (cyclic) stabilizers. Since (S, L) is
assumed to be Levitt-free, the action of L on Yinf ty contains (possibly) only IET
components and (possibly) a simplicial part.
Let Γ∞ be the abelian decomposition that is associated with the action of L on
Y∞ . Γinf ty has to be compatible with the stable dominant abelian decomposition
Θi0 , i.e., every elliptic element in Θi0 must be elliptic in Γ∞ . Our goal is to show
that Γ∞ can be further refined, by restricting the homomorphisms, {fi }, to non-QH
vertex groups in Γ∞ , and passing to a further subsequence, to the stable dominant
abelian decomposition, Θi0 .
We start with the following claim that is similar to lemma 6.5 in the freely
indecomposable case.
Lemma 7.9. Let Q be a QH vertex group in Θi0 that does not appear in any of
the abelian decompositions Λi , for i ≥ i0 .
If there is a non-peripheral element in Q that fixes a point in Y∞ , then the entire
QH vertex group Q fixes a point in Y∞ .
Proof: Suppose that a non-peripheral element q ∈ Q fixes a point in Y . q is
contained in some ball Bm in the Cayley graph of L w.r.t. s1 , . . . , sr . Θi0 is the
stable dominant abelian decomposition of the sequence of abelian decompositions:
Λ1 , . . . . Hence, there must exist an abelian decomposition Λi1 , for some i1 >
max(i0 , m), for which either:
(i) q is a non-peripheral element in some QH vertex group in Λi1 .
(ii) q is hyperbolic in the abelian decomposition Λi1 .
This implies that q is either non-peripheral element in QH vertex groups, or a hyperbolic element in all the abelian decompositions, Λi , for i ≥ i1 . These abelian
113
decompositions, Λi , i > i1 , may contain both stable QH vertex groups and stable
edge groups. By the construction of the procedure that produces the abelian decompositions, {Λi }, for such a non-peripheral element q in a QH vertex group, or
such an element q that is hyperbolic w.r.t. Λi1 , for which q ∈ Bm , m ≤ i1 , there
must exist an index i2 ≥ i1 , such that:
(i) q is hyperbolic w.r.t. all the abelian decompositions, Λi , for i ≥ i2 .
(ii) q is hyperbolic w.r.t. all the abelian decompositions that are obtained by
collapsing all the edges and all the QH vertex groups in Λi , i ≥ i2 , except
for a single QH vertex group or a single edge group that can be conjugated
into Q. In particular, q is hyperbolic in the abelian decomposition sΛi for
i ≥ i2 .
According to the procedure, if either (i) or (ii) hold for Λi2 , for some i2 ≥ i1 ,
then it remains true for all i ≥ i2 . Now, as in the proof of lemma 6.5, the traces
and the lengths of q in its actions on the trees that are associated with the abelian
decompositions sΛi2 +1 , . . . , are bounded below by either a (global) positive constant
times the lengths of the fixed set of generators of s QH vertex groups in Λi′ and by
lengths of the generators of edge groups in Λi′ , i2 ≤ i′ < i when acting on the tree
that is associated with sΛi .
This implies that if q is elliptic in the action of L on the limit tree Y∞ , then all
the QH vertex groups and all the edge groups that can be conjugated into Q in
all the abelian decompositions, Λi , i ≥ i2 , must be elliptic as well. Hence, by the
structure of the stable dominant abelian decomposition, Θi0 , Q must be elliptic as
well.
Lemma 7.9 guarantees that in the action of L on Y∞ , a QH vertex group in the
stable dominant abelian decomposition, Θi0 , is either elliptic or is associated with
an IET component. The next proposition proves that the abelian decomposition
Γ∞ that is associated with the action of L on the limit tree Y∞ , is dominated by
Θi0 . i.e., every QH vertex group in Γ∞ is an edge group in Θi0 , and every edge
(with non-trivial stabilizer) in Γ∞ is conjugate to an edge in Θi0 , which by part (ii)
of lemma 7.7 implies that such an edge appears in all the abelian decompositions
Λi and sΛi for large enough i.
Lemma 7.10. Let Q be a QH vertex group, and let E be an edge group in Γ∞ ,
the abelian decomposition that is associated with the action of L on Y∞ . Then Q
is conjugate to a QH vertex group in Θi0 , and E is conjugate to an edge group in
Θi0 .
Proof: Lemma 7.7 proves that for every large index i, every QH vertex group Q
in Θi0 , and every edge group E in Θi0 , there exists a QH vertex group with stable
weight in Λi that is conjugate to a subsurface in Q or an edge group with stable
weight in Λi that is conjugate to either a s.c.c. in Q or to the edge group E.
By lemma 7.9 a QH vertex group in Θi0 is either elliptic in Γ∞ , or it is conjugate
to a QH vertex group in Γ∞ . An edge e in Θ∞ may appear as an edge in Γ∞ , and
all the edge groups in Θi0 are elliptic in Γ∞ .
Suppose that for every large index i all the the QH vertex groups in Λi , and
all the edge groups in Λi , can be conjugated to subsurfaces of QH vertex groups
in Θi0 , or to s.c.c. in QH vertex group in Θi0 , or to edge groups (with non-trivial
stabilizers) in Θi0 .
114
By the procedure for the construction of the abelian decompositions, {Λi }, if
all the QH vertex groups in Θi0 fix points in Y∞ , then Γ∞ contains only some of
the edge groups with non-trivial stabilizers in Θi0 . These appear as edge groups
with stable weight in all the abelian decompositions, Λi , for large index i. Hence,
we may assume that at least one QH vertex group in Θi0 appears as a QH vertex
group in Γ∞ .
Let Q1 , . . . , Qv be the QH vertex groups in Θi0 that appear as QH vertex groups
in Γ∞ . Since the action of L on Y∞ is geometric, and for large i, Λi contains only
QH vertex groups and edge groups that can be conjugated into QH vertex groups
and edge groups in Θi0 , for every u ∈ L, the path from the base point in Y∞ to the
image of the base point under the action of u, is composed from (possibly) finitely
many subpaths that are contained in the IET components that are associated with
Q1 , . . . , Qv , and (possibly) finitely many edges with non-trivial stabilizers, that are
in the simplicial part of Y∞ . This implies that Γ∞ contains only conjugates the
QH vertex groups, Q1 , . . . , Qv , that are also QH vertex groups in Θi0 , and finitely
many edges, that are all edges in Θi0 , and appear as edges in Λi for large enough i.
Suppose that not all the edge groups and the QH vertex groups in Λi , for large i,
can be conjugated into edge groups and QH vertex groups in Θi0 . In that case, for
large i, there exist QH vertex groups with stable weight, or edges with non-trivial
edge groups with stable weight in Λi , that can not be conjugated into QH vertex
groups nor into edge groups in Θi0 .
Since the pair (S, L) is Levitt free, the action of L on Y∞ contains no Levitt
components. Since the limit action is constructed from a sequence of (gradually)
non-separable homomorphisms, the limit action of L on Y∞ must be geometric.
Therefore, Y∞ contains only IET components and a simplicial part, with which
there are associated QH vertex groups and edges with non-trivial stabilizers in Γ∞ .
The procedure for the construction of the abelian decompositions, {Λi }, forces
finite equivariance on generators of dominant QH vertex groups and edge groups,
that guarantees that an edge with non-trivial stabilizer exists in Γ∞ if and only
if conjugates of that edge exist in all the abelian decompositions, Λi , for large i.
In particular, such an edge must be conjugate to an edge in the stable dominant
abelian decomposition, Θi0 .
The finite equivariance that is forced on generators of dominant QH vertex groups
and dominant edge groups, also implies that for large enough i, all the QH vertex
groups in Λi , and all the edge groups in Λi , are either elliptic in Γ∞ , or they can
be conjugated into QH vertex groups or s.c.c. in QH vertex groups in Γ∞ . This
clearly implies that the boundaries of all the QH vertex groups in Γ∞ are elliptic
in all the abelian decompositions, Λi , for i large enough. Hence, all the QH vertex
groups in Γ∞ are in fact conjugate to QH vertex groups in Θi0 .
By lemma 7.10 all the QH vertex groups and all the edge groups in Γ∞ , are
(conjugates of) QH vertex groups and edge groups in the stable dominant abelian
decomposition Θi0 . Suppose that Θi0 contains QH vertex groups or edge groups
that don’t have conjugates in Γ∞ .
In that case we restrict the sequence of homomorphisms, {fi }, to the (elliptic)
vertex groups in Y∞ . Since Θi0 contains QH vertex groups that are not conjugate
to QH vertex groups in Γ∞ , for all large i, the abelian decompositions Λi contain
QH vertex groups or edge groups with stable weight that are elliptic in Γ∞ . Hence,
from the restrictions of the homomorphisms {fi } to the point stabilizers in Y∞ it
115
is possible to associate a non-trivial abelian decomposition with at least one of the
point stabilizers.
Therefore, we pass to a convergent subsequence of the sequnce {fi }, and associate with the non-QH vertex groups in Γ∞ abelian decompositions, that at least
one of them is non-trivial. Since the QH vertex groups and the edge groups in Γ∞
are conjugates of QH vertex groups and edge groups in the stable dominant abelian
decomposition, Θi0 , all the edge groups in Γ∞ are elliptic in the abelian decompositions of the various elliptic vertex groups. Hence, the abelian decompositions of
the various vertex groups further refine the abelian decomposition Γ∞ .
Let Γ1∞ be the obtained refinement of Γ∞ . By the same argument that was used
in proving lemma 7.10, Γ1∞ contains (new) QH vertex groups and edge groups that
are not conjugate to QH vertex groups and edge groups in Γ∞ , but they are all
conjugates of QH vertex groups and edge groups in Θi0 .
If Γ1∞ does not contain conjugates of all the QH vertex groups and all the edge
groups in Θi0 , we repeat the refinement process, by restricting the sequence of
homomorphisms, {fi }, to the elliptic vertex groups in Γ1∞ . After finitely many
iterations we obtain an abelian decomposition Γf , such that the QH vertex groups
and the edge groups in Γf and in Θi0 are conjugate. Furthermore, the (elliptic)
non-QH vertex groups in Γf are conjugates of the non-QH vertex groups in Θi0 .
By the properties of the procedures for the constructions of Γf and the sequence
of abelian decompositions, {Λi }, every element that is hyperbolic in Θi0 is hyperbolic in Γf . Hence, the collections of hyperbolic and elliptic elements in Γf and
Θi0 are identical. To be able to replace a suffix of the sequence of the abelian
decompositions, {Λi }, with the abelian decomposition Γf , we still need to prove
the following proposition.
Proposition 7.11. There exists an index i1 ≥ i0 , such that for every i ≥ i1 , the
modular group M od(Λi ) is contained in the modular group M od(Γf ).
Proof: Note that the modular groups, M od(Λi ) and M od(Γf ), are generated by
Dehn twists along edge groups and modular groups of QH vertex groups in the
two abelian decompositions. We have already argued that for large i the edge
groups in Λi can be conjugated into either edge groups in Γf , or into s.c.c. in QH
vertex groups in Γf , and the QH vertex groups in Λi can be conjugated into QH
vertex groups in Γf . Therefore, to prove the proposition we just need to analyze
the branching points in IET components in the trees that are associated with the
decompositions {Λi } and Γf for large i.
Suppose first that there are no edges with trivial stabilizers that are connected to
the QH vertex groups in Γf . This means that the IET components in the limit trees
from which Γf was obtained, contain no branching points that are also branching
points in other components, except for the orbits of points that are stabilized by
peripheral elements in the QH vertex groups.
Recall that the abelian decomposition Γf was obtained using a (finite) successive
refinement of an abelian decomposition Γ∞ . Let Y∞ be the tree from which the
abelian decomposition Γ∞ was obtained. The action of L on the real tree Y∞ is
geometric, hence, to analyze the branching points in Y∞ it is enough to look at the
segments, [y∞ , sj (y∞ )], where y∞ is the base point in Y∞ , and sj , 1 ≤ j ≤ r, are
the fixed set of generators of the semigroups S < L. Since the actions of L on each
of the trees Yi , from which the abelian decompositions Λi were obtained, are all
116
geometric, the same conclusion holds for these actions.
Since the action of L on Y∞ is geometric, the path [y∞ , sj (y∞ )], is divided
into (possibly) finitely many segments that are contained in IET components, and
(possibly) finitely many segments with non-trivial stabilizers in the simplicial part of
Γ∞ , where the last segments are associated with edges with non-trivial edge groups
in Γ∞ . Once again, the actions of L on the real trees Yi are geometric, hence, the
same conclusion holds for the segments, [yi , sj (yi )], where yi is the basepoint of the
tree Yi .
There exists an index i2 ≥ i0 , so that for every i ≥ i2 every QH vertex group in
Λi is a subgroup of a conjugate of a QH vertex group in Γf , and every edge group
in Λi is either conjugate to an edge group in Γf or it can be conjugated into a s.c.c.
in a QH vertex group in Γf .
Suppose that there is a subsequence of indices i, such that QH vertex groups
in Λi contain branching points that are not stabilized by one of their peripheral
elements. We can pass to a subsequence of the indices, for which such branching
points occur along the segments, [yi , sj (yi )], for some fixed j, 1 ≤ j ≤ r. Suppose
that we can pass to a further subsequence, so that there exist such branching
points that are not stabilized by peripheral elements in IET components that are
associated with conjugates of QH vertex groups in Λi , and these QH vertex groups
can be conjugated into QH vertex groups in Γ∞ (according to our assumptions
there exist such branching points in IET components that are associated with QH
vertex groups that can be conjugated into QH vertex groups in Γf . We assume that
there is a subsequence in which the QH vertex groups in Λi that their associated
IET components contain these branching points can be conjugated into QH vertex
groups in Γ∞ ).
The path [y∞ , sj (y∞ )] is divided into subpaths in IET components in Y∞ , and
segments with non-trivial stabilizers in the simplicial part of Y∞ . The paths
[yi , sj (yi )] can be divided into subpaths in IET components in Yi and segments
with non-trivial stabilizers in the simplicial part of Yi . Furthermore, the sequence
of subpaths in [yi , sj (yi )], can be divided into finitely many consecutive subsequences, such that the QH vertex groups and the edge groups in every subsequence
can be jointly conjugated into the same QH vertex group or edge group in Γf .
By our assumption [yi , sj (yi )] contains a branching point in an IET component,
that is not stabilized by a peripheral element in the QH vertex group that is associated with that IET component, and that QH vertex group can be conjugated
into a QH vertex group in Γ∞ . By passing to a further subsequence we may assume that such a branching point exists in a subsequence of subpaths in [yi , sj (yi )]
that is mapped into the same subpath in [y∞ , sj (y∞ )] that is contained in an IET
component in Y∞ .
Let Q be the QH vertex group in Γ∞ that is associated with that IET component.
Let u ∈ L be a peripheral element in Q that stabilizes the endpoint of the subpath
in [y∞ , sj (y∞ )] that is contained in an IET component that is associated with
Q. The point that is stabilized by u in that IET component is a branching point
in the division of [y∞ , sj (y∞ )]. By our assumption, for every index i from the
subsequence, the subpath in [yi , sj (yi )] that is mapped into the subpath in the IET
that is associated with Q in Y∞ , ends in a branching point that is not stabilized by
u.
The element u is contained in a ball of radius i3 in the Cayley graph of L w.r.t. the
generators s1 , . . . , sr . Hence, for all i > max(i3 , i2 ) the procedure that constructs
117
the sequences of homomorphisms, {hin }, takes into consideration the actions of the
images of all the elements in the ball of radius i3 in the Cayley graph of L. Since
there exists an index i > max(i3 , i2 ) for which u does not stabilize the branching
point at the end of the subpath in [yi , sj (yi )] ⊂ Yi , that is mapped into the subpath
of [y∞ , sj (y∞ )] ⊂ Y∞ , the procedure that constructs the homomorphisms, {hin },
guarantees that the endpoint of the subpath in Y∞ into which the corresponding
subpath in [yi , sj (yi )] is mapped can not be stabilized by the element u. This clearly
contradicts the assumption that u does stabilize the endpoint of that subpath in
Y∞ .
This argument proves that for large i, the subpaths of [yi , sj (yi )] that are mapped
into IET components in Y∞ , start and end by either points that are stabilized by
peripheral elements in conjugates of QH vertex groups in Λi , that can be conjugated
into the associated QH vertex group in Γ∞ , or in non-cyclic vertex groups that are
connected to edge groups in Λi , where these edge groups can be conjugated into
s.c.c. in the associated QH vertex group in Γ∞ .
The abelian decomposition Γf was obtained from Γ∞ by a finite refinement
procedure. By following the steps of this refinement procedure, and repeating
the argument that was used for subpaths that are mapped into subpaths in IET
components in Y∞ , it follows that for large i the IET components in Λi contain no
branching points that are not stabilized by peripheral elements. Hence proposition
7.11 follows in case the QH vertex groups in Γf are not connected to edges with
trivial stabilizers.
Now, Suppose that QH vertex groups in Γf are connected to edges with trivial stabilizers. First, suppose that there exists a QH vertex group in Γ∞ that
is connected to edges with trivial stabilizers. As we already argued, the paths,
[y∞ , sj (y∞ )], can be divided into finitely many subpaths that are either contained
in IET components in Y∞ , or they are segments with non-trivial stabilizers in the
simplicial part of Y∞ . The starting and ending points of these subpaths are either:
(1) associated with the beginning or the end of an edge in Γ∞ .
(2) stabilized by a peripheral element in a QH vertex group in Γ∞ .
(3) contained in an IET component but not stabilized by a peripheral element
of that IET component.
The argument that we applied in case there were no branching points of type
(3), implies that for large i, the branching points in Yi that start or end subpaths
of [yi , sj (yi )], 1 ≤ j ≤ r, that are mapped into subpaths in Γ∞ that start or end in
branching points of type (1) or (2) in Γ∞ , have to be of type (1) or (2) in Yi .
The germs of the (finitely many) branching points of type (3) in the segments,
[y∞ , sj (y∞ )], belong to finitely many orbits (under the action of L on Y∞ ). By the
same argument that we used in order to analyze the branching points of types (1)
and (2), if two starting or ending points of two sequences of consecutive subpaths
in [yi , sj (yi )], 1 ≤ j ≤ r, are mapped into starting or ending points of subpaths in
QH components in [y∞ , sj (y∞ )], so that the germs of these branching points in Y∞
are in the same orbit under the action of L, then the germs of the pair of branching
(starting or ending) points in [yi , sj (yi )] are in the same orbit in Yi under the action
of L.
Using the finite refinement procedure that led from Γ∞ into Γf , for large i, the
same hold for starting and ending points subpaths of [yi , sj (yi )], 1 ≤ j ≤ r, that are
mapped into subpaths in IET components or segments with non-trivial stabilizers
118
in one of the finitely many trees that were used to refine Γ∞ and obtain Γf . In
particular, the preimages of an orbit of branching points in Γ∞ in Yi , is in the
same orbit under the action of L. Using the refinement process the same holds for
preimages of orbits of branching points in Γf . This equivariance of the preimages
of branching points in Γf , guarantees that for large i, the modular groups M od(Λi )
are contained in the modular group M od(Γf ).
In the freely indecomposable case, in the proof of theorem 6.1 proposition 6.4
enables us to replace an infinite sequence of abelian decompositions by a finite
sequence that terminates in the stable dominant abelian decomposition, where the
modular group of the stable dominant abelian decomposition was guaranteed to
contain the modular groups of all the abelain decompositions of the (infinite) suffix
of abelian decompositions that was removed from the original sequence.
Proposition 7.11 enables us to proceed in the same way in the freely decomposable, Levitt-free case. i.e., given the infinite sequence of abelian decompositions,
{Λi }, we remove a suffix of it that is replaced by the abelian decomposition Γf .
Hence, the sequence, {Λi }, is replaced by the finite sequence: Λ1 , . . . , Λi1 −1 , Γf .
Note that unlike the freely indecomposable case, we replace the removed suffix
by the abelian decomposition Γf , that is obtained from a limit of the sequence
of homomorphisms, {fi }, and not by the stable dominant abelian decomposition,
Θi0 . In the freely decomposable case, these two abelian decompositions are equivalent, but in the freely decomposable, Levitt-free case, the modular groups that are
associated with the two abelian decompositions may be different.
We continue to the next steps of the construction with the sequence of homomorphisms {fi } and the abelian decomposition that was obtained from their limit Γf .
Recall, that the sequence {f i } was obtained from the sequences of pair homomorphisms that we denoted {hin }. The sequences {hin } were constructed by shortening
using elements from the dominant modular groups of the abelian decompositions,
{Λi }. By proposition 7.11 the modular groups of the abelian decompositions, Λi ,
for i ≥ i1 are all contained in the modular group of the abelian decomposition
Γf . Hence, the sequence of homomorphisms, {fi }, is obtained from the sequence,
{hni1 −1 }, by precompositions with automorphisms from the modular group of Γf .
We continue to the next step starting with the sequence of homomorphisms {f i },
and the associated abelian decomposition Γf . Note that this abelian decomposition
contains at least one QH vertex group. At this point we repeat the whole construction of a sequence of abelian decompositions. If the sequence terminates after a
finite number of steps, the conclusion of theorem 7.2 follows. If it ends up with
an infinite sequence of abelian decompositions, we replace it by a finite sequence
of abelian decompositions using the procedure that we described and proposition
7.11.
We continue iteratively as we did in the proof of theorem 6.1. If this iterative
procedure terminates after finitely many steps, the conclusion of theorem 7.2 follows. Otherwise we obtained an infinite sequence of abelian decompositions that
do all contain QH vertex groups. In each of these abelian decompositions, either:
(1) there exists a dominant edge group that is not elliptic in an abelian decomposition that appears afterwards in the sequence.
(2) there exists a QH vertex group Q in the abelian decomposition, such that the
abelian decomposition collapses to an abelian decomposition ΓQ that con119
tains one QH vertex group, Q, and possible several non-QH vertex groups
that are connected only to the vertex stabilized by Q by edges with trivial and cyclic edge groups. ΓQ and an abelian decomposition that appears
afterwards in the sequence do not have a common refinement (see the construction of the original sequence of abelian decompositions Λ1 , . . . ).
Now, we apply the procedure that analyzed the original sequence of abelian decompositions, {Λi }, to the sequence of abelian decompositions that we constructed.
By proposition 7.11 a suffix of the sequence can be replaced with an abelian decompositions that has the same elliptic subgroups as the stable dominant abelian
decomposition of the sequence, and this abelian decomposition contains either
(i) more than one QH vertex group, or a QH vertex group with higher (topological) complexity (see the proof of theorem 6.1).
(ii) only a single QH vertex group, possibly of the same topological complexity
(e.g. a once punctured torus), but this QH vertex has to be connected to
the other vertex groups in the abelian decomposition with at least one edge
with trivial stabilizer.
As in the proof of theorem 6.1, we repeat the whole construction starting with the
(higher complexity) abelian decomposition that we obtained and the subsequence
of homomorphisms that is associated with it. Either the construction terminates
in finitely many steps, or a suffix of an infinite sequence of abelian decompositions
can be replaced with an abelian decomposition of higher complexity. i.e., the total
topological complexity of the QH vertex groups that appear in the abelian decompositions is bounded below by a higher lower bound, or the minimum number of
edges with trivial stabilizers that are connected to the QH vertex groups is bigger. Repeating this construction itartively, by the accessibility of f.p. groups, or
by acylindrical accessibility, we are left with a finite resolution that satisfies the
conclusion of theorem 7.2 (see the proof of theorem 6.1).
So far we assumed that the pairs that we consider (S, L) are Levitt free, and L
contains no non-cyclic abelian subgroups. To deal with pairs that are not Levitt
free, we need the abelian decompositions that we construct to have special vertex
groups that are free factors of the ambient limit group L, and we call Levitt vertex
groups. If a limit group L admits a superstable action on a real tree, the action
is geometric, and every non-degenerate segment in the simplicial part of the tree
can be divided into finitely many segments with non-trivial abelian stabilizers,
then using the Rips’ machine, it is possible to associate with the action a graph
of groups (see [Gu]). The vertex groups in this graph of groups are either QH or
axial (abelian) or Levitt (thin) or point stabilizers. The stabilizers of an edge in
this graph of groups are either trivial or abelian. In particular, the edges that are
connected to Levitt vertex groups must have trivial stabilizers.
To generalize theorem 7.2 to pairs that are not necessarily Levitt free, we need to
allow the abelian decompositions that are associated with pairs along a resolution
to include Levitt vertex groups. This means that to the generators of the modular
groups that are associated with such abelian decompositions we need to add automorphisms of the free factors (that are free groups) that are Levitt vertex groups.
Furthermore, given a sequence of pair homomorphisms, to extract a subsequence
that factor through a resolution, similar to the one that appears in theorem 7.2, we
will need to generalize the JSJ machine that was used in the proof of theorem 7.2, to
120
construct larger and larger Levitt vertex groups. i.e., the JSJ type decompositions
that we consider need to include a new type of vertex groups (that was so far not
needed over groups), the Levitt vertex groups and their modular groups.
Theorem 7.12. Let (S, L) be a pair, and suppose that the limit group L contains
no non-cyclic abelian subgroup. Let {hn : (S, L) → (F Sk , Fk )} be a sequence of pair
homomorphisms that converges into a faithful action of L on a real tree Y .
Then there exists a resolution:
(S1 , L1 ) → . . . → (Sf , Lf )
that satisfies the following properties:
(1) (S1 , L1 ) = (S, L), and ηi : (Si , Li ) → (Si+1 , Li+1 ) is an isomorphism for
i = 1, . . . , f − 2 and ηf −1 : (Sf −1 , Lf −1 ) → (Sf , Lf ) is a quotient map.
(2) with each of the pairs (Si , Li ), 1 ≤ i ≤ f , there is an associated abelian
decomposition that we denote Λi . The abelian decompositions Λ1 , . . . , Λf −1
contain edges with trivial or cyclic edge stabilizers, and QH, Levitt and rigid
vertex groups.
(3) either ηf −1 is a proper quotient map, or the abelian decomposition Λf contains separating edges with trivial edge groups. Each separating edge is
oriented.
(4) there exists a subsequence of the homomorphisms {hn } that factors through
the resolution. i.e., each homomorphism hnr from the subsequence, can be
written in the form:
hnr = ĥr ◦ ϕrf −1 ◦ . . . ◦ ϕ1r
where ĥr : (Sf , Lf ) → (F Sk , Fk ), each of the automorphisms ϕir ∈ M od(Λi ),
where M od(Λi ) is generated by the modular groups that are associated with
the QH vertex groups, modular groups that are associated with Levitt vertex
groups (these are automorphisms of the Levitt free factors), and by Dehn
twists along cyclic edge groups in Λi .
Each of the homomorphisms:
i
hinr = ĥr ◦ ϕm
r ◦ . . . ◦ ϕr
is a pair homomorphism hinr : (Si , Li ) → (F Sk , Fk ).
(5) if (Sf , Lf ) is not a proper quotient of (S, L), then the pair homomorphisms
ĥr are compatible with Λf . Let R1 , . . . , Rv be the connected components of
Λf after deleting its (oriented) separating edges. The homomorphisms ĥr
are composed from homomorphisms of the fundamental groups of the connected components R1 , . . . , Rv , together with assignments of values from
F Sk to the oriented separating edges. The homomorphisms of the fundamental groups of the connected components R1 , . . . , Rv converge into a
faithful action of these groups on real trees with associated abelian decompositions: R1 , . . . , Rv .
Proof: We modify the same procedure that was used in proving theorem 7.2 to
include Levitt components. Recall that we started the proof of theorem 7.2 by
121
iteratively constructing abelian decompositions, Λ1 , . . . . As we pointed in the construction of these abelian decompositions in the Levitt free case, we may assume
that all the faithful actions of L on real trees that we consider are geometric, and in
the simplicial parts of these actions every non-degenerate segment can be divided
into finitely many subsegments with non-trivial stabilizers.
Let Λ be the abelian decomposition that is associated with the faithful action of
L on the real tree Y , that is obtained from a convergent subsequence of the given
sequence of homomorphisms, {hn }. Λ may contain rigid, QH and Levitt vertex
groups, and the stabilizers of edge groups may be trivial or cyclic. A Levitt vertex
group is adjacent only to edges with trivial stabilizers.
We start by refining the abelian decomposition Λ in a similar way to what we did
in the proof of theorem 7.2. First we replace each non-QH, non-Levitt vertex group
that is connected by periodic edge groups only to non-QH non-Levitt vertex groups,
by restricting the homomorphisms {hn } to such a vertex group, and replace the
vertex group with the obtained (non-trivial) abelian decomposition. Repeating this
refinement procedure iteratively, we get an abelian decomposition that we denote
Λ1 .
We fix finite generating sets of all the edge groups and all the non-QH, nonLevitt vertex groups in Λ1 . We divide the edge and non-QH, non Levitt vertex
groups in Λ1 into finitely many equivalence classes of their growth rates as we did
the proof of theorems 6.1 and 7.2. After possibly passing to a subsequence, there
exists a class that dominates all the other classes, that we call the dominant class,
that includes (possibly) dominant edge groups and (possibly) dominant non-QH,
non-Levitt vertex groups.
As in the freely indecomposable case, we denote by M od(Λ1 ) the modular group
that is associated with Λ1 . We set M XM od(Λ1 ) to be the dominant subgroup
that is generated by Dehn twists along dominant edge groups, and modular groups
of dominant QH and Levitt vertex groups, i.e., modular groups of those QH and
Levitt vertex groups that the lengths of the images of their fixed sets of generators
grow faster than the lengths of the images of the fixed generators of dominant edge
and non-QH, non-Levitt vertex groups. Note that since a Levitt vertex group is a
free factor of the ambient group L, the modular group of a Levitt vertex group is
defined to be the natural extension of the automorphism group of that free factor
to the ambient group L (acting by appropriate conjugations on all the other vertex
and edge groups).
We start by using the full modular group M od(Λ1 ). For each index n, we set the
pair homomorphism: h1n : (S, L) → (F Sk , Fk ), h1n = hn ◦ ϕn , where ϕn ∈ M od(Λ1 ),
to be a shortest pair homomorphism that is obtained from hn by precomposing it
with a modular automorphism from M od(Λ1 ). If there exists a subsequence of the
homomorphisms {h1n } that converges into a proper quotient of the pair (S, L), or
that converges into a non-geometric action of L on a real tree, or into an action that
contains a segment in its simplicial part, and this segment has a trivial stabilizer,
we set the limit of this subsequence to be (Sf , Lf ), and the conclusion of theorem
7.12 follows with a resolution of length 1.
Therefore, we may assume that every convergent subsequence of the homomorphisms {h1n } converges into a faithful geometric action of the limit group L on
some real tree. In that case we use only elements from the dominant modular
group, M XM od(Λ1 ).
We modify what we did in the proofs of theorems 6.1 and 7.2. First, we shorten
122
the action of each of the dominant QH and Levitt vertex groups using the procedure
that is used in the proofs of propositions 2.7, 2.8 and 2.18, 2.19. For each of
the dominant QH and Levitt vertex groups, the procedures that are used in the
proofs of these propositions give us an infinite collections of positive generators,
m
um
1 , . . . , ug , with similar presentations of the corresponding QH vertex groups,
which means that the sequence of sets of generators belong to the same isomorphism
class. Furthermore, with each of these sets of generators there associated words,
wjm , j = 1, . . . , r, of lengths that increase with m, such that a given (fixed) set of
m
m
positive elements can be presented as: yj = wjm (um
1 , . . . , ug ). The words wj are
words in the generators um
i and their inverses. However, they can be presented as
positive words in the generators um
j , and unique appearances of words tℓ , that are
and
their inverses (i.e., the words do not depend on
fixed words in the elements um
j
m m
m), and these elements tℓ (u1 , . . . , um
g ) are positive for every m.
With each of the dominant QH and Levitt vertex groups we associate a system of
generators u11 , . . . , u1g ). Since an IET action of a QH vertex group, and and a Levitt
action are indecomposable in the sense of [Gu], finitely many (fixed) translates of
each of the positive paths that are associated with the positive paths, u1i1 , cover
the positive path that is associated with u1i2 , and cover the positive paths that are
associated with the words tℓ (u11 , . . . , u1g ).
As in the proofs of theorems 6.1 and 7.2, we shorten the homomorphisms {hn }
using the dominant modular group M XM od(Λ1 ). For each homomorphism hn ,
we pick the shortest homomorphism after precomposing with an element from
M XM od(Λ1 ) that keeps the positivity of the given set of the images of the generators s1 , . . . , sr , and the elements u11 , . . . , u1g and tℓ (for each dominant QH and
Levitt vertex groups), and keeps their lengths to be at least the maximal length of a
constant multiple c1 of the maximal length of the image of a generator of a dominant
edge or non-QH, non-Levitt vertex group (the generators are chosen from the fixed
finite sets of generators of each of these vertex and edge groups). We further require
that after the shortening, the image of each of the elements u11 , . . . , u1g and tℓ , will
be covered by the finitely many translates that cover them in the limit action that
is associated with Λ1 . We (still) denote the obtained (shortened) homomorphisms
{h1n }.
By the shortening arguments that are proved in propositions 2.7, 2.8 and 2.18,
2.19, the lengths of the images, under the shortened homomorphisms {h1n }, of the
elements u11 , . . . , u1g and tℓ , that are associated with the various dominant QH and
Levitt vertex groups, are bounded by some constant c2 (that is independent of n)
times the maximal length of the images of the fixed generators of the dominant
vertex and edge groups.
We pass to a subsequence of the homomorphisms {h1n } that converges into an
action of L on some real tree with an associated abelian decomposition ∆2 . If the
action of L is not faithful, or non-geometric, or contains a non-degenerate segment
in its simplicial part that can not be divided into finitely many segments with nontrivial stabilizers, the conclusions of theorem 7.12 follow. Hence, we may assume
that the action of L is faithful, and the action is geometric. We further check if the
sequence {h1n } contains a separable subsequence (definition 7.3). If it does, we pass
to this subsequence, and the conclusion of theorem 7.12 follows (cf. proposition
7.4).
We further refine ∆2 , precisely as we refined the first abelian decomposition Λ.
123
ˆ 2.
We denote the obtained (possibly) refined abelian decomposition ∆
ˆ 2 in a similar way to what we did in the Levitt free case.
We further refine ∆
Let Q be a dominant QH vertex group in Λ1 , so that all its boundary elements are
ˆ 2 . Each shortened homomorphism, h1n , is obtained from hn , by preelliptic in ∆
composition with a (shortened) automorphism from the dominant modular group,
ϕn ∈ M XM od(Λ1 ). ϕn is a composition of elements from automorphisms of free
factors that are dominant Levitt components, modular groups of dominant QH vertex group in Λ, and Dehn twists along edges with dominant edge groups. We set
τn ∈ M XM od(Λ1 ) to be a composition of the same elements from automorphisms
of Levitt components, and from the modular groups of dominant QH vertex groups
in Λ1 , that are not the dominant QH vertex group Q, and the same Dehn twists
along edges with dominant edge group as the shortened homomorphism ϕn . Let
µn = hn ◦ τn .
The sequence of homomorphisms µn , converges into a faithful action of L on a
real tree, with an associated abelian decomposition ΓQ , that has a single QH vertex
group, a conjugate of Q, and possibly several edges with non-trivial edge groups,
ˆ 2.
that are edges in both Λ1 and ∆
ˆ 2 and ΓQ , have a common refineSuppose that the abelian decompositions, ∆
ment, in which a conjugate of Q appears as a QH vertex group, and all the Levitt
ˆ 2 that do not correcomponents and all the the edges and QH vertex groups in ∆
spond to s.c.c. or proper QH subgroups of Q also appear in the common refinement
(the elliptic elements in the common refinement are precisely those elements that
ˆ 2 ). This, in particular, implies that all the QH vertex
are elliptic in both ΓQ and ∆
groups that can not be conjugated into Q, all the Levitt components, and all the
ˆ 2 that can not be conjugated into non-peripheral elements in Q,
edge groups in ∆
ˆ2
are elliptic in ΓQ . In case there exists such a common refinement, we replace ∆
with this common refinement. We repeat this possible refinement for all the QH
vertex groups in Λ1 that satisfy these conditions (the refinement procedure does
not depend on the order of the QH vertex groups in Λ1 that satisfy the refinement
˜ 2.
conditions). We denote the obtained abelian decomposition, ∆
˜ 2 to possibly include Levitt vertex groups that appear in Λ1 .
We further refine ∆
˜ 2 that
Let B be a dominant Levitt vertex group in Λ1 . We look for a refinement of ∆
will include B as a vertex group, and we do it in a similar way to what we did with
dominant QH vertex group in Λ1 . For each index n, let ϕn ∈ M XM od(Λ1 ) be the
automorphism that was used to shorten hn , i.e., h1n = hn ◦ϕn . ϕn is a composition of
elements from automorphisms of free factors that are dominant Levitt components,
modular groups of dominant QH vertex group in Λ, and Dehn twists along edges
with dominant edge groups. We set γn ∈ M XM od(Λ1 ) to be a composition of
the same elements from automorphisms of dominant Levitt components, except for
the dominant Levitt component B, and from the modular groups of dominant QH
vertex groups in Λ1 , and the same Dehn twists along edges with dominant edge
group as the shortened homomorphism ϕn . Let νn = hn ◦ γn .
The sequence of homomorphisms νn , converges into a faithful action of L on
a real tree, with an associated abelian decomposition ΓB , that has a single Levitt
component, a conjugate of B, no QH vertex groups, and possibly several edges with
ˆ 2 (hence, also in ∆
˜ 2 ).
non-trivial edge groups, that are edges in both Λ1 and ∆
˜ 2 and ΓB , have a common refineSuppose that the abelian decompositions, ∆
ment, in which a conjugate of B appears as a Levitt component, and all the Levitt
124
˜ 2 that
components and all the the edge groups and the QH vertex groups in ∆
can not be conjugated into B also appear in the common refinement (the elliptic
elements in the common refinement are precisely those elements that are elliptic
˜ 2 ). This, in particular, implies that all the Levitt components,
in both ΓB and ∆
all the QH vertex groups and all the edge groups that can not be conjugated into
B, are elliptic in ΓB . In case there exists such a common refinement, we replace
˜ 2 with this common refinement. We repeat this possible refinement for all the
∆
Levitt vertex groups in Λ1 that satisfy these conditions (the refinement procedure
does not depend on the order of the Levitt vertex groups in Λ1 that satisfy the
refinement conditions). We denote the obtained abelian decomposition, Λ2 .
We continue as in the Levitt-free case. First we precompose the sequence of
homomorphisms, {h1n }, with a fixed automorphism ψ1 ∈ M XM od(Λ1 ), in the
same way it was done in the Levitt free case. Then we associate weights with the
edge groups, the QH vertex groups, and the Levitt vertex groups in Λ2 .
We proceed iteratively. First we shorten using the full modular group, M od(Λi ).
If the obtained sequence of homomorphisms has a separable subsequence, or a
subsequence that convergence into a proper quotient of the limit group L, the
conclusion of theorem 7.12 follows. Otherwise we use only the dominant modular
group, M XM od(Λi ).
We shorten the homomorphisms, {hi−1
n }, using the dominant modular group
M XM od(Λi ). We denote the obtained shortened sequence, {hin }. If the sequence
{hin } contains a separable subsequence, or a sequence that converges into a proper
quotient of L, the conclusion of theorem 7.12 follows. Otherwise we pass to a
convergent subsequence, refine the obtained associated decomposition precsiely as
we did in the first shortening step, and precompose the sequence, {hin }, with an
automorphism psii ∈ M XM od(Λi). The automorphism ψi is constructed in the
same way it was constructed in the Levitt-free case. We still denote the compositions, {hin ◦ ψi }, {hin }. Finally, precisely as we did in the Levitt-free case, we assign
weights with the edge groups with non-trivial stabilizers in Λi , and with QH and
Levitt vertex groups in Λi , precsiely as we did it in the Levitt-free case.
If in all steps the obtained actions are faithful, the actions are all geometric, and
the sequences of homomorphisms contain no separable subsequences, we get an
infinite sequence of abelian decompositions, Λ1 , Λ2 , . . . . As in the Levitt free case,
our goal is to replace a suffix of this sequence with a single abelian decomposition
that is obtained as a limit from a sequence of pair homomorphisms.
Given the sequence of abelian decompositions, Λ1 , . . . , we associate with it its
stably dominant decomposition Θi0 . Note that since the construction of the stably
dominant decomposition (definition 7.5) does not consider and does not encode the
precise factorization of the free factor which is a free group, one can not expect to
replace a suffix of the sequence of abelian decompositions: Λ1 , . . . , with Θi0 itself,
but rather with a modification of it, that associates a further (free) decomposition
with the free factor which is a free group (and in particular, specifies exactly the
Levitt components up to conjugacy).
First, we associate with the sequence of abelian decompositions, Λ1 , . . . , its finite
set of stable weights (see definition 7.6). As in the Levitt free case, given the stable
weights we modify the sequences of homomorphisms, {hin }. Recall that the unstable modular group of an abelian decomposition, Λi , is generated by the modular
groups of QH and Levitt vertex groups in Λi that have unstable weights, and Dehn
twists along edges with non-trivial stabilizers in Λi that have unstable weights (see
125
definition 7.6). For each index i, we precompose the homomorphisms {hin } with
automorphisms τni ∈ U SM od(Λi ), such that hin ◦ τni preserve all the positivity and
the (finite) equivariance properties that hin was supposed to preserve, and is the
shortest among all the homomorphisms of the form: hin ◦ τ that preserve the positivity and the finite equivariance requirements, and in which: τ ∈ U SM od(Λi ).
We denote the homomorphisms hin ◦ τn , shin .
If for some index i, the sequence {shin } contains a separable subsequence, or if
it contains a subsequence that converges into a proper quotient of the limit group
L, the conclusions of theorem 7.2 follow. Hence, in the sequel we may assume that
{shin } contain no such subsequences.
Lemmas 7.7 and 7.8 (and their proof) remain valid in the presence of Levitt
vertex groups. Hence, for all large i, and for every QH vertex group Q in the stable
dominant abelian decomposition, Θi0 , and for every edge group E with non-trivial
stabilizer in Θi0 , there exist QH vertex groups and edge groups with non-trivial
stabilizers with stable weights in Λi , such that these QH vertex groups and edge
groups with stable weights can be conjugated into Q or E.
As in the Levitt free case the sequence, {shin }, subconverges into an action on
a real tree from which an abelian decomposition, sΛi , can be obtained (using the
refinement procedure that was used in construction Λi ), where sΛi is obtained from
Λi by collapsing the following:
(1) Levitt and QH vertex groups and edge groups with unstable weight.
(2) non-dominant Levitt and QH vertex groups of stable weight that are not
conjugate to Levitt and QH vertex groups in Θi0 .
(3) edge groups of stable weight that are not conjugate to edges in Θi0 , and
for which the length of the corresponding edge in the real tree into which
{hin } converges, is bounded by a constant times the length of a generator
of a dominant edge group or a dominant non-QH vertex group.
At this point we are ready to choose a sequence of homomorphisms, {fi }, that
will subconverge to an abelian decomposition that is going to replace a suffix of
the sequence of decompositions, Λ1 , . . . . For each index i, it is possible to choose a
homomorphism, fi , from the sequence {hin }, that satisfies similar properties as in
the Levitt free case.
For each index i ≥ i0 , we choose the homomorphism fi , to be a homomorphism
from the sequence {shin } that is not separable with respect to the abelian decompositions: ∆1 , . . . , ∆i , where the sequence, {∆i }, enumerates all the possible graphs
of groups with fundamental group L and trivial edge groups (see definition 7.3).
The sequence of homomorphisms {shin } subconverges into a faithful action of L
on the limit tree sYi+1 . We require fi to approximate the action on the limit tree
sYi+1 , of all the elements in a ball of radius i in the Cayley graph of L (w.r.t. the
given generating set s1 , . . . , sr ), of all the fixed sets of generators of the QH and
Levitt vertex groups, and of edge groups with non-trivial stabilizers in Λ1 , . . . , Λi ,
and of all the (finitely many) elements that were chosen to demonstrate the mixing
property of the dominant QH and Levitt vertex groups and dominant edge groups
in Λ1 , . . . , Λi .
The sequence of homomorphisms, {fi }, subconverges into a faithful action of
the limit group L on a real tree Y∞ . Since the sequence {fi } contains no separable
subsequence, the action of L on Y∞ must be geometric, and contains no nondegenerate segments in its simplicial part that can not be divided into finitely
126
many segments with non-trivial (cyclic) stabilizers.
Let Γ∞ be the abelian decomposition that is associated with the action of L on
Y∞ . Γ∞ may have Levitt and QH vertex groups, non-QH non-Levitt vertex groups,
and edges with trivial and with non-trivial stabilizers. Γinf ty has to be compatible
with the stable dominant abelian decomposition Θi0 , i.e., every elliptic element in
Θi0 must be elliptic in Γ∞ . Our goal is to show that Γ∞ can be further refined, by
restricting the homomorphisms, {fi }, to non-QH vertex groups in Γ∞ , and passing
to a further subsequence, to an abelian decomposition that have the same elliptic
elements as the stable dominant abelian decomposition, Θi0 . This refinement of
Γ∞ is going to be used to replace a suffix of the sequence of abelian decompositions,
Λ1 , . . . .
Lemma 7.9 (and its proof) remain valid in the presence of Levitt components.
Hence, a QH vertex group in the stable dominant abelian decomposition, Θi0 , is
either elliptic in Γ∞ or it is conjugate to a QH vertex group in Γ∞ . Lemma 7.10
is also valid in the presence of Levitt components, and its proof requires a slight
modification.
Lemma 7.13. Let Q be a QH vertex group, and let E be an edge group in Γ∞ ,
the abelian decomposition that is associated with the action of L on Y∞ . Then Q
is conjugate to a QH vertex group in Θi0 , and E is conjugate to an edge group in
Θi0 .
Proof: By lemma 7.9 a QH vertex group in Θi0 is either elliptic in Γ∞ , or it is
conjugate to a QH vertex group in Γ∞ . An edge e in Θ∞ may appear as an edge
in Γ∞ , and all the edge groups in Θi0 are elliptic in Γ∞ .
Let Q1 , . . . , Qv be the QH vertex groups in Θi0 that are conjugate to QH vertex
groups in Γ∞ . Suppose that for large i, Λi contains only QH vertex groups and edge
groups that can be conjugated into QH vertex groups and edge groups in Θi0 . In
that case for large i, Λi does not contain Levitt components, and Γ∞ contains only
conjugates of the QH vertex groups, Q1 , . . . , Qv , that are also QH vertex groups
in Θi0 , and finitely many edges, that are all edges in Θi0 , and appear as edges in
Λi for large enough i.
Suppose that not all the edge groups and the QH and Levitt vertex groups in Λi ,
for large i, can be conjugated into edge groups and QH vertex groups in Θi0 (Levitt
vertex groups in Λi for large i can never be conjugated into QH vertex groups in
Θi0 . In that case, for large i, there exist QH or Levitt vertex groups with stable
weight, or edges with non-trivial edge groups with stable weight in Λi , that can not
be conjugated into QH vertex groups nor into edge groups in Θi0 .
Since the limit action is constructed from a sequence of (gradually) non-separable
homomorphisms, the limit action of L on Y∞ must be geometric. Therefore, Y∞
(possibly) contains only IET and Levitt components and a simplicial part, with
which there are associated Levitt and QH vertex groups and edges with non-trivial
stabilizers in Γ∞ .
The procedure for the construction of the abelian decompositions, {Λi }, forces
finite equivariance on generators of dominant Levitt and QH vertex groups and
edge groups, that guarantees that an edge with non-trivial stabilizer exists in Γ∞
if and only if conjugates of that edge exist in all the abelian decompositions, Λi ,
for large i. In particular, such an edge must be conjugate to an edge in the stable
dominant abelian decomposition, Θi0 .
127
The finite equivariance that is forced on generators of dominant QH and Levitt
vertex groups and dominant edge groups, also implies that for large enough i, all
the QH and Levitt vertex groups in Λi , and all the edge groups in Λi , are either
elliptic in Γ∞ , or they can be conjugated into QH or Levitt vertex groups or s.c.c.
in Γ∞ . This clearly implies that the boundaries of all the QH vertex groups in Γ∞
are elliptic in all the abelian decompositions, Λi , for i large enough. Hence, all the
QH vertex groups in Γ∞ are in fact conjugate to QH vertex groups in Θi0 .
By lemma 7.13 all the QH vertex groups and all the edge groups in Γ∞ , are
(conjugates of) QH vertex groups and edge groups in the stable dominant abelian
decomposition Θi0 . Γ∞ may also contain some Levitt vertex groups. As we have
already indicated in the proof of lemma 7.13, for large i, every Levitt vertex group
in Λi is either elliptic in Γ∞ or it can be conjugated into a Levitt vertex group in
Γ∞ . Every QH vertex group in Λi is either elliptic in Γ∞ or it can be conjugated
into either a Levitt vertex group or into a QH vertex group in Γ∞ . Every dominant
edge group in Λi is either elliptic in Γ∞ , or it can be conjugated into a s.c.c. in a
QH vertex group in Γ∞ or into a Levitt component in Γ∞ . In the last case the
stabilizer of the Levitt vertex group inherits a non-trivial abelian decomposition
from Λi , an inherited abelian decomposition in which the edge with the dominant
edge group in Λi appears as an edge.
Suppose that not all the elliptic elements in Γ∞ are elliptic in Θi0 (i.e., suppose
that non-QH non-Levitt vertex groups in Γ∞ are not conjugates of non-QH vertex
group in Θi0 ).
In that case we restrict the sequence of homomorphisms, {fi }, to the (elliptic)
vertex groups in Y∞ . Since these elliptic subgroups are not elliptic in Θi0 , for large
enough i, the abelian decompositions Λi contain QH or Levitt vertex groups or
edge groups with stable weight that are elliptic in Γ∞ . Hence, from the restrictions
of the homomorphisms {fi } to the point stabilizers in Y∞ it is possible to associate
a non-trivial abelian decomposition with at least one of the point stabilizers.
Therefore, we pass to a convergent subsequence of the sequnce {fi }, and associate
with the elliptic subgroups in Γ∞ abelian decompositions, that at least one of them
is non-trivial. Since the QH vertex groups and the edge groups in Γ∞ are conjugates
of QH vertex groups and edge groups in the stable dominant abelian decomposition,
Θi0 , all the edge groups in Γ∞ are elliptic in the abelian decompositions of the
various elliptic vertex groups. Hence, the abelian decompositions of the various
vertex groups further refine the abelian decomposition Γ∞ .
Let Γ1∞ be the obtained refinement of Γ∞ . Γ1∞ contains (possibly) new QH or
Levitt vertex groups and edge groups that are not conjugate to QH and Levitt
vertex groups and edge groups in Γ∞ . By the proof of lemma 7.13, all the new QH
vertex groups in Γ1∞ y are conjugates of QH vertex groups in Θi0 , and all the new
edge groups in Γ1∞ are conjugates of edge groups in Θi0 .
If the elliptic vertex groups in Γ1∞ are not elliptic in Θi−0 we repeat the refinement
process, by restricting the sequence of homomorphisms, {fi }, to the elliptic vertex
groups in Γ1∞ . After finitely many iterations we obtain an abelian decomposition
Γf . The QH vertex groups and the edge groups in Γf are conjugates of the QH
vertex groups and the edge groups in Θi0 . Γf may contain Levitt components.
The elliptic (non-QH, non-Levitt) vertex groups in Γf are conjugates of the elliptic
vertex groups in Θi0 .
As in the Levitt free case, to be able to replace a suffix of the sequence of the
128
abelian decompositions, {Λi }, with the abelian decomposition Γf , we still need to
prove an analogue of proposition 7.11 in the presence of Levitt components.
Proposition 7.14. There exists an index i1 ≥ i0 , such that for every i ≥ i1 , the
modular group M od(Λi ) is contained in the modular group M od(Γf ).
Proof: The modular groups, M od(Λi ) and M od(Γf ), are generated by automorphisms of Levitt factors, Dehn twists along edge groups and modular groups of
QH vertex groups in the two abelian decompositions. We have already argued that
for large i the edge groups in Λi can be conjugated into either edge groups in Γf ,
or into s.c.c. in QH vertex groups or into Levitt vertex groups in Γf , and in case
it can be conjugated into a Levitt factor, the Levitt factor inherits a non-trivial
decomposition along that edge group that it inherits from Λi . QH vertex groups
in Λi can be conjugated into QH vertex groups or into Levitt factors in Γf , and
Levitt factors in Λi can be conjugated into Levitt factors in Γf . Therefore, as in
the Levitt free case (proposition 7.11), to prove the proposition we just need to
analyze the branching points in IET and Levitt components in the trees that are
associated with the decompositions {Λi } and Γf .
The abelian decomposition Γf was obtained using a (finite) successive refinement
of an abelian decomposition Γ∞ . Let Y∞ be the tree from which the abelian
decomposition Γ∞ was obtained. The action of L on the real tree Y∞ is geometric,
hence, to analyze the branching points in Y∞ it is enough to look at the segments,
[y∞ , sj (y∞ )], where y∞ is the base point in Y∞ , and sj , 1 ≤ j ≤ r, are the fixed set
of generators of the semigroups S < L. Since the actions of L on each of the trees
Yi , from which the abelian decompositions Λi were obtained, are all geometric, the
same conclusion holds for these actions.
Since the action of L on Y∞ is geometric, the path [y∞ , sj (y∞ )], is divided into
(possibly) finitely many segments that are contained in Levitt components, IET
components, and (possibly) finitely many segments with non-trivial stabilizers in
the simplicial part of Γ∞ , where the last segments are associated with edges with
non-trivial edge groups in Γ∞ . The actions of L on the real trees Yi are geometric,
hence, the same conclusion holds for the segments, [yi , sj (yi )], where yi is the
basepoint of the tree Yi .
There exists an index i2 ≥ i0 , so that for every i ≥ i2 every Levitt vertex vertex
group in Λi is a subgroup of a conjugate of a Levitt vertex group in Γf , every QH
vertex group in Λi is a subgroup of a conjugate of a Levitt or a QH vertex group
in Γf , and every edge group in Λi is either conjugate to an edge group in Γf or it
can be conjugated into a s.c.c. in a QH vertex group in Γf , or it can be conjugated
into a Levitt component in Γf .
The path [y∞ , sj (y∞ )] is divided into subpaths in Levitt and IET components
in Y∞ , and segments with non-trivial stabilizers in the simplicial part of Y∞ . The
paths [yi , sj (yi )] can be divided into subpaths in Levitt and IET components in Yi
and segments with non-trivial stabilizers in the simplicial part of Yi . Furthermore,
the sequence of subpaths in [yi , sj (yi )], can be divided into finitely many consecutive
subsequences, such that the Levitt and QH vertex groups and the edge groups in
every subsequence can be jointly conjugated into the same Levitt or QH vertex
group or edge group in Γf .
We first look at those sequences of consecutive subpaths in [yi , sj (yi )] that are
not mapped to points in Y∞ . The starting and ending points of these consecutive
129
subpaths are either:
(1) associated with the beginning or the end of an edge in Γ∞ .
(2) stabilized by a peripheral element in a QH vertex group in Γ∞ .
(3) contained in an IET component but not stabilized by a peripheral element
of that IET component.
(4) contained in a Levitt component in Y∞ .
The argument that we applied in the Levitt free case, implies that for large i,
the branching points in Yi that start or end subpaths of [yi , sj (yi )], 1 ≤ j ≤ r, that
are mapped into subpaths in Γ∞ that start or end in branching points of type (1)
or (2) in Γ∞ , have to be of type (1) or (2) in Yi .
The germs of the (finitely many) branching points of types (3) and (4) in the
segments, [y∞ , sj (y∞ )], belong to finitely many orbits (under the action of L on
Y∞ ). By the same argument that we used in order to analyze the branching points
of types (1) and (2) in the Levitt free case, if two starting or ending points of
two sequences of consecutive subpaths in [yi , sj (yi )], 1 ≤ j ≤ r, are mapped into
starting or ending points of subpaths in QH or Levitt components in [y∞ , sj (y∞ )],
so that the germs of these branching points in Y∞ are in the same orbit under the
action of L, then the germs of the pair of branching (starting or ending) points in
[yi , sj (yi )] are in the same orbit in Yi under the action of L.
Using the finite refinement procedure that led from Γ∞ into Γf , for large i, the
same hold for starting and ending points subpaths of [yi , sj (yi )], 1 ≤ j ≤ r, that are
mapped into subpaths in IET or Levitt components or segments with non-trivial
stabilizers in one of the finitely many trees that were used to refine Γ∞ and obtain
Γf . In particular, the preimages of an orbit of branching points in Γ∞ in Yi , is in the
same orbit under the action of L. Using the refinement process the same holds for
preimages of orbits of branching points in Γf . This equivariance of the preimages
of branching points in Γf , guarantees that for large i, the modular groups M od(Λi )
are contained in the modular group M od(Γf ).
As in the Levitt free case, proposition 7.14 enables us to remove a suffix of the
sequence of abelian decompositions, Λ1 , . . . , and replace it by the finite sequence:
Λ1 , . . . , Λi1 −1 , Γf . We continue to the next steps with the abelian decomposition,
Γf , and with the sequence of pair homomorphisms, {f i }, that converges to a limit
action from which Γf was obtained.
The terminal abelian decomposition of that finite resolution contains at least
one QH vertex group or one Levitt component. We repeat the whole construction of a sequence of abelian decompositions starting with the sequence of pair
homomorphisms {f i }. If the sequence terminates after a finite number of steps,
the conclusion of theorem 7.12 follows. If it ends up with an infinite sequence of
abelian decompositions, we use the same construction as the one that was used in
proving proposition 7.14.
We continue iteratively as we did in the proof of theorems 6.1 and 7.2. If this iterative procedure terminates after finitely many steps, the conclusion of theorem 7.12
follows. Otherwise we obtained an infinite sequence of abelian decompositions that
do all contain QH or Levitt vertex groups. In each of these abelian decompositions,
either:
(1) there exists a dominant edge group that is not elliptic in an abelian decomposition that appears afterwards in the sequence.
130
(2) there exists a QH vertex group Q in the abelian decomposition, such that
the abelian decomposition collapses to an abelian decomposition ΓQ that
contains one QH vertex group, Q, and possibly several non-QH, non Levitt
vertex groups that are connected only to the vertex stabilized by Q by edges
with trivial and cyclic edge groups. ΓQ and an abelian decomposition that
appears afterwards in the sequence do not have a common refinement (see
the construction of the original sequence of abelian decompositions Λ1 , . . . ).
(3) there exists a Levitt vertex group B in the abelian decomposition, such
that the abelian decomposition collapses to an abelian decomposition Γ̃B
that contains one Levitt vertex group, B, and possibly several non-QH, non
Levitt vertex groups that are connected only to the vertex stabilized by B
by edges with trivial edge groups. ΓB and an abelian decomposition that
appears afterwards in the sequence do not have a common refinement.
Now, we apply proposition 7.14 to the sequence of abelian decompositions that
we constructed. By proposition 7.14 a suffix of the sequence can be replaced with
the abelian decomposition, such that all the modular groups that are associated
with the abelian decompositions from the suffix are contained in the modular group
of the abelian decomposition that the proposition produces and dominates the
suffix. Hence, the abelian decomposition that is obtained using proposition 7.14
contains either:
(i) a Levitt vertex group.
(ii) a Levitt vertex vertex group and a QH vertex group, or two Levitt vertex
groups, or two QH vertex groups.
(iii) more than one QH vertex group, or a QH vertex group with higher (topological) complexity (see the proof of theorem 6.1).
(iv) only a single QH vertex group, possibly of the same topological complexity
as QH vertex groups that appear in the abelian decompositions in the suffix
(e.g. a once punctured torus), but this QH vertex has to be connected to
the other vertex groups in the abelian decomposition with at least one edge
with trivial stabilizer.
(v) suppose that a subsequence of the abelian decompositions contain Levitt
vertex groups. Then either the abelian decomposition that dominates a
suffix of the entire sequence contains two Levitt vertex groups, or a Levitt
vertex group of rank at least 3, or a single Levitt vertex group that is
connected to other vertex groups in the abelian decomposition with at least
two edge groups with trivial stabilizers.
As in the proof of theorems 6.1 and 7.2, we repeat the whole construction starting with the (higher complexity) abelian decomposition that we obtained and the
subsequence of homomorphisms that is associated with it. Either the construction
terminates in finitely many steps, or a suffix of an infinite sequence of abelian decompositions can be replaced with an abelian decomposition of higher complexity.
i.e., either the ranks of the Levitt vertex groups (ordered lexichographically from
highest to lowest ranks), or the topological complexities of the QH vertex groups
(ordered lexichographically from high to low), that appear in the abelian decompositions, is bounded below by a higher lower bound, or the minimum number of
edges with trivial stabilizers that are connected to the Levitt and QH vertex groups
is bigger. Repeating this construction itartively, by the accessibility of f.p. groups,
or by acylindrical accessibility, we are left with a finite resolution that satisfies the
131
conclusion of theorem 7.12 (see the proof of theorem 6.1).
For presentation purposes, in theorems 7.2 and 7.12 we assumed that the limit
group L contains no non-cyclic abelian groups. To include abelian vertex groups
we use the same modifications as we used in the freely indecomposable case (see
definition 6.6 and theorem 6.7).
Theorem 7.15. Let (S, L) be a pair, and let s1 , . . . , sr be a fixed generating set of
the semigroup S. Let {hn : (S, L) → (F Sk , Fk )} be a sequence of pair homomorphisms that converges into a faithful action of L on a real tree Y .
Then there exists a resolution:
(S1 , L1 ) → (S2 , L2 ) → . . . → (Sm , Lm ) → (Sf , Lf )
that satisfies the following properties:
(1) (S1 , L1 ) = (S, L), and ηi : (Si , Li ) → (Si+1 , Li+1 ) is an isomorphism for
i = 1, . . . , m − 1 and ηm : (Sm , Lm ) → (Sf , Lf ) is a proper quotient map.
(2) with each of the pairs (Si , Li ), 1 ≤ i ≤ f , there is an associated abelian
decomposition that we denote Λi . The abelian decompositions Λ1 , . . . , Λf −1
contain edges with trivial or abelian edge stabilizers, and QH, Levitt, abelian
(axial), and rigid vertex groups.
(3) either ηf −1 is a proper quotient map, or the abelian decomposition Λf contains separating edges with trivial edge groups. Each separating edge is
oriented.
(4) there exists a subsequence of the homomorphisms {hn } that factors through
the resolution. i.e., each homomorphism hnr from the subsequence, is obtained from a homomorphism of the terminal pair (Sf , Lf ) using a composition of a modification that uses generalized Dehn twists that are associated with Λm , and modular automorphisms that are associated with
Λ1 , . . . , Λm−1 .
(5) if (Sf , Lf ) is not a proper quotient of (S, L), then the (shortened) pair homomorphisms {hfn } of (Sf , Lf ) (that are obtained by shortening a subsequence
of the homomorphisms {hn }) are compatible with Λf . Let R1 , . . . , Rv be the
connected components of Λf after deleting its (oriented) separating edges.
The homomorphisms ĥfn are composed from homomorphisms of the fundamental groups of the connected components R1 , . . . , Rv , together with
assignments of values from F Sk to the oriented separating edges. The
homomorphisms of the fundamental groups of the connected components
R1 , . . . , Rv converge into a faithful action of these groups on real trees with
associated abelian decompositions: R1 , . . . , Rv .
Proof: The theorem follows from theorem 7.7 using the same argument that was
used to prove theorem 6.7 from theorem 6.1.
Theorem 7.15 generalizes theorem 6.7 to general pairs. Given a sequence of pair
homomorphisms of a pair (S, L), and a sequence of pair homomorphisms {hn :
(S, L) → (F Sk , Fk )}, that converges into a faithful action of L on some real tree,
it proves that it is possible to extract a subsequence that factors through a finite
resolution of the pair (S, L) that terminates in a proper quotient of the pair (S, L), or
132
with a sequence of separable homomorphisms of (S, L). This allows one to continue
the construction of a resolution iteratively and get a resolution that terminates with
a graph of groups in which all the edges are separating edges and all the vertex
groups are trivial.
Proposition 7.16. Let (S, L) be a pair, and let s1 , . . . , sr be a fixed generating
set of the semigroup S. Let {hn : (S, L) → (F Sk , Fk )} be a sequence of pair
homomorphisms that converges into a faithful action of L on a real tree Y .
Then there exists a resolution:
(S1 , L1 ) → (S2 , L2 ) → . . . → (Sf , Lf )
that satisfies the following properties:
(1) (Si+1 , Li+1 ) is a quotient of (Si , Li ), but not necessarily a proper quotient.
Lf is a free group.
(2) with each of the pairs (Si , Li ), 1 ≤ i ≤ f , there is an associated abelian
decomposition that we denote Λi . The abelian decompositions Λ1 , . . . , Λf −1
contain edges with trivial or abelian edge stabilizers, and QH, Levitt, abelian
(axial), and rigid vertex groups. Some edges in these abelian decompositions
may be separating edges, and separating edges in Λi are canonically mapped
into separating edges in Λi+1 .
(3) Λf that is associated with the terminal pair, (Sf , Lf ), is a graph (of groups)
with trivial vertex groups, and all its edges are separating edges. In particular, they are oriented and with trivial stabilizers. The fundamental group
of this graph is Lf , and the generators of the semigroup S, s1 , . . . , sr , represent positive closed paths in the graph.
(4) there exists a subsequence of the homomorphisms {hn } that factors through
the resolution. i.e., each homomorphism hnr from the subsequence, is obtained from a homomorphism of the terminal pair (Sf , Lf ) using a composition of generalized Dehn twists and modular automorphisms that are
associated with Λ1 , . . . , Λf −1 .
Proof: We start with the sequence of homomorphisms {hn }. By theorem 7.15 it is
possible to extract a subsequence that factors to a finite resolution Res1 . The finite
resolution Res1 terminates in either a proper quotient of the original pair (S, L),
or it ends with a pair that is isomorphic to (S, L), but the abelian decomposition
that is associated with the terminal pair has separating edges (see definition 7.3).
Let {hm
nk } be the sequence of homomorphisms that is obtained from a subsequence of the original sequence of homomorphisms {hn } after factoring through
the resolution Res1 (this subsequence, {hm
nk } was used to construct the abelian
decomposition that is associated with the terminal level of Res1 ). If the terminal
pair that is associated with the resolution of Res1 is a proper quotient of (S, L), we
continue by applying theorem 7.11 to the subsequence {hm
nk }.
Suppose that the terminal pair of Res1 is isomorphic to (S, L). In that sequence
m
{hnk } is separable with respect to an abelian decomposition ∆1 , in which all the
edges have trivial edge groups, and they are all separating edges. In that case
we replace the homomorphisms, {hm
nk }, with a sequence of homomorphisms into a
bigger coefficient group as described in part (5) of definition 7.3. Recall that n that
part of definition 7.3, given each of the homomorphisms hm
nk , we replace the values
133
that are assigned with each of the labels that are associated with the edges in ∆1 , to
values that contain a single (positively oriented) appearance of the generator that
is associated with each label, and no appearances of generators that are associated
with the other labels, without changing the homomorphisms of the vertex groups
m
in ∆1 to obtain a homomorphism ĥm
nk . Each such homomorphism ĥnk is a pair
homomorphism: ĥm
nk : (Sm , Lm ) → (F̂ S, F̂ ), where (Sm , Lm ) is the terminal limit
group of Res1 , and (F̂s , F̂ ) is a standard extension of the standard coefficient pair,
(F Sk , Fk ), that is obtained by adding a finite set of new generators.
Now, we apply theorem 7.15 to the modified homomorphisms {hm
nk }, where we
analyze and modify (shorten) only the actions of the vertex groups in ∆. The
conclusion of theorem 7.11 gives another resolution Res2 . Either the terminal pair
of Res2 is a proper quotient of the original pair (S, L), or the resolution terminates
with an isomorphic pair with an associated separating abelian decomposition ∆2 ,
and the number of separating edges in ∆2 is strictly bigger than the number of
separating edges in ∆1 .
We proceed iteratively, and modify the homomorphisms that are the output of
the construction of a resolution, whenever a new separating edges are associated
with the terminal limit group of a resolution. By the d.c.c. property of limit groups,
the construction terminates after finitely many steps. By construction, the terminal
group has to be a free group, that is associated with a graph (of groups), in which
all edge and vertex groups are trivial and all the edges are separating edges.
Proposition 7.16 enables us to apply a compactness argument, and associate a
Makanin-Razborov diagram with a pair (S, L).
Theorem 7.17. Let (S, G) be a pair of a group G and a f.g. subsemigroup S that
generates G as a group. Then there exist finitely many resolutions of the form that
is constructed in proposition 7.11:
(S0 , L0 ) → (S2 , L2 ) → . . . → (Sf , Lf )
where (S0 , L0 ) is a quotient pair of (S, G), and each Li is a limit group, such that:
(1) every pair homomorphism, h : (S, L) → (F Sk , Fk ), factors through at least
one of these finitely many resolutions.
(2) for each of the resolutions in the collection, there exists a sequence of homomorphisms: {hn : (S, L) → (F Sk , Fk )}, that converges into a faithful
action of the initial pair (S0 , L0 ) on a real tree with an associated abelian
decomposition (after the refinement that is used in the proof of proposition
6.4) Λ0 .
Furthermore, the sequence of homomorphisms {hn } can be modified using modular automorphisms and generalized Dehn twists that are associated
with the abelian decompositions: Λ0 , . . . , Λm , and by adding separators to
separating edges, to get sequences of pair homomorphisms {h1n }, . . . , {hfn }.
Each of these modified sequences of homomorphisms {hin } converges into a
faithful action of the pair (Si , Li ) on a real tree with an associated abelian
decomposition (after an appropriate refinement) Λi .
Proof: Proposition 7.16 shows that given any sequence of pair homomorphisms of
(S, G) into (F Sk , Fk ), it is possible to extract a subsequence that factors through a
134
(finite) resolution that satisfies part (2) of the theorem. Hence, the theorem follows
by the same compactness argument that was used in the proof of theorem 6.8.
Theorem 7.17 associates a Makanin-Razborov diagram with a pair, and theorem
6.8 associates such a diagram with a freely indecomposable restricted pair. We end
this chapter by associating a Makanin-Razborov diagram with a general restricted
pair.
Theorem 7.18. Let (S, G) be a restricted pair of a group G and a f.g. subsemigroup
S that generates G as a group. i.e., the standard pair (F Sk , Fk ) is a subpair of
(S, G). Then there exist finitely many resolutions of the form:
(S0 , L0 ) → (S2 , L2 ) → . . . → (Sf , Lf )
where (S0 , L0 ) is a quotient restricted pair of (S, G), and each Li is a restricted
limit group, such that:
(1) (Si+1 , Li+1 ) is a restricted quotient pair of (Si , Li ), but not necessarily a
proper quotient. Lf is a restricted free product of a (possibly trivial) free
group and the coefficient group Fk and the coefficient group Fk .
(2) with each of the pairs (Si , Li ), 1 ≤ i ≤ f , there is an associated abelian
decomposition that we denote Λi that satisfies part (2) in proposition 7.8.
The coefficient group Fk is contained in a (distinguished) vertex group in
each of the abelian decomposition Λi .
(3) Λf that is associated with the terminal pair, (Sf , Lf ), is a graph of groups
with a single non-trivial vertex group, which is the coefficient group Fk . All
its edges are separating edges. In particular, they are oriented and with
trivial stabilizers. The fundamental group of this graph is Lf = Ff ∗ F .
(4) every restricted pair homomorphism, h : (S, G) → (F Sk , Fk ), factors through
at least one of these finitely many resolutions.
(5) part (2) in the statement of theorem 7.9 holds for the restricted diagram,
where sequences of homomorphisms and their shortenings and modifications
are replaced by sequences of restricted homomorphisms.
Proof:
Identical to the proof of theorem 7.17.
References
[Be-Fe1]
M. Bestvina and M. Feighn, Stable actions of groups on real trees, Inventiones Math.
121 (1995), 287-321.
[Be-Fe2]
, Bounding the complexity of simplicial group actions, Inventiones Math. 103
(1991), 449-469.
[De]
T. Delzant, Sur l’accessibilite acylindrique des groupes de presentation finie, Annals
institute Fourier (Grenoble) 49 (1999), 1215-1224.
[Di]
V. Diekert, More than 1700 years of word equations, LNCS proceedings of 2015 CAI
Stuttgart, Germany (2015), 1-4.
135
[DJP]
V. Diekert, A. Jez and W. Plandowski, Finding all solutions to equations in free
groups and monoids with involution, preprint.
[Du-Sa]
M. Dunwoody and M. Sageev, JSJ-splittings for finitely presented groups over slender
groups, Inventiones Math. 135 (1999), 25-44.
[Du1]
V. G. Durnev, On the positive theory of free semigroup, Dokl. Akad. Nauk. SSSR 211
(1973), 772-774.
[Du2]
, Undecidability of the positive ∀∃3 theory of a free semigroup, Siberian Math.
Journal 36 (1995), 917-929.
[Fu-Pa]
K. Fujiwara and P. Papasoglu, JSJ decompositions of finitely presented groups and
complexes of groups, GAFA 16 (2006), 70-125.
[Gu]
V. Guirardel, Actions of finitely generated groups on R-trees, Annals Inst. Fourier
(Grenoble) 58 (2008), 159-211.
[Je]
A. Jez, Recompression: a simple and powerful technique for word equations, STACS
vol. 20 of LIPIcs, N. Portier and T. Wilke, editors, 233-244.
[Kh-My]
O. Kharlampovich and A. Myasnikov, Irreducible affine varieties over a free group II,
Jour. of Algebra 200 (1998), 517-570.
[Khm]
W. V. O. Khmelevskii, Equations in a free semigroup, Trudy Math. institute Steklov
(LOMI) 107 (1971), 21-28.
[Ly]
I. G. Lysenok, Quadratic equations in free monoids with involution and surface train
tracks, preprint.
[Ma1]
G. S. Makanin, The problem of solvability of equations in a free semigroup, Math.
Sbornik 103 (1977), 147-236.
[Ma2]
, Equations in a free group, Math. USSR Izvestiya 21 (1983), 449-469.
[Ma3]
On general solution of equations in a free semigroup, Word equations and
related topics (Tubingen 1990), Lecture notes in computer science 572, Springer 1992,
pp. 1-11.
[Ma4]
, Finite parametrization of solutions of equations in a free monoid I, Math.
Sbornik 195 (2004), 41-90.
[Ma5]
, Parametrization of solutions of the equation x1 x2 . . . xn = xn xn−1 . . . x1 in
a free monoid, Mat. Zametki 89 (2011), 839-844.
[Mar]
S. S. Marchenkov, Undecidability of the positive ∀∃ theory of a free semigroup, Sibirsk
Math. Zh. 23 (1982), 196-198.
136
[Me]
Yu. I. Merzlyakov, Positive formulae on free groups, Algebra i Logika 5 (1966), 257266.
[Mo]
J. W. Morgan, Ergodic theory and free actions of groups on R-trees, Inventiones Math.
94 (1988), 605-622.
[Qu]
W. V. O. Quine, Concatenation as a basis for arithmetic, Journal of symbolic logic
11 (1946), 105-114.
[Ra]
A. A. Razborov, On systems of equations in a free group, Ph.D. thesis, Steklov Math.
institute, 1987.
[Re]
P. Reynolds, On indecomposable trees in the boundary of outer space, Geom. Dedicata
153 (2011), 59-71.
[Ri-Se2]
E. Rips and Z. Sela, Cyclic splittings of finitely presented groups and the canonical
JSJ decomposition, Annals of Mathematics 146 (1997), 53-104.
[Se1]
Z. Sela, Diophantine geometry over groups I: Makanin-Razborov diagrams, Publications Mathematique de l’IHES 93 (2001), 31-105.
[Se2]
, Diophantine geometry over groups II: Completions, closures and formal solutions, Israel jour. of Math. 134 (2003), 173-254.
[Se3]
, Diophantine geometry over groups V1 : Quantifier elimination I, Israel jour.
of Mathematics 150 (2005), 1-197.
[Se4]
, Diophantine geometry over groups V2 : Quantifier elimination II, GAFA 16
(2006), 537-706.
[Se5]
, Diophantine geometry over groups VI: The elementary theory of a free group,
GAFA 16 (2006), 707-730.
[Se6]
, Acylindrical accessibility for groups, Inventiones Mathematicae 129 (1997),
527-565.
[Se7]
, Structure and rigidity in (Gromov) hyperbolic groups and discrete groups in
rank 1 Lie Groups II, GAFA 7 (1997), 561-593.
[Sk]
R. Skora, Splittings of surfaces, Jour. of the AMS 9 (1996), 605-616.
[We]
R. Weidmann, The Nielsen method for groups acting on trees, Proc. of London Math.
Society 85 (2002), 93-118.
137
| 4math.GR
|
Algorithms for the Majority Rule (+) Consensus
Tree and the Frequency Difference Consensus
Tree
Jesper Jansson1 , Chuanqi Shen2 , and Wing-Kin Sung3,4
arXiv:1307.7821v2 [cs.DS] 6 Aug 2013
1
Laboratory of Mathematical Bioinformatics (Akutsu Laboratory),
Institute for Chemical Research,
Kyoto University, Gokasho, Uji, Kyoto 611-0011, Japan.
E-mail: [email protected]
Funded by The Hakubi Project and KAKENHI grant number 23700011.
2
Stanford University, 450 Serra Mall, Stanford, CA 94305-2004, U.S.A.
E-mail: [email protected]
3
School of Computing, National University of Singapore, 13 Computing Drive,
Singapore 117417.
E-mail: [email protected]
4
Genome Institute of Singapore, 60 Biopolis Street, Genome, Singapore 138672.
Abstract. This paper presents two new deterministic algorithms for
constructing consensus trees. Given an input of k phylogenetic trees with
identical leaf label sets and n leaves each, the first algorithm constructs
the majority rule (+) consensus tree in O(kn) time, which is optimal
since the input size is Ω(kn), and the second one constructs the frequency
difference consensus tree in min{O(kn2 ), O(kn(k + log2 n))} time.
1
Introduction
A consensus tree is a phylogenetic tree that summarizes a given collection of
phylogenetic trees having the same leaf labels but different branching structures. Consensus trees are used to resolve structural differences between two or
more existing phylogenetic trees arising from conflicts in the raw data, to find
strongly supported groupings, and to summarize large sets of candidate trees
obtained by bootstrapping when trying to infer a new phylogenetic tree accurately [2,10,12,27].
Since the first type of consensus tree was proposed by Adams III [1] in 1972,
many others have been defined and analyzed. See, e.g., [5], Chapter 30 in [12],
or Chapter 8.4 in [27] for some surveys. Which particular type of consensus
tree to use in practice depends on the context. For example, the strict consensus
tree [25] is very intuitive and easy to compute [9] and may be sufficient when there
is not so much disagreement in the data, the majority rule consensus tree [21]
is “the optimal tree to report if we view the cost of reporting an estimate of
the phylogeny to be a linear function of the number of incorrect clades in the
estimate and the number of true clades that are missing from the estimate and we
T1 :
T2 :
T3 :
e
e
d
a
d
b c d
c
a
T4 :
c
a
e
a
b
b
b
c
Majority rule:
Majority rule (+):
Frequency diff.:
e
e
c d e
a
e
b
c
a
d
d
b
a
b c
d
Fig. 1. Let S = {T1 , T2 , T3 , T4 } as shown above with L = Λ(T1 ) = Λ(T2 ) = Λ(T3 ) =
Λ(T4 ) = {a, b, c, d, e}. The only non-trivial majority cluster of S is {a, b}, the nontrivial majority (+) clusters of S are {a, b} and {a, b, c, d}, and the non-trivial frequency
difference clusters of S are {a, b}, {a, b, c, d}, and {c, d}. The majority rule, majority
rule (+), and frequency difference consensus trees of S are displayed.
view the reporting of an incorrect grouping as a more serious error than missing
a clade” [16], and the R* consensus tree [5] provides a statistically consistent
estimator of the species tree topology when combining gene trees [10]. Therefore,
scientists need efficient algorithms for constructing a broad range of different
consensus trees.
In a recent series of papers [8,17,18,19], we have developed fast algorithms
for computing the majority rule consensus tree [21], the loose consensus tree [4]
(also known in the literature as the combinable component consensus tree or
the semi-strict consensus tree), a greedy consensus tree [5,13], the R* consensus
tree [5], and consensus trees for so-called multi-labeled phylogenetic trees (MULtrees) [20]. In this paper, we study two relatively new types of consensus trees
called the majority rule (+) consensus tree [7,11] and the frequency difference
consensus tree [14], and give algorithms for constructing them efficiently.
1.1
Definitions and notation
We shall use the following basic definitions. A phylogenetic tree is a rooted,
unordered, leaf-labeled tree in which every internal node has at least two children
and all leaves have different labels. (Below, phylogenetic trees are referred to as
“trees” for short). For any tree T , the set of all nodes in T is denoted by V (T )
and the set of all leaf labels in T by Λ(T ). Any nonempty subset C of Λ(T ) is
called a cluster of Λ(T ); if |C| = 1 or C = Λ(T ) then C is trivial, and otherwise,
C is non-trivial. For any u ∈ V (T ), T [u] denotes the subtree of T rooted at the
node u. Observe that if uS
is the root of T or if u is a leaf then Λ(T [u]) is a trivial
cluster. The set C(T ) = u∈V (T ) {Λ(T [u])} is called the cluster collection of T ,
and any cluster C ⊆ Λ(T ) is said to occur in T if C ∈ C(T ).
Two clusters C1 , C2 ⊆ Λ(T ) are compatible if C1 ⊆ C2 , C2 ⊆ C1 , or C1 ∩C2 =
∅. If C1 and C2 are compatible, we write C1 ^ C2 ; otherwise, C1 6^ C2 . A cluster
C ⊆ Λ(T ) is compatible with T if C ^ Λ(T [u]) holds for every node u ∈ V (T ).
In this case, we write C ^ T , and C 6^ T otherwise. If T1 and T2 are two trees
with Λ(T1 ) = Λ(T2 ) such that every cluster in C(T1 ) is compatible with T2 then
it follows that every cluster in C(T2 ) is compatible with T1 , and we say that T1
and T2 are compatible. Any two clusters or trees that are not compatible are
called incompatible.
Next, let S = {T1 , T2 , . . . , Tk } be a set of trees satisfying Λ(T1 ) = Λ(T2 ) =
· · · = Λ(Tk ) = L for some leaf label set L. For any cluster C of L, denote the
set of all trees in S in which C occurs by KC (S) and the set of all trees in S
that are incompatible with C by QC (S). Thus, KC (S) = {Ti : C ∈ C(Ti )} and
QC (S) = {Ti : C 6^ Ti }. Define three special types of clusters:
• If |KC (S)| >
k
2
then C is a majority cluster of S.
• If |KC (S)| > |QC (S)| then C is a majority (+) cluster of S.
• If |KC (S)| > max{|KD (S)| : D ⊆ L and C 6^ D} then C is a frequency
difference cluster of S.
(Informally, a frequency difference cluster is a cluster that occurs more frequently
than each of the clusters that is incompatible with it.) Note that a majority
cluster of S is always a majority (+) cluster of S and that a majority (+) cluster
of S is always a frequency difference cluster of S, but not the other way around.
The majority rule consensus tree of S [21] is the tree T such that Λ(T ) = L
and C(T ) consists of all majority clusters of S. Similarly, the majority rule (+)
consensus tree of S [7,11] is the tree T such that Λ(T ) = L and C(T ) consists
of all majority (+) clusters of S, and the frequency difference consensus tree
of S [14] is the tree T such that Λ(T ) = L and C(T ) consists of all frequency
difference clusters of S. See Fig. 1 for some examples.
From here on, S is assumed to be an input set of identically leaf-labeled trees,
and the leaf label set of S is denoted by L. To express the size of the input, we
define k = |S| and n = |L|.
1.2
Previous work
Margush and McMorris [21] introduced the majority rule consensus tree in 1981,
and a deterministic algorithm for constructing it in optimal O(kn) worst-case
running time was presented recently in [18]. (A randomized algorithm with
O(kn) expected running time and unbounded worst-case running time was given
earlier by Amenta et al. [2].) The majority rule consensus tree has several desirable mathematical properties [3,16,22], and algorithms for constructing it
have been implemented in popular computational phylogenetics packages like
PHYLIP [13], TNT [15], COMPONENT [23], MrBayes [24], SumTrees in DendroPy [26], and PAUP* [28]. Consequently, it is one of the most widely used consensus trees in practice [7, p. 450]. One drawback of the majority rule consensus
tree is that it may be too harsh and discard valuable branching information. For
example, in Fig. 1, even though the cluster {a, b, c, d} is compatible with 75% of
the input trees, it is not included in the majority rule consensus tree. For this
reason, people have become interested in alternative types of consensus trees
that include all the majority clusters and at the same time, also include other
meaningful, well-defined kinds of clusters. The majority rule (+) consensus tree
and the frequency difference consensus tree are two such consensus trees.
The majority rule (+) consensus tree was defined by Dong et al. [11] in 2010.
It was obtained as a special case of an attempted generalization by Cotton and
Wilkinson [7] of the majority rule consensus tree. According to [11], Cotton and
Wilkinson [7] suggested two types of supertrees5 called majority-rule (-) and
majority-rule (+) that were supposed to generalize the majority rule consensus
tree. Unexpectedly, only the first one did, and by restricting the second one to the
consensus tree case, [11] arrived at the majority rule (+) consensus tree. Dong
et al. [11] established some fundamental properties of this type of consensus tree
and pointed out the existence of a polynomial-time algorithm for constructing it,
but left the task of finding the best possible such algorithm as an open problem.
As far as we know, no implementation for computing the majority rule (+)
consensus tree is publicly available.
Goloboff et al. [14] initially proposed the frequency difference consensus tree
as a way to improve methods for evaluating group support in parsimony analysis.
Its relationships to other consensus trees have been studied in [11]. A method
for constructing it has been implemented in the free software package TNT [15]
but the algorithm used is not documented and its time complexity is unknown.
We note that since the number of clusters occurring in S may be Ω(kn), a naive
algorithm that compares every cluster in S to every other cluster in S directly
would require Ω(k 2 n2 ) time.
1.3
Organization of the paper and new results
Due to space constraints, some proofs have been omitted from the conference
version of this paper. Please see the journal version for the complete proofs.
The paper is organized as follows. Section 2 summarizes some results from the
literature that are needed later. In Section 3, we modify the techniques from [18]
to obtain an O(kn)-time algorithm for the majority rule (+) consensus tree. Its
running time is optimal because the size of the input is Ω(kn); hence, we resolve
the open problem of Dong et al. [11] mentioned above. Next, Section 4 gives a
min{O(kn2 ), O(kn(k + log2 n))}-time algorithm for constructing the frequency
difference consensus tree (here, the second term is smaller than the first term
5
A supertree is a generalization of a consensus tree that does not require the input
trees to have identical leaf label sets.
if k = o(n); e.g., if k = O(1) then the running time reduces to O(n log2 n)).
Our algorithms are fully deterministic and do not need to use hashing. Finally,
Section 5 discusses implementations.
2
2.1
Preliminaries
The delete and insert operations
The delete and insert operations are two operations that modify the structure
of a tree. They are defined in the following way.
Let T be a tree and let u be any non-root, internal node in T . The delete
operation on u makes all of u’s children become children of the parent of u,
and then removes u and the edge between u and its parent. (See, e.g., Figure 2
in [17] for an illustration.) The time needed for this operation is proportional
to the number of children of u, and the effect of applying it is that the cluster
collection of T is changed to C(T ) \ {Λ(T [u])}. Conversely, the insert operation
creates a new node u that becomes: (1) a child of an existing internal node v, and
(2) the parent of a proper subset X of v’s children satisfying |X|
S ≥ 2; the effect
is that C(T ) is changed to C(T ) ∪ {Λ(T [u])}, where Λ(T [u]) = vi ∈X Λ(T [vi ]).
2.2
Subroutines
The new algorithms in this paper use the following algorithms from the literature as subroutines: Day’s algorithm [9], Procedure One-Way Compatible [17],
and Procedure Merge Trees [17]. Day’s algorithm [9] is used to efficiently check
whether any specified cluster that occurs in a tree T also occurs in another
tree Tref , and can be applied to find the set of all clusters that occur in both T
and Tref in linear time. Procedure One-Way Compatible takes as input two
trees TA and TB with identical leaf label sets and outputs a copy of TA in which
every cluster that is not compatible with TB has been removed. (The procedure is
asymmetric; e.g., if TA consists of n leaves attached to a root node and TB 6= TA
then One-Way Compatible(TA , TB ) = TA , while One-Way Compatible(TB , TA )
= TB .) Procedure Merge Trees takes as input two compatible trees with identical leaf label sets and outputs a tree that combines their cluster collections.
Their properties are summarized below; for details, see references [9] and [17].
Lemma 1. (Day [9]) Let Tref and T be two given trees with Λ(Tref ) = Λ(T ) = L
and let n = |L|. After O(n) time preprocessing, it is possible to determine, for
any u ∈ V (T ), if Λ(T [u]) ∈ C(Tref ) in O(1) time.
Lemma 2. ([17]) Let TA and TB be two given trees with Λ(TA ) = Λ(TB ) = L
and let n = |L|. Procedure One-Way Compatible(TA , TB ) returns a tree T with
Λ(T ) = L such that C(T ) = {C ∈ C(TA ) : C is compatible with TB } in O(n)
time.
Lemma 3. ([17]) Let TA and TB be two given trees with Λ(TA ) = Λ(TB ) = L
that are compatible and let n = |L|. Procedure Merge Trees(TA , TB ) returns a
tree T with Λ(T ) = L and C(T ) = C(TA ) ∪ C(TB ) in O(n) time.
3
Constructing the majority rule (+) consensus tree
This section presents an algorithm named Maj Rule Plus for computing the
majority rule (+) consensus tree of S in (optimal) O(kn) time.
The pseudocode of Maj Rule Plus is given in Fig. 2. The algorithm has
two phases. Phase 1 examines the input trees, one by one, to construct a set of
candidate clusters that includes all majority (+) clusters. Then, Phase 2 removes
all candidate clusters that are not majority (+) clusters.6
During Phase 1, the current candidate clusters are stored as nodes in a tree T .
Every node v in T represents a current candidate cluster Λ(T [v]) and has a
counter count(v) that, starting from the iteration at which Λ(T [v]) became a
candidate cluster, keeps track of the number of input trees in which it occurs
minus the number of input trees that are incompatible with it. More precisely,
while treating the tree Tj for any j ∈ {2, 3, . . . , k} in Step 3.1, count(v) for
each current candidate cluster Λ(T [v]) is updated as follows: if Λ(T [v]) occurs
in Tj then count(v) is incremented by 1, if Λ(T [v]) does not occur in Tj and is
not compatible with Tj then count(v) is decremented by 1, and otherwise (i.e.,
Λ(T [v]) does not occur in Tj but is compatible with Tj ) count(v) is unchanged.
Furthermore, if any count(v) reaches 0 then the node v is deleted from T so
that Λ(T [v]) is no longer a current candidate cluster. Next, in Step 3.3, every
cluster occurring in Tj that is not a current candidate but compatible with T is
inserted into T (thus becoming a current candidate cluster) and its counter is
initialized to 1. Lemma 4 below proves that the set of majority (+) clusters of S
is contained in the set of candidate clusters at the end of Phase 1.
Lemma 4. For any C ⊆ L, if C is a majority (+) cluster of S then C ∈ C(T )
at the end of Phase 1.
Proof. Suppose that C is a majority (+) cluster of S. Let Tx be any tree in QC (S)
and consider iteration x in Step 3: If C is a current candidate at the beginning of
iteration x then its counter will be decremented, cancelling out the occurrence
of C in one tree Tj where 1 ≤ j < x; otherwise, C may be prevented from being
inserted into T in at most one later iteration j (where x < j ≤ k and C ∈ C(Tj ))
because of some cluster occurring in Tx . It follows from |KC (S)| − |QC (S)| > 0
that C’s counter will be greater than 0 at the end of Phase 1, and therefore
C ∈ C(T ).
t
u
In Phase 2, Step 5 of the algorithm computes the values of |KC (S)| and
|QC (S)| for every candidate cluster C and stores them in K (v) and Q(v), respectively, where C = Λ(T [v]). Finally, Step 6 removes every candidate cluster C
that does not satisfy the condition |KC (S)| > |QC (S)|. By definition, the clusters
that remain in T are the majority (+) clusters.
Theorem 1. Algorithm Maj Rule Plus constructs the majority rule (+) consensus tree of S in O(kn) time.
6
This basic strategy was previously used in the O(kn)-time algorithm in [18] for
computing the majority rule consensus tree.
Algorithm
Input:
Maj Rule Plus
A set S = {T1 , T2 , . . . , Tk } of trees with Λ(T1 ) = Λ(T2 ) = · · · = Λ(Tk ).
Output: The majority rule (+) consensus tree of S.
/* Phase 1 */
1 T := T1
2 for each v ∈ V (T ) do count(v) := 1
3 for j := 2 to k do
3.1
for each v ∈ V (T ) do
if Λ(T [v]) occurs in Tj then count(v) := count(v) + 1
else if Λ(T [v]) is not compatible with Tj then count(v) := count(v)−1
endfor
3.2
for each v ∈ V (T ) in top-down order do
if count(v) = 0 then delete node v.
3.3
for every C ∈ C(Tj ) that is compatible with T but does not occur in T do
Insert C into T .
Initialize count(v) := 1 for the new node v satisfying Λ(T [v]) = C.
endfor
endfor
/* Phase 2 */
4 for each v ∈ V (T ) do K (v) := 0; Q(v) := 0
5 for j := 1 to k do
5.1
for each v ∈ V (T ) do
if Λ(T [v]) occurs in Tj then K (v) := K (v) + 1
else if Λ(T [v]) is not compatible with Tj then Q(v) := Q(v) + 1
endfor
6 for each v ∈ V (T ) in top-down order do
if K (v) ≤ Q(v) then perform a delete operation on v.
7 return T
End Maj Rule Plus
Fig. 2. Algorithm Maj Rule Plus for constructing the majority rule (+) consensus tree.
Proof. The correctness follows from Lemma 4 and the above discussion.
The time complexity analysis is analogous to the proof of Theorem 4 in [18].
First consider Phase 1. Step 3.1 takes O(n) time by: (1) running Day’s algorithm
with Tref = Tj and then checking each node v in V (T ) to see if Λ(T [v]) occurs
in Tj (according to Lemma 1, this requires O(n) time for preprocessing, and each
of the O(n) nodes in V (T ) may be checked in O(1) time), and (2) computing
X := One-Way Compatible(T, Tj ) and then checking for each node v in V (T )
if v does not exist in X to determine if Λ(T [v]) 6^ Tj (this takes O(n) time
by Lemma 2). The delete operations in Step 3.2 take O(n) time because the
nodes are handled in top-down order, which means that for every node, its
parent will change at most once in each iteration. In Step 3.3, define Y :=
One-Way Compatible(Tj , T ) and Z := Merge Trees(Y, T ). Then by Lemmas 2
and 3, the cluster collection of Y consists of the clusters occurring in Tj that are
compatible with the set of current candidates, and Z is the result of inserting
these clusters into T . Thus, Step 3.3 can be implemented by computing Y and Z,
updating T ’s structure according to Z, and setting the counters of all new nodes
to 1, so Step 3.3 takes O(n) time. The main loop in Step 3 consists of O(k)
iterations, and Phase 1 therefore takes O(kn) time in total.
Next, Phase 2 also takes O(kn) time because Step 5.1 can be implemented in
O(n) time with the same techniques as in Step 3.1, and Step 6 is performed in
O(n) time by handling the nodes in top-down order so that each node’s parent
is changed at most once, as in Step 3.2.
t
u
4
Constructing the frequency difference consensus tree
Here, we present an algorithm for finding the frequency consensus tree of S in
min{O(kn2 ), O(kn(k + log2 n))} time. It is called Frequency Difference and is
described in Section 4.1 below. The algorithm uses the procedure Merge Trees
as well as a new procedure named Filter Clusters whose details are given in
Section 4.2.
For each tree Tj ∈ S and each node u ∈ V (Tj ), define the weight of u as the
value |KΛ(Tj [u]) (S)|, i.e., the number of trees from S in which the cluster Λ(Tj [u])
occurs, and denote it by w(u). For convenience, also define w(C) = w(u), where
C = Λ(Tj [u]). The input to Procedure Filter Clusters is two trees TA , TB with
Λ(TA ) = Λ(TB ) = L such that every cluster occurring in TA or TB also occurs
in at least one tree in S, and the output is a copy of TA in which every cluster
that is incompatible with some cluster in TB with a higher weight has been
removed. Formally, the output of Filter Clusters is a tree T with Λ(T ) =
L such that C(T ) = {Λ(TA [u]) : u ∈ V (TA ) and w(u) > w(x) for every x ∈
V (TB ) with Λ(TA [u]) 6^ Λ(TB [x])}.
4.1
Algorithm Frequency Difference
We first describe Algorithm Frequency Difference. Refer to Fig. 3 for the
pseudocode.
The algorithm starts by computing the weight w(C) of every cluster C occurring in S in a preprocessing
step (Step 1). Next, let C(S) for any set S
S
of trees denote the union Ti ∈S C(Ti ), and for any j ∈ {1, 2, . . . , k}, define a
forward frequency difference consensus tree of {T1 , T2 , . . . , Tj } as any tree that
includes every cluster C in C({T1 , T2 , . . . , Tj }) satisfying w(C) > w(X) for all
X ∈ C({T1 , T2 , . . . , Tj }) with C 6^ X. Steps 2–3 use Procedure Filter Clusters
from Section 4.2 to build a tree T that, after any iteration j ∈ {1, 2, . . . , k}, is
a forward frequency difference consensus tree of {T1 , T2 , . . . , Tj }, as proved in
Lemma 5 below. After iteration k, C(T ) contains all frequency difference clusters
of S but possibly some other clusters as well, so Step 4 applies Filter Clusters
again to remove all non-frequency difference clusters of S from T .
Algorithm
Input:
Frequency Difference
A set S = {T1 , T2 , . . . , Tk } of trees with Λ(T1 ) = Λ(T2 ) = · · · = Λ(Tk ).
Output: The frequency difference consensus tree of S.
/* Preprocessing */
1 Compute w(C) for every cluster C occurring in S.
/* Main algorithm */
2 T := T1
3 for j := 2 to k do
A := Filter Clusters(T, Tj ); B := Filter Clusters(Tj , T )
T := Merge Trees(A, B)
endfor
4 for j := 1 to k do T := Filter Clusters(T, Tj )
5 return T
End Frequency Difference
Fig. 3. Algorithm Frequency Difference for constructing the frequency difference consensus tree.
Lemma 5. For any j ∈ {2, 3, . . . , k}, suppose that T is a forward frequency difference consensus tree of {T1 , T2 , . . . , Tj−1 }. Let A := Filter Clusters(T, Tj )
and B := Filter Clusters(Tj , T ). Then Merge Trees(A, B) is a forward frequency difference consensus tree of {T1 , T2 , . . . , Tj }.
Proof. (Omitted from the conference version due to space constraints.)
t
u
Theorem 2. Algorithm Frequency Difference constructs the frequency difference consensus tree of S in min{O(kn2 ), O(k 2 n)} + O(k · f (n)) time, where
f (n) is the running time of Procedure Filter Clusters.
Proof. After completing iteration k of Step 3, C(T ) is a superset of the set of all
frequency difference clusters of S by Lemma 5. Next, Step 4 removes all nonfrequency difference clusters of S, so the output will be the frequency difference
consensus tree of S.
To analyze the time complexity, first consider how to compute all the weights
in Step 1. One method is to first fix an arbitrary ordering of L and represent every
cluster C of L as a bit vector of length n (for every i ∈ {1, 2, . . . , n}, the ith bit is
set to 1 if and only if the ith leaf label belongs to C). Then, spend O(kn2 ) time to
construct a list of bit vectors for all O(kn) clusters occurring in S by a bottom-up
traversal of each tree in S, sort the resulting list of bit vectors by radix sort, and
traverse the sorted list to identify the number of occurrences of each cluster. All
this takes O(kn2 ) time. An alternative method, which uses O(k 2 n) time, is to
initialize the weight of every node in S to 1 and then, for j ∈ {1, 2, . . . , k}, apply
Day’s algorithm (see Lemma 1) with Tref = Tj and T ranging over all Ti with
1 ≤ i ≤ k, i 6= j to find all clusters in T that also occur in Tj and increase the
weights of their nodes in T by 1. Therefore, Step 1 takes min{O(kn2 ), O(k 2 n)}
time. Next, Steps 3 and 4 make O(k) calls to the procedures Merge Trees and
Filter Clusters. The running time of Merge Trees is O(n) by Lemma 3 and
the running time of Filter Clusters is f (n) = Ω(n), so Steps 3 and 4 take
O(k · f (n)) time.
t
u
Lemma 7 in the next subsection shows that f (n) = O(n log2 n) is possible,
which yields:
Corollary 1. Algorithm Frequency Difference constructs the frequency difference consensus tree of S in min{O(kn2 ), O(kn(k + log2 n))} time.
4.2
Procedure Filter Clusters
Recall that for any node u in any input tree Tj , its weight w(u) is |KΛ(Tj [u]) (S)|.
Also, w(C) = w(u), where C = Λ(Tj [u]). We assume that all w(u)-values have
been computed in a preprocessing step and are available.
Let T be a tree. For every nonempty X ⊆ V (T ), lca T (X) denotes the lowest
common ancestor of X in T . To obtain a fast solution for Filter Clusters, we
need the next lemma.
Lemma 6. Let T be a tree, let X be any cluster of Λ(T ), and let rX = lca T (X).
For any v ∈ V (T ), it holds that X 6^ Λ(T [v]) if and only if: (1) v lies on a path
from a child of rX to some leaf belonging to X; and (2) Λ(T [v]) 6⊆ X.
Proof. Given T , X, rX , and v as in the lemma statement, there are four possible
cases: (i) v is a proper ancestor of rX or equal to rX ; (ii) v lies on a path from
a child of rX to some leaf in X and all leaf descendants of v belong to X; (iii) v
lies on a path from a child of rX to some leaf in X and not all leaf descendants
of v belong to X; or (iv) v is a proper descendant of rX that does not lie on any
path from a leaf in X to rX . In case (i), X ⊆ Λ(T [v]). In case (ii), Λ(T [v]) ⊆ X.
In case (iii), Λ(T [v]) 6⊆ X while X ∩ Λ(T [v]) 6= ∅. In case (iv), X ∩ Λ(T [v]) = ∅.
By the definition of compatible clusters, X 6^ Λ(T [v]) if and only if case (iii)
occurs.
t
u
Lemma 6 leads to an O(n2 )-time method for Filter Clusters, which we now
briefly describe. For each node u ∈ V (TA ) in top-down order, do the following:
Let X = Λ(TA [u]) and find all v ∈ V (TB ) such that X 6^ Λ(TB [v]) in O(n) time
by doing bottom-up traversals of TB to first mark all ancestors of leaves belonging
to X that are proper descendants of the lowest common ancestor of X in TB ,
and then unmarking all marked nodes that have no leaf descendants outside
of X. By Lemma 6, X 6^ Λ(TB [v]) if and only if v is one of the resulting marked
nodes. If w(u) ≤ w(v) for any such v then do a delete operation on u in TA .
Clearly, the total running time is O(n2 ). (This simple method gives f (n) = O(n2 )
in Theorem 2 in Section 4.1, and hence a total running time of O(kn2 ) for
Algorithm Frequency Difference.) Below, we refine this idea to get an even
faster solution for Filter Clusters.
High-level description: We use the centroid path decomposition technique [6]
to divide the nodes of TA into a so-called centroid path and a set of side trees. A
centroid path of TA is defined as a path in TA of the form π = hpα , pα−1 , . . . , p1 i,
where pα is the root of TA , the node pi−1 for every i ∈ {2, . . . , α} is any child
of pi with the maximum number of leaf descendants, and p1 is a leaf. Given
a centroid path π, removing π and all its incident edges from TA produces a
set σ(π) of disjoint trees whose root nodes are children of nodes belonging to π
in TA ; these trees are called the side trees of π. Importantly, |Λ(τ )| ≤ n/2 for
every side tree τ of π. Also, {Λ(τ ) : τ ∈ σ(π)} forms a partition of L \ {p1 }.
Furthermore, if π is a centroid path
S of TA then the
S cluster collection C(TA ) can
be written recursively as C(TA ) = τ ∈σ(π) C(τ ) ∪ pi ∈π {Λ(TA [pi ])}. Intuitively,
this allows the cluster collection of TA to be broken into smaller sets that can
be checked more easily, and then put together again at the end.
The fast version of Filter Clusters is shown in Fig. 4. It first computes a
centroid path π = hpα , pα−1 , . . . , p1 i of TA and the set σ(π) of side trees of π
in Step 1. Then, in Steps 2–3,
S it applies itself recursively to each side tree of π
to get rid of any cluster in τ ∈σ(π) C(τ ) that is incompatible with some cluster
in TB with a higher weight than itself, and the remaining clusters
are inserted
S
into a temporary tree Rs . Next, Steps 4–5 check all clusters in pi ∈π {Λ(TA [pi ])}
to determine which of them are not incompatible with any cluster in TB with a
higher weight, and create a temporary tree Rc whose cluster collection consists
of all those clusters that pass this test. Finally, Step 6 combines the cluster
collections of Rs and Rc by applying the procedure Merge Trees. The details of
Procedure Filter Clusters are discussed next.
Steps 2–3 (handling the side trees): For every nonempty C ⊆ Λ(T ), define
T |C (“the subtree of T induced by C”; see, e.g., [6]) as the tree T 0 with leaf label
set C and internal node set {lca T ({a, b}) : a, b ∈ C} which preserves the ancestor
0
relations from T , i.e., which satisfies lca T (C 0 ) = lca T (C 0 ) for all nonempty
C 0 ⊆ C. Now, let σ(π) be the set of side trees of the centroid path π computed
in Step 1. For each τ ∈ σ(π), define a weighted tree TB ||Λ(τ ) as follows. First,
construct TB |Λ(τ ) and let the weight of each node in this tree equal its weight
in TB . Next, for each edge (u, v) in TB |Λ(τ ), let P be the path in TB between u
and v, excluding u and v; if P is not empty then create a new node z in TB |Λ(τ ),
replace the edge (u, v) by the two edges (u, z) and (z, v), and set the weight of z
to the maximum weight of all nodes belonging to P . Each such z is called a
special node and has exactly one child. See Fig. 5 for an example. We extend
the concept of “compatible” to special nodes as follows: if C ⊆ L and z is a
special node in TB ||Λ(τ ) then C ^ z if and only if C and Λ((TB ||Λ(τ ))[z]) are
disjoint or (TB ||Λ(τ ))[z] has no proper descendant that is a special node. The
obtained tree TB ||Λ(τ ) satisfies Λ(τ ) = Λ(TB ||Λ(τ )) and has the property that
for every cluster C in C(τ ), max{w(X) : X ∈ C(TB ) and C 6^ X} is equal to
max{w(X) : X ∈ C(TB ||Λ(τ )) and C 6^ X}.
After constructing TB ||Λ(τ ), Filter Clusters is applied to (τ, TB ||Λ(τ ))
recursively to remove all bad clusters from τ . For each τ ∈ σ(π), the resulting
tree is denoted by τ 0 . All the clusters of τ 0 are inserted into the tree Rs by di-
Algorithm Filter Clusters
Input: Two trees TA , TB with Λ(TA ) = Λ(TB ) = L such that every cluster
occurring in TA or TB also occurs in at least one tree in S.
Output: A tree T with Λ(T ) = L such that C(T ) = {Λ(TA [u]) : u ∈ V (TA ) and
w(u) > w(x) for every x ∈ V (TB ) with Λ(TA [u]) 6^ Λ(TB [x])}.
1 Compute a centroid path π = hpα , pα−1 , . . . , p1 i of TA , where pα is the root
of TA and p1 is a leaf, and compute the set σ(π) of side trees of π.
/* Handle the side trees. */
2 Let Rs be a tree consisting only of a root node and a single leaf labeled by p1 .
3 for each side tree τ ∈ σ(π) do
τ 0 := Filter Clusters(τ, TB ||Λ(τ ))
Attach the root of τ 0 to the root of Rs .
endfor
/* Handle the centroid path. */
4 Let Rc be a tree with Λ(Rc ) = L where every leaf is directly attached to the
root. Let BT be an empty binary search tree. For every x ∈ V (TB ), initialize
counter (x) := 0. Do a bottom-up traversal of TB to precompute |Λ(TB [x])| for
every x ∈ V (TB ). Preprocess TB for answering lca-queries. Let β1 := 0.
5 for i := 2 to α do
5.1
Let D be the set of leaves in Λ(TA [pi ]) \ Λ(TA [pi−1 ]).
5.2
Compute ri := lca TB ({ri−1 } ∪ D). /* ri now equals lca TB (Λ(TA [pi ]). */
5.3
Insert every node belonging to the path from ri to ri−1 , except ri , into BT .
5.4
for each x ∈ D do
Insert x into BT .
while (parent(x) is not in BT and parent(x) 6= ri ) do
x := parent(x ); insert x into BT .
endfor
5.5
for each x ∈ D do
counter (x) := counter (x) + 1
while (counter (x) = |Λ(TB [x])|) do
counter (parent(x)) := counter (parent(x)) + |Λ(TB [x])|
Remove x from BT ; x := parent(x)
endwhile
endfor
5.6
Let M := maximum weight of a node in BT ; if BT is empty then M := 0.
5.7
Compute βi , and if βi > M then let M := βi .
5.8
if (w(Λ(TA [pi ])) > M ) then put Λ(TA [pi ]) in Rc by an insert operation.
endfor
/* Combine the surviving clusters. */
6 T := Merge Trees(Rs , Rc )
7 return T
End Filter Clusters
Fig. 4. The procedure Filter Clusters.
9
9
2
5
8
7
9
i
9
a
9
5
3
8
9
7
9
j
4
a
9
g
b
9
9
c
d
6
9
9
h
4
9 9
9
e f g h
Fig. 5. Let TB be the tree (with node weights) on the left. The tree TB ||{a, g, h} is
shown on the right.
rectly attaching τ 0 to the root of Rs . Since {Λ(τ 0 ) : τ ∈ σ(π)} forms a partition
of L \ {p1 }, every leaf label in L appears exactly once in Rs and we have C(Rs ) =
{Λ(TA [u]) : u ∈ V (τ ) for some τ ∈ σ(π) and w(u) > w(x) for every x ∈
V (TB ) with Λ(TA [u]) 6^ Λ(TB [x])} ∪ {L} after Step 3 is finished.
S
Steps 4–5 (handling the centroid path): The clusters pi ∈π {Λ(TA [pi ])} on
the centroid path are nested because pi is the parent of pi−1 , so Λ(TA [pi−1 ]) ⊆
Λ(TA [pi ]) for every i ∈ {2, 3, . . . , α}. The main loop (Step 5) checks each of these
clusters in order of increasing cardinality.
The algorithm maintains a binary search tree BT that, right after Step 5.5
in any iteration i of the main loop is complete, contains every node x from TB
with Λ(TA [pi ]) 6^ Λ(TB [x]). Whenever a node x is inserted into BT , its key
is set to the weight w(TB [x]). Using BT , Step 5.6 retrieves the weight M of
the heaviest cluster in TB that is incompatible with Λ(TA [pi ]) (if any). Then,
Step 5.7 computes a value βi , defined as the maximum weight of all special nodes
in TB (if any) that are incompatible with the current TA [pi ]; if βi > M then
M is set to βi . Step 5.8 saves Λ(TA [pi ]) by inserting it into the tree Rc if its
weight is strictly greater than M . After Step 5 is done, C(Rc ) = {Λ(TA [u]) : u ∈
π and w(u) > w(x) for every x ∈ V (TB ) with Λ(TA [u]) 6^ Λ(TB [x])}.
In order to update BT correctly while moving upwards along π in Step 5, the
algorithm relies on Lemma 6. In each iteration i ∈ {2, 3, . . . , α} of Step 5, ri is
the lowest common ancestor in TB of Λ(TA [pi ]). By Lemma 6, the clusters in TB
that are incompatible with Λ(TA [pi ]) are of the form TB [v] where: (1) v lies on a
path in TB from a child of ri to a leaf in Λ(TA [pi ]); and (2) Λ(T [v]) 6⊆ Λ(TA [pi ]).
Accordingly, BT is updated in Steps 5.3–5.5 as follows. Condition (1) is taken
care of by first inserting all nodes from TB between ri−1 and ri except ri into BT
in Step 5.3 and then inserting all leaf descendants of pi that are not descendants
of pi−1 , along with any of their ancestors in TB that were not already in BT ,
into BT in Step 5.4. Finally, Step 5.5 enforces condition (2) by using counters to
locate and remove all nodes from BT (if any) whose clusters are proper subsets
of Λ(TA [pi ]). To do this, counter (x) for every node x in TB is updated so that
it stores the number of leaves in Λ(TB [x]) ∩ Λ(TA [pi ]) for the current i, and if
counter (x) reaches the value |Λ(TB [x])| then x is removed from BT .
To compute βi in Step 5.7, take the maximum of: (i) βi−1 ; (ii) the weights
of all special nodes on the path between ri and ri−1 in TB ; and (iii) the weights
of all special nodes that belong to a path between ri and a leaf in D.
Lemma 7. Procedure Filter Clusters runs in O(n log2 n) time.
Proof. (Omitted from the conference version due to space constraints.)
5
t
u
Implementations
As noted in Section 1.2, there does not seem to be any publicly available implementation for the majority rule (+) consensus tree. To fill this void, we implemented Algorithm Maj Rule Plus from Section 3 in C++ and included it in the
source code of the FACT (Fast Algorithms for Consensus Trees) package [17] at:
http://compbio.ddns.comp.nus.edu.sg/~consensus.tree/
To test the implementation, we repeatedly applied it to 10 random sets of trees
for various specified values of (k, n), generated with the method described in
Section 6.2 of [17]. The following worst-case running times (in seconds) were
obtained using Ubuntu Nutty Narwhal, a 64-bit operating system with 8.00 GB
RAM, and a 2.20 GHz CPU:
(k, n) (100, 500) (100, 1000) (100, 2000) (100, 5000) (500, 100) (1000, 100) (2000, 100) (5000, 100) (1000, 2000)
Time
0.63
1.51
2.99
6.78
0.65
1.29
2.72
6.66
27.29
The situation for the frequency difference consensus tree is less critical as
there already exist implementations, e.g., in the software package TNT [15]. Nevertheless, it could be useful to implement our algorithm Frequency Difference
from Section 4 in the future and compare its practical performance to TNT.
Before doing that, one should try to simplify the procedure Filter Clusters.
References
1. E. N. Adams III. Consensus techniques and the comparison of taxonomic trees.
Systematic Zoology, 21(4):390–397, 1972.
2. N. Amenta, F. Clarke, and K. St. John. A linear-time majority tree algorithm. In
Proceedings of WABI 2003, volume 2812 of LNCS, pages 216–227. Springer-Verlag,
2003.
3. J.-P. Barthélemy and F. R. McMorris. The median procedure for n-trees. Journal
of Classification, 3(2):329–334, 1986.
4. K. Bremer. Combinable component consensus. Cladistics, 6(4):369–372, 1990.
5. D. Bryant. A classification of consensus methods for phylogenetics. In M. F.
Janowitz, F.-J. Lapointe, F. R. McMorris, B. Mirkin, and F. S. Roberts, editors,
Bioconsensus, volume 61 of DIMACS Series in Discrete Mathematics and Theoretical Computer Science, pages 163–184. American Mathematical Society, 2003.
6. R. Cole, M. Farach-Colton, R. Hariharan, T. Przytycka, and M. Thorup. An
O(n log n) algorithm for the maximum agreement subtree problem for binary trees.
SIAM Journal on Computing, 30(5):1385–1404, 2000.
7. J. A. Cotton and M. Wilkinson. Majority-rule supertrees. Systematic Biology,
56(3):445–452, 2007.
8. Y. Cui, J. Jansson, and W.-K. Sung. Polynomial-time algorithms for building a
consensus MUL-tree. Journal of Computational Biology, 19(9):1073–1088, 2012.
9. W. H. E. Day. Optimal algorithms for comparing trees with labeled leaves. Journal
of Classification, 2(1):7–28, 1985.
10. J. H. Degnan, M. DeGiorgio, D. Bryant, and N. A. Rosenberg. Properties of
consensus methods for inferring species trees from gene trees. Systematic Biology,
58(1):35–54, 2009.
11. J. Dong, D. Fernández-Baca, F. R. McMorris, and R. C. Powers. Majority-rule (+)
consensus trees. Mathematical Biosciences, 228(1):10–15, 2010.
12. J. Felsenstein. Inferring Phylogenies. Sinauer Associates, Inc., Sunderland, Massachusetts, 2004.
13. J. Felsenstein. PHYLIP, version 3.6. Software package, Department of Genome
Sciences, University of Washington, Seattle, U.S.A., 2005.
14. P. A. Goloboff, J. S. Farris, M. Källersjö, B. Oxelman, M. J. Ramı́rez, and C. A.
Szumik. Improvements to resampling measures of group support. Cladistics,
19(4):324–332, 2003.
15. P. A. Goloboff, J. S. Farris, and K. C. Nixon. TNT, a free program for phylogenetic
analysis. Cladistics, 24(5):774–786, 2008.
16. M. T. Holder, J. Sukumaran, and P. O. Lewis. A justification for reporting
the majority-rule consensus tree in Bayesian phylogenetics. Systematic Biology,
57(5):814–821, 2008.
17. J. Jansson, C. Shen, and W.-K. Sung. Improved algorithms for constructing consensus trees. In Proceedings of SODA 2013, pages 1800–1813. SIAM, 2013.
18. J. Jansson, C. Shen, and W.-K. Sung. An optimal algorithm for building the
majority rule consensus tree. In Proceedings of RECOMB 2013, volume 7821 of
LNCS, pages 88–99. Springer-Verlag, 2013.
19. J. Jansson and W.-K. Sung. Constructing the R* consensus tree of two trees in
subcubic time. Algorithmica, 66(2):329–345, 2013.
20. M. Lott, A. Spillner, K. T. Huber, A. Petri, B. Oxelman, and V. Moulton. Inferring
polyploid phylogenies from multiply-labeled gene trees. BMC Evolutionary Biology,
9:216, 2009.
21. T. Margush and F. R. McMorris. Consensus n-Trees. Bulletin of Mathematical
Biology, 43(2):239–244, 1981.
22. F. R. McMorris and R. C. Powers. A characterization of majority rule for hierarchies. Journal of Classification, 25(2):153–158, 2008.
23. R. Page. COMPONENT, version 2.0. Software package, University of Glasgow,
U.K., 1993.
24. F. Ronquist and J. P. Huelsenbeck. MrBayes 3: Bayesian phylogenetic inference
under mixed models. Bioinformatics, 19(12):1572–1574, 2003.
25. R. R. Sokal and F. J. Rohlf. Taxonomic congruence in the Leptopodomorpha
re-examined. Systematic Zoology, 30(3):309–325, 1981.
26. J. Sukumaran and M. T. Holder. DendroPy: a Python library for phylogenetic
computing. Bioinformatics, 26(12):1569–1571, 2010.
27. W.-K. Sung. Algorithms in Bioinformatics: A Practical Introduction. Chapman &
Hall/CRC, 2010.
28. D. L. Swofford. PAUP*, version 4.0. Software package, Sinauer Associates, Inc.,
Sunderland, Massachusetts, 2003.
| 5cs.CE
|
OUTER ACTIONS OF Out(Fn ) ON SMALL RIGHT-ANGLED
ARTIN GROUPS
arXiv:1607.07031v2 [math.GR] 30 Jan 2018
DAWID KIELAK
Abstract. We determine the precise conditions under which SOut(Fn ), the
unique index two subgroup of Out(Fn ), can act non-trivially via outer automorphisms on a RAAG whose defining graph has fewer than 21 n
vertices.
2
We also show that the outer automorphism group of a RAAG cannot act
faithfully via outer automorphisms on a RAAG with a strictly smaller (in
number of vertices) defining graph.
Along the way we determine the minimal dimensions of non-trivial linear
representations of congruence quotients of the integral special linear groups
over algebraically closed fields of characteristic zero, and provide a new lower
bound on the cardinality of a set on which SOut(Fn ) can act non-trivially.
1. Introduction
The main purpose of this article is to study the ways in which Out(Fn ) can act
via outer automorphisms on a right-angled Artin group AΓ with defining graph Γ.
(Recall that AΓ is given by a presentation with generators being the vertices of Γ,
and relators being commutators of vertices which span an edge in Γ.) Such actions
have previously been studied for the extremal cases: when the graph Γ is discrete,
we have Out(AΓ ) = Out(Fm ) for some m, and homomorphisms
Out(Fn ) → Out(Fm )
have been investigated by Bogopolski–Puga [BP], Khramtsov [Khr2], Bridson–
Vogtmann [BV2], and the author [Kie1, Kie2]. When the graph Γ is complete,
we have Out(AΓ ) = GLm (Z), and homomorphisms
Out(Fn ) → GLm (Z)
or more general representation theory of Out(Fn ) have been studied by Grunewald–
Lubotzky [GL], Potapchik–Rapinchuk [PR], Turchin–Wilwacher [TW], and the author [Kie1, Kie2].
There are two natural ways of constructing non-trivial homomorphisms
φ : Out(Fn ) → Out(AΓ )
When Γ is a join of two graphs, ∆ and Σ say, then Out(AΓ ) contains
Out(A∆ ) × Out(AΣ )
as a finite index subgroup. When additionally ∆ is isomorphic to the discrete graph
with n vertices, then Out(A∆ ) = Out(Fn ), and so we have an obvious embedding
φ.
In fact this method works also for a discrete ∆ with a very large number of vertices, since there are injective maps Out(Fn ) → Out(Fm ) constructed by Bridson–
Vogtmann [BV2] for specific values of m growing exponentially with n.
Date: January 31, 2018.
1
2
DAWID KIELAK
The other way of constructing non-trivial homomorphisms φ becomes possible
when Γ contains n vertices with identical stars. In this case it is immediate that
these vertices form a clique Θ, and we have a map
GLn (Z) = Aut(AΘ ) → Aut(AΓ ) → Out(AΓ )
We also have the projection
Out(Fn ) → Out(H1 (Fn )) = GLn (Z)
and combining these two maps gives us a non-trivial (though also non-injective) φ.
This second method does not work in other situations, due to the following result
of Wade.
Theorem 1.1 ([Wad]). Let n > 3. Every homomorphism
SLn (Z) → Out(AΓ )
has finite image if and only if Γ does not contain n distinct vertices with equal stars.
In fact Wade proved a much more general result, in which the domain of the
homomorphism is allowed to be any irreducible lattice in a real semisimple Lie
group with finite centre and without compact factors, and with real rank n − 1.
The aim of this paper is to prove
3.7. Let n > 6. Suppose that Γ is a simplicial graph with fewer than
Theorem
1 n
vertices,
which does not contain n distinct vertices with equal stars, and is
2 2
not a join of the discrete graph with n vertices and another (possibly empty) graph.
Then every homomorphism SOut(Fn ) → Out(AΓ ) is trivial.
Here SOut(Fn ) denotes the unique index two subgroup of Out(Fn ).
The proof is an induction, based on an observation present in a paper of Charney–
Crisp–Vogtmann [CCV], elaborated further in a paper of Hensel and the author [HK],
which states that, typically, the graph Γ contains many induced subgraphs Σ which
are invariant up to symmetry, in the sense that the subgroup of AΓ the vertices of
Σ generate is invariant under any outer action up to an automorphism induced by
a symmetry of Γ (and up to conjugacy).
To use the induction we need to show that such subgraphs are really invariant,
that is that we do not need to worry about the symmetries of Γ. To achieve this
we prove
Theorem 2.28. Every action of Out(Fn ) (with n > 6) on a set of cardinality
m 6 n+1
factors through Z/2Z.
2
Since SOut(Fn ) is the unique index two subgroup of Out(Fn ), the conclusion of
this theorem is equivalent to saying that SOut(Fn ) lies in the kernel of the action.
A crucial ingredient in the proof of this theorem is the following.
Theorem 2.27. Let V be a non-trivial, irreducible K-linear representation of
SLn (Z/qZ)
where n > 3, q is a power of a prime p, and where K is an algebraically closed field
of characteristic 0. Then
2
if
(n, p) = (3, 2)
dim V >
pn−1 − 1 otherwise
This result seems not to be present in the literature; it extends a theorem of
Landazuri–Seitz [LS] yielding a very similar statement for q = p (see Theorem 2.26).
At the end of the paper we also offer
OUTER ACTIONS OF Out(Fn ) ON SMALL RAAGS
3
Theorem 4.1. There are no injective homomorphisms Out(AΓ ) → Out(AΓ′ ) when
Γ′ has fewer vertices than Γ.
This theorem follows from looking at the Z/2Z-rank, i.e. the largest subgroup
isomorphic to (Z/2Z)k .
2. The tools
2.1. Automorphisms of free groups.
Definition 2.1 (SOut(Fn )). Consider the composition
Aut(Fn ) → GLn (Z) → Z/2Z
where the first map is obtained by abelianising Fn , and the second map is the
determinant. We define SAut(Fn ) to be the kernel of this map; we define SOut(Fn )
to be the image of SAut(Fn ) in Out(Fn ).
It is easy to see that both SAut(Fn ) and SOut(Fn ) are index two subgroups of,
respectively, Aut(Fn ) and Out(Fn ).
The group SAut(Fn ) has a finite presentation given by Gersten [Ger], and from
this presentation one can immediately obtain the following result.
Proposition 2.2 (Gersten [Ger]). The abelianisation of SAut(Fn ), and hence of
SOut(Fn ), is trivial for all n > 3.
It follows that SOut(Fn ) is the unique subgroup of Out(Fn ) of index two.
We will now look at symmetric and alternating subgroups of Out(Fn ), and list
some corollaries of their existence.
Proposition 2.3 ([BV1, Proposition 1]). Let n > 3. There exists a symmetric
subgroup of rank n
Symn < Out(Fn )
such that any homomorphism φ : Out(Fn ) → G that is not injective on Symn has
image of cardinality at most 2.
The symmetric group is precisely the symmetric group operating on some fixed
basis of Fn . It is easy to see that it intersects SOut(Fn ) in an alternating group
Altn . Whenever we talk about the alternating subgroup Altn of SOut(Fn ), we mean
this subgroup. Note that SOut(Fn ) actually contains an alternating subgroup of
rank n + 1, which is a supergroup of our Altn ; we will denote it by Altn+1 . There
is also a symmetric supergroup Symn+1 of Altn+1 contained in Out(Fn ).
The proof of [BV1, Proposition 1] actually allows one to prove the following
proposition.
Proposition 2.4. Let n > 3. Then SOut(Fn ) is the normal closure of any nontrivial element of Altn .
Following the proof of [BV1, Theorem A], we can now conclude
Corollary 2.5. Let
φ : SOut(Fn ) → GLk (Z)
be a homomorphism, with n > 6 and k < n. Then φ is trivial.
Proof. For n > 6, the alternating group Altn+1 does not have non-trivial complex
representations below dimension n. Thus φ|Altn+1 is not injective, and therefore
trivial, as Altn+1 is simple. Now we apply Proposition 2.4.
More can be said about linear representations of Out(Fn ) in somewhat larger
dimensions – see [Kie1, Kie2, TW].
Another related result that we will use is the following.
4
DAWID KIELAK
Theorem 2.6 ([Kie1]). Let n > 6 and m < n2 . Then every homomorphism
Out(Fn ) → Out(Fm ) has image of cardinality at most 2, provided that m 6= n.
In fact, we will need to go back to the proof of the above theorem and show:
Theorem 2.7. Let n > 6 and m < 12 n2 . Then every homomorphism
SOut(Fn ) → Out(Fm )
is trivial, provided that m 6= n.
The proof of this result forms the content of the next section.
2.2. Homomorphisms SOut(Fn ) → Out(Fm ). To study such homomorphisms we
need to introduce finite subgroups Bn and B of SOut(Fn ) that will be of particular
use. Let Fn be freely generated by {a1 , . . . , an }.
Definition 2.8. Let us define δ ∈ Out(Fn ) by δ(ai ) = ai −1 for each i. (Formally
speaking, this defines an element in Aut(Fn ); we take δ to be the image of this
element in Out(Fn ).) Define σ12 ∈ Symn < Out(Fn ) to be the transposition
swapping a1 with a2 . Define ξ ∈ SOut(Fn ) by
δ
if n is even
ξ=
δσ12 if n is odd
and set Bn = hAltn+1 , ξi 6 SOut(Fn ).
We also set A to be either Altn−1 , the pointwise stabiliser of {1, 2} when Altn+1
acts on {1, 2, . . . , n + 1} in the natural way (in the case of odd n), or Altn+1 (in the
case of even n). Furthermore, we set B = hA, ξi.
It is easy to see that Bn is a finite group – it is a subgroup of the automorphism
group of the (suitably marked) (n + 1)-cage graph, that is a graph with 2 vertices
and n + 1 edges connecting one to another.
To prove Theorem 2.7 we need to introduce some more notation from [Kie1].
Throughout, when we talk about modules or representations, we work over the
complex numbers.
Definition 2.9. A B-module V admits a convenient split if and only if V splits as
a B-module into
V = U ⊕ U′
where U is a sum of trivial A-modules and ξ acts as minus the identity on U ′ .
Definition 2.10. A graph X with a G-action is called G-admissible if and only
if it is connected, has no vertices of valence 2, and any G-invariant forest in X
contains no edges. Here by ‘invariant’ we mean setwise invariant.
Proposition 2.11 ([Kie1]).
Let n > 6. Suppose that X is a Bn -admissible graph
of rank smaller than n+1
such that
2
(1) the B-module H1 (X; C) admits a convenient split; and
(2) any vector in H1 (X; C) which is fixed by Altn+1 is also fixed by ξ; and
(3) the action of Bn on X restricted to A is non-trivial.
Then X is the (n + 1)-cage.
The above proposition does not (unfortunately) feature in this form in [Kie1] –
it does however follow from the proof of [Kie1, Proposition 6.7].
Proof of Theorem 2.7. Let φ : SOut(Fn ) → Out(Fm ) be a homomorphism. Using
Nielsen realisation for free groups (due to, independently, Culler [Cul], Khramtsov [Khr1]
and Zimmermann [Zim]) we construct a finite connected graph X with fundamental group Fm , on which Bn acts in a way realising the outer action φ|Bn . We
OUTER ACTIONS OF Out(Fn ) ON SMALL RAAGS
5
easily arrange for X to be Bn -admissible by collapsing invariant forests. Note that
V = H1 (Fm ; C) is naturally isomorphic to H1 (X; C) as a Bn -module.
We have a linear representation
SOut(Fn ) → Out(Fm ) → GL(V )
where the first map is φ. We can induce it to a linear representation
Out(Fn ) → GL(W )
of dimension dim W = 2 dim V = 2m. Since we are assuming that
1 n
m<
2 2
the combination of [Kie1, Lemma 3.8 and Proposition 3.11] tells us that W splits
as an Out(Fn )-module as
W = W0 ⊕ W1 ⊕ Wn−1 ⊕ Wn
where the action of Out(Fn ) is trivial on W0 but not on Wn , and the action of the
subgroup SOut(Fn ) is trivial on both. Moreover, as Symn+1 modules, W1 is the
sum of standard and Wn−1 of signed standard representations. We also know that
δ acts on Wi as multiplication by (−1)i .
When n is even this immediately tells us that, as a B = Bn -module, we have
W = U ⊕ U′
where U = W0 ⊕ Wn is sum of trivial A = Altn+1 -modules, and ξ = δ acts on
U ′ = W1 ⊕ Wn−1
as minus the identity.
When n is odd we can still write
W = U ⊕ U′
as a B-module, with A acting trivially on U and ξ acting as minus the identity
on U ′ . Here we have W0 ⊕ Wn < U , but U also contains the trivial A-modules
contained in W1 ⊕ Wn−1 . The module U ′ is the sum of the standard A-modules.
Thus W admits a convenient split.
Now we claim that V also admits a convenient split as a B-module. To define the
induced Out(Fn )-module W we need to pick en element Out(Fn ) r SOut(Fn ); we
have already defined such an element, namely σ12 . The involution σ12 commutes
with ξ and conjugates A to itself. Thus, as an A module, V could only consist of
the trivial and standard representations, since these are the only A-modules present
in W . Moreover, any trivial A-module in V is still a trivial A-module in W , and
so ξ acts as minus the identity on it. Therefore V also admits a convenient split as
a B-module. This way we have verified assumption (1) of Proposition 2.11.
Observe that the SOut(Fn )-module V embeds into W . In W every Altn+1 -fixed
vector lies in W0 ⊕ Wn , and here ξ acts as the identity. Thus assumption (2) of
Proposition 2.11 is satisfied in W , and therefore also in V .
We have verified the assumptions (1) and (2) of Proposition 2.11; we also know
that the conclusion of Proposition 2.11 fails, since the n + 1-cage has rank n, which
would force m = n, contradicting the hypothesis of the theorem. Hence we know
that assumption (3) of Proposition 2.11 fails, and so A acts trivially on X. But
this implies that A 6 ker φ.
Note that A is a subgroup of the simple group Altn+1 , and so we have
Altn+1 6 ker φ
But then Proposition 2.4 tells us that φ is trivial.
6
DAWID KIELAK
2.3. Automorphisms of RAAGs. Throughout the paper, Γ will be a simplicial
graph, and AΓ will be the associated RAAG, that is the group generated by the
vertices of Γ, with a relation of two vertices commuting if and only if they are joined
by an edge in Γ.
We will often look at subgraphs of Γ, and we always take them to be induced
subgraphs. Thus we will make no distinction between a subgraph of Γ and a subset
of the vertex set of Γ.
Given an induced subgraph Σ ⊆ Γ we define AΣ to be the subgroup of AΓ generated by (the vertices of) Σ. Abstractly, AΣ is isomorphic to the RAAG associated
to Σ (since Σ is an induced subgraph).
Definition 2.12 (Links, stars, and extended stars). Given a subgraph Σ ⊆ Γ we
define
• lk(Σ) = {w ∈ Γ | w is adjacent to v for all v ∈ Σ};
• st(Σ) = Σ ∪ lk(Σ);
b
• st(Σ)
= lk(Σ) ∪ lk(lk(Σ)).
Definition 2.13 (Joins and cones). We say that two subgraphs Σ, ∆ ⊆ Γ form a
join Σ ∗ ∆ ⊆ Γ if and only if Σ ⊆ lk(∆) and ∆ ⊆ lk(Σ).
A subgraph Σ ⊆ Γ is a cone if and only if there exists a vertex v ∈ Σ such that
Σ = v ∗ (Σ r {v}). In particular, a singleton is a cone.
Definition 2.14 (Join decomposition). Given a graph Σ we say that
Σ = Σ1 ∗ · · · ∗ Σk
is the join decomposition of Σ when each Σi is non-empty, and is not a join of two
non-empty subgraphs.
Each of the graphs Σi is called a factor, and the join of all the factors which are
singletons is called the clique factor.
We will often focus on a specific finite index subgroup Out0 (AΓ ) of Out(AΓ ),
called the group of pure outer automorphisms of AΓ . To define it we need to discuss
a generating set of Out(AΓ ) due to Laurence [Lau] (it was earlier conjectured to be
a generating set by Servatius [Ser]).
Aut(AΓ ) is generated by the following classes of automorphisms:
(1)
(2)
(3)
(4)
Inversions
Partial conjugations
Transvections
Graph symmetries
Here, an inversion maps one generator of AΓ to its inverse, fixing all other generators.
A partial conjugation needs a vertex v; it conjugates all generators in one connected component of Γ r st(v) by v, and fixes all other generators.
A transvection requires vertices v, w with st(v) ⊇ lk(w). For such v and w, a
transvection is the automorphism which maps w to wv, and fixes all other generators.
A graph symmetry is an automorphism of AΓ which permutes the generators
according to a combinatorial automorphism of Γ.
The group Aut0 (AΓ ) of pure automorphisms is defined to be the subgroup generated by generators of the first three types, i.e. without graph symmetries. The
group Out0 (AΓ ) of pure outer automorphisms is the quotient of Aut0 (AΓ ) by the
inner automorphisms.
Let us quote the following result of Charney–Crisp–Vogtmann:
OUTER ACTIONS OF Out(Fn ) ON SMALL RAAGS
7
Proposition 2.15 ([CCV, Corollary 3.3]). There exists a finite subgroup
Q < Out(AΓ )
consisting solely of graph symmetries, such that
Out(AΓ ) = Out0 (AΓ ) ⋊ Q
Corollary 2.16. Suppose that any action of G on a set of cardinality at most k is
trivial, and assume that Γ has k vertices. Then any homomorphism
φ : G → Out(AΓ )
0
has image contained in Out (AΓ ).
Proof. Proposition 2.15 tells us that
Out(AΓ ) = Out0 (AΓ ) ⋊ Q
for some group Q acting faithfully on Γ. Hence we can postcompose φ with the
quotient map
Out0 (AΓ ) ⋊ Q → Q
and obtain an action of G on the set of vertices of Γ. By assumption this action
has to be trivial, and thus φ(G) lies in the kernel of this quotient map, which is
Out0 (AΓ ).
Definition 2.17 (G-invariant subgraphs). Given a homomorphism G → Out(AΓ )
we say that a subgraph Σ ⊆ Γ is G-invariant if and only if the conjugacy class of
AΣ is preserved (setwise) by G.
Definition 2.18. Having an invariant subgraph Σ ⊆ Γ allows us to discuss two
additional actions:
• Since, for any subgraph Σ, the normaliser of AΣ in AΓ is equal to AΣ C(AΣ ),
where C(AΣ ) is the centraliser of AΣ (see e.g. [CSV, Proposition 2.2]), any
invariant subgraph Σ gives us an induced (outer) action G → Out(AΣ ).
• When Σ is invariant, we also have the induced quotient action
G → Out(AΓ /hhAΣ ii) ≃ Out(AΓrΣ )
Let us quote the following.
Lemma 2.19 ([HK, Lemmata 4.2 and 4.3]). For any homomorphism G → Out0 (AΓ )
we have:
(1) for every subgraph Σ ⊆ Γ which is not a cone, lk(Σ) is G-invariant;
(2) connected components of Γ which are not singletons are G-invariant;
b
(3) st(Σ)
is G-invariant for every subgraph Σ;
(4) if Σ and ∆ are G-invariant, then so is Σ ∩ ∆;
(5) if Σ is G-invariant, then so is st(Σ).
Definition 2.20 (Trivialised subgraphs). Let φ : G → Out(AΓ ) be given. We say
that a subgraph Σ ⊆ Γ is trivialised if and only if Σ is G-invariant, and the induced
action is trivial.
Lemma 2.21. Let φ : G → Out(AΓ ) be a homomorphism. Suppose that Σ is a
connected component of Γ which is trivialised by G. Consider the graph
Γ′ = (Γ r Σ) ⊔ {s}
were s denotes a new vertex not present in Γ. There exists an action
ψ : G → Out(AΓ′ )
for which {s} is invariant, and such that the quotient actions
G → Out(AΓrΣ )
8
DAWID KIELAK
induced by φ and ψ by removing, respectively, Σ and s, coincide.
Proof. Consider an epimorphism f : AΓ → AΓ′ defined on vertices of Γ by
v if v 6∈ Σ
f (v) =
s if v ∈ Σ
The kernel of f is normally generated by elements vu−1 , where v, u ∈ Σ are vertices.
Since the induced action of G on AΣ is trivialised, the action preserves each element
vu−1 up to conjugacy. But this in particular means that G preserves the (conjugacy
class of) the kernel of f , and hence φ induces an action
G → Out(AΓ′ )
which we call ψ. It is now immediate that ψ is as required.
2.4. Finite groups acting on RAAGs.
Definition 2.22. Suppose that Γ has k vertices. Then the abelianisation of AΓ is
isomorphic to Zk , and we have the natural map
Out(AΓ ) → Out(H1 (AΓ )) = GLk (Z)
We will refer to the kernel of this map as the Torelli subgroup.
We will need the following consequence of independent (and more general) results
of Toinet [Toi] and Wade [Wad].
Theorem 2.23 (Toinet [Toi]; Wade [Wad]). The Torelli group is torsion free.
Lemma 2.24. Let φ : H → Out(AΓ ) be a homomorphism with a finite domain.
Suppose that Γ = Σ1 ∪ · · · ∪ Σm , and each Σi is trivialised by H. Then so is Γ.
Proof. Consider the action
ψ : H → Out(H1 (AΓ )) = GLk (Z)
obtained by abelianising AΓ , where k is the number of vertices of Γ. This Z-linear
representation ψ preserves the images of the subgroups AΣi , and is trivial on each
of them. Thus the representation is trivial, and so φ(H) lies in the Torelli group.
But the Torelli subgroup is torsion free. Hence φ is trivial.
Lemma 2.25. Let φ : G → Out(AΓ ) be a homomorphism. Let
Γ = (Γ1 ∪ · · · ∪ Γn ) ⊔ Θ
where n > 1, each Γi is trivialised by G, and where Θ is a discrete graph with m
vertices. Suppose that for some l ∈ {m, m + 1} any homomorphism
G → Out(Fl )
is trivial. Then Γ is trivialised, provided that G is the normal closure of a finite
subgroup H, and that G contains a perfect subgroup P , which in turn contains H.
Proof. We can quotient out all of the groups AΓi , and obtain an induced quotient
action
(∗)
G → Out(AΘ )
We claim that this map is trivial. To prove the claim we have to consider two
cases: the first case occurs when l = m in the hypothesis of our lemma, that is
every homomorphism
G → Out(Fm )
is trivial. Since Θ is a discrete graph with m vertices, we have Out(AΘ ) = Out(Fm )
and so the homomorphism (∗) is trivial.
OUTER ACTIONS OF Out(Fn ) ON SMALL RAAGS
9
The second case occurs when l = m + 1 in the hypothesis of our lemma. In this
situation we quotient AΓ by each subgroup AΓi for i > 1, but instead of quotienting
out AΓ1 , we use Lemma 2.21. This way we obtain an outer action on a free group
with m + 1 generators, and such an action has to be trivial by assumption. Thus
we can take a further quotient and conclude again that the induced quotient action
(∗) on AΘ is trivial. This proves the claim.
Now consider the action of G on the abelianisation of AΓ . We obtain a map
ψ : G → GLk (Z)
where k is the number of vertices of Γ. Since each Γi is trivialised, and the induced
quotient action on AΘ is trivial, we see that ψ(G) lies in the abelian subgroup
of GLn (Z) formed by block-upper triangular matrices with identity blocks on the
diagonal, and a single non-trivial block of fixed size above the diagonal. But P is
perfect, and so ψ(P ) must lie in the Torelli subgroup of Out(AΓ ). This is however
torsion free by Theorem 2.23, and so H must in fact lie in the kernel of φ. We
conclude that the action of G on Γ is also trivial, since G is the normal closure of
H.
2.5. Some representation theory. Let us mention a result about representations
of PSLn (Z/pZ), for prime p, due to Landazuri and Seitz:
Theorem 2.26 ([LS]). Suppose that we have a non-trivial, irreducible projective
representation PSLn (Z/pZ) → PGL(V ), where n > 3, p is prime, and V is a vector
space over a field K of characteristic other than p. Then
2
if
(n, p) = (3, 2)
dim V >
pn−1 − 1 otherwise
We offer an extension of their theorem for algebraically closed fields of characteristic 0, which we will need to discuss actions of Out(Fn ) and SOut(Fn ) on finite
sets.
Theorem 2.27. Let V be a non-trivial, irreducible K-linear representation of
SLn (Z/qZ), where n > 3, q is a power of a prime p, and where K is an algebraically closed field of characteristic 0. Then
2
if
(n, p) = (3, 2)
dim V >
pn−1 − 1 otherwise
Proof. Let φ : SLn (Z/qZ) → GL(V ) denote our representation. Consider Z, the
subgroup of SLn (Z/qZ) generated by diagonal matrices with all non-zero entries
equal. Note that Z is the centre of SLn (Z/qZ). Hence V splits as an SLn (Z/qZ)module into intersections of eigenspaces of all elements of Z. Since V is irreducible,
we conclude that φ(Z) lies in the centre of GL(V ).
First suppose that q = p. Consider the composition
SLn (Z/qZ) → GL(V ) → PGL(V )
We have just showed that Z lies in the kernel of this composition, and so our representation descends to a representation of PSLn (Z/pZ) ∼
= SLn (Z/pZ)/Z. This new,
projective representation is still irreducible. It is also non-trivial, as otherwise V
would have to be a 1-dimensional non-trivial SLn (Z/qZ)-representation. There are
no such representations since SLn (Z/qZ) is perfect when p = q. Now Theorem 2.26
yields the result.
Suppose now that q = pα , where α > 1. Let N E SLn (Z/qZ) be the kernel of the
natural map SLn (Z/qZ) → SLn (Z/pZ). As an N -module, by Maschke’s Theorem,
10
DAWID KIELAK
V splits as
V =
k
M
Ui
i=1
where each Ui 6= {0} is a direct sum of irreducible N -modules, and irreducible
submodules W 6 Ui , W ′ 6 Uj are isomorphic if and only if i = j.
Observe that we get an induced action of SLn (Z/qZ)/N ∼
= SLn (Z/pZ) on the set
{Ui , U2 , . . . , Uk }. As V is an irreducible SLn (Z/qZ)-module, the action is transitive.
Note that an action of a group on a finite set S induces a representation on the
vector space with basis S. If k > 1 then this representation is not the sum of trivial
ones, because of the transitivity just described, and so
2
if
(n, p) = (3, 2)
k>
pn−1 − 1 otherwise
since our theorem holds for SLn (Z/pZ). Since dim Ui > 1 for all i, we get dim V > k
and our result follows.
Let us henceforth assume that k = 1. We have
l
M
W
V = U1 =
j=1
where W is an irreducible N -module.
Note that we have an alternating group Altn < SLn (Z/qZ) satisfying
Altn ∩N = {1}
Let σ ∈ Altn be an element of order o(σ) equal to 2 or 3.
Consider the group M = hN, σi < SLn (Z/qZ). Note that M ∼
= N ⋊ Zo(σ) . The
module V splits as a direct sum of irreducible M -modules by Maschke’s theorem.
Let X be such an irreducible M -module.
Note that X as an N -module is a direct sum of, say, m copies of the N -module
W (with m > 1). Frobenius Reciprocity (see e.g. [Wei, Corollary 4.1.17]) tells us
that the multiplicity m of W (as an N -module) in X is equal to the multiplicity
of the M -module X in the M -module induced from the N -module W . Hence the
multiplicity of W in the M -module induced from the N -module W is at least m2 .
But it is bounded above by o(σ) and o(σ) 6 3, which forces m = 1, as m > 1.
This shows in particular that X as an N -module is isomorphic to W . It also
shows that the M -module induced from W contains a submodule isomorphic to X.
Since
M∼
= N ⋊ Zo(σ)
an easy calculation shows that σ acts on this copy of X as a scalar multiple of
the identity matrix, i.e. via a central matrix. This is true for every irreducible
M -submodule X of V , and hence σ commutes with N when acting on V . Since
the above statement is true for each σ ∈ Altn of order 2 or 3, we conclude that φ
factors through SLn (Z/qZ)/[N, Altn ]. Note that we need to consider elements σ of
order 3 when we are dealing with the case n = 4.
Mennicke’s proof of the Congruence Subgroup Property [Men] tells us that N
is normally generated (as a subgroup of SLn (Z/qZ)) by the pth powers of the
elementary matrices. Now SLn (Z/qZ) itself is generated by elementary matrices;
let us denote such a matrix by Eij with the usual convention. Observe that for all
σ ∈ Altn we have
p
−1 p
−1 p
−1
φ(Eαβ
Eij Eαβ ) = φ(σ −1 Eαβ
Eij Eαβ σ) = φ(Eσ(α)σ(β)
Eij
Eσ(α)σ(β) )
Choose σ ∈ An such that σ(α)
= i and σ(β) = j. We conclude that φ(N ) lies in
the centre of φ SLn (Z/qZ) . In particular, φ(N ) is abelian, and hence (as K is
OUTER ACTIONS OF Out(Fn ) ON SMALL RAAGS
11
algebraically closed) dim W = 1, as W is an irreducible N -module. Since V is a
direct sum of N -modules isomorphic to W , the group N acts via matrices in the
centre of GL(V ). Hence N lies in the kernel of the composition
SLn (Z/qZ)
φ
/ GL(V )
/ PGL(V )
We have already shown that Z lies in this kernel, and so our representation descends to a projective representation of PSLn (Z/pZ). If we can show that this
representation is non-trivial, we can then apply Theorem 2.26 and our proof will
be finished.
Suppose that this projective representation is trivial. This means that V is a
1-dimensional, non-trivial SLn (Z/qZ)-representation. This is however impossible,
since the abelianisation of SLn (Z/qZ) is trivial when n > 3.
2.6. Actions of Out(Fn ) on finite sets.
Theorem 2.28. Every action of Out(Fn ) (with n > 6) on a set of cardinality
m 6 n+1
factors through Z/2Z.
2
Proof. Suppose that we are given such an action. It gives us
Out(Fn ) → Symm ֒→ GLm−1 (C)
where Symm denotes the symmetric group of rank m, and the second map is the
standard irreducible representation of Symm . Since
n+1
m−1<
2
the composition factors through the natural map Out(Fn ) → GLn (Z) induced by
abelianising Fn , by [Kie1, Theorem 3.13]. Thus we have
Out(Fn ) → GLn (Z) → GLm−1 (C)
with finite image. The Congruence Subgroup Property [Men] tells us that the map
GLn (Z) → GLm−1 (C) factors through a congruence map
GLn (Z) → GLn (Z/pα Z)
for some positive integer α and some prime p. Now
m − 1 < 2n−1 − 1 6 pn−1 − 1
and so the restricted map SLn (Z/pα Z) → GLm−1 (C) must be trivial by Theorem 2.27. Thus the given action factors through GLn (Z/pα Z)/ SLn (Z/pα Z), which
is an abelian group. Therefore SOut(Fn ) lies in the kernel of φ, since it is perfect
(Proposition 2.2), and we are finished.
Corollary 2.29. Every action of SOut(Fn ) (with n > 6) on a set of cardinality
is trivial.
m 6 21 n+1
2
Proof. Every action of an index k subgroup of a group G on a set of cardinality m
can be induced to an action of G on a set of cardinality km.
3. The main result
Definition 3.1. Let Dn denote the discrete graph with n vertices.
Definition 3.2. Let φ : G → Out(AΓ ) be a homomorphism, and let n be fixed.
We define two properties of the action (with respect to n):
C For every G-invariant clique Σ in Γ with at least n vertices there exists
a G-invariant subgraph Θ of Γ, such that Θ ∩ Σ is a proper non-empty
subgraph of Σ.
12
DAWID KIELAK
D For every G-invariant subgraph ∆ of Γ isomorphic to Dn , there exists a Ginvariant subgraph Θ of Γ, such that Θ ∩∆ is a proper non-empty subgraph
of ∆.
The notation C stands for ‘clique’, and D for ‘discrete’.
Lemma 3.3. Let φ : G → Out(AΓ ) be an action satisfying C and D. Let Ω be a
G-invariant subgraph of Γ. Then both the induced action and the induced quotient
action satisfy C and D.
Proof. Starting with a subgraph Σ or ∆ in either Ω or Γ r Ω, we observe that
the subgraph is a subgraph of Γ, and so using the relevant property we obtain a
G-invariant subgraph Θ. We now only need to observe that Θ ∩ Ω is G-invariant
by Lemma 2.19(4), and the image of Θ in Γ r Ω is invariant under the induced
quotient action
G → Out(AΓrΩ )
Theorem 3.4. Let us fix positive integers n and m > n. Suppose that a group G
satisfies all of the following:
(1) G is the normal closure of a finite subgroup H.
(2) All homomorphisms
G → Out(Fk )
are trivial when k 6= n and k < m.
(3) All homomorphisms
G → GLk (Z)
are trivial when k < n.
(4) Any action of G on a set of cardinality smaller than m is trivial.
Let
φ : G → Out(AΓ )
be a homomorphism, where Γ has fewer than m vertices. Then φ is trivial, provided
that the action satisfies properties C and D (with respect to n).
Proof. Formally, the proof is an induction on the number of vertices of Γ, and splits
into two cases.
Before we proceed, let us observe that assumption (4) allows us to apply Corollary 2.16, and hence to use Lemma 2.19 whenever we need to.
Case 1: Suppose that Γ does not admit proper non-empty G-invariant subgraphs.
Note that this is in particular the case when Γ is a single vertex, which is the
base case of our induction.
We claim that Γ is either discrete, or a clique. To prove the claim, let us suppose
that Γ is not discrete.
b
Let v be a vertex of Γ with a non-empty link. Lemma 2.19(3) tells us that st(v)
is G-invariant, and thus it must be equal to Γ. Hence Γ is a join, and therefore
admits a join decomposition.
If each factor of the decomposition is a singleton, then Γ is a clique as claimed.
Otherwise, the decomposition contains a factor Σ which is not a singleton and not a
join, and so in particular not a cone. Thus Lemma 2.19(1) informs us that lk(Σ) is
G-invariant. This is a contradiction, since this link is a proper non-empty subgraph.
We have thus shown the claim.
Suppose that Γ is a clique, with, say, k vertices. Property C immediately tells
us that k < n, and so we are dealing with a homomorphism
φ : G → Out(AΓ ) = GLk (Z)
where k < n. Such a homomorphism is trivial by assumption (3).
OUTER ACTIONS OF Out(Fn ) ON SMALL RAAGS
13
Suppose that Γ is a discrete graph, with, say, k vertices. Property D immediately
tells us that k 6= n, and so we are dealing with a homomorphism
φ : G → Out(AΓ ) = Out(Fk )
where k 6= n and k < m. Such a homomorphism is trivial by assumption (2).
Case 2: Suppose that Γ admits a proper non-empty G-invariant subgraph Σ.
Lemma 3.3 guarantees that the induced action
G → Out(AΣ )
satisfies the assumptions of our theorem, and thus, using the inductive hypothesis,
we conclude that this induced action is trivial.
We argue in an identical manner for the induced quotient action
G → Out(AΓrΣ )
and conclude that it is also trivial.
These two observations imply that in particular the restriction of these two
actions to the finite group H from assumption (1) is trivial. Now Lemma 2.24 tells
us that H lies in the kernel of φ, and hence so does G, as it is a normal closure of
H by assumption (1).
Lemma 3.5. Suppose that Γ does not contain n distinct vertices with identical
stars. Then property C holds for any action G → Out0 (AΓ ).
Proof. Let Σ be a G-invariant clique in Γ with at least n vertices. Since we know
that no n vertices of Γ have identical stars, we need to have distinct vertices of
Σ, say v and w, with st(v) 6= st(w). Without loss of generality we may assume
that there exists u ∈ st(v) r st(w). In particular this implies that u and w are not
adjacent.
Consider Λ = lk({u, w}): it is invariant by Lemma 2.19(1), since {u, w} is not a
cone; it intersects Σ non-trivially, since the intersection contains v; the intersection
is also proper, since w 6∈ Λ. Thus property C is satisfied.
Proposition 3.6. In Theorem 3.4, we can replace the assumption on the action
satisfying D by the assumption that Γ is not a join of Dn and another (possibly
empty) graph, provided that G satisfies additionally
(5) G contains a perfect subgroup P , which in turn contains H.
Proof. We are going to proceed by induction on the number of vertices of Γ, as
before. Assuming the inductive hypothesis, we will either show the conclusion of
the theorem directly, or we will show that in fact property D holds.
Note that the base case of induction (Σ being a singleton) always satisfies D.
Let ∆ be as in property D, and suppose that the property fails for this subgraph.
Case 1: suppose that there exists a vertex u of ∆ with a non-empty link.
b
Let v be a vertex of Γ r ∆ joined to some vertex of ∆. Consider st(v);
this
b
subgraph is G-invariant by Lemma 2.19(3). If st(v)
intersects ∆ and does not
b
contain it, then ∆ does satisfy property D. We may thus assume that ∆ ⊆ st(v).
b
We would like to apply induction to st(v), and conclude that this subgraph, and
hence ∆, are trivialised. This would force ∆ to satisfy property D.
b
There are two cases in which we cannot apply the inductive hypothesis to st(v):
this subgraph might be equal to Γ, or it might be a join of a subgraph isomorphic
to Dn and another subgraph.
In the former case, Γ is a join of two non-empty graphs. If there exists a factor Θ
of the join decomposition of Γ which is not a singleton, and which does not contain
∆, then let us look at lk(Θ). This is a proper subgraph of Γ, it is G-invariant by
14
DAWID KIELAK
Lemma 2.19(1), and is not a join of Dn and another graph since Γ is not. Thus
we may apply the inductive hypothesis to lk(Θ) and conclude that it is trivialised.
But ∆ ⊆ lk(Θ), and so ∆ is also trivialised, and thus satisfies D.
If Γ has no such factor Θ in its join decomposition, then Γ = st(Σ), where Σ is a
non-empty clique. The clique Σ is a proper subgraph, since it does not contain ∆.
It is G-invariant by Lemma 2.19(1) and so the inductive hypothesis tells us that it
is trivialised.
The induced quotient action G → Out(AΓrΣ ) is also trivialised by induction,
as Γ r Σ cannot be a join of Dn and another graph as before. We now apply
Lemma 2.24 for the subgroup H, and conclude that H, and hence its normal closure
G, act trivially.
b
Now we need to look at the situation in which st(v)
is a proper subgraph of Γ,
but it is a join of Dn and another graph.
b
Let us look at Λ, the intersection of st(v)
with the link of all factors of the
b
join decomposition of st(v)
isomorphic to Dn . The subgraph Λ is G-invariant
by Lemma 2.19(1) and (4). It is a proper subgraph of Γ, and so the inductive
hypothesis tells us that Λ is trivialised. If Λ contains ∆ then we are done.
The graph Λ does not contain ∆ if and only if ∆ is a factor of the join decompob
sition of st(v).
Observe that we can actually use another vertex of Γ r ∆ in place of
v, provided that this other vertex is joined by an edge to some vertex of ∆. Thus
b
we may assume that ∆ is a factor of the join decomposition of every st(v)
where v
is as described. This is however only possible when st(∆) is a connected component
of Γ. There must be at least one more component, since Γ is not a join of ∆ and
another graph.
Note that the component st(∆) is invariant by Lemma 2.19(5).
Suppose that the clique factor Σ of lk(∆) is non-trivial. As before, Σ is trivialised.
Observing that Γ r Σ is disconnected, and if it is discrete then it is has more than n
vertices, allows us to apply the inductive hypothesis to the quotient action induced
by Σ, and so, arguing as before, we see that Γ is trivialised.
Now suppose that lk(∆) has a trivial clique component. The join decomposition
of the component st(∆) consists of at least two factors, each of which is invariant by
Lemma 2.19(1). Let Θ be such a factor. Removing Θ leaves us with a disconnected
graph smaller than Γ. Thus, we may apply the inductive hypothesis, provided that
Γ r Θ is not Dn . This might however occur: in this situation st(∆) r Θ fulfils
the role of the graph Θ from the definition of D, and so we can use the inductive
hypothesis nevertheless.
We now apply Lemma 2.24 to the subgroup H and the induced quotient actions
determined by removing two distinct factors of st(∆), and conclude that H, and
hence its normal closure G, act trivially on AΓ .
Case 2: lk(u) = ∅ for every vertex u of ∆.
We write Γ = Γ1 ⊔· · ·⊔Γk ⊔Θ where the subgraphs Γi are non-discrete connected
components of Γ, and Θ is discrete. By assumption ∆ ⊆ Θ.
If k > 2, then removing any component Γi leaves us with a smaller graph, to
which we can apply the inductive hypothesis. Then we use Lemma 2.25.
If k = 0 then Θ is not isomorphic to Dn by assumption. Then we know that the
action φ is trivial by assumption (2).
If k = 1, then we need to look more closely at Γ1 . If Γ1 does not have factors
isomorphic to Dn in its join decomposition, then by induction we know that Γ1 is
trivialised. Now we use Lemma 2.25.
Suppose that Γ1 contains a subgraph Ω isomorphic to Dn in its join decomposition. If Γ1 has a non-trivial clique factor, then this factor is invariant, induction
tells us that it is trivialised, and the induced quotient action is also trivial. Thus
OUTER ACTIONS OF Out(Fn ) ON SMALL RAAGS
15
the entire action of H is trivial, thanks to Lemma 2.24, and thus the action of G is
trivial, as G is the normal closure of H.
If the clique factor is trivial, then taking links of different factors of the join
decomposition of Γ1 allows us to repeat the argument we just used, and conclude
that H, and thus G, act trivially.
Theorem
3.7. Let n > 6. Suppose that Γ is a simplicial graph with fewer than
1 n
vertices.
Let φ : SOut(Fn ) → Out(AΓ ) be a homomorphism. Then φ is trivial,
2 2
provided that there are no n vertices in Γ with identical stars, and that Γ is not a
join of the discrete graph with n vertices and another (possibly empty) graph.
Proof. We start by showing that G = SOut(Fn ) satisfiesthe assumptions (1)–(4)
of Theorem 3.4 and (5) of Proposition 3.6, with m = 12 n2 .
(1) Let H = Altn . The group G is the normal closure of H by Proposition 2.4.
(2) All homomorphisms
G → Out(Fk )
are trivial when k 6= n and k < m by Theorem 2.7.
(3) All homomorphisms
G → GLk (Z)
are trivial when k < n by Corollary 2.5.
(4) Any action of G on a set of cardinality smaller than m is trivial by Corollary 2.29.
(5) G is perfect by Proposition 2.2.
To verify property C we use Lemma 3.5, and property D we replace using Proposition 3.6. Now we apply Theorem 3.4.
4. From larger to smaller RAAGs
In this section we will look at homomorphisms Out(AΓ ) → Out(AΓ′ ), where Γ′
has fewer vertices than Γ.
Theorem 4.1. There are no injective homomorphisms Out(AΓ ) → Out(AΓ′ ) when
Γ′ has fewer vertices than Γ.
n
Proof. For a group G we define its Z2 -rank to be the largest n such that (Z2 )
embeds into G.
We claim that the Z2 -rank of Out(AΓ ) is equal to |Γ|, the number of vertices of
Γ.
Firstly, note that for every vertex of Γ we have the corresponding inversion in
Out(AΓ ), and these inversions commute; hence the Z2 -rank of Out(AΓ ) is at least
|Γ|.
For the upper bound, observe that the Z2 -rank of GLn (R) is equal to n, since
we can simultaneously diagonalise commuting involutions in GLn (R). Thus, the
Z2 -rank of GLn (Z) is equal to n as well (since it is easy to produce a subgroup of
this rank).
Finally, note that the kernel of the natural map Out(AΓ ) → GLn (Z) with n = |Γ|
is torsion free by Theorem 2.23, and so the Z2 -rank of GLn (Z) is bounded below
by the Z2 -rank of Out(AΓ ).
Remark 4.2. The proof of the above theorem works for many subgroups of Out(AΓ )
as well; specifically it applies to Out0 (AΓ ), the group of untwisted outer automorphisms U(AΓ ), and the intersection U0 (AΓ ) = U(AΓ ) ∩ Out0 (AΓ ).
It also works when the domain of the homomorphisms is Aut(AΓ ), or more
generally any group with Z2 -rank larger than the number of vertices of Γ′ .
16
DAWID KIELAK
References
[BP]
[BV1]
[BV2]
[CCV]
[CSV]
[Cul]
[Ger]
[GL]
[HK]
[Khr1]
[Khr2]
[Kie1]
[Kie2]
[LS]
[Lau]
[Men]
[PR]
[Ser]
[Toi]
[TW]
[Wad]
[Wei]
[Zim]
O.V. Bogopol’skiı̆ and D.V. Puga. On embeddings of Out(Fn ), the outer automorphism
group of the free group of rank n, into Out(Fm ) for m > n. Algebra and Logic 41(2002),
69–73.
Martin R. Bridson and Karen Vogtmann. Homomorphisms from automorphism groups of
free groups. Bull. London Math. Soc. 35(2003), 785–792.
Martin R. Bridson and Karen Vogtmann. Abelian covers of graphs and maps between outer
automorphism groups of free groups. Mathematische Annalen 353(2012), 1069–1102.
Ruth Charney, John Crisp, and Karen Vogtmann. Automorphisms of 2-dimensional rightangled Artin groups. Geom. Topol. 11(2007), 2227–2264.
Ruth Charney, Nathaniel Stambaugh, and Karen Vogtmann. Outer space for untwisted
automorphisms of right-angled Artin groups. arXiv:1212.4791.
Marc Culler. Finite groups of outer automorphisms of a free group. In Contributions to
group theory, volume 33 of Contemp. Math., pages 197–207. Amer. Math. Soc., Providence,
RI, 1984.
S. M. Gersten. A presentation for the special automorphism group of a free group. J. Pure
Appl. Algebra 33(1984), 269–279.
F. Grunewald and A. Lubotzky. Linear Representations of the Automorphism Group of a
Free Group. Geometric and Functional Analysis 18(June 2006), 1564–1608.
Sebastian Hensel and Dawid Kielak. Nielsen realisation for untwisted automorphisms of
right-angled Artin groups. arXiv:1410.1618v2.
D. G. Khramtsov. Finite groups of automorphisms of free groups. Mat. Zametki 38(1985),
386–392, 476.
D. G. Khramtsov. Outer automorphisms of free groups. In Group-theoretic investigations
(Russian), pages 95–127. Akad. Nauk SSSR Ural. Otdel., Sverdlovsk, 1990.
Dawid Kielak. Outer automorphism groups of free groups: linear and free representations.
J. London Math. Soc. 87(2013), 917–942.
Dawid Kielak. Low-dimensional free and linear representations of Out(F3 ). J. Group Theory 18(2015), 913–949.
Vicente Landazuri and Gary M. Seitz. On the minimal degrees of projective representations
of the finite Chevalley groups. J. Algebra 32(1974), 418–443.
Michael R. Laurence. A generating set for the automorphism group of a graph group. J.
London Math. Soc. (2) 52(1995), 318–334.
Jens L. Mennicke. Finite factor groups of the unimodular group. Ann. of Math. (2)
81(1965), 31–37.
A. Potapchik and A. Rapinchuk. Low-dimensional linear representations of Aut Fn , n ≥ 3.
Trans. Amer. Math. Soc. 352(2000), 1437–1451.
Herman Servatius. Automorphisms of graph groups. J. Algebra 126(1989), 34–60.
Emmanuel Toinet. Conjugacy p-separability of right-angled Artin groups and applications.
Groups Geom. Dyn. 7(2013), 751–790.
Victor Turchin and Thomas Willwacher. Hochschild–Pirashvili homology on suspensions
and representations of Out(Fn ). arXiv:1507.08483.
Richard D. Wade. Johnson homomorphisms and actions of higher-rank lattices on rightangled Artin groups. J. Lond. Math. Soc. (2) 88(2013), 860–882.
S.H. Weintraub. Representation theory of finite groups: algebra and arithmetic. Graduate
studies in mathematics. American Mathematical Society, 2003.
Bruno Zimmermann. Über Homöomorphismen n-dimensionaler Henkelkörper und endliche
Erweiterungen von Schottky-Gruppen. Comment. Math. Helv. 56(1981), 474–486.
Dawid Kielak
Fakultät für Mathematik
Universität Bielefeld
Postfach 100131
D-33501 Bielefeld
Germany
[email protected]
| 4math.GR
|
arXiv:1703.00635v1 [math.GR] 2 Mar 2017
Classification of finite Cθθ-groups with even order and its
application
Ali Mahmoudifar
Department of Mathematics, Tehran-North Branch, Islamic Azad University, Tehran, Iran
e-mail: [email protected]
Abstract
A finite group of order divisible by 3 in which centralizers of 3-elements are 3-subgroups
will be called a Cθθ-group. The prime graph (or Gruenberg-Kegel graph) of a finite group
G is denoted by Γ(G) (or GK(G)) and its a familiar. Also the degrees sequence of Γ(G) is
called the degree pattern of G and is denoted by D(G). In this paper, first we classify the
finite Cθθ-groups with even order. Then we show that there are infinitely many Cθθ-groups
with the same degree pattern. Finally, we proved that the simple group PSL(2, q) and the
almost simple group PGL(2, q), where q > 9 is a power of 3, are determined by their degree
pattern.
2000 AMS Subject Classification: 20D05, 20D60, 20D08.
Keywords : linear group, prime graph, degree pattern, Cθθ-group.
1
Introduction
If n is a natural number, then we denote by π(n), the set of all prime divisors of n. Throughout
this paper by G we mean a finite group. The set π(|G|) is denoted by π(G). Also the set of
element orders of G is denoted by πe (G). We denote by µ(S), the maximal numbers of πe (G)
under the divisibility relation. The prime graph (or Gruenberg-Kegel graph) of G is a simple
graph whose vertex set is π(G) and two distinct primes p and q are joined by an edge (and we
write p ∼ q), whenever G contains an element of order pq. The prime graph of G is denoted by
Γ(G). If S is a finite non-abelian simple group and G is a finite group such that S ≤ G ≤ Aut(S),
then we say that G is an almost simple group related to the simple group S.The degree pattern
of a finite group is defined as follows:
Definition 1.1. ([12]) Let G be a group and π(G) = {p1 , p2 , . . . , pk } where p1 < p2 < · · · < pk .
Then the degree pattern of G is defined as follows:
D(G) := (deg(p1 ), deg(p2 ), . . . , deg(pk )),
where deg(pi ), 1 ≤ i ≤ k, is the degree of vertex pi in Γ(G).
Following Higman [9], a finite group of order divisible by 3 in which centralizers of 3-elements
are 3-groups will be called a Cθθ-group. A finite group with a Sylow 3-subgroup M containing
the centralizer of each of its nonidentity elements will be called a 3CC-group. So by the above
notations, a finite group G is a Cθθ-group whenever deg(3) = 0 in Γ(G).
We say that a finite group H is determined by its degree pattern of its prime graph if and
only if for each finite group G such that D(G) = D(H), then G is isomorphic to H. It is
1
clear that the relation D(G) = D(H) does not imply that π(G) = π(H) (for example we have
D(S3 ) = D(D10 ) = (0, 0) but π(S3 ) = {2, 3} and π(D10 ) = {2, 5}). In this paper, first we
classify the finite Cθθ-groups with even order. Then we show that there are infinitely many
Cθθ-groups with the same degree pattern. Also we proved that the simple group PSL(2, q) and
the almost simple group PGL(2, q), where q > 9 is a power of 3, PSL(3, 4) and the simple group
PSL(2, p) where either (p − 1)/2 or (p + 1)/2 is a power of 3, are determined by their degree
pattern. Finally, we ask a question about this type determination of finite groups.
2
Preliminary Results
Lemma 2.1. [3, Theram A] Let G be a 3CC-group with an S3 -subgroup M . Then one of the
following statements is true:
(1) M E G;
(2) G has a normal nilpotent subgroup N such that G/N ∼
= NG (M ), M cyclic;
(3) G has a normal elementary abelian 2-subgroup N such that G/N ∼
= PSL(2, 2α ), α ≥ 2;
(4) G ∼
= PSL(2, q) for some q > 3;
(5) G ∼
= PSL(3, 4);
(6) G contains a simple normal subgroup PSL(2, 3n ), n > 1, of index 2.
Lemma 2.2. [10, 13] Let G be a Frobenius group of even order and let H and K be the Frobenius
complement and Frobenius kernel of G, respectively. Then the following assertions hold:
(a) K is a nilpotent group;
(b) |K| ≡ 1 (mod |H|);
(c) Every subgroup of H of order pq, with p, q (not necessarily distinct) primes, is cyclic.
In particular, every Sylow Subgroup of H of odd order is cyclic and a 2-Sylow subgroup of H
is either cyclic or a generalized quaternion group. If H is a non-solvable group, then H has a
subgroup of index at most 2 isomorphic to Z × SL(2, 5), where Z has cyclic Sylow p-subgroups
and π(Z) ∩ {2, 3, 5} = ∅. If H is solvable and O(H) = 1, then either H is a 2-group or H has a
subgroup of index at most 2 isomorphic to SL(2, 3).
Lemma 2.3. [6] Let G be a 2-Frobenius group, i.e., G has a normal series 1 ⊳ H ⊳ K ⊳ G, such
that K and G/H are Frobenius groups with kernels H and K/H, respectively. If G has even
order, then
(i) G/K and K/H are cyclic, |G/K| |Aut(K/H)| and (|G/K|, |K/H|) = 1;
(ii) H is a nilpotent group and G is a solvable group.
Lemma 2.4. [5, 1] If p is an odd prime number, then µ(P GL(2, pk )) = {p, pk − 1, pk + 1} and
µ(P SL(2, pk )) = {p, (pk − 1)/2, (pk + 1)/2}. Also we have µ(P GL(2, 2k )) = µ(P SL(2, 2k )) =
{2, 2k − 1, 2k + 1}.
Lemma 2.5. [8] The equation pm − q n = 1, where p and q are primes and m, n > 1 has only
one solution, namely 32 − 23 = 1.
Lemma 2.6. [8] With the exceptions of the relations (239)2 − 2(13)4 = −1 and 35 − 2(11)2 = 1
every solution of the equation
pm − 2q n = ±1; p, q prime; m, n > 1
has exponents m = n = 2; i.e. it comes from a unit p − q · 21/2 of the quadratic field Q(21/2 ) for
which the coefficients p and q are primes.
2
3
Main Results
Lemma 3.1. Let G be an almost simple group related to the simple group P SL(2, pn ) where p is
a prime number and pn > 3. If deg(p) = 0 in the prime graph of G, then either G ∼
= PSL(2, pn )
or G ∼
= PGL(2, pn ).
Proof. By [11], we know that Out(PSL(2, pn )) ∼
= hδi × hσi, where δ
with order 2 and σ is field automorphism of PSL(2, pn ) with order
automorphisms acts on SL(2, pn ) as follows:
σ
δ
a b
a αb
a b
=
&
=
c d
αc d
c d
is a diagonal automorphism
n. Also we know that these
ap bp
cp dp
,
where α is the involution in the multiplicative group of the field GF (pn ), i.e. GF (pn )# . So if in
the above relations, we put b = c = 0 and a = d such that the order of a in GF (pn )# is p, then
we deduce that any field automorphism σ f and any diagonal-field automorphism δσ f such that
1 ≤ f ≤ n has a fixed point in PSL(2, pn ) with order p.
The above discussion implies that when G is an almost simple group related to PSL(2, pn ) in
which deg(p) = 0 in its prime graph, then G is an extension of this simple group by a diagonal
automorphism of PSL(2, pn ). Therefore, G is isomorphic to PGL(2, pn ). Finally, by Lemma 3.1,
we get the result.
Lemma 3.2. Let G = K : C be a Frobenius group with kernel K and complement C. If C is
solvable, then the prime graph of C is a complete graph.
Proof. First, let p and q be two different prime numbers included in π(C). Also let p and q are
not adjacent in Γ(C). Since C is solvable, C has a Hall {p, q}-subgroup H. Without lose of
generality, we assume that Op (H) 6= 1. So if Hq is a Sylow q-subgroup of H, then Op (H) : Hq is
a Frobenius subgroup with kernel Op (H). This implies that K : (Op (H) : Hq ) is a 2-Frobenius
subgroup of G. So by Lemma 2.3, we get that Hq has a fixed point in K which is a contradiction
since Hq is a subgroup of complement C.
Lemma 3.3. Let G = K : C be a Frobenius group with kernel K and complement C. If either
3 | |K| or deg(3) = 0 in Γ(G), then G is solvable and Γ(G) has two connected components π(K)
and π(C) which are complete.
Proof. By Lemma 2.2, it is sufficient to show that Γ(C) is a complete graph. First, let 3 | |K|.
Thus by Lemma 2.2, we get that C is solvable and so by Lemma 3.2, Γ(C) is complete.
Now let deg(3) = 0 in Γ(G) and 3 ∤ |K|. If C has even order, then Z(C) 6= 1 which is
impossible since deg(3) = 0 in Γ(G). Hence C has odd order and so C is solvable and by
Lemma 3.2, we get the result.
Lemma 3.4. Let G be a 2-Frobenius group with normal series 1 ⊳ H ⊳ K ⊳ G, such that K
and G/H are Frobenius groups with kernels H and K/H, respectively. Then the connected
components of the prime graph of G are complete.
Proof. Since H is nilpotent and G/K and K/H are cyclic groups, we may assume that K is a
p-subgroup and G/K is a q-subgroup of G such that p and q are some prime divisors of |G|.
Hence by Lemma 2.3, we get the result.
Lemma 3.5. Let G be a finite group with even order. If 3 | |G| and deg(3) = 0 in the prime
graph of G, then G is a 3CC-group.
3
Proof. Since deg(3) = 0, so by the definition G is a Cθθ-group. Also since G has an even order,
by [3], it follows that G is 3CC-group.
Theorem 3.6. Let G be a finite group with even order. If 3 is an isolated vertex in the prime
graph of G and D(G) is the degree pattern of G, then:
a) If G is solvable, then one of the following cases holds:
a-1) G = K : C is a Frobenius group with kernel K and complement C. In this case one of
the subgroups K or C is a 3-subgroup of G.
a-2) G is a 2-Frobenius group with normal series 1⊳ H ⊳ K ⊳ G such that N is a 3′ -subgroup,
K is a cyclic 3-subgroup and C ∼
= Z2 is cyclic.
Moreover, in these two cases, if we put n = |π(G)|, then we have
D(G) = (n − 2, 0, n − 2, . . . , n − 2)
b) if G is non-solvable, then one of the following cases holds:
b-1) has a normal elementary abelian 2-subgroup N such that G/N is isomorphic to either
P SL(2, 4) or P SL(2, 8) and we have
D(P SL(2, 4)) = D(P SL(2, 8)) = (0, 0, 0)
b-2) G is isomorphic the simple group P SL(3, 4) and we have
D(P SL(3, 4)) = (0, 0, 0, 0),
b-3) G is isomorphic the simple group P SL(2, p) where p is a prime number and either
(p + 1)/2 or (p − 1)/2 is a power of 3. In this case, if we put n = |π(P SL(2, p))|, then we have
D(P SL(2, p)) = (n − 3, 0, n − 3, · · · , n − 3, 0),
b-4) G is isomorphic the simple group P SL(2, 3n ) where n > 1. In this case, if we put
n1 = |π((3n − 1)/2)| and n2 = |π((3n + 1)/2)|, then we have
D(P SL(2, 3n )) = (n1 − 2 or n2 − 2, 0, n1 − 2 or n2 − 2, . . . , n1 − 2 or n2 − 2),
b-5) G is isomorphic to the almost simple group P GL(2, 3n ), where n > 1. In this case, if
we put n1 = |π(3n − 1)| and n2 = |π(3n + 1)|, then we have
D(P GL(2, 3n )) = (n1 + n2 − 2, 0, n1 − 1 or n2 − 1, . . . , n1 − 1 or n2 − 1).
Proof. First we note that since we assume that 3 is an isolated vertex in the prime graph of G,
so G is a Cθθ-group. Also since G has an even order, then by Lemma 2.1, one of the following
assertions holds:
Case 1: If M is a Sylow 3-subgroup of G, then M is normal. Hence by Schur-Zassenhaus’s
Theorem, M has a complement H and so G = M : H in which H is a Hall 3′ -subgroup of G.
Since 3 is an isolated vertex in the prime graph of G, we get that H acts fixed point freely on M .
This implies that G is a Frobenius group with kernel M and complement H. So by Lemma 3.3,
we get (a-1).
Case 2: G has a normal nilpotent subgroup N such that G/N ∼
= NG (M ), where M is a
Sylow 3-subgroup of G and M is cyclic.
Subcase 2-1: If M = NG (M ), then G = N : M is a Frobenius group with kernel N . Also
in this case, since N is nilpotent and M is a 3-subgroup, then G is solvable and again we get
(a-1).
4
Subcase 2-2: If M is a pure subgroup of NG (M ), then NG (M ) = M : C is a Frobenius
group with kernel M . Since M is cyclic, then C ∼
= Z2 . In this case, G is a 2-Frobenius group
with normal series 1 E N E N : M E G which implies (a-2).
Case 3: G has a normal elementary abelian 2-subgroup N such that G/N ∼
= P SL(2, 2α ),
α
α
α
where α ≥ 2. By Lemma 2.4, we have µ(P SL(2, 2 )) = {2, 2 − 1, 2 − 1}. So either 2α − 1 or
2α + 1 is a power of 3. Thus by Lemma 2.5, the only possible cases are either α = 2 or α = 3.
This asserts that G/N is isomorphic to either P SL(2, 4) or P SL(2, 8), which gets (b-1).
Case 4: G ∼
= P SL(2, q) where q is a power of odd prime number p > 3. In this case, by
Lemma 2.4, µ(P SL(2, q)) = {p, q − 1, q + 1}. Since 3 is an isolated vertex in the prime graph
of PSL(2, q), we get that either (q − 1)/2 or (q + 1)/2 is a power of 3. Now by Lemma 2.6, it
follows that q = p. So either p − 1 or p + 1 is a power of 3, Case (b-3).
Case 5: G is isomorphic to P SL(2, 3n ) : hθi, where θ is an involution. This means that G
is an almost simple group related to PSL(2, 3n ) such that deg(3) = 0 in the prime graph of G.
So by Lemma 3.1, we conclude that G ∼
= P GL(2, 3n ), Case (b-5).
Case 6: G is isomorphic to P SL(2, 3n ) where n > 1 or P SL(3, 4). By Lemma 2.4 and [7],
we get that the vertex 3 is an isolated vertex in the prime graph of these simple groups, Cases
(b-2) and (b-4).
By [7] and Lemma 2.4, we can easily compute the degree pattern of the above groups, which
completes the proof.
We remark that the above theorem does not show that the finite groups explained in that
theorem exist. it is obvious that when p is a prime number such that p ∈ {5, 7, 17, 19, 53, 163},
the simple group P SL(2, p) satisfies the condition of Case (b-3) of Theorem 3.6. However, it is
not clear that the number of these finite simple groups is finite or infinite. In the following, we
show that there is infinitely many solvable groups with even order such that deg(3) = 0 in their
prime graph.
Lemma 3.7. There exist infinitely many Frobenius groups of even order whose Fitting subgroups
are 3-subgroups.
Proof. Let F be an arbitrary field of characteristic 3. We know that the multiplicative group
F # has a unique involution γ. Also we know that F # acts fixed point freely on the additive
group F . So the semidirect product of F by the subgroup generated by γ is a Frobenius group
of even order whose Fitting subgroup is a 3-subgroup.
similarly to the previous lemma we get the following result:
Lemma 3.8. There exist infinitely many Frobenius groups of even order whose complement is
a cyclic 3-subgroup.
Lemma 3.9. There exist infinitely many 2-Frobenius groups G with normal series 1EN EH EG
such that H/N is a 3-group and deg(3) = 0 in Γ(G).
Proof. Let F be the Galois Field GF (22 ) = {0, 1, α, α + 1}, where α2 = 1. Similarly to the
previous lemma, we can construct a Frobenius group with Fitting subgroup F and cyclic component F # = {1, α, α + 1}. If we put C := {0, 1}, then C is a subgroup of Galois Filed F . Now
we define an action of C on F # as follows if x ∈ C and y ∈ F # \ {1}, then y x := x + y. This
action shows that the semidirect product group G := F # : C is a Frobenius group with Fitting
subgroup F # .
L L L
Now let V = F
F
· · · F be a vector space with dimension n over the field F . So G
acts on the additive group V as follows if ξ1 + ξ2 + · · · + ξn ∈ V and yx ∈ G, where y ∈ F # and
5
x ∈ C, then (ξ1 + ξ2 + · · · + ξn )yx := (ξ1 y) + (ξ2 y + · · · + (ξn y). This implies that V : F # is a
Frobenius group with fitting subgroup V . So the finite group V : G is a 2-Frobenius group in
which 3 is an isolated vertex in its prime graph.
Corollary 3.10. Let G be finite solvable group with even order such that deg(3) = 0 in the
prime graph of G. If H is a finite group such that D(H) = D(G), then H is solvable.
Proof. By Theorem 3.6, it is straightforward.
The next lemma shows that there are infinitely manu non-solvable groups such that deg(3) =
0 in their prime graph.
Lemma 3.11. Let L be a finite group isomorphic to one of the simple groups P SL(2, 4) or
P SL(2, 8). Then there exist infinitely many groups G such that deg(3) = 0 in Γ(G) and G/O2 (G)
is isomorphic to L.
Proof. Let L ∼
= P SL(2, 4)
2
GF (2 ) as follows:
a
G := c
0
∼
= SL(2, 4). there is a modular representation of L over Galois Field
b u
d v | where ad − bc = 1 and u, v ∈ GF (22 )
o 1
L
So L acts fixed point freely on the vector space V := GF (22 ) GF (22 ). Hence the finite group
V : L is a group in which 3 is an L
isolated
we similarly, we can
L vertex
L in its prime graph. Now ∼
construct a finite group G := (V
V
· · · V ) : L such that G/O2 (G) = L and deg(3) = 0
in Γ(G). With a similar argument we can make such a finite group when L ∼
= P SL(2, 8)
Corollary 3.12. Let G be finite non-solvable group such that deg(3) = 0 in the prime graph of
G. If H is a finite group such that D(H) = D(G), then one of the following cases holds:
1) H has a normal 2-subgroup N , such that H/N is isomorphic to either PSL(2, 4) or
PSL(2, 8). Moreover, in this case we have D(H) = (0, 0, 0).
2) H is isomorphic to G.
Proof. Using the degree pattern of groups in Theorem 3.6, it is straightforward.
Corollary 3.13. If G is a finite group such that D(G) = D(PGL(2, 3k )), where k > 2 is a
natural number, then G ∼
= PGL(2, 3k ).
Corollary 3.14. If G is a finite group such that Γ(G) = Γ(PGL(2, 3k )), where k > 2 is a
natural number, then G ∼
= PGL(2, 3k ).
References
[1] Z. Akhlaghi, M. Khatami and B. Khosravi, Characterization by prime graph of PGL(2, pk )
where p and k are odd, International Journal of Algebra and Computation 20 (7) (2010)
847–873.
[2] Arad, Zvi, and David Chillag. ”On finite groups with conditions on the centralizers of
p-elements.” Journal of Algebra 51.1 (1978): 164–172.
6
[3] Z. ARAD, A classification of 3CC-groups and application to Glauberman-Goldschmidt
theorem, J. Algebra 43 (1976), 176-180.
[4] Z. ARAD AND M. HERZOG, On fundamental subgroups of order divisible by three, Houston J. Math. 3 (1977), 309-313.
[5] G. Y. Chen, V. D. Mazurov, W. J. Shi, A. V. Vasilev and A. Kh. Zhurtov, Recognition of
the finite almost simple groups PGL2(q) by their spectrum, J. Group Theory 10(1) (2007)
7185.
[6] G. Y. Chen, On structure of Frobenius and 2-Frobenius group, Jornal of Southwest China
Normal University, 20(5), 485-487(1995).(in Chinese)
[7] J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker and R. A. Wilson, Atlas of Finite
Groups (Oxford University Press, Oxford, 1985).
[8] P. Crescenzo, A diophantine equation which arises in the theory of finite groups, Adv. Math.
17(1) (1975) 2529.
[9] G. HIGMAN, Odd characterization of finite simple groups, lecture notes, University of
Michigan, 1968.
[10] D. Gorenstein, Finite Groups, New York, Harpar and Row, (1980).
[11] Kleidman, P. and Liebeck, M., The subgroup structure of finite classical groups, Cambridge
University Press, 1990.
[12] R. Kogani-Moghadam and A. R. Moghaddamfar, Groups with the same order and degree
pattern, Science China Mathematics, 55(4), (2012), 701–720.
[13] D. S. Passman, Permutation Groups, New York, Benjamin Inc., (1968).
[14] K. Zsigmondy, Zur theorie der potenzreste, Monatsh. Math. Phys. 3 (1892) 265–284.
7
| 4math.GR
|
Localization of Networked Robot Systems Subject
to Random Delay and Packet Loss
Manh Duong Phung, Thi Thanh Van Nguyen, Thuan Hoang Tran, and Quang Vinh Tran
Abstract— This paper deals with the localization problem of
mobile robot subject to communication delay and packet loss.
The delay and loss may appear in a random fashion in both
control inputs and observation measurements. A unified statespace representation is constructed to describe these mixed
uncertainties. Based on it, the optimal linear estimator is
developed. The main idea is the derivation of a relevance factor
to incorporate delayed measurements to the being estimate.
The estimator is then extended for nonlinear systems. The
performance of this method is tested within the simulations in
MATLAB and the experiments in a real robot system. The
good localization results prove the efficiency of the method for
the purpose of localization of networked mobile robot.
Index Terms—Networked robot systems, robot localization,
Kalman filter, random delay, packet loss.
I. INTRODUCTION
Networked robot systems (NRSs) have gained the
research interest recently due to its ability to support a wide
range of applications including but not limited to
telemedicine, telehomecare, virtual laboratory, and disaster
rescue. A number of projects have been proposed to deal
with the problems involved in the development of networked
robots. Some works focus on hardware configuration and
control architecture for specific applications [1]–[4]. The
others deal with the navigation and transport protocols [5]–
[8].
In this paper, the problem of localization is investigated.
Differing from classic robots, localization of NRSs faces
new challenges introducing by the network such as the
inevitable communication delays, the out-of-sequence data
arriving, the limited available bandwidth, and the partial
intermittent observations. In the literature, several
approaches have been proposed to cope with those changes.
In [9], four cameras are set in the robot field as external
visual sensors to serve the localization and navigation task.
The cameras are connected to the Internet and fixed on roof
to formulate four adjacent grids of vision without the dead
zone. A recognition algorithm is implemented to recognize
and give out the relative location of the robot, target,
landmark and obstacle symbols. In [10], the robot pose is
M. D. Phung, T. T. V. Nguyen, T. H. Tran, and Q. V. Tran are with the
Department of Electronics and Computer Engineering, VNU University of
Engineering and Technology, Vietnam National University, Hanoi, Vietnam
(Phone: +84437549272; email: [email protected]).
estimated at the local site using odometry, sonar and
compass sensors. The information is then transmitted to the
remote site as the robot pose at the receiving time without
considering the change of the robot during the
communication time. A map-based localization method for
Internet-based personal robot system is introduced in [11].
The absolute position of the robot is determined by
comparing a reference map of the local site with the one built
by a map building technique at the remote site.
It is recognizable from the proposed studies that the data
transmission between the remote controller and the actuator
was treated as a given condition rather than being modeled
and analyzed to provide more theoretic approaches. In order
to overcome this, several works have been proposed to
address the communication problems such as the H filter
for systems with random delays, the optimal filter with multi
packet dropout, or the adaptive Kalman filter with random
sensor delays, multiple packet dropouts, and missing
measurements [12]–[15]. Though these works are efficient
for NRSs, the systems have to be linear. Further
modifications are required to implement them to nonlinear
systems such as mobile robot.
In this work, we develop a state estimator for the problem
of robot localization subject to random delay and packet loss.
An augmented filter called past observation-based extended
Kalman filter (PO-EKF) inspired by the well-known optimal
filter, the Kalman filter, is introduced as the compensator and
estimator. The main idea behind the filter is the
determination of a “relevance factor” which describes the
relevance of observations from the past to the present. This
factor is employed as a multiplier to enable the incorporation
of delayed measurements to the posteriori estimation.
Simulations have been carried out in MATLAB and
experiments have been implemented in a real NRS. The
results confirm the effectiveness of the proposed approach.
II. SYSTEM MODELING AND PROBLEM FORMULATION
Consider the following discrete time state-space robot
system:
x k 1 f (x k , u k , w k )
(1)
z k h( x k , v k )
(2)
where x [ x y ] is the state vector described the robot’s
pose (position and orientation), zk is the measured output, f
T
and h are the system functions, and wk and vk are
independent, zero-mean, white noise processes with normal
probability distributions: w k ~ N(0, Qk ) ; v k ~ N(0, Rk );
E (w i v jT ) 0 .
When distributing over communication networks, the
system is decentralized and its functioning operation depends
on a number of network parameters such as delay, loss, and
out of order. The networks are in general very complex and
can greatly differ in their architecture and implementation. In
this work, a network is modeled as a module between the
plant and controller which delivers input signals and
observation measurements with possible delay and loss. The
delay is assumed to be random but measurable. The packet
loss is modeled as a binary random variable k defined as
follows:
1, if a packet arrives at time k
k
0, otherwise
(3)
Let n be the time delay (in number of sampling periods)
between the controller and the actuator, m be the time delay
between the sensor and the controller, kca be the binary
random variable described the arrival of inputs from the
controller to the actuator, ksc be the binary random variable
described the arrival of measurements from the sensor to the
controller. The NRS and its signal timing can be described as
in Fig. 1.
correction. In each phase, it propagates an estimate value and
an error covariance reflecting its belief to the estimate. When
operating over the network, both inputs and observations
suffer from the communication delay and loss. These data
cannot be fused using the standard Kalman filter but require
some modifications in the structure of the filter. In this
section, we first derive the optimal filter for the linear NRS.
The optimality is in sense that it minimizes the covariance of
the estimation error. We then extend the filter for the
nonlinear case.
A. Optimal filter for linear NRSs
Consider linear NRSs. Equation (4)–(5) can be rewritten
as:
x k Ak 1x k 1 kca n 1 Bk 1u k n 1 w k 1
Ak 1x k 1 Bk 1u k n 1 w k 1
(6)
z ik ksc m z k m ksc m H k m x k m ksc m v k m
H x v
(7)
i
i
i
where u k , z ik , H i, v i , and i are defined by the above equations.
The optimal filter for system (6)–(7) is derived as follows.
Priori State Estimate: The priori estimate, xˆ k , is defined
as the expectation of the state x k given all measurements up
to and including the last step k-1. From (6), it is given by:
xˆ k E (x k ) Ak 1 E (x k 1 ) Bk 1 E (u k n 1 ) E (w k 1 ) (8)
As E ( x k 1 ) is the posteriori state estimate at time k-1, u k n 1
is a known input, and w k 1 is zero-mean, (8) becomes:
xˆ k Ak 1xˆ k 1 Bk 1u k n 1
(9)
Priori Error Covariance: Let e k and e k be the priori
and posteriori estimate errors, respectively:
e k x k xˆ k
e k x k xˆ k
Fig. 1. Model of networked robot systems with signal timing
From Fig. 1, the plant at time k is driven by the input at
time k-n-1. The measurement received at the controller at
time k is taken at time k-m. Thus, the system is time-varying
and can be rewritten as:
From (6) and (9), the covariance of the priori estimate error
is given by:
Pk E (e k ek T )
x k f (x k 1 , kca n 1u k n 1 , w k 1 )
(4)
E ( Ak 1ek 1ek T1 AkT1 w k 1wTk 1
z k ksc m z k m ksc m h(x k m , v k m )
(5)
Ak 1ek1wTk 1 w k 1ek T1 AkT1 )
The localization is the problem of state estimation of
system (4)–(5). Our approach is the development of an
optimal filter based on Kalman filter’s theory [16].
III. LOCALIZATION WITH PAST OBSERVATION-BASED
EXTENDED KALMAN FILTER
The Kalman filter is considered as one of the most
efficient methods for mobile robot localization. It estimates
the robot’s state through two phases: prediction and
(10)
(11)
(12)
Ak 1 Pk1 AkT1 Qk 1
Posteriori State Estimate: From (7) the measurement z ik
received at time k would update the system state at a
previous time i rather than the present time k. Due to the
delay, this measurement could not reach the estimator until
time k. We therefore construct the data update equation of
the form:
xˆ k xˆ k Kk (z ik H i xˆ i )
(13)
and recompute the Kalman gain and error covariance to
ensure the optimality of the new data update equation.
Kalman gain and Posteriori Error Covariance:
Assume that the measurement is fused using (13) with an
arbitrary gain Kk. The covariance of the posteriori estimate
error, Pk, is determined as:
k
L E (ei ek T ) Pi F T
Substituting (22) in (18) and (17) yields:
and
T
k k
P E (e e )
E[e k e kT e k eiT ( K k H i )T e k v Ti K kT K k H i ei e k T
K H e e T ( K H )T K H e v T K T K v e T
k
i i
i
k
i
k
i i
i
k
k
(14)
Pk E (e k e kT ) E (e k eiT )( K k H i )T K k H i E (ei e k T )
K H E (e e T )( K H )T K E ( v v T ) K T ]
k
i
k
i
i
(24)
(15)
k
Pk LT H iT K kT K k H i L K k H i Pi H iT K kT K k Ri K kT
(25)
where Ki* is the Kalman gain at time i of the standard
Kalman filter. It turns out that the past residual
(z k H i xˆ i ) in (13) can be normally updated to the
posteriori estimate at time k as at time i but the Kalman gain needs to be changed by the factor F . This
factor describes the relevant of the measurement updated at time i to the state at time k.
Due to the independence between e and v , (14) can be
simplified to:
i
K k FPi H iT [ H i Pi H iT Ri ]1
K k FKi*
i
(23)
i k
K k v e ( K k H i )T K k v i v Ti K kT ]
i
Pk Pk K k H i Pi F T
Remark 1: Equation (24) can be rewritten as:
T
i i
k
(22)
Remark 2: (13) can be rewritten as:
where L E(eiek T ).
As the matrix Kk is chosen to be the gain or blending
factor that minimizes the posteriori error covariance, this
minimization is accomplished by taking the derivative of the
trace of the posteriori error covariance with respect to Kk,
setting that result equal to zero, and then solving for Kk.
Applying this process to (15) obtains:
It implies that if a measurement is not arrived
(i 0), the estimator does not use any information
of the “dummy” observation to the estimate. It simply
sets the posteriori estimate to the value of the priori
estimate.
tr ( Pk )
2( LT H iT K k H i Pi H iT K k Ri ) 0 (16)
K k
Remark 3: When the delays n and m are zero, the new
filter reduces to the standard form of the Kalman filter.
K k LT H iT [ H i Pi H iT Ri ]1
(17)
Inserting (17) in (15) leads to a simpler form of Pk:
Pk Pk K k H i L
(18)
In order to compute L, the priori state estimate at time k
needs determining from the estimate at time i. Through the
time update (9) and the data update (13), e becomes:
e k x k xˆ k
Ak 1e k 1 w k 1
(19)
Ak 1[( I K k 1 H k 1 )e k 1 K k 1 v k 1 ] w k 1
After m updating steps, the estimation error becomes:
where
e k Fei 1 (w i ,..., w k 1 ) 2 ( v i ,..., v k 1 )
(26)
B. Optimal filter for nonlinear NRSs
Though the filter derived in previous section is capable
for NRSs, the system has to be linear. As practical robot
system is often nonlinear, further modification needs to be
accomplished. In this section, the derivation of the extended
Kalman filter is inherited to extend the derived filter for
nonlinear systems. The main idea is the linearization of a
nonlinear system around its previous estimated states.
Performing a Taylor series expansion of the state
equation around (xˆ k 1 , u k n1 ,0) gives:
x k f (xˆ k 1 , u k n 1 , 0)
f
w
( xˆ k1 , u k n1 ,0)
f
(x k 1 xˆ k 1 )
x ( xˆ k 1 ,uk n1 ,0)
w k 1
(20)
f (xˆ k 1 , u k n 1 , 0) Ak 1 (x k 1 xˆ k 1 ) Wk 1w k 1
(21)
Ak 1x k 1 u
m
F Ak j ( I K k j H k j )
xˆ k xˆ k i K k (z ik H i xˆ i )
Ak 1x k 1 [ f (xˆ
k 1
*
k n 1
k 1 k 1
, u k n 1 , 0) A xˆ
w
(27)
] Wk 1w k 1
*
k 1
j 1
and 1 and 2 are the functions of noise sequences w and v .
From (20) and the independence between e and noise
sequences, the covariance L becomes:
where Ak 1, Wk 1 are defined by the above equation. Similarly,
the measurement equation is linearized around (xˆ i ,0) to
obtain
z ik i [h(xˆ i , 0)
h
h
vi ]
(xi xˆ i )
x ( xˆ i ,0)
v ( xˆ i ,0)
h (xˆ i , 0) H i (xi xˆ i ) Vi v i
H x [h (xˆ , 0) H xˆ ] V v
i
i
i
i
i
i
2 (k ) 0
Qk R
L2 (k )
0
(28)
i
H i xi ε*i v *i
where hi, H i, Vi are defined by the above equation. The
system (27) and the measurement (28) now become linear
and the proposed filter can be applied to obtain the new filter
for nonlinear NRSs as follows:
The time update equations at prediction phase:
xˆ k f ( xˆ k 1 , u k n 1 , 0)
(29)
Pk Ak 1 Pk1 AkT1 Wk 1Qk 1WkT1
(30)
where is a constant with the value 0.01 determined by
experiments. In the simulations, it is supposed that the robot
has a sensor system that can measure the robot position and
orientation in the motion plane. The measurements suffer
from a Gaussian noise with zero mean and the covariance:
0
0.01 0
0
Rk = 0 0.01
0
0 0.018
F Ak j ( I K k j H k j )
(31)
K k FPi H iT ( H i Pi H iT Vi RiVi T ) 1
(32)
xˆ k xˆ k K k [ z ik h ( xˆ i , 0)]
(33)
P P K k H i Pi F
(34)
Ak 1
f k
x
( xˆ k , u k , 0 )
j 1
k
k
T
We call this filter past observation-based extended Kalman
filter (PO-EKF).
IV. SIMULATIONS
In order to evaluate the efficiency of the PO-EKF for the
localization of NRSs, simulations have been carried out in
MATLAB.
A. Simulation Setup
Parameters for simulations are extracted from the real
NRS described in the next section. The robot is a two
wheeled, differential-drive mobile robot with the kinematics
given by:
R
Ts (L (k ) R (k )) cos k
2
R
yk 1 yk Ts (L (k ) R (k )) sin k
2
R
k 1 k Ts (L (k ) R (k ))
L
xk 1 xk
(35)
where Ts is the sampling period, R is the radius of driven
wheels, L is the distance between the wheels, and L (k ) and
R (k ) are the rotational speeds of the left and right wheels,
respectively. The input noise is modeled as being
proportional to the angular speed L ( k ) and R ( k ) of the left
and right wheels, respectively. The covariance matrix Qk is
defined as:
(37)
Remaining parameters are retrieved from the kinematics
of the robot as follows:
The data update equations at correction phase:
m
(36)
Wk 1
f k
w
( xˆ k , u k , 0 )
1 0 Ts vc sin ˆk
0 1 Ts vc cos ˆk
1
0 0
(38)
cos ˆk
R
Ts sin ˆk
2
2
L
(39)
H k Vk I
cos ˆk
sin ˆk
2
L
(40)
In simulations, the performance of the PO-EKF is
compared with the EKF and the optimal filter proposed in
[15]. The EKF is implemented with the assumption that it
does not know if a measurement is delayed or not. It
incorporates every received measurement as there is no
delay. The optimal filter in [15] is called LEKF in this work.
B. Simulation Results
The first simulation is conducted with network
parameters as follows: the time delay is between 100ms and
800ms, and the loss rate is 1%. These values are determined
based on experimental measurements of the Internet (as will
be described in next section). Fig. 2a gives the tracking
performance of the filters in motion plane. Fig. 2b, c, d give
the comparison curves of the root mean square errors
(RMSEs) simulated by 100-times Monte Carlo tests for
filters. We see that EKF has the worst accuracy while the
PO-EKF and LEKF introduce similar property. Table I
shows the amount of floating point operations and the
execution time of the filters, scaled with respect to the EKF.
The PO-EKF is around two times higher than the EKF but
hundred times smaller than the LEKF.
EKF
LEKF
PO-EKF
True value
10
5
0
2
0.1
0.05
0
5
10
0
1
0.2
EKF
LEKF
PO-EKF
0.15
MSE in Y(m )
0.15
RMSE in X(m)
Y(m)
15
0.2
EKF
LEKF
PO-EKF
X(m)
0.1
0.05
0
200
400 600
Time(100ms)
800
0
10
0
200
400 600 800
Time(100ms)
MSE in Theta(rad)
0.2
20
0.15
0.1
0.05
0
10
EKF
LEKF
PO-EKF
0
200
400 600
Time(100ms)
800
100
Fig. 2. Comparison between our filter (PO-EKF) and the EKF and the LEKF with m n [1,8], p ( ca 0) p ( sc 0) 0.01
(a) Trajectories in motion plan; (b) RMSE in X direction; (c) RMSE in Y direction; (d) RMSE in orientation
EKF
LEKF
PO-EKF
True value
12
10
8
6
0.25
0.2
0.15
0.1
10
20
X(m)
30
0
4
0
500
1000 1500 2000
Time(100ms)
0.3
0.2
0
25
EKF
LEKF
PO-EKF
0.3
0.1
0.05
4
0
EKF
LEKF
PO-EKF
0.4
RMSE in Y(m)
Y(m)
14
0.3
RMSE in X(m)
16
0.35
0.5
EKF
LEKF
PO-EKF
RMSE in Theta(rad)
0.35
18
0.25
0.2
0.15
0.1
0.05
0
500
1000 1500 2000
Time(100ms)
250
0
0
500
1000 1500 2000
Time(100ms)
25
0.2
1.5
0
0.1
1
0.5
EKF
LEKF
PO-EKF
No-delay EKF
0
-0.5
0
2
4
X(m)
6
-0.1
-0.2
EKF
LEKF
PO-EKF
-0.3
8
-0.4
0
50
100
Time(100ms)
0.1
0.05
0
-0.1
EKF
LEKF
PO-EKF
-0.2
15
-0.3
0
50
100
Time(100ms)
Error in Theta(rad)
0.1
Error in Y(m)
2
Error in X(m)
Y(m)
Fig. 3. Comparison between our filter (PO-EKF) and the EKF and the LEKF with m n [8,15], p ( ca 0) p ( sc 0) 0.1
(a) Trajectories in motion plan; (b) RMSEs in X direction; (c) RMSEs in Y direction; (d) RMSEs in orientation
0
-0.05
-0.1
EKF
LEKF
PO-EKF
-0.15
-0.2
-0.25
15
0
50
100
Time(100ms)
15
Fig. 4. Comparison between our filter (PO-EKF) and the EKF and the LEKF with local configuration: m n [3, 5] , p ( ca 0) p ( sc 0) 0.015
(a) Trajectories in motion plan; (b) Errors in X direction; (c) Errors in Y direction; (d) Errors in orientation
0.1
0
0.5
EKF
LEKF
PO-EKF
No-delay EKF
0
0
2
4
X(m)
6
8
-0.1
-0.2
-0.3
-0.4
EKF
LEKF
PO-EKF
-0.5
-0.6
0.2
0.1
0.2
Error in Y(m)
1
Error in X(m)
Y(m)
1.5
-0.5
0.4
0
50
100
Time(100ms)
0
-0.2
EKF
LEKF
PO-EKF
-0.4
150
Error in Theta(rad)
2
-0.6
0
50
100
Time(100ms)
0
-0.1
-0.2
-0.3
EKF
LEKF
PO-EKF
-0.4
150
-0.5
0
50
100
Time(100ms)
Fig. 5. Comparison between our filter (PO-EKF) and the EKF and the LEKF with VPN configuration: m n [6,8], p ( ca 0) p ( sc 0) 0.02
(a) Trajectories in motion plan; (b) Errors in X direction; (c) Errors in Y direction; (d) Errors in orientation
15
TABLE I. NORMALIZED COMPUTATIONAL BURDEN OF FILTERS
EKF
LEKF
PO-EKF
Floating point operations
1.0
478.7
2.4
Execution time
1.0
532.3
2.0
In the next simulation, an extreme scenario is considered
in which the time delay is between 800ms and 1500ms and
the loss rate is 10%. The robot follows a sinusoidal path. The
LEKF uses a finite buffer with 50 slots. Fig. 2 shows the
tracking performance and the RMSEs. Table II shows the
computational burden of filters. We see that the PO-EKF has
better accuracy than the EKF and the same accuracy as the
LEKF at steady state. Though the LEKF reduces the
computation (based on the finite buffer), it is still high
compared to the PO-EKF.
TABLE II. NORMALIZED COMPUTATIONAL BURDEN OF FILTERS
EKF
LEKF
PO-EKF
Floating point operations
1.0
36.5
4.7
Execution time
1.0
33.7
2.4
V. EXPERIMENTS
Experiments have been carried out in a real NRS. Details
of the NRS can be referred from our previous work [17].
Two network configurations were employed in experiments.
One is the local configuration in which the robot and the
controller are connected to local Internet service providers.
The other is the VPN configuration in which the robot and
the controller are connected (via VPNs) to servers located at
the United State. The purpose is to capture the low and high
delay of the network. Due to the fact that every attempt to
measure the true trajectory in experiments is influenced by
measurement errors, trajectories estimated by filters are
compared with the trajectory estimated by the EKF with nodelay data.
Fig. 4–5 show the localization results in local and VPN
configurations, respectively. We see that the PO-EKF
introduces better accuracy than the EKF and the same
accuracy as the LEKF at steady state.
VI. CONCLUSION
In this paper, we introduce a new state estimator called
PO-EKF for the localization of NRSs subject to random
delay and packet loss. The optimality of the estimator was
theoretically proven. The good performance in term of
accuracy and computational demand was confirmed through
a number of simulations, comparisons, and experiments.
ACKNOWLEDGMENT
This work was technically supported by the project
CN.12.15 of VNU University of Engineering and
Technology. The travel grant was supported by the
Foundation for Science and Technology Development
(NAFOSTED).
REFERENCES
[1] A. Sanfeliu, N. Hagita, A. Saffiotti, “Network robot systems,”
J. Robotics and Autonomous Systems, vol. 56, issues 10, pp.
793–797, 2008.
[2] P. X. Liu, M. Q. Meng, C. Hu, J. Sheng, “A Modular
Structure for Internet Mobile Robots,” Proc. IEEE/RSJ Int.
Conf. Intelligent Robots and Systems, 2003.
[3] K. Han, S. Kim, Y. Kim, J. Kim, “Internet Control
Architecture for Internet-Based Personal Robot,” J.
Autonomous Robots, vol. 10, issue 2, pp. 135–147, 2001.
[4] H. Hu, L. Yu, P. W. Tsui, Q. Zhou, “Internet-based Robotic
Systems for Teleoperation,” International J. Assembly
Automation, vol. 21, no. 2, pp. 143–152, 2002.
[5] R. Wirz, R. Marína, J. M. Claver, M. Ferre, R. Aracil, J.
Fernández, “End-to-end congestion control protocols for
remote programming of robots using heterogeneous networks:
A comparative analysis,” J. Robotics and Autonomous
Systems, vol. 56, issues 10 , pp. 865–874, 2008.
[6] P. X. Liu, M. Q. Meng, P. R. Liu, S. X. Yang, “An End-toEnd Transmission Architecture for the Remote Control of
Robots Over IP Networks,” IEEE/ASME Trans.
Mechatronics, vol. 10, issues 5, pp. 560–570, 2005.
[7] M. Wargui, A. Tayebi, M. Tadjine and A. Rachid, “On the
stability of an autonomous mobile robot subject to network
induced delay,” Proc. IEEE Int. Conf. Control Applications,
1997.
[8] P. A. Nino-Suarez, E. Aranda-Bricaire and M. Velasco-Villa,
“Discrete-time sliding mode path-tracking control for a
wheeled mobile robot,” Proc. 45th IEEE Conf. Decision &
Control, 2006.
[9] X. Hou, J. Su, “A Distributed Architecture for Internet
Robot,“ Proc. IEEE Int. Conf. Robotics and Automation
(ICRA), 2004.
[10] K. Han, Y. Kim, J. Kim, S. Hsia, “Internet Control of
Personal Robot between KAIST and UC Davis,” Proc. IEEE
Int. Conf. Robotics and Automation (ICRA), 2002.
[11] J. Kim, K. Han, S. Kim, Y. Kim, “Internet-Based Personal
Robot System using Map-Based Localization,” Proc. 32nd Int.
Symp. Robotics (ISR), 2001.
[12] H. Song, L. Yu, W. Zhang, “H∞ filtering of network-based
systems with random delays,” Proc. 17th World Congress,
The International Federation of Automatic Control, 2008.
[13] S. Sun, L. Xie, W. Xiao, N. Xiao, “Optimal Filtering for
Systems With Multiple Packet Dropouts,” IEEE Trans.
Circuits and Systems, vol. 55, no. 7, pp. 2008.
[14] M. Moayedi, Y. K. Foo, Y. C. Soh, “Adaptive Kalman
Filtering in Networked Systems with Random Sensor Delays,
Multiple Packet Dropouts and Missing Measurements,” IEEE
Trans. Signal Processing, vol. 58, no. 3, 2010.
[15] Luca Schenato, “Optimal Estimation in Networked Control
Systems Subject to Random Delay and Packet Drop,” IEEE
Trans. Automatic Control, vol. 53, no. 5, pp. 1311-1317,
2008.
[16] G. Welch, G. Bishop, “An Introduction to the Kalman Filter,”
Proc. SIGGRAPH, 2001.
[17] P. M. Duong, T. T. Hoang, N. T. T. Van, D. A. Viet and T. Q.
Vinh, “A Novel Platform for Internet-based Mobile Robot
Systems, Proc. 7th IEEE Conf. Industrial Electronics &
Applications
(ICIEA),
2012.
| 3cs.SY
|
arXiv:1707.06864v1 [math.GR] 21 Jul 2017
Interval structures for braid groups B(e, e, n)
Georges Neaime
Laboratoire de Mathématiques Nicolas Oresme
Université de Caen Normandie
[email protected]
March 18, 2018
Abstract
Complex braid groups are a generalization of Artin-Tits groups. The general
goal is to extend what is known for Artin-Tits groups to other complex braid
groups. We are interested in Garside structures that derive from intervals. Actually, we construct intervals in the complex reflection group G(e, e, n) which
gives rise to Garside groups. Some of these groups correspond to the complex
braid group B(e, e, n). For the other Garside groups that appear, we give some
of their properties in order to understand these new structures.
Contents
1 Introduction
2
2 The groups G(e, e, n) and B(e, e, n)
6
3 Reduced words in G(e, e, n)
9
4 Balanced elements of maximal length
15
5 Interval structures
20
6 About the interval structures
24
1
1
Introduction
A reflection is an element s of GLn (C) with n ≥ 1 such that ker(s− 1) is a hyperplane
and s2 = 1. Relaxing the last condition to s of finite order defines the notion of pseudoreflection. Let W be a finite subgroup of GLn (C) and R be the set of reflections of
W . We say that W is a complex reflection group if W is generated by R. Since
every complex reflection group is a direct product of irreducible ones, we restrict our
work to irreducible complex reflection groups. These groups have been classified by
Shephard and Todd [17] in 1954. The classification consists of two families and 15
exceptions. For e, n ≥ 1, the first family is denoted by G(e, e, n) and defined as the
group of n × n matrices consisting of
• monomial matrices (each row and column has a unique nonzero entry),
• with all nonzero entries lying in µe , the group of e-th roots of unity, and
• for which the product of the nonzero entries is 1.
The second family is denoted by G(2e, e, n) and defined as the group of n × n matrices
consisting of monomial matrices, with all nonzero entries lying in µ2e , and for which
the product of the nonzero entries is 1 or −1. For the definition of the exceptional
groups, the reader may check [17].
For every complex reflection group W , there exists a corresponding hyperplane
arrangement
and hyperplane complement: A = {Ker(s − 1) | s ∈ R} and X =
S
Cn \ A. The pure braid group is defined as P := π1 (X) and the braid group (or
complex braid group) as B := π1 (X/W ). Note that we have the short exact sequence:
1 −→ P −→ B −→ W −→ 1.
If W = G(de, e, n) with d = 1 or 2, then we denote by B(de, e, n) the associated
braid group. This construction of the braid group is also valid for finite complex
pseudo-reflection groups. However, using the classification of Shephard and Todd for
irreducible complex pseudo-reflection groups and case-by-case results of [3], we may
restrict our work to complex reflection groups as far as group-theoretic properties of
the braid group are concerned.
The previous definitions are a generalization of the well-known Coxeter and ArtinTits groups that we recall now. One way of defining a finite Coxeter group W is by
a presentation with
• a generating set S and
• relations:
– quadratic relations: s2 = 1 for all s ∈ S and
– braid relations: sts
| {z· ·}· for s 6= t ∈ S where mst is the order of
| {z· ·}· = tst
mst
mst
st ∈ W .
2
The Artin-Tits group B(W ) corresponding to W is the group of fractions of the
monoid B + (W ) defined by a presentation with generators: a set Se in bijection with
e
the set S and the relations are only the braid relations: s̃|t̃s̃{z· ·}· = |t̃s̃{z
t̃ · ·}· for s̃ 6= t̃ ∈ S.
mst
mst
The seminal example of these groups is when W = Sn , the symmetric group with
n ≥ 2. It has a presentation with generators s1 , s2 , · · · , sn−1 and relations:
• s2i = 1 for 1 ≤ i ≤ n − 1,
• si si+1 si = si+1 si si+1 for 1 ≤ i ≤ n − 2, and
• si sj = sj si for |i − j| > 1.
The Artin-Tits group associated with Sn is the ‘classical’ braid group denoted by Bn .
The following diagram presentation encodes all the information about the generators
and relations of the presentation of Bn .
s̃1
s̃n−2 s̃n−1
s̃2
Figure 1: Diagram for the presentation of Bn .
The link with the first definitions is as follows. Consider W a real reflection group
meaning that W < GLn (R) < GLn (C). By a theorem of Coxeter, every real reflection group corresponds to a Coxeter group. Furthermore, by a theorem of Brieskorn,
the Artin-Tits group corresponding to W is the braid group π1 (X/W ) attached to W .
It is widely believed that complex braid groups share similar properties with ArtinTits groups. One would like to extend what is known for Artin-Tits groups to other
complex braid groups. For instance, it is shown by Bessis and Corran [1] in 2006,
and by Corran and Picantin [6] in 2009 that the complex braid group B(e, e, n) admits some Garside structures. We are interested in constructing Garside structures
for B(e, e, n) that derive from intervals in the associated complex reflection group
G(e, e, n). For instance, the complex braid group B(e, e, n) admits interval structures
that derive from the non-crossing partitions of type (e, e, n). See [1] for a detailed
study. The aim of this paper is to provide new interval structures for B(e, e, n). In
Section 3, we compute the length of all elements of G(e, e, n) over an appropriate
generating set. This allows us to construct intervals in the complex reflection group
G(e, e, n) and show that they are lattices which automatically gives rise to Garside
structures. This is done in Theorem 5.15. In the last section, we provide group presentations for these Garside structures which we denote by B (k) (e, e, n) for 1 ≤ k ≤ e − 1
and identify which of them correspond to B(e, e, n). For the other Garside structures
that appear, we give some of their properties in order to understand them. One of
the important results obtained is Theorem 6.14:
B (k) (e, e, n) is isomorphic to B(e, e, n) if and only if k ∧ e = 1.
In the remaining part of this section, we include the necessary preliminaries to
accurately describe these Garside structures.
3
1.1
Garside monoids and groups
In his PhD thesis, defended in 1965 [12], and in the article that followed [13], Garside
solved the Conjugacy Problem for the classical braid group Bn by introducing a submonoid Bn+ of Bn and an element ∆n of Bn+ that he called fundamental, and then
showing that there exists a normal form for every element in Bn . In the beginning of
the 1970’s, it was realized by Brieskorn and Saito [2] and Deligne [11] that Garside’s
results extend to all Artin-Tits groups. At the end of the 1990’s, after listing the
abstract properties of Bn+ and the fundamental element ∆n , Dehornoy and Paris [10]
defined the notion of Gaussian groups and Garside groups which leads, in “a natural,
but slowly emerging program” as stated in [8], to Garside theory. For a complete
study about Garside structures that is still far from complete, we refer the reader to
[8].
We start by defining Garside monoids and groups. Let M be a monoid. Under
some assumptions about M , more precisely the assumptions 1 and 2 of Definition 1.2,
one can define a partial order relation on M as follows.
Definition 1.1. Let f, g ∈ M . We say that f left-divides g or simply f divides g when
there is no confusion, written f g, if f g ′ = g holds for some g ′ ∈ M . Similarly, we
say that f right-divides g, written f r g, if g ′ f = g holds for some g ′ ∈ M .
We are ready to define Garside monoids and groups.
Definition 1.2. We say that M is a Garside monoid if
1. M is cancellative, that is f g = f h =⇒ g = h and gf = hf =⇒ g = h for
f, g, h ∈ M ,
2. there exists λ : M −→ N s.t. λ(f g) ≥ λ(f ) + λ(g) and g 6= 1 =⇒ λ(g) 6= 0,
3. any two elements of M have a gcd and an lcm for and r , and
4. there exists an element ∆ ∈ M such that the set of its left divisors coincides
with the set of its right divisors, generate M , and is finite.
The element ∆ is called a Garside element of M and the divisors of ∆ are called
the simples of the Garside structure.
Assumptions 1 and 3 of Definition 1.2 ensure that Ore’s conditions are satisfied.
Hence there exists a group of fractions of the monoid M in which it embeds. This
allows us to give the following definition.
Definition 1.3. A Garside group is the group of fractions of a Garside monoid.
Note that one of the important aspects of a Garside structure is the existence of
a normal form for all elements of the Garside group. Furthermore, many problems
like the Word and Conjugacy problems can be solved in Garside groups which makes
their study interesting.
4
1.2
Interval structures
Let G be a finite group generated by a finite set S. There is a way to construct
Garside structures from intervals in G.
We start by defining a partial order relation on G.
Definition 1.4. Let f, g ∈ G. We say that g is a divisor of f or f is a multiple of
g, and write g f , if f = gh with h ∈ G and ℓ(f ) = ℓ(g) + ℓ(h), where ℓ(f ) is the
length over S of f ∈ G.
Definition 1.5. For w ∈ G, define a monoid M ([1, w]) by the presentation of monoid
with
• generating set P in bijection with the interval
[1, w] := {f ∈ G | 1 f w} and
• relations: f g = h if f, g, h ∈ [1, w], f g = h, and f h, that is ℓ(f ) + ℓ(g) = ℓ(h).
Similarly, one can define the partial order relation on G
g r f if and only if ℓ(f g −1 ) + ℓ(g) = ℓ(f ),
then define the interval [1, w]r and the monoid M ([1, w]r ).
Definition 1.6. Let w be in G. We say that w is a balanced element of G if [1, w] =
[1, w]r .
We have the following theorem (see Section 10 of [15] for a proof).
Theorem 1.7. If w ∈ G is balanced and both posets ([1, w], ) and ([1, w]r , r ) are
lattices, then M ([1, w]) is a Garside monoid with Garside element w and with simples
[1, w].
The previous construction gives rise to an interval structure. The interval monoid
is M ([1, w]). When M ([1, w]) is a Garside monoid, its group of fractions exists and
is denoted by G(M ([1, w])). We call it the interval group. We will give a seminal
example of this structure. It shows that Artin-Tits groups admit interval structures.
Example 1.8. Let W be a finite coxeter group
W =< S | s2 = 1, sts
| {z· ·}· = tst
| {z· ·}· for s 6= t ∈ S > .
mst
mst
Take G = W and g = w0 the longest element over S in W . We have [1, w0 ] = W .
Construct the interval monoid M ([1, w0 ]). We have M ([1, w0 ]) is the Artin-Tits
monoid B + (W ). Hence B + (W ) is generated by a copy W of W with f g = h if
f g = h and ℓ(f ) + ℓ(g) = ℓ(h); f, g, and h ∈ W . We have the following result.
Theorem 1.9. B + (W ) is a Garside monoid with Garside element w0 and with simples W .
5
The groups G(e, e, n) and B(e, e, n)
2
In this section, we recall some presentations by generators and relations of G(e, e, n)
and B(e, e, n) based on the results of [3] and [6].
2.1
Presentations
Recall that for e, n ≥ 1, G(e, e, n) is the group of n × n matrices consisting of monomial matrices, with all nonzero entries lying in µe , the e-th roots of unity, and for
which the product of the nonzero entries is 1. Note that this family includes three
families of finite Coxeter groups: G(1, 1, n) is the symmetric group, G(e, e, 2) is the
dihedral group, and G(2, 2, n) is type Dn .
Define a group by a presentation with generators and relations that can be described by the following diagram.
t1
2
e
t0
s3
s4
sn−1
sn
2
2
2
2
2
Figure 2: Diagram for the presentation of BMR of G(e, e, n).
The generators of this presentation are t0 , t1 , s3 , s4 , · · · , sn−1 , sn and the relations
are:
• quadratic relations for all generators,
• the relations of the symmetric group for s3 , s4 , · · · , sn−1 ,
• the relation of the dihedral group I2 (e) for t0 and t1 : t0 t1 t0 · · · = t1 t0 t1 · · ·,
| {z } | {z }
e
e
• s3 ti s3 = ti s3 ti for i = 0, 1, and
• sj ti = ti sj for i = 0, 1 and 4 ≤ j ≤ n.
It is shown in [3] that this group is isomorphic to the complex reflection
group G(e,
e, n)
I
0
0
0
j−2
0 ζe−i
0
0
0 1
0
0
0 for i = 0, 1 and sj 7−→ sj :=
with ti 7−→ ti := ζei
0
1 0
0
0
0 In−2
0
0 0 In−j
for 3 ≤ j ≤ n, where Ik is the identity matrix with 1 ≤ k ≤ n.
This presentation is called the presentation of BMR (Broué Malle Rouquier) of
G(e, e, n).
6
By [3], removing the quadratic relations gives a presentation of the complex braid
group B(e, e, n) attached to G(e, e, n) with diagram as follows.
t̃1
s̃3
s̃n−1
s̃4
s̃n
e
t̃0
Figure 3: Diagram for the presentation of BMR of B(e, e, n).
Note that the set of generators of this presentation is in bijection with
{t0 , t1 , s3 , s4 , · · · , sn }. This presentation is called the presentation of BMR of B(e, e, n).
2.2
Other presentations
Consider the presentation of BMR for the complex braid group B(e, e, n). For e ≥ 3
and n ≥ 3, it is shown in [5], p.122, that the monoid defined by this presentation
fails to embed in B(e, e, n). Thus, this presentation does not give rise to a Garside
structure for B(e, e, n). Considering the interest of the Garside groups given earlier,
it is legitimate to look for a Garside structure for B(e, e, n).
Let t̃i := t̃i−1 t̃i−2 t̃−1
i−1 for 2 ≤ i ≤ e − 1. Consider the following diagram presentation (the kite).
t̃i
t̃2
t̃1
t̃e−1
t̃0
s̃3
s̃4
s̃n
···
s̃n−1
Figure 4: Diagram for the presentation of CP of B(e, e, n).
The generators of this presentation are t̃0 , t̃1 , · · · , t̃e−1 , s̃3 , s̃4 , · · · , s̃n−1 , s̃n and
the relations are:
• the relations of the symmetric group for s̃3 , s̃4 , · · · , s̃n−1 ,
7
• the relation of the ‘dual’ dihedral group for t̃0 , t̃1 , · · · , t̃e−1 : t̃i t̃i−1 = t̃j t̃j−1 for
i, j ∈ Z/eZ,
• s̃3 t̃i s̃3 = t̃i s̃3 t̃i for 0 ≤ i ≤ e − 1, and
• s̃j t̃i = t̃i s̃j for 0 ≤ i ≤ e − 1 and 4 ≤ j ≤ n.
It is shown in [6] that the group defined by this presentation is isomorphic to
B(e, e, n). We call it the presentation of CP (Corran Picantin) of B(e, e, n). It is also
shown in [6] that:
Proposition 2.1. The presentation of CP gives rise to a Garside structure for
B(e, e, n) with
t se · · · sen and
t1 e
t se · · · sen · · · se3 e
t e
t0 se3 e
Garside element: ∆ = e
t1 e
|
{z0 3
}
| 1{z0 }3
|{z}
∆3
∆2
∆n
simples: the elements of the form δ2 δ3 · · · δn where δi is a divisor of ∆i for
2 ≤ i ≤ n.
It is stated in [6] that if one adds the relations x2 = 1 for all generators x of the
presentation of CP, one obtains a presentation of a group isomorphic to G(e, e, n). It
is called the presentation of CP of G(e, e, n) with diagram as follows.
ti
t2
2
2
t1
te−1
t0
2
2
2
2
s3
2
s4
···
2
sn
2
sn−1
Figure 5: Diagram for the presentation of CP of G(e, e, n).
The generators of this presentation belong to the set X := {t0 , t1 , · · · , te−1 , s3 , · · · , sn }
and the diagram presentation encodes the same relations as the diagram of Figure 2.2
with quadratic relations for all the generators in X.
0 ζe−i
0
0
0
The isomorphism with G(e, e, n) is given by ti 7−→ ti := ζei
0
0
I
n−2
Ij−2 0 0
0
0
0 1
0
for 3 ≤ j ≤ n. Denote
for 0 ≤ i ≤ e − 1, and sj 7−→ sj :=
0
1 0
0
0
0 0 In−j
X the set {t0 , t1 , · · · , te−1 , s3 , · · · , sn }.
8
In the next section, we compute the length over X of each element w in G(e, e, n)
by providing a minimal word representative over X of w.
3
Reduced words in G(e, e, n)
In this section, we represent each element of G(e, e, n) by a reduced word over X.
After some preparations, this is done in Proposition 3.10. Also, we characterize all
the elements of G(e, e, n) that are of maximal length over X.
Recall that an element w ∈ X∗ is called word over X. We denote by ℓ(w) the
length over X of the word w.
Definition 3.1. Let w be an element of G(e, e, n). We define ℓ(w) to be the minimal
word length ℓ(w) of a word w over X that represents w. A reduced expression of w
is any word representative of w of word length ℓ(w).
We introduce an algorithm that produces a word RE(w) over X for a given matrix
w in G(e, e, n). Later on, we prove that RE(w) is a reduced expression over X of w.
Input : w, a matrix in G(e, e, n).
Output: RE(w), a word over X.
Local variables: w′ , RE(w), i, U , c, k.
Initialisation: U := [1, ζe , ζe2 , ..., ζee−1 ], s2 := t0 , s2 := t0 ,
RE(w) = ε: the empty word, w′ := w.
for i from n down to 2 do
c := 1; k := 0;
while w′ [i, c] = 0 do
c := c + 1;
end
Then w′ [i, c] is the root of unity on row i;
while U [k + 1] 6= w′ [i, c] do
k := k + 1;
end
Then w′ [i, c] = ζek .
if k 6= 0 then
w′ := w′ sc sc−1 · · · s3 s2 tk ; Then w′ [i, 2] = 1;
RE(w) := tk s2 s3 · · · sc RE(w);
c := 2;
end
w′ := w′ sc+1 · · · si ; Then w′ [i, i] = 1;
RE(w) := si · · · sc+1 RE(w);
end
Return RE(w);
Algorithm 1: A word over X corresponding to an element w ∈ G(e, e, n).
9
0
0
Example 3.2. We apply Algorithm 1 to w :=
0
1
0
ζ32
Step 1 (i = 4, k = 0, c = 1): w′ := ws2 s3 s4 =
0
0
0 0
0 ζ32
Step 2 (i = 3, k = 1, c = 2): w′ := w′ s2 =
ζ3 0
0 0
0 0 1 0
1 0 0 0
′
′
, then w′ := w′ s3
then w := w t1 =
0 1 0 0
0 0 0 1
Step 3 (i = 2, k = 0, c = 1): w′ := w′ s2 = I4 .
Hence RE(w) = s2 s3 t1 s2 s4 s3 s2 = t0 s3 t1 t0 s4 s3 t0 .
0 0
ζ32 0
0 ζ3
0 0
0
1
0
0
ζ3 0
0
0
1 0
0 0
,
0 0
0 1
0
1
=
0
0
1
0
∈ G(3, 3, 4).
0
0
0
0
.
0
1
1 0
0 0
0 1
0 0
0
0
.
0
1
Let wn := w ∈ G(e, e, n). For i from n to 2,
the i-th step of Algorithm 1
0
wi
into a block diagonal matrix
transforms the block diagonal matrix
0 In−i
wi−1
0
∈ G(e, e, n) with w1 = I1 . Actually, for 2 ≤ i ≤ n, there exists
0
In−i+1
a unique c with 1 ≤ c ≤ n such that wi [i, c] 6= 0. At each step i of Algorithm 1, if
wi [i, c] = 1, we shift it into the diagonal position [i, i] by right multiplication by transpositions of the symmetric group Sn . If wi [i, c] 6= 1, we shift it into the first column
by right multiplication by transpositions, transform it into 1 by right multiplication
by an element of {t0 , t1 , · · · , te−1 }, and then shift the 1 obtained into the diagonal
position [i, i]. The following lemma is straightforward.
Lemma 3.3. For 2 ≤ i ≤ n, the block wi−1 is obtained by
• removing the row i and the column c from wi , then by
• multiplying the first column of the new matrix by wi [i, c].
Example 3.4. Let w be as in Example 3.2. The block
wn−1 isobtained by removing
0 0 1
the row n and the column 1 from wn = w to obtain ζ32 0 0, then by multiplying
0 ζ3 0
the first column of this matrix by 1. Check that the block wn−1 obtained in Example
3.2, after multiplying w by t0 s3 s4 , is the same as the one obtained here. The same
can be said for the other blocks wi with 2 ≤ i ≤ n − 1.
Definition 3.5. At each step i from n to 2,
10
• if wi [i, c] = ζek with k
si · · · s3 tk t0 s3 · · · sc
si · · · s3 tk t0
si · · · s3 tk
6= 0, we define REi (w) to be the word
if c ≥ 3,
if c = 2,
if c = 1, and
• if wi [i, c] = 1, we define REi (w) to be the word
si · · · sc+1 .
Remark that for 2 ≤ i ≤ n, the word REi (w) is either the empty word (when
wi [i, i] = 1) or a word that contains si necessarily but does not contain any of
si+1 , si+2 , · · · , sn .
Lemma 3.6. We have RE(w) = RE2 (w)RE3 (w) · · · REn (w).
Proof. The output RE(w) of Algorithm 1 is a concatenation of the words
RE2 (w), RE3 (w), · · · , and REn (w) obtained at each step i from n to 2 of Algorithm 1.
Example 3.7. If w is defined as in Example 3.2, we have
RE(w) = t0
s t t s s t .
|{z} |3 {z1 }0 |4 {z3 }0
RE2 (w) RE3 (w) RE4 (w)
Proposition 3.8. The word RE(w) given by Algorithm 1 is a word representative
over X of w ∈ G(e, e, n).
Proof. Algorithm 1 transforms the matrix w into In by multiplying it on the right
by elements of X and RE(w) is a concatenation (in reverse order) of elements of X
corresponding to the matrices of X used to transform w into In . Hence RE(w) is a
word representative over X of w ∈ G(e, e, n).
The following proposition will prepare us to prove that the output of Algorithm 1
is a reduced expression over X of a given element w ∈ G(e, e, n).
Proposition 3.9. Let w be an element of G(e, e, n). For all x ∈ X, we have
|ℓ(RE(xw)) − ℓ(RE(w))| = 1.
Proof. For 1 ≤ i ≤ n, there exists unique ci such that w[i, ci ] 6= 0. We denote w[i, ci ]
by ai .
Case 1: Suppose x = si for 3 ≤ i ≤ n.
Since s2i w = w, we can assume without restriction that ci−1 < ci .
Set w′ := si w. Since the left multiplication by the matrix x exchanges the rows i − 1
and i of w and the other rows remain the same, by Definition 3.5 and Lemma 3.3, we
have:
REi+1 (xw)REi+2 (xw) · · · REn (xw) = REi+1 (w)REi+2 (w) · · · REn (w) and
RE2 (xw)RE3 (xw) · · · REi−2 (xw) = RE2 (w)RE3 (w) · · · REi−2 (w).
Then, in order to prove our property, we should compare ℓ1 := ℓ(REi−1 (w)REi (w))
11
and ℓ2 := ℓ(REi−1 (xw)REi (xw)).
Since ci−1 < ci , by Lemma 3.3, the rows i − 1 and i of the blocks wi and wi′ are of
the form:
.. c .. c′ .. i
bi−1
i−1
wi :
ai
i
.. c
wi′ :
i−1
i
.. c′ ..
ai
i
bi−1
with c < c′ and where we write bi−1 instead of ai−1 since ai−1 is likely to change
when applying Algorithm 1 if ci−1 = 1, that is ai−1 on the first column of w.
We will discuss different cases depending on the values of ai and bi−1 .
• Suppose ai = 1.
– If bi−1 = 1,
we have REi (w) = si · · · sc′ +2 sc′ +1 and REi−1 (w) = si−1 · · · sc+2 sc+1 .
Furthermore, we have REi (xw) = si · · · sc+2 sc+1
and REi−1 (xw) = si−1 · · · sc′ +1 sc′ .
It follows that ℓ1 = ((i − 1) − (c + 1) + 1) + (i − (c′ + 1) + 1) = 2i − c − c′ − 1
and ℓ2 = ((i − 1) − c′ + 1) + (i − (c + 1) + 1) = 2i − c − c′ hence ℓ2 = ℓ1 + 1.
– If bi−1 = ζek with 1 ≤ k ≤ e − 1,
we have REi (w) = si · · · sc′ +2 sc′ +1 and REi−1 (w) = si−1 · · · s3 tk t0 s3 sc .
Furthermore, we have REi (xw) = si · · · s3 tk t0 s3 sc and REi−1 (xw) = sc′ · · · si−1 .
It follows that ℓ1 = (((i−1)−3+1)+2+(c−3+1))+(i−(c′ +1)+1) = 2i+
c−c′ −3 and ℓ2 = ((i−1)−c′ +1)+((i−3+1)+2+(c−3+1)) = 2i+c−c′ −2
hence ℓ2 = ℓ1 + 1.
It follows that
if ai = 1, then ℓ(RE(si w)) = ℓ(RE(w)) + 1.
(a)
• Suppose now that ai = ζek with 1 ≤ k ≤ e − 1.
– If bi−1 = 1,
we have REi (w) = si · · · s3 tk t0 s3 sc′ and REi−1 (w) = si−1 · · · sc+1 .
Also, we have REi (xw) = si · · · sc+1 and
REi−1 (xw) = si−1 · · · s3 tk t0 s3 sc′ −1 .
It follows that ℓ1 = ((i − 1) − (c + 1) − 1) + ((i − 3 + 1) + 2 + (c′ − 3 + 1)) =
2i−c+c′ −5 and ℓ2 = (((i−1)−3+1)+2+((c′−1)−3+1))+(i−(c+1)−1) =
2i − c + c′ − 6 hence ℓ2 = ℓ1 − 1.
12
′
– If bi−1 = ζek with 1 ≤ k ′ ≤ e − 1,
we have REi (w) = si · · · s3 tk t0 s3 sc′ and
REi−1 (w) = si−1 · · · s3 tk′ t0 s3 sc .
Also, we have REi (xw) = si · · · s3 tk′ t0 s3 sc and
REi−1 (xw) = si−1 · · · s3 tk t0 s3 sc′ −1 .
It follows that ℓ1 = ((i−1)−3+1)+2+(c−3+1)+(i−3+1)+2+(c′−3+1) =
2i + c + c′ − 5 and ℓ2 = ((i − 1) − 3 + 1) + 2 + ((c′ − 1) − 3 + 1) + (i − 3 +
1) + 2 + (c − 3 + 1) = 2i + c + c′ − 6 hence ℓ2 = ℓ1 − 1.
It follows that
if ai 6= 1, then ℓ(RE(si w)) = ℓ(RE(w)) − 1.
(b)
Case 2: Suppose x = ti for 0 ≤ i ≤ e − 1.
Set w′ := ti w. By definition of the left multiplication by ti , we have that the last
n − 2 rows of w and w′ are the same. Hence, by Definition 3.5 and Lemma 3.3, we
have:
RE3 (xw)RE4 (xw) · · · REn (xw) = RE3 (w)RE4 (w) · · · REn (w). In order to prove our
property in this case, we should compare ℓ1 := ℓ(RE2 (w)) and ℓ2 := ℓ(RE2 (xw)).
• Consider the case where c1 < c2 .
Since c1 < c2 , by Lemma 3.3, the blocks w2 and w2′ are of the form:
0
ζe−i a2
b
0
and w2′ =
w2 = 1
with b1 instead of a1 since a1 is likely
ζei b1
0
0 a2
to change when applying Algorithm 1 if c1 = 1.
– Suppose a2 = 1,
we have b1 = 1 necessarily hence ℓ1 = 0. Since RE2 (xw) = ti , we have
ℓ2 = 1. It follows that when c1 < c2 ,
if a2 = 1, then ℓ(RE(ti w)) = ℓ(RE(w)) + 1.
(c)
– Suppose a2 = ζek with 1 ≤ k ≤ e − 1,
we have RE2 (w) = tk t0 . Thus ℓ1 = 2. We also have ℓ2 = 1 for any b1 . It
follows that when c1 < c2 ,
if a2 6= 1, then ℓ(RE(ti w)) = ℓ(RE(w)) − 1.
(d)
• Now, consider the case where c1 > c2 .
Since c1 > c2 , by Lemma 3.3, the blocks w2 and w2′ are of the form:
−i
0 a1
ζ b
0
w2 =
with b2 instead of a2 since a2 is likely
and w2′ = e 2
b2 0
0
ζei a1
to change when applying Algorithm 1 if c2 = 1.
13
– Suppose a1 6= ζe−i ,
we have ℓ1 = 1 necessarily for any b2 , and since ζei a1 6= 1, we have ℓ2 = 2.
Hence when c1 > c2 ,
if a1 6= ζe−i , then ℓ(RE(ti w)) = ℓ(RE(w)) + 1.
(e)
– Suppose a1 = ζe−i ,
we have ℓ1 = 1 and ℓ2 = 0 for any b2 . Hence when c1 > c2 ,
if a1 = ζe−i , then ℓ(RE(ti w)) = ℓ(RE(w)) − 1.
(f )
This finishes our proof.
Proposition 3.10. Let w be an element of G(e, e, n). The word RE(w) is a reduced
expression over X of w.
Proof. We must prove that ℓ(w) = ℓ(RE(w)).
Let x1 x2 · · · xr be a reduced expression over X of w. Hence ℓ(w) = ℓ(x1 x2 · · · xr ) = r.
Since RE(w) is a word representative over X of w, we have ℓ(RE(w)) ≥ ℓ(x1 x2 · · · xr ) =
r. We prove that ℓ(RE(w)) ≤ r. Observe that we can write w as x1 x2 · · · xr where
x1 , x2 , · · · , xr are the matrices of G(e, e, n) corresponding to x1 , x2 , · · · , xr .
By Proposition 3.9, we have: ℓ(RE(w)) = ℓ(RE(x1 x2 · · · xr )) ≤ ℓ(RE(x2 x3 · · · xr )) +
1 ≤ ℓ(RE(x3 · · · xr )) + 2 ≤ · · · ≤ r. Hence ℓ(RE(w)) = r = ℓ(w) and we are done.
The following proposition is useful in the other sections. It summarizes the proof
of Proposition 3.9.
Proposition 3.11. Let w be an element of G(e, e, n). Denote by ai the unique
nonzero entry w[i, ci ] on the row i of w where 1 ≤ i, ci ≤ n.
1. For 3 ≤ i ≤ n, we have:
(a) if ci−1 < ci , then
ℓ(si w) = ℓ(w) − 1 if and only if ai 6= 1.
(b) if ci−1 > ci , then
ℓ(si w) = ℓ(w) − 1 if and only if ai−1 = 1.
2. If c1 < c2 , then ∀ 0 ≤ k ≤ e − 1, we have
ℓ(tk w) = ℓ(w) − 1 if and only if a2 6= 1.
3. If c1 > c2 , then ∀ 0 ≤ k ≤ e − 1, we have
ℓ(tk w) = ℓ(w) − 1 if and only if a1 = ζe−k .
Proof. The claim 1(a) is deduced from (a) and (b), 2 is deduced from (c) and (d), and
3 is deduced from (e) and (f ) where (a), (b), (c), (d), (e), and (f ) are given in the
proof of Proposition 3.9. Since s2i = 1, 1(b) can be deduced from 1(a).
14
The next proposition is about elements of G(e, e, n) that are of maximal length.
Proposition 3.12. The maximal length of an element of G(e, e, n) is n(n − 1). It
is realized for diagonal matrices w such that w[i, i] is an e-th root of unity different
from 1 for 2 ≤ i ≤ n. A minimal word representative of such an element is then of
the form tk2 t0 s3 tk3 t0 s3 · · · sn · · · s3 tkn t0 s3 · · · sn with 1 ≤ k2 , · · · , kn ≤ (e − 1), and
|
{z
}
| {z } | {z }
the number of elements that are of maximal length is (e − 1)(n−1) .
Proof. By Algorithm 1, an element w in G(e, e, n) is of maximal length when wi [i, i] =
ζek for 2 ≤ i ≤ n and ζek 6= 1. By Lemma 3.3, this condition is satisfied when w is a
diagonal matrix such that w[i, i] is an e-th root of unity different from 1 for 2 ≤ i ≤ n.
A minimal word representative given by Algorithm 1 for such an element is of the
form tk2 t0 s3 tk3 t0 s3 · · · sn · · · s3 tkn t0 s3 · · · sn (1 ≤ k2 , · · · , kn ≤ (e − 1)) which is of
| {z } | {z }
|
{z
}
length 2 + 4 + ... + 2(n − 1) = n(n − 1). The number of elements of this form is
(e − 1)(n−1) .
Example 3.13. Consider λ :=
(ζe−1 )(n−1)
ζe
..
.
∈ G(e, e, n). We have
ζe
RE(λ) = t1 t0 s3 t1 t0 s3 · · · sn · · · s3 t1 t0 s3 · · · sn . Hence ℓ(λ) = n(n − 1) which is the
|
{z
}
|{z} | {z }
maximal length of an element of G(e, e, n).
Note that λ is the image in G(e, e, n) of Λ, the Garside element of the presentation
of Corran and Picantin of the complex braid group B(e, e, n).
Once the length function over X is understood, we can construct intervals in
G(e, e, n) and characterize the balanced elements that are of maximal length.
4
Balanced elements of maximal length
The aim of this section is to prove that the only balanced elements of G(e, e, n) that
are of maximal length are λk with 1 ≤ k ≤ e − 1, where λ was given in Section 3. This
is done by characterizing the intervals of the elements that are of maximal length.
We start by recalling the partial order relations on G(e, e, n).
Definition 4.1. Let w, w′ ∈ G(e, e, n). We say that w′ is a divisor of w or w is
a multiple of w′ , and write w′ w, if w = w′ w′′ with w′′ ∈ G(e, e, n) and ℓ(w) =
ℓ(w′ ) + ℓ(w′′ ). This defines a partial order relation on G(e, e, n).
Similarly, we have another partial order relation on G(e, e, n).
Definition 4.2. Let w, w′ ∈ G(e, e, n). We say that w′ is a right divisor of w or w
is a left multiple of w′ , and write w′ r w, if there exists w′′ ∈ G(e, e, n) such that
w = w′′ w′ and ℓ(w) = ℓ(w′′ ) + ℓ(w′ ).
The following lemma is straightforward.
15
Lemma 4.3. Let w, w′ ∈ G(e, e, n) and let x1 x2 · · · xr be a reduced expression of
w′ over X. We have w′ w if and only if ∀ 1 ≤ i ≤ r, ℓ(xi xi−1 · · · x1 w) =
ℓ(xi−1 · · · x1 w) − 1.
Proof. On the one hand, we have w′ w′′ = w with w′′ = xr xr−1 · · · x1 w and the condition ∀ 1 ≤ i ≤ r, ℓ(xi xi−1 · · · x1 w) = ℓ(xi−1 · · · x1 w)−1 implies that ℓ(w′′ ) = ℓ(w) − r.
So we get ℓ(w′′ ) + ℓ(w′ ) = ℓ(w). Hence w′ w.
On the other hand, since x2 = 1 for all x ∈ X, we have ℓ(xw) = ℓ(w) ± 1 for all
w ∈ G(e, e, n). If there exists i such that ℓ(xi xi−1 · · · x1 w) = ℓ(xi−1 · · · x1 w) + 1
with 1 ≤ i ≤ r, then ℓ(w′′ ) = ℓ(xr xr−1 · · · x1 w) > ℓ(w) − r. It follows that
ℓ(w′ ) + ℓ(w′′ ) > ℓ(w). Hence w′ w.
Consider the homomorphism − : X∗ −→ G(e, e, n) : x 7−→ x := x ∈ X. If
RE(w) = x1 x2 · · · xr with w ∈ G(e, e, n) and x1 , x2 , · · · , xr ∈ X, then RE(w) =
x1 x2 · · · xr = w where x1 , x2 , · · · , xr ∈ X.
Definition 4.4. We define D to be the set
o
n
w ∈ G(e, e, n) s.t. REi (w) REi (λ) f or 2 ≤ i ≤ n .
Proposition 4.5. The set D consists of elements w of G(e, e, n) such that for all
2 ≤ i ≤ n, REi (w) is one of the following words
si · · · s3 t1 t0 s3 · · · si′ with 3 ≤ i′ ≤ i,
si · · · s3 t1 t0 ,
si · · · s3 tk
with 0 ≤ k ≤ e − 1, and
with 3 ≤ i′ ≤ i.
si · · · si′
Proof. We have REi (λ) = si · · · s3 t1 t0 s3 · · · si . Let w ∈ G(e, e, n). Note that REi (w)
is necessarily one of the words given in the the first column of the following table. For each REi (w), there exists unique w′ ∈ G(e, e, n) with RE(w′ ) given in the
second
column,
such that REi (w)w′ = REi(λ). In the last column, we compute
ℓ REi (w) + ℓ(w′ ) that is equal to ℓ REi (λ) = 2(i − 1) only for the first four cases.
The result follows immediately.
REi (w)
si · · · s3 t1 t0 s3 · · · si′ with 3 ≤ i′ ≤ i
si · · · s3 t1 t0
si · · · s3 tk with 0 ≤ k ≤ e − 1
si · · · si′ with 3 ≤ i′ ≤ i
si · · · s3 tk t0 with 2 ≤ k ≤ e − 1
si · · · s3 tk t0 s3 · · · si′ with 2 ≤ k ≤ e − 1,
and 3 ≤ i′ ≤ i
RE(w′ )
si′ +1 · · · si
s3 · · · si
tk−1 s3 · · · si
si′ −1 · · · s3 t1 t0 s3 · · · si
t0 tk−1 s3 · · · si
si′ · · · s3 t0 tk−1 s3 · · · si
The next proposition characterizes the divisors of λ in G(e, e, n).
Proposition 4.6. The set D is equal to the interval [1, λ] where
[1, λ] = {w ∈ G(e, e, n) s.t. 1 w λ}.
16
2(i − 1)
2(i − 1)
2(i − 1)
2(i − 1)
2i
2(i − 1)+
2(i′ − 1)
Proof. Let w ∈ G(e, e, n). We have RE(w) = RE2 (w)RE3 (w) · · · REn (w). Let w ∈ X∗
− ∈ X∗ the word obtained by reading w
be a word representative of w. Denote by ←
w
←−−−−−−
←−−−−−
from right to left. For 3 ≤ i ≤ n, we denote by αi the element REi−1 (w) · · · RE2 (w) in
G(e, e, n).
Suppose that w ∈ D. We apply Lemma 4.3 to prove that w λ. Fix 2 ≤ i ≤ n. By
Proposition 4.5, we have four different possibilities for REi (w).
First, consider the cases REi (w) = si · · · s3 t1 t0 s3 · · · si′ , si · · · s3 t1 t0 , or si · · · si′ with
←−−−−
3 ≤ i′ ≤ i. Hence REi (w) = si′ · · · s3 t0 t1 s3 · · · si , t0 t1 s3 · · · si , or si′ · · · si , respectively.
Note that the left multiplication of the matrix λ by αi produces permutations only
in the block consisting of the first i − 1 rows and the first i − 1 columns of λ. Since
λ[i, i] = ζe (6= 1), by 1(a) of Proposition 3.11, the left multiplication of αi λ by
s3 · · · si decreases the length. Also, by 2 of Proposition 3.11, the left multiplication
of s3 · · · si αi λ by t1 decreases the length. Note that by these left multiplications,
λ[i, i] = ζe is shifted to the first row then transformed to ζe ζe−1 = 1. Hence, by 1(b) of
Proposition 3.11, the left multiplication of t1 s3 · · · si αi λ by si′ · · · s3 t0 decreases the
length, as desired.
←−−−−
Suppose that REi (w) = si · · · s3 tk with 0 ≤ k ≤ e − 1. We have REi (w) = tk s3 · · · si .
Since λ[i, i] = ζe (6= 1), by 1(a) of Proposition 3.11, the left multiplication of αi λ
by s3 · · · si decreases the length. By 2 of Proposition 3.11, the left multiplication of
s3 · · · si αi λ by tk also decreases the length. Hence, applying Lemma 4.3, we have
w λ.
Conversely, suppose that w ∈
/ D, we prove that w λ. If RE(w) = x1 · · · xr ,
by Lemma 4.3, we show that there exists 1 ≤ i ≤ r such that ℓ(xi xi−1 · · · x1 λ) =
ℓ(xi−1 · · · x1 λ) + 1. Since w ∈
/ D, by Proposition 4.5, we may consider the first
REi (w) that appears in RE(w) = RE2 (w) · · · REn (w) such that REi (w) = si · · · s3 tk t0
←−−−−−
or si · · · s3 tk t0 s3 · · · si′ with 2 ≤ k ≤ e − 1 and 3 ≤ i′ ≤ i. Thus, we have RE i (w) =
t0 tk s3 · · · si or si′ · · · s3 t0 tk s3 · · · si , respectively. Since λ[i, i] = ζe (6= 1), by 1(a) of
Proposition 3.11, the left multiplication of αi λ by s3 · · · si decreases the length. By
2 of Proposition 3.11, the left multiplication of s3 · · · si αi λ by tk also decreases the
length. Note that by these left multiplications, λ[i, i] = ζe is shifted to the first row
then transformed to ζe ζe−k = ζe1−k . Since 2 ≤ k ≤ e − 1, we have ζe1−k 6= 1. By 3 of
Proposition 3.11, it follows that the left multiplication of tk s3 · · · si αi λ by t0 increases
the length. Hence w λ.
We want to recognize if an element w ∈ G(e, e, n) is in the set D directly from its
matrix form. For this purpose, we start by describing a matrix form for each element
w ∈ G(e, e, n). Since w is a monomial matrix, there exists nonzero entries that we
refer to as bullets such that the upper left-hand side of w with respect to the bullets,
that we denote by Z(w), have all its entries zero. We denote the other side of the
matrix by Z ′ (w).
17
0
0
ζ32 0
0
0 0 0 0 ζ
3
Example 4.7. Let w =
0 0 ζ3 0 0 ∈ G(3, 3, 5). The bullets are the
1 0 0 0 0
2
0 ζ3 0 0 0
encircled elements and the drawn path separates Z(w) from Z ′ (w).
Remark 4.8. Let w[i, c] be one of the bullets of w ∈ G(e, e, n). We have
w[i − 1, c] ∈ Z(w) and w[i, c − 1] ∈ Z(w).
Also the bullets are the only nonzero entries of w that satisfy this condition.
The following proposition gives a nice description of the divisors of λ in G(e, e, n).
Proposition 4.9. Let w ∈ G(e, e, n). We have that w ∈ D if and only if all nonzero
entries of Z ′ (w) are either 1 or ζe .
Proof. Let w ∈ D and let w[i, c] be a nonzero entry of Z ′ (w). Since w ∈ D, by
Proposition 4.5, we have REi (w) = si · · · s3 t1 t0 , si · · · s3 t1 t0 s3 · · · si′ , or si · · · si′ with
3 ≤ i′ ≤ i. By Lemma 3.3, we have wi [i, i′ ] = w[i, c] for 1 < i′ ≤ i. Hence, by
Definition 3.5, we have w[i, c] = 1 or ζe .
Conversely, let w[i, c] ∈ Z ′ (w). We have w[i, c] = 1 or ζe . Then we have again
REi (w) = si · · · s3 t1 t0 , si · · · s3 t1 t0 s3 · · · si′ , or si · · · si′ for 3 ≤ i′ ≤ i. If w[i, c] is a
bullet of w, by Lemma 3.3, we have wi [i, 1] = ζek for some 0 ≤ k ≤ e − 1, for which
case REi (w) = si · · · s3 tk . Hence, by Proposition 4.5, we have w ∈ D.
ζ32 0 0
0
0 0 ζ 0
3
Example 4.10. Let w =
0 ζ3 0 0
0 0 0 ζ32
it follows immediately that w ∈
/ D.
∈ G(3, 3, 4). Since Z ′ (w) contains ζ32 ,
Our description of the interval [1, λ] allows us to prove easily that λ is balanced.
Let us recall the definition of a balanced element.
Definition 4.11. A balanced element in G(e, e, n) is an element w such that w′ w
holds precisely when w r w′ .
The next lemma is obvious. It will be useful later.
Lemma 4.12. Let g be a balanced element and let w, w′ ∈ [1, g]. If w′ w, then
(w′ )−1 w g.
In order to prove that λ is balanced, we first check the following.
Lemma 4.13. If w ∈ D, we have w−1 λ ∈ D and λw−1 ∈ D.
18
Proof. We show that w−1 λ = tw λ ∈ D and λw−1 = λtw ∈ D, where tw is the complex
conjugate of the transpose tw of the matrix w. We use the matrix form of an element
of D. Actually, by Remark 4.8, the bullets of tw correspond to the complex conjugate
of the bullets of w. Also, all nonzero entries of w−1 , apart from its bullets, are in
{1, ζe−1 }. Multiplying w−1 by λ, we have that all nonzero entries of w−1 λ and of
λw−1 , apart from their bullets, are in {ζe , 1}. Hence w−1 λ ∈ D and λw−1 ∈ D.
Example 4.14. We illustrate the idea of the proof of Lemma
0
0
Consider w ∈ D as follows and show that tw λ ∈ D: w = 0
ζ3
0
0 0 0 ζ32 0
0 0 0 ζ3 0
0
0
0
0
1
0 0 0 0 1
tw λ
tw
2
0 0 ζ3 0 0 −→ 0 0 ζ3 0 0 −→
1 0 0 0 0
1 0 0 0 0
2
0 ζ3 0 0 0
0 ζ3 0 0 0
4.13.
0 0
1
0
0
0
0
ζ3
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
1
ζ32
0
0
0
1
0
0
ζ3
tw
0 −→
0
0
1 0
0 ζ3
0 0 .
0 0
0 0
Proposition 4.15. We have that λ is balanced.
Proof. Suppose that w λ. We have w ∈ D then λw−1 ∈ D by Lemma 4.13. Hence
λ = (λw−1 )w satisfies ℓ(λw−1 ) + ℓ(w) = ℓ(λ), namely w r λ. Conversely, suppose
that w r λ. We have λ = w′ w with w′ ∈ G(e, e, n) and ℓ(w′ ) + ℓ(w) = ℓ(λ). It
−1
−1
follows that w′ ∈ D then w′ λ ∈ D by Lemma 4.13. Since w = w′ λ, we have
w ∈ D, namely w λ.
We state without proof a similar result for the powers of λ in G(e, e, n).
Definition 4.16. For 1 ≤ k ≤ e − 1, let Dk be the set of w ∈ G(e, e, n) such that all
nonzero entries of Z ′ (w) are either 1 or ζek .
Proposition 4.17. For 1 ≤ k ≤ e − 1,
0 0 0
0 0 0
Example 4.18. Let w = 0 0 ζ3
1 0 0
0 ζ32 0
we have [1, λk ] = Dk and λk is balanced.
1 0
0 1
0 0
∈ G(3, 3, 5). We have w ∈ [1, λ2 ].
0 0
0 0
19
More generally, let w ∈ G(e, e, n) be an element of maximal length, namely by
Proposition 3.12, a diagonal matrix such that for 2 ≤ i ≤ n, w[i, i] is an e-th root
of unity different from 1. As previously, we can prove that a divisor w′ of w satisfies
that for all 2 ≤ i ≤ n, if w′ [i, c] 6= 0 is not a bullet of w′ , then w′ [i, c] = 1 or w[i, i].
Suppose that w is of maximal length such that w[i, i] 6= w[j, j] with 2 ≤ i 6= j ≤ n.
We have (i, j) w where (i, j) is the transposition matrix. Hence w′ := (i, j)−1 w =
(i, j)w r w. We have w′ [j, i] = w[i, i]. Thus, w′ [j, i] that is not a bullet of w′ is
different from 1 and w[j, j], since it is equal to w[i, i]. Hence w′ w. We thus get
the following.
Proposition 4.19. The balanced elements of G(e, e, n) that are of maximal length
over X are precisely λk where 1 ≤ k ≤ e − 1.
We are ready to study the interval structures associated with the intervals [1, λk ]
where 1 ≤ k ≤ e − 1.
5
Interval structures
In this section, we construct the monoid M ([1, λk ]) associated with each of the intervals [1, λk ] constructed in Section 4 with 1 ≤ k ≤ e − 1. By Proposition 4.19, λk
is balanced. Hence, by Theorem 1.7, in order to prove that M ([1, λk ]) is a Garside
monoid, it remains to show that both posets ([1, λk ], ) and ([1, λk ], r ) are lattices.
This is stated in Corollary 5.13. The interval structure is given in Theorem 5.15.
Let 1 ≤ k ≤ e − 1 and let w ∈ [1, λk ]. For each 1 ≤ i ≤ n there exists a unique
ci such that w[i, ci ] 6= 0. We denote w[i, ci ] by ai . We apply Lemma 4.3 to prove the
following lemmas. The reader is invited to write the matrix form of w to illustrate
each step of the proof.
Lemma 5.1. Let ti w where i ∈ Z/eZ.
• If c1 < c2 , then tk t0 w and
• if c2 < c1 , then tj w for all j with j 6= i.
Hence if ti w and tj w with i, j ∈ Z/eZ and i 6= j, then tk t0 w.
Proof. Suppose c1 < c2 .
Since w ∈ [1, λk ], by Proposition 4.9, a2 = 1 or ζek . Since ti w, we have ℓ(ti w) =
ℓ(w) − 1. Hence by 2 of Proposition 3.11, we get a2 6= 1. Hence a2 = ζek . Again by 2
of Proposition 3.11, since a2 6= 1, we have ℓ(tk w) = ℓ(w) − 1. Let w′ := tk w. We have
w′ [1, c2 ] = ζe−k a2 = ζe−k ζek = 1. Hence by 3 of Proposition 3.11, ℓ(t0 w′ ) = ℓ(w′ ) − 1.
It follows that tk t0 w.
Suppose c2 < c1 .
Since ti w, we have ℓ(ti w) = ℓ(w) − 1. Hence by 3 of Proposition 3.11, we have
a1 = ζe−i . If there exists j ∈ Z/eZ with j 6= i such that tj w, then ℓ(tj w) = ℓ(w)−1.
Again by 3 of Proposition 3.11, we have a1 = ζe−j . Thus, i = j which is not allowed.
The last statement of the lemma follows immediately.
20
Lemma 5.2. If ti w with i ∈ Z/eZ and s3 w, then s3 ti s3 = ti s3 ti w.
Proof. Set w′ := s3 w and w′′ := ti w.
Suppose c1 < c2 .
Since w ∈ [1, λk ], by Proposition 4.9, we have a2 = 1 or ζek . Since ti w, we have
ℓ(ti w) = ℓ(w) − 1. Thus, by 2 of Proposition 3.11, we get a2 6= 1. Hence a2 = ζek .
Suppose that c3 < c2 . Since s3 w, we have ℓ(s3 w) = ℓ(w) − 1. Hence by 1(b)
of Proposition 3.11, a2 = 1 which is not allowed. Then, assume c2 < c3 . Since
w ∈ [1, λk ], we have a3 = 1 or ζek . By 1(a) of Proposition 3.11, we have a3 6= 1.
Hence a3 = ζek . Now, we prove that s3 ti s3 w by applying Lemma 4.3. Indeed, since
a3 6= 1, by 2 of Proposition 3.11, ℓ(ti w′ ) = ℓ(w′ ) − 1, and since a2 6= 1, by 1(a) of
Proposition 3.11, we have ℓ(s3 w′′ ) = ℓ(w′′ ) − 1.
Suppose c2 < c1 .
Since ℓ(ti w) = ℓ(w) − 1, by 3 of Proposition 3.11, we have a1 = ζe−i .
• Assume c2 < c3 .
Since ℓ(s3 w) = ℓ(w) − 1, by 1(a) of Proposition 3.11, we have a3 6= 1. We have
ℓ(ti w′ ) = ℓ(w′ ) − 1 for both cases c1 < c3 and c3 < c1 . Actually, if c1 < c3 , since
a3 6= 1, by 2 of Proposition 3.11, we have ℓ(ti w′ ) = ℓ(w′ ) − 1, and if c3 < c1 ,
since a1 = ζe−i , by 3 of Proposition 3.11, ℓ(ti w′ ) = ℓ(w′ ) − 1. Now, since
ζei a1 = ζei ζe−i = 1, by 1(b) of Proposition 3.11, we have ℓ(s3 w′′ ) = ℓ(w′′ ) − 1.
• Assume c3 < c2 .
Since a1 = ζe−i , by 3 of Proposition 3.11, ℓ(ti w′ ) = ℓ(w′ ) − 1. Since ζei a1 = 1,
by 1(b) of Proposition 3.11, we have ℓ(s3 w′′ ) = ℓ(w′′ ) − 1.
Lemma 5.3. If ti w with i ∈ Z/eZ and sj w with 4 ≤ j ≤ n, then ti sj = sj ti w.
Proof. We distinguish four different cases: case 1: c1 < c2 and cj−1 < cj , case 2:
c1 < c2 and cj < cj−1 , case 3: c2 < c1 and cj−1 < cj , and case 4: c2 < c1 and
cj < cj−1 . The proof is similar to the proofs of Lemmas 5.1 and 5.2 so we prove that
sj ti w only for the first case. Suppose that c1 < c2 and cj−1 < cj . Since ti w,
we have ℓ(ti w) = ℓ(w) − 1. Hence, by 2 of Proposition 3.11, we have a2 6= 1. Also,
since sj w, we have ℓ(sj w) = ℓ(w) − 1. Hence, by 1(a) of Proposition 3.11, we have
aj 6= 1. Set w′ := sj w. Since a2 6= 1, then ℓ(ti w′ ) = ℓ(w′ ) − 1. Hence sj ti w.
The proof of the following lemma is similar to the proofs of Lemmas 5.2 and 5.3
and is left to the reader.
Lemma 5.4. If si w and si+1 w for 3 ≤ i ≤ n−1, then si si+1 si = si+1 si si+1 w,
and if si w and sj w for 3 ≤ i, j ≤ n and |i − j| > 1, then si sj = sj si w.
The following proposition is a direct consequence of all the preceding lemmas.
Proposition 5.5. Let x, y ∈ X = {t0 , t1 , · · · , te−1 , s3 , · · · , sn }. The least common
multiple in ([1, λk ], ) of x and y, denoted by x∨y, exists and is given by the following
identities:
21
• ti ∨ tj = tk t0 = ti ti−k = tj tj−k for i 6= j ∈ Z/eZ,
• ti ∨ s3 = ti s3 ti = s3 ti s3 for i ∈ Z/eZ,
• ti ∨ sj = ti sj = sj ti for i ∈ Z/eZ and 4 ≤ j ≤ n,
• si ∨ si+1 = si si+1 si = si+1 si si+1 for 3 ≤ i ≤ n − 1, and
• si ∨ sj = si sj = sj si for 3 ≤ i 6= j ≤ n and |i − j| > 1.
We have a similar result for ([1, λk ], r ).
Proposition 5.6. Let x, y ∈ X. The least common multiple in ([1, λk ], r ) of x and
y, denoted by x ∨r y, exists and is equal to x ∨ y.
Proof. Define an antihomomorphism φ : G(e, e, n) −→ G(e, e, n) : ti 7−→ t−i , sj 7−→ sj
with i ∈ Z/eZ and 3 ≤ j ≤ n. It is obvious that φ2 is the identity. Also, if
w ∈ G(e, e, n), we have ℓ(φ(w)) = ℓ(w). Let x, y ∈ X and w ∈ [1, λk ]. We prove that
if x r w and y r w, then x ∨ y r w. Actually, we have w = vx and w = v ′ y
with v, v ′ ∈ G(e, e, n). Thus, φ(w) = φ(x)φ(v) and φ(w) = φ(y)φ(v ′ ). Since φ respects the length over X in G(e, e, n), we have φ(x) φ(w) and φ(y) φ(w). Hence
φ(x) ∨ φ(y) φ(w). Moreover, one can check that φ(x) ∨ φ(y) = φ(x ∨ y) for all
x, y ∈ X. Thus, φ(x ∨ y) φ(w), that is φ(w) = φ(x ∨ y)u for some u ∈ G(e, e, n).
We get w = φ(u)(x ∨ y) and since φ respects the length function, we have x ∨ y r w.
Note that Propositions 5.5 and 5.6 are important to prove that both posets
([1, λk ], ) and ([1, λk ], r ) are lattices. Actually, they will make possible an induction proof of Proposition 5.12 below. For now, let us state some general properties
about lattices that will be useful in our proof. Let (S, ) be a finite poset.
Definition 5.7. Say that (S, ) is a meet-semilattice if and only if f ∧ g := gcd(f, g)
exists for any f, g ∈ S.
V
Equivalently, (S, ) is a meet-semilattice if and only if T exists for any finite
nonempty subset T of S.
Definition 5.8. Say that (S, ) is a join-semilattice if and only if f ∨ g := lcm(f, g)
exists for any f, g ∈ S.
W
Equivalently, (S, ) is a join-semilattice if and only if T exists for any finite
nonempty subset T of S.
Proposition 5.9. (S, ) is a meet-semilattice if and only if for any f, g ∈ S, either
f ∨ g exists, or f and g have no common multiples.
Proof. Let f, g ∈ S and suppose that f and g have at least one common multiple.
Let A := {h ∈ S | f h and g h} be the set of the common multiples
V of f and g.
Since
S
is
finite,
A
is
also
finite.
Since
(S,
)
is
a
meet-semilattice,
A exists and
V
A = lcm(f, g) = f ∨ g.
Conversely, let f, g ∈ S and let B := {h ∈ S | h f and h g} be the set ofWall
common divisors ofWf and g. Since all elements of B have common multiples, B
exists and we have B = gcd(f, g) = f ∧ g.
22
Definition 5.10. The poset (S, ) is a lattice if and only if it is both a meet- and
join- semilattice.
The following lemma is a consequence of Proposition 5.9.
W
Lemma 5.11. If (S, ) is a meet-semilattice such that S exists, then (S, ) is a
lattice.
We will prove that ([1, λk ], ) is a meet-semilattice by applying Proposition 5.9.
On occasion, for 1 ≤ m ≤ ℓ(λk ) with ℓ(λk ) = n(n − 1), we introduce
([1, λk ])m := {w ∈ [1, λk ] s.t. ℓ(w) ≤ m}.
Proposition 5.12. Let 0 ≤ k ≤ e − 1. For 1 ≤ m ≤ n(n − 1) and u, v in ([1, λk ])m ,
either u∨v exists in ([1, λk ])m , or u and v do not have common multiples in ([1, λk ])m .
Proof. Let 1 ≤ m ≤ n(n − 1). We make a proof by induction on m. By Proposition
5.5, our claim holds for m = 1. Suppose m > 1. Assume that the claim holds for
all 1 ≤ m′ ≤ m − 1. We want to prove it for m′ = m. The proof is illustrated in
Figure 6 below. Let u, v be in ([1, λk ])m such that u and v have at least one common
multiple in ([1, λk ])m which we denote by w. Write u = xu1 and v = yv1 such that
x, y ∈ X and ℓ(u) = ℓ(u1 ) + 1, ℓ(v) = ℓ(v1 ) + 1. By Proposition 5.5, x ∨ y exists.
Since x w and y w, x ∨ y divides w. We can write x ∨ y = xy1 = yx1 with
ℓ(x ∨ y) = ℓ(x1 ) + 1 = ℓ(y1 ) + 1. By Lemma 4.12, we have x1 , v1 ∈ [1, λk ]. Also, we
have ℓ(x1 ) < m, ℓ(v1 ) < m and x1 , v1 have a common multiple in ([1, λk ])m−1 . Thus,
by the induction hypothesis, x1 ∨ v1 exists in ([1, λk ])m−1 . Similarly, y1 ∨ u1 exists in
([1, λk ])m−1 . Write x1 ∨v1 = v1 x2 = x1 v2 with ℓ(x1 ∨v1 ) = ℓ(v1 )+ℓ(x2 ) = ℓ(v2 )+ℓ(x1 )
and write y1 ∨ u1 = u1 y2 = y1 u2 with ℓ(y1 ∨ u1 ) = ℓ(y1 ) + ℓ(u2 ) = ℓ(u1 ) + ℓ(y2 ). By
Lemma 4.12, we have u2 , v2 ∈ [1, λk ]. Also, we have ℓ(u2 ) < m, ℓ(v2 ) < m and u2 , v2
have a common multiple in ([1, λk ])m−1 . Thus, by the induction hypothesis, u2 ∨ v2
exists in ([1, λk ])m−1 . Write u2 ∨ v2 = v2 u3 = u2 v3 with ℓ(u2 ∨ v2 ) = ℓ(v2 ) + ℓ(u3 ) =
ℓ(u2 ) + ℓ(v3 ). Since uy2 v3 = vx2 u3 is a common multiple of u and v that divides
every common multiple w of u and v, we deduce that u ∨ v = uy2 v3 = vx2 u3 and we
are done.
y
x
y1
u1
u2
y2
v1
x1
v2
x2
u3
v3
w
Figure 6: The proof of Proposition 5.12.
23
Similarly, applying Proposition 5.6, we obtain the same results for ([1, λk ], r ).
We thus proved the following.
Corollary 5.13. Both posets ([1, λk ], ) and ([1, λk ], r ) are lattices.
Proof. Applying Proposition 5.9, ([1,Wλk ], ) is a meet-semilattice. Also, by definition of the interval [1, λk ], we have [1, λk ] = λk . Thus, applying Proposition 5.9,
([1, λk ], ) is a lattice. The same can be done for ([1, λk ], r ).
We are ready to define the interval monoid M ([1, λk ]).
Definition 5.14. Let Dk be a set in bijection with Dk = [1, λk ] with
[1, λk ] −→ Dk : w 7−→ w.
We define the monoid M ([1, λk ]) by the following presentation of monoid with
• generating set: Dk (a copy of the interval [1, λk ]) and
• relations: w = w′ w′′ whenever w, w′ and w′′ ∈ [1, λk ], w = w′ w′′ and ℓ(w) =
ℓ(w′ ) + ℓ(w′′ ).
We have that λk is balanced. Also, by Corollary 5.13, both posets ([1, λk ], ) and
([1, λk ], r ) are lattices. Hence, by Theorem 1.7, we have:
Theorem 5.15. M ([1, λk ]) is an interval monoid with Garside element λk and with
simples Dk . Its group of fractions exists and is denoted by G(M ([1, λk ])).
These interval structures have been implemented into GAP3, Contribution to the
chevie package (see [16]). The next section is devoted to the study of these interval
structures.
6
About the interval structures
In this section, we provide a new presentation of the interval monoid M ([1, λk ]).
Furthurmore, we prove that G(M ([1, λk ])) is isomorphic to the complex braid group
B(e, e, n) if and only if k ∧ e = 1. When k ∧ e 6= 1, we describe these new structures
and show some of their properties.
6.1
Presentations
Our first aim is to prove that the interval monoid M ([1, λk ]) is isomorphic to the
monoid B ⊕k (e, e, n) defined as follows.
Definition 6.1. For 1 ≤ k ≤ e − 1, we define the monoid B ⊕k (e, e, n) by a presentation of monoid with
e = {t̃0 , t̃1 , · · · , t̃e−1 , s̃3 , · · · , s̃n } and
• generating set: X
24
• relations:
for
for
for
for
for
s̃i s̃j s̃i = s̃j s̃i s̃j
s̃i s̃j = s̃j s̃i
s̃3 t̃i s̃3 = t̃i s̃3 t̃i
s̃j t̃i = t̃i s̃j
t̃i t̃i−k = t̃j t̃j−k
|i − j| = 1,
|i − j| > 1,
i ∈ Z/eZ,
i ∈ Z/eZ and 4 ≤ j ≤ n, and
i, j ∈ Z/eZ.
Note that the monoid B ⊕1 (e, e, n) is the monoid B ⊕ (e, e, n) of Corran and Picantin introduced in Section 2.
The following result is similar to Matsumoto’s property in the case of real reflection
groups.
Proposition 6.2. There exists a map F : [1, λk ] −→ B ⊕k (e, e, n) defined by
F (w) = x̃1 x̃2 · · · x̃r whenever x1 x2 · · · xr is a reduced expression over X of w, where
e for 1 ≤ i ≤ r.
x̃i ∈ X
B
w2 if w2 is obtained from w1 by
Proof. Let w1 and w2 be in X∗ . We write w1
applying only the relations of the presentation of B ⊕k (e, e, n) where we replace t̃i by
ti and s̃j by tj for all i ∈ Z/eZ and 3 ≤ j ≤ n.
Let w be in [1, λk ] and suppose that w1 and w2 are two reduced expressions over X
B
of w. We prove that w1
w2 by induction on ℓ(w1 ).
The result holds vacuously for ℓ(w1 ) = 0 and ℓ(w1 ) = 1. Suppose that ℓ(w1 ) > 1.
Write
w1 = x1 w′ 1 and w2 = x2 w′ 2 , with x1 , x2 ∈ X.
If x1 = x2 , we have x1 w1′ = x2 w2′ in G(e, e, n) from which we get w1′ = w2′ . Then, by
B
B
the induction hypothesis, we have w′ 1
w′ 2 . Hence w1
w2 .
If x1 6= x2 , since x1 w and x2 w, we have x1 ∨ x2 w where x1 ∨ x2 is the lcm
of x1 and x2 in ([1, λk ], ) given in Proposition 5.5. Write w = (x1 ∨ x2 )w′ . Also,
B
write x1 ∨ x2 = x1 v1 and x1 ∨ x2 = x2 v2 where we can check that x1 v1
x2 v2 for
all possible cases for x1 and x2 . All the words x1 w′ 1 , x2 w′ 2 , x1 v1 w′ , and x2 v2 w′
represent w. In particular, x1 w′ 1 and x1 v1 w′ represent w. Hence w1′ = v1 w′ and by
the induction hypothesis, we have w′ 1
B
x1 w′ 1
v1 w′ . Thus, we have
B
x1 v1 w′ .
Similarly, since x2 w′ 2 and x2 v2 w′ represent w, we get
x2 v2 w′
B
Since x1 v1
B
x2 w′ 2 .
x2 v2 , we have
x1 v1 w′
B
x2 v2 w′ .
x1 v1 w′
B
x2 v2 w′
We obtain:
w1
Hence w1
B
B
x1 w′ 1
B
w2 and we are done.
25
B
x2 w′ 2
B
w2 .
By the following proposition, we provide an alternative presentation of the interval
monoid M ([1, λk ]) given in Definition 5.14.
Proposition 6.3. The monoid B ⊕k (e, e, n) is isomorphic to M ([1, λk ]).
Proof. Consider the map ρ : Dk −→ B ⊕k (e, e, n) : w 7−→ F (w) where F is defined
in Proposition 6.2. Let w = w ′ w ′′ be a defining relation of M ([1, λk ]). Since ℓ(w) =
ℓ(w′ ) + ℓ(w′′ ), a reduced expression for w′ w′′ is obtained by concatenating reduced
expressions for w′ and w′′ . It follows that F (w′ w′′ ) = F (w′ )F (w′′ ). We conclude that
ρ has a unique extension to a monoid homomorphism M ([1, λk ]) −→ B ⊕k (e, e, n),
which we denote by the same symbol.
e −→ M ([1, λk ]) : x
Conversely, consider the map ρ′ : X
e 7−→ x. In order to prove
′
that ρ extends to a unique monoid homomorphism B ⊕k (e, e, n) −→ M ([1, λk ]), we
e1 = w
e2 of
have to check that w1 = w2 in M ([1, λk ]) for any defining relation w
B ⊕k (e, e, n). Given a relation w
e1 = w
e2 = x̃1 x̃2 · · · x̃r of B ⊕k (e, e, n), we have w1 =
w2 = x1 x2 · · · xr a reduced word over X. On the other hand, applying repeatedly
the defining relations in M ([1, λk ]) yields to w = x1 x2 · · · xr if w = x1 x2 · · · xr is a
reduced expression over X. Thus, we can conclude that w1 = w2 , as desired.
e −→
Hence we have defined two homomorphisms ρ : Dk −→ B ⊕k (e, e, n) and ρ′ : X
k
′
′
M ([1, λ ]) such that ρ ◦ ρ = idB ⊕k (e,e,n) and ρ ◦ ρ = idM([1,λk ]) . It follows that
B ⊕k (e, e, n) is isomorphic to M ([1, λk ]).
We deduce that B ⊕k (e, e, n) is a Garside monoid and we denote by B (k) (e, e, n)
its group of fractions.
Fix k such that 1 ≤ k ≤ e − 1. A diagram presentation for B (k) (e, e, n) is the same
as the diagram corresponding to the presentation of Corran and Picantin of B(e, e, n)
given in Figure 2.2, with a dashed edge between t̃i and t̃i−k and between t̃j and t̃j−k
for each relation of the form t̃i t̃i−k = t̃j t̃j−k , i, j ∈ Z/eZ. For example, the diagram
corresponding to B (2) (8, 8, 2) is as follows.
3
•
2
•
1
•
0
•
4
•
•
5
•
6
•
7
Figure 7: Diagram for the presentation of B (2) (8, 8, 2).
26
6.2
Identifying B(e, e, n)
Now, we want to check which of the monoids B ⊕k (e, e, n) are isomorphic to B ⊕ (e, e, n).
Assume there exists an isomorphism φ : B ⊕k (e, e, n) −→ B ⊕ (e, e, n) for a given k with
0 ≤ k ≤ e − 1. We start with the following lemma.
Lemma 6.4. The isomorphism φ fixes s̃3 , s̃4 , · · · , s̃n and permutes the t̃i where
i ∈ Z/eZ.
e ∗ . We have ℓ(f ) ≤ ℓ(φ(f )). Thus, we have ℓ(x̃) ≤ ℓ(φ(x̃)) for
Proof. Let f be in X
e Also, ℓ(φ(x̃)) ≤ ℓ(φ−1 (φ(x̃))) = ℓ(x). Hence ℓ(x̃) = ℓ(φ(x̃)) = 1. It follows
x̃ ∈ X.
that φ(x̃) ∈ {t̃0 , t̃1 , · · · , t̃e−1 , s̃3 , · · · , s̃n }.
Furthurmore, the only generator of B ⊕k (e, e, n) that commutes with all other generators except for one of them is s̃n . On the other hand, s̃n is the only generator of
B ⊕ (e, e, n) that satisfies the latter property. Hence φ(s̃n ) = s̃n . Next, s̃n−1 is the
only generator of B ⊕k (e, e, n) that does not commute with s̃n . The only generator
of B ⊕ (e, e, n) that does not commute with s̃n is also s̃n−1 . Hence φ(s̃n−1 ) = s̃n−1 .
Next, the only generator of B ⊕k (e, e, n) different from s̃n and that does not commute
with s̃n−1 is s̃n−2 . And so on, we get φ(s̃j ) = s̃j for 3 ≤ j ≤ n. It remains that
φ({t̃i | 0 ≤ i ≤ e − 1}) = {t̃i | 0 ≤ i ≤ e − 1}.
Proposition 6.5. The monoids B ⊕k (e, e, n) and B ⊕ (e, e, n) are isomorphic if and
only if k ∧ e = 1.
Proof. Assume there exists an isomorphism φ between the monoids B ⊕k (e, e, n) and
B ⊕ (e, e, n). By Lemma 6.4, we have φ(s̃j ) = s̃j for 3 ≤ j ≤ n and φ({t̃i | 0 ≤ i ≤ e − 1})
= {e
ti | 0 ≤ i ≤ e − 1}. A diagram presentation Γk of B ⊕k (e, e, 2) for 1 ≤ k ≤ e − 1
can be viewed as the same diagram presentation of B (k) (e, e, 2) given earlier. The
isomorphism φ implies that we have only one connected component in Γk or, in another words, a closed chain. Hence k is a generator of the cyclic group Z/eZ. Then
k satisfies the condition k ∧ e = 1.
Conversely, let 1 ≤ k ≤ e − 1 such that k ∧ e = 1. We define a map φ : B ⊕ (e, e, n) −→
B ⊕k (e, e, n) where φ(s̃j ) = s̃j for 3 ≤ j ≤ n and φ(t̃0 ) = t̃k , φ(t̃1 ) = t̃2k , φ(t̃2 ) = t̃3k ,
· · · , φ(t̃e−1 ) = e
tek . The map φ is a well-defined monoid homomorphism, which is both
surjective (as it corresponds a generator of B ⊕ (e, e, n) to a generator of B ⊕k (e, e, n))
and injective (as it is bijective on the relations). Hence φ defines an isomorphism of
monoids.
When k ∧ e = 1, since B ⊕k (e, e, n) is isomorphic to B ⊕ (e, e, n), we have the
following.
Corollary 6.6. B (k) (e, e, n) is isomorphic to the complex braid group B(e, e, n) for
k ∧ e = 1.
The reason that the proof of Proposition 6.5 fails in the case k ∧ e 6= 1 is that
we have more than one connected component in Γk that link t̃0 , t̃1 , · · · , and t̃e−1
together, as we can see in Figure 7. Actually, it is easy to check that the number of
connected components that link t̃0 , t̃1 , · · · , and t̃e−1 together is the number of cosets
of the subgroup of Z/eZ generated by the class of k, that is equal to k ∧ e, and each
of these cosets have e′ = e/k ∧ e elements. This will be useful in the next subsection.
27
6.3
New Garside groups
When k ∧ e 6= 1, we describe B (k) (e, e, n) as an amalgamated product of k ∧ e copies
of the complex braid group B(e′ , e′ , n) with e′ = e/e ∧ k, over a common subgroup
which is the Artin-Tits group B(2, 1, n − 1). This allows us to compute the center of
B (k) (e, e, n). Finally, using the Garside structure of B (k) (e, e, n), we compute its first
and second integral homology groups using the Dehornoy-Lafont complex [9] and the
method used in [4].
By an adaptation of the results of Crisp in [7] as in Lemma 5.2 of [4], we have the
following embedding. Let B := B(2, 1, n − 1) be the Artin-Tits group defined by the
following diagram presentation.
q1
q2
qn−2
q3
qn−1
Proposition 6.7. The group B injects in B (k) (e, e, n).
Proof. Define a monoid homomorphism φ : B + −→ B ⊕k (e, e, n) : q1 7−→ t̃i t̃i−k ,
q2 7−→ s̃3 , · · · , qn−1 7−→ s̃n . It is easy to check that for all x, y ∈ {q1 , q2 , · · · , qn−1 }, we
have lcm(φ(x), φ(y)) = φ(lcm(x, y)). Hence by applying Lemma 5.2 of [4], B(2, 1, n − 1)
injects in B (k) (e, e, n).
We construct B (k) (e, e, n) as follows.
Proposition 6.8. Let B(1) := B(e′ , e′ , n) ∗B B(e′ , e′ , n) be the amalgamated product
of two copies of B(e′ , e′ , n) over B = B(2, 1, n − 1) with e′ = e/e ∧ k. Define B(2) :=
B(e′ , e′ , n) ∗B (B(e′ , e′ , n) ∗B B(e′ , e′ , n)) and so on until defining B((e ∧ k) − 1).
We have B((e ∧ k) − 1) = B (k) (e, e, n).
Proof. Due to the presentation of B (k) (e, e, n) given in Definition 6.1 and to the
presentation of the amalgamated products (see Section 4.2 of [14]), one can deduce
that B((e ∧ k) − 1) is equal to B (k) (e, e, n).
28
B((e ∧ k) − 1) = B (k) (e, e, n)
B(e′ , e′ , n)
B(3)
B(2)
B(e′ , e′ , n)
B(e′ , e′ , n)
B(e′ , e′ , n)
Figure 8: The construction of B (k) (e, e, n).
Example 6.9. Consider the case of B (2) (6, 6, 3). It is an amalgamated product of
k ∧ e = 2 copies of B(e′ , e′ , 3) with e′ = e/(k ∧ e) = 3 over the Artin-Tits group
B(2, 1, 2). Consider the following diagram of this amalgamation.
The presentation of B(3, 3, 3) ∗ B(3, 3, 3) over B(2, 1, 2) is as follows:
• the generators are the union of the generators of the two copies of B(3, 3, 3),
• the relations are the union of the relations of the two copies of B(3, 3, 3) with
the additional relations s̃3 = s̃′3 and t̃2 t̃0 = t̃3 t̃1
This is exactly the presentation of B (2) (6, 6, 3) given in Definition 6.1.
s̃3
t̃3 t̃1
∼
s̃′3
t̃2 t̃0
−→
t̃5
t̃4
t̃3
t̃2
t̃0
t̃1
s̃′3
s̃3
B (2) (6, 6, 3)
Proposition 6.10. The center of B (k) (e, e, n) is infinite cyclic isomorphic to Z.
Proof. By Corollary 4.5 of [14] that computes the center of an amalgamated product,
the center of B (k) (e, e, n) is the intersection of the centers of B and B(e′ , e′ , n). Since
the center of B and B(e′ , e′ , n) is infinite cyclic [3], the center of B (k) (e, e, n) is infinite
cyclic isomorphic to Z.
29
Since the center of B(e, e, n) is also isomorphic to Z (see [3]), in order to distinguish these structures from the braid groups B(e, e, n), we compute their first and
second integral homology groups. We recall the Dehornoy-Lafont complex and follow
the method in [4] where the second integral homology group of B(e, e, n) is computed.
e by considering s̃n < s̃n−1 < · · · < s̃3 < t̃0 < t̃1 <
We order the elements of X
e which
· · · < t̃e−1 . For f ∈ B ⊕k (e, e, n), denote by d(f ) the smaller element in X
e
divides f on the right. An r-cell is an r-tuple [x1 , · · · , xr ] of elements in X such that
x1 < x2 < · · · < xr and xi = d(lcm(xi , xi+1 , · · · , xr )). The set Cr of r-chains is
er , the set of all r-cells with the convention
the free ZB ⊕k (e, e, n)-module with basis X
e0 = {[∅]}. We provide the definition of the differential ∂r : Cr −→ Cr−1 .
X
e and A an r-cell. Denote
Definition 6.11. Let [α, A] be an (r + 1)-cell, with α ∈ X
⊕k
α/A the unique element of B (e, e, n) such that (α/A )lcm(A) = lcm(α, A). Define
the differential ∂r : Cr −→ Cr−1 recursively through two Z-module homomorphisms
sr : Cr −→ Cr+1 and ur : Cr −→ Cr as follows.
∂r+1 [α, A] = α/A [A] − ur (α/A [A]),
with ur+1 = sr ◦ ∂r+1 where u0 (f [∅]) = [∅], for all f ∈ B ⊕k (e, e, n), and
sr ([∅]) = 0, sr (x[A]) = 0 if α := d(xlcm(A)) coincides with the first coefficient in A,
and otherwise sr (x[A]) = y[α, A] + sr (yur (α/A [A])) with x = yα/A .
We provide the final result of the computation of ∂1 , ∂2 , and ∂3 for all 1, 2, and
e we have
3-cells, respectively. For all x ∈ X,
∂1 [x] = (x − 1)[∅],
for all 1 ≤ i ≤ e − 1,
∂2 [t̃0 , t̃i ] = t̃i+k [t̃i ] − t̃k [t̃0 ] − [t̃k ] + [t̃i+k ],
e with xyx = yxy,
for x, y ∈ X
∂2 [x, y] = (yx + 1 − x)[y] + (y − xy − 1)[x], and
e with xy = yx,
for x, y ∈ X
∂2 [x, y] = (x − 1)[y] − (y − 1)[x].
For j 6= −k mod e, we have:
∂3 [s̃3 , t̃0 , t̃j ] = (s̃3 t̃k t̃0 s̃3 − t̃k t̃0 s̃3 + t̃j+2k s̃3 )[t̃0 , t̃j ] − t̃j+2k s̃3 t̃j+k [s̃3 , t̃j ] + (t̃j+2k −
s̃3 t̃j+2k )[s̃3 , t̃j+k ] + (s̃3 − t̃j+2k s̃3 − 1)[t̃0 , t̃j+k ] + (s̃3 t̃2k − t̃2k )[s̃3 , t̃k ] + (t̃2k s̃3 + 1 −
s̃3 )[t̃0 , t̃k ] + [s̃3 , t̃j+2k ] + t̃2k s̃3 t̃k [s̃3 , t̃0 ] − [s̃3 , t̃2k ] and
∂3 [s̃3 , t̃0 , t̃−k ] = (s̃3 t̃k t̃0 s̃3 − t̃k t̃0 s̃3 + t̃k s̃3 )[t̃0 , t̃−k ] − t̃k s̃3 t̃0 [s̃3 , t̃−k ] + (1 − t̃2k +
s̃3 t̃2k )[s̃3 , t̃k ] + (1 + t̃2k s̃3 − s̃3 )[t̃0 , t̃k ] + (t̃k − s̃3 t̃k + t̃2k s̃3 t̃k )[s̃3 , t̃0 ] − [s̃3 , t̃2k ].
Also, for 1 ≤ i ≤ e − 1 and 4 ≤ j ≤ n, we have:
30
∂3 [s̃j , t̃0 , t̃i ] = (s̃j − 1)[t̃0 , t̃i ] − t̃i+k [s̃j , t̃i ] + t̃k [s̃j , t̃0 ] − [s̃j , t̃i+k ] + [s̃j , t̃k ],
e with xyx = yxy, xz = zx, and yzy = zyz,
for x, y, z ∈ X
∂3 [x, y, z] =
(z + xyz − yz − 1)[x, y] − [x, z] + (xz − z − x + 1 − yxz)y[x, z] + (x − 1 − yx + zyx)[y, z],
e with xyx = yxy, xz = zx, and yz = zy,
for x, y, z ∈ X
∂3 [x, y, z] = (1 − x + yx)[y, z] + (y − 1 − xy)[x, z] + (z − 1)[x, y],
e with xy = yx, xz = zx, and yzy = zyz,
for x, y, z ∈ X
∂3 [x, y, z] = (1 + yz − z)[x, y] + (y − 1 − zy)[x, z] + (x − 1)[y, z], and
e with xy = yx, xz = zx, and yz = zy,
for x, y, z ∈ X
∂3 [x, y, z] = (1 − y)[x, z] + (z − 1)[x, y] + (x − 1)[y, z].
Let dr = ∂r ⊗ZB ⊕k (e,e,n) Z : Cr ⊗ZB ⊕k (e,e,n) Z −→ Cr−1 ⊗ZB ⊕k (e,e,n) Z be the
differential with trivial coefficients. For example, for d2 , we have: for all 1 ≤ i ≤ e − 1,
d2 [t̃0 , t̃i ] = [t̃i ] − [t̃0 ] − [t̃k ] + [t̃i+k ],
e with xyx = yxy,
for x, y ∈ X
e with xy = yx,
for x, y ∈ X
d2 [x, y] = [y] − [x], and
d2 [x, y] = 0.
The same can be done for ∂3 .
Definition 6.12. Define the integral homology group of order r to be
Hr (B (k) (e, e, n), Z) = ker(dr )/Im(dr+1 ).
We are ready to compute the first and second integral homology groups of B (k) (e, e, n).
Using the presentation of B (k) (e, e, n) given in 6.1, one can check that the abelianization of B (k) (e, e, n) is isomorphic to Z. Since H1 (B (k) (e, e, n), Z) is isomorphic to the
abalianization of B (k) (e, e, n), we deduce that H1 (B (k) (e, e, n), Z) is isomorphic to Z.
Since H1 (B(e, e, n), Z) is also isomorphic to Z (see [3]), the first integral homology
group does not give any additional information weither these groups are isomorphic
to some B(e, e, n) or not.
Recall that by [4], we have
• H2 (B(e, e, 3), Z) ≃ Z/eZ where e ≥ 2,
• H2 (B(e, e, 4), Z) ≃ Z/eZ × Z/2Z when e is odd and H2 (B(e, e, 4), Z) ≃ Z/eZ ×
(Z/2Z)2 when e is even, and
• H2 (B(e, e, n), Z) ≃ Z/eZ × Z/2Z when n ≥ 5 and e ≥ 2.
31
In order to compute H2 (B (k) (e, e, n), Z), we follow exactly the proof of Theorem 6.4
in [4]. Using the same notations as in [4], we set
vi = [t̃0 , t̃i ] + [s̃3 , t̃0 ] + [s̃3 , t̃k ] − [s̃3 , t̃i ] − [s̃3 , t̃i+k ] where 1 ≤ i ≤ e − 1,
and we also have
H2 (B (k) (e, e, n), Z) = (K1 /d3 (C1 )) ⊕ (K2 /d3 (C2 )).
We have d3 [s̃3 , t̃0 , t̃j ] = vj − vj+k + vk if j 6= −k, and d3 [s̃3 , t̃0 , t̃−k ] = v−k + vk . Denote
ui = [s̃3 , t̃0 , t̃i ] for 1 ≤ i ≤ e − 1. We define a basis of C2 as follows. For each coset of
the subgroup of Z/eZ generated by the class of k, say {t̃x , t̃x+k , · · · , t̃x−k } such that
1 ≤ x ≤ e − 1, we define wx+ik = ux+ik + ux+(i+1)k + · · · + ux−k for 0 ≤ i ≤ e − 1,
and when x = 0, we define wik = uik + u(i+1)k + · · · + u−k for 1 ≤ i ≤ e − 1. In
the Z-basis (wk , w2k , · · · , w−k , w1 , w1+k , · · · , w1−k , · · · , wx , wx+k , · · · , wx−k , · · · ) and
(vk , v2k , · · · , v−k , v1 , v1+k , · · · , v1−k , · · · , vx , vx+k , · · · , vx−k , · · · ), d3 is in triangular
form with (e ∧ k) − 1 diagonal coefficients that are zero, all other diagonal coefficients are equal to 1 except one of them that is equal to e′ = e/(e ∧ k). In this case,
we have H2 (B (k) (e, e, 3), Z) = Z(e∧k)−1 × Z/e′ Z. The rest of the proof is essentially
similar to the proof of Theorem 6.4 in [4]. When n = 4, we get [s̃4 , t̃i ] ≡ [s̃4 , t̃i+2k ]
for every i and since 2[s̃4 , t̃i ] ≡ 0 for every i, we get K2 /d3 (C2 ) ≃ (Z/2Z)c , where c is
the number of cosets of the subgroup of Z/eZ generated by the class of 2k.
In the following proposition, we provide the second integral homology group of
B (k) (e, e, n).
Proposition 6.13. Let n ≥ 3 and e ≥ 2.
• When n = 3, we have H2 (B (k) (e, e, 3), Z) ≃ Z(e∧k)−1 × Z/e′ Z,
• when n = 4, H2 (B (k) (e, e, 4), Z) ≃ Z(e∧k)−1 × Z/e′ Z × (Z/2Z)c , where c is the
number of cosets of the subgroup of Z/eZ generated by the class of 2k, and
• when n ≥ 5, H2 (B (k) (e, e, n), Z) ≃ Z(e∧k)−1 × Z/e′ Z × Z/2Z.
Comparing this result with H2 (B(e, e, n), Z), one can check that if k ∧ e 6= 1,
B (k) (e, e, n) is not isomorphic to a complex braid group of type B(d, d, n) with d ≥ 2.
Thus, we conclude by the following theorem.
Theorem 6.14. B (k) (e, e, n) is isomorphic to B(e, e, n) if and only if k ∧ e = 1.
Acknowledgements
I would like to thank my PhD supervisors, Eddy Godelle and Ivan Marin for fruitful
discussions. Also, I would like to thank all the members of Laboratoire de Mathématiques Nicols Oresme and Laboratoire Amiénois de Mathémathiques Fondamentales
et Appliquées for providing me the right environment to finish this work.
32
References
[1] D. Bessis and R. Corran. Non-crossing partitions of type (e, e, r). Adv. Math.,
202:1–49, 2006.
[2] E. Brieskorn and K. Saito. Artin-gruppen und coxeter-gruppen. Invent. Math.,
17:245–271, 1972.
[3] M. Broué, G. Malle, and R. Rouquier. Complex reflection groups, braid groups,
hecke algebras. J. Reine Angew. Math., 500:127–190, 1998.
[4] F. Callegaro and I. Marin. Homology computations for complex braid groups. J.
European Math. Soc., (16):103–164, 2014.
[5] R. Corran. On monoids related to braid groups. PhD thesis, University of Sydney,
May 2000.
[6] R. Corran and M. Picantin. A new garside structure for braid groups of type
(e, e, r). J. London Math. Soc., 84(3):689–711, 2011.
[7] J. Crisp. Injective maps between artin groups. “Geometric Group Theory Down
Under” ed. J. Cossey et al., de Gruyter Verlag, pages 119–137, 1999.
[8] P. Dehornoy, F. Digne, E. Godelle, D. Krammer, and J. Michel. Foundations of
Garside Theory, volume 22. EMS Tracts in Mathematics, 2015.
[9] P. Dehornoy and Y. Lafont. Homology of gaussian groups. Ann. Inst. Fourier,
53(2):1001–1052, 2003.
[10] P. Dehornoy and L. Paris. Gaussian groups and garside groups, two generalizations of artin-groups. Proc. London Math. Soc., 79(3):569–604, 1999.
[11] P. Deligne. Les immeubles des groupes de tresses généralisés. Invent. Math.,
17:273–302, 1972.
[12] F.A. Garside. The theory of knots and associated problems. PhD thesis, Oxford
University, 1965.
[13] F.A. Garside. The braid group and other groups. Quart. J. Math. Oxford,
20:235–254, 1969.
[14] A. Karrass, W. Magnus, and D. Solitar. Combinatorial Group Theory. Reprint
of the Interscience Publishers, New York, 1976.
[15] J. Michel. Groupes finis de réflexion. University notes, Université Paris Diderot,
2004. http://webusers.imj-prg.fr/ jean.michel/papiers/cours2004.pdf.
[16] J. Michel and G. Neaime. Contribution to the chevie package, cp.g.
http://webusers.imj-prg.fr/ jean.michel/gap3.
see
[17] G.C. Shephard and J.A. Todd. Finite unitary reflection groups. Canad. J. Math.,
6(2):274–304, 1954.
33
| 4math.GR
|
EQUIVARIANT COMPRESSION OF CERTAIN DIRECT
LIMIT GROUPS AND AMALGAMATED FREE PRODUCTS
arXiv:1309.4636v3 [math.GR] 7 Feb 2016
CHRIS CAVE AND DENNIS DREESEN
Abstract. We give a means of estimating the equivariant compression of a group G in terms of properties of open subgroups Gi ⊂ G
whose direct limit is G. Quantifying a result by Gal, we also study
the behaviour of the equivariant compression under amalgamated free
products G1 ∗H G2 where H is of finite index in both G1 and G2 .
1. Introduction
The Haagerup property, which is a strong converse of Kazhdan’s property
(T), has translations and applications in various fields of mathematics such
as representation theory, harmonic analysis, operator K-theory and so on.
It implies the Baum–Connes conjecture and related Novikov conjecture [7].
We use the following definition of the Haagerup property.
Definition 1.1. A locally compact second countable group G is said to
satisfy the Haagerup property if it admits a continuous proper affine
isometric action α on some Hilbert space H. Here, proper means that for
every M > 0, there exists a compact set K ⊂ G such that kα(g)(0)k ≥ M
whenever g ∈ G \ K. We say that the action is continuous if the associated
map G × H → H, (g, v) 7→ α(g)(v) is jointly continuous.
Convention 1.2. Throughout this paper, all actions are assumed continuous and all groups will be second countable and locally compact.
Recall that any affine isometric action α can be written as π + b where π
is a unitary representation of G and where b : G → H, g 7→ α(g)(0) satisfies
(1)
∀g, h ∈ G : b(gh) = π(g)b(h) + b(g).
In other words, b is a 1-cocycle associated to π.
In [13], the authors define compression as a means to quantify how strongly
a finitely generated group satisfies the Haagerup property. More generally,
assume that G is a compactly generated group. Denote by S some compact
generating subset and equip G with the word length metric relative to S.
Using the triangle inequality, one checks easily that any 1-cocycle b associated to a unitary action of G on a Hilbert space is Lipschitz. On the other
MSC2010: 20F65 (geometric group theory), 22D10 (Unitary representations of locally
compact groups)
The first author is sponsored by the EPSRC, grant number EP/I016945/1. The second
author is a Marie Curie Intra-European Fellow within the 7th European Community
Framework Programme.
1
2
CHRIS CAVE AND DENNIS DREESEN
hand, one can look for the supremum of r ∈ [0, 1] such that there exists
C, D > 0 with
∀g ∈ G :
1 r
|g| − D ≤ kb(g)k ≤ C|g| + D.
C
Definition 1.3. The above supremum, denoted R(b), is called the compression of b and taking the supremum over all proper affine isometric actions
of G on all Hilbert spaces leads to the equivariant Hilbert space compression of G, denoted α#
2 (G). Suppose now that G is no longer compactly
generated but still has a proper length function. Then define α#
2 (G) to be
the supremum of R(b) but over all large-scale Lipschitz 1-cocycles.
The equivariant Hilbert space compression contains information on the
group. First of all, if α#
2 (G) > 0, then G is Haagerup. The converse
was disproved by T. Austin in [4], where the author proves the existence of
finitely generated amenable groups with equivariant compression 0. Further,
it was shown in [13] that if for a finitely generated group α#
2 (G) > 1/2, then
G is amenable. This result was generalized to compactly generated groups
in [9] and it provides some sort of converse for the well-known fact that
amenability implies the Haagerup property. Much effort has been done to
calculate the explicit equivariant compression value of several groups and
classes of groups, see e.g. [19], [2], [12], [20], [5].
L
Given two finitely generated group G and
L H the group H G is no longer
finitelyL
generated. However we can view H G as a subspace of G≀H and so
equip H G with a natural proper
L metric. In this article we are motivated
by comparing the compression of H G with G≀H. We assume that a given
group G, equipped with a proper length function l, can be viewed as a direct
limit of open (hence closed) subgroups G1 ⊂ G2 ⊂ G3 ⊂ . . . ⊂ G. We equip
each Gi with the subspace metric from G. Our main objective will be to
find bounds on α#
2 (G) in terms of properties of the Gi . Note that, as each
#
Gi is a metric subspace of G, we have α#
2 (G) ≤ inf i∈N α2 (Gi ). The main
challenge is to find a sensible lower bound on α#
2 (G). The key property
that we introduce is the (α, l, q) polynomial property, which we shorten to
(α, l, q)-PP (see Definition 2.5 below). Precisely, we obtain the following
result.
Theorem 1.4. Let G be a locally compact, second countable group equipped
with a proper length function l. Suppose there exists a sequence of open
subgroups (Gi )i∈N , each equipped with the restriction of l to Gi , such that
lim Gi = G and α = inf{α#
2 (Gi )} > 0. If (Gi )i∈N has (α, l, q)-PP, then there
−→
are the following two cases:
l ≥ q ⇒ α#
2 (G) ≥
α
2l + 1
or,
l ≤ q ⇒ α#
2 (G) ≥
α
.
l+q+1
EQUIVARIANT COMPRESSION THROUGH SUBGROUPS
3
We use this result to obtain a lower bound of the compression of the
following examples. Let F : [0, 1] × R≥0 → R be the function
(
d(2α − 1) if 2α ≥ 1
F (α, d) =
0
otherwise.
Theorem 1.5. Let G and H be finitely generated groups where H has polynomial growth of degree d ≥ 1. Then
!
M
α#
2 (G)
α#
G
≥
2
#
1 + F (α2 (G), d) + 2α#
2 (G)(1 + d)
H
L
where H G is equipped with the subspace metric from G ≀ H.
L
Our result also allows to consider spaces H Gh where Gh actually depends on the parameter h ∈ H. For example,
L we take a collection of finite
groups Fi with F0 = {0} and look at G = i∈N Fi . This is the first available
lower bound for the equivariant compression of groups of this type.
Theorem
1.6. Let {Fi }i∈N be a collection of finite groups. Equip G =
L
F
with
the length function l(g) = min {n ∈ N : g ∈ ⊕ni=0 Fi }. Then
i
i∈N
α#
2 (G, l) > 1/3.
We give a proof of Theorem 1.4 in Section 2.2 and apply to these concrete examples in Section 2.3. Note that our result can also be viewed as a
study of the behaviour of equivariant compression under direct limits. The
behaviour of the Haagerup property and the equivariant compression under
group constructions has been studied extensively (see e.g. [18], [11], Chapter
6 of [7], [8], [1]).
In Section 3, we quantify part of [12] to study the behaviour of the equivariant compression under certain amalgamated free products G1 ∗H G2
where H is of finite index in both G1 and G2 . Suppose H is a closed finite index subgroup inside groups compactly generated groups G1 and G2
and there exists proper affine isometric actions βi : Gi → Aff(Vi ) on Hilbert
spaces Vi . In [12], the author shows that if there exists a non-trivial closed
subspace W ⊂ V1 ∩ V2 that is fixed by the restricted actions βi |H then the
product G1 ∗H G2 also admits a proper affine isometric action on a Hilbert
space. We quantify this result.
Theorem 1.7. With the above assumptions α#
2 (G1 ∗H G2 ) ≥
α#
2 (H)
2
2. The equivariant compression of direct limits of groups
2.1. Preliminaries and formulation of the main result. Suppose G is
a locally compact second countable group equipped with a proper length
function l, i.e. closed l-balls are compact. Assume that there exists a sequence of open subgroups Gi ⊂ G such that lim Gi = G, i.e. G is the direct
−→
limit of the Gi . We equip each Gi with the restriction of l to Gi . It will be
#
our goal to find bounds on α#
2 (G) in terms of the α2 (Gi ). Clearly, as the
Gi are subgroups then an upper bound of the equivariant compression is the
infimum of the equivariant compressions of the Gi . The challenge is to find
a sensible lower bound. The next example will show that it is not enough
to only consider the α#
2 (Gi ).
4
CHRIS CAVE AND DENNIS DREESEN
Example 2.1. Consider the wreath product Z ≀ Z equipped with the standard word metric relative to {(δ1 , 0), (0, 1)}, where δ1 is the characteristic function of {0}. Let Z(Z) = {f : Z → Z : f is has finite support} be
equipped with the subspace metric from Z ≀ Z. Consider the direct limit of
groups
Z ֒→ Z3 ֒→ Z5 · · · ֒→ Z(Z)
where Z2n+1 has the subspace metric from Z(Z) . This metric is quasiisometric to the standard word metric on Z2n+1 and so each term has equivariant compression 1. So Z(Z) is a direct limit of groups with equivariant
compression 1 but by [2] has equivariant compression less than 3/4. On the
other hand the sequence
Z → Z → ··· → Z
is a sequence of groups with equivariant compression 1 and the equivariant
compression of the direct limit is 1.
Given a sequence of 1-cocycles bi of Gi , then in order to predict the
equivariant compression of the direct limit, it will be necessary to incorporate more information on the growth behaviour of the bi than merely the
compression exponent R(bi ). The growth behaviour of 1-cocycles can be
completely caught by so called conditionally negative definite functions on
the group (See Proposition 2.3 and Theorem 2.4 below).
Definition 2.2. A continuous map ψ : G → R+ is called conditionally
negative definite if ψ(g) = ψ(g−1 ) for every g ∈ G and
Pnif for all n ∈
N, ∀g1 , g2 , . . . , gn ∈ G and all a1 , a2 , . . . , an ∈ R with
i=1 ai = 0, we
have
X
ai aj ψ(gi−1 gj ) ≤ 0.
i,j
Proposition 2.3 (Example 13, page 62 of [10]). Let H be a Hilbert space
and b : G → H a 1-cocycle associated to a unitary representation. Then the
map ψ : G → R, g 7→ kb(g)k2 is a conditionally negative definite function on
G.
Theorem 2.4 (Proposition 14, page 63 of [10]). Let ψ : G → R be a conditionally negative definite function on a group G. Then there exists an affine
isometric action α on a Hilbert space H such that the associated 1-cocycle
satisfies ψ(g) = kb(g)k2 .
These two results imply that we can pass between conditionally negative
definite functions and 1-cocycles associated to unitary actions.
Definition 2.5. Let G be a group equipped with a proper length function l
and suppose that (Gi )i∈N is a normalized nested sequence of open subgroups
such that lim Gi = G. Assume that α := inf i∈N α#
2 (Gi ) ∈ (0, 1] and l, q ≥ 0.
−→
The sequence (Gi )i has the (α, l, q)-polynomial property ((α, l, q)-PP) if there
exists:
(1) a sequence (ηi )i ⊂ R+ converging to 0 such that ηi < α for each
i ∈ N,
(2) (Ai , Bi )i∈N ⊂ R>0 × R≥0 ,
EQUIVARIANT COMPRESSION THROUGH SUBGROUPS
5
(3) a sequence of 1-cocycles (bi : Gi → Hi )i∈N , where each bi is associated
to a unitary action πi of Gi on a Hilbert space Hi
such that
1 2α−ηi
|g|
− Bi ≤ kbi (g)k2 ≤ Ai |g|2 + Bi ∀g ∈ Gi , ∀i ∈ N
Ai
and there is C, D > 0 such that Ai ≤ Cil , Bi ≤ Diq for all i ∈ N.
Note that the only real restrictions are the inequalities Ai ≤ Cil , Bi ≤
Diq : we exclude sequences Ai , Bi that grow faster than any polynomial.
The intuition is that equivariant compression is a polynomial property (this
follows immediately from its definition), so that sequences Ai , Bi growing
faster than any polynomial would be too dominant and one would lose all
hope of obtaining a lower bound on α#
2 (G). On the other hand, if the Ai and
Bi grow polynomially, then one can use compression to somehow compensate
for this growth. One then obtains a strictly positive lower bound on α#
2 (G)
which may decrease depending on how big l and q are. We have the following
useful characterisation of (α, l, q)-polynomial property.
Lemma 2.6. Let G be a locally compact second countable group and l is
a proper length metric. Suppose there exists a sequence of open subgroups
(Gi )i∈N such that lim Gi = G. If each Gi are equipped with the restricted
−→
length metric from G then (Gi )i∈N has the (α, l, q)-polynomical property if
and only if there exists C, D > 0 such that for all ε > 0 there exists
(1) a sequence (Ai , Bi )i∈N ⊂ R>0 × R≥0 such that Ai ≤ Cil and Bi ≤
Diq ;
(2) a sequence of 1-cocycles (bi : Gi → Hi )i∈N
such that
1 2α−ε
|g|
− Bi ≤ kbi (g)k2 ≤ Ai |g|2 + Bi ∀g ∈ Gi , ∀i ∈ N
Ai
Proof. The “if” direction is obvious. For the “only if” direction fix ε > 0
and suppose (Gi )i∈N has the (α, l, q)-polynomial property with respect to
sequences (ηi )i∈N and (bi : Gi → Hi )i∈N . Choose N ∈ N large enough so
that ηk < ε for all k ≥ N . Thus bk : Gk → Hk satisfies the above conditions
for all k ≥ N . For k ≤ N we take the restriction of bN to Gk to obtain the
sequence satisfying the above conditions for all k ∈ N.
Proposition 2.7. Let G be a locally compact second countable group and
suppose there exists a sequence of open subgroups (Gi )i∈N such that lim Gi =
−→
G. If α := α#
(G)
>
0
then
(G
)
has
(α,
0,
0)-polynomical
property.
i
i∈N
2
Proof. For all 0 < ε < α there exists a 1-cocycle b such that
1 α−ε
|g|
− B ≤ kb(g)k ∀g ∈ G
A
The restriction of b to each Gi is a 1-cocycle and gives (Gi )i∈N the (α, 0, 0)polynomial property.
Combining this with Theorem 1.4 we have the following consequence
which confirms our intuition.
6
CHRIS CAVE AND DENNIS DREESEN
Corollary 2.8. Let G be a locally compact second countable group with a
proper length function l. If there exists a sequence of open subgroups (Gi )i∈N
such that lim Gi = G then (Gi )i∈N has the (α, l, q)-polynomial property for
−→
some α ∈ (0, 1] and l, q ≥ 0 if and only if α#
2 (G) > 0
2.2.
The proof of Theorem 1.4.
Proof of Theorem 1.4. First we can assume that l is uniformly discrete.
That is there exists a c > 0 such that l(x) > c for all x ∈ G \ {e}. This
is because given a length function l one can define a new length function l′
such that l′ (x) = 1 whenever 0 < l(x) ≤ 1 and l′ (x) = l(x) when l(x) ≥ 1.
Hence l′ will be quasi-isometric to l and so will not change the compression
of G or Gi .
Take sequences (ψi : Gi → R)i∈N , (ηi )i and (A, B) = (Ai , Bi )i∈N ⊂ R>0 ×
≥0
R satisfying the conditions of (α, l, q)-PP (see Definition 2.5). We assume
here, without loss of generality, that the sequences (Ai )i , (Bi )i are nondecreasing.
For each Gi , define a sequence of maps (ϕik : Gi → R)k∈N by
(
−ψi (g)
if g ∈ Gi
exp
k
ϕik (g) =
0
otherwise.
Note that each ϕik is continuous as Gi is open and also closed, being the
complement of ∪g∈G
/ i gGi . By (α, l, q)-PP, for all i, k ∈ N, we have
−Ai |g|2 − Bi
≤ ϕik (g) ∀g ∈ Gi , and
exp
k
−|g|2α−ηi + Ai Bi
i
∀g ∈ G.
ϕk (g) ≤ exp
Ai k
Fix some p > 0, set J(i) = (Ai + Bi )i1+p and define ψ : G → R by
X
ψ(g) =
1 − Φi (g),
i∈N
where Φi (g) :=
ϕiJ(i) (g).
To check that ψ is well defined, choose any g ∈ G
and note that for i > |g|, we have g ∈ Gi and so ϕik (g) ≥ exp( −Ai |g|k
Hence
X
X
−Ai |g|2 − Bi
1 − Φi (g) ≤
1 − exp
(Ai + Bi )i1+p
i>|g|
i>|g|
X
−|g|2
≤
1 − exp
i1+p
i>|g|
≤
X |g|2
X 1
2
=
|g|
i1+p
i1+p
i>|g|
i>|g|
As
ψ(g) =
|g|
X
i=1
1 − Φi (g) +
X
i>|g|
1 − Φi (g),
2 −B
i
).
EQUIVARIANT COMPRESSION THROUGH SUBGROUPS
7
we see that ψ is well defined and that it can be written as a limit of continuous functions converging uniformly over compact sets. Consequently, it
is itself continuous. By Schoenberg’s theorem (see [10, Theorem 5.16]), all
of the maps ϕik are positive definite on Gi and hence on G (see [15, Section
32.43(a)]). In other words,
∀n ∈ N, ∀a1 , a2 , . . . , an ∈ R, ∀g1 , g2 , . . . , gn ∈ G :
n
X
ai aj ϕik (gi−1 gj ) ≥ 0.
i,j=1
Hence, ψ is a conditionally negative definite map. Moreover, using that l is
uniformly discrete, we can find a constant E > 0 such that
X 1
(2)
≤ E|g|2
ψ(g) ≤ |g| + |g|2
i1+p
i>|g|
so the 1-cocycle associated to ψ via Theorem 2.4 is large-scale Lipschitz.
Let us now try to find the compression of this 1-cocycle. Set V I : N → R
to be the function
1
V I(i) = (Ai J(i) ln(2) + Ai Bi ) 2α−ηi
One checks easily that
(3)
|g| ≥ V I(i) ⇒ Φi (g) = ϕiJ(i) (g) ≤
1
.
2
To make the function V I more concrete, let us look at the values of Ai , Bi
and J(i). Recall that by assumption, we have Ai ≤ Cil , Bi ≤ Diq . Hence
for i sufficiently large, we have J(i) ≤ (Cil + Diq )i1+p ≤ F iX where F is
some constant and X = 1 + p + max(l, q). We thus obtain that there is a
constant K > 0 such that for every i sufficiently large (say i > I for some
I ∈ N0 ),
V I(i) ≤ KiY /(2α−ηi ) ,
where
Y = max(X + l, l + q)
= max(1 + p + 2l, 1 + p + l + q)
As the sequence ηi converges to 0, we can choose any δ > 0 and take I > 0
such that in addition ηi < δ for i > I. We then have for all i > I that
V I(i) ≤ KiY /(2α−δ) .
Together with Equation (3), this implies that for i > I,
1
|g| ≥ KiY /(2α−δ) ⇒ Φi (g) = ϕiJ(i) (g) ≤ .
2
For every g ∈ G, set
n
o
c(g)p,δ = sup i ∈ N | KiY /(2α−δ) ≤ |g| .
(4)
We then have for every g ∈ G with |g| large enough, that
8
CHRIS CAVE AND DENNIS DREESEN
c(g)p,δ
ψ(g) ≥
X
1 − ϕiJ(i) (g)
i=1
c(g)p,δ
≥
X
1/2 =
i=I+1
c(g)p,δ − I
2
2α−δ
(2α−δ)/Y − 1, we conclude that R(b) ≥
As c(g)p,δ ≥ ( |g|
K)
2 max(1+p+2l,1+p+l+q) .
As this is true for any small p, δ > 0, we can take the limit for p, δ → 0 to
α
obtain α#
2 (G) ≥ max(1+2l,1+l+q) . Hence, we have the following two cases:
l ≥ q ⇒ α#
2 (G) ≥
or,
l ≤ q ⇒ α#
2 (G) ≥
2.3.
α
1 + 2l
α
.
l+q+1
Examples. Let F : [0, 1] × R≥0 → R be the function
(
d(2α − 1) if 2α ≥ 1
F (α, d) =
0
otherwise.
Theorem 2.9. Let G and H be finitely generated groups where H has polynomial growth of degree d ≥ 1. Then
!
M
α#
#
2 (G)
α2
G ≥
#
1 + F (α2 (G), d) + 2α#
2 (G)(1 + d)
H
L
where H G is equipped with the subspace metric from G ≀ H.
Remark 2.10. Theorem 1.3. from [17] provides a lower bound to the compression of G ≀ H. Under the assumptions in Theorem 2.9, Theorem 1.3. in
#
[17] gives a lower bound α#
2 (G ≀ H) ≥ α1 (G)/2. As this bound is in terms
of L1 -compression, this makes comparison between between the bound in
Theorem 2.9 and [17, Theorem 1.3.] difficult. However it is known that
#
#
α#
2 (G) ≤ α1 (G) ≤ 2α2 (G) for all finitely generated groups G, see the proof
of Theorem 1.1. and Theorem 1.3. in [17] and [18, Lemma 2.3.].
We use this to show that under some circumstances the above lower bound
is larger than the bound provided in [17, Theorem 1.3.]. Suppose that
#
α#
1 (G)/2 < α2 (G). Then there exists a c > 0 such that
n
o
c
α#
2 (G) ≤ min 2(1+d) , 1/2 then by Theorem 2.9
α#
2 (⊕H G) ≥
2α#
2 (G)
α#
1 (G)
> 1 + c. If
α# (G)
α#
2 (G)
> 1
1+c
2
Unfortunately the values of α#
2 are not so well understood and at the time
1
of writing the only know values for α#
2 are 1, 1/2, 0 and 2−21−k for k ∈ N
[2, 18, 4]. In the non-equivariant case any value for compression can be
achieved [3]. It is likely that there exists groups such that α#
2 takes values
EQUIVARIANT COMPRESSION THROUGH SUBGROUPS
9
strictly between 0 and 1/2 in which case our theorem can be applied to
provide larger lower bounds than α#
1 (G)/2.
L
Proof. We consider HL
G to be the group of functions f : H → G that have
finite support. Let f ∈ H G and let Supp(f ) = {h1 , . . . , hn } ⊂ H. Set the
length of f as follows
!
n
X
dH (hσ(i) , hσ(i+1) ) + dH (hσ(n) , 1)
|f |G≀H = inf dH (1, hσ(1) ) +
σ∈Sn
i=1
+
X
|f (h)|G .
h∈H
This is the induced
length metric from G ≀ H and so this is a proper length
L
function on H G. Consider the following group
Gi = {f : H → G : Supp(f ) ⊂ B(1, i)}
and set ni = |B(1, i)|. Each Gi is finitely generated and the restricted wreath
metric to Gi is proper and left invariant so the wreath metric and the word
metric are quasi-isometric. In particular
X
|f |G≀H − 2i|B(1, i)| ≤
|f (h)|G ≤ |f |G≀H
h∈B(1,i)
for all f ∈ Gi . By [14, Proposition 4.1. and Corollary 2.13.] it follows that
#
#
α#
2 (Gi ) = α2 (G) for all i ∈ N. Set 0 < α < α2 (G) and consider a 1-cocyle
b : G → H such that
1 2α
|g| ≤ kb(g)k2 ≤ C|g|2G .
C G
Enumerate B(1, i) so that {h1 , . . . , hni } = B(1, i) and define a 1-cocycle
bi : Gi → Hni , where bi (f ) = (b(f (h1 )), . . . , b(f (hni ))). If |f |G≀H > 4i|B(1, i)|,
then
α
α
i
i
X
X
1
1/α
|f (hnj )|G
kbi (f )k1/α =
b(f (hnj ))
≥ 1/α
C
j=1
j=1
≥
1
C 1/α
(|f |G≀H − 2i|B(1, i)|) α ≥
1
|f |α .
2C 1/α G≀H
If 2α < 1 then kbi (f )k2 ≥ kbi (f )k1/α for all f ∈ Gi and so it follows that
i2α
1
2α
|f
|
−
|B(1, i)|2α ≤ kbi (f )k22
C
4C 2/α G≀H
for all f ∈ Gi . Hence (Gi )i∈N has the (α, 0, 2α(1 + d)) polynomial property.
1−2α
2
If 2α ≥ 1 then by Hölder’s inequality kbi (f )k2 ≥ ni
f ∈ Gi and so it follows that
kbi (f )k1/α for all
i2α
1
2α
−
|B(1, i)|2α ≤ kbi (f )k22 .
|f
|
G≀H
2/α
2α−1
C
4C |B(1, i)|
10
CHRIS CAVE AND DENNIS DREESEN
for all f ∈ Gi . Hence (Gi )i∈N has the (α, d(2α − 1), 2α(1 + d)) polynomial
property. Thus by Theorem 1.4 and that α, d ≥ 0 it follows that
M
α
α#
G) ≥
2 (
1 + F (α, d) + 2α(1 + d)
H
for all α < α#
2 (G) and so the statement of the theorem holds.
Theorem 2.11.L
Let {Fi }i∈N be a collection of finite groups such that F0 =
{1}. Let G =
i∈N Fi be equip with the proper length function l(g) =
n
min {n ∈ N : g ∈ ⊕i=0 Fi }. Then α#
2 (G) ≥ 1/3.
L
Proof. Set Gi = ij=0 Fj and observe that α#
2 (Gi ) = 1 as Gi is finite for all
i ∈ N. Define fi : Gi → R to be the 0-map. This is clearly a 1-cocycle and
satisfies
∀g ∈ Gi : l(g)2 − i2 ≤ |fi (g)|2 ≤ l(g)2 + i2 .
Hence (Gi )i∈N has the (1,0,2)-polynomial property. Thus α#
2 (G) ≥ 1/3.
Example 2.12. We will use [3] to provide an example of a sequence that
does not have (α, l, q)-polynomial property for any α ∈ (0, 1] and l, q > 0.
Let Πk , k ≥ 1 be a sequence of Lafforgue expanders that do not embed into
any uniformly convex Banach space [16]. These are finite factor groups Mk
of a lattice Γ of SL3 (F ) for a local field F .
For every α ∈ [0, 1] there exists a finitely generated group G and a sequence of scaling constants λk such that λk Πk has compression α and G
is quasi-isometric to λk Πk . Furthermore G contains the free product ∗k Mk
as a subgroup. Let α = 0 and let G and the scaling constants λk be such
that G has compression 0. We can equip ∗k Mk with a proper left invariant
metric coming from G. Hence we have a sequence
M1 ֒→ M1 ∗ M2 ֒→ · · · ֒→ ∗nk=1 Mk ֒→ · · · ֒→ ∗k Mk
For each n > 0, ∗nk=1 Mk has equivariant compression 1/2 [11, Theorem 1.4.]
however the limit group ∗k Mk contains a quasi-isometric copy of λk Πk and so
has compression 0. Thus this sequence can not have the (α, l, q)-polynomial
property for any α ∈ (0, 1] and l, q > 0.
3. The behaviour of compression under free products
amalgamated over finite index subgroups
It is known that the Haagerup property is not preserved under amalgamated free products. Indeed, (SL2 (Z) ⋊ Z2 , Z2 ) has the relative property
(T ). So SL2 (Z) ⋊ Z2 = (Z6 ⋊ Z2 ) ∗(Z2 ⋊Z2 ) (Z4 ⋊ Z2 ) is not Haagerup. In
[12], S.R. Gal proves the following result.
Theorem 3.1. Let G1 and G2 be finitely generated groups with the Haagerup
property that have a common finite index subgroup H. For each i = 1, 2,
let βi be a proper affine isometric action of Gi on a Hilbert space Vi (=
l2 (Z)). Assume that W < V1 ∩ V2 is invariant under the actions (βi )|H and
moreover that both these (restricted) actions coincide on W . Then G1 ∗H G2
is Haagerup.
EQUIVARIANT COMPRESSION THROUGH SUBGROUPS
11
Under the same conditions as above, we want to give estimates on α#
2 (G1 ∗H
G2 ) in terms of the equivariant Hilbert space compressions of G1 , G2 (see
Theorem 3.3 below). Note that the following lemma shows that α#
2 (G1 ) =
#
#
α2 (H) = α2 (G2 ) when H is of finite index in both G1 and G2 . We are
indebted to Alain Valette for this lemma and its proof. The notation α#
p
refers to the equivariant Lp -compression for some p ≥ 1. It is defined
in exactly the same way as α#
2 except that one considers affine isometric
actions on Lp -spaces instead of L2 -spaces.
Lemma 3.2. Let G be a compactly generated, locally compact group, and
#
let H be an open, finite-index subgroup of G. Then α#
p (H) = αp (G).
Proof. As H is embedded H-equivariantly, quasi-isometrically in G, we have
#
α#
p (H) ≥ αp (G). To prove the converse inequality, we may assume that
α#
p (H) > 0. Let S be a compact generating subset of H. Let A(h)v =
π(h)v + b(h) be an affine isometric action of H on Lp , such that for some
α
α < α#
p (H) we have kb(h)kp ≥ C|h|S , for every h ∈ H. Now we induce up
the action A from H to G, as on p.91 of [6]1. The affine space of the induced
action is
E := {f : G → Lp : f (gh) = A(h)−1 f (g), ∀h ∈ H and almost every g ∈ G},
P
with distance given by kf1 − f2 kpp = x∈G/H kf1 (x) − f2 (x)kpp . The induced
affine isometric action à of G on E is then given by (Ã(g))f (g′ ) = f (g −1 g′ ),
for f ∈ E, g, g ′ ∈ G.
A function ξ0 ∈ E is then defined as follows. Let s1 = e, s2 , , ..., sn be a
set of representatives for the left cosets of H in G. Set ξ0 (si h) = b(h−1 ), for
h ∈ H, i = 1, ..., n. Define the 1-cocycle b̃ on G by b̃(g) = Ã(g)ξ0 − ξ0 , for
g ∈ G. For an h ∈ H, we then have:
kb̃(h)kpp =
n
X
i=1
kξ0 (h−1 si )−ξ0 (si )kpp =
n
X
kξ0 (h−1 si )kpp ≥ kξ0 (h−1 )kpp = kb(h)kpp .
i=1
Set K = max1≤i≤n kb̃(si )kp . Take T = S ∪ {s1 , ..., sn } as a compact generating set of G. For g ∈ G, write g = si h for 1 ≤ i ≤ n, h ∈ H. Then
kb̃(g)kp ≥ kb̃(h)kp − K ≥ kb(h)kp − K ≥ C|h|αS − K ≥ C|h|αT − K
≥ C(|g|T − 1)α − K ≥ C ′ |g|αT − K ′ .
#
So the compression of the 1-cocycle b̃ is at least α, hence α#
p (G) ≥ αp (H).
The following proof uses a construction by S.R. Gal, see page 4 of [12].
Theorem 3.3. Let V1 and V2 be closed subspaces of a Hilbert space. Suppose
H is a finite index subgroup of G1 and G2 and suppose there are proper affine
isometric actions βi (with compression αi ) of each Gi on Vi . Assume that
W < V1 ∩ V2 is invariant under the actions (βi |H ) and moreover that both
1We seize this opportunity to correct a misprint in the definition of the vector ξ in
0
that construction in p.91 of [6].
12
CHRIS CAVE AND DENNIS DREESEN
these (restricted) actions coincide on W . Then α#
2 (G1 ∗H G2 ) ≥
In particular, α#
2 (G1 ∗H G2 ) ≥
α#
2 (H)
.
2
min(α1 ,α2 )
.
2
Proof. Following [12], let us build a Hilbert space WΓ on which Γ = G1 ∗H G2
acts affinely and isometrically. Let ω be a finite alternating sequence of 1’s
and 2’s and suppose π is a linear action of H on some Hilbert space denoted
Hω . One can induce up the linear action from H to Gi , obtaining a Hilbert
space
V := f : Gi → Hω | ∀h ∈ H, f (gh) = π(h−1 )f (g)
and an orthogonal action πi : Gi → O(V ) defined by πi (g)f (g′ ) = f (g −1 g′ ).
The subspace
f : Gi → Hω | ∀h ∈ H, f (h) = π(h−1 )f (1), f|Gi \H = 0
can be identified with Hω by letting an element f correspond to f (1). It
is clear that the action πi restricted to H coincides with the original linear
action π via this identification.
So, starting from any linear H-action on a Hilbert space Hω , we can
obtain a linear action of say G1 on a Hilbert space that can be written as
Hω ⊕ H1ω for some H1ω . We can restrict this action to a linear H-action
on H1ω and we can lift this to an action of G2 on a space H1ω ⊕ H21ω and
so on, repeating the process indefinitely. Here, we will execute this infinite
process twice.
The first linear H-action on which we apply the process is obtained as
follows. As βi (H)(W ) = W for each i = 1, 2, the restriction to H of β1 , gives
naturally a linear H-action on H1 := V1 /W . The second linear H-action is
obtained by similarly noting that the restriction to H of β2 gives a linear
H-action on H2 := V2 /W . We then apply the above process indefinitely.
G2 y
G2 y
H1•
}|
{
z }| { z
:= H1 ⊕ H21 ⊕ H121 ⊕H2121 ⊕ · · ·,
{z
} |
{z
}
|
G1 y
G1 y
G1 y
H2•
z }| { z
}|
{
:= H2 ⊕ H12 ⊕ H212 ⊕H1212 ⊕ · · ·,
{z
} |
{z
}
|
G2 y
where for ω a sequence of alternating 1’s and 2’s, Gi acts on Hω ⊕ Hiω . Note
that there are two H-actions on H1• as H acts on the first term H1 . One can
verify that both H-actions coincide (this fact is also mentioned in [12],page
4). The same is true for H2• .
Denote H1◦ = H1• ⊖ H1 and similarly, set H2◦ = H2• ⊖ H2 . We denote
WΓ = W ⊕ H1• ⊕ H2• = V1 ⊕ H1◦ ⊕ H2• = V2 ⊕ H2◦ ⊕ H1• .
The above formula, which writes W as a direct sum in three distinct ways,
shows that both G1 and G2 act on WΓ . As mentioned before, the actions
coincide on H and so we obtain an affine isometric action of Γ on WΓ . Note
that the corresponding 1-cocycle, when restricted to G1 (or G2 ), coincides
with the 1-cocycle of β1 (or β2 ).
We inductively define a length function ψT : Γ → N by ψT (h) = 0 for
all h ∈ H and ψT (γ) = min {ψT (η) + 1 | γ = ηg, where g ∈ G1 ∪ G2 }. By
applying Proposition 2 in [10] to the Bass–Serre tree of G1 ∗H G2 , we see that
this map is conditionally negative definite and thus the normed square of a
1-cocycle associated to an affine isometric action of Γ on a Hilbert space.
EQUIVARIANT COMPRESSION THROUGH SUBGROUPS
13
Let ψΓ be the conditionally negative definite function associated to the action of Γ on WΓ . We now find the compression of the conditionally negative
definite map ψ = ψΓ + ψT . First set
M = max |tij |Gi : i = 1, 2 and 1 ≤ j ≤ [Gi : H] ,
where tij are right coset representatives of H in Gi such that ti1 = 1Gi for
i = 1, 2.
Denote α = min(α1 , α2 ) and fix some ε > 0 arbitrarily small. Let γ ∈ Γ
and suppose in normal form γ = gtij11 · · · tijkk , where g ∈ Gi for some i = 1, 2.
Assume first that ψT (γ) ≥
that ψT (γ) <
we have
|γ|α−ε
M
|γ|α−ε
M .
In that case, ψ(γ) ≥
|γ|α−ε
M .
Else, we have
and so for all γ ∈ Γ such that |γ| is sufficiently large,
ψ(γ) ≥ ψΓ (γ) = kγ · 0k2
≥ (kg · 0k − ψT (γ)M )2
& ((|γ| − ψT (γ)M )α−ε/2 − ψT (γ)M )2
≥ ((|γ| − |γ|α−ε )α−ε/2 − |γ|α−ε )2
& |γ|2α−ε ,
where & represents inequality up to a multiplicative constant; we use here
that one can always assume, without loss of generality, that the 1-cocycles
associated to β1 and β2 satisfy kbi (gi )k & |gi |α−ε (see Lemma 3.4 in [1]).
So now, by the first case, ψ(γ) ≥ |γ|α−ε for all γ ∈ Γ that are sufficiently
#
large. Hence, we obtain the lower bound α#
2 (Γ) ≥ α2 (H)/2.
Acknowledgement
The authors would like to thank Jacek Brodzki for interesting discussions
and the referee for their helpful recommendations and for suggesting a simplier example in Example 2.12. The first author thanks Martin Finn-Sell
for interesting conversations on affine subspaces. The second author thanks
Michal Marcinkowski for interesting discussions related to Gal’s paper [12].
References
[1] Y. Antolin and D. Dreesen. The Haagerup property is stable under graph products.
preprint, 2013.
[2] G. N. Arzhantseva, V. S. Guba, and M. V. Sapir. Metrics on diagram groups and
uniform embeddings in a Hilbert space. Comment. Math. Helv., 81(4):911–929, 2006.
[3] Goulnara Arzhantseva, Cornelia Druţu, and Mark Sapir. Compression functions of
uniform embeddings of groups into Hilbert and Banach spaces. J. Reine Angew.
Math., 633:213–235, 2009.
[4] Tim Austin. Amenable groups with very poor compression into Lebesgue spaces.
Duke Math. J., 159(2):187–222, 2011.
[5] Tim Austin, Assaf Naor, and Yuval Peres. The wreath product of Z with Z has Hilbert
compression exponent 23 . Proc. Amer. Math. Soc., 137(1):85–90, 2009.
[6] Bachir Bekka, Pierre de la Harpe, and Alain Valette. Kazhdan’s property (T), volume 11 of New Mathematical Monographs. Cambridge University Press, Cambridge,
2008.
14
CHRIS CAVE AND DENNIS DREESEN
[7] Pierre-Alain Cherix, Michael Cowling, Paul Jolissaint, Pierre Julg, and Alain
Valette. Groups with the Haagerup property, volume 197 of Progress in Mathematics.
Birkhäuser Verlag, Basel, 2001. Gromov’s a-T-menability.
[8] Yves Cornulier, Yves Stalder, and Alain Valette. Proper actions of wreath products
and generalizations. Trans. Amer. Math. Soc., 364(6):3159–3184, 2012.
[9] Yves de Cornulier, Romain Tessera, and Alain Valette. Isometric group actions on
Hilbert spaces: growth of cocycles. Geom. Funct. Anal., 17(3):770–792, 2007.
[10] Pierre de la Harpe and Alain Valette. La propriété (T ) de Kazhdan pour les groupes
localement compacts (avec un appendice de Marc Burger). Number 175. 1989. With
an appendix by M. Burger.
[11] Dennis Dreesen. Hilbert space compression for free products and HNN-extensions. J.
Funct. Anal., 261(12):3585–3611, 2011.
[12] Światoslaw R. Gal. a-T-menability of groups acting on trees. Bull. Austral. Math.
Soc., 69(2):297–303, 2004.
[13] E. Guentner and J. Kaminker. Exactness and the Novikov conjecture. Topology,
41(2):411 – 418, 2002.
[14] Erik Guentner and Jerome Kaminker. Exactness and uniform embeddability of discrete groups. J. London Math. Soc. (2), 70(3):703–718, 2004.
[15] Edwin Hewitt and Kenneth A. Ross. Abstract harmonic analysis. Vol. II: Structure
and analysis for compact groups. Analysis on locally compact Abelian groups. Die
Grundlehren der mathematischen Wissenschaften, Band 152. Springer-Verlag, New
York, 1970.
[16] Vincent Lafforgue. Un renforcement de la propriété (T). Duke Math. J., 143(3):559–
602, 2008.
[17] Sean Li. Compression bounds for wreath products. Proc. Amer. Math. Soc.,
138(8):2701–2714, 2010.
[18] A. Naor and Y. Peres. Embeddings of discrete groups and the speed of random walks.
Int. Math. Res. Not., 2008.
[19] Yves Stalder and Alain Valette. Wreath products with the integers, proper actions
and Hilbert space compression. Geom. Dedicata, 124:199–211, 2007.
[20] Romain Tessera. Asymptotic isoperimetry on groups and uniform embeddings into
Banach spaces. Comment. Math. Helv., 86(3):499–535, 2011.
School of Mathematics, University of Southampton, Highfield, Southampton, SO17 1BJ, United Kingdom.
E-mail address, Chris Cave: [email protected]
School of Mathematics, University of Southampton, Highfield, Southampton, SO17 1BJ, United Kingdom.
E-mail address, Dennis Dreesen: [email protected]
| 4math.GR
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.